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 compat_log;
 110 extern int latencytop_enabled;
 111 extern int sysctl_nr_open_min, sysctl_nr_open_max;
 112 #ifndef CONFIG_MMU
 113 extern int sysctl_nr_trim_pages;
 114 #endif
 115
 116 /* Constants used for minimum and  maximum */
 117 #ifdef CONFIG_LOCKUP_DETECTOR
 118 static int sixty = 60;
 119 #endif
 120
 121 static int __maybe_unused neg_one = -1;
 122
 123 static int zero;
 124 static int __maybe_unused one = 1;
 125 static int __maybe_unused two = 2;
 126 static int __maybe_unused four = 4;
 127 static unsigned long one_ul = 1;
 128 static int one_hundred = 100;
 129 static int one_thousand = 1000;
 130 #ifdef CONFIG_PRINTK
 131 static int ten_thousand = 10000;
 132 #endif
 133 #ifdef CONFIG_PERF_EVENTS
 134 static int six_hundred_forty_kb = 640 * 1024;
 135 #endif
 136
 137 /* this is needed for the proc_doulongvec_minmax of vm_dirty_bytes */
 138 static unsigned long dirty_bytes_min = 2 * PAGE_SIZE;
 139
 140 /* this is needed for the proc_dointvec_minmax for [fs_]overflow UID and GID */
 141 static int maxolduid = 65535;
 142 static int minolduid;
 143
 144 static int ngroups_max = NGROUPS_MAX;
 145 static const int cap_last_cap = CAP_LAST_CAP;
 146
 147 /*this is needed for proc_doulongvec_minmax of sysctl_hung_task_timeout_secs */
 148 #ifdef CONFIG_DETECT_HUNG_TASK
 149 static unsigned long hung_task_timeout_max = (LONG_MAX/HZ);
 150 #endif
 151
 152 #ifdef CONFIG_INOTIFY_USER
 153 #include <linux/inotify.h>
 154 #endif
 155 #ifdef CONFIG_SPARC
 156 #endif
 157
 158 #ifdef __hppa__
 159 extern int pwrsw_enabled;
 160 #endif
 161
 162 #ifdef CONFIG_SYSCTL_ARCH_UNALIGN_ALLOW
 163 extern int unaligned_enabled;
 164 #endif
 165
 166 #ifdef CONFIG_IA64
 167 extern int unaligned_dump_stack;
 168 #endif
 169
 170 #ifdef CONFIG_SYSCTL_ARCH_UNALIGN_NO_WARN
 171 extern int no_unaligned_warning;
 172 #endif
 173
 174 #ifdef CONFIG_PROC_SYSCTL
 175
 176 #define SYSCTL_WRITES_LEGACY    -1
 177 #define SYSCTL_WRITES_WARN       0
 178 #define SYSCTL_WRITES_STRICT     1
 179
 180 static int sysctl_writes_strict = SYSCTL_WRITES_STRICT;
 181
 182 static int proc_do_cad_pid(struct ctl_table *table, int write,
 183                   void __user *buffer, size_t *lenp, loff_t *ppos);
 184 static int proc_taint(struct ctl_table *table, int write,
 185                                void __user *buffer, size_t *lenp, loff_t *ppos);
 186 #endif
 187
 188 #ifdef CONFIG_PRINTK
 189 static int proc_dointvec_minmax_sysadmin(struct ctl_table *table, int write,
 190                                 void __user *buffer, size_t *lenp, loff_t *ppos);
 191 #endif
 192
 193 static int proc_dointvec_minmax_coredump(struct ctl_table *table, int write,
 194                 void __user *buffer, size_t *lenp, loff_t *ppos);
 195 #ifdef CONFIG_COREDUMP
 196 static int proc_dostring_coredump(struct ctl_table *table, int write,
 197                 void __user *buffer, size_t *lenp, loff_t *ppos);
 198 #endif
 199
 200 #ifdef CONFIG_MAGIC_SYSRQ
 201 /* Note: sysrq code uses it's own private copy */
 202 static int __sysrq_enabled = CONFIG_MAGIC_SYSRQ_DEFAULT_ENABLE;
 203
 204 static int sysrq_sysctl_handler(struct ctl_table *table, int write,
 205                                 void __user *buffer, size_t *lenp,
 206                                 loff_t *ppos)
 207 {
 208         int error;
 209
 210         error = proc_dointvec(table, write, buffer, lenp, ppos);
 211         if (error)
 212                 return error;
 213
 214         if (write)
 215                 sysrq_toggle_support(__sysrq_enabled);
 216
 217         return 0;
 218 }
 219
 220 #endif
 221
 222 static struct ctl_table kern_table[];
 223 static struct ctl_table vm_table[];
 224 static struct ctl_table fs_table[];
 225 static struct ctl_table debug_table[];
 226 static struct ctl_table dev_table[];
 227 extern struct ctl_table random_table[];
 228 #ifdef CONFIG_EPOLL
 229 extern struct ctl_table epoll_table[];
 230 #endif
 231
 232 #ifdef HAVE_ARCH_PICK_MMAP_LAYOUT
 233 int sysctl_legacy_va_layout;
 234 #endif
 235
 236 /* The default sysctl tables: */
 237
 238 static struct ctl_table sysctl_base_table[] = {
 239         {
 240                 .procname       = "kernel",
 241                 .mode           = 0555,
 242                 .child          = kern_table,
 243         },
 244         {
 245                 .procname       = "vm",
 246                 .mode           = 0555,
 247                 .child          = vm_table,
 248         },
 249         {
 250                 .procname       = "fs",
 251                 .mode           = 0555,
 252                 .child          = fs_table,
 253         },
 254         {
 255                 .procname       = "debug",
 256                 .mode           = 0555,
 257                 .child          = debug_table,
 258         },
 259         {
 260                 .procname       = "dev",
 261                 .mode           = 0555,
 262                 .child          = dev_table,
 263         },
 264         { }
 265 };
 266
 267 #ifdef CONFIG_SCHED_DEBUG
 268 static int min_sched_granularity_ns = 100000;           /* 100 usecs */
 269 static int max_sched_granularity_ns = NSEC_PER_SEC;     /* 1 second */
 270 static int min_wakeup_granularity_ns;                   /* 0 usecs */
 271 static int max_wakeup_granularity_ns = NSEC_PER_SEC;    /* 1 second */
 272 #ifdef CONFIG_SMP
 273 static int min_sched_tunable_scaling = SCHED_TUNABLESCALING_NONE;
 274 static int max_sched_tunable_scaling = SCHED_TUNABLESCALING_END-1;
 275 #endif /* CONFIG_SMP */
 276 #endif /* CONFIG_SCHED_DEBUG */
 277
 278 #ifdef CONFIG_COMPACTION
 279 static int min_extfrag_threshold;
 280 static int max_extfrag_threshold = 1000;
 281 #endif
 282
 283 static struct ctl_table kern_table[] = {
 284         {
 285                 .procname       = "sched_child_runs_first",
 286                 .data           = &sysctl_sched_child_runs_first,
 287                 .maxlen         = sizeof(unsigned int),
 288                 .mode           = 0644,
 289                 .proc_handler   = proc_dointvec,
 290         },
 291 #ifdef CONFIG_SCHED_DEBUG
 292         {
 293                 .procname       = "sched_min_granularity_ns",
 294                 .data           = &sysctl_sched_min_granularity,
 295                 .maxlen         = sizeof(unsigned int),
 296                 .mode           = 0644,
 297                 .proc_handler   = sched_proc_update_handler,
 298                 .extra1         = &min_sched_granularity_ns,
 299                 .extra2         = &max_sched_granularity_ns,
 300         },
 301         {
 302                 .procname       = "sched_latency_ns",
 303                 .data           = &sysctl_sched_latency,
 304                 .maxlen         = sizeof(unsigned int),
 305                 .mode           = 0644,
 306                 .proc_handler   = sched_proc_update_handler,
 307                 .extra1         = &min_sched_granularity_ns,
 308                 .extra2         = &max_sched_granularity_ns,
 309         },
 310         {
 311                 .procname       = "sched_wakeup_granularity_ns",
 312                 .data           = &sysctl_sched_wakeup_granularity,
 313                 .maxlen         = sizeof(unsigned int),
 314                 .mode           = 0644,
 315                 .proc_handler   = sched_proc_update_handler,
 316                 .extra1         = &min_wakeup_granularity_ns,
 317                 .extra2         = &max_wakeup_granularity_ns,
 318         },
 319 #ifdef CONFIG_SMP
 320         {
 321                 .procname       = "sched_tunable_scaling",
 322                 .data           = &sysctl_sched_tunable_scaling,
 323                 .maxlen         = sizeof(enum sched_tunable_scaling),
 324                 .mode           = 0644,
 325                 .proc_handler   = sched_proc_update_handler,
 326                 .extra1         = &min_sched_tunable_scaling,
 327                 .extra2         = &max_sched_tunable_scaling,
 328         },
 329         {
 330                 .procname       = "sched_migration_cost_ns",
 331                 .data           = &sysctl_sched_migration_cost,
 332                 .maxlen         = sizeof(unsigned int),
 333                 .mode           = 0644,
 334                 .proc_handler   = proc_dointvec,
 335         },
 336         {
 337                 .procname       = "sched_nr_migrate",
 338                 .data           = &sysctl_sched_nr_migrate,
 339                 .maxlen         = sizeof(unsigned int),
 340                 .mode           = 0644,
 341                 .proc_handler   = proc_dointvec,
 342         },
 343         {
 344                 .procname       = "sched_time_avg_ms",
 345                 .data           = &sysctl_sched_time_avg,
 346                 .maxlen         = sizeof(unsigned int),
 347                 .mode           = 0644,
 348                 .proc_handler   = proc_dointvec,
 349         },
 350         {
 351                 .procname       = "sched_shares_window_ns",
 352                 .data           = &sysctl_sched_shares_window,
 353                 .maxlen         = sizeof(unsigned int),
 354                 .mode           = 0644,
 355                 .proc_handler   = proc_dointvec,
 356         },
 357 #ifdef CONFIG_SCHEDSTATS
 358         {
 359                 .procname       = "sched_schedstats",
 360                 .data           = NULL,
 361                 .maxlen         = sizeof(unsigned int),
 362                 .mode           = 0644,
 363                 .proc_handler   = sysctl_schedstats,
 364                 .extra1         = &zero,
 365                 .extra2         = &one,
 366         },
 367 #endif /* CONFIG_SCHEDSTATS */
 368 #endif /* CONFIG_SMP */
 369 #ifdef CONFIG_NUMA_BALANCING
 370         {
 371                 .procname       = "numa_balancing_scan_delay_ms",
 372                 .data           = &sysctl_numa_balancing_scan_delay,
 373                 .maxlen         = sizeof(unsigned int),
 374                 .mode           = 0644,
 375                 .proc_handler   = proc_dointvec,
 376         },
 377         {
 378                 .procname       = "numa_balancing_scan_period_min_ms",
 379                 .data           = &sysctl_numa_balancing_scan_period_min,
 380                 .maxlen         = sizeof(unsigned int),
 381                 .mode           = 0644,
 382                 .proc_handler   = proc_dointvec,
 383         },
 384         {
 385                 .procname       = "numa_balancing_scan_period_max_ms",
 386                 .data           = &sysctl_numa_balancing_scan_period_max,
 387                 .maxlen         = sizeof(unsigned int),
 388                 .mode           = 0644,
 389                 .proc_handler   = proc_dointvec,
 390         },
 391         {
 392                 .procname       = "numa_balancing_scan_size_mb",
 393                 .data           = &sysctl_numa_balancing_scan_size,
 394                 .maxlen         = sizeof(unsigned int),
 395                 .mode           = 0644,
 396                 .proc_handler   = proc_dointvec_minmax,
 397                 .extra1         = &one,
 398         },
 399         {
 400                 .procname       = "numa_balancing",
 401                 .data           = NULL, /* filled in by handler */
 402                 .maxlen         = sizeof(unsigned int),
 403                 .mode           = 0644,
 404                 .proc_handler   = sysctl_numa_balancing,
 405                 .extra1         = &zero,
 406                 .extra2         = &one,
 407         },
 408 #endif /* CONFIG_NUMA_BALANCING */
 409 #endif /* CONFIG_SCHED_DEBUG */
 410         {
 411                 .procname       = "sched_rt_period_us",
 412                 .data           = &sysctl_sched_rt_period,
 413                 .maxlen         = sizeof(unsigned int),
 414                 .mode           = 0644,
 415                 .proc_handler   = sched_rt_handler,
 416         },
 417         {
 418                 .procname       = "sched_rt_runtime_us",
 419                 .data           = &sysctl_sched_rt_runtime,
 420                 .maxlen         = sizeof(int),
 421                 .mode           = 0644,
 422                 .proc_handler   = sched_rt_handler,
 423         },
 424         {
 425                 .procname       = "sched_rr_timeslice_ms",
 426                 .data           = &sched_rr_timeslice,
 427                 .maxlen         = sizeof(int),
 428                 .mode           = 0644,
 429                 .proc_handler   = sched_rr_handler,
 430         },
 431 #ifdef CONFIG_SCHED_AUTOGROUP
 432         {
 433                 .procname       = "sched_autogroup_enabled",
 434                 .data           = &sysctl_sched_autogroup_enabled,
 435                 .maxlen         = sizeof(unsigned int),
 436                 .mode           = 0644,
 437                 .proc_handler   = proc_dointvec_minmax,
 438                 .extra1         = &zero,
 439                 .extra2         = &one,
 440         },
 441 #endif
 442 #ifdef CONFIG_CFS_BANDWIDTH
 443         {
 444                 .procname       = "sched_cfs_bandwidth_slice_us",
 445                 .data           = &sysctl_sched_cfs_bandwidth_slice,
 446                 .maxlen         = sizeof(unsigned int),
 447                 .mode           = 0644,
 448                 .proc_handler   = proc_dointvec_minmax,
 449                 .extra1         = &one,
 450         },
 451 #endif
 452 #ifdef CONFIG_PROVE_LOCKING
 453         {
 454                 .procname       = "prove_locking",
 455                 .data           = &prove_locking,
 456                 .maxlen         = sizeof(int),
 457                 .mode           = 0644,
 458                 .proc_handler   = proc_dointvec,
 459         },
 460 #endif
 461 #ifdef CONFIG_LOCK_STAT
 462         {
 463                 .procname       = "lock_stat",
 464                 .data           = &lock_stat,
 465                 .maxlen         = sizeof(int),
 466                 .mode           = 0644,
 467                 .proc_handler   = proc_dointvec,
 468         },
 469 #endif
 470         {
 471                 .procname       = "panic",
 472                 .data           = &panic_timeout,
 473                 .maxlen         = sizeof(int),
 474                 .mode           = 0644,
 475                 .proc_handler   = proc_dointvec,
 476         },
 477 #ifdef CONFIG_COREDUMP
 478         {
 479                 .procname       = "core_uses_pid",
 480                 .data           = &core_uses_pid,
 481                 .maxlen         = sizeof(int),
 482                 .mode           = 0644,
 483                 .proc_handler   = proc_dointvec,
 484         },
 485         {
 486                 .procname       = "core_pattern",
 487                 .data           = core_pattern,
 488                 .maxlen         = CORENAME_MAX_SIZE,
 489                 .mode           = 0644,
 490                 .proc_handler   = proc_dostring_coredump,
 491         },
 492         {
 493                 .procname       = "core_pipe_limit",
 494                 .data           = &core_pipe_limit,
 495                 .maxlen         = sizeof(unsigned int),
 496                 .mode           = 0644,
 497                 .proc_handler   = proc_dointvec,
 498         },
 499 #endif
 500 #ifdef CONFIG_PROC_SYSCTL
 501         {
 502                 .procname       = "tainted",
 503                 .maxlen         = sizeof(long),
 504                 .mode           = 0644,
 505                 .proc_handler   = proc_taint,
 506         },
 507         {
 508                 .procname       = "sysctl_writes_strict",
 509                 .data           = &sysctl_writes_strict,
 510                 .maxlen         = sizeof(int),
 511                 .mode           = 0644,
 512                 .proc_handler   = proc_dointvec_minmax,
 513                 .extra1         = &neg_one,
 514                 .extra2         = &one,
 515         },
 516 #endif
 517 #ifdef CONFIG_LATENCYTOP
 518         {
 519                 .procname       = "latencytop",
 520                 .data           = &latencytop_enabled,
 521                 .maxlen         = sizeof(int),
 522                 .mode           = 0644,
 523                 .proc_handler   = sysctl_latencytop,
 524         },
 525 #endif
 526 #ifdef CONFIG_BLK_DEV_INITRD
 527         {
 528                 .procname       = "real-root-dev",
 529                 .data           = &real_root_dev,
 530                 .maxlen         = sizeof(int),
 531                 .mode           = 0644,
 532                 .proc_handler   = proc_dointvec,
 533         },
 534 #endif
 535         {
 536                 .procname       = "print-fatal-signals",
 537                 .data           = &print_fatal_signals,
 538                 .maxlen         = sizeof(int),
 539                 .mode           = 0644,
 540                 .proc_handler   = proc_dointvec,
 541         },
 542 #ifdef CONFIG_SPARC
 543         {
 544                 .procname       = "reboot-cmd",
 545                 .data           = reboot_command,
 546                 .maxlen         = 256,
 547                 .mode           = 0644,
 548                 .proc_handler   = proc_dostring,
 549         },
 550         {
 551                 .procname       = "stop-a",
 552                 .data           = &stop_a_enabled,
 553                 .maxlen         = sizeof (int),
 554                 .mode           = 0644,
 555                 .proc_handler   = proc_dointvec,
 556         },
 557         {
 558                 .procname       = "scons-poweroff",
 559                 .data           = &scons_pwroff,
 560                 .maxlen         = sizeof (int),
 561                 .mode           = 0644,
 562                 .proc_handler   = proc_dointvec,
 563         },
 564 #endif
 565 #ifdef CONFIG_SPARC64
 566         {
 567                 .procname       = "tsb-ratio",
 568                 .data           = &sysctl_tsb_ratio,
 569                 .maxlen         = sizeof (int),
 570                 .mode           = 0644,
 571                 .proc_handler   = proc_dointvec,
 572         },
 573 #endif
 574 #ifdef __hppa__
 575         {
 576                 .procname       = "soft-power",
 577                 .data           = &pwrsw_enabled,
 578                 .maxlen         = sizeof (int),
 579                 .mode           = 0644,
 580                 .proc_handler   = proc_dointvec,
 581         },
 582 #endif
 583 #ifdef CONFIG_SYSCTL_ARCH_UNALIGN_ALLOW
 584         {
 585                 .procname       = "unaligned-trap",
 586                 .data           = &unaligned_enabled,
 587                 .maxlen         = sizeof (int),
 588                 .mode           = 0644,
 589                 .proc_handler   = proc_dointvec,
 590         },
 591 #endif
 592         {
 593                 .procname       = "ctrl-alt-del",
 594                 .data           = &C_A_D,
 595                 .maxlen         = sizeof(int),
 596                 .mode           = 0644,
 597                 .proc_handler   = proc_dointvec,
 598         },
 599 #ifdef CONFIG_FUNCTION_TRACER
 600         {
 601                 .procname       = "ftrace_enabled",
 602                 .data           = &ftrace_enabled,
 603                 .maxlen         = sizeof(int),
 604                 .mode           = 0644,
 605                 .proc_handler   = ftrace_enable_sysctl,
 606         },
 607 #endif
 608 #ifdef CONFIG_STACK_TRACER
 609         {
 610                 .procname       = "stack_tracer_enabled",
 611                 .data           = &stack_tracer_enabled,
 612                 .maxlen         = sizeof(int),
 613                 .mode           = 0644,
 614                 .proc_handler   = stack_trace_sysctl,
 615         },
 616 #endif
 617 #ifdef CONFIG_TRACING
 618         {
 619                 .procname       = "ftrace_dump_on_oops",
 620                 .data           = &ftrace_dump_on_oops,
 621                 .maxlen         = sizeof(int),
 622                 .mode           = 0644,
 623                 .proc_handler   = proc_dointvec,
 624         },
 625         {
 626                 .procname       = "traceoff_on_warning",
 627                 .data           = &__disable_trace_on_warning,
 628                 .maxlen         = sizeof(__disable_trace_on_warning),
 629                 .mode           = 0644,
 630                 .proc_handler   = proc_dointvec,
 631         },
 632         {
 633                 .procname       = "tracepoint_printk",
 634                 .data           = &tracepoint_printk,
 635                 .maxlen         = sizeof(tracepoint_printk),
 636                 .mode           = 0644,
 637                 .proc_handler   = proc_dointvec,
 638         },
 639 #endif
 640 #ifdef CONFIG_KEXEC_CORE
 641         {
 642                 .procname       = "kexec_load_disabled",
 643                 .data           = &kexec_load_disabled,
 644                 .maxlen         = sizeof(int),
 645                 .mode           = 0644,
 646                 /* only handle a transition from default "0" to "1" */
 647                 .proc_handler   = proc_dointvec_minmax,
 648                 .extra1         = &one,
 649                 .extra2         = &one,
 650         },
 651 #endif
 652 #ifdef CONFIG_MODULES
 653         {
 654                 .procname       = "modprobe",
 655                 .data           = &modprobe_path,
 656                 .maxlen         = KMOD_PATH_LEN,
 657                 .mode           = 0644,
 658                 .proc_handler   = proc_dostring,
 659         },
 660         {
 661                 .procname       = "modules_disabled",
 662                 .data           = &modules_disabled,
 663                 .maxlen         = sizeof(int),
 664                 .mode           = 0644,
 665                 /* only handle a transition from default "0" to "1" */
 666                 .proc_handler   = proc_dointvec_minmax,
 667                 .extra1         = &one,
 668                 .extra2         = &one,
 669         },
 670 #endif
 671 #ifdef CONFIG_UEVENT_HELPER
 672         {
 673                 .procname       = "hotplug",
 674                 .data           = &uevent_helper,
 675                 .maxlen         = UEVENT_HELPER_PATH_LEN,
 676                 .mode           = 0644,
 677                 .proc_handler   = proc_dostring,
 678         },
 679 #endif
 680 #ifdef CONFIG_CHR_DEV_SG
 681         {
 682                 .procname       = "sg-big-buff",
 683                 .data           = &sg_big_buff,
 684                 .maxlen         = sizeof (int),
 685                 .mode           = 0444,
 686                 .proc_handler   = proc_dointvec,
 687         },
 688 #endif
 689 #ifdef CONFIG_BSD_PROCESS_ACCT
 690         {
 691                 .procname       = "acct",
 692                 .data           = &acct_parm,
 693                 .maxlen         = 3*sizeof(int),
 694                 .mode           = 0644,
 695                 .proc_handler   = proc_dointvec,
 696         },
 697 #endif
 698 #ifdef CONFIG_MAGIC_SYSRQ
 699         {
 700                 .procname       = "sysrq",
 701                 .data           = &__sysrq_enabled,
 702                 .maxlen         = sizeof (int),
 703                 .mode           = 0644,
 704                 .proc_handler   = sysrq_sysctl_handler,
 705         },
 706 #endif
 707 #ifdef CONFIG_PROC_SYSCTL
 708         {
 709                 .procname       = "cad_pid",
 710                 .data           = NULL,
 711                 .maxlen         = sizeof (int),
 712                 .mode           = 0600,
 713                 .proc_handler   = proc_do_cad_pid,
 714         },
 715 #endif
 716         {
 717                 .procname       = "threads-max",
 718                 .data           = NULL,
 719                 .maxlen         = sizeof(int),
 720                 .mode           = 0644,
 721                 .proc_handler   = sysctl_max_threads,
 722         },
 723         {
 724                 .procname       = "random",
 725                 .mode           = 0555,
 726                 .child          = random_table,
 727         },
 728         {
 729                 .procname       = "usermodehelper",
 730                 .mode           = 0555,
 731                 .child          = usermodehelper_table,
 732         },
 733         {
 734                 .procname       = "overflowuid",
 735                 .data           = &overflowuid,
 736                 .maxlen         = sizeof(int),
 737                 .mode           = 0644,
 738                 .proc_handler   = proc_dointvec_minmax,
 739                 .extra1         = &minolduid,
 740                 .extra2         = &maxolduid,
 741         },
 742         {
 743                 .procname       = "overflowgid",
 744                 .data           = &overflowgid,
 745                 .maxlen         = sizeof(int),
 746                 .mode           = 0644,
 747                 .proc_handler   = proc_dointvec_minmax,
 748                 .extra1         = &minolduid,
 749                 .extra2         = &maxolduid,
 750         },
 751 #ifdef CONFIG_S390
 752 #ifdef CONFIG_MATHEMU
 753         {
 754                 .procname       = "ieee_emulation_warnings",
 755                 .data           = &sysctl_ieee_emulation_warnings,
 756                 .maxlen         = sizeof(int),
 757                 .mode           = 0644,
 758                 .proc_handler   = proc_dointvec,
 759         },
 760 #endif
 761         {
 762                 .procname       = "userprocess_debug",
 763                 .data           = &show_unhandled_signals,
 764                 .maxlen         = sizeof(int),
 765                 .mode           = 0644,
 766                 .proc_handler   = proc_dointvec,
 767         },
 768 #endif
 769         {
 770                 .procname       = "pid_max",
 771                 .data           = &pid_max,
 772                 .maxlen         = sizeof (int),
 773                 .mode           = 0644,
 774                 .proc_handler   = proc_dointvec_minmax,
 775                 .extra1         = &pid_max_min,
 776                 .extra2         = &pid_max_max,
 777         },
 778         {
 779                 .procname       = "panic_on_oops",
 780                 .data           = &panic_on_oops,
 781                 .maxlen         = sizeof(int),
 782                 .mode           = 0644,
 783                 .proc_handler   = proc_dointvec,
 784         },
 785 #if defined CONFIG_PRINTK
 786         {
 787                 .procname       = "printk",
 788                 .data           = &console_loglevel,
 789                 .maxlen         = 4*sizeof(int),
 790                 .mode           = 0644,
 791                 .proc_handler   = proc_dointvec,
 792         },
 793         {
 794                 .procname       = "printk_ratelimit",
 795                 .data           = &printk_ratelimit_state.interval,
 796                 .maxlen         = sizeof(int),
 797                 .mode           = 0644,
 798                 .proc_handler   = proc_dointvec_jiffies,
 799         },
 800         {
 801                 .procname       = "printk_ratelimit_burst",
 802                 .data           = &printk_ratelimit_state.burst,
 803                 .maxlen         = sizeof(int),
 804                 .mode           = 0644,
 805                 .proc_handler   = proc_dointvec,
 806         },
 807         {
 808                 .procname       = "printk_delay",
 809                 .data           = &printk_delay_msec,
 810                 .maxlen         = sizeof(int),
 811                 .mode           = 0644,
 812                 .proc_handler   = proc_dointvec_minmax,
 813                 .extra1         = &zero,
 814                 .extra2         = &ten_thousand,
 815         },
 816         {
 817                 .procname       = "dmesg_restrict",
 818                 .data           = &dmesg_restrict,
 819                 .maxlen         = sizeof(int),
 820                 .mode           = 0644,
 821                 .proc_handler   = proc_dointvec_minmax_sysadmin,
 822                 .extra1         = &zero,
 823                 .extra2         = &one,
 824         },
 825         {
 826                 .procname       = "kptr_restrict",
 827                 .data           = &kptr_restrict,
 828                 .maxlen         = sizeof(int),
 829                 .mode           = 0644,
 830                 .proc_handler   = proc_dointvec_minmax_sysadmin,
 831                 .extra1         = &zero,
 832                 .extra2         = &two,
 833         },
 834 #endif
 835         {
 836                 .procname       = "ngroups_max",
 837                 .data           = &ngroups_max,
 838                 .maxlen         = sizeof (int),
 839                 .mode           = 0444,
 840                 .proc_handler   = proc_dointvec,
 841         },
 842         {
 843                 .procname       = "cap_last_cap",
 844                 .data           = (void *)&cap_last_cap,
 845                 .maxlen         = sizeof(int),
 846                 .mode           = 0444,
 847                 .proc_handler   = proc_dointvec,
 848         },
 849 #if defined(CONFIG_LOCKUP_DETECTOR)
 850         {
 851                 .procname       = "watchdog",
 852                 .data           = &watchdog_user_enabled,
 853                 .maxlen         = sizeof (int),
 854                 .mode           = 0644,
 855                 .proc_handler   = proc_watchdog,
 856                 .extra1         = &zero,
 857                 .extra2         = &one,
 858         },
 859         {
 860                 .procname       = "watchdog_thresh",
 861                 .data           = &watchdog_thresh,
 862                 .maxlen         = sizeof(int),
 863                 .mode           = 0644,
 864                 .proc_handler   = proc_watchdog_thresh,
 865                 .extra1         = &zero,
 866                 .extra2         = &sixty,
 867         },
 868         {
 869                 .procname       = "nmi_watchdog",
 870                 .data           = &nmi_watchdog_enabled,
 871                 .maxlen         = sizeof (int),
 872                 .mode           = 0644,
 873                 .proc_handler   = proc_nmi_watchdog,
 874                 .extra1         = &zero,
 875 #if defined(CONFIG_HAVE_NMI_WATCHDOG) || defined(CONFIG_HARDLOCKUP_DETECTOR)
 876                 .extra2         = &one,
 877 #else
 878                 .extra2         = &zero,
 879 #endif
 880         },
 881         {
 882                 .procname       = "soft_watchdog",
 883                 .data           = &soft_watchdog_enabled,
 884                 .maxlen         = sizeof (int),
 885                 .mode           = 0644,
 886                 .proc_handler   = proc_soft_watchdog,
 887                 .extra1         = &zero,
 888                 .extra2         = &one,
 889         },
 890         {
 891                 .procname       = "watchdog_cpumask",
 892                 .data           = &watchdog_cpumask_bits,
 893                 .maxlen         = NR_CPUS,
 894                 .mode           = 0644,
 895                 .proc_handler   = proc_watchdog_cpumask,
 896         },
 897         {
 898                 .procname       = "softlockup_panic",
 899                 .data           = &softlockup_panic,
 900                 .maxlen         = sizeof(int),
 901                 .mode           = 0644,
 902                 .proc_handler   = proc_dointvec_minmax,
 903                 .extra1         = &zero,
 904                 .extra2         = &one,
 905         },
 906 #ifdef CONFIG_HARDLOCKUP_DETECTOR
 907         {
 908                 .procname       = "hardlockup_panic",
 909                 .data           = &hardlockup_panic,
 910                 .maxlen         = sizeof(int),
 911                 .mode           = 0644,
 912                 .proc_handler   = proc_dointvec_minmax,
 913                 .extra1         = &zero,
 914                 .extra2         = &one,
 915         },
 916 #endif
 917 #ifdef CONFIG_SMP
 918         {
 919                 .procname       = "softlockup_all_cpu_backtrace",
 920                 .data           = &sysctl_softlockup_all_cpu_backtrace,
 921                 .maxlen         = sizeof(int),
 922                 .mode           = 0644,
 923                 .proc_handler   = proc_dointvec_minmax,
 924                 .extra1         = &zero,
 925                 .extra2         = &one,
 926         },
 927         {
 928                 .procname       = "hardlockup_all_cpu_backtrace",
 929                 .data           = &sysctl_hardlockup_all_cpu_backtrace,
 930                 .maxlen         = sizeof(int),
 931                 .mode           = 0644,
 932                 .proc_handler   = proc_dointvec_minmax,
 933                 .extra1         = &zero,
 934                 .extra2         = &one,
 935         },
 936 #endif /* CONFIG_SMP */
 937 #endif
 938 #if defined(CONFIG_X86_LOCAL_APIC) && defined(CONFIG_X86)
 939         {
 940                 .procname       = "unknown_nmi_panic",
 941                 .data           = &unknown_nmi_panic,
 942                 .maxlen         = sizeof (int),
 943                 .mode           = 0644,
 944                 .proc_handler   = proc_dointvec,
 945         },
 946 #endif
 947 #if defined(CONFIG_X86)
 948         {
 949                 .procname       = "panic_on_unrecovered_nmi",
 950                 .data           = &panic_on_unrecovered_nmi,
 951                 .maxlen         = sizeof(int),
 952                 .mode           = 0644,
 953                 .proc_handler   = proc_dointvec,
 954         },
 955         {
 956                 .procname       = "panic_on_io_nmi",
 957                 .data           = &panic_on_io_nmi,
 958                 .maxlen         = sizeof(int),
 959                 .mode           = 0644,
 960                 .proc_handler   = proc_dointvec,
 961         },
 962 #ifdef CONFIG_DEBUG_STACKOVERFLOW
 963         {
 964                 .procname       = "panic_on_stackoverflow",
 965                 .data           = &sysctl_panic_on_stackoverflow,
 966                 .maxlen         = sizeof(int),
 967                 .mode           = 0644,
 968                 .proc_handler   = proc_dointvec,
 969         },
 970 #endif
 971         {
 972                 .procname       = "bootloader_type",
 973                 .data           = &bootloader_type,
 974                 .maxlen         = sizeof (int),
 975                 .mode           = 0444,
 976                 .proc_handler   = proc_dointvec,
 977         },
 978         {
 979                 .procname       = "bootloader_version",
 980                 .data           = &bootloader_version,
 981                 .maxlen         = sizeof (int),
 982                 .mode           = 0444,
 983                 .proc_handler   = proc_dointvec,
 984         },
 985         {
 986                 .procname       = "kstack_depth_to_print",
 987                 .data           = &kstack_depth_to_print,
 988                 .maxlen         = sizeof(int),
 989                 .mode           = 0644,
 990                 .proc_handler   = proc_dointvec,
 991         },
 992         {
 993                 .procname       = "io_delay_type",
 994                 .data           = &io_delay_type,
 995                 .maxlen         = sizeof(int),
 996                 .mode           = 0644,
 997                 .proc_handler   = proc_dointvec,
 998         },
 999 #endif
1000 #if defined(CONFIG_MMU)
1001         {
1002                 .procname       = "randomize_va_space",
1003                 .data           = &randomize_va_space,
1004                 .maxlen         = sizeof(int),
1005                 .mode           = 0644,
1006                 .proc_handler   = proc_dointvec,
1007         },
1008 #endif
1009 #if defined(CONFIG_S390) && defined(CONFIG_SMP)
1010         {
1011                 .procname       = "spin_retry",
1012                 .data           = &spin_retry,
1013                 .maxlen         = sizeof (int),
1014                 .mode           = 0644,
1015                 .proc_handler   = proc_dointvec,
1016         },
1017 #endif
1018 #if     defined(CONFIG_ACPI_SLEEP) && defined(CONFIG_X86)
1019         {
1020                 .procname       = "acpi_video_flags",
1021                 .data           = &acpi_realmode_flags,
1022                 .maxlen         = sizeof (unsigned long),
1023                 .mode           = 0644,
1024                 .proc_handler   = proc_doulongvec_minmax,
1025         },
1026 #endif
1027 #ifdef CONFIG_SYSCTL_ARCH_UNALIGN_NO_WARN
1028         {
1029                 .procname       = "ignore-unaligned-usertrap",
1030                 .data           = &no_unaligned_warning,
1031                 .maxlen         = sizeof (int),
1032                 .mode           = 0644,
1033                 .proc_handler   = proc_dointvec,
1034         },
1035 #endif
1036 #ifdef CONFIG_IA64
1037         {
1038                 .procname       = "unaligned-dump-stack",
1039                 .data           = &unaligned_dump_stack,
1040                 .maxlen         = sizeof (int),
1041                 .mode           = 0644,
1042                 .proc_handler   = proc_dointvec,
1043         },
1044 #endif
1045 #ifdef CONFIG_DETECT_HUNG_TASK
1046         {
1047                 .procname       = "hung_task_panic",
1048                 .data           = &sysctl_hung_task_panic,
1049                 .maxlen         = sizeof(int),
1050                 .mode           = 0644,
1051                 .proc_handler   = proc_dointvec_minmax,
1052                 .extra1         = &zero,
1053                 .extra2         = &one,
1054         },
1055         {
1056                 .procname       = "hung_task_check_count",
1057                 .data           = &sysctl_hung_task_check_count,
1058                 .maxlen         = sizeof(int),
1059                 .mode           = 0644,
1060                 .proc_handler   = proc_dointvec_minmax,
1061                 .extra1         = &zero,
1062         },
1063         {
1064                 .procname       = "hung_task_timeout_secs",
1065                 .data           = &sysctl_hung_task_timeout_secs,
1066                 .maxlen         = sizeof(unsigned long),
1067                 .mode           = 0644,
1068                 .proc_handler   = proc_dohung_task_timeout_secs,
1069                 .extra2         = &hung_task_timeout_max,
1070         },
1071         {
1072                 .procname       = "hung_task_warnings",
1073                 .data           = &sysctl_hung_task_warnings,
1074                 .maxlen         = sizeof(int),
1075                 .mode           = 0644,
1076                 .proc_handler   = proc_dointvec_minmax,
1077                 .extra1         = &neg_one,
1078         },
1079 #endif
1080 #ifdef CONFIG_COMPAT
1081         {
1082                 .procname       = "compat-log",
1083                 .data           = &compat_log,
1084                 .maxlen         = sizeof (int),
1085                 .mode           = 0644,
1086                 .proc_handler   = proc_dointvec,
1087         },
1088 #endif
1089 #ifdef CONFIG_RT_MUTEXES
1090         {
1091                 .procname       = "max_lock_depth",
1092                 .data           = &max_lock_depth,
1093                 .maxlen         = sizeof(int),
1094                 .mode           = 0644,
1095                 .proc_handler   = proc_dointvec,
1096         },
1097 #endif
1098         {
1099                 .procname       = "poweroff_cmd",
1100                 .data           = &poweroff_cmd,
1101                 .maxlen         = POWEROFF_CMD_PATH_LEN,
1102                 .mode           = 0644,
1103                 .proc_handler   = proc_dostring,
1104         },
1105 #ifdef CONFIG_KEYS
1106         {
1107                 .procname       = "keys",
1108                 .mode           = 0555,
1109                 .child          = key_sysctls,
1110         },
1111 #endif
1112 #ifdef CONFIG_PERF_EVENTS
1113         /*
1114          * User-space scripts rely on the existence of this file
1115          * as a feature check for perf_events being enabled.
1116          *
1117          * So it's an ABI, do not remove!
1118          */
1119         {
1120                 .procname       = "perf_event_paranoid",
1121                 .data           = &sysctl_perf_event_paranoid,
1122                 .maxlen         = sizeof(sysctl_perf_event_paranoid),
1123                 .mode           = 0644,
1124                 .proc_handler   = proc_dointvec,
1125         },
1126         {
1127                 .procname       = "perf_event_mlock_kb",
1128                 .data           = &sysctl_perf_event_mlock,
1129                 .maxlen         = sizeof(sysctl_perf_event_mlock),
1130                 .mode           = 0644,
1131                 .proc_handler   = proc_dointvec,
1132         },
1133         {
1134                 .procname       = "perf_event_max_sample_rate",
1135                 .data           = &sysctl_perf_event_sample_rate,
1136                 .maxlen         = sizeof(sysctl_perf_event_sample_rate),
1137                 .mode           = 0644,
1138                 .proc_handler   = perf_proc_update_handler,
1139                 .extra1         = &one,
1140         },
1141         {
1142                 .procname       = "perf_cpu_time_max_percent",
1143                 .data           = &sysctl_perf_cpu_time_max_percent,
1144                 .maxlen         = sizeof(sysctl_perf_cpu_time_max_percent),
1145                 .mode           = 0644,
1146                 .proc_handler   = perf_cpu_time_max_percent_handler,
1147                 .extra1         = &zero,
1148                 .extra2         = &one_hundred,
1149         },
1150         {
1151                 .procname       = "perf_event_max_stack",
1152                 .data           = &sysctl_perf_event_max_stack,
1153                 .maxlen         = sizeof(sysctl_perf_event_max_stack),
1154                 .mode           = 0644,
1155                 .proc_handler   = perf_event_max_stack_handler,
1156                 .extra1         = &zero,
1157                 .extra2         = &six_hundred_forty_kb,
1158         },
1159         {
1160                 .procname       = "perf_event_max_contexts_per_stack",
1161                 .data           = &sysctl_perf_event_max_contexts_per_stack,
1162                 .maxlen         = sizeof(sysctl_perf_event_max_contexts_per_stack),
1163                 .mode           = 0644,
1164                 .proc_handler   = perf_event_max_stack_handler,
1165                 .extra1         = &zero,
1166                 .extra2         = &one_thousand,
1167         },
1168 #endif
1169 #ifdef CONFIG_KMEMCHECK
1170         {
1171                 .procname       = "kmemcheck",
1172                 .data           = &kmemcheck_enabled,
1173                 .maxlen         = sizeof(int),
1174                 .mode           = 0644,
1175                 .proc_handler   = proc_dointvec,
1176         },
1177 #endif
1178         {
1179                 .procname       = "panic_on_warn",
1180                 .data           = &panic_on_warn,
1181                 .maxlen         = sizeof(int),
1182                 .mode           = 0644,
1183                 .proc_handler   = proc_dointvec_minmax,
1184                 .extra1         = &zero,
1185                 .extra2         = &one,
1186         },
1187 #if defined(CONFIG_SMP) && defined(CONFIG_NO_HZ_COMMON)
1188         {
1189                 .procname       = "timer_migration",
1190                 .data           = &sysctl_timer_migration,
1191                 .maxlen         = sizeof(unsigned int),
1192                 .mode           = 0644,
1193                 .proc_handler   = timer_migration_handler,
1194         },
1195 #endif
1196 #ifdef CONFIG_BPF_SYSCALL
1197         {
1198                 .procname       = "unprivileged_bpf_disabled",
1199                 .data           = &sysctl_unprivileged_bpf_disabled,
1200                 .maxlen         = sizeof(sysctl_unprivileged_bpf_disabled),
1201                 .mode           = 0644,
1202                 /* only handle a transition from default "0" to "1" */
1203                 .proc_handler   = proc_dointvec_minmax,
1204                 .extra1         = &one,
1205                 .extra2         = &one,
1206         },
1207 #endif
1208         { }
1209 };
1210
1211 static struct ctl_table vm_table[] = {
1212         {
1213                 .procname       = "overcommit_memory",
1214                 .data           = &sysctl_overcommit_memory,
1215                 .maxlen         = sizeof(sysctl_overcommit_memory),
1216                 .mode           = 0644,
1217                 .proc_handler   = proc_dointvec_minmax,
1218                 .extra1         = &zero,
1219                 .extra2         = &two,
1220         },
1221         {
1222                 .procname       = "panic_on_oom",
1223                 .data           = &sysctl_panic_on_oom,
1224                 .maxlen         = sizeof(sysctl_panic_on_oom),
1225                 .mode           = 0644,
1226                 .proc_handler   = proc_dointvec_minmax,
1227                 .extra1         = &zero,
1228                 .extra2         = &two,
1229         },
1230         {
1231                 .procname       = "oom_kill_allocating_task",
1232                 .data           = &sysctl_oom_kill_allocating_task,
1233                 .maxlen         = sizeof(sysctl_oom_kill_allocating_task),
1234                 .mode           = 0644,
1235                 .proc_handler   = proc_dointvec,
1236         },
1237         {
1238                 .procname       = "oom_dump_tasks",
1239                 .data           = &sysctl_oom_dump_tasks,
1240                 .maxlen         = sizeof(sysctl_oom_dump_tasks),
1241                 .mode           = 0644,
1242                 .proc_handler   = proc_dointvec,
1243         },
1244         {
1245                 .procname       = "overcommit_ratio",
1246                 .data           = &sysctl_overcommit_ratio,
1247                 .maxlen         = sizeof(sysctl_overcommit_ratio),
1248                 .mode           = 0644,
1249                 .proc_handler   = overcommit_ratio_handler,
1250         },
1251         {
1252                 .procname       = "overcommit_kbytes",
1253                 .data           = &sysctl_overcommit_kbytes,
1254                 .maxlen         = sizeof(sysctl_overcommit_kbytes),
1255                 .mode           = 0644,
1256                 .proc_handler   = overcommit_kbytes_handler,
1257         },
1258         {
1259                 .procname       = "page-cluster",
1260                 .data           = &page_cluster,
1261                 .maxlen         = sizeof(int),
1262                 .mode           = 0644,
1263                 .proc_handler   = proc_dointvec_minmax,
1264                 .extra1         = &zero,
1265         },
1266         {
1267                 .procname       = "dirty_background_ratio",
1268                 .data           = &dirty_background_ratio,
1269                 .maxlen         = sizeof(dirty_background_ratio),
1270                 .mode           = 0644,
1271                 .proc_handler   = dirty_background_ratio_handler,
1272                 .extra1         = &zero,
1273                 .extra2         = &one_hundred,
1274         },
1275         {
1276                 .procname       = "dirty_background_bytes",
1277                 .data           = &dirty_background_bytes,
1278                 .maxlen         = sizeof(dirty_background_bytes),
1279                 .mode           = 0644,
1280                 .proc_handler   = dirty_background_bytes_handler,
1281                 .extra1         = &one_ul,
1282         },
1283         {
1284                 .procname       = "dirty_ratio",
1285                 .data           = &vm_dirty_ratio,
1286                 .maxlen         = sizeof(vm_dirty_ratio),
1287                 .mode           = 0644,
1288                 .proc_handler   = dirty_ratio_handler,
1289                 .extra1         = &zero,
1290                 .extra2         = &one_hundred,
1291         },
1292         {
1293                 .procname       = "dirty_bytes",
1294                 .data           = &vm_dirty_bytes,
1295                 .maxlen         = sizeof(vm_dirty_bytes),
1296                 .mode           = 0644,
1297                 .proc_handler   = dirty_bytes_handler,
1298                 .extra1         = &dirty_bytes_min,
1299         },
1300         {
1301                 .procname       = "dirty_writeback_centisecs",
1302                 .data           = &dirty_writeback_interval,
1303                 .maxlen         = sizeof(dirty_writeback_interval),
1304                 .mode           = 0644,
1305                 .proc_handler   = dirty_writeback_centisecs_handler,
1306         },
1307         {
1308                 .procname       = "dirty_expire_centisecs",
1309                 .data           = &dirty_expire_interval,
1310                 .maxlen         = sizeof(dirty_expire_interval),
1311                 .mode           = 0644,
1312                 .proc_handler   = proc_dointvec_minmax,
1313                 .extra1         = &zero,
1314         },
1315         {
1316                 .procname       = "dirtytime_expire_seconds",
1317                 .data           = &dirtytime_expire_interval,
1318                 .maxlen         = sizeof(dirty_expire_interval),
1319                 .mode           = 0644,
1320                 .proc_handler   = dirtytime_interval_handler,
1321                 .extra1         = &zero,
1322         },
1323         {
1324                 .procname       = "nr_pdflush_threads",
1325                 .mode           = 0444 /* read-only */,
1326                 .proc_handler   = pdflush_proc_obsolete,
1327         },
1328         {
1329                 .procname       = "swappiness",
1330                 .data           = &vm_swappiness,
1331                 .maxlen         = sizeof(vm_swappiness),
1332                 .mode           = 0644,
1333                 .proc_handler   = proc_dointvec_minmax,
1334                 .extra1         = &zero,
1335                 .extra2         = &one_hundred,
1336         },
1337 #ifdef CONFIG_HUGETLB_PAGE
1338         {
1339                 .procname       = "nr_hugepages",
1340                 .data           = NULL,
1341                 .maxlen         = sizeof(unsigned long),
1342                 .mode           = 0644,
1343                 .proc_handler   = hugetlb_sysctl_handler,
1344         },
1345 #ifdef CONFIG_NUMA
1346         {
1347                 .procname       = "nr_hugepages_mempolicy",
1348                 .data           = NULL,
1349                 .maxlen         = sizeof(unsigned long),
1350                 .mode           = 0644,
1351                 .proc_handler   = &hugetlb_mempolicy_sysctl_handler,
1352         },
1353 #endif
1354          {
1355                 .procname       = "hugetlb_shm_group",
1356                 .data           = &sysctl_hugetlb_shm_group,
1357                 .maxlen         = sizeof(gid_t),
1358                 .mode           = 0644,
1359                 .proc_handler   = proc_dointvec,
1360          },
1361          {
1362                 .procname       = "hugepages_treat_as_movable",
1363                 .data           = &hugepages_treat_as_movable,
1364                 .maxlen         = sizeof(int),
1365                 .mode           = 0644,
1366                 .proc_handler   = proc_dointvec,
1367         },
1368         {
1369                 .procname       = "nr_overcommit_hugepages",
1370                 .data           = NULL,
1371                 .maxlen         = sizeof(unsigned long),
1372                 .mode           = 0644,
1373                 .proc_handler   = hugetlb_overcommit_handler,
1374         },
1375 #endif
1376         {
1377                 .procname       = "lowmem_reserve_ratio",
1378                 .data           = &sysctl_lowmem_reserve_ratio,
1379                 .maxlen         = sizeof(sysctl_lowmem_reserve_ratio),
1380                 .mode           = 0644,
1381                 .proc_handler   = lowmem_reserve_ratio_sysctl_handler,
1382         },
1383         {
1384                 .procname       = "drop_caches",
1385                 .data           = &sysctl_drop_caches,
1386                 .maxlen         = sizeof(int),
1387                 .mode           = 0644,
1388                 .proc_handler   = drop_caches_sysctl_handler,
1389                 .extra1         = &one,
1390                 .extra2         = &four,
1391         },
1392 #ifdef CONFIG_COMPACTION
1393         {
1394                 .procname       = "compact_memory",
1395                 .data           = &sysctl_compact_memory,
1396                 .maxlen         = sizeof(int),
1397                 .mode           = 0200,
1398                 .proc_handler   = sysctl_compaction_handler,
1399         },
1400         {
1401                 .procname       = "extfrag_threshold",
1402                 .data           = &sysctl_extfrag_threshold,
1403                 .maxlen         = sizeof(int),
1404                 .mode           = 0644,
1405                 .proc_handler   = sysctl_extfrag_handler,
1406                 .extra1         = &min_extfrag_threshold,
1407                 .extra2         = &max_extfrag_threshold,
1408         },
1409         {
1410                 .procname       = "compact_unevictable_allowed",
1411                 .data           = &sysctl_compact_unevictable_allowed,
1412                 .maxlen         = sizeof(int),
1413                 .mode           = 0644,
1414                 .proc_handler   = proc_dointvec,
1415                 .extra1         = &zero,
1416                 .extra2         = &one,
1417         },
1418
1419 #endif /* CONFIG_COMPACTION */
1420         {
1421                 .procname       = "min_free_kbytes",
1422                 .data           = &min_free_kbytes,
1423                 .maxlen         = sizeof(min_free_kbytes),
1424                 .mode           = 0644,
1425                 .proc_handler   = min_free_kbytes_sysctl_handler,
1426                 .extra1         = &zero,
1427         },
1428         {
1429                 .procname       = "watermark_scale_factor",
1430                 .data           = &watermark_scale_factor,
1431                 .maxlen         = sizeof(watermark_scale_factor),
1432                 .mode           = 0644,
1433                 .proc_handler   = watermark_scale_factor_sysctl_handler,
1434                 .extra1         = &one,
1435                 .extra2         = &one_thousand,
1436         },
1437         {
1438                 .procname       = "percpu_pagelist_fraction",
1439                 .data           = &percpu_pagelist_fraction,
1440                 .maxlen         = sizeof(percpu_pagelist_fraction),
1441                 .mode           = 0644,
1442                 .proc_handler   = percpu_pagelist_fraction_sysctl_handler,
1443                 .extra1         = &zero,
1444         },
1445 #ifdef CONFIG_MMU
1446         {
1447                 .procname       = "max_map_count",
1448                 .data           = &sysctl_max_map_count,
1449                 .maxlen         = sizeof(sysctl_max_map_count),
1450                 .mode           = 0644,
1451                 .proc_handler   = proc_dointvec_minmax,
1452                 .extra1         = &zero,
1453         },
1454 #else
1455         {
1456                 .procname       = "nr_trim_pages",
1457                 .data           = &sysctl_nr_trim_pages,
1458                 .maxlen         = sizeof(sysctl_nr_trim_pages),
1459                 .mode           = 0644,
1460                 .proc_handler   = proc_dointvec_minmax,
1461                 .extra1         = &zero,
1462         },
1463 #endif
1464         {
1465                 .procname       = "laptop_mode",
1466                 .data           = &laptop_mode,
1467                 .maxlen         = sizeof(laptop_mode),
1468                 .mode           = 0644,
1469                 .proc_handler   = proc_dointvec_jiffies,
1470         },
1471         {
1472                 .procname       = "block_dump",
1473                 .data           = &block_dump,
1474                 .maxlen         = sizeof(block_dump),
1475                 .mode           = 0644,
1476                 .proc_handler   = proc_dointvec,
1477                 .extra1         = &zero,
1478         },
1479         {
1480                 .procname       = "vfs_cache_pressure",
1481                 .data           = &sysctl_vfs_cache_pressure,
1482                 .maxlen         = sizeof(sysctl_vfs_cache_pressure),
1483                 .mode           = 0644,
1484                 .proc_handler   = proc_dointvec,
1485                 .extra1         = &zero,
1486         },
1487 #ifdef HAVE_ARCH_PICK_MMAP_LAYOUT
1488         {
1489                 .procname       = "legacy_va_layout",
1490                 .data           = &sysctl_legacy_va_layout,
1491                 .maxlen         = sizeof(sysctl_legacy_va_layout),
1492                 .mode           = 0644,
1493                 .proc_handler   = proc_dointvec,
1494                 .extra1         = &zero,
1495         },
1496 #endif
1497 #ifdef CONFIG_NUMA
1498         {
1499                 .procname       = "zone_reclaim_mode",
1500                 .data           = &zone_reclaim_mode,
1501                 .maxlen         = sizeof(zone_reclaim_mode),
1502                 .mode           = 0644,
1503                 .proc_handler   = proc_dointvec,
1504                 .extra1         = &zero,
1505         },
1506         {
1507                 .procname       = "min_unmapped_ratio",
1508                 .data           = &sysctl_min_unmapped_ratio,
1509                 .maxlen         = sizeof(sysctl_min_unmapped_ratio),
1510                 .mode           = 0644,
1511                 .proc_handler   = sysctl_min_unmapped_ratio_sysctl_handler,
1512                 .extra1         = &zero,
1513                 .extra2         = &one_hundred,
1514         },
1515         {
1516                 .procname       = "min_slab_ratio",
1517                 .data           = &sysctl_min_slab_ratio,
1518                 .maxlen         = sizeof(sysctl_min_slab_ratio),
1519                 .mode           = 0644,
1520                 .proc_handler   = sysctl_min_slab_ratio_sysctl_handler,
1521                 .extra1         = &zero,
1522                 .extra2         = &one_hundred,
1523         },
1524 #endif
1525 #ifdef CONFIG_SMP
1526         {
1527                 .procname       = "stat_interval",
1528                 .data           = &sysctl_stat_interval,
1529                 .maxlen         = sizeof(sysctl_stat_interval),
1530                 .mode           = 0644,
1531                 .proc_handler   = proc_dointvec_jiffies,
1532         },
1533         {
1534                 .procname       = "stat_refresh",
1535                 .data           = NULL,
1536                 .maxlen         = 0,
1537                 .mode           = 0600,
1538                 .proc_handler   = vmstat_refresh,
1539         },
1540 #endif
1541 #ifdef CONFIG_MMU
1542         {
1543                 .procname       = "mmap_min_addr",
1544                 .data           = &dac_mmap_min_addr,
1545                 .maxlen         = sizeof(unsigned long),
1546                 .mode           = 0644,
1547                 .proc_handler   = mmap_min_addr_handler,
1548         },
1549 #endif
1550 #ifdef CONFIG_NUMA
1551         {
1552                 .procname       = "numa_zonelist_order",
1553                 .data           = &numa_zonelist_order,
1554                 .maxlen         = NUMA_ZONELIST_ORDER_LEN,
1555                 .mode           = 0644,
1556                 .proc_handler   = numa_zonelist_order_handler,
1557         },
1558 #endif
1559 #if (defined(CONFIG_X86_32) && !defined(CONFIG_UML))|| \
1560    (defined(CONFIG_SUPERH) && defined(CONFIG_VSYSCALL))
1561         {
1562                 .procname       = "vdso_enabled",
1563 #ifdef CONFIG_X86_32
1564                 .data           = &vdso32_enabled,
1565                 .maxlen         = sizeof(vdso32_enabled),
1566 #else
1567                 .data           = &vdso_enabled,
1568                 .maxlen         = sizeof(vdso_enabled),
1569 #endif
1570                 .mode           = 0644,
1571                 .proc_handler   = proc_dointvec,
1572                 .extra1         = &zero,
1573         },
1574 #endif
1575 #ifdef CONFIG_HIGHMEM
1576         {
1577                 .procname       = "highmem_is_dirtyable",
1578                 .data           = &vm_highmem_is_dirtyable,
1579                 .maxlen         = sizeof(vm_highmem_is_dirtyable),
1580                 .mode           = 0644,
1581                 .proc_handler   = proc_dointvec_minmax,
1582                 .extra1         = &zero,
1583                 .extra2         = &one,
1584         },
1585 #endif
1586 #ifdef CONFIG_MEMORY_FAILURE
1587         {
1588                 .procname       = "memory_failure_early_kill",
1589                 .data           = &sysctl_memory_failure_early_kill,
1590                 .maxlen         = sizeof(sysctl_memory_failure_early_kill),
1591                 .mode           = 0644,
1592                 .proc_handler   = proc_dointvec_minmax,
1593                 .extra1         = &zero,
1594                 .extra2         = &one,
1595         },
1596         {
1597                 .procname       = "memory_failure_recovery",
1598                 .data           = &sysctl_memory_failure_recovery,
1599                 .maxlen         = sizeof(sysctl_memory_failure_recovery),
1600                 .mode           = 0644,
1601                 .proc_handler   = proc_dointvec_minmax,
1602                 .extra1         = &zero,
1603                 .extra2         = &one,
1604         },
1605 #endif
1606         {
1607                 .procname       = "user_reserve_kbytes",
1608                 .data           = &sysctl_user_reserve_kbytes,
1609                 .maxlen         = sizeof(sysctl_user_reserve_kbytes),
1610                 .mode           = 0644,
1611                 .proc_handler   = proc_doulongvec_minmax,
1612         },
1613         {
1614                 .procname       = "admin_reserve_kbytes",
1615                 .data           = &sysctl_admin_reserve_kbytes,
1616                 .maxlen         = sizeof(sysctl_admin_reserve_kbytes),
1617                 .mode           = 0644,
1618                 .proc_handler   = proc_doulongvec_minmax,
1619         },
1620 #ifdef CONFIG_HAVE_ARCH_MMAP_RND_BITS
1621         {
1622                 .procname       = "mmap_rnd_bits",
1623                 .data           = &mmap_rnd_bits,
1624                 .maxlen         = sizeof(mmap_rnd_bits),
1625                 .mode           = 0600,
1626                 .proc_handler   = proc_dointvec_minmax,
1627                 .extra1         = (void *)&mmap_rnd_bits_min,
1628                 .extra2         = (void *)&mmap_rnd_bits_max,
1629         },
1630 #endif
1631 #ifdef CONFIG_HAVE_ARCH_MMAP_RND_COMPAT_BITS
1632         {
1633                 .procname       = "mmap_rnd_compat_bits",
1634                 .data           = &mmap_rnd_compat_bits,
1635                 .maxlen         = sizeof(mmap_rnd_compat_bits),
1636                 .mode           = 0600,
1637                 .proc_handler   = proc_dointvec_minmax,
1638                 .extra1         = (void *)&mmap_rnd_compat_bits_min,
1639                 .extra2         = (void *)&mmap_rnd_compat_bits_max,
1640         },
1641 #endif
1642         { }
1643 };
1644
1645 static struct ctl_table fs_table[] = {
1646         {
1647                 .procname       = "inode-nr",
1648                 .data           = &inodes_stat,
1649                 .maxlen         = 2*sizeof(long),
1650                 .mode           = 0444,
1651                 .proc_handler   = proc_nr_inodes,
1652         },
1653         {
1654                 .procname       = "inode-state",
1655                 .data           = &inodes_stat,
1656                 .maxlen         = 7*sizeof(long),
1657                 .mode           = 0444,
1658                 .proc_handler   = proc_nr_inodes,
1659         },
1660         {
1661                 .procname       = "file-nr",
1662                 .data           = &files_stat,
1663                 .maxlen         = sizeof(files_stat),
1664                 .mode           = 0444,
1665                 .proc_handler   = proc_nr_files,
1666         },
1667         {
1668                 .procname       = "file-max",
1669                 .data           = &files_stat.max_files,
1670                 .maxlen         = sizeof(files_stat.max_files),
1671                 .mode           = 0644,
1672                 .proc_handler   = proc_doulongvec_minmax,
1673         },
1674         {
1675                 .procname       = "nr_open",
1676                 .data           = &sysctl_nr_open,
1677                 .maxlen         = sizeof(int),
1678                 .mode           = 0644,
1679                 .proc_handler   = proc_dointvec_minmax,
1680                 .extra1         = &sysctl_nr_open_min,
1681                 .extra2         = &sysctl_nr_open_max,
1682         },
1683         {
1684                 .procname       = "dentry-state",
1685                 .data           = &dentry_stat,
1686                 .maxlen         = 6*sizeof(long),
1687                 .mode           = 0444,
1688                 .proc_handler   = proc_nr_dentry,
1689         },
1690         {
1691                 .procname       = "overflowuid",
1692                 .data           = &fs_overflowuid,
1693                 .maxlen         = sizeof(int),
1694                 .mode           = 0644,
1695                 .proc_handler   = proc_dointvec_minmax,
1696                 .extra1         = &minolduid,
1697                 .extra2         = &maxolduid,
1698         },
1699         {
1700                 .procname       = "overflowgid",
1701                 .data           = &fs_overflowgid,
1702                 .maxlen         = sizeof(int),
1703                 .mode           = 0644,
1704                 .proc_handler   = proc_dointvec_minmax,
1705                 .extra1         = &minolduid,
1706                 .extra2         = &maxolduid,
1707         },
1708 #ifdef CONFIG_FILE_LOCKING
1709         {
1710                 .procname       = "leases-enable",
1711                 .data           = &leases_enable,
1712                 .maxlen         = sizeof(int),
1713                 .mode           = 0644,
1714                 .proc_handler   = proc_dointvec,
1715         },
1716 #endif
1717 #ifdef CONFIG_DNOTIFY
1718         {
1719                 .procname       = "dir-notify-enable",
1720                 .data           = &dir_notify_enable,
1721                 .maxlen         = sizeof(int),
1722                 .mode           = 0644,
1723                 .proc_handler   = proc_dointvec,
1724         },
1725 #endif
1726 #ifdef CONFIG_MMU
1727 #ifdef CONFIG_FILE_LOCKING
1728         {
1729                 .procname       = "lease-break-time",
1730                 .data           = &lease_break_time,
1731                 .maxlen         = sizeof(int),
1732                 .mode           = 0644,
1733                 .proc_handler   = proc_dointvec,
1734         },
1735 #endif
1736 #ifdef CONFIG_AIO
1737         {
1738                 .procname       = "aio-nr",
1739                 .data           = &aio_nr,
1740                 .maxlen         = sizeof(aio_nr),
1741                 .mode           = 0444,
1742                 .proc_handler   = proc_doulongvec_minmax,
1743         },
1744         {
1745                 .procname       = "aio-max-nr",
1746                 .data           = &aio_max_nr,
1747                 .maxlen         = sizeof(aio_max_nr),
1748                 .mode           = 0644,
1749                 .proc_handler   = proc_doulongvec_minmax,
1750         },
1751 #endif /* CONFIG_AIO */
1752 #ifdef CONFIG_INOTIFY_USER
1753         {
1754                 .procname       = "inotify",
1755                 .mode           = 0555,
1756                 .child          = inotify_table,
1757         },
1758 #endif
1759 #ifdef CONFIG_EPOLL
1760         {
1761                 .procname       = "epoll",
1762                 .mode           = 0555,
1763                 .child          = epoll_table,
1764         },
1765 #endif
1766 #endif
1767         {
1768                 .procname       = "protected_symlinks",
1769                 .data           = &sysctl_protected_symlinks,
1770                 .maxlen         = sizeof(int),
1771                 .mode           = 0600,
1772                 .proc_handler   = proc_dointvec_minmax,
1773                 .extra1         = &zero,
1774                 .extra2         = &one,
1775         },
1776         {
1777                 .procname       = "protected_hardlinks",
1778                 .data           = &sysctl_protected_hardlinks,
1779                 .maxlen         = sizeof(int),
1780                 .mode           = 0600,
1781                 .proc_handler   = proc_dointvec_minmax,
1782                 .extra1         = &zero,
1783                 .extra2         = &one,
1784         },
1785         {
1786                 .procname       = "suid_dumpable",
1787                 .data           = &suid_dumpable,
1788                 .maxlen         = sizeof(int),
1789                 .mode           = 0644,
1790                 .proc_handler   = proc_dointvec_minmax_coredump,
1791                 .extra1         = &zero,
1792                 .extra2         = &two,
1793         },
1794 #if defined(CONFIG_BINFMT_MISC) || defined(CONFIG_BINFMT_MISC_MODULE)
1795         {
1796                 .procname       = "binfmt_misc",
1797                 .mode           = 0555,
1798                 .child          = sysctl_mount_point,
1799         },
1800 #endif
1801         {
1802                 .procname       = "pipe-max-size",
1803                 .data           = &pipe_max_size,
1804                 .maxlen         = sizeof(int),
1805                 .mode           = 0644,
1806                 .proc_handler   = &pipe_proc_fn,
1807                 .extra1         = &pipe_min_size,
1808         },
1809         {
1810                 .procname       = "pipe-user-pages-hard",
1811                 .data           = &pipe_user_pages_hard,
1812                 .maxlen         = sizeof(pipe_user_pages_hard),
1813                 .mode           = 0644,
1814                 .proc_handler   = proc_doulongvec_minmax,
1815         },
1816         {
1817                 .procname       = "pipe-user-pages-soft",
1818                 .data           = &pipe_user_pages_soft,
1819                 .maxlen         = sizeof(pipe_user_pages_soft),
1820                 .mode           = 0644,
1821                 .proc_handler   = proc_doulongvec_minmax,
1822         },
1823         { }
1824 };
1825
1826 static struct ctl_table debug_table[] = {
1827 #ifdef CONFIG_SYSCTL_EXCEPTION_TRACE
1828         {
1829                 .procname       = "exception-trace",
1830                 .data           = &show_unhandled_signals,
1831                 .maxlen         = sizeof(int),
1832                 .mode           = 0644,
1833                 .proc_handler   = proc_dointvec
1834         },
1835 #endif
1836 #if defined(CONFIG_OPTPROBES)
1837         {
1838                 .procname       = "kprobes-optimization",
1839                 .data           = &sysctl_kprobes_optimization,
1840                 .maxlen         = sizeof(int),
1841                 .mode           = 0644,
1842                 .proc_handler   = proc_kprobes_optimization_handler,
1843                 .extra1         = &zero,
1844                 .extra2         = &one,
1845         },
1846 #endif
1847         { }
1848 };
1849
1850 static struct ctl_table dev_table[] = {
1851         { }
1852 };
1853
1854 int __init sysctl_init(void)
1855 {
1856         struct ctl_table_header *hdr;
1857
1858         hdr = register_sysctl_table(sysctl_base_table);
1859         kmemleak_not_leak(hdr);
1860         return 0;
1861 }
1862
1863 #endif /* CONFIG_SYSCTL */
1864
1865 /*
1866  * /proc/sys support
1867  */
1868
1869 #ifdef CONFIG_PROC_SYSCTL
1870
1871 static int _proc_do_string(char *data, int maxlen, int write,
1872                            char __user *buffer,
1873                            size_t *lenp, loff_t *ppos)
1874 {
1875         size_t len;
1876         char __user *p;
1877         char c;
1878
1879         if (!data || !maxlen || !*lenp) {
1880                 *lenp = 0;
1881                 return 0;
1882         }
1883
1884         if (write) {
1885                 if (sysctl_writes_strict == SYSCTL_WRITES_STRICT) {
1886                         /* Only continue writes not past the end of buffer. */
1887                         len = strlen(data);
1888                         if (len > maxlen - 1)
1889                                 len = maxlen - 1;
1890
1891                         if (*ppos > len)
1892                                 return 0;
1893                         len = *ppos;
1894                 } else {
1895                         /* Start writing from beginning of buffer. */
1896                         len = 0;
1897                 }
1898
1899                 *ppos += *lenp;
1900                 p = buffer;
1901                 while ((p - buffer) < *lenp && len < maxlen - 1) {
1902                         if (get_user(c, p++))
1903                                 return -EFAULT;
1904                         if (c == 0 || c == '\n')
1905                                 break;
1906                         data[len++] = c;
1907                 }
1908                 data[len] = 0;
1909         } else {
1910                 len = strlen(data);
1911                 if (len > maxlen)
1912                         len = maxlen;
1913
1914                 if (*ppos > len) {
1915                         *lenp = 0;
1916                         return 0;
1917                 }
1918
1919                 data += *ppos;
1920                 len  -= *ppos;
1921
1922                 if (len > *lenp)
1923                         len = *lenp;
1924                 if (len)
1925                         if (copy_to_user(buffer, data, len))
1926                                 return -EFAULT;
1927                 if (len < *lenp) {
1928                         if (put_user('\n', buffer + len))
1929                                 return -EFAULT;
1930                         len++;
1931                 }
1932                 *lenp = len;
1933                 *ppos += len;
1934         }
1935         return 0;
1936 }
1937
1938 static void warn_sysctl_write(struct ctl_table *table)
1939 {
1940         pr_warn_once("%s wrote to %s when file position was not 0!\n"
1941                 "This will not be supported in the future. To silence this\n"
1942                 "warning, set kernel.sysctl_writes_strict = -1\n",
1943                 current->comm, table->procname);
1944 }
1945
1946 /**
1947  * proc_dostring - read a string sysctl
1948  * @table: the sysctl table
1949  * @write: %TRUE if this is a write to the sysctl file
1950  * @buffer: the user buffer
1951  * @lenp: the size of the user buffer
1952  * @ppos: file position
1953  *
1954  * Reads/writes a string from/to the user buffer. If the kernel
1955  * buffer provided is not large enough to hold the string, the
1956  * string is truncated. The copied string is %NULL-terminated.
1957  * If the string is being read by the user process, it is copied
1958  * and a newline '\n' is added. It is truncated if the buffer is
1959  * not large enough.
1960  *
1961  * Returns 0 on success.
1962  */
1963 int proc_dostring(struct ctl_table *table, int write,
1964                   void __user *buffer, size_t *lenp, loff_t *ppos)
1965 {
1966         if (write && *ppos && sysctl_writes_strict == SYSCTL_WRITES_WARN)
1967                 warn_sysctl_write(table);
1968
1969         return _proc_do_string((char *)(table->data), table->maxlen, write,
1970                                (char __user *)buffer, lenp, ppos);
1971 }
1972
1973 static size_t proc_skip_spaces(char **buf)
1974 {
1975         size_t ret;
1976         char *tmp = skip_spaces(*buf);
1977         ret = tmp - *buf;
1978         *buf = tmp;
1979         return ret;
1980 }
1981
1982 static void proc_skip_char(char **buf, size_t *size, const char v)
1983 {
1984         while (*size) {
1985                 if (**buf != v)
1986                         break;
1987                 (*size)--;
1988                 (*buf)++;
1989         }
1990 }
1991
1992 #define TMPBUFLEN 22
1993 /**
1994  * proc_get_long - reads an ASCII formatted integer from a user buffer
1995  *
1996  * @buf: a kernel buffer
1997  * @size: size of the kernel buffer
1998  * @val: this is where the number will be stored
1999  * @neg: set to %TRUE if number is negative
2000  * @perm_tr: a vector which contains the allowed trailers
2001  * @perm_tr_len: size of the perm_tr vector
2002  * @tr: pointer to store the trailer character
2003  *
2004  * In case of success %0 is returned and @buf and @size are updated with
2005  * the amount of bytes read. If @tr is non-NULL and a trailing
2006  * character exists (size is non-zero after returning from this
2007  * function), @tr is updated with the trailing character.
2008  */
2009 static int proc_get_long(char **buf, size_t *size,
2010                           unsigned long *val, bool *neg,
2011                           const char *perm_tr, unsigned perm_tr_len, char *tr)
2012 {
2013         int len;
2014         char *p, tmp[TMPBUFLEN];
2015
2016         if (!*size)
2017                 return -EINVAL;
2018
2019         len = *size;
2020         if (len > TMPBUFLEN - 1)
2021                 len = TMPBUFLEN - 1;
2022
2023         memcpy(tmp, *buf, len);
2024
2025         tmp[len] = 0;
2026         p = tmp;
2027         if (*p == '-' && *size > 1) {
2028                 *neg = true;
2029                 p++;
2030         } else
2031                 *neg = false;
2032         if (!isdigit(*p))
2033                 return -EINVAL;
2034
2035         *val = simple_strtoul(p, &p, 0);
2036
2037         len = p - tmp;
2038
2039         /* We don't know if the next char is whitespace thus we may accept
2040          * invalid integers (e.g. 1234...a) or two integers instead of one
2041          * (e.g. 123...1). So lets not allow such large numbers. */
2042         if (len == TMPBUFLEN - 1)
2043                 return -EINVAL;
2044
2045         if (len < *size && perm_tr_len && !memchr(perm_tr, *p, perm_tr_len))
2046                 return -EINVAL;
2047
2048         if (tr && (len < *size))
2049                 *tr = *p;
2050
2051         *buf += len;
2052         *size -= len;
2053
2054         return 0;
2055 }
2056
2057 /**
2058  * proc_put_long - converts an integer to a decimal ASCII formatted string
2059  *
2060  * @buf: the user buffer
2061  * @size: the size of the user buffer
2062  * @val: the integer to be converted
2063  * @neg: sign of the number, %TRUE for negative
2064  *
2065  * In case of success %0 is returned and @buf and @size are updated with
2066  * the amount of bytes written.
2067  */
2068 static int proc_put_long(void __user **buf, size_t *size, unsigned long val,
2069                           bool neg)
2070 {
2071         int len;
2072         char tmp[TMPBUFLEN], *p = tmp;
2073
2074         sprintf(p, "%s%lu", neg ? "-" : "", val);
2075         len = strlen(tmp);
2076         if (len > *size)
2077                 len = *size;
2078         if (copy_to_user(*buf, tmp, len))
2079                 return -EFAULT;
2080         *size -= len;
2081         *buf += len;
2082         return 0;
2083 }
2084 #undef TMPBUFLEN
2085
2086 static int proc_put_char(void __user **buf, size_t *size, char c)
2087 {
2088         if (*size) {
2089                 char __user **buffer = (char __user **)buf;
2090                 if (put_user(c, *buffer))
2091                         return -EFAULT;
2092                 (*size)--, (*buffer)++;
2093                 *buf = *buffer;
2094         }
2095         return 0;
2096 }
2097
2098 static int do_proc_dointvec_conv(bool *negp, unsigned long *lvalp,
2099                                  int *valp,
2100                                  int write, void *data)
2101 {
2102         if (write) {
2103                 if (*negp) {
2104                         if (*lvalp > (unsigned long) INT_MAX + 1)
2105                                 return -EINVAL;
2106                         *valp = -*lvalp;
2107                 } else {
2108                         if (*lvalp > (unsigned long) INT_MAX)
2109                                 return -EINVAL;
2110                         *valp = *lvalp;
2111                 }
2112         } else {
2113                 int val = *valp;
2114                 if (val < 0) {
2115                         *negp = true;
2116                         *lvalp = -(unsigned long)val;
2117                 } else {
2118                         *negp = false;
2119                         *lvalp = (unsigned long)val;
2120                 }
2121         }
2122         return 0;
2123 }
2124
2125 static const char proc_wspace_sep[] = { ' ', '\t', '\n' };
2126
2127 static int __do_proc_dointvec(void *tbl_data, struct ctl_table *table,
2128                   int write, void __user *buffer,
2129                   size_t *lenp, loff_t *ppos,
2130                   int (*conv)(bool *negp, unsigned long *lvalp, int *valp,
2131                               int write, void *data),
2132                   void *data)
2133 {
2134         int *i, vleft, first = 1, err = 0;
2135         size_t left;
2136         char *kbuf = NULL, *p;
2137
2138         if (!tbl_data || !table->maxlen || !*lenp || (*ppos && !write)) {
2139                 *lenp = 0;
2140                 return 0;
2141         }
2142
2143         i = (int *) tbl_data;
2144         vleft = table->maxlen / sizeof(*i);
2145         left = *lenp;
2146
2147         if (!conv)
2148                 conv = do_proc_dointvec_conv;
2149
2150         if (write) {
2151                 if (*ppos) {
2152                         switch (sysctl_writes_strict) {
2153                         case SYSCTL_WRITES_STRICT:
2154                                 goto out;
2155                         case SYSCTL_WRITES_WARN:
2156                                 warn_sysctl_write(table);
2157                                 break;
2158                         default:
2159                                 break;
2160                         }
2161                 }
2162
2163                 if (left > PAGE_SIZE - 1)
2164                         left = PAGE_SIZE - 1;
2165                 p = kbuf = memdup_user_nul(buffer, left);
2166                 if (IS_ERR(kbuf))
2167                         return PTR_ERR(kbuf);
2168         }
2169
2170         for (; left && vleft--; i++, first=0) {
2171                 unsigned long lval;
2172                 bool neg;
2173
2174                 if (write) {
2175                         left -= proc_skip_spaces(&p);
2176
2177                         if (!left)
2178                                 break;
2179                         err = proc_get_long(&p, &left, &lval, &neg,
2180                                              proc_wspace_sep,
2181                                              sizeof(proc_wspace_sep), NULL);
2182                         if (err)
2183                                 break;
2184                         if (conv(&neg, &lval, i, 1, data)) {
2185                                 err = -EINVAL;
2186                                 break;
2187                         }
2188                 } else {
2189                         if (conv(&neg, &lval, i, 0, data)) {
2190                                 err = -EINVAL;
2191                                 break;
2192                         }
2193                         if (!first)
2194                                 err = proc_put_char(&buffer, &left, '\t');
2195                         if (err)
2196                                 break;
2197                         err = proc_put_long(&buffer, &left, lval, neg);
2198                         if (err)
2199                                 break;
2200                 }
2201         }
2202
2203         if (!write && !first && left && !err)
2204                 err = proc_put_char(&buffer, &left, '\n');
2205         if (write && !err && left)
2206                 left -= proc_skip_spaces(&p);
2207         if (write) {
2208                 kfree(kbuf);
2209                 if (first)
2210                         return err ? : -EINVAL;
2211         }
2212         *lenp -= left;
2213 out:
2214         *ppos += *lenp;
2215         return err;
2216 }
2217
2218 static int do_proc_dointvec(struct ctl_table *table, int write,
2219                   void __user *buffer, size_t *lenp, loff_t *ppos,
2220                   int (*conv)(bool *negp, unsigned long *lvalp, int *valp,
2221                               int write, void *data),
2222                   void *data)
2223 {
2224         return __do_proc_dointvec(table->data, table, write,
2225                         buffer, lenp, ppos, conv, data);
2226 }
2227
2228 /**
2229  * proc_dointvec - read a vector of integers
2230  * @table: the sysctl table
2231  * @write: %TRUE if this is a write to the sysctl file
2232  * @buffer: the user buffer
2233  * @lenp: the size of the user buffer
2234  * @ppos: file position
2235  *
2236  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2237  * values from/to the user buffer, treated as an ASCII string.
2238  *
2239  * Returns 0 on success.
2240  */
2241 int proc_dointvec(struct ctl_table *table, int write,
2242                      void __user *buffer, size_t *lenp, loff_t *ppos)
2243 {
2244     return do_proc_dointvec(table,write,buffer,lenp,ppos,
2245                             NULL,NULL);
2246 }
2247
2248 /*
2249  * Taint values can only be increased
2250  * This means we can safely use a temporary.
2251  */
2252 static int proc_taint(struct ctl_table *table, int write,
2253                                void __user *buffer, size_t *lenp, loff_t *ppos)
2254 {
2255         struct ctl_table t;
2256         unsigned long tmptaint = get_taint();
2257         int err;
2258
2259         if (write && !capable(CAP_SYS_ADMIN))
2260                 return -EPERM;
2261
2262         t = *table;
2263         t.data = &tmptaint;
2264         err = proc_doulongvec_minmax(&t, write, buffer, lenp, ppos);
2265         if (err < 0)
2266                 return err;
2267
2268         if (write) {
2269                 /*
2270                  * Poor man's atomic or. Not worth adding a primitive
2271                  * to everyone's atomic.h for this
2272                  */
2273                 int i;
2274                 for (i = 0; i < BITS_PER_LONG && tmptaint >> i; i++) {
2275                         if ((tmptaint >> i) & 1)
2276                                 add_taint(i, LOCKDEP_STILL_OK);
2277                 }
2278         }
2279
2280         return err;
2281 }
2282
2283 #ifdef CONFIG_PRINTK
2284 static int proc_dointvec_minmax_sysadmin(struct ctl_table *table, int write,
2285                                 void __user *buffer, size_t *lenp, loff_t *ppos)
2286 {
2287         if (write && !capable(CAP_SYS_ADMIN))
2288                 return -EPERM;
2289
2290         return proc_dointvec_minmax(table, write, buffer, lenp, ppos);
2291 }
2292 #endif
2293
2294 struct do_proc_dointvec_minmax_conv_param {
2295         int *min;
2296         int *max;
2297 };
2298
2299 static int do_proc_dointvec_minmax_conv(bool *negp, unsigned long *lvalp,
2300                                         int *valp,
2301                                         int write, void *data)
2302 {
2303         struct do_proc_dointvec_minmax_conv_param *param = data;
2304         if (write) {
2305                 int val = *negp ? -*lvalp : *lvalp;
2306                 if ((param->min && *param->min > val) ||
2307                     (param->max && *param->max < val))
2308                         return -EINVAL;
2309                 *valp = val;
2310         } else {
2311                 int val = *valp;
2312                 if (val < 0) {
2313                         *negp = true;
2314                         *lvalp = -(unsigned long)val;
2315                 } else {
2316                         *negp = false;
2317                         *lvalp = (unsigned long)val;
2318                 }
2319         }
2320         return 0;
2321 }
2322
2323 /**
2324  * proc_dointvec_minmax - read a vector of integers with min/max values
2325  * @table: the sysctl table
2326  * @write: %TRUE if this is a write to the sysctl file
2327  * @buffer: the user buffer
2328  * @lenp: the size of the user buffer
2329  * @ppos: file position
2330  *
2331  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2332  * values from/to the user buffer, treated as an ASCII string.
2333  *
2334  * This routine will ensure the values are within the range specified by
2335  * table->extra1 (min) and table->extra2 (max).
2336  *
2337  * Returns 0 on success.
2338  */
2339 int proc_dointvec_minmax(struct ctl_table *table, int write,
2340                   void __user *buffer, size_t *lenp, loff_t *ppos)
2341 {
2342         struct do_proc_dointvec_minmax_conv_param param = {
2343                 .min = (int *) table->extra1,
2344                 .max = (int *) table->extra2,
2345         };
2346         return do_proc_dointvec(table, write, buffer, lenp, ppos,
2347                                 do_proc_dointvec_minmax_conv, &param);
2348 }
2349
2350 static void validate_coredump_safety(void)
2351 {
2352 #ifdef CONFIG_COREDUMP
2353         if (suid_dumpable == SUID_DUMP_ROOT &&
2354             core_pattern[0] != '/' && core_pattern[0] != '|') {
2355                 printk(KERN_WARNING "Unsafe core_pattern used with "\
2356                         "suid_dumpable=2. Pipe handler or fully qualified "\
2357                         "core dump path required.\n");
2358         }
2359 #endif
2360 }
2361
2362 static int proc_dointvec_minmax_coredump(struct ctl_table *table, int write,
2363                 void __user *buffer, size_t *lenp, loff_t *ppos)
2364 {
2365         int error = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
2366         if (!error)
2367                 validate_coredump_safety();
2368         return error;
2369 }
2370
2371 #ifdef CONFIG_COREDUMP
2372 static int proc_dostring_coredump(struct ctl_table *table, int write,
2373                   void __user *buffer, size_t *lenp, loff_t *ppos)
2374 {
2375         int error = proc_dostring(table, write, buffer, lenp, ppos);
2376         if (!error)
2377                 validate_coredump_safety();
2378         return error;
2379 }
2380 #endif
2381
2382 static int __do_proc_doulongvec_minmax(void *data, struct ctl_table *table, int write,
2383                                      void __user *buffer,
2384                                      size_t *lenp, loff_t *ppos,
2385                                      unsigned long convmul,
2386                                      unsigned long convdiv)
2387 {
2388         unsigned long *i, *min, *max;
2389         int vleft, first = 1, err = 0;
2390         size_t left;
2391         char *kbuf = NULL, *p;
2392
2393         if (!data || !table->maxlen || !*lenp || (*ppos && !write)) {
2394                 *lenp = 0;
2395                 return 0;
2396         }
2397
2398         i = (unsigned long *) data;
2399         min = (unsigned long *) table->extra1;
2400         max = (unsigned long *) table->extra2;
2401         vleft = table->maxlen / sizeof(unsigned long);
2402         left = *lenp;
2403
2404         if (write) {
2405                 if (*ppos) {
2406                         switch (sysctl_writes_strict) {
2407                         case SYSCTL_WRITES_STRICT:
2408                                 goto out;
2409                         case SYSCTL_WRITES_WARN:
2410                                 warn_sysctl_write(table);
2411                                 break;
2412                         default:
2413                                 break;
2414                         }
2415                 }
2416
2417                 if (left > PAGE_SIZE - 1)
2418                         left = PAGE_SIZE - 1;
2419                 p = kbuf = memdup_user_nul(buffer, left);
2420                 if (IS_ERR(kbuf))
2421                         return PTR_ERR(kbuf);
2422         }
2423
2424         for (; left && vleft--; i++, first = 0) {
2425                 unsigned long val;
2426
2427                 if (write) {
2428                         bool neg;
2429
2430                         left -= proc_skip_spaces(&p);
2431
2432                         err = proc_get_long(&p, &left, &val, &neg,
2433                                              proc_wspace_sep,
2434                                              sizeof(proc_wspace_sep), NULL);
2435                         if (err)
2436                                 break;
2437                         if (neg)
2438                                 continue;
2439                         if ((min && val < *min) || (max && val > *max))
2440                                 continue;
2441                         *i = val;
2442                 } else {
2443                         val = convdiv * (*i) / convmul;
2444                         if (!first) {
2445                                 err = proc_put_char(&buffer, &left, '\t');
2446                                 if (err)
2447                                         break;
2448                         }
2449                         err = proc_put_long(&buffer, &left, val, false);
2450                         if (err)
2451                                 break;
2452                 }
2453         }
2454
2455         if (!write && !first && left && !err)
2456                 err = proc_put_char(&buffer, &left, '\n');
2457         if (write && !err)
2458                 left -= proc_skip_spaces(&p);
2459         if (write) {
2460                 kfree(kbuf);
2461                 if (first)
2462                         return err ? : -EINVAL;
2463         }
2464         *lenp -= left;
2465 out:
2466         *ppos += *lenp;
2467         return err;
2468 }
2469
2470 static int do_proc_doulongvec_minmax(struct ctl_table *table, int write,
2471                                      void __user *buffer,
2472                                      size_t *lenp, loff_t *ppos,
2473                                      unsigned long convmul,
2474                                      unsigned long convdiv)
2475 {
2476         return __do_proc_doulongvec_minmax(table->data, table, write,
2477                         buffer, lenp, ppos, convmul, convdiv);
2478 }
2479
2480 /**
2481  * proc_doulongvec_minmax - read a vector of long integers with min/max values
2482  * @table: the sysctl table
2483  * @write: %TRUE if this is a write to the sysctl file
2484  * @buffer: the user buffer
2485  * @lenp: the size of the user buffer
2486  * @ppos: file position
2487  *
2488  * Reads/writes up to table->maxlen/sizeof(unsigned long) unsigned long
2489  * values from/to the user buffer, treated as an ASCII string.
2490  *
2491  * This routine will ensure the values are within the range specified by
2492  * table->extra1 (min) and table->extra2 (max).
2493  *
2494  * Returns 0 on success.
2495  */
2496 int proc_doulongvec_minmax(struct ctl_table *table, int write,
2497                            void __user *buffer, size_t *lenp, loff_t *ppos)
2498 {
2499     return do_proc_doulongvec_minmax(table, write, buffer, lenp, ppos, 1l, 1l);
2500 }
2501
2502 /**
2503  * proc_doulongvec_ms_jiffies_minmax - read a vector of millisecond values with min/max values
2504  * @table: the sysctl table
2505  * @write: %TRUE if this is a write to the sysctl file
2506  * @buffer: the user buffer
2507  * @lenp: the size of the user buffer
2508  * @ppos: file position
2509  *
2510  * Reads/writes up to table->maxlen/sizeof(unsigned long) unsigned long
2511  * values from/to the user buffer, treated as an ASCII string. The values
2512  * are treated as milliseconds, and converted to jiffies when they are stored.
2513  *
2514  * This routine will ensure the values are within the range specified by
2515  * table->extra1 (min) and table->extra2 (max).
2516  *
2517  * Returns 0 on success.
2518  */
2519 int proc_doulongvec_ms_jiffies_minmax(struct ctl_table *table, int write,
2520                                       void __user *buffer,
2521                                       size_t *lenp, loff_t *ppos)
2522 {
2523     return do_proc_doulongvec_minmax(table, write, buffer,
2524                                      lenp, ppos, HZ, 1000l);
2525 }
2526
2527
2528 static int do_proc_dointvec_jiffies_conv(bool *negp, unsigned long *lvalp,
2529                                          int *valp,
2530                                          int write, void *data)
2531 {
2532         if (write) {
2533                 if (*lvalp > LONG_MAX / HZ)
2534                         return 1;
2535                 *valp = *negp ? -(*lvalp*HZ) : (*lvalp*HZ);
2536         } else {
2537                 int val = *valp;
2538                 unsigned long lval;
2539                 if (val < 0) {
2540                         *negp = true;
2541                         lval = -(unsigned long)val;
2542                 } else {
2543                         *negp = false;
2544                         lval = (unsigned long)val;
2545                 }
2546                 *lvalp = lval / HZ;
2547         }
2548         return 0;
2549 }
2550
2551 static int do_proc_dointvec_userhz_jiffies_conv(bool *negp, unsigned long *lvalp,
2552                                                 int *valp,
2553                                                 int write, void *data)
2554 {
2555         if (write) {
2556                 if (USER_HZ < HZ && *lvalp > (LONG_MAX / HZ) * USER_HZ)
2557                         return 1;
2558                 *valp = clock_t_to_jiffies(*negp ? -*lvalp : *lvalp);
2559         } else {
2560                 int val = *valp;
2561                 unsigned long lval;
2562                 if (val < 0) {
2563                         *negp = true;
2564                         lval = -(unsigned long)val;
2565                 } else {
2566                         *negp = false;
2567                         lval = (unsigned long)val;
2568                 }
2569                 *lvalp = jiffies_to_clock_t(lval);
2570         }
2571         return 0;
2572 }
2573
2574 static int do_proc_dointvec_ms_jiffies_conv(bool *negp, unsigned long *lvalp,
2575                                             int *valp,
2576                                             int write, void *data)
2577 {
2578         if (write) {
2579                 unsigned long jif = msecs_to_jiffies(*negp ? -*lvalp : *lvalp);
2580
2581                 if (jif > INT_MAX)
2582                         return 1;
2583                 *valp = (int)jif;
2584         } else {
2585                 int val = *valp;
2586                 unsigned long lval;
2587                 if (val < 0) {
2588                         *negp = true;
2589                         lval = -(unsigned long)val;
2590                 } else {
2591                         *negp = false;
2592                         lval = (unsigned long)val;
2593                 }
2594                 *lvalp = jiffies_to_msecs(lval);
2595         }
2596         return 0;
2597 }
2598
2599 /**
2600  * proc_dointvec_jiffies - read a vector of integers as seconds
2601  * @table: the sysctl table
2602  * @write: %TRUE if this is a write to the sysctl file
2603  * @buffer: the user buffer
2604  * @lenp: the size of the user buffer
2605  * @ppos: file position
2606  *
2607  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2608  * values from/to the user buffer, treated as an ASCII string.
2609  * The values read are assumed to be in seconds, and are converted into
2610  * jiffies.
2611  *
2612  * Returns 0 on success.
2613  */
2614 int proc_dointvec_jiffies(struct ctl_table *table, int write,
2615                           void __user *buffer, size_t *lenp, loff_t *ppos)
2616 {
2617     return do_proc_dointvec(table,write,buffer,lenp,ppos,
2618                             do_proc_dointvec_jiffies_conv,NULL);
2619 }
2620
2621 /**
2622  * proc_dointvec_userhz_jiffies - read a vector of integers as 1/USER_HZ seconds
2623  * @table: the sysctl table
2624  * @write: %TRUE if this is a write to the sysctl file
2625  * @buffer: the user buffer
2626  * @lenp: the size of the user buffer
2627  * @ppos: pointer to the file position
2628  *
2629  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2630  * values from/to the user buffer, treated as an ASCII string.
2631  * The values read are assumed to be in 1/USER_HZ seconds, and
2632  * are converted into jiffies.
2633  *
2634  * Returns 0 on success.
2635  */
2636 int proc_dointvec_userhz_jiffies(struct ctl_table *table, int write,
2637                                  void __user *buffer, size_t *lenp, loff_t *ppos)
2638 {
2639     return do_proc_dointvec(table,write,buffer,lenp,ppos,
2640                             do_proc_dointvec_userhz_jiffies_conv,NULL);
2641 }
2642
2643 /**
2644  * proc_dointvec_ms_jiffies - read a vector of integers as 1 milliseconds
2645  * @table: the sysctl table
2646  * @write: %TRUE if this is a write to the sysctl file
2647  * @buffer: the user buffer
2648  * @lenp: the size of the user buffer
2649  * @ppos: file position
2650  * @ppos: the current position in the file
2651  *
2652  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2653  * values from/to the user buffer, treated as an ASCII string.
2654  * The values read are assumed to be in 1/1000 seconds, and
2655  * are converted into jiffies.
2656  *
2657  * Returns 0 on success.
2658  */
2659 int proc_dointvec_ms_jiffies(struct ctl_table *table, int write,
2660                              void __user *buffer, size_t *lenp, loff_t *ppos)
2661 {
2662         return do_proc_dointvec(table, write, buffer, lenp, ppos,
2663                                 do_proc_dointvec_ms_jiffies_conv, NULL);
2664 }
2665
2666 static int proc_do_cad_pid(struct ctl_table *table, int write,
2667                            void __user *buffer, size_t *lenp, loff_t *ppos)
2668 {
2669         struct pid *new_pid;
2670         pid_t tmp;
2671         int r;
2672
2673         tmp = pid_vnr(cad_pid);
2674
2675         r = __do_proc_dointvec(&tmp, table, write, buffer,
2676                                lenp, ppos, NULL, NULL);
2677         if (r || !write)
2678                 return r;
2679
2680         new_pid = find_get_pid(tmp);
2681         if (!new_pid)
2682                 return -ESRCH;
2683
2684         put_pid(xchg(&cad_pid, new_pid));
2685         return 0;
2686 }
2687
2688 /**
2689  * proc_do_large_bitmap - read/write from/to a large bitmap
2690  * @table: the sysctl table
2691  * @write: %TRUE if this is a write to the sysctl file
2692  * @buffer: the user buffer
2693  * @lenp: the size of the user buffer
2694  * @ppos: file position
2695  *
2696  * The bitmap is stored at table->data and the bitmap length (in bits)
2697  * in table->maxlen.
2698  *
2699  * We use a range comma separated format (e.g. 1,3-4,10-10) so that
2700  * large bitmaps may be represented in a compact manner. Writing into
2701  * the file will clear the bitmap then update it with the given input.
2702  *
2703  * Returns 0 on success.
2704  */
2705 int proc_do_large_bitmap(struct ctl_table *table, int write,
2706                          void __user *buffer, size_t *lenp, loff_t *ppos)
2707 {
2708         int err = 0;
2709         bool first = 1;
2710         size_t left = *lenp;
2711         unsigned long bitmap_len = table->maxlen;
2712         unsigned long *bitmap = *(unsigned long **) table->data;
2713         unsigned long *tmp_bitmap = NULL;
2714         char tr_a[] = { '-', ',', '\n' }, tr_b[] = { ',', '\n', 0 }, c;
2715
2716         if (!bitmap || !bitmap_len || !left || (*ppos && !write)) {
2717                 *lenp = 0;
2718                 return 0;
2719         }
2720
2721         if (write) {
2722                 char *kbuf, *p;
2723
2724                 if (left > PAGE_SIZE - 1)
2725                         left = PAGE_SIZE - 1;
2726
2727                 p = kbuf = memdup_user_nul(buffer, left);
2728                 if (IS_ERR(kbuf))
2729                         return PTR_ERR(kbuf);
2730
2731                 tmp_bitmap = kzalloc(BITS_TO_LONGS(bitmap_len) * sizeof(unsigned long),
2732                                      GFP_KERNEL);
2733                 if (!tmp_bitmap) {
2734                         kfree(kbuf);
2735                         return -ENOMEM;
2736                 }
2737                 proc_skip_char(&p, &left, '\n');
2738                 while (!err && left) {
2739                         unsigned long val_a, val_b;
2740                         bool neg;
2741
2742                         err = proc_get_long(&p, &left, &val_a, &neg, tr_a,
2743                                              sizeof(tr_a), &c);
2744                         if (err)
2745                                 break;
2746                         if (val_a >= bitmap_len || neg) {
2747                                 err = -EINVAL;
2748                                 break;
2749                         }
2750
2751                         val_b = val_a;
2752                         if (left) {
2753                                 p++;
2754                                 left--;
2755                         }
2756
2757                         if (c == '-') {
2758                                 err = proc_get_long(&p, &left, &val_b,
2759                                                      &neg, tr_b, sizeof(tr_b),
2760                                                      &c);
2761                                 if (err)
2762                                         break;
2763                                 if (val_b >= bitmap_len || neg ||
2764                                     val_a > val_b) {
2765                                         err = -EINVAL;
2766                                         break;
2767                                 }
2768                                 if (left) {
2769                                         p++;
2770                                         left--;
2771                                 }
2772                         }
2773
2774                         bitmap_set(tmp_bitmap, val_a, val_b - val_a + 1);
2775                         first = 0;
2776                         proc_skip_char(&p, &left, '\n');
2777                 }
2778                 kfree(kbuf);
2779         } else {
2780                 unsigned long bit_a, bit_b = 0;
2781
2782                 while (left) {
2783                         bit_a = find_next_bit(bitmap, bitmap_len, bit_b);
2784                         if (bit_a >= bitmap_len)
2785                                 break;
2786                         bit_b = find_next_zero_bit(bitmap, bitmap_len,
2787                                                    bit_a + 1) - 1;
2788
2789                         if (!first) {
2790                                 err = proc_put_char(&buffer, &left, ',');
2791                                 if (err)
2792                                         break;
2793                         }
2794                         err = proc_put_long(&buffer, &left, bit_a, false);
2795                         if (err)
2796                                 break;
2797                         if (bit_a != bit_b) {
2798                                 err = proc_put_char(&buffer, &left, '-');
2799                                 if (err)
2800                                         break;
2801                                 err = proc_put_long(&buffer, &left, bit_b, false);
2802                                 if (err)
2803                                         break;
2804                         }
2805
2806                         first = 0; bit_b++;
2807                 }
2808                 if (!err)
2809                         err = proc_put_char(&buffer, &left, '\n');
2810         }
2811
2812         if (!err) {
2813                 if (write) {
2814                         if (*ppos)
2815                                 bitmap_or(bitmap, bitmap, tmp_bitmap, bitmap_len);
2816                         else
2817                                 bitmap_copy(bitmap, tmp_bitmap, bitmap_len);
2818                 }
2819                 kfree(tmp_bitmap);
2820                 *lenp -= left;
2821                 *ppos += *lenp;
2822                 return 0;
2823         } else {
2824                 kfree(tmp_bitmap);
2825                 return err;
2826         }
2827 }
2828
2829 #else /* CONFIG_PROC_SYSCTL */
2830
2831 int proc_dostring(struct ctl_table *table, int write,
2832                   void __user *buffer, size_t *lenp, loff_t *ppos)
2833 {
2834         return -ENOSYS;
2835 }
2836
2837 int proc_dointvec(struct ctl_table *table, int write,
2838                   void __user *buffer, size_t *lenp, loff_t *ppos)
2839 {
2840         return -ENOSYS;
2841 }
2842
2843 int proc_dointvec_minmax(struct ctl_table *table, int write,
2844                     void __user *buffer, size_t *lenp, loff_t *ppos)
2845 {
2846         return -ENOSYS;
2847 }
2848
2849 int proc_dointvec_jiffies(struct ctl_table *table, int write,
2850                     void __user *buffer, size_t *lenp, loff_t *ppos)
2851 {
2852         return -ENOSYS;
2853 }
2854
2855 int proc_dointvec_userhz_jiffies(struct ctl_table *table, int write,
2856                     void __user *buffer, size_t *lenp, loff_t *ppos)
2857 {
2858         return -ENOSYS;
2859 }
2860
2861 int proc_dointvec_ms_jiffies(struct ctl_table *table, int write,
2862                              void __user *buffer, size_t *lenp, loff_t *ppos)
2863 {
2864         return -ENOSYS;
2865 }
2866
2867 int proc_doulongvec_minmax(struct ctl_table *table, int write,
2868                     void __user *buffer, size_t *lenp, loff_t *ppos)
2869 {
2870         return -ENOSYS;
2871 }
2872
2873 int proc_doulongvec_ms_jiffies_minmax(struct ctl_table *table, int write,
2874                                       void __user *buffer,
2875                                       size_t *lenp, loff_t *ppos)
2876 {
2877     return -ENOSYS;
2878 }
2879
2880
2881 #endif /* CONFIG_PROC_SYSCTL */
2882
2883 /*
2884  * No sense putting this after each symbol definition, twice,
2885  * exception granted :-)
2886  */
2887 EXPORT_SYMBOL(proc_dointvec);
2888 EXPORT_SYMBOL(proc_dointvec_jiffies);
2889 EXPORT_SYMBOL(proc_dointvec_minmax);
2890 EXPORT_SYMBOL(proc_dointvec_userhz_jiffies);
2891 EXPORT_SYMBOL(proc_dointvec_ms_jiffies);
2892 EXPORT_SYMBOL(proc_dostring);
2893 EXPORT_SYMBOL(proc_doulongvec_minmax);
2894 EXPORT_SYMBOL(proc_doulongvec_ms_jiffies_minmax);