kernel/sysctl.c

   1 /*
   2  * sysctl.c: General linux system control interface
   3  *
   4  * Begun 24 March 1995, Stephen Tweedie
   5  * Added /proc support, Dec 1995
   6  * Added bdflush entry and intvec min/max checking, 2/23/96, Tom Dyas.
   7  * Added hooks for /proc/sys/net (minor, minor patch), 96/4/1, Mike Shaver.
   8  * Added kernel/java-{interpreter,appletviewer}, 96/5/10, Mike Shaver.
   9  * Dynamic registration fixes, Stephen Tweedie.
  10  * Added kswapd-interval, ctrl-alt-del, printk stuff, 1/8/97, Chris Horn.
  11  * Made sysctl support optional via CONFIG_SYSCTL, 1/10/97, Chris
  12  *  Horn.
  13  * Added proc_doulongvec_ms_jiffies_minmax, 09/08/99, Carlos H. Bauer.
  14  * Added proc_doulongvec_minmax, 09/08/99, Carlos H. Bauer.
  15  * Changed linked lists to use list.h instead of lists.h, 02/24/00, Bill
  16  *  Wendling.
  17  * The list_for_each() macro wasn't appropriate for the sysctl loop.
  18  *  Removed it and replaced it with older style, 03/23/00, Bill Wendling
  19  */
  20
  21 #include <linux/module.h>
  22 #include <linux/aio.h>
  23 #include <linux/mm.h>
  24 #include <linux/swap.h>
  25 #include <linux/slab.h>
  26 #include <linux/sysctl.h>
  27 #include <linux/bitmap.h>
  28 #include <linux/signal.h>
  29 #include <linux/printk.h>
  30 #include <linux/proc_fs.h>
  31 #include <linux/security.h>
  32 #include <linux/ctype.h>
  33 #include <linux/kmemcheck.h>
  34 #include <linux/kmemleak.h>
  35 #include <linux/fs.h>
  36 #include <linux/init.h>
  37 #include <linux/kernel.h>
  38 #include <linux/kobject.h>
  39 #include <linux/net.h>
  40 #include <linux/sysrq.h>
  41 #include <linux/highuid.h>
  42 #include <linux/writeback.h>
  43 #include <linux/ratelimit.h>
  44 #include <linux/compaction.h>
  45 #include <linux/hugetlb.h>
  46 #include <linux/initrd.h>
  47 #include <linux/key.h>
  48 #include <linux/times.h>
  49 #include <linux/limits.h>
  50 #include <linux/dcache.h>
  51 #include <linux/dnotify.h>
  52 #include <linux/syscalls.h>
  53 #include <linux/vmstat.h>
  54 #include <linux/nfs_fs.h>
  55 #include <linux/acpi.h>
  56 #include <linux/reboot.h>
  57 #include <linux/ftrace.h>
  58 #include <linux/perf_event.h>
  59 #include <linux/kprobes.h>
  60 #include <linux/pipe_fs_i.h>
  61 #include <linux/oom.h>
  62 #include <linux/kmod.h>
  63 #include <linux/capability.h>
  64 #include <linux/binfmts.h>
  65 #include <linux/sched/sysctl.h>
  66 #include <linux/kexec.h>
  67 #include <linux/bpf.h>
  68
  69 #include <asm/uaccess.h>
  70 #include <asm/processor.h>
  71
  72 #ifdef CONFIG_X86
  73 #include <asm/nmi.h>
  74 #include <asm/stacktrace.h>
  75 #include <asm/io.h>
  76 #endif
  77 #ifdef CONFIG_SPARC
  78 #include <asm/setup.h>
  79 #endif
  80 #ifdef CONFIG_BSD_PROCESS_ACCT
  81 #include <linux/acct.h>
  82 #endif
  83 #ifdef CONFIG_RT_MUTEXES
  84 #include <linux/rtmutex.h>
  85 #endif
  86 #if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_LOCK_STAT)
  87 #include <linux/lockdep.h>
  88 #endif
  89 #ifdef CONFIG_CHR_DEV_SG
  90 #include <scsi/sg.h>
  91 #endif
  92
  93 #ifdef CONFIG_LOCKUP_DETECTOR
  94 #include <linux/nmi.h>
  95 #endif
  96
  97 #if defined(CONFIG_SYSCTL)
  98
  99 /* External variables not in a header file. */
 100 extern int suid_dumpable;
 101 #ifdef CONFIG_COREDUMP
 102 extern int core_uses_pid;
 103 extern char core_pattern[];
 104 extern unsigned int core_pipe_limit;
 105 #endif
 106 extern int pid_max;
 107 extern int pid_max_min, pid_max_max;
 108 extern int percpu_pagelist_fraction;
 109 extern int latencytop_enabled;
 110 extern unsigned int sysctl_nr_open_min, sysctl_nr_open_max;
 111 #ifndef CONFIG_MMU
 112 extern int sysctl_nr_trim_pages;
 113 #endif
 114
 115 /* Constants used for minimum and  maximum */
 116 #ifdef CONFIG_LOCKUP_DETECTOR
 117 static int sixty = 60;
 118 #endif
 119
 120 static int __maybe_unused neg_one = -1;
 121
 122 static int zero;
 123 static int __maybe_unused one = 1;
 124 static int __maybe_unused two = 2;
 125 static int __maybe_unused four = 4;
 126 static unsigned long one_ul = 1;
 127 static int one_hundred = 100;
 128 static int one_thousand = 1000;
 129 #ifdef CONFIG_PRINTK
 130 static int ten_thousand = 10000;
 131 #endif
 132 #ifdef CONFIG_PERF_EVENTS
 133 static int six_hundred_forty_kb = 640 * 1024;
 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
 143 static int ngroups_max = NGROUPS_MAX;
 144 static const int cap_last_cap = CAP_LAST_CAP;
 145
 146 /*this is needed for proc_doulongvec_minmax of sysctl_hung_task_timeout_secs */
 147 #ifdef CONFIG_DETECT_HUNG_TASK
 148 static unsigned long hung_task_timeout_max = (LONG_MAX/HZ);
 149 #endif
 150
 151 #ifdef CONFIG_INOTIFY_USER
 152 #include <linux/inotify.h>
 153 #endif
 154 #ifdef CONFIG_SPARC
 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
 175 #define SYSCTL_WRITES_LEGACY    -1
 176 #define SYSCTL_WRITES_WARN       0
 177 #define SYSCTL_WRITES_STRICT     1
 178
 179 static int sysctl_writes_strict = SYSCTL_WRITES_STRICT;
 180
 181 static int proc_do_cad_pid(struct ctl_table *table, int write,
 182                   void __user *buffer, size_t *lenp, loff_t *ppos);
 183 static int proc_taint(struct ctl_table *table, int write,
 184                                void __user *buffer, size_t *lenp, loff_t *ppos);
 185 #endif
 186
 187 #ifdef CONFIG_PRINTK
 188 static int proc_dointvec_minmax_sysadmin(struct ctl_table *table, int write,
 189                                 void __user *buffer, size_t *lenp, loff_t *ppos);
 190 #endif
 191
 192 static int proc_dointvec_minmax_coredump(struct ctl_table *table, int write,
 193                 void __user *buffer, size_t *lenp, loff_t *ppos);
 194 #ifdef CONFIG_COREDUMP
 195 static int proc_dostring_coredump(struct ctl_table *table, int write,
 196                 void __user *buffer, size_t *lenp, loff_t *ppos);
 197 #endif
 198
 199 #ifdef CONFIG_MAGIC_SYSRQ
 200 /* Note: sysrq code uses it's own private copy */
 201 static int __sysrq_enabled = CONFIG_MAGIC_SYSRQ_DEFAULT_ENABLE;
 202
 203 static int sysrq_sysctl_handler(struct ctl_table *table, int write,
 204                                 void __user *buffer, size_t *lenp,
 205                                 loff_t *ppos)
 206 {
 207         int error;
 208
 209         error = proc_dointvec(table, write, buffer, lenp, ppos);
 210         if (error)
 211                 return error;
 212
 213         if (write)
 214                 sysrq_toggle_support(__sysrq_enabled);
 215
 216         return 0;
 217 }
 218
 219 #endif
 220
 221 static struct ctl_table kern_table[];
 222 static struct ctl_table vm_table[];
 223 static struct ctl_table fs_table[];
 224 static struct ctl_table debug_table[];
 225 static struct ctl_table dev_table[];
 226 extern struct ctl_table random_table[];
 227 #ifdef CONFIG_EPOLL
 228 extern struct ctl_table epoll_table[];
 229 #endif
 230
 231 #ifdef HAVE_ARCH_PICK_MMAP_LAYOUT
 232 int sysctl_legacy_va_layout;
 233 #endif
 234
 235 /* The default sysctl tables: */
 236
 237 static struct ctl_table sysctl_base_table[] = {
 238         {
 239                 .procname       = "kernel",
 240                 .mode           = 0555,
 241                 .child          = kern_table,
 242         },
 243         {
 244                 .procname       = "vm",
 245                 .mode           = 0555,
 246                 .child          = vm_table,
 247         },
 248         {
 249                 .procname       = "fs",
 250                 .mode           = 0555,
 251                 .child          = fs_table,
 252         },
 253         {
 254                 .procname       = "debug",
 255                 .mode           = 0555,
 256                 .child          = debug_table,
 257         },
 258         {
 259                 .procname       = "dev",
 260                 .mode           = 0555,
 261                 .child          = dev_table,
 262         },
 263         { }
 264 };
 265
 266 #ifdef CONFIG_SCHED_DEBUG
 267 static int min_sched_granularity_ns = 100000;           /* 100 usecs */
 268 static int max_sched_granularity_ns = NSEC_PER_SEC;     /* 1 second */
 269 static int min_wakeup_granularity_ns;                   /* 0 usecs */
 270 static int max_wakeup_granularity_ns = NSEC_PER_SEC;    /* 1 second */
 271 #ifdef CONFIG_SMP
 272 static int min_sched_tunable_scaling = SCHED_TUNABLESCALING_NONE;
 273 static int max_sched_tunable_scaling = SCHED_TUNABLESCALING_END-1;
 274 #endif /* CONFIG_SMP */
 275 #endif /* CONFIG_SCHED_DEBUG */
 276
 277 #ifdef CONFIG_COMPACTION
 278 static int min_extfrag_threshold;
 279 static int max_extfrag_threshold = 1000;
 280 #endif
 281
 282 static struct ctl_table kern_table[] = {
 283         {
 284                 .procname       = "sched_child_runs_first",
 285                 .data           = &sysctl_sched_child_runs_first,
 286                 .maxlen         = sizeof(unsigned int),
 287                 .mode           = 0644,
 288                 .proc_handler   = proc_dointvec,
 289         },
 290 #ifdef CONFIG_SCHED_DEBUG
 291         {
 292                 .procname       = "sched_min_granularity_ns",
 293                 .data           = &sysctl_sched_min_granularity,
 294                 .maxlen         = sizeof(unsigned int),
 295                 .mode           = 0644,
 296                 .proc_handler   = sched_proc_update_handler,
 297                 .extra1         = &min_sched_granularity_ns,
 298                 .extra2         = &max_sched_granularity_ns,
 299         },
 300         {
 301                 .procname       = "sched_latency_ns",
 302                 .data           = &sysctl_sched_latency,
 303                 .maxlen         = sizeof(unsigned int),
 304                 .mode           = 0644,
 305                 .proc_handler   = sched_proc_update_handler,
 306                 .extra1         = &min_sched_granularity_ns,
 307                 .extra2         = &max_sched_granularity_ns,
 308         },
 309         {
 310                 .procname       = "sched_wakeup_granularity_ns",
 311                 .data           = &sysctl_sched_wakeup_granularity,
 312                 .maxlen         = sizeof(unsigned int),
 313                 .mode           = 0644,
 314                 .proc_handler   = sched_proc_update_handler,
 315                 .extra1         = &min_wakeup_granularity_ns,
 316                 .extra2         = &max_wakeup_granularity_ns,
 317         },
 318 #ifdef CONFIG_SMP
 319         {
 320                 .procname       = "sched_tunable_scaling",
 321                 .data           = &sysctl_sched_tunable_scaling,
 322                 .maxlen         = sizeof(enum sched_tunable_scaling),
 323                 .mode           = 0644,
 324                 .proc_handler   = sched_proc_update_handler,
 325                 .extra1         = &min_sched_tunable_scaling,
 326                 .extra2         = &max_sched_tunable_scaling,
 327         },
 328         {
 329                 .procname       = "sched_migration_cost_ns",
 330                 .data           = &sysctl_sched_migration_cost,
 331                 .maxlen         = sizeof(unsigned int),
 332                 .mode           = 0644,
 333                 .proc_handler   = proc_dointvec,
 334         },
 335         {
 336                 .procname       = "sched_nr_migrate",
 337                 .data           = &sysctl_sched_nr_migrate,
 338                 .maxlen         = sizeof(unsigned int),
 339                 .mode           = 0644,
 340                 .proc_handler   = proc_dointvec,
 341         },
 342         {
 343                 .procname       = "sched_time_avg_ms",
 344                 .data           = &sysctl_sched_time_avg,
 345                 .maxlen         = sizeof(unsigned int),
 346                 .mode           = 0644,
 347                 .proc_handler   = proc_dointvec,
 348         },
 349         {
 350                 .procname       = "sched_shares_window_ns",
 351                 .data           = &sysctl_sched_shares_window,
 352                 .maxlen         = sizeof(unsigned int),
 353                 .mode           = 0644,
 354                 .proc_handler   = proc_dointvec,
 355         },
 356 #ifdef CONFIG_SCHEDSTATS
 357         {
 358                 .procname       = "sched_schedstats",
 359                 .data           = NULL,
 360                 .maxlen         = sizeof(unsigned int),
 361                 .mode           = 0644,
 362                 .proc_handler   = sysctl_schedstats,
 363                 .extra1         = &zero,
 364                 .extra2         = &one,
 365         },
 366 #endif /* CONFIG_SCHEDSTATS */
 367 #endif /* CONFIG_SMP */
 368 #ifdef CONFIG_NUMA_BALANCING
 369         {
 370                 .procname       = "numa_balancing_scan_delay_ms",
 371                 .data           = &sysctl_numa_balancing_scan_delay,
 372                 .maxlen         = sizeof(unsigned int),
 373                 .mode           = 0644,
 374                 .proc_handler   = proc_dointvec,
 375         },
 376         {
 377                 .procname       = "numa_balancing_scan_period_min_ms",
 378                 .data           = &sysctl_numa_balancing_scan_period_min,
 379                 .maxlen         = sizeof(unsigned int),
 380                 .mode           = 0644,
 381                 .proc_handler   = proc_dointvec,
 382         },
 383         {
 384                 .procname       = "numa_balancing_scan_period_max_ms",
 385                 .data           = &sysctl_numa_balancing_scan_period_max,
 386                 .maxlen         = sizeof(unsigned int),
 387                 .mode           = 0644,
 388                 .proc_handler   = proc_dointvec,
 389         },
 390         {
 391                 .procname       = "numa_balancing_scan_size_mb",
 392                 .data           = &sysctl_numa_balancing_scan_size,
 393                 .maxlen         = sizeof(unsigned int),
 394                 .mode           = 0644,
 395                 .proc_handler   = proc_dointvec_minmax,
 396                 .extra1         = &one,
 397         },
 398         {
 399                 .procname       = "numa_balancing",
 400                 .data           = NULL, /* filled in by handler */
 401                 .maxlen         = sizeof(unsigned int),
 402                 .mode           = 0644,
 403                 .proc_handler   = sysctl_numa_balancing,
 404                 .extra1         = &zero,
 405                 .extra2         = &one,
 406         },
 407 #endif /* CONFIG_NUMA_BALANCING */
 408 #endif /* CONFIG_SCHED_DEBUG */
 409         {
 410                 .procname       = "sched_rt_period_us",
 411                 .data           = &sysctl_sched_rt_period,
 412                 .maxlen         = sizeof(unsigned int),
 413                 .mode           = 0644,
 414                 .proc_handler   = sched_rt_handler,
 415         },
 416         {
 417                 .procname       = "sched_rt_runtime_us",
 418                 .data           = &sysctl_sched_rt_runtime,
 419                 .maxlen         = sizeof(int),
 420                 .mode           = 0644,
 421                 .proc_handler   = sched_rt_handler,
 422         },
 423         {
 424                 .procname       = "sched_rr_timeslice_ms",
 425                 .data           = &sched_rr_timeslice,
 426                 .maxlen         = sizeof(int),
 427                 .mode           = 0644,
 428                 .proc_handler   = sched_rr_handler,
 429         },
 430 #ifdef CONFIG_SCHED_AUTOGROUP
 431         {
 432                 .procname       = "sched_autogroup_enabled",
 433                 .data           = &sysctl_sched_autogroup_enabled,
 434                 .maxlen         = sizeof(unsigned int),
 435                 .mode           = 0644,
 436                 .proc_handler   = proc_dointvec_minmax,
 437                 .extra1         = &zero,
 438                 .extra2         = &one,
 439         },
 440 #endif
 441 #ifdef CONFIG_CFS_BANDWIDTH
 442         {
 443                 .procname       = "sched_cfs_bandwidth_slice_us",
 444                 .data           = &sysctl_sched_cfs_bandwidth_slice,
 445                 .maxlen         = sizeof(unsigned int),
 446                 .mode           = 0644,
 447                 .proc_handler   = proc_dointvec_minmax,
 448                 .extra1         = &one,
 449         },
 450 #endif
 451 #ifdef CONFIG_PROVE_LOCKING
 452         {
 453                 .procname       = "prove_locking",
 454                 .data           = &prove_locking,
 455                 .maxlen         = sizeof(int),
 456                 .mode           = 0644,
 457                 .proc_handler   = proc_dointvec,
 458         },
 459 #endif
 460 #ifdef CONFIG_LOCK_STAT
 461         {
 462                 .procname       = "lock_stat",
 463                 .data           = &lock_stat,
 464                 .maxlen         = sizeof(int),
 465                 .mode           = 0644,
 466                 .proc_handler   = proc_dointvec,
 467         },
 468 #endif
 469         {
 470                 .procname       = "panic",
 471                 .data           = &panic_timeout,
 472                 .maxlen         = sizeof(int),
 473                 .mode           = 0644,
 474                 .proc_handler   = proc_dointvec,
 475         },
 476 #ifdef CONFIG_COREDUMP
 477         {
 478                 .procname       = "core_uses_pid",
 479                 .data           = &core_uses_pid,
 480                 .maxlen         = sizeof(int),
 481                 .mode           = 0644,
 482                 .proc_handler   = proc_dointvec,
 483         },
 484         {
 485                 .procname       = "core_pattern",
 486                 .data           = core_pattern,
 487                 .maxlen         = CORENAME_MAX_SIZE,
 488                 .mode           = 0644,
 489                 .proc_handler   = proc_dostring_coredump,
 490         },
 491         {
 492                 .procname       = "core_pipe_limit",
 493                 .data           = &core_pipe_limit,
 494                 .maxlen         = sizeof(unsigned int),
 495                 .mode           = 0644,
 496                 .proc_handler   = proc_dointvec,
 497         },
 498 #endif
 499 #ifdef CONFIG_PROC_SYSCTL
 500         {
 501                 .procname       = "tainted",
 502                 .maxlen         = sizeof(long),
 503                 .mode           = 0644,
 504                 .proc_handler   = proc_taint,
 505         },
 506         {
 507                 .procname       = "sysctl_writes_strict",
 508                 .data           = &sysctl_writes_strict,
 509                 .maxlen         = sizeof(int),
 510                 .mode           = 0644,
 511                 .proc_handler   = proc_dointvec_minmax,
 512                 .extra1         = &neg_one,
 513                 .extra2         = &one,
 514         },
 515 #endif
 516 #ifdef CONFIG_LATENCYTOP
 517         {
 518                 .procname       = "latencytop",
 519                 .data           = &latencytop_enabled,
 520                 .maxlen         = sizeof(int),
 521                 .mode           = 0644,
 522                 .proc_handler   = sysctl_latencytop,
 523         },
 524 #endif
 525 #ifdef CONFIG_BLK_DEV_INITRD
 526         {
 527                 .procname       = "real-root-dev",
 528                 .data           = &real_root_dev,
 529                 .maxlen         = sizeof(int),
 530                 .mode           = 0644,
 531                 .proc_handler   = proc_dointvec,
 532         },
 533 #endif
 534         {
 535                 .procname       = "print-fatal-signals",
 536                 .data           = &print_fatal_signals,
 537                 .maxlen         = sizeof(int),
 538                 .mode           = 0644,
 539                 .proc_handler   = proc_dointvec,
 540         },
 541 #ifdef CONFIG_SPARC
 542         {
 543                 .procname       = "reboot-cmd",
 544                 .data           = reboot_command,
 545                 .maxlen         = 256,
 546                 .mode           = 0644,
 547                 .proc_handler   = proc_dostring,
 548         },
 549         {
 550                 .procname       = "stop-a",
 551                 .data           = &stop_a_enabled,
 552                 .maxlen         = sizeof (int),
 553                 .mode           = 0644,
 554                 .proc_handler   = proc_dointvec,
 555         },
 556         {
 557                 .procname       = "scons-poweroff",
 558                 .data           = &scons_pwroff,
 559                 .maxlen         = sizeof (int),
 560                 .mode           = 0644,
 561                 .proc_handler   = proc_dointvec,
 562         },
 563 #endif
 564 #ifdef CONFIG_SPARC64
 565         {
 566                 .procname       = "tsb-ratio",
 567                 .data           = &sysctl_tsb_ratio,
 568                 .maxlen         = sizeof (int),
 569                 .mode           = 0644,
 570                 .proc_handler   = proc_dointvec,
 571         },
 572 #endif
 573 #ifdef __hppa__
 574         {
 575                 .procname       = "soft-power",
 576                 .data           = &pwrsw_enabled,
 577                 .maxlen         = sizeof (int),
 578                 .mode           = 0644,
 579                 .proc_handler   = proc_dointvec,
 580         },
 581 #endif
 582 #ifdef CONFIG_SYSCTL_ARCH_UNALIGN_ALLOW
 583         {
 584                 .procname       = "unaligned-trap",
 585                 .data           = &unaligned_enabled,
 586                 .maxlen         = sizeof (int),
 587                 .mode           = 0644,
 588                 .proc_handler   = proc_dointvec,
 589         },
 590 #endif
 591         {
 592                 .procname       = "ctrl-alt-del",
 593                 .data           = &C_A_D,
 594                 .maxlen         = sizeof(int),
 595                 .mode           = 0644,
 596                 .proc_handler   = proc_dointvec,
 597         },
 598 #ifdef CONFIG_FUNCTION_TRACER
 599         {
 600                 .procname       = "ftrace_enabled",
 601                 .data           = &ftrace_enabled,
 602                 .maxlen         = sizeof(int),
 603                 .mode           = 0644,
 604                 .proc_handler   = ftrace_enable_sysctl,
 605         },
 606 #endif
 607 #ifdef CONFIG_STACK_TRACER
 608         {
 609                 .procname       = "stack_tracer_enabled",
 610                 .data           = &stack_tracer_enabled,
 611                 .maxlen         = sizeof(int),
 612                 .mode           = 0644,
 613                 .proc_handler   = stack_trace_sysctl,
 614         },
 615 #endif
 616 #ifdef CONFIG_TRACING
 617         {
 618                 .procname       = "ftrace_dump_on_oops",
 619                 .data           = &ftrace_dump_on_oops,
 620                 .maxlen         = sizeof(int),
 621                 .mode           = 0644,
 622                 .proc_handler   = proc_dointvec,
 623         },
 624         {
 625                 .procname       = "traceoff_on_warning",
 626                 .data           = &__disable_trace_on_warning,
 627                 .maxlen         = sizeof(__disable_trace_on_warning),
 628                 .mode           = 0644,
 629                 .proc_handler   = proc_dointvec,
 630         },
 631         {
 632                 .procname       = "tracepoint_printk",
 633                 .data           = &tracepoint_printk,
 634                 .maxlen         = sizeof(tracepoint_printk),
 635                 .mode           = 0644,
 636                 .proc_handler   = proc_dointvec,
 637         },
 638 #endif
 639 #ifdef CONFIG_KEXEC_CORE
 640         {
 641                 .procname       = "kexec_load_disabled",
 642                 .data           = &kexec_load_disabled,
 643                 .maxlen         = sizeof(int),
 644                 .mode           = 0644,
 645                 /* only handle a transition from default "0" to "1" */
 646                 .proc_handler   = proc_dointvec_minmax,
 647                 .extra1         = &one,
 648                 .extra2         = &one,
 649         },
 650 #endif
 651 #ifdef CONFIG_MODULES
 652         {
 653                 .procname       = "modprobe",
 654                 .data           = &modprobe_path,
 655                 .maxlen         = KMOD_PATH_LEN,
 656                 .mode           = 0644,
 657                 .proc_handler   = proc_dostring,
 658         },
 659         {
 660                 .procname       = "modules_disabled",
 661                 .data           = &modules_disabled,
 662                 .maxlen         = sizeof(int),
 663                 .mode           = 0644,
 664                 /* only handle a transition from default "0" to "1" */
 665                 .proc_handler   = proc_dointvec_minmax,
 666                 .extra1         = &one,
 667                 .extra2         = &one,
 668         },
 669 #endif
 670 #ifdef CONFIG_UEVENT_HELPER
 671         {
 672                 .procname       = "hotplug",
 673                 .data           = &uevent_helper,
 674                 .maxlen         = UEVENT_HELPER_PATH_LEN,
 675                 .mode           = 0644,
 676                 .proc_handler   = proc_dostring,
 677         },
 678 #endif
 679 #ifdef CONFIG_CHR_DEV_SG
 680         {
 681                 .procname       = "sg-big-buff",
 682                 .data           = &sg_big_buff,
 683                 .maxlen         = sizeof (int),
 684                 .mode           = 0444,
 685                 .proc_handler   = proc_dointvec,
 686         },
 687 #endif
 688 #ifdef CONFIG_BSD_PROCESS_ACCT
 689         {
 690                 .procname       = "acct",
 691                 .data           = &acct_parm,
 692                 .maxlen         = 3*sizeof(int),
 693                 .mode           = 0644,
 694                 .proc_handler   = proc_dointvec,
 695         },
 696 #endif
 697 #ifdef CONFIG_MAGIC_SYSRQ
 698         {
 699                 .procname       = "sysrq",
 700                 .data           = &__sysrq_enabled,
 701                 .maxlen         = sizeof (int),
 702                 .mode           = 0644,
 703                 .proc_handler   = sysrq_sysctl_handler,
 704         },
 705 #endif
 706 #ifdef CONFIG_PROC_SYSCTL
 707         {
 708                 .procname       = "cad_pid",
 709                 .data           = NULL,
 710                 .maxlen         = sizeof (int),
 711                 .mode           = 0600,
 712                 .proc_handler   = proc_do_cad_pid,
 713         },
 714 #endif
 715         {
 716                 .procname       = "threads-max",
 717                 .data           = NULL,
 718                 .maxlen         = sizeof(int),
 719                 .mode           = 0644,
 720                 .proc_handler   = sysctl_max_threads,
 721         },
 722         {
 723                 .procname       = "random",
 724                 .mode           = 0555,
 725                 .child          = random_table,
 726         },
 727         {
 728                 .procname       = "usermodehelper",
 729                 .mode           = 0555,
 730                 .child          = usermodehelper_table,
 731         },
 732         {
 733                 .procname       = "overflowuid",
 734                 .data           = &overflowuid,
 735                 .maxlen         = sizeof(int),
 736                 .mode           = 0644,
 737                 .proc_handler   = proc_dointvec_minmax,
 738                 .extra1         = &minolduid,
 739                 .extra2         = &maxolduid,
 740         },
 741         {
 742                 .procname       = "overflowgid",
 743                 .data           = &overflowgid,
 744                 .maxlen         = sizeof(int),
 745                 .mode           = 0644,
 746                 .proc_handler   = proc_dointvec_minmax,
 747                 .extra1         = &minolduid,
 748                 .extra2         = &maxolduid,
 749         },
 750 #ifdef CONFIG_S390
 751 #ifdef CONFIG_MATHEMU
 752         {
 753                 .procname       = "ieee_emulation_warnings",
 754                 .data           = &sysctl_ieee_emulation_warnings,
 755                 .maxlen         = sizeof(int),
 756                 .mode           = 0644,
 757                 .proc_handler   = proc_dointvec,
 758         },
 759 #endif
 760         {
 761                 .procname       = "userprocess_debug",
 762                 .data           = &show_unhandled_signals,
 763                 .maxlen         = sizeof(int),
 764                 .mode           = 0644,
 765                 .proc_handler   = proc_dointvec,
 766         },
 767 #endif
 768         {
 769                 .procname       = "pid_max",
 770                 .data           = &pid_max,
 771                 .maxlen         = sizeof (int),
 772                 .mode           = 0644,
 773                 .proc_handler   = proc_dointvec_minmax,
 774                 .extra1         = &pid_max_min,
 775                 .extra2         = &pid_max_max,
 776         },
 777         {
 778                 .procname       = "panic_on_oops",
 779                 .data           = &panic_on_oops,
 780                 .maxlen         = sizeof(int),
 781                 .mode           = 0644,
 782                 .proc_handler   = proc_dointvec,
 783         },
 784 #if defined CONFIG_PRINTK
 785         {
 786                 .procname       = "printk",
 787                 .data           = &console_loglevel,
 788                 .maxlen         = 4*sizeof(int),
 789                 .mode           = 0644,
 790                 .proc_handler   = proc_dointvec,
 791         },
 792         {
 793                 .procname       = "printk_ratelimit",
 794                 .data           = &printk_ratelimit_state.interval,
 795                 .maxlen         = sizeof(int),
 796                 .mode           = 0644,
 797                 .proc_handler   = proc_dointvec_jiffies,
 798         },
 799         {
 800                 .procname       = "printk_ratelimit_burst",
 801                 .data           = &printk_ratelimit_state.burst,
 802                 .maxlen         = sizeof(int),
 803                 .mode           = 0644,
 804                 .proc_handler   = proc_dointvec,
 805         },
 806         {
 807                 .procname       = "printk_delay",
 808                 .data           = &printk_delay_msec,
 809                 .maxlen         = sizeof(int),
 810                 .mode           = 0644,
 811                 .proc_handler   = proc_dointvec_minmax,
 812                 .extra1         = &zero,
 813                 .extra2         = &ten_thousand,
 814         },
 815         {
 816                 .procname       = "printk_devkmsg",
 817                 .data           = devkmsg_log_str,
 818                 .maxlen         = DEVKMSG_STR_MAX_SIZE,
 819                 .mode           = 0644,
 820                 .proc_handler   = devkmsg_sysctl_set_loglvl,
 821         },
 822         {
 823                 .procname       = "dmesg_restrict",
 824                 .data           = &dmesg_restrict,
 825                 .maxlen         = sizeof(int),
 826                 .mode           = 0644,
 827                 .proc_handler   = proc_dointvec_minmax_sysadmin,
 828                 .extra1         = &zero,
 829                 .extra2         = &one,
 830         },
 831         {
 832                 .procname       = "kptr_restrict",
 833                 .data           = &kptr_restrict,
 834                 .maxlen         = sizeof(int),
 835                 .mode           = 0644,
 836                 .proc_handler   = proc_dointvec_minmax_sysadmin,
 837                 .extra1         = &zero,
 838                 .extra2         = &two,
 839         },
 840 #endif
 841         {
 842                 .procname       = "ngroups_max",
 843                 .data           = &ngroups_max,
 844                 .maxlen         = sizeof (int),
 845                 .mode           = 0444,
 846                 .proc_handler   = proc_dointvec,
 847         },
 848         {
 849                 .procname       = "cap_last_cap",
 850                 .data           = (void *)&cap_last_cap,
 851                 .maxlen         = sizeof(int),
 852                 .mode           = 0444,
 853                 .proc_handler   = proc_dointvec,
 854         },
 855 #if defined(CONFIG_LOCKUP_DETECTOR)
 856         {
 857                 .procname       = "watchdog",
 858                 .data           = &watchdog_user_enabled,
 859                 .maxlen         = sizeof (int),
 860                 .mode           = 0644,
 861                 .proc_handler   = proc_watchdog,
 862                 .extra1         = &zero,
 863                 .extra2         = &one,
 864         },
 865         {
 866                 .procname       = "watchdog_thresh",
 867                 .data           = &watchdog_thresh,
 868                 .maxlen         = sizeof(int),
 869                 .mode           = 0644,
 870                 .proc_handler   = proc_watchdog_thresh,
 871                 .extra1         = &zero,
 872                 .extra2         = &sixty,
 873         },
 874         {
 875                 .procname       = "nmi_watchdog",
 876                 .data           = &nmi_watchdog_enabled,
 877                 .maxlen         = sizeof (int),
 878                 .mode           = 0644,
 879                 .proc_handler   = proc_nmi_watchdog,
 880                 .extra1         = &zero,
 881 #if defined(CONFIG_HAVE_NMI_WATCHDOG) || defined(CONFIG_HARDLOCKUP_DETECTOR)
 882                 .extra2         = &one,
 883 #else
 884                 .extra2         = &zero,
 885 #endif
 886         },
 887         {
 888                 .procname       = "soft_watchdog",
 889                 .data           = &soft_watchdog_enabled,
 890                 .maxlen         = sizeof (int),
 891                 .mode           = 0644,
 892                 .proc_handler   = proc_soft_watchdog,
 893                 .extra1         = &zero,
 894                 .extra2         = &one,
 895         },
 896         {
 897                 .procname       = "watchdog_cpumask",
 898                 .data           = &watchdog_cpumask_bits,
 899                 .maxlen         = NR_CPUS,
 900                 .mode           = 0644,
 901                 .proc_handler   = proc_watchdog_cpumask,
 902         },
 903         {
 904                 .procname       = "softlockup_panic",
 905                 .data           = &softlockup_panic,
 906                 .maxlen         = sizeof(int),
 907                 .mode           = 0644,
 908                 .proc_handler   = proc_dointvec_minmax,
 909                 .extra1         = &zero,
 910                 .extra2         = &one,
 911         },
 912 #ifdef CONFIG_HARDLOCKUP_DETECTOR
 913         {
 914                 .procname       = "hardlockup_panic",
 915                 .data           = &hardlockup_panic,
 916                 .maxlen         = sizeof(int),
 917                 .mode           = 0644,
 918                 .proc_handler   = proc_dointvec_minmax,
 919                 .extra1         = &zero,
 920                 .extra2         = &one,
 921         },
 922 #endif
 923 #ifdef CONFIG_SMP
 924         {
 925                 .procname       = "softlockup_all_cpu_backtrace",
 926                 .data           = &sysctl_softlockup_all_cpu_backtrace,
 927                 .maxlen         = sizeof(int),
 928                 .mode           = 0644,
 929                 .proc_handler   = proc_dointvec_minmax,
 930                 .extra1         = &zero,
 931                 .extra2         = &one,
 932         },
 933         {
 934                 .procname       = "hardlockup_all_cpu_backtrace",
 935                 .data           = &sysctl_hardlockup_all_cpu_backtrace,
 936                 .maxlen         = sizeof(int),
 937                 .mode           = 0644,
 938                 .proc_handler   = proc_dointvec_minmax,
 939                 .extra1         = &zero,
 940                 .extra2         = &one,
 941         },
 942 #endif /* CONFIG_SMP */
 943 #endif
 944 #if defined(CONFIG_X86_LOCAL_APIC) && defined(CONFIG_X86)
 945         {
 946                 .procname       = "unknown_nmi_panic",
 947                 .data           = &unknown_nmi_panic,
 948                 .maxlen         = sizeof (int),
 949                 .mode           = 0644,
 950                 .proc_handler   = proc_dointvec,
 951         },
 952 #endif
 953 #if defined(CONFIG_X86)
 954         {
 955                 .procname       = "panic_on_unrecovered_nmi",
 956                 .data           = &panic_on_unrecovered_nmi,
 957                 .maxlen         = sizeof(int),
 958                 .mode           = 0644,
 959                 .proc_handler   = proc_dointvec,
 960         },
 961         {
 962                 .procname       = "panic_on_io_nmi",
 963                 .data           = &panic_on_io_nmi,
 964                 .maxlen         = sizeof(int),
 965                 .mode           = 0644,
 966                 .proc_handler   = proc_dointvec,
 967         },
 968 #ifdef CONFIG_DEBUG_STACKOVERFLOW
 969         {
 970                 .procname       = "panic_on_stackoverflow",
 971                 .data           = &sysctl_panic_on_stackoverflow,
 972                 .maxlen         = sizeof(int),
 973                 .mode           = 0644,
 974                 .proc_handler   = proc_dointvec,
 975         },
 976 #endif
 977         {
 978                 .procname       = "bootloader_type",
 979                 .data           = &bootloader_type,
 980                 .maxlen         = sizeof (int),
 981                 .mode           = 0444,
 982                 .proc_handler   = proc_dointvec,
 983         },
 984         {
 985                 .procname       = "bootloader_version",
 986                 .data           = &bootloader_version,
 987                 .maxlen         = sizeof (int),
 988                 .mode           = 0444,
 989                 .proc_handler   = proc_dointvec,
 990         },
 991         {
 992                 .procname       = "kstack_depth_to_print",
 993                 .data           = &kstack_depth_to_print,
 994                 .maxlen         = sizeof(int),
 995                 .mode           = 0644,
 996                 .proc_handler   = proc_dointvec,
 997         },
 998         {
 999                 .procname       = "io_delay_type",
1000                 .data           = &io_delay_type,
1001                 .maxlen         = sizeof(int),
1002                 .mode           = 0644,
1003                 .proc_handler   = proc_dointvec,
1004         },
1005 #endif
1006 #if defined(CONFIG_MMU)
1007         {
1008                 .procname       = "randomize_va_space",
1009                 .data           = &randomize_va_space,
1010                 .maxlen         = sizeof(int),
1011                 .mode           = 0644,
1012                 .proc_handler   = proc_dointvec,
1013         },
1014 #endif
1015 #if defined(CONFIG_S390) && defined(CONFIG_SMP)
1016         {
1017                 .procname       = "spin_retry",
1018                 .data           = &spin_retry,
1019                 .maxlen         = sizeof (int),
1020                 .mode           = 0644,
1021                 .proc_handler   = proc_dointvec,
1022         },
1023 #endif
1024 #if     defined(CONFIG_ACPI_SLEEP) && defined(CONFIG_X86)
1025         {
1026                 .procname       = "acpi_video_flags",
1027                 .data           = &acpi_realmode_flags,
1028                 .maxlen         = sizeof (unsigned long),
1029                 .mode           = 0644,
1030                 .proc_handler   = proc_doulongvec_minmax,
1031         },
1032 #endif
1033 #ifdef CONFIG_SYSCTL_ARCH_UNALIGN_NO_WARN
1034         {
1035                 .procname       = "ignore-unaligned-usertrap",
1036                 .data           = &no_unaligned_warning,
1037                 .maxlen         = sizeof (int),
1038                 .mode           = 0644,
1039                 .proc_handler   = proc_dointvec,
1040         },
1041 #endif
1042 #ifdef CONFIG_IA64
1043         {
1044                 .procname       = "unaligned-dump-stack",
1045                 .data           = &unaligned_dump_stack,
1046                 .maxlen         = sizeof (int),
1047                 .mode           = 0644,
1048                 .proc_handler   = proc_dointvec,
1049         },
1050 #endif
1051 #ifdef CONFIG_DETECT_HUNG_TASK
1052         {
1053                 .procname       = "hung_task_panic",
1054                 .data           = &sysctl_hung_task_panic,
1055                 .maxlen         = sizeof(int),
1056                 .mode           = 0644,
1057                 .proc_handler   = proc_dointvec_minmax,
1058                 .extra1         = &zero,
1059                 .extra2         = &one,
1060         },
1061         {
1062                 .procname       = "hung_task_check_count",
1063                 .data           = &sysctl_hung_task_check_count,
1064                 .maxlen         = sizeof(int),
1065                 .mode           = 0644,
1066                 .proc_handler   = proc_dointvec_minmax,
1067                 .extra1         = &zero,
1068         },
1069         {
1070                 .procname       = "hung_task_timeout_secs",
1071                 .data           = &sysctl_hung_task_timeout_secs,
1072                 .maxlen         = sizeof(unsigned long),
1073                 .mode           = 0644,
1074                 .proc_handler   = proc_dohung_task_timeout_secs,
1075                 .extra2         = &hung_task_timeout_max,
1076         },
1077         {
1078                 .procname       = "hung_task_warnings",
1079                 .data           = &sysctl_hung_task_warnings,
1080                 .maxlen         = sizeof(int),
1081                 .mode           = 0644,
1082                 .proc_handler   = proc_dointvec_minmax,
1083                 .extra1         = &neg_one,
1084         },
1085 #endif
1086 #ifdef CONFIG_RT_MUTEXES
1087         {
1088                 .procname       = "max_lock_depth",
1089                 .data           = &max_lock_depth,
1090                 .maxlen         = sizeof(int),
1091                 .mode           = 0644,
1092                 .proc_handler   = proc_dointvec,
1093         },
1094 #endif
1095         {
1096                 .procname       = "poweroff_cmd",
1097                 .data           = &poweroff_cmd,
1098                 .maxlen         = POWEROFF_CMD_PATH_LEN,
1099                 .mode           = 0644,
1100                 .proc_handler   = proc_dostring,
1101         },
1102 #ifdef CONFIG_KEYS
1103         {
1104                 .procname       = "keys",
1105                 .mode           = 0555,
1106                 .child          = key_sysctls,
1107         },
1108 #endif
1109 #ifdef CONFIG_PERF_EVENTS
1110         /*
1111          * User-space scripts rely on the existence of this file
1112          * as a feature check for perf_events being enabled.
1113          *
1114          * So it's an ABI, do not remove!
1115          */
1116         {
1117                 .procname       = "perf_event_paranoid",
1118                 .data           = &sysctl_perf_event_paranoid,
1119                 .maxlen         = sizeof(sysctl_perf_event_paranoid),
1120                 .mode           = 0644,
1121                 .proc_handler   = proc_dointvec,
1122         },
1123         {
1124                 .procname       = "perf_event_mlock_kb",
1125                 .data           = &sysctl_perf_event_mlock,
1126                 .maxlen         = sizeof(sysctl_perf_event_mlock),
1127                 .mode           = 0644,
1128                 .proc_handler   = proc_dointvec,
1129         },
1130         {
1131                 .procname       = "perf_event_max_sample_rate",
1132                 .data           = &sysctl_perf_event_sample_rate,
1133                 .maxlen         = sizeof(sysctl_perf_event_sample_rate),
1134                 .mode           = 0644,
1135                 .proc_handler   = perf_proc_update_handler,
1136                 .extra1         = &one,
1137         },
1138         {
1139                 .procname       = "perf_cpu_time_max_percent",
1140                 .data           = &sysctl_perf_cpu_time_max_percent,
1141                 .maxlen         = sizeof(sysctl_perf_cpu_time_max_percent),
1142                 .mode           = 0644,
1143                 .proc_handler   = perf_cpu_time_max_percent_handler,
1144                 .extra1         = &zero,
1145                 .extra2         = &one_hundred,
1146         },
1147         {
1148                 .procname       = "perf_event_max_stack",
1149                 .data           = &sysctl_perf_event_max_stack,
1150                 .maxlen         = sizeof(sysctl_perf_event_max_stack),
1151                 .mode           = 0644,
1152                 .proc_handler   = perf_event_max_stack_handler,
1153                 .extra1         = &zero,
1154                 .extra2         = &six_hundred_forty_kb,
1155         },
1156         {
1157                 .procname       = "perf_event_max_contexts_per_stack",
1158                 .data           = &sysctl_perf_event_max_contexts_per_stack,
1159                 .maxlen         = sizeof(sysctl_perf_event_max_contexts_per_stack),
1160                 .mode           = 0644,
1161                 .proc_handler   = perf_event_max_stack_handler,
1162                 .extra1         = &zero,
1163                 .extra2         = &one_thousand,
1164         },
1165 #endif
1166 #ifdef CONFIG_KMEMCHECK
1167         {
1168                 .procname       = "kmemcheck",
1169                 .data           = &kmemcheck_enabled,
1170                 .maxlen         = sizeof(int),
1171                 .mode           = 0644,
1172                 .proc_handler   = proc_dointvec,
1173         },
1174 #endif
1175         {
1176                 .procname       = "panic_on_warn",
1177                 .data           = &panic_on_warn,
1178                 .maxlen         = sizeof(int),
1179                 .mode           = 0644,
1180                 .proc_handler   = proc_dointvec_minmax,
1181                 .extra1         = &zero,
1182                 .extra2         = &one,
1183         },
1184 #if defined(CONFIG_SMP) && defined(CONFIG_NO_HZ_COMMON)
1185         {
1186                 .procname       = "timer_migration",
1187                 .data           = &sysctl_timer_migration,
1188                 .maxlen         = sizeof(unsigned int),
1189                 .mode           = 0644,
1190                 .proc_handler   = timer_migration_handler,
1191         },
1192 #endif
1193 #ifdef CONFIG_BPF_SYSCALL
1194         {
1195                 .procname       = "unprivileged_bpf_disabled",
1196                 .data           = &sysctl_unprivileged_bpf_disabled,
1197                 .maxlen         = sizeof(sysctl_unprivileged_bpf_disabled),
1198                 .mode           = 0644,
1199                 /* only handle a transition from default "0" to "1" */
1200                 .proc_handler   = proc_dointvec_minmax,
1201                 .extra1         = &one,
1202                 .extra2         = &one,
1203         },
1204 #endif
1205 #if defined(CONFIG_TREE_RCU) || defined(CONFIG_PREEMPT_RCU)
1206         {
1207                 .procname       = "panic_on_rcu_stall",
1208                 .data           = &sysctl_panic_on_rcu_stall,
1209                 .maxlen         = sizeof(sysctl_panic_on_rcu_stall),
1210                 .mode           = 0644,
1211                 .proc_handler   = proc_dointvec_minmax,
1212                 .extra1         = &zero,
1213                 .extra2         = &one,
1214         },
1215 #endif
1216         { }
1217 };
1218
1219 static struct ctl_table vm_table[] = {
1220         {
1221                 .procname       = "overcommit_memory",
1222                 .data           = &sysctl_overcommit_memory,
1223                 .maxlen         = sizeof(sysctl_overcommit_memory),
1224                 .mode           = 0644,
1225                 .proc_handler   = proc_dointvec_minmax,
1226                 .extra1         = &zero,
1227                 .extra2         = &two,
1228         },
1229         {
1230                 .procname       = "panic_on_oom",
1231                 .data           = &sysctl_panic_on_oom,
1232                 .maxlen         = sizeof(sysctl_panic_on_oom),
1233                 .mode           = 0644,
1234                 .proc_handler   = proc_dointvec_minmax,
1235                 .extra1         = &zero,
1236                 .extra2         = &two,
1237         },
1238         {
1239                 .procname       = "oom_kill_allocating_task",
1240                 .data           = &sysctl_oom_kill_allocating_task,
1241                 .maxlen         = sizeof(sysctl_oom_kill_allocating_task),
1242                 .mode           = 0644,
1243                 .proc_handler   = proc_dointvec,
1244         },
1245         {
1246                 .procname       = "oom_dump_tasks",
1247                 .data           = &sysctl_oom_dump_tasks,
1248                 .maxlen         = sizeof(sysctl_oom_dump_tasks),
1249                 .mode           = 0644,
1250                 .proc_handler   = proc_dointvec,
1251         },
1252         {
1253                 .procname       = "overcommit_ratio",
1254                 .data           = &sysctl_overcommit_ratio,
1255                 .maxlen         = sizeof(sysctl_overcommit_ratio),
1256                 .mode           = 0644,
1257                 .proc_handler   = overcommit_ratio_handler,
1258         },
1259         {
1260                 .procname       = "overcommit_kbytes",
1261                 .data           = &sysctl_overcommit_kbytes,
1262                 .maxlen         = sizeof(sysctl_overcommit_kbytes),
1263                 .mode           = 0644,
1264                 .proc_handler   = overcommit_kbytes_handler,
1265         },
1266         {
1267                 .procname       = "page-cluster",
1268                 .data           = &page_cluster,
1269                 .maxlen         = sizeof(int),
1270                 .mode           = 0644,
1271                 .proc_handler   = proc_dointvec_minmax,
1272                 .extra1         = &zero,
1273         },
1274         {
1275                 .procname       = "dirty_background_ratio",
1276                 .data           = &dirty_background_ratio,
1277                 .maxlen         = sizeof(dirty_background_ratio),
1278                 .mode           = 0644,
1279                 .proc_handler   = dirty_background_ratio_handler,
1280                 .extra1         = &zero,
1281                 .extra2         = &one_hundred,
1282         },
1283         {
1284                 .procname       = "dirty_background_bytes",
1285                 .data           = &dirty_background_bytes,
1286                 .maxlen         = sizeof(dirty_background_bytes),
1287                 .mode           = 0644,
1288                 .proc_handler   = dirty_background_bytes_handler,
1289                 .extra1         = &one_ul,
1290         },
1291         {
1292                 .procname       = "dirty_ratio",
1293                 .data           = &vm_dirty_ratio,
1294                 .maxlen         = sizeof(vm_dirty_ratio),
1295                 .mode           = 0644,
1296                 .proc_handler   = dirty_ratio_handler,
1297                 .extra1         = &zero,
1298                 .extra2         = &one_hundred,
1299         },
1300         {
1301                 .procname       = "dirty_bytes",
1302                 .data           = &vm_dirty_bytes,
1303                 .maxlen         = sizeof(vm_dirty_bytes),
1304                 .mode           = 0644,
1305                 .proc_handler   = dirty_bytes_handler,
1306                 .extra1         = &dirty_bytes_min,
1307         },
1308         {
1309                 .procname       = "dirty_writeback_centisecs",
1310                 .data           = &dirty_writeback_interval,
1311                 .maxlen         = sizeof(dirty_writeback_interval),
1312                 .mode           = 0644,
1313                 .proc_handler   = dirty_writeback_centisecs_handler,
1314         },
1315         {
1316                 .procname       = "dirty_expire_centisecs",
1317                 .data           = &dirty_expire_interval,
1318                 .maxlen         = sizeof(dirty_expire_interval),
1319                 .mode           = 0644,
1320                 .proc_handler   = proc_dointvec_minmax,
1321                 .extra1         = &zero,
1322         },
1323         {
1324                 .procname       = "dirtytime_expire_seconds",
1325                 .data           = &dirtytime_expire_interval,
1326                 .maxlen         = sizeof(dirty_expire_interval),
1327                 .mode           = 0644,
1328                 .proc_handler   = dirtytime_interval_handler,
1329                 .extra1         = &zero,
1330         },
1331         {
1332                 .procname       = "nr_pdflush_threads",
1333                 .mode           = 0444 /* read-only */,
1334                 .proc_handler   = pdflush_proc_obsolete,
1335         },
1336         {
1337                 .procname       = "swappiness",
1338                 .data           = &vm_swappiness,
1339                 .maxlen         = sizeof(vm_swappiness),
1340                 .mode           = 0644,
1341                 .proc_handler   = proc_dointvec_minmax,
1342                 .extra1         = &zero,
1343                 .extra2         = &one_hundred,
1344         },
1345 #ifdef CONFIG_HUGETLB_PAGE
1346         {
1347                 .procname       = "nr_hugepages",
1348                 .data           = NULL,
1349                 .maxlen         = sizeof(unsigned long),
1350                 .mode           = 0644,
1351                 .proc_handler   = hugetlb_sysctl_handler,
1352         },
1353 #ifdef CONFIG_NUMA
1354         {
1355                 .procname       = "nr_hugepages_mempolicy",
1356                 .data           = NULL,
1357                 .maxlen         = sizeof(unsigned long),
1358                 .mode           = 0644,
1359                 .proc_handler   = &hugetlb_mempolicy_sysctl_handler,
1360         },
1361 #endif
1362          {
1363                 .procname       = "hugetlb_shm_group",
1364                 .data           = &sysctl_hugetlb_shm_group,
1365                 .maxlen         = sizeof(gid_t),
1366                 .mode           = 0644,
1367                 .proc_handler   = proc_dointvec,
1368          },
1369          {
1370                 .procname       = "hugepages_treat_as_movable",
1371                 .data           = &hugepages_treat_as_movable,
1372                 .maxlen         = sizeof(int),
1373                 .mode           = 0644,
1374                 .proc_handler   = proc_dointvec,
1375         },
1376         {
1377                 .procname       = "nr_overcommit_hugepages",
1378                 .data           = NULL,
1379                 .maxlen         = sizeof(unsigned long),
1380                 .mode           = 0644,
1381                 .proc_handler   = hugetlb_overcommit_handler,
1382         },
1383 #endif
1384         {
1385                 .procname       = "lowmem_reserve_ratio",
1386                 .data           = &sysctl_lowmem_reserve_ratio,
1387                 .maxlen         = sizeof(sysctl_lowmem_reserve_ratio),
1388                 .mode           = 0644,
1389                 .proc_handler   = lowmem_reserve_ratio_sysctl_handler,
1390         },
1391         {
1392                 .procname       = "drop_caches",
1393                 .data           = &sysctl_drop_caches,
1394                 .maxlen         = sizeof(int),
1395                 .mode           = 0644,
1396                 .proc_handler   = drop_caches_sysctl_handler,
1397                 .extra1         = &one,
1398                 .extra2         = &four,
1399         },
1400 #ifdef CONFIG_COMPACTION
1401         {
1402                 .procname       = "compact_memory",
1403                 .data           = &sysctl_compact_memory,
1404                 .maxlen         = sizeof(int),
1405                 .mode           = 0200,
1406                 .proc_handler   = sysctl_compaction_handler,
1407         },
1408         {
1409                 .procname       = "extfrag_threshold",
1410                 .data           = &sysctl_extfrag_threshold,
1411                 .maxlen         = sizeof(int),
1412                 .mode           = 0644,
1413                 .proc_handler   = sysctl_extfrag_handler,
1414                 .extra1         = &min_extfrag_threshold,
1415                 .extra2         = &max_extfrag_threshold,
1416         },
1417         {
1418                 .procname       = "compact_unevictable_allowed",
1419                 .data           = &sysctl_compact_unevictable_allowed,
1420                 .maxlen         = sizeof(int),
1421                 .mode           = 0644,
1422                 .proc_handler   = proc_dointvec,
1423                 .extra1         = &zero,
1424                 .extra2         = &one,
1425         },
1426
1427 #endif /* CONFIG_COMPACTION */
1428         {
1429                 .procname       = "min_free_kbytes",
1430                 .data           = &min_free_kbytes,
1431                 .maxlen         = sizeof(min_free_kbytes),
1432                 .mode           = 0644,
1433                 .proc_handler   = min_free_kbytes_sysctl_handler,
1434                 .extra1         = &zero,
1435         },
1436         {
1437                 .procname       = "watermark_scale_factor",
1438                 .data           = &watermark_scale_factor,
1439                 .maxlen         = sizeof(watermark_scale_factor),
1440                 .mode           = 0644,
1441                 .proc_handler   = watermark_scale_factor_sysctl_handler,
1442                 .extra1         = &one,
1443                 .extra2         = &one_thousand,
1444         },
1445         {
1446                 .procname       = "percpu_pagelist_fraction",
1447                 .data           = &percpu_pagelist_fraction,
1448                 .maxlen         = sizeof(percpu_pagelist_fraction),
1449                 .mode           = 0644,
1450                 .proc_handler   = percpu_pagelist_fraction_sysctl_handler,
1451                 .extra1         = &zero,
1452         },
1453 #ifdef CONFIG_MMU
1454         {
1455                 .procname       = "max_map_count",
1456                 .data           = &sysctl_max_map_count,
1457                 .maxlen         = sizeof(sysctl_max_map_count),
1458                 .mode           = 0644,
1459                 .proc_handler   = proc_dointvec_minmax,
1460                 .extra1         = &zero,
1461         },
1462 #else
1463         {
1464                 .procname       = "nr_trim_pages",
1465                 .data           = &sysctl_nr_trim_pages,
1466                 .maxlen         = sizeof(sysctl_nr_trim_pages),
1467                 .mode           = 0644,
1468                 .proc_handler   = proc_dointvec_minmax,
1469                 .extra1         = &zero,
1470         },
1471 #endif
1472         {
1473                 .procname       = "laptop_mode",
1474                 .data           = &laptop_mode,
1475                 .maxlen         = sizeof(laptop_mode),
1476                 .mode           = 0644,
1477                 .proc_handler   = proc_dointvec_jiffies,
1478         },
1479         {
1480                 .procname       = "block_dump",
1481                 .data           = &block_dump,
1482                 .maxlen         = sizeof(block_dump),
1483                 .mode           = 0644,
1484                 .proc_handler   = proc_dointvec,
1485                 .extra1         = &zero,
1486         },
1487         {
1488                 .procname       = "vfs_cache_pressure",
1489                 .data           = &sysctl_vfs_cache_pressure,
1490                 .maxlen         = sizeof(sysctl_vfs_cache_pressure),
1491                 .mode           = 0644,
1492                 .proc_handler   = proc_dointvec,
1493                 .extra1         = &zero,
1494         },
1495 #ifdef HAVE_ARCH_PICK_MMAP_LAYOUT
1496         {
1497                 .procname       = "legacy_va_layout",
1498                 .data           = &sysctl_legacy_va_layout,
1499                 .maxlen         = sizeof(sysctl_legacy_va_layout),
1500                 .mode           = 0644,
1501                 .proc_handler   = proc_dointvec,
1502                 .extra1         = &zero,
1503         },
1504 #endif
1505 #ifdef CONFIG_NUMA
1506         {
1507                 .procname       = "zone_reclaim_mode",
1508                 .data           = &node_reclaim_mode,
1509                 .maxlen         = sizeof(node_reclaim_mode),
1510                 .mode           = 0644,
1511                 .proc_handler   = proc_dointvec,
1512                 .extra1         = &zero,
1513         },
1514         {
1515                 .procname       = "min_unmapped_ratio",
1516                 .data           = &sysctl_min_unmapped_ratio,
1517                 .maxlen         = sizeof(sysctl_min_unmapped_ratio),
1518                 .mode           = 0644,
1519                 .proc_handler   = sysctl_min_unmapped_ratio_sysctl_handler,
1520                 .extra1         = &zero,
1521                 .extra2         = &one_hundred,
1522         },
1523         {
1524                 .procname       = "min_slab_ratio",
1525                 .data           = &sysctl_min_slab_ratio,
1526                 .maxlen         = sizeof(sysctl_min_slab_ratio),
1527                 .mode           = 0644,
1528                 .proc_handler   = sysctl_min_slab_ratio_sysctl_handler,
1529                 .extra1         = &zero,
1530                 .extra2         = &one_hundred,
1531         },
1532 #endif
1533 #ifdef CONFIG_SMP
1534         {
1535                 .procname       = "stat_interval",
1536                 .data           = &sysctl_stat_interval,
1537                 .maxlen         = sizeof(sysctl_stat_interval),
1538                 .mode           = 0644,
1539                 .proc_handler   = proc_dointvec_jiffies,
1540         },
1541         {
1542                 .procname       = "stat_refresh",
1543                 .data           = NULL,
1544                 .maxlen         = 0,
1545                 .mode           = 0600,
1546                 .proc_handler   = vmstat_refresh,
1547         },
1548 #endif
1549 #ifdef CONFIG_MMU
1550         {
1551                 .procname       = "mmap_min_addr",
1552                 .data           = &dac_mmap_min_addr,
1553                 .maxlen         = sizeof(unsigned long),
1554                 .mode           = 0644,
1555                 .proc_handler   = mmap_min_addr_handler,
1556         },
1557 #endif
1558 #ifdef CONFIG_NUMA
1559         {
1560                 .procname       = "numa_zonelist_order",
1561                 .data           = &numa_zonelist_order,
1562                 .maxlen         = NUMA_ZONELIST_ORDER_LEN,
1563                 .mode           = 0644,
1564                 .proc_handler   = numa_zonelist_order_handler,
1565         },
1566 #endif
1567 #if (defined(CONFIG_X86_32) && !defined(CONFIG_UML))|| \
1568    (defined(CONFIG_SUPERH) && defined(CONFIG_VSYSCALL))
1569         {
1570                 .procname       = "vdso_enabled",
1571 #ifdef CONFIG_X86_32
1572                 .data           = &vdso32_enabled,
1573                 .maxlen         = sizeof(vdso32_enabled),
1574 #else
1575                 .data           = &vdso_enabled,
1576                 .maxlen         = sizeof(vdso_enabled),
1577 #endif
1578                 .mode           = 0644,
1579                 .proc_handler   = proc_dointvec,
1580                 .extra1         = &zero,
1581         },
1582 #endif
1583 #ifdef CONFIG_HIGHMEM
1584         {
1585                 .procname       = "highmem_is_dirtyable",
1586                 .data           = &vm_highmem_is_dirtyable,
1587                 .maxlen         = sizeof(vm_highmem_is_dirtyable),
1588                 .mode           = 0644,
1589                 .proc_handler   = proc_dointvec_minmax,
1590                 .extra1         = &zero,
1591                 .extra2         = &one,
1592         },
1593 #endif
1594 #ifdef CONFIG_MEMORY_FAILURE
1595         {
1596                 .procname       = "memory_failure_early_kill",
1597                 .data           = &sysctl_memory_failure_early_kill,
1598                 .maxlen         = sizeof(sysctl_memory_failure_early_kill),
1599                 .mode           = 0644,
1600                 .proc_handler   = proc_dointvec_minmax,
1601                 .extra1         = &zero,
1602                 .extra2         = &one,
1603         },
1604         {
1605                 .procname       = "memory_failure_recovery",
1606                 .data           = &sysctl_memory_failure_recovery,
1607                 .maxlen         = sizeof(sysctl_memory_failure_recovery),
1608                 .mode           = 0644,
1609                 .proc_handler   = proc_dointvec_minmax,
1610                 .extra1         = &zero,
1611                 .extra2         = &one,
1612         },
1613 #endif
1614         {
1615                 .procname       = "user_reserve_kbytes",
1616                 .data           = &sysctl_user_reserve_kbytes,
1617                 .maxlen         = sizeof(sysctl_user_reserve_kbytes),
1618                 .mode           = 0644,
1619                 .proc_handler   = proc_doulongvec_minmax,
1620         },
1621         {
1622                 .procname       = "admin_reserve_kbytes",
1623                 .data           = &sysctl_admin_reserve_kbytes,
1624                 .maxlen         = sizeof(sysctl_admin_reserve_kbytes),
1625                 .mode           = 0644,
1626                 .proc_handler   = proc_doulongvec_minmax,
1627         },
1628 #ifdef CONFIG_HAVE_ARCH_MMAP_RND_BITS
1629         {
1630                 .procname       = "mmap_rnd_bits",
1631                 .data           = &mmap_rnd_bits,
1632                 .maxlen         = sizeof(mmap_rnd_bits),
1633                 .mode           = 0600,
1634                 .proc_handler   = proc_dointvec_minmax,
1635                 .extra1         = (void *)&mmap_rnd_bits_min,
1636                 .extra2         = (void *)&mmap_rnd_bits_max,
1637         },
1638 #endif
1639 #ifdef CONFIG_HAVE_ARCH_MMAP_RND_COMPAT_BITS
1640         {
1641                 .procname       = "mmap_rnd_compat_bits",
1642                 .data           = &mmap_rnd_compat_bits,
1643                 .maxlen         = sizeof(mmap_rnd_compat_bits),
1644                 .mode           = 0600,
1645                 .proc_handler   = proc_dointvec_minmax,
1646                 .extra1         = (void *)&mmap_rnd_compat_bits_min,
1647                 .extra2         = (void *)&mmap_rnd_compat_bits_max,
1648         },
1649 #endif
1650         { }
1651 };
1652
1653 static struct ctl_table fs_table[] = {
1654         {
1655                 .procname       = "inode-nr",
1656                 .data           = &inodes_stat,
1657                 .maxlen         = 2*sizeof(long),
1658                 .mode           = 0444,
1659                 .proc_handler   = proc_nr_inodes,
1660         },
1661         {
1662                 .procname       = "inode-state",
1663                 .data           = &inodes_stat,
1664                 .maxlen         = 7*sizeof(long),
1665                 .mode           = 0444,
1666                 .proc_handler   = proc_nr_inodes,
1667         },
1668         {
1669                 .procname       = "file-nr",
1670                 .data           = &files_stat,
1671                 .maxlen         = sizeof(files_stat),
1672                 .mode           = 0444,
1673                 .proc_handler   = proc_nr_files,
1674         },
1675         {
1676                 .procname       = "file-max",
1677                 .data           = &files_stat.max_files,
1678                 .maxlen         = sizeof(files_stat.max_files),
1679                 .mode           = 0644,
1680                 .proc_handler   = proc_doulongvec_minmax,
1681         },
1682         {
1683                 .procname       = "nr_open",
1684                 .data           = &sysctl_nr_open,
1685                 .maxlen         = sizeof(unsigned int),
1686                 .mode           = 0644,
1687                 .proc_handler   = proc_dointvec_minmax,
1688                 .extra1         = &sysctl_nr_open_min,
1689                 .extra2         = &sysctl_nr_open_max,
1690         },
1691         {
1692                 .procname       = "dentry-state",
1693                 .data           = &dentry_stat,
1694                 .maxlen         = 6*sizeof(long),
1695                 .mode           = 0444,
1696                 .proc_handler   = proc_nr_dentry,
1697         },
1698         {
1699                 .procname       = "overflowuid",
1700                 .data           = &fs_overflowuid,
1701                 .maxlen         = sizeof(int),
1702                 .mode           = 0644,
1703                 .proc_handler   = proc_dointvec_minmax,
1704                 .extra1         = &minolduid,
1705                 .extra2         = &maxolduid,
1706         },
1707         {
1708                 .procname       = "overflowgid",
1709                 .data           = &fs_overflowgid,
1710                 .maxlen         = sizeof(int),
1711                 .mode           = 0644,
1712                 .proc_handler   = proc_dointvec_minmax,
1713                 .extra1         = &minolduid,
1714                 .extra2         = &maxolduid,
1715         },
1716 #ifdef CONFIG_FILE_LOCKING
1717         {
1718                 .procname       = "leases-enable",
1719                 .data           = &leases_enable,
1720                 .maxlen         = sizeof(int),
1721                 .mode           = 0644,
1722                 .proc_handler   = proc_dointvec,
1723         },
1724 #endif
1725 #ifdef CONFIG_DNOTIFY
1726         {
1727                 .procname       = "dir-notify-enable",
1728                 .data           = &dir_notify_enable,
1729                 .maxlen         = sizeof(int),
1730                 .mode           = 0644,
1731                 .proc_handler   = proc_dointvec,
1732         },
1733 #endif
1734 #ifdef CONFIG_MMU
1735 #ifdef CONFIG_FILE_LOCKING
1736         {
1737                 .procname       = "lease-break-time",
1738                 .data           = &lease_break_time,
1739                 .maxlen         = sizeof(int),
1740                 .mode           = 0644,
1741                 .proc_handler   = proc_dointvec,
1742         },
1743 #endif
1744 #ifdef CONFIG_AIO
1745         {
1746                 .procname       = "aio-nr",
1747                 .data           = &aio_nr,
1748                 .maxlen         = sizeof(aio_nr),
1749                 .mode           = 0444,
1750                 .proc_handler   = proc_doulongvec_minmax,
1751         },
1752         {
1753                 .procname       = "aio-max-nr",
1754                 .data           = &aio_max_nr,
1755                 .maxlen         = sizeof(aio_max_nr),
1756                 .mode           = 0644,
1757                 .proc_handler   = proc_doulongvec_minmax,
1758         },
1759 #endif /* CONFIG_AIO */
1760 #ifdef CONFIG_INOTIFY_USER
1761         {
1762                 .procname       = "inotify",
1763                 .mode           = 0555,
1764                 .child          = inotify_table,
1765         },
1766 #endif
1767 #ifdef CONFIG_EPOLL
1768         {
1769                 .procname       = "epoll",
1770                 .mode           = 0555,
1771                 .child          = epoll_table,
1772         },
1773 #endif
1774 #endif
1775         {
1776                 .procname       = "protected_symlinks",
1777                 .data           = &sysctl_protected_symlinks,
1778                 .maxlen         = sizeof(int),
1779                 .mode           = 0600,
1780                 .proc_handler   = proc_dointvec_minmax,
1781                 .extra1         = &zero,
1782                 .extra2         = &one,
1783         },
1784         {
1785                 .procname       = "protected_hardlinks",
1786                 .data           = &sysctl_protected_hardlinks,
1787                 .maxlen         = sizeof(int),
1788                 .mode           = 0600,
1789                 .proc_handler   = proc_dointvec_minmax,
1790                 .extra1         = &zero,
1791                 .extra2         = &one,
1792         },
1793         {
1794                 .procname       = "suid_dumpable",
1795                 .data           = &suid_dumpable,
1796                 .maxlen         = sizeof(int),
1797                 .mode           = 0644,
1798                 .proc_handler   = proc_dointvec_minmax_coredump,
1799                 .extra1         = &zero,
1800                 .extra2         = &two,
1801         },
1802 #if defined(CONFIG_BINFMT_MISC) || defined(CONFIG_BINFMT_MISC_MODULE)
1803         {
1804                 .procname       = "binfmt_misc",
1805                 .mode           = 0555,
1806                 .child          = sysctl_mount_point,
1807         },
1808 #endif
1809         {
1810                 .procname       = "pipe-max-size",
1811                 .data           = &pipe_max_size,
1812                 .maxlen         = sizeof(int),
1813                 .mode           = 0644,
1814                 .proc_handler   = &pipe_proc_fn,
1815                 .extra1         = &pipe_min_size,
1816         },
1817         {
1818                 .procname       = "pipe-user-pages-hard",
1819                 .data           = &pipe_user_pages_hard,
1820                 .maxlen         = sizeof(pipe_user_pages_hard),
1821                 .mode           = 0644,
1822                 .proc_handler   = proc_doulongvec_minmax,
1823         },
1824         {
1825                 .procname       = "pipe-user-pages-soft",
1826                 .data           = &pipe_user_pages_soft,
1827                 .maxlen         = sizeof(pipe_user_pages_soft),
1828                 .mode           = 0644,
1829                 .proc_handler   = proc_doulongvec_minmax,
1830         },
1831         { }
1832 };
1833
1834 static struct ctl_table debug_table[] = {
1835 #ifdef CONFIG_SYSCTL_EXCEPTION_TRACE
1836         {
1837                 .procname       = "exception-trace",
1838                 .data           = &show_unhandled_signals,
1839                 .maxlen         = sizeof(int),
1840                 .mode           = 0644,
1841                 .proc_handler   = proc_dointvec
1842         },
1843 #endif
1844 #if defined(CONFIG_OPTPROBES)
1845         {
1846                 .procname       = "kprobes-optimization",
1847                 .data           = &sysctl_kprobes_optimization,
1848                 .maxlen         = sizeof(int),
1849                 .mode           = 0644,
1850                 .proc_handler   = proc_kprobes_optimization_handler,
1851                 .extra1         = &zero,
1852                 .extra2         = &one,
1853         },
1854 #endif
1855         { }
1856 };
1857
1858 static struct ctl_table dev_table[] = {
1859         { }
1860 };
1861
1862 int __init sysctl_init(void)
1863 {
1864         struct ctl_table_header *hdr;
1865
1866         hdr = register_sysctl_table(sysctl_base_table);
1867         kmemleak_not_leak(hdr);
1868         return 0;
1869 }
1870
1871 #endif /* CONFIG_SYSCTL */
1872
1873 /*
1874  * /proc/sys support
1875  */
1876
1877 #ifdef CONFIG_PROC_SYSCTL
1878
1879 static int _proc_do_string(char *data, int maxlen, int write,
1880                            char __user *buffer,
1881                            size_t *lenp, loff_t *ppos)
1882 {
1883         size_t len;
1884         char __user *p;
1885         char c;
1886
1887         if (!data || !maxlen || !*lenp) {
1888                 *lenp = 0;
1889                 return 0;
1890         }
1891
1892         if (write) {
1893                 if (sysctl_writes_strict == SYSCTL_WRITES_STRICT) {
1894                         /* Only continue writes not past the end of buffer. */
1895                         len = strlen(data);
1896                         if (len > maxlen - 1)
1897                                 len = maxlen - 1;
1898
1899                         if (*ppos > len)
1900                                 return 0;
1901                         len = *ppos;
1902                 } else {
1903                         /* Start writing from beginning of buffer. */
1904                         len = 0;
1905                 }
1906
1907                 *ppos += *lenp;
1908                 p = buffer;
1909                 while ((p - buffer) < *lenp && len < maxlen - 1) {
1910                         if (get_user(c, p++))
1911                                 return -EFAULT;
1912                         if (c == 0 || c == '\n')
1913                                 break;
1914                         data[len++] = c;
1915                 }
1916                 data[len] = 0;
1917         } else {
1918                 len = strlen(data);
1919                 if (len > maxlen)
1920                         len = maxlen;
1921
1922                 if (*ppos > len) {
1923                         *lenp = 0;
1924                         return 0;
1925                 }
1926
1927                 data += *ppos;
1928                 len  -= *ppos;
1929
1930                 if (len > *lenp)
1931                         len = *lenp;
1932                 if (len)
1933                         if (copy_to_user(buffer, data, len))
1934                                 return -EFAULT;
1935                 if (len < *lenp) {
1936                         if (put_user('\n', buffer + len))
1937                                 return -EFAULT;
1938                         len++;
1939                 }
1940                 *lenp = len;
1941                 *ppos += len;
1942         }
1943         return 0;
1944 }
1945
1946 static void warn_sysctl_write(struct ctl_table *table)
1947 {
1948         pr_warn_once("%s wrote to %s when file position was not 0!\n"
1949                 "This will not be supported in the future. To silence this\n"
1950                 "warning, set kernel.sysctl_writes_strict = -1\n",
1951                 current->comm, table->procname);
1952 }
1953
1954 /**
1955  * proc_dostring - read a string sysctl
1956  * @table: the sysctl table
1957  * @write: %TRUE if this is a write to the sysctl file
1958  * @buffer: the user buffer
1959  * @lenp: the size of the user buffer
1960  * @ppos: file position
1961  *
1962  * Reads/writes a string from/to the user buffer. If the kernel
1963  * buffer provided is not large enough to hold the string, the
1964  * string is truncated. The copied string is %NULL-terminated.
1965  * If the string is being read by the user process, it is copied
1966  * and a newline '\n' is added. It is truncated if the buffer is
1967  * not large enough.
1968  *
1969  * Returns 0 on success.
1970  */
1971 int proc_dostring(struct ctl_table *table, int write,
1972                   void __user *buffer, size_t *lenp, loff_t *ppos)
1973 {
1974         if (write && *ppos && sysctl_writes_strict == SYSCTL_WRITES_WARN)
1975                 warn_sysctl_write(table);
1976
1977         return _proc_do_string((char *)(table->data), table->maxlen, write,
1978                                (char __user *)buffer, lenp, ppos);
1979 }
1980
1981 static size_t proc_skip_spaces(char **buf)
1982 {
1983         size_t ret;
1984         char *tmp = skip_spaces(*buf);
1985         ret = tmp - *buf;
1986         *buf = tmp;
1987         return ret;
1988 }
1989
1990 static void proc_skip_char(char **buf, size_t *size, const char v)
1991 {
1992         while (*size) {
1993                 if (**buf != v)
1994                         break;
1995                 (*size)--;
1996                 (*buf)++;
1997         }
1998 }
1999
2000 #define TMPBUFLEN 22
2001 /**
2002  * proc_get_long - reads an ASCII formatted integer from a user buffer
2003  *
2004  * @buf: a kernel buffer
2005  * @size: size of the kernel buffer
2006  * @val: this is where the number will be stored
2007  * @neg: set to %TRUE if number is negative
2008  * @perm_tr: a vector which contains the allowed trailers
2009  * @perm_tr_len: size of the perm_tr vector
2010  * @tr: pointer to store the trailer character
2011  *
2012  * In case of success %0 is returned and @buf and @size are updated with
2013  * the amount of bytes read. If @tr is non-NULL and a trailing
2014  * character exists (size is non-zero after returning from this
2015  * function), @tr is updated with the trailing character.
2016  */
2017 static int proc_get_long(char **buf, size_t *size,
2018                           unsigned long *val, bool *neg,
2019                           const char *perm_tr, unsigned perm_tr_len, char *tr)
2020 {
2021         int len;
2022         char *p, tmp[TMPBUFLEN];
2023
2024         if (!*size)
2025                 return -EINVAL;
2026
2027         len = *size;
2028         if (len > TMPBUFLEN - 1)
2029                 len = TMPBUFLEN - 1;
2030
2031         memcpy(tmp, *buf, len);
2032
2033         tmp[len] = 0;
2034         p = tmp;
2035         if (*p == '-' && *size > 1) {
2036                 *neg = true;
2037                 p++;
2038         } else
2039                 *neg = false;
2040         if (!isdigit(*p))
2041                 return -EINVAL;
2042
2043         *val = simple_strtoul(p, &p, 0);
2044
2045         len = p - tmp;
2046
2047         /* We don't know if the next char is whitespace thus we may accept
2048          * invalid integers (e.g. 1234...a) or two integers instead of one
2049          * (e.g. 123...1). So lets not allow such large numbers. */
2050         if (len == TMPBUFLEN - 1)
2051                 return -EINVAL;
2052
2053         if (len < *size && perm_tr_len && !memchr(perm_tr, *p, perm_tr_len))
2054                 return -EINVAL;
2055
2056         if (tr && (len < *size))
2057                 *tr = *p;
2058
2059         *buf += len;
2060         *size -= len;
2061
2062         return 0;
2063 }
2064
2065 /**
2066  * proc_put_long - converts an integer to a decimal ASCII formatted string
2067  *
2068  * @buf: the user buffer
2069  * @size: the size of the user buffer
2070  * @val: the integer to be converted
2071  * @neg: sign of the number, %TRUE for negative
2072  *
2073  * In case of success %0 is returned and @buf and @size are updated with
2074  * the amount of bytes written.
2075  */
2076 static int proc_put_long(void __user **buf, size_t *size, unsigned long val,
2077                           bool neg)
2078 {
2079         int len;
2080         char tmp[TMPBUFLEN], *p = tmp;
2081
2082         sprintf(p, "%s%lu", neg ? "-" : "", val);
2083         len = strlen(tmp);
2084         if (len > *size)
2085                 len = *size;
2086         if (copy_to_user(*buf, tmp, len))
2087                 return -EFAULT;
2088         *size -= len;
2089         *buf += len;
2090         return 0;
2091 }
2092 #undef TMPBUFLEN
2093
2094 static int proc_put_char(void __user **buf, size_t *size, char c)
2095 {
2096         if (*size) {
2097                 char __user **buffer = (char __user **)buf;
2098                 if (put_user(c, *buffer))
2099                         return -EFAULT;
2100                 (*size)--, (*buffer)++;
2101                 *buf = *buffer;
2102         }
2103         return 0;
2104 }
2105
2106 static int do_proc_dointvec_conv(bool *negp, unsigned long *lvalp,
2107                                  int *valp,
2108                                  int write, void *data)
2109 {
2110         if (write) {
2111                 if (*negp) {
2112                         if (*lvalp > (unsigned long) INT_MAX + 1)
2113                                 return -EINVAL;
2114                         *valp = -*lvalp;
2115                 } else {
2116                         if (*lvalp > (unsigned long) INT_MAX)
2117                                 return -EINVAL;
2118                         *valp = *lvalp;
2119                 }
2120         } else {
2121                 int val = *valp;
2122                 if (val < 0) {
2123                         *negp = true;
2124                         *lvalp = -(unsigned long)val;
2125                 } else {
2126                         *negp = false;
2127                         *lvalp = (unsigned long)val;
2128                 }
2129         }
2130         return 0;
2131 }
2132
2133 static int do_proc_douintvec_conv(bool *negp, unsigned long *lvalp,
2134                                  int *valp,
2135                                  int write, void *data)
2136 {
2137         if (write) {
2138                 if (*negp)
2139                         return -EINVAL;
2140                 *valp = *lvalp;
2141         } else {
2142                 unsigned int val = *valp;
2143                 *lvalp = (unsigned long)val;
2144         }
2145         return 0;
2146 }
2147
2148 static const char proc_wspace_sep[] = { ' ', '\t', '\n' };
2149
2150 static int __do_proc_dointvec(void *tbl_data, struct ctl_table *table,
2151                   int write, void __user *buffer,
2152                   size_t *lenp, loff_t *ppos,
2153                   int (*conv)(bool *negp, unsigned long *lvalp, int *valp,
2154                               int write, void *data),
2155                   void *data)
2156 {
2157         int *i, vleft, first = 1, err = 0;
2158         size_t left;
2159         char *kbuf = NULL, *p;
2160
2161         if (!tbl_data || !table->maxlen || !*lenp || (*ppos && !write)) {
2162                 *lenp = 0;
2163                 return 0;
2164         }
2165
2166         i = (int *) tbl_data;
2167         vleft = table->maxlen / sizeof(*i);
2168         left = *lenp;
2169
2170         if (!conv)
2171                 conv = do_proc_dointvec_conv;
2172
2173         if (write) {
2174                 if (*ppos) {
2175                         switch (sysctl_writes_strict) {
2176                         case SYSCTL_WRITES_STRICT:
2177                                 goto out;
2178                         case SYSCTL_WRITES_WARN:
2179                                 warn_sysctl_write(table);
2180                                 break;
2181                         default:
2182                                 break;
2183                         }
2184                 }
2185
2186                 if (left > PAGE_SIZE - 1)
2187                         left = PAGE_SIZE - 1;
2188                 p = kbuf = memdup_user_nul(buffer, left);
2189                 if (IS_ERR(kbuf))
2190                         return PTR_ERR(kbuf);
2191         }
2192
2193         for (; left && vleft--; i++, first=0) {
2194                 unsigned long lval;
2195                 bool neg;
2196
2197                 if (write) {
2198                         left -= proc_skip_spaces(&p);
2199
2200                         if (!left)
2201                                 break;
2202                         err = proc_get_long(&p, &left, &lval, &neg,
2203                                              proc_wspace_sep,
2204                                              sizeof(proc_wspace_sep), NULL);
2205                         if (err)
2206                                 break;
2207                         if (conv(&neg, &lval, i, 1, data)) {
2208                                 err = -EINVAL;
2209                                 break;
2210                         }
2211                 } else {
2212                         if (conv(&neg, &lval, i, 0, data)) {
2213                                 err = -EINVAL;
2214                                 break;
2215                         }
2216                         if (!first)
2217                                 err = proc_put_char(&buffer, &left, '\t');
2218                         if (err)
2219                                 break;
2220                         err = proc_put_long(&buffer, &left, lval, neg);
2221                         if (err)
2222                                 break;
2223                 }
2224         }
2225
2226         if (!write && !first && left && !err)
2227                 err = proc_put_char(&buffer, &left, '\n');
2228         if (write && !err && left)
2229                 left -= proc_skip_spaces(&p);
2230         if (write) {
2231                 kfree(kbuf);
2232                 if (first)
2233                         return err ? : -EINVAL;
2234         }
2235         *lenp -= left;
2236 out:
2237         *ppos += *lenp;
2238         return err;
2239 }
2240
2241 static int do_proc_dointvec(struct ctl_table *table, int write,
2242                   void __user *buffer, size_t *lenp, loff_t *ppos,
2243                   int (*conv)(bool *negp, unsigned long *lvalp, int *valp,
2244                               int write, void *data),
2245                   void *data)
2246 {
2247         return __do_proc_dointvec(table->data, table, write,
2248                         buffer, lenp, ppos, conv, data);
2249 }
2250
2251 /**
2252  * proc_dointvec - read a vector of integers
2253  * @table: the sysctl table
2254  * @write: %TRUE if this is a write to the sysctl file
2255  * @buffer: the user buffer
2256  * @lenp: the size of the user buffer
2257  * @ppos: file position
2258  *
2259  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2260  * values from/to the user buffer, treated as an ASCII string.
2261  *
2262  * Returns 0 on success.
2263  */
2264 int proc_dointvec(struct ctl_table *table, int write,
2265                      void __user *buffer, size_t *lenp, loff_t *ppos)
2266 {
2267         return do_proc_dointvec(table, write, buffer, lenp, ppos, NULL, NULL);
2268 }
2269
2270 /**
2271  * proc_douintvec - read a vector of unsigned integers
2272  * @table: the sysctl table
2273  * @write: %TRUE if this is a write to the sysctl file
2274  * @buffer: the user buffer
2275  * @lenp: the size of the user buffer
2276  * @ppos: file position
2277  *
2278  * Reads/writes up to table->maxlen/sizeof(unsigned int) unsigned integer
2279  * values from/to the user buffer, treated as an ASCII string.
2280  *
2281  * Returns 0 on success.
2282  */
2283 int proc_douintvec(struct ctl_table *table, int write,
2284                      void __user *buffer, size_t *lenp, loff_t *ppos)
2285 {
2286         return do_proc_dointvec(table, write, buffer, lenp, ppos,
2287                                 do_proc_douintvec_conv, NULL);
2288 }
2289
2290 /*
2291  * Taint values can only be increased
2292  * This means we can safely use a temporary.
2293  */
2294 static int proc_taint(struct ctl_table *table, int write,
2295                                void __user *buffer, size_t *lenp, loff_t *ppos)
2296 {
2297         struct ctl_table t;
2298         unsigned long tmptaint = get_taint();
2299         int err;
2300
2301         if (write && !capable(CAP_SYS_ADMIN))
2302                 return -EPERM;
2303
2304         t = *table;
2305         t.data = &tmptaint;
2306         err = proc_doulongvec_minmax(&t, write, buffer, lenp, ppos);
2307         if (err < 0)
2308                 return err;
2309
2310         if (write) {
2311                 /*
2312                  * Poor man's atomic or. Not worth adding a primitive
2313                  * to everyone's atomic.h for this
2314                  */
2315                 int i;
2316                 for (i = 0; i < BITS_PER_LONG && tmptaint >> i; i++) {
2317                         if ((tmptaint >> i) & 1)
2318                                 add_taint(i, LOCKDEP_STILL_OK);
2319                 }
2320         }
2321
2322         return err;
2323 }
2324
2325 #ifdef CONFIG_PRINTK
2326 static int proc_dointvec_minmax_sysadmin(struct ctl_table *table, int write,
2327                                 void __user *buffer, size_t *lenp, loff_t *ppos)
2328 {
2329         if (write && !capable(CAP_SYS_ADMIN))
2330                 return -EPERM;
2331
2332         return proc_dointvec_minmax(table, write, buffer, lenp, ppos);
2333 }
2334 #endif
2335
2336 struct do_proc_dointvec_minmax_conv_param {
2337         int *min;
2338         int *max;
2339 };
2340
2341 static int do_proc_dointvec_minmax_conv(bool *negp, unsigned long *lvalp,
2342                                         int *valp,
2343                                         int write, void *data)
2344 {
2345         struct do_proc_dointvec_minmax_conv_param *param = data;
2346         if (write) {
2347                 int val = *negp ? -*lvalp : *lvalp;
2348                 if ((param->min && *param->min > val) ||
2349                     (param->max && *param->max < val))
2350                         return -EINVAL;
2351                 *valp = val;
2352         } else {
2353                 int val = *valp;
2354                 if (val < 0) {
2355                         *negp = true;
2356                         *lvalp = -(unsigned long)val;
2357                 } else {
2358                         *negp = false;
2359                         *lvalp = (unsigned long)val;
2360                 }
2361         }
2362         return 0;
2363 }
2364
2365 /**
2366  * proc_dointvec_minmax - read a vector of integers with min/max values
2367  * @table: the sysctl table
2368  * @write: %TRUE if this is a write to the sysctl file
2369  * @buffer: the user buffer
2370  * @lenp: the size of the user buffer
2371  * @ppos: file position
2372  *
2373  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2374  * values from/to the user buffer, treated as an ASCII string.
2375  *
2376  * This routine will ensure the values are within the range specified by
2377  * table->extra1 (min) and table->extra2 (max).
2378  *
2379  * Returns 0 on success.
2380  */
2381 int proc_dointvec_minmax(struct ctl_table *table, int write,
2382                   void __user *buffer, size_t *lenp, loff_t *ppos)
2383 {
2384         struct do_proc_dointvec_minmax_conv_param param = {
2385                 .min = (int *) table->extra1,
2386                 .max = (int *) table->extra2,
2387         };
2388         return do_proc_dointvec(table, write, buffer, lenp, ppos,
2389                                 do_proc_dointvec_minmax_conv, &param);
2390 }
2391
2392 static void validate_coredump_safety(void)
2393 {
2394 #ifdef CONFIG_COREDUMP
2395         if (suid_dumpable == SUID_DUMP_ROOT &&
2396             core_pattern[0] != '/' && core_pattern[0] != '|') {
2397                 printk(KERN_WARNING "Unsafe core_pattern used with "\
2398                         "suid_dumpable=2. Pipe handler or fully qualified "\
2399                         "core dump path required.\n");
2400         }
2401 #endif
2402 }
2403
2404 static int proc_dointvec_minmax_coredump(struct ctl_table *table, int write,
2405                 void __user *buffer, size_t *lenp, loff_t *ppos)
2406 {
2407         int error = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
2408         if (!error)
2409                 validate_coredump_safety();
2410         return error;
2411 }
2412
2413 #ifdef CONFIG_COREDUMP
2414 static int proc_dostring_coredump(struct ctl_table *table, int write,
2415                   void __user *buffer, size_t *lenp, loff_t *ppos)
2416 {
2417         int error = proc_dostring(table, write, buffer, lenp, ppos);
2418         if (!error)
2419                 validate_coredump_safety();
2420         return error;
2421 }
2422 #endif
2423
2424 static int __do_proc_doulongvec_minmax(void *data, struct ctl_table *table, int write,
2425                                      void __user *buffer,
2426                                      size_t *lenp, loff_t *ppos,
2427                                      unsigned long convmul,
2428                                      unsigned long convdiv)
2429 {
2430         unsigned long *i, *min, *max;
2431         int vleft, first = 1, err = 0;
2432         size_t left;
2433         char *kbuf = NULL, *p;
2434
2435         if (!data || !table->maxlen || !*lenp || (*ppos && !write)) {
2436                 *lenp = 0;
2437                 return 0;
2438         }
2439
2440         i = (unsigned long *) data;
2441         min = (unsigned long *) table->extra1;
2442         max = (unsigned long *) table->extra2;
2443         vleft = table->maxlen / sizeof(unsigned long);
2444         left = *lenp;
2445
2446         if (write) {
2447                 if (*ppos) {
2448                         switch (sysctl_writes_strict) {
2449                         case SYSCTL_WRITES_STRICT:
2450                                 goto out;
2451                         case SYSCTL_WRITES_WARN:
2452                                 warn_sysctl_write(table);
2453                                 break;
2454                         default:
2455                                 break;
2456                         }
2457                 }
2458
2459                 if (left > PAGE_SIZE - 1)
2460                         left = PAGE_SIZE - 1;
2461                 p = kbuf = memdup_user_nul(buffer, left);
2462                 if (IS_ERR(kbuf))
2463                         return PTR_ERR(kbuf);
2464         }
2465
2466         for (; left && vleft--; i++, first = 0) {
2467                 unsigned long val;
2468
2469                 if (write) {
2470                         bool neg;
2471
2472                         left -= proc_skip_spaces(&p);
2473
2474                         err = proc_get_long(&p, &left, &val, &neg,
2475                                              proc_wspace_sep,
2476                                              sizeof(proc_wspace_sep), NULL);
2477                         if (err)
2478                                 break;
2479                         if (neg)
2480                                 continue;
2481                         if ((min && val < *min) || (max && val > *max))
2482                                 continue;
2483                         *i = val;
2484                 } else {
2485                         val = convdiv * (*i) / convmul;
2486                         if (!first) {
2487                                 err = proc_put_char(&buffer, &left, '\t');
2488                                 if (err)
2489                                         break;
2490                         }
2491                         err = proc_put_long(&buffer, &left, val, false);
2492                         if (err)
2493                                 break;
2494                 }
2495         }
2496
2497         if (!write && !first && left && !err)
2498                 err = proc_put_char(&buffer, &left, '\n');
2499         if (write && !err)
2500                 left -= proc_skip_spaces(&p);
2501         if (write) {
2502                 kfree(kbuf);
2503                 if (first)
2504                         return err ? : -EINVAL;
2505         }
2506         *lenp -= left;
2507 out:
2508         *ppos += *lenp;
2509         return err;
2510 }
2511
2512 static int do_proc_doulongvec_minmax(struct ctl_table *table, int write,
2513                                      void __user *buffer,
2514                                      size_t *lenp, loff_t *ppos,
2515                                      unsigned long convmul,
2516                                      unsigned long convdiv)
2517 {
2518         return __do_proc_doulongvec_minmax(table->data, table, write,
2519                         buffer, lenp, ppos, convmul, convdiv);
2520 }
2521
2522 /**
2523  * proc_doulongvec_minmax - read a vector of long integers with min/max values
2524  * @table: the sysctl table
2525  * @write: %TRUE if this is a write to the sysctl file
2526  * @buffer: the user buffer
2527  * @lenp: the size of the user buffer
2528  * @ppos: file position
2529  *
2530  * Reads/writes up to table->maxlen/sizeof(unsigned long) unsigned long
2531  * values from/to the user buffer, treated as an ASCII string.
2532  *
2533  * This routine will ensure the values are within the range specified by
2534  * table->extra1 (min) and table->extra2 (max).
2535  *
2536  * Returns 0 on success.
2537  */
2538 int proc_doulongvec_minmax(struct ctl_table *table, int write,
2539                            void __user *buffer, size_t *lenp, loff_t *ppos)
2540 {
2541     return do_proc_doulongvec_minmax(table, write, buffer, lenp, ppos, 1l, 1l);
2542 }
2543
2544 /**
2545  * proc_doulongvec_ms_jiffies_minmax - read a vector of millisecond values with min/max values
2546  * @table: the sysctl table
2547  * @write: %TRUE if this is a write to the sysctl file
2548  * @buffer: the user buffer
2549  * @lenp: the size of the user buffer
2550  * @ppos: file position
2551  *
2552  * Reads/writes up to table->maxlen/sizeof(unsigned long) unsigned long
2553  * values from/to the user buffer, treated as an ASCII string. The values
2554  * are treated as milliseconds, and converted to jiffies when they are stored.
2555  *
2556  * This routine will ensure the values are within the range specified by
2557  * table->extra1 (min) and table->extra2 (max).
2558  *
2559  * Returns 0 on success.
2560  */
2561 int proc_doulongvec_ms_jiffies_minmax(struct ctl_table *table, int write,
2562                                       void __user *buffer,
2563                                       size_t *lenp, loff_t *ppos)
2564 {
2565     return do_proc_doulongvec_minmax(table, write, buffer,
2566                                      lenp, ppos, HZ, 1000l);
2567 }
2568
2569
2570 static int do_proc_dointvec_jiffies_conv(bool *negp, unsigned long *lvalp,
2571                                          int *valp,
2572                                          int write, void *data)
2573 {
2574         if (write) {
2575                 if (*lvalp > LONG_MAX / HZ)
2576                         return 1;
2577                 *valp = *negp ? -(*lvalp*HZ) : (*lvalp*HZ);
2578         } else {
2579                 int val = *valp;
2580                 unsigned long lval;
2581                 if (val < 0) {
2582                         *negp = true;
2583                         lval = -(unsigned long)val;
2584                 } else {
2585                         *negp = false;
2586                         lval = (unsigned long)val;
2587                 }
2588                 *lvalp = lval / HZ;
2589         }
2590         return 0;
2591 }
2592
2593 static int do_proc_dointvec_userhz_jiffies_conv(bool *negp, unsigned long *lvalp,
2594                                                 int *valp,
2595                                                 int write, void *data)
2596 {
2597         if (write) {
2598                 if (USER_HZ < HZ && *lvalp > (LONG_MAX / HZ) * USER_HZ)
2599                         return 1;
2600                 *valp = clock_t_to_jiffies(*negp ? -*lvalp : *lvalp);
2601         } else {
2602                 int val = *valp;
2603                 unsigned long lval;
2604                 if (val < 0) {
2605                         *negp = true;
2606                         lval = -(unsigned long)val;
2607                 } else {
2608                         *negp = false;
2609                         lval = (unsigned long)val;
2610                 }
2611                 *lvalp = jiffies_to_clock_t(lval);
2612         }
2613         return 0;
2614 }
2615
2616 static int do_proc_dointvec_ms_jiffies_conv(bool *negp, unsigned long *lvalp,
2617                                             int *valp,
2618                                             int write, void *data)
2619 {
2620         if (write) {
2621                 unsigned long jif = msecs_to_jiffies(*negp ? -*lvalp : *lvalp);
2622
2623                 if (jif > INT_MAX)
2624                         return 1;
2625                 *valp = (int)jif;
2626         } else {
2627                 int val = *valp;
2628                 unsigned long lval;
2629                 if (val < 0) {
2630                         *negp = true;
2631                         lval = -(unsigned long)val;
2632                 } else {
2633                         *negp = false;
2634                         lval = (unsigned long)val;
2635                 }
2636                 *lvalp = jiffies_to_msecs(lval);
2637         }
2638         return 0;
2639 }
2640
2641 /**
2642  * proc_dointvec_jiffies - read a vector of integers as seconds
2643  * @table: the sysctl table
2644  * @write: %TRUE if this is a write to the sysctl file
2645  * @buffer: the user buffer
2646  * @lenp: the size of the user buffer
2647  * @ppos: file position
2648  *
2649  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2650  * values from/to the user buffer, treated as an ASCII string.
2651  * The values read are assumed to be in seconds, and are converted into
2652  * jiffies.
2653  *
2654  * Returns 0 on success.
2655  */
2656 int proc_dointvec_jiffies(struct ctl_table *table, int write,
2657                           void __user *buffer, size_t *lenp, loff_t *ppos)
2658 {
2659     return do_proc_dointvec(table,write,buffer,lenp,ppos,
2660                             do_proc_dointvec_jiffies_conv,NULL);
2661 }
2662
2663 /**
2664  * proc_dointvec_userhz_jiffies - read a vector of integers as 1/USER_HZ seconds
2665  * @table: the sysctl table
2666  * @write: %TRUE if this is a write to the sysctl file
2667  * @buffer: the user buffer
2668  * @lenp: the size of the user buffer
2669  * @ppos: pointer to the file position
2670  *
2671  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2672  * values from/to the user buffer, treated as an ASCII string.
2673  * The values read are assumed to be in 1/USER_HZ seconds, and
2674  * are converted into jiffies.
2675  *
2676  * Returns 0 on success.
2677  */
2678 int proc_dointvec_userhz_jiffies(struct ctl_table *table, int write,
2679                                  void __user *buffer, size_t *lenp, loff_t *ppos)
2680 {
2681     return do_proc_dointvec(table,write,buffer,lenp,ppos,
2682                             do_proc_dointvec_userhz_jiffies_conv,NULL);
2683 }
2684
2685 /**
2686  * proc_dointvec_ms_jiffies - read a vector of integers as 1 milliseconds
2687  * @table: the sysctl table
2688  * @write: %TRUE if this is a write to the sysctl file
2689  * @buffer: the user buffer
2690  * @lenp: the size of the user buffer
2691  * @ppos: file position
2692  * @ppos: the current position in the file
2693  *
2694  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2695  * values from/to the user buffer, treated as an ASCII string.
2696  * The values read are assumed to be in 1/1000 seconds, and
2697  * are converted into jiffies.
2698  *
2699  * Returns 0 on success.
2700  */
2701 int proc_dointvec_ms_jiffies(struct ctl_table *table, int write,
2702                              void __user *buffer, size_t *lenp, loff_t *ppos)
2703 {
2704         return do_proc_dointvec(table, write, buffer, lenp, ppos,
2705                                 do_proc_dointvec_ms_jiffies_conv, NULL);
2706 }
2707
2708 static int proc_do_cad_pid(struct ctl_table *table, int write,
2709                            void __user *buffer, size_t *lenp, loff_t *ppos)
2710 {
2711         struct pid *new_pid;
2712         pid_t tmp;
2713         int r;
2714
2715         tmp = pid_vnr(cad_pid);
2716
2717         r = __do_proc_dointvec(&tmp, table, write, buffer,
2718                                lenp, ppos, NULL, NULL);
2719         if (r || !write)
2720                 return r;
2721
2722         new_pid = find_get_pid(tmp);
2723         if (!new_pid)
2724                 return -ESRCH;
2725
2726         put_pid(xchg(&cad_pid, new_pid));
2727         return 0;
2728 }
2729
2730 /**
2731  * proc_do_large_bitmap - read/write from/to a large bitmap
2732  * @table: the sysctl table
2733  * @write: %TRUE if this is a write to the sysctl file
2734  * @buffer: the user buffer
2735  * @lenp: the size of the user buffer
2736  * @ppos: file position
2737  *
2738  * The bitmap is stored at table->data and the bitmap length (in bits)
2739  * in table->maxlen.
2740  *
2741  * We use a range comma separated format (e.g. 1,3-4,10-10) so that
2742  * large bitmaps may be represented in a compact manner. Writing into
2743  * the file will clear the bitmap then update it with the given input.
2744  *
2745  * Returns 0 on success.
2746  */
2747 int proc_do_large_bitmap(struct ctl_table *table, int write,
2748                          void __user *buffer, size_t *lenp, loff_t *ppos)
2749 {
2750         int err = 0;
2751         bool first = 1;
2752         size_t left = *lenp;
2753         unsigned long bitmap_len = table->maxlen;
2754         unsigned long *bitmap = *(unsigned long **) table->data;
2755         unsigned long *tmp_bitmap = NULL;
2756         char tr_a[] = { '-', ',', '\n' }, tr_b[] = { ',', '\n', 0 }, c;
2757
2758         if (!bitmap || !bitmap_len || !left || (*ppos && !write)) {
2759                 *lenp = 0;
2760                 return 0;
2761         }
2762
2763         if (write) {
2764                 char *kbuf, *p;
2765
2766                 if (left > PAGE_SIZE - 1)
2767                         left = PAGE_SIZE - 1;
2768
2769                 p = kbuf = memdup_user_nul(buffer, left);
2770                 if (IS_ERR(kbuf))
2771                         return PTR_ERR(kbuf);
2772
2773                 tmp_bitmap = kzalloc(BITS_TO_LONGS(bitmap_len) * sizeof(unsigned long),
2774                                      GFP_KERNEL);
2775                 if (!tmp_bitmap) {
2776                         kfree(kbuf);
2777                         return -ENOMEM;
2778                 }
2779                 proc_skip_char(&p, &left, '\n');
2780                 while (!err && left) {
2781                         unsigned long val_a, val_b;
2782                         bool neg;
2783
2784                         err = proc_get_long(&p, &left, &val_a, &neg, tr_a,
2785                                              sizeof(tr_a), &c);
2786                         if (err)
2787                                 break;
2788                         if (val_a >= bitmap_len || neg) {
2789                                 err = -EINVAL;
2790                                 break;
2791                         }
2792
2793                         val_b = val_a;
2794                         if (left) {
2795                                 p++;
2796                                 left--;
2797                         }
2798
2799                         if (c == '-') {
2800                                 err = proc_get_long(&p, &left, &val_b,
2801                                                      &neg, tr_b, sizeof(tr_b),
2802                                                      &c);
2803                                 if (err)
2804                                         break;
2805                                 if (val_b >= bitmap_len || neg ||
2806                                     val_a > val_b) {
2807                                         err = -EINVAL;
2808                                         break;
2809                                 }
2810                                 if (left) {
2811                                         p++;
2812                                         left--;
2813                                 }
2814                         }
2815
2816                         bitmap_set(tmp_bitmap, val_a, val_b - val_a + 1);
2817                         first = 0;
2818                         proc_skip_char(&p, &left, '\n');
2819                 }
2820                 kfree(kbuf);
2821         } else {
2822                 unsigned long bit_a, bit_b = 0;
2823
2824                 while (left) {
2825                         bit_a = find_next_bit(bitmap, bitmap_len, bit_b);
2826                         if (bit_a >= bitmap_len)
2827                                 break;
2828                         bit_b = find_next_zero_bit(bitmap, bitmap_len,
2829                                                    bit_a + 1) - 1;
2830
2831                         if (!first) {
2832                                 err = proc_put_char(&buffer, &left, ',');
2833                                 if (err)
2834                                         break;
2835                         }
2836                         err = proc_put_long(&buffer, &left, bit_a, false);
2837                         if (err)
2838                                 break;
2839                         if (bit_a != bit_b) {
2840                                 err = proc_put_char(&buffer, &left, '-');
2841                                 if (err)
2842                                         break;
2843                                 err = proc_put_long(&buffer, &left, bit_b, false);
2844                                 if (err)
2845                                         break;
2846                         }
2847
2848                         first = 0; bit_b++;
2849                 }
2850                 if (!err)
2851                         err = proc_put_char(&buffer, &left, '\n');
2852         }
2853
2854         if (!err) {
2855                 if (write) {
2856                         if (*ppos)
2857                                 bitmap_or(bitmap, bitmap, tmp_bitmap, bitmap_len);
2858                         else
2859                                 bitmap_copy(bitmap, tmp_bitmap, bitmap_len);
2860                 }
2861                 kfree(tmp_bitmap);
2862                 *lenp -= left;
2863                 *ppos += *lenp;
2864                 return 0;
2865         } else {
2866                 kfree(tmp_bitmap);
2867                 return err;
2868         }
2869 }
2870
2871 #else /* CONFIG_PROC_SYSCTL */
2872
2873 int proc_dostring(struct ctl_table *table, int write,
2874                   void __user *buffer, size_t *lenp, loff_t *ppos)
2875 {
2876         return -ENOSYS;
2877 }
2878
2879 int proc_dointvec(struct ctl_table *table, int write,
2880                   void __user *buffer, size_t *lenp, loff_t *ppos)
2881 {
2882         return -ENOSYS;
2883 }
2884
2885 int proc_douintvec(struct ctl_table *table, int write,
2886                   void __user *buffer, size_t *lenp, loff_t *ppos)
2887 {
2888         return -ENOSYS;
2889 }
2890
2891 int proc_dointvec_minmax(struct ctl_table *table, int write,
2892                     void __user *buffer, size_t *lenp, loff_t *ppos)
2893 {
2894         return -ENOSYS;
2895 }
2896
2897 int proc_dointvec_jiffies(struct ctl_table *table, int write,
2898                     void __user *buffer, size_t *lenp, loff_t *ppos)
2899 {
2900         return -ENOSYS;
2901 }
2902
2903 int proc_dointvec_userhz_jiffies(struct ctl_table *table, int write,
2904                     void __user *buffer, size_t *lenp, loff_t *ppos)
2905 {
2906         return -ENOSYS;
2907 }
2908
2909 int proc_dointvec_ms_jiffies(struct ctl_table *table, int write,
2910                              void __user *buffer, size_t *lenp, loff_t *ppos)
2911 {
2912         return -ENOSYS;
2913 }
2914
2915 int proc_doulongvec_minmax(struct ctl_table *table, int write,
2916                     void __user *buffer, size_t *lenp, loff_t *ppos)
2917 {
2918         return -ENOSYS;
2919 }
2920
2921 int proc_doulongvec_ms_jiffies_minmax(struct ctl_table *table, int write,
2922                                       void __user *buffer,
2923                                       size_t *lenp, loff_t *ppos)
2924 {
2925     return -ENOSYS;
2926 }
2927
2928
2929 #endif /* CONFIG_PROC_SYSCTL */
2930
2931 /*
2932  * No sense putting this after each symbol definition, twice,
2933  * exception granted :-)
2934  */
2935 EXPORT_SYMBOL(proc_dointvec);
2936 EXPORT_SYMBOL(proc_douintvec);
2937 EXPORT_SYMBOL(proc_dointvec_jiffies);
2938 EXPORT_SYMBOL(proc_dointvec_minmax);
2939 EXPORT_SYMBOL(proc_dointvec_userhz_jiffies);
2940 EXPORT_SYMBOL(proc_dointvec_ms_jiffies);
2941 EXPORT_SYMBOL(proc_dostring);
2942 EXPORT_SYMBOL(proc_doulongvec_minmax);
2943 EXPORT_SYMBOL(proc_doulongvec_ms_jiffies_minmax);