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 #include <linux/mount.h>
  69
  70 #include <asm/uaccess.h>
  71 #include <asm/processor.h>
  72
  73 #ifdef CONFIG_X86
  74 #include <asm/nmi.h>
  75 #include <asm/stacktrace.h>
  76 #include <asm/io.h>
  77 #endif
  78 #ifdef CONFIG_SPARC
  79 #include <asm/setup.h>
  80 #endif
  81 #ifdef CONFIG_BSD_PROCESS_ACCT
  82 #include <linux/acct.h>
  83 #endif
  84 #ifdef CONFIG_RT_MUTEXES
  85 #include <linux/rtmutex.h>
  86 #endif
  87 #if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_LOCK_STAT)
  88 #include <linux/lockdep.h>
  89 #endif
  90 #ifdef CONFIG_CHR_DEV_SG
  91 #include <scsi/sg.h>
  92 #endif
  93
  94 #ifdef CONFIG_LOCKUP_DETECTOR
  95 #include <linux/nmi.h>
  96 #endif
  97
  98 #if defined(CONFIG_SYSCTL)
  99
 100 /* External variables not in a header file. */
 101 extern int suid_dumpable;
 102 #ifdef CONFIG_COREDUMP
 103 extern int core_uses_pid;
 104 extern char core_pattern[];
 105 extern unsigned int core_pipe_limit;
 106 #endif
 107 extern int pid_max;
 108 extern int pid_max_min, pid_max_max;
 109 extern int percpu_pagelist_fraction;
 110 extern int latencytop_enabled;
 111 extern unsigned 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 zero_ul;
 128 static unsigned long one_ul = 1;
 129 static unsigned long long_max = LONG_MAX;
 130 static int one_hundred = 100;
 131 static int one_thousand = 1000;
 132 #ifdef CONFIG_PRINTK
 133 static int ten_thousand = 10000;
 134 #endif
 135 #ifdef CONFIG_PERF_EVENTS
 136 static int six_hundred_forty_kb = 640 * 1024;
 137 #endif
 138
 139 /* this is needed for the proc_doulongvec_minmax of vm_dirty_bytes */
 140 static unsigned long dirty_bytes_min = 2 * PAGE_SIZE;
 141
 142 /* this is needed for the proc_dointvec_minmax for [fs_]overflow UID and GID */
 143 static int maxolduid = 65535;
 144 static int minolduid;
 145
 146 static int ngroups_max = NGROUPS_MAX;
 147 static const int cap_last_cap = CAP_LAST_CAP;
 148
 149 /*this is needed for proc_doulongvec_minmax of sysctl_hung_task_timeout_secs */
 150 #ifdef CONFIG_DETECT_HUNG_TASK
 151 static unsigned long hung_task_timeout_max = (LONG_MAX/HZ);
 152 #endif
 153
 154 #ifdef CONFIG_INOTIFY_USER
 155 #include <linux/inotify.h>
 156 #endif
 157 #ifdef CONFIG_SPARC
 158 #endif
 159
 160 #ifdef __hppa__
 161 extern int pwrsw_enabled;
 162 #endif
 163
 164 #ifdef CONFIG_SYSCTL_ARCH_UNALIGN_ALLOW
 165 extern int unaligned_enabled;
 166 #endif
 167
 168 #ifdef CONFIG_IA64
 169 extern int unaligned_dump_stack;
 170 #endif
 171
 172 #ifdef CONFIG_SYSCTL_ARCH_UNALIGN_NO_WARN
 173 extern int no_unaligned_warning;
 174 #endif
 175
 176 #ifdef CONFIG_PROC_SYSCTL
 177
 178 #define SYSCTL_WRITES_LEGACY    -1
 179 #define SYSCTL_WRITES_WARN       0
 180 #define SYSCTL_WRITES_STRICT     1
 181
 182 static int sysctl_writes_strict = SYSCTL_WRITES_STRICT;
 183
 184 static int proc_do_cad_pid(struct ctl_table *table, int write,
 185                   void __user *buffer, size_t *lenp, loff_t *ppos);
 186 static int proc_taint(struct ctl_table *table, int write,
 187                                void __user *buffer, size_t *lenp, loff_t *ppos);
 188 #endif
 189
 190 #ifdef CONFIG_PRINTK
 191 static int proc_dointvec_minmax_sysadmin(struct ctl_table *table, int write,
 192                                 void __user *buffer, size_t *lenp, loff_t *ppos);
 193 #endif
 194
 195 static int proc_dointvec_minmax_coredump(struct ctl_table *table, int write,
 196                 void __user *buffer, size_t *lenp, loff_t *ppos);
 197 #ifdef CONFIG_COREDUMP
 198 static int proc_dostring_coredump(struct ctl_table *table, int write,
 199                 void __user *buffer, size_t *lenp, loff_t *ppos);
 200 #endif
 201
 202 #ifdef CONFIG_MAGIC_SYSRQ
 203 /* Note: sysrq code uses it's own private copy */
 204 static int __sysrq_enabled = CONFIG_MAGIC_SYSRQ_DEFAULT_ENABLE;
 205
 206 static int sysrq_sysctl_handler(struct ctl_table *table, int write,
 207                                 void __user *buffer, size_t *lenp,
 208                                 loff_t *ppos)
 209 {
 210         int error;
 211
 212         error = proc_dointvec(table, write, buffer, lenp, ppos);
 213         if (error)
 214                 return error;
 215
 216         if (write)
 217                 sysrq_toggle_support(__sysrq_enabled);
 218
 219         return 0;
 220 }
 221
 222 #endif
 223
 224 static struct ctl_table kern_table[];
 225 static struct ctl_table vm_table[];
 226 static struct ctl_table fs_table[];
 227 static struct ctl_table debug_table[];
 228 static struct ctl_table dev_table[];
 229 extern struct ctl_table random_table[];
 230 #ifdef CONFIG_EPOLL
 231 extern struct ctl_table epoll_table[];
 232 #endif
 233
 234 #ifdef HAVE_ARCH_PICK_MMAP_LAYOUT
 235 int sysctl_legacy_va_layout;
 236 #endif
 237
 238 /* The default sysctl tables: */
 239
 240 static struct ctl_table sysctl_base_table[] = {
 241         {
 242                 .procname       = "kernel",
 243                 .mode           = 0555,
 244                 .child          = kern_table,
 245         },
 246         {
 247                 .procname       = "vm",
 248                 .mode           = 0555,
 249                 .child          = vm_table,
 250         },
 251         {
 252                 .procname       = "fs",
 253                 .mode           = 0555,
 254                 .child          = fs_table,
 255         },
 256         {
 257                 .procname       = "debug",
 258                 .mode           = 0555,
 259                 .child          = debug_table,
 260         },
 261         {
 262                 .procname       = "dev",
 263                 .mode           = 0555,
 264                 .child          = dev_table,
 265         },
 266         { }
 267 };
 268
 269 #ifdef CONFIG_SCHED_DEBUG
 270 static int min_sched_granularity_ns = 100000;           /* 100 usecs */
 271 static int max_sched_granularity_ns = NSEC_PER_SEC;     /* 1 second */
 272 static int min_wakeup_granularity_ns;                   /* 0 usecs */
 273 static int max_wakeup_granularity_ns = NSEC_PER_SEC;    /* 1 second */
 274 #ifdef CONFIG_SMP
 275 static int min_sched_tunable_scaling = SCHED_TUNABLESCALING_NONE;
 276 static int max_sched_tunable_scaling = SCHED_TUNABLESCALING_END-1;
 277 #endif /* CONFIG_SMP */
 278 #endif /* CONFIG_SCHED_DEBUG */
 279
 280 #ifdef CONFIG_COMPACTION
 281 static int min_extfrag_threshold;
 282 static int max_extfrag_threshold = 1000;
 283 #endif
 284
 285 static struct ctl_table kern_table[] = {
 286         {
 287                 .procname       = "sched_child_runs_first",
 288                 .data           = &sysctl_sched_child_runs_first,
 289                 .maxlen         = sizeof(unsigned int),
 290                 .mode           = 0644,
 291                 .proc_handler   = proc_dointvec,
 292         },
 293 #ifdef CONFIG_SCHED_DEBUG
 294         {
 295                 .procname       = "sched_min_granularity_ns",
 296                 .data           = &sysctl_sched_min_granularity,
 297                 .maxlen         = sizeof(unsigned int),
 298                 .mode           = 0644,
 299                 .proc_handler   = sched_proc_update_handler,
 300                 .extra1         = &min_sched_granularity_ns,
 301                 .extra2         = &max_sched_granularity_ns,
 302         },
 303         {
 304                 .procname       = "sched_latency_ns",
 305                 .data           = &sysctl_sched_latency,
 306                 .maxlen         = sizeof(unsigned int),
 307                 .mode           = 0644,
 308                 .proc_handler   = sched_proc_update_handler,
 309                 .extra1         = &min_sched_granularity_ns,
 310                 .extra2         = &max_sched_granularity_ns,
 311         },
 312         {
 313                 .procname       = "sched_wakeup_granularity_ns",
 314                 .data           = &sysctl_sched_wakeup_granularity,
 315                 .maxlen         = sizeof(unsigned int),
 316                 .mode           = 0644,
 317                 .proc_handler   = sched_proc_update_handler,
 318                 .extra1         = &min_wakeup_granularity_ns,
 319                 .extra2         = &max_wakeup_granularity_ns,
 320         },
 321 #ifdef CONFIG_SMP
 322         {
 323                 .procname       = "sched_tunable_scaling",
 324                 .data           = &sysctl_sched_tunable_scaling,
 325                 .maxlen         = sizeof(enum sched_tunable_scaling),
 326                 .mode           = 0644,
 327                 .proc_handler   = sched_proc_update_handler,
 328                 .extra1         = &min_sched_tunable_scaling,
 329                 .extra2         = &max_sched_tunable_scaling,
 330         },
 331         {
 332                 .procname       = "sched_migration_cost_ns",
 333                 .data           = &sysctl_sched_migration_cost,
 334                 .maxlen         = sizeof(unsigned int),
 335                 .mode           = 0644,
 336                 .proc_handler   = proc_dointvec,
 337         },
 338         {
 339                 .procname       = "sched_nr_migrate",
 340                 .data           = &sysctl_sched_nr_migrate,
 341                 .maxlen         = sizeof(unsigned int),
 342                 .mode           = 0644,
 343                 .proc_handler   = proc_dointvec,
 344         },
 345         {
 346                 .procname       = "sched_time_avg_ms",
 347                 .data           = &sysctl_sched_time_avg,
 348                 .maxlen         = sizeof(unsigned int),
 349                 .mode           = 0644,
 350                 .proc_handler   = proc_dointvec_minmax,
 351                 .extra1         = &one,
 352         },
 353         {
 354                 .procname       = "sched_shares_window_ns",
 355                 .data           = &sysctl_sched_shares_window,
 356                 .maxlen         = sizeof(unsigned int),
 357                 .mode           = 0644,
 358                 .proc_handler   = proc_dointvec,
 359         },
 360 #ifdef CONFIG_SCHEDSTATS
 361         {
 362                 .procname       = "sched_schedstats",
 363                 .data           = NULL,
 364                 .maxlen         = sizeof(unsigned int),
 365                 .mode           = 0644,
 366                 .proc_handler   = sysctl_schedstats,
 367                 .extra1         = &zero,
 368                 .extra2         = &one,
 369         },
 370 #endif /* CONFIG_SCHEDSTATS */
 371 #endif /* CONFIG_SMP */
 372 #ifdef CONFIG_NUMA_BALANCING
 373         {
 374                 .procname       = "numa_balancing_scan_delay_ms",
 375                 .data           = &sysctl_numa_balancing_scan_delay,
 376                 .maxlen         = sizeof(unsigned int),
 377                 .mode           = 0644,
 378                 .proc_handler   = proc_dointvec,
 379         },
 380         {
 381                 .procname       = "numa_balancing_scan_period_min_ms",
 382                 .data           = &sysctl_numa_balancing_scan_period_min,
 383                 .maxlen         = sizeof(unsigned int),
 384                 .mode           = 0644,
 385                 .proc_handler   = proc_dointvec,
 386         },
 387         {
 388                 .procname       = "numa_balancing_scan_period_max_ms",
 389                 .data           = &sysctl_numa_balancing_scan_period_max,
 390                 .maxlen         = sizeof(unsigned int),
 391                 .mode           = 0644,
 392                 .proc_handler   = proc_dointvec,
 393         },
 394         {
 395                 .procname       = "numa_balancing_scan_size_mb",
 396                 .data           = &sysctl_numa_balancing_scan_size,
 397                 .maxlen         = sizeof(unsigned int),
 398                 .mode           = 0644,
 399                 .proc_handler   = proc_dointvec_minmax,
 400                 .extra1         = &one,
 401         },
 402         {
 403                 .procname       = "numa_balancing",
 404                 .data           = NULL, /* filled in by handler */
 405                 .maxlen         = sizeof(unsigned int),
 406                 .mode           = 0644,
 407                 .proc_handler   = sysctl_numa_balancing,
 408                 .extra1         = &zero,
 409                 .extra2         = &one,
 410         },
 411 #endif /* CONFIG_NUMA_BALANCING */
 412 #endif /* CONFIG_SCHED_DEBUG */
 413         {
 414                 .procname       = "sched_rt_period_us",
 415                 .data           = &sysctl_sched_rt_period,
 416                 .maxlen         = sizeof(unsigned int),
 417                 .mode           = 0644,
 418                 .proc_handler   = sched_rt_handler,
 419         },
 420         {
 421                 .procname       = "sched_rt_runtime_us",
 422                 .data           = &sysctl_sched_rt_runtime,
 423                 .maxlen         = sizeof(int),
 424                 .mode           = 0644,
 425                 .proc_handler   = sched_rt_handler,
 426         },
 427         {
 428                 .procname       = "sched_rr_timeslice_ms",
 429                 .data           = &sched_rr_timeslice,
 430                 .maxlen         = sizeof(int),
 431                 .mode           = 0644,
 432                 .proc_handler   = sched_rr_handler,
 433         },
 434 #ifdef CONFIG_SCHED_AUTOGROUP
 435         {
 436                 .procname       = "sched_autogroup_enabled",
 437                 .data           = &sysctl_sched_autogroup_enabled,
 438                 .maxlen         = sizeof(unsigned int),
 439                 .mode           = 0644,
 440                 .proc_handler   = proc_dointvec_minmax,
 441                 .extra1         = &zero,
 442                 .extra2         = &one,
 443         },
 444 #endif
 445 #ifdef CONFIG_CFS_BANDWIDTH
 446         {
 447                 .procname       = "sched_cfs_bandwidth_slice_us",
 448                 .data           = &sysctl_sched_cfs_bandwidth_slice,
 449                 .maxlen         = sizeof(unsigned int),
 450                 .mode           = 0644,
 451                 .proc_handler   = proc_dointvec_minmax,
 452                 .extra1         = &one,
 453         },
 454 #endif
 455 #ifdef CONFIG_PROVE_LOCKING
 456         {
 457                 .procname       = "prove_locking",
 458                 .data           = &prove_locking,
 459                 .maxlen         = sizeof(int),
 460                 .mode           = 0644,
 461                 .proc_handler   = proc_dointvec,
 462         },
 463 #endif
 464 #ifdef CONFIG_LOCK_STAT
 465         {
 466                 .procname       = "lock_stat",
 467                 .data           = &lock_stat,
 468                 .maxlen         = sizeof(int),
 469                 .mode           = 0644,
 470                 .proc_handler   = proc_dointvec,
 471         },
 472 #endif
 473         {
 474                 .procname       = "panic",
 475                 .data           = &panic_timeout,
 476                 .maxlen         = sizeof(int),
 477                 .mode           = 0644,
 478                 .proc_handler   = proc_dointvec,
 479         },
 480 #ifdef CONFIG_COREDUMP
 481         {
 482                 .procname       = "core_uses_pid",
 483                 .data           = &core_uses_pid,
 484                 .maxlen         = sizeof(int),
 485                 .mode           = 0644,
 486                 .proc_handler   = proc_dointvec,
 487         },
 488         {
 489                 .procname       = "core_pattern",
 490                 .data           = core_pattern,
 491                 .maxlen         = CORENAME_MAX_SIZE,
 492                 .mode           = 0644,
 493                 .proc_handler   = proc_dostring_coredump,
 494         },
 495         {
 496                 .procname       = "core_pipe_limit",
 497                 .data           = &core_pipe_limit,
 498                 .maxlen         = sizeof(unsigned int),
 499                 .mode           = 0644,
 500                 .proc_handler   = proc_dointvec,
 501         },
 502 #endif
 503 #ifdef CONFIG_PROC_SYSCTL
 504         {
 505                 .procname       = "tainted",
 506                 .maxlen         = sizeof(long),
 507                 .mode           = 0644,
 508                 .proc_handler   = proc_taint,
 509         },
 510         {
 511                 .procname       = "sysctl_writes_strict",
 512                 .data           = &sysctl_writes_strict,
 513                 .maxlen         = sizeof(int),
 514                 .mode           = 0644,
 515                 .proc_handler   = proc_dointvec_minmax,
 516                 .extra1         = &neg_one,
 517                 .extra2         = &one,
 518         },
 519 #endif
 520 #ifdef CONFIG_LATENCYTOP
 521         {
 522                 .procname       = "latencytop",
 523                 .data           = &latencytop_enabled,
 524                 .maxlen         = sizeof(int),
 525                 .mode           = 0644,
 526                 .proc_handler   = sysctl_latencytop,
 527         },
 528 #endif
 529 #ifdef CONFIG_BLK_DEV_INITRD
 530         {
 531                 .procname       = "real-root-dev",
 532                 .data           = &real_root_dev,
 533                 .maxlen         = sizeof(int),
 534                 .mode           = 0644,
 535                 .proc_handler   = proc_dointvec,
 536         },
 537 #endif
 538         {
 539                 .procname       = "print-fatal-signals",
 540                 .data           = &print_fatal_signals,
 541                 .maxlen         = sizeof(int),
 542                 .mode           = 0644,
 543                 .proc_handler   = proc_dointvec,
 544         },
 545 #ifdef CONFIG_SPARC
 546         {
 547                 .procname       = "reboot-cmd",
 548                 .data           = reboot_command,
 549                 .maxlen         = 256,
 550                 .mode           = 0644,
 551                 .proc_handler   = proc_dostring,
 552         },
 553         {
 554                 .procname       = "stop-a",
 555                 .data           = &stop_a_enabled,
 556                 .maxlen         = sizeof (int),
 557                 .mode           = 0644,
 558                 .proc_handler   = proc_dointvec,
 559         },
 560         {
 561                 .procname       = "scons-poweroff",
 562                 .data           = &scons_pwroff,
 563                 .maxlen         = sizeof (int),
 564                 .mode           = 0644,
 565                 .proc_handler   = proc_dointvec,
 566         },
 567 #endif
 568 #ifdef CONFIG_SPARC64
 569         {
 570                 .procname       = "tsb-ratio",
 571                 .data           = &sysctl_tsb_ratio,
 572                 .maxlen         = sizeof (int),
 573                 .mode           = 0644,
 574                 .proc_handler   = proc_dointvec,
 575         },
 576 #endif
 577 #ifdef __hppa__
 578         {
 579                 .procname       = "soft-power",
 580                 .data           = &pwrsw_enabled,
 581                 .maxlen         = sizeof (int),
 582                 .mode           = 0644,
 583                 .proc_handler   = proc_dointvec,
 584         },
 585 #endif
 586 #ifdef CONFIG_SYSCTL_ARCH_UNALIGN_ALLOW
 587         {
 588                 .procname       = "unaligned-trap",
 589                 .data           = &unaligned_enabled,
 590                 .maxlen         = sizeof (int),
 591                 .mode           = 0644,
 592                 .proc_handler   = proc_dointvec,
 593         },
 594 #endif
 595         {
 596                 .procname       = "ctrl-alt-del",
 597                 .data           = &C_A_D,
 598                 .maxlen         = sizeof(int),
 599                 .mode           = 0644,
 600                 .proc_handler   = proc_dointvec,
 601         },
 602 #ifdef CONFIG_FUNCTION_TRACER
 603         {
 604                 .procname       = "ftrace_enabled",
 605                 .data           = &ftrace_enabled,
 606                 .maxlen         = sizeof(int),
 607                 .mode           = 0644,
 608                 .proc_handler   = ftrace_enable_sysctl,
 609         },
 610 #endif
 611 #ifdef CONFIG_STACK_TRACER
 612         {
 613                 .procname       = "stack_tracer_enabled",
 614                 .data           = &stack_tracer_enabled,
 615                 .maxlen         = sizeof(int),
 616                 .mode           = 0644,
 617                 .proc_handler   = stack_trace_sysctl,
 618         },
 619 #endif
 620 #ifdef CONFIG_TRACING
 621         {
 622                 .procname       = "ftrace_dump_on_oops",
 623                 .data           = &ftrace_dump_on_oops,
 624                 .maxlen         = sizeof(int),
 625                 .mode           = 0644,
 626                 .proc_handler   = proc_dointvec,
 627         },
 628         {
 629                 .procname       = "traceoff_on_warning",
 630                 .data           = &__disable_trace_on_warning,
 631                 .maxlen         = sizeof(__disable_trace_on_warning),
 632                 .mode           = 0644,
 633                 .proc_handler   = proc_dointvec,
 634         },
 635         {
 636                 .procname       = "tracepoint_printk",
 637                 .data           = &tracepoint_printk,
 638                 .maxlen         = sizeof(tracepoint_printk),
 639                 .mode           = 0644,
 640                 .proc_handler   = proc_dointvec,
 641         },
 642 #endif
 643 #ifdef CONFIG_KEXEC_CORE
 644         {
 645                 .procname       = "kexec_load_disabled",
 646                 .data           = &kexec_load_disabled,
 647                 .maxlen         = sizeof(int),
 648                 .mode           = 0644,
 649                 /* only handle a transition from default "0" to "1" */
 650                 .proc_handler   = proc_dointvec_minmax,
 651                 .extra1         = &one,
 652                 .extra2         = &one,
 653         },
 654 #endif
 655 #ifdef CONFIG_MODULES
 656         {
 657                 .procname       = "modprobe",
 658                 .data           = &modprobe_path,
 659                 .maxlen         = KMOD_PATH_LEN,
 660                 .mode           = 0644,
 661                 .proc_handler   = proc_dostring,
 662         },
 663         {
 664                 .procname       = "modules_disabled",
 665                 .data           = &modules_disabled,
 666                 .maxlen         = sizeof(int),
 667                 .mode           = 0644,
 668                 /* only handle a transition from default "0" to "1" */
 669                 .proc_handler   = proc_dointvec_minmax,
 670                 .extra1         = &one,
 671                 .extra2         = &one,
 672         },
 673 #endif
 674 #ifdef CONFIG_UEVENT_HELPER
 675         {
 676                 .procname       = "hotplug",
 677                 .data           = &uevent_helper,
 678                 .maxlen         = UEVENT_HELPER_PATH_LEN,
 679                 .mode           = 0644,
 680                 .proc_handler   = proc_dostring,
 681         },
 682 #endif
 683 #ifdef CONFIG_CHR_DEV_SG
 684         {
 685                 .procname       = "sg-big-buff",
 686                 .data           = &sg_big_buff,
 687                 .maxlen         = sizeof (int),
 688                 .mode           = 0444,
 689                 .proc_handler   = proc_dointvec,
 690         },
 691 #endif
 692 #ifdef CONFIG_BSD_PROCESS_ACCT
 693         {
 694                 .procname       = "acct",
 695                 .data           = &acct_parm,
 696                 .maxlen         = 3*sizeof(int),
 697                 .mode           = 0644,
 698                 .proc_handler   = proc_dointvec,
 699         },
 700 #endif
 701 #ifdef CONFIG_MAGIC_SYSRQ
 702         {
 703                 .procname       = "sysrq",
 704                 .data           = &__sysrq_enabled,
 705                 .maxlen         = sizeof (int),
 706                 .mode           = 0644,
 707                 .proc_handler   = sysrq_sysctl_handler,
 708         },
 709 #endif
 710 #ifdef CONFIG_PROC_SYSCTL
 711         {
 712                 .procname       = "cad_pid",
 713                 .data           = NULL,
 714                 .maxlen         = sizeof (int),
 715                 .mode           = 0600,
 716                 .proc_handler   = proc_do_cad_pid,
 717         },
 718 #endif
 719         {
 720                 .procname       = "threads-max",
 721                 .data           = NULL,
 722                 .maxlen         = sizeof(int),
 723                 .mode           = 0644,
 724                 .proc_handler   = sysctl_max_threads,
 725         },
 726         {
 727                 .procname       = "random",
 728                 .mode           = 0555,
 729                 .child          = random_table,
 730         },
 731         {
 732                 .procname       = "usermodehelper",
 733                 .mode           = 0555,
 734                 .child          = usermodehelper_table,
 735         },
 736         {
 737                 .procname       = "overflowuid",
 738                 .data           = &overflowuid,
 739                 .maxlen         = sizeof(int),
 740                 .mode           = 0644,
 741                 .proc_handler   = proc_dointvec_minmax,
 742                 .extra1         = &minolduid,
 743                 .extra2         = &maxolduid,
 744         },
 745         {
 746                 .procname       = "overflowgid",
 747                 .data           = &overflowgid,
 748                 .maxlen         = sizeof(int),
 749                 .mode           = 0644,
 750                 .proc_handler   = proc_dointvec_minmax,
 751                 .extra1         = &minolduid,
 752                 .extra2         = &maxolduid,
 753         },
 754 #ifdef CONFIG_S390
 755 #ifdef CONFIG_MATHEMU
 756         {
 757                 .procname       = "ieee_emulation_warnings",
 758                 .data           = &sysctl_ieee_emulation_warnings,
 759                 .maxlen         = sizeof(int),
 760                 .mode           = 0644,
 761                 .proc_handler   = proc_dointvec,
 762         },
 763 #endif
 764         {
 765                 .procname       = "userprocess_debug",
 766                 .data           = &show_unhandled_signals,
 767                 .maxlen         = sizeof(int),
 768                 .mode           = 0644,
 769                 .proc_handler   = proc_dointvec,
 770         },
 771 #endif
 772         {
 773                 .procname       = "pid_max",
 774                 .data           = &pid_max,
 775                 .maxlen         = sizeof (int),
 776                 .mode           = 0644,
 777                 .proc_handler   = proc_dointvec_minmax,
 778                 .extra1         = &pid_max_min,
 779                 .extra2         = &pid_max_max,
 780         },
 781         {
 782                 .procname       = "panic_on_oops",
 783                 .data           = &panic_on_oops,
 784                 .maxlen         = sizeof(int),
 785                 .mode           = 0644,
 786                 .proc_handler   = proc_dointvec,
 787         },
 788 #if defined CONFIG_PRINTK
 789         {
 790                 .procname       = "printk",
 791                 .data           = &console_loglevel,
 792                 .maxlen         = 4*sizeof(int),
 793                 .mode           = 0644,
 794                 .proc_handler   = proc_dointvec,
 795         },
 796         {
 797                 .procname       = "printk_ratelimit",
 798                 .data           = &printk_ratelimit_state.interval,
 799                 .maxlen         = sizeof(int),
 800                 .mode           = 0644,
 801                 .proc_handler   = proc_dointvec_jiffies,
 802         },
 803         {
 804                 .procname       = "printk_ratelimit_burst",
 805                 .data           = &printk_ratelimit_state.burst,
 806                 .maxlen         = sizeof(int),
 807                 .mode           = 0644,
 808                 .proc_handler   = proc_dointvec,
 809         },
 810         {
 811                 .procname       = "printk_delay",
 812                 .data           = &printk_delay_msec,
 813                 .maxlen         = sizeof(int),
 814                 .mode           = 0644,
 815                 .proc_handler   = proc_dointvec_minmax,
 816                 .extra1         = &zero,
 817                 .extra2         = &ten_thousand,
 818         },
 819         {
 820                 .procname       = "printk_devkmsg",
 821                 .data           = devkmsg_log_str,
 822                 .maxlen         = DEVKMSG_STR_MAX_SIZE,
 823                 .mode           = 0644,
 824                 .proc_handler   = devkmsg_sysctl_set_loglvl,
 825         },
 826         {
 827                 .procname       = "dmesg_restrict",
 828                 .data           = &dmesg_restrict,
 829                 .maxlen         = sizeof(int),
 830                 .mode           = 0644,
 831                 .proc_handler   = proc_dointvec_minmax_sysadmin,
 832                 .extra1         = &zero,
 833                 .extra2         = &one,
 834         },
 835         {
 836                 .procname       = "kptr_restrict",
 837                 .data           = &kptr_restrict,
 838                 .maxlen         = sizeof(int),
 839                 .mode           = 0644,
 840                 .proc_handler   = proc_dointvec_minmax_sysadmin,
 841                 .extra1         = &zero,
 842                 .extra2         = &two,
 843         },
 844 #endif
 845         {
 846                 .procname       = "ngroups_max",
 847                 .data           = &ngroups_max,
 848                 .maxlen         = sizeof (int),
 849                 .mode           = 0444,
 850                 .proc_handler   = proc_dointvec,
 851         },
 852         {
 853                 .procname       = "cap_last_cap",
 854                 .data           = (void *)&cap_last_cap,
 855                 .maxlen         = sizeof(int),
 856                 .mode           = 0444,
 857                 .proc_handler   = proc_dointvec,
 858         },
 859 #if defined(CONFIG_LOCKUP_DETECTOR)
 860         {
 861                 .procname       = "watchdog",
 862                 .data           = &watchdog_user_enabled,
 863                 .maxlen         = sizeof (int),
 864                 .mode           = 0644,
 865                 .proc_handler   = proc_watchdog,
 866                 .extra1         = &zero,
 867                 .extra2         = &one,
 868         },
 869         {
 870                 .procname       = "watchdog_thresh",
 871                 .data           = &watchdog_thresh,
 872                 .maxlen         = sizeof(int),
 873                 .mode           = 0644,
 874                 .proc_handler   = proc_watchdog_thresh,
 875                 .extra1         = &zero,
 876                 .extra2         = &sixty,
 877         },
 878         {
 879                 .procname       = "nmi_watchdog",
 880                 .data           = &nmi_watchdog_enabled,
 881                 .maxlen         = sizeof (int),
 882                 .mode           = 0644,
 883                 .proc_handler   = proc_nmi_watchdog,
 884                 .extra1         = &zero,
 885 #if defined(CONFIG_HAVE_NMI_WATCHDOG) || defined(CONFIG_HARDLOCKUP_DETECTOR)
 886                 .extra2         = &one,
 887 #else
 888                 .extra2         = &zero,
 889 #endif
 890         },
 891         {
 892                 .procname       = "soft_watchdog",
 893                 .data           = &soft_watchdog_enabled,
 894                 .maxlen         = sizeof (int),
 895                 .mode           = 0644,
 896                 .proc_handler   = proc_soft_watchdog,
 897                 .extra1         = &zero,
 898                 .extra2         = &one,
 899         },
 900         {
 901                 .procname       = "watchdog_cpumask",
 902                 .data           = &watchdog_cpumask_bits,
 903                 .maxlen         = NR_CPUS,
 904                 .mode           = 0644,
 905                 .proc_handler   = proc_watchdog_cpumask,
 906         },
 907         {
 908                 .procname       = "softlockup_panic",
 909                 .data           = &softlockup_panic,
 910                 .maxlen         = sizeof(int),
 911                 .mode           = 0644,
 912                 .proc_handler   = proc_dointvec_minmax,
 913                 .extra1         = &zero,
 914                 .extra2         = &one,
 915         },
 916 #ifdef CONFIG_HARDLOCKUP_DETECTOR
 917         {
 918                 .procname       = "hardlockup_panic",
 919                 .data           = &hardlockup_panic,
 920                 .maxlen         = sizeof(int),
 921                 .mode           = 0644,
 922                 .proc_handler   = proc_dointvec_minmax,
 923                 .extra1         = &zero,
 924                 .extra2         = &one,
 925         },
 926 #endif
 927 #ifdef CONFIG_SMP
 928         {
 929                 .procname       = "softlockup_all_cpu_backtrace",
 930                 .data           = &sysctl_softlockup_all_cpu_backtrace,
 931                 .maxlen         = sizeof(int),
 932                 .mode           = 0644,
 933                 .proc_handler   = proc_dointvec_minmax,
 934                 .extra1         = &zero,
 935                 .extra2         = &one,
 936         },
 937         {
 938                 .procname       = "hardlockup_all_cpu_backtrace",
 939                 .data           = &sysctl_hardlockup_all_cpu_backtrace,
 940                 .maxlen         = sizeof(int),
 941                 .mode           = 0644,
 942                 .proc_handler   = proc_dointvec_minmax,
 943                 .extra1         = &zero,
 944                 .extra2         = &one,
 945         },
 946 #endif /* CONFIG_SMP */
 947 #endif
 948 #if defined(CONFIG_X86_LOCAL_APIC) && defined(CONFIG_X86)
 949         {
 950                 .procname       = "unknown_nmi_panic",
 951                 .data           = &unknown_nmi_panic,
 952                 .maxlen         = sizeof (int),
 953                 .mode           = 0644,
 954                 .proc_handler   = proc_dointvec,
 955         },
 956 #endif
 957 #if defined(CONFIG_X86)
 958         {
 959                 .procname       = "panic_on_unrecovered_nmi",
 960                 .data           = &panic_on_unrecovered_nmi,
 961                 .maxlen         = sizeof(int),
 962                 .mode           = 0644,
 963                 .proc_handler   = proc_dointvec,
 964         },
 965         {
 966                 .procname       = "panic_on_io_nmi",
 967                 .data           = &panic_on_io_nmi,
 968                 .maxlen         = sizeof(int),
 969                 .mode           = 0644,
 970                 .proc_handler   = proc_dointvec,
 971         },
 972 #ifdef CONFIG_DEBUG_STACKOVERFLOW
 973         {
 974                 .procname       = "panic_on_stackoverflow",
 975                 .data           = &sysctl_panic_on_stackoverflow,
 976                 .maxlen         = sizeof(int),
 977                 .mode           = 0644,
 978                 .proc_handler   = proc_dointvec,
 979         },
 980 #endif
 981         {
 982                 .procname       = "bootloader_type",
 983                 .data           = &bootloader_type,
 984                 .maxlen         = sizeof (int),
 985                 .mode           = 0444,
 986                 .proc_handler   = proc_dointvec,
 987         },
 988         {
 989                 .procname       = "bootloader_version",
 990                 .data           = &bootloader_version,
 991                 .maxlen         = sizeof (int),
 992                 .mode           = 0444,
 993                 .proc_handler   = proc_dointvec,
 994         },
 995         {
 996                 .procname       = "kstack_depth_to_print",
 997                 .data           = &kstack_depth_to_print,
 998                 .maxlen         = sizeof(int),
 999                 .mode           = 0644,
1000                 .proc_handler   = proc_dointvec,
1001         },
1002         {
1003                 .procname       = "io_delay_type",
1004                 .data           = &io_delay_type,
1005                 .maxlen         = sizeof(int),
1006                 .mode           = 0644,
1007                 .proc_handler   = proc_dointvec,
1008         },
1009 #endif
1010 #if defined(CONFIG_MMU)
1011         {
1012                 .procname       = "randomize_va_space",
1013                 .data           = &randomize_va_space,
1014                 .maxlen         = sizeof(int),
1015                 .mode           = 0644,
1016                 .proc_handler   = proc_dointvec,
1017         },
1018 #endif
1019 #if defined(CONFIG_S390) && defined(CONFIG_SMP)
1020         {
1021                 .procname       = "spin_retry",
1022                 .data           = &spin_retry,
1023                 .maxlen         = sizeof (int),
1024                 .mode           = 0644,
1025                 .proc_handler   = proc_dointvec,
1026         },
1027 #endif
1028 #if     defined(CONFIG_ACPI_SLEEP) && defined(CONFIG_X86)
1029         {
1030                 .procname       = "acpi_video_flags",
1031                 .data           = &acpi_realmode_flags,
1032                 .maxlen         = sizeof (unsigned long),
1033                 .mode           = 0644,
1034                 .proc_handler   = proc_doulongvec_minmax,
1035         },
1036 #endif
1037 #ifdef CONFIG_SYSCTL_ARCH_UNALIGN_NO_WARN
1038         {
1039                 .procname       = "ignore-unaligned-usertrap",
1040                 .data           = &no_unaligned_warning,
1041                 .maxlen         = sizeof (int),
1042                 .mode           = 0644,
1043                 .proc_handler   = proc_dointvec,
1044         },
1045 #endif
1046 #ifdef CONFIG_IA64
1047         {
1048                 .procname       = "unaligned-dump-stack",
1049                 .data           = &unaligned_dump_stack,
1050                 .maxlen         = sizeof (int),
1051                 .mode           = 0644,
1052                 .proc_handler   = proc_dointvec,
1053         },
1054 #endif
1055 #ifdef CONFIG_DETECT_HUNG_TASK
1056         {
1057                 .procname       = "hung_task_panic",
1058                 .data           = &sysctl_hung_task_panic,
1059                 .maxlen         = sizeof(int),
1060                 .mode           = 0644,
1061                 .proc_handler   = proc_dointvec_minmax,
1062                 .extra1         = &zero,
1063                 .extra2         = &one,
1064         },
1065         {
1066                 .procname       = "hung_task_check_count",
1067                 .data           = &sysctl_hung_task_check_count,
1068                 .maxlen         = sizeof(int),
1069                 .mode           = 0644,
1070                 .proc_handler   = proc_dointvec_minmax,
1071                 .extra1         = &zero,
1072         },
1073         {
1074                 .procname       = "hung_task_timeout_secs",
1075                 .data           = &sysctl_hung_task_timeout_secs,
1076                 .maxlen         = sizeof(unsigned long),
1077                 .mode           = 0644,
1078                 .proc_handler   = proc_dohung_task_timeout_secs,
1079                 .extra2         = &hung_task_timeout_max,
1080         },
1081         {
1082                 .procname       = "hung_task_warnings",
1083                 .data           = &sysctl_hung_task_warnings,
1084                 .maxlen         = sizeof(int),
1085                 .mode           = 0644,
1086                 .proc_handler   = proc_dointvec_minmax,
1087                 .extra1         = &neg_one,
1088         },
1089 #endif
1090 #ifdef CONFIG_RT_MUTEXES
1091         {
1092                 .procname       = "max_lock_depth",
1093                 .data           = &max_lock_depth,
1094                 .maxlen         = sizeof(int),
1095                 .mode           = 0644,
1096                 .proc_handler   = proc_dointvec,
1097         },
1098 #endif
1099         {
1100                 .procname       = "poweroff_cmd",
1101                 .data           = &poweroff_cmd,
1102                 .maxlen         = POWEROFF_CMD_PATH_LEN,
1103                 .mode           = 0644,
1104                 .proc_handler   = proc_dostring,
1105         },
1106 #ifdef CONFIG_KEYS
1107         {
1108                 .procname       = "keys",
1109                 .mode           = 0555,
1110                 .child          = key_sysctls,
1111         },
1112 #endif
1113 #ifdef CONFIG_PERF_EVENTS
1114         /*
1115          * User-space scripts rely on the existence of this file
1116          * as a feature check for perf_events being enabled.
1117          *
1118          * So it's an ABI, do not remove!
1119          */
1120         {
1121                 .procname       = "perf_event_paranoid",
1122                 .data           = &sysctl_perf_event_paranoid,
1123                 .maxlen         = sizeof(sysctl_perf_event_paranoid),
1124                 .mode           = 0644,
1125                 .proc_handler   = proc_dointvec,
1126         },
1127         {
1128                 .procname       = "perf_event_mlock_kb",
1129                 .data           = &sysctl_perf_event_mlock,
1130                 .maxlen         = sizeof(sysctl_perf_event_mlock),
1131                 .mode           = 0644,
1132                 .proc_handler   = proc_dointvec,
1133         },
1134         {
1135                 .procname       = "perf_event_max_sample_rate",
1136                 .data           = &sysctl_perf_event_sample_rate,
1137                 .maxlen         = sizeof(sysctl_perf_event_sample_rate),
1138                 .mode           = 0644,
1139                 .proc_handler   = perf_proc_update_handler,
1140                 .extra1         = &one,
1141         },
1142         {
1143                 .procname       = "perf_cpu_time_max_percent",
1144                 .data           = &sysctl_perf_cpu_time_max_percent,
1145                 .maxlen         = sizeof(sysctl_perf_cpu_time_max_percent),
1146                 .mode           = 0644,
1147                 .proc_handler   = perf_cpu_time_max_percent_handler,
1148                 .extra1         = &zero,
1149                 .extra2         = &one_hundred,
1150         },
1151         {
1152                 .procname       = "perf_event_max_stack",
1153                 .data           = &sysctl_perf_event_max_stack,
1154                 .maxlen         = sizeof(sysctl_perf_event_max_stack),
1155                 .mode           = 0644,
1156                 .proc_handler   = perf_event_max_stack_handler,
1157                 .extra1         = &zero,
1158                 .extra2         = &six_hundred_forty_kb,
1159         },
1160         {
1161                 .procname       = "perf_event_max_contexts_per_stack",
1162                 .data           = &sysctl_perf_event_max_contexts_per_stack,
1163                 .maxlen         = sizeof(sysctl_perf_event_max_contexts_per_stack),
1164                 .mode           = 0644,
1165                 .proc_handler   = perf_event_max_stack_handler,
1166                 .extra1         = &zero,
1167                 .extra2         = &one_thousand,
1168         },
1169 #endif
1170 #ifdef CONFIG_KMEMCHECK
1171         {
1172                 .procname       = "kmemcheck",
1173                 .data           = &kmemcheck_enabled,
1174                 .maxlen         = sizeof(int),
1175                 .mode           = 0644,
1176                 .proc_handler   = proc_dointvec,
1177         },
1178 #endif
1179         {
1180                 .procname       = "panic_on_warn",
1181                 .data           = &panic_on_warn,
1182                 .maxlen         = sizeof(int),
1183                 .mode           = 0644,
1184                 .proc_handler   = proc_dointvec_minmax,
1185                 .extra1         = &zero,
1186                 .extra2         = &one,
1187         },
1188 #if defined(CONFIG_SMP) && defined(CONFIG_NO_HZ_COMMON)
1189         {
1190                 .procname       = "timer_migration",
1191                 .data           = &sysctl_timer_migration,
1192                 .maxlen         = sizeof(unsigned int),
1193                 .mode           = 0644,
1194                 .proc_handler   = timer_migration_handler,
1195                 .extra1         = &zero,
1196                 .extra2         = &one,
1197         },
1198 #endif
1199 #ifdef CONFIG_BPF_SYSCALL
1200         {
1201                 .procname       = "unprivileged_bpf_disabled",
1202                 .data           = &sysctl_unprivileged_bpf_disabled,
1203                 .maxlen         = sizeof(sysctl_unprivileged_bpf_disabled),
1204                 .mode           = 0644,
1205                 /* only handle a transition from default "0" to "1" */
1206                 .proc_handler   = proc_dointvec_minmax,
1207                 .extra1         = &one,
1208                 .extra2         = &one,
1209         },
1210 #endif
1211 #if defined(CONFIG_TREE_RCU) || defined(CONFIG_PREEMPT_RCU)
1212         {
1213                 .procname       = "panic_on_rcu_stall",
1214                 .data           = &sysctl_panic_on_rcu_stall,
1215                 .maxlen         = sizeof(sysctl_panic_on_rcu_stall),
1216                 .mode           = 0644,
1217                 .proc_handler   = proc_dointvec_minmax,
1218                 .extra1         = &zero,
1219                 .extra2         = &one,
1220         },
1221 #endif
1222         { }
1223 };
1224
1225 static struct ctl_table vm_table[] = {
1226         {
1227                 .procname       = "overcommit_memory",
1228                 .data           = &sysctl_overcommit_memory,
1229                 .maxlen         = sizeof(sysctl_overcommit_memory),
1230                 .mode           = 0644,
1231                 .proc_handler   = proc_dointvec_minmax,
1232                 .extra1         = &zero,
1233                 .extra2         = &two,
1234         },
1235         {
1236                 .procname       = "panic_on_oom",
1237                 .data           = &sysctl_panic_on_oom,
1238                 .maxlen         = sizeof(sysctl_panic_on_oom),
1239                 .mode           = 0644,
1240                 .proc_handler   = proc_dointvec_minmax,
1241                 .extra1         = &zero,
1242                 .extra2         = &two,
1243         },
1244         {
1245                 .procname       = "oom_kill_allocating_task",
1246                 .data           = &sysctl_oom_kill_allocating_task,
1247                 .maxlen         = sizeof(sysctl_oom_kill_allocating_task),
1248                 .mode           = 0644,
1249                 .proc_handler   = proc_dointvec,
1250         },
1251         {
1252                 .procname       = "oom_dump_tasks",
1253                 .data           = &sysctl_oom_dump_tasks,
1254                 .maxlen         = sizeof(sysctl_oom_dump_tasks),
1255                 .mode           = 0644,
1256                 .proc_handler   = proc_dointvec,
1257         },
1258         {
1259                 .procname       = "overcommit_ratio",
1260                 .data           = &sysctl_overcommit_ratio,
1261                 .maxlen         = sizeof(sysctl_overcommit_ratio),
1262                 .mode           = 0644,
1263                 .proc_handler   = overcommit_ratio_handler,
1264         },
1265         {
1266                 .procname       = "overcommit_kbytes",
1267                 .data           = &sysctl_overcommit_kbytes,
1268                 .maxlen         = sizeof(sysctl_overcommit_kbytes),
1269                 .mode           = 0644,
1270                 .proc_handler   = overcommit_kbytes_handler,
1271         },
1272         {
1273                 .procname       = "page-cluster",
1274                 .data           = &page_cluster,
1275                 .maxlen         = sizeof(int),
1276                 .mode           = 0644,
1277                 .proc_handler   = proc_dointvec_minmax,
1278                 .extra1         = &zero,
1279         },
1280         {
1281                 .procname       = "dirty_background_ratio",
1282                 .data           = &dirty_background_ratio,
1283                 .maxlen         = sizeof(dirty_background_ratio),
1284                 .mode           = 0644,
1285                 .proc_handler   = dirty_background_ratio_handler,
1286                 .extra1         = &zero,
1287                 .extra2         = &one_hundred,
1288         },
1289         {
1290                 .procname       = "dirty_background_bytes",
1291                 .data           = &dirty_background_bytes,
1292                 .maxlen         = sizeof(dirty_background_bytes),
1293                 .mode           = 0644,
1294                 .proc_handler   = dirty_background_bytes_handler,
1295                 .extra1         = &one_ul,
1296         },
1297         {
1298                 .procname       = "dirty_ratio",
1299                 .data           = &vm_dirty_ratio,
1300                 .maxlen         = sizeof(vm_dirty_ratio),
1301                 .mode           = 0644,
1302                 .proc_handler   = dirty_ratio_handler,
1303                 .extra1         = &zero,
1304                 .extra2         = &one_hundred,
1305         },
1306         {
1307                 .procname       = "dirty_bytes",
1308                 .data           = &vm_dirty_bytes,
1309                 .maxlen         = sizeof(vm_dirty_bytes),
1310                 .mode           = 0644,
1311                 .proc_handler   = dirty_bytes_handler,
1312                 .extra1         = &dirty_bytes_min,
1313         },
1314         {
1315                 .procname       = "dirty_writeback_centisecs",
1316                 .data           = &dirty_writeback_interval,
1317                 .maxlen         = sizeof(dirty_writeback_interval),
1318                 .mode           = 0644,
1319                 .proc_handler   = dirty_writeback_centisecs_handler,
1320         },
1321         {
1322                 .procname       = "dirty_expire_centisecs",
1323                 .data           = &dirty_expire_interval,
1324                 .maxlen         = sizeof(dirty_expire_interval),
1325                 .mode           = 0644,
1326                 .proc_handler   = proc_dointvec_minmax,
1327                 .extra1         = &zero,
1328         },
1329         {
1330                 .procname       = "dirtytime_expire_seconds",
1331                 .data           = &dirtytime_expire_interval,
1332                 .maxlen         = sizeof(dirty_expire_interval),
1333                 .mode           = 0644,
1334                 .proc_handler   = dirtytime_interval_handler,
1335                 .extra1         = &zero,
1336         },
1337         {
1338                 .procname       = "nr_pdflush_threads",
1339                 .mode           = 0444 /* read-only */,
1340                 .proc_handler   = pdflush_proc_obsolete,
1341         },
1342         {
1343                 .procname       = "swappiness",
1344                 .data           = &vm_swappiness,
1345                 .maxlen         = sizeof(vm_swappiness),
1346                 .mode           = 0644,
1347                 .proc_handler   = proc_dointvec_minmax,
1348                 .extra1         = &zero,
1349                 .extra2         = &one_hundred,
1350         },
1351 #ifdef CONFIG_HUGETLB_PAGE
1352         {
1353                 .procname       = "nr_hugepages",
1354                 .data           = NULL,
1355                 .maxlen         = sizeof(unsigned long),
1356                 .mode           = 0644,
1357                 .proc_handler   = hugetlb_sysctl_handler,
1358         },
1359 #ifdef CONFIG_NUMA
1360         {
1361                 .procname       = "nr_hugepages_mempolicy",
1362                 .data           = NULL,
1363                 .maxlen         = sizeof(unsigned long),
1364                 .mode           = 0644,
1365                 .proc_handler   = &hugetlb_mempolicy_sysctl_handler,
1366         },
1367 #endif
1368          {
1369                 .procname       = "hugetlb_shm_group",
1370                 .data           = &sysctl_hugetlb_shm_group,
1371                 .maxlen         = sizeof(gid_t),
1372                 .mode           = 0644,
1373                 .proc_handler   = proc_dointvec,
1374          },
1375          {
1376                 .procname       = "hugepages_treat_as_movable",
1377                 .data           = &hugepages_treat_as_movable,
1378                 .maxlen         = sizeof(int),
1379                 .mode           = 0644,
1380                 .proc_handler   = proc_dointvec,
1381         },
1382         {
1383                 .procname       = "nr_overcommit_hugepages",
1384                 .data           = NULL,
1385                 .maxlen         = sizeof(unsigned long),
1386                 .mode           = 0644,
1387                 .proc_handler   = hugetlb_overcommit_handler,
1388         },
1389 #endif
1390         {
1391                 .procname       = "lowmem_reserve_ratio",
1392                 .data           = &sysctl_lowmem_reserve_ratio,
1393                 .maxlen         = sizeof(sysctl_lowmem_reserve_ratio),
1394                 .mode           = 0644,
1395                 .proc_handler   = lowmem_reserve_ratio_sysctl_handler,
1396         },
1397         {
1398                 .procname       = "drop_caches",
1399                 .data           = &sysctl_drop_caches,
1400                 .maxlen         = sizeof(int),
1401                 .mode           = 0644,
1402                 .proc_handler   = drop_caches_sysctl_handler,
1403                 .extra1         = &one,
1404                 .extra2         = &four,
1405         },
1406 #ifdef CONFIG_COMPACTION
1407         {
1408                 .procname       = "compact_memory",
1409                 .data           = &sysctl_compact_memory,
1410                 .maxlen         = sizeof(int),
1411                 .mode           = 0200,
1412                 .proc_handler   = sysctl_compaction_handler,
1413         },
1414         {
1415                 .procname       = "extfrag_threshold",
1416                 .data           = &sysctl_extfrag_threshold,
1417                 .maxlen         = sizeof(int),
1418                 .mode           = 0644,
1419                 .proc_handler   = sysctl_extfrag_handler,
1420                 .extra1         = &min_extfrag_threshold,
1421                 .extra2         = &max_extfrag_threshold,
1422         },
1423         {
1424                 .procname       = "compact_unevictable_allowed",
1425                 .data           = &sysctl_compact_unevictable_allowed,
1426                 .maxlen         = sizeof(int),
1427                 .mode           = 0644,
1428                 .proc_handler   = proc_dointvec,
1429                 .extra1         = &zero,
1430                 .extra2         = &one,
1431         },
1432
1433 #endif /* CONFIG_COMPACTION */
1434         {
1435                 .procname       = "min_free_kbytes",
1436                 .data           = &min_free_kbytes,
1437                 .maxlen         = sizeof(min_free_kbytes),
1438                 .mode           = 0644,
1439                 .proc_handler   = min_free_kbytes_sysctl_handler,
1440                 .extra1         = &zero,
1441         },
1442         {
1443                 .procname       = "watermark_scale_factor",
1444                 .data           = &watermark_scale_factor,
1445                 .maxlen         = sizeof(watermark_scale_factor),
1446                 .mode           = 0644,
1447                 .proc_handler   = watermark_scale_factor_sysctl_handler,
1448                 .extra1         = &one,
1449                 .extra2         = &one_thousand,
1450         },
1451         {
1452                 .procname       = "percpu_pagelist_fraction",
1453                 .data           = &percpu_pagelist_fraction,
1454                 .maxlen         = sizeof(percpu_pagelist_fraction),
1455                 .mode           = 0644,
1456                 .proc_handler   = percpu_pagelist_fraction_sysctl_handler,
1457                 .extra1         = &zero,
1458         },
1459 #ifdef CONFIG_MMU
1460         {
1461                 .procname       = "max_map_count",
1462                 .data           = &sysctl_max_map_count,
1463                 .maxlen         = sizeof(sysctl_max_map_count),
1464                 .mode           = 0644,
1465                 .proc_handler   = proc_dointvec_minmax,
1466                 .extra1         = &zero,
1467         },
1468 #else
1469         {
1470                 .procname       = "nr_trim_pages",
1471                 .data           = &sysctl_nr_trim_pages,
1472                 .maxlen         = sizeof(sysctl_nr_trim_pages),
1473                 .mode           = 0644,
1474                 .proc_handler   = proc_dointvec_minmax,
1475                 .extra1         = &zero,
1476         },
1477 #endif
1478         {
1479                 .procname       = "laptop_mode",
1480                 .data           = &laptop_mode,
1481                 .maxlen         = sizeof(laptop_mode),
1482                 .mode           = 0644,
1483                 .proc_handler   = proc_dointvec_jiffies,
1484         },
1485         {
1486                 .procname       = "block_dump",
1487                 .data           = &block_dump,
1488                 .maxlen         = sizeof(block_dump),
1489                 .mode           = 0644,
1490                 .proc_handler   = proc_dointvec,
1491                 .extra1         = &zero,
1492         },
1493         {
1494                 .procname       = "vfs_cache_pressure",
1495                 .data           = &sysctl_vfs_cache_pressure,
1496                 .maxlen         = sizeof(sysctl_vfs_cache_pressure),
1497                 .mode           = 0644,
1498                 .proc_handler   = proc_dointvec,
1499                 .extra1         = &zero,
1500         },
1501 #ifdef HAVE_ARCH_PICK_MMAP_LAYOUT
1502         {
1503                 .procname       = "legacy_va_layout",
1504                 .data           = &sysctl_legacy_va_layout,
1505                 .maxlen         = sizeof(sysctl_legacy_va_layout),
1506                 .mode           = 0644,
1507                 .proc_handler   = proc_dointvec,
1508                 .extra1         = &zero,
1509         },
1510 #endif
1511 #ifdef CONFIG_NUMA
1512         {
1513                 .procname       = "zone_reclaim_mode",
1514                 .data           = &node_reclaim_mode,
1515                 .maxlen         = sizeof(node_reclaim_mode),
1516                 .mode           = 0644,
1517                 .proc_handler   = proc_dointvec,
1518                 .extra1         = &zero,
1519         },
1520         {
1521                 .procname       = "min_unmapped_ratio",
1522                 .data           = &sysctl_min_unmapped_ratio,
1523                 .maxlen         = sizeof(sysctl_min_unmapped_ratio),
1524                 .mode           = 0644,
1525                 .proc_handler   = sysctl_min_unmapped_ratio_sysctl_handler,
1526                 .extra1         = &zero,
1527                 .extra2         = &one_hundred,
1528         },
1529         {
1530                 .procname       = "min_slab_ratio",
1531                 .data           = &sysctl_min_slab_ratio,
1532                 .maxlen         = sizeof(sysctl_min_slab_ratio),
1533                 .mode           = 0644,
1534                 .proc_handler   = sysctl_min_slab_ratio_sysctl_handler,
1535                 .extra1         = &zero,
1536                 .extra2         = &one_hundred,
1537         },
1538 #endif
1539 #ifdef CONFIG_SMP
1540         {
1541                 .procname       = "stat_interval",
1542                 .data           = &sysctl_stat_interval,
1543                 .maxlen         = sizeof(sysctl_stat_interval),
1544                 .mode           = 0644,
1545                 .proc_handler   = proc_dointvec_jiffies,
1546         },
1547         {
1548                 .procname       = "stat_refresh",
1549                 .data           = NULL,
1550                 .maxlen         = 0,
1551                 .mode           = 0600,
1552                 .proc_handler   = vmstat_refresh,
1553         },
1554 #endif
1555 #ifdef CONFIG_MMU
1556         {
1557                 .procname       = "mmap_min_addr",
1558                 .data           = &dac_mmap_min_addr,
1559                 .maxlen         = sizeof(unsigned long),
1560                 .mode           = 0644,
1561                 .proc_handler   = mmap_min_addr_handler,
1562         },
1563 #endif
1564 #ifdef CONFIG_NUMA
1565         {
1566                 .procname       = "numa_zonelist_order",
1567                 .data           = &numa_zonelist_order,
1568                 .maxlen         = NUMA_ZONELIST_ORDER_LEN,
1569                 .mode           = 0644,
1570                 .proc_handler   = numa_zonelist_order_handler,
1571         },
1572 #endif
1573 #if (defined(CONFIG_X86_32) && !defined(CONFIG_UML))|| \
1574    (defined(CONFIG_SUPERH) && defined(CONFIG_VSYSCALL))
1575         {
1576                 .procname       = "vdso_enabled",
1577 #ifdef CONFIG_X86_32
1578                 .data           = &vdso32_enabled,
1579                 .maxlen         = sizeof(vdso32_enabled),
1580 #else
1581                 .data           = &vdso_enabled,
1582                 .maxlen         = sizeof(vdso_enabled),
1583 #endif
1584                 .mode           = 0644,
1585                 .proc_handler   = proc_dointvec,
1586                 .extra1         = &zero,
1587         },
1588 #endif
1589 #ifdef CONFIG_HIGHMEM
1590         {
1591                 .procname       = "highmem_is_dirtyable",
1592                 .data           = &vm_highmem_is_dirtyable,
1593                 .maxlen         = sizeof(vm_highmem_is_dirtyable),
1594                 .mode           = 0644,
1595                 .proc_handler   = proc_dointvec_minmax,
1596                 .extra1         = &zero,
1597                 .extra2         = &one,
1598         },
1599 #endif
1600 #ifdef CONFIG_MEMORY_FAILURE
1601         {
1602                 .procname       = "memory_failure_early_kill",
1603                 .data           = &sysctl_memory_failure_early_kill,
1604                 .maxlen         = sizeof(sysctl_memory_failure_early_kill),
1605                 .mode           = 0644,
1606                 .proc_handler   = proc_dointvec_minmax,
1607                 .extra1         = &zero,
1608                 .extra2         = &one,
1609         },
1610         {
1611                 .procname       = "memory_failure_recovery",
1612                 .data           = &sysctl_memory_failure_recovery,
1613                 .maxlen         = sizeof(sysctl_memory_failure_recovery),
1614                 .mode           = 0644,
1615                 .proc_handler   = proc_dointvec_minmax,
1616                 .extra1         = &zero,
1617                 .extra2         = &one,
1618         },
1619 #endif
1620         {
1621                 .procname       = "user_reserve_kbytes",
1622                 .data           = &sysctl_user_reserve_kbytes,
1623                 .maxlen         = sizeof(sysctl_user_reserve_kbytes),
1624                 .mode           = 0644,
1625                 .proc_handler   = proc_doulongvec_minmax,
1626         },
1627         {
1628                 .procname       = "admin_reserve_kbytes",
1629                 .data           = &sysctl_admin_reserve_kbytes,
1630                 .maxlen         = sizeof(sysctl_admin_reserve_kbytes),
1631                 .mode           = 0644,
1632                 .proc_handler   = proc_doulongvec_minmax,
1633         },
1634 #ifdef CONFIG_HAVE_ARCH_MMAP_RND_BITS
1635         {
1636                 .procname       = "mmap_rnd_bits",
1637                 .data           = &mmap_rnd_bits,
1638                 .maxlen         = sizeof(mmap_rnd_bits),
1639                 .mode           = 0600,
1640                 .proc_handler   = proc_dointvec_minmax,
1641                 .extra1         = (void *)&mmap_rnd_bits_min,
1642                 .extra2         = (void *)&mmap_rnd_bits_max,
1643         },
1644 #endif
1645 #ifdef CONFIG_HAVE_ARCH_MMAP_RND_COMPAT_BITS
1646         {
1647                 .procname       = "mmap_rnd_compat_bits",
1648                 .data           = &mmap_rnd_compat_bits,
1649                 .maxlen         = sizeof(mmap_rnd_compat_bits),
1650                 .mode           = 0600,
1651                 .proc_handler   = proc_dointvec_minmax,
1652                 .extra1         = (void *)&mmap_rnd_compat_bits_min,
1653                 .extra2         = (void *)&mmap_rnd_compat_bits_max,
1654         },
1655 #endif
1656         { }
1657 };
1658
1659 static struct ctl_table fs_table[] = {
1660         {
1661                 .procname       = "inode-nr",
1662                 .data           = &inodes_stat,
1663                 .maxlen         = 2*sizeof(long),
1664                 .mode           = 0444,
1665                 .proc_handler   = proc_nr_inodes,
1666         },
1667         {
1668                 .procname       = "inode-state",
1669                 .data           = &inodes_stat,
1670                 .maxlen         = 7*sizeof(long),
1671                 .mode           = 0444,
1672                 .proc_handler   = proc_nr_inodes,
1673         },
1674         {
1675                 .procname       = "file-nr",
1676                 .data           = &files_stat,
1677                 .maxlen         = sizeof(files_stat),
1678                 .mode           = 0444,
1679                 .proc_handler   = proc_nr_files,
1680         },
1681         {
1682                 .procname       = "file-max",
1683                 .data           = &files_stat.max_files,
1684                 .maxlen         = sizeof(files_stat.max_files),
1685                 .mode           = 0644,
1686                 .proc_handler   = proc_doulongvec_minmax,
1687                 .extra1         = &zero_ul,
1688                 .extra2         = &long_max,
1689         },
1690         {
1691                 .procname       = "nr_open",
1692                 .data           = &sysctl_nr_open,
1693                 .maxlen         = sizeof(unsigned int),
1694                 .mode           = 0644,
1695                 .proc_handler   = proc_dointvec_minmax,
1696                 .extra1         = &sysctl_nr_open_min,
1697                 .extra2         = &sysctl_nr_open_max,
1698         },
1699         {
1700                 .procname       = "dentry-state",
1701                 .data           = &dentry_stat,
1702                 .maxlen         = 6*sizeof(long),
1703                 .mode           = 0444,
1704                 .proc_handler   = proc_nr_dentry,
1705         },
1706         {
1707                 .procname       = "overflowuid",
1708                 .data           = &fs_overflowuid,
1709                 .maxlen         = sizeof(int),
1710                 .mode           = 0644,
1711                 .proc_handler   = proc_dointvec_minmax,
1712                 .extra1         = &minolduid,
1713                 .extra2         = &maxolduid,
1714         },
1715         {
1716                 .procname       = "overflowgid",
1717                 .data           = &fs_overflowgid,
1718                 .maxlen         = sizeof(int),
1719                 .mode           = 0644,
1720                 .proc_handler   = proc_dointvec_minmax,
1721                 .extra1         = &minolduid,
1722                 .extra2         = &maxolduid,
1723         },
1724 #ifdef CONFIG_FILE_LOCKING
1725         {
1726                 .procname       = "leases-enable",
1727                 .data           = &leases_enable,
1728                 .maxlen         = sizeof(int),
1729                 .mode           = 0644,
1730                 .proc_handler   = proc_dointvec,
1731         },
1732 #endif
1733 #ifdef CONFIG_DNOTIFY
1734         {
1735                 .procname       = "dir-notify-enable",
1736                 .data           = &dir_notify_enable,
1737                 .maxlen         = sizeof(int),
1738                 .mode           = 0644,
1739                 .proc_handler   = proc_dointvec,
1740         },
1741 #endif
1742 #ifdef CONFIG_MMU
1743 #ifdef CONFIG_FILE_LOCKING
1744         {
1745                 .procname       = "lease-break-time",
1746                 .data           = &lease_break_time,
1747                 .maxlen         = sizeof(int),
1748                 .mode           = 0644,
1749                 .proc_handler   = proc_dointvec,
1750         },
1751 #endif
1752 #ifdef CONFIG_AIO
1753         {
1754                 .procname       = "aio-nr",
1755                 .data           = &aio_nr,
1756                 .maxlen         = sizeof(aio_nr),
1757                 .mode           = 0444,
1758                 .proc_handler   = proc_doulongvec_minmax,
1759         },
1760         {
1761                 .procname       = "aio-max-nr",
1762                 .data           = &aio_max_nr,
1763                 .maxlen         = sizeof(aio_max_nr),
1764                 .mode           = 0644,
1765                 .proc_handler   = proc_doulongvec_minmax,
1766         },
1767 #endif /* CONFIG_AIO */
1768 #ifdef CONFIG_INOTIFY_USER
1769         {
1770                 .procname       = "inotify",
1771                 .mode           = 0555,
1772                 .child          = inotify_table,
1773         },
1774 #endif
1775 #ifdef CONFIG_EPOLL
1776         {
1777                 .procname       = "epoll",
1778                 .mode           = 0555,
1779                 .child          = epoll_table,
1780         },
1781 #endif
1782 #endif
1783         {
1784                 .procname       = "protected_symlinks",
1785                 .data           = &sysctl_protected_symlinks,
1786                 .maxlen         = sizeof(int),
1787                 .mode           = 0600,
1788                 .proc_handler   = proc_dointvec_minmax,
1789                 .extra1         = &zero,
1790                 .extra2         = &one,
1791         },
1792         {
1793                 .procname       = "protected_hardlinks",
1794                 .data           = &sysctl_protected_hardlinks,
1795                 .maxlen         = sizeof(int),
1796                 .mode           = 0600,
1797                 .proc_handler   = proc_dointvec_minmax,
1798                 .extra1         = &zero,
1799                 .extra2         = &one,
1800         },
1801         {
1802                 .procname       = "protected_fifos",
1803                 .data           = &sysctl_protected_fifos,
1804                 .maxlen         = sizeof(int),
1805                 .mode           = 0600,
1806                 .proc_handler   = proc_dointvec_minmax,
1807                 .extra1         = &zero,
1808                 .extra2         = &two,
1809         },
1810         {
1811                 .procname       = "protected_regular",
1812                 .data           = &sysctl_protected_regular,
1813                 .maxlen         = sizeof(int),
1814                 .mode           = 0600,
1815                 .proc_handler   = proc_dointvec_minmax,
1816                 .extra1         = &zero,
1817                 .extra2         = &two,
1818         },
1819         {
1820                 .procname       = "suid_dumpable",
1821                 .data           = &suid_dumpable,
1822                 .maxlen         = sizeof(int),
1823                 .mode           = 0644,
1824                 .proc_handler   = proc_dointvec_minmax_coredump,
1825                 .extra1         = &zero,
1826                 .extra2         = &two,
1827         },
1828 #if defined(CONFIG_BINFMT_MISC) || defined(CONFIG_BINFMT_MISC_MODULE)
1829         {
1830                 .procname       = "binfmt_misc",
1831                 .mode           = 0555,
1832                 .child          = sysctl_mount_point,
1833         },
1834 #endif
1835         {
1836                 .procname       = "pipe-max-size",
1837                 .data           = &pipe_max_size,
1838                 .maxlen         = sizeof(int),
1839                 .mode           = 0644,
1840                 .proc_handler   = &pipe_proc_fn,
1841                 .extra1         = &pipe_min_size,
1842         },
1843         {
1844                 .procname       = "pipe-user-pages-hard",
1845                 .data           = &pipe_user_pages_hard,
1846                 .maxlen         = sizeof(pipe_user_pages_hard),
1847                 .mode           = 0644,
1848                 .proc_handler   = proc_doulongvec_minmax,
1849         },
1850         {
1851                 .procname       = "pipe-user-pages-soft",
1852                 .data           = &pipe_user_pages_soft,
1853                 .maxlen         = sizeof(pipe_user_pages_soft),
1854                 .mode           = 0644,
1855                 .proc_handler   = proc_doulongvec_minmax,
1856         },
1857         {
1858                 .procname       = "mount-max",
1859                 .data           = &sysctl_mount_max,
1860                 .maxlen         = sizeof(unsigned int),
1861                 .mode           = 0644,
1862                 .proc_handler   = proc_dointvec_minmax,
1863                 .extra1         = &one,
1864         },
1865         { }
1866 };
1867
1868 static struct ctl_table debug_table[] = {
1869 #ifdef CONFIG_SYSCTL_EXCEPTION_TRACE
1870         {
1871                 .procname       = "exception-trace",
1872                 .data           = &show_unhandled_signals,
1873                 .maxlen         = sizeof(int),
1874                 .mode           = 0644,
1875                 .proc_handler   = proc_dointvec
1876         },
1877 #endif
1878 #if defined(CONFIG_OPTPROBES)
1879         {
1880                 .procname       = "kprobes-optimization",
1881                 .data           = &sysctl_kprobes_optimization,
1882                 .maxlen         = sizeof(int),
1883                 .mode           = 0644,
1884                 .proc_handler   = proc_kprobes_optimization_handler,
1885                 .extra1         = &zero,
1886                 .extra2         = &one,
1887         },
1888 #endif
1889         { }
1890 };
1891
1892 static struct ctl_table dev_table[] = {
1893         { }
1894 };
1895
1896 int __init sysctl_init(void)
1897 {
1898         struct ctl_table_header *hdr;
1899
1900         hdr = register_sysctl_table(sysctl_base_table);
1901         kmemleak_not_leak(hdr);
1902         return 0;
1903 }
1904
1905 #endif /* CONFIG_SYSCTL */
1906
1907 /*
1908  * /proc/sys support
1909  */
1910
1911 #ifdef CONFIG_PROC_SYSCTL
1912
1913 static int _proc_do_string(char *data, int maxlen, int write,
1914                            char __user *buffer,
1915                            size_t *lenp, loff_t *ppos)
1916 {
1917         size_t len;
1918         char __user *p;
1919         char c;
1920
1921         if (!data || !maxlen || !*lenp) {
1922                 *lenp = 0;
1923                 return 0;
1924         }
1925
1926         if (write) {
1927                 if (sysctl_writes_strict == SYSCTL_WRITES_STRICT) {
1928                         /* Only continue writes not past the end of buffer. */
1929                         len = strlen(data);
1930                         if (len > maxlen - 1)
1931                                 len = maxlen - 1;
1932
1933                         if (*ppos > len)
1934                                 return 0;
1935                         len = *ppos;
1936                 } else {
1937                         /* Start writing from beginning of buffer. */
1938                         len = 0;
1939                 }
1940
1941                 *ppos += *lenp;
1942                 p = buffer;
1943                 while ((p - buffer) < *lenp && len < maxlen - 1) {
1944                         if (get_user(c, p++))
1945                                 return -EFAULT;
1946                         if (c == 0 || c == '\n')
1947                                 break;
1948                         data[len++] = c;
1949                 }
1950                 data[len] = 0;
1951         } else {
1952                 len = strlen(data);
1953                 if (len > maxlen)
1954                         len = maxlen;
1955
1956                 if (*ppos > len) {
1957                         *lenp = 0;
1958                         return 0;
1959                 }
1960
1961                 data += *ppos;
1962                 len  -= *ppos;
1963
1964                 if (len > *lenp)
1965                         len = *lenp;
1966                 if (len)
1967                         if (copy_to_user(buffer, data, len))
1968                                 return -EFAULT;
1969                 if (len < *lenp) {
1970                         if (put_user('\n', buffer + len))
1971                                 return -EFAULT;
1972                         len++;
1973                 }
1974                 *lenp = len;
1975                 *ppos += len;
1976         }
1977         return 0;
1978 }
1979
1980 static void warn_sysctl_write(struct ctl_table *table)
1981 {
1982         pr_warn_once("%s wrote to %s when file position was not 0!\n"
1983                 "This will not be supported in the future. To silence this\n"
1984                 "warning, set kernel.sysctl_writes_strict = -1\n",
1985                 current->comm, table->procname);
1986 }
1987
1988 /**
1989  * proc_dostring - read a string sysctl
1990  * @table: the sysctl table
1991  * @write: %TRUE if this is a write to the sysctl file
1992  * @buffer: the user buffer
1993  * @lenp: the size of the user buffer
1994  * @ppos: file position
1995  *
1996  * Reads/writes a string from/to the user buffer. If the kernel
1997  * buffer provided is not large enough to hold the string, the
1998  * string is truncated. The copied string is %NULL-terminated.
1999  * If the string is being read by the user process, it is copied
2000  * and a newline '\n' is added. It is truncated if the buffer is
2001  * not large enough.
2002  *
2003  * Returns 0 on success.
2004  */
2005 int proc_dostring(struct ctl_table *table, int write,
2006                   void __user *buffer, size_t *lenp, loff_t *ppos)
2007 {
2008         if (write && *ppos && sysctl_writes_strict == SYSCTL_WRITES_WARN)
2009                 warn_sysctl_write(table);
2010
2011         return _proc_do_string((char *)(table->data), table->maxlen, write,
2012                                (char __user *)buffer, lenp, ppos);
2013 }
2014
2015 static size_t proc_skip_spaces(char **buf)
2016 {
2017         size_t ret;
2018         char *tmp = skip_spaces(*buf);
2019         ret = tmp - *buf;
2020         *buf = tmp;
2021         return ret;
2022 }
2023
2024 static void proc_skip_char(char **buf, size_t *size, const char v)
2025 {
2026         while (*size) {
2027                 if (**buf != v)
2028                         break;
2029                 (*size)--;
2030                 (*buf)++;
2031         }
2032 }
2033
2034 #define TMPBUFLEN 22
2035 /**
2036  * proc_get_long - reads an ASCII formatted integer from a user buffer
2037  *
2038  * @buf: a kernel buffer
2039  * @size: size of the kernel buffer
2040  * @val: this is where the number will be stored
2041  * @neg: set to %TRUE if number is negative
2042  * @perm_tr: a vector which contains the allowed trailers
2043  * @perm_tr_len: size of the perm_tr vector
2044  * @tr: pointer to store the trailer character
2045  *
2046  * In case of success %0 is returned and @buf and @size are updated with
2047  * the amount of bytes read. If @tr is non-NULL and a trailing
2048  * character exists (size is non-zero after returning from this
2049  * function), @tr is updated with the trailing character.
2050  */
2051 static int proc_get_long(char **buf, size_t *size,
2052                           unsigned long *val, bool *neg,
2053                           const char *perm_tr, unsigned perm_tr_len, char *tr)
2054 {
2055         int len;
2056         char *p, tmp[TMPBUFLEN];
2057
2058         if (!*size)
2059                 return -EINVAL;
2060
2061         len = *size;
2062         if (len > TMPBUFLEN - 1)
2063                 len = TMPBUFLEN - 1;
2064
2065         memcpy(tmp, *buf, len);
2066
2067         tmp[len] = 0;
2068         p = tmp;
2069         if (*p == '-' && *size > 1) {
2070                 *neg = true;
2071                 p++;
2072         } else
2073                 *neg = false;
2074         if (!isdigit(*p))
2075                 return -EINVAL;
2076
2077         *val = simple_strtoul(p, &p, 0);
2078
2079         len = p - tmp;
2080
2081         /* We don't know if the next char is whitespace thus we may accept
2082          * invalid integers (e.g. 1234...a) or two integers instead of one
2083          * (e.g. 123...1). So lets not allow such large numbers. */
2084         if (len == TMPBUFLEN - 1)
2085                 return -EINVAL;
2086
2087         if (len < *size && perm_tr_len && !memchr(perm_tr, *p, perm_tr_len))
2088                 return -EINVAL;
2089
2090         if (tr && (len < *size))
2091                 *tr = *p;
2092
2093         *buf += len;
2094         *size -= len;
2095
2096         return 0;
2097 }
2098
2099 /**
2100  * proc_put_long - converts an integer to a decimal ASCII formatted string
2101  *
2102  * @buf: the user buffer
2103  * @size: the size of the user buffer
2104  * @val: the integer to be converted
2105  * @neg: sign of the number, %TRUE for negative
2106  *
2107  * In case of success %0 is returned and @buf and @size are updated with
2108  * the amount of bytes written.
2109  */
2110 static int proc_put_long(void __user **buf, size_t *size, unsigned long val,
2111                           bool neg)
2112 {
2113         int len;
2114         char tmp[TMPBUFLEN], *p = tmp;
2115
2116         sprintf(p, "%s%lu", neg ? "-" : "", val);
2117         len = strlen(tmp);
2118         if (len > *size)
2119                 len = *size;
2120         if (copy_to_user(*buf, tmp, len))
2121                 return -EFAULT;
2122         *size -= len;
2123         *buf += len;
2124         return 0;
2125 }
2126 #undef TMPBUFLEN
2127
2128 static int proc_put_char(void __user **buf, size_t *size, char c)
2129 {
2130         if (*size) {
2131                 char __user **buffer = (char __user **)buf;
2132                 if (put_user(c, *buffer))
2133                         return -EFAULT;
2134                 (*size)--, (*buffer)++;
2135                 *buf = *buffer;
2136         }
2137         return 0;
2138 }
2139
2140 static int do_proc_dointvec_conv(bool *negp, unsigned long *lvalp,
2141                                  int *valp,
2142                                  int write, void *data)
2143 {
2144         if (write) {
2145                 if (*negp) {
2146                         if (*lvalp > (unsigned long) INT_MAX + 1)
2147                                 return -EINVAL;
2148                         *valp = -*lvalp;
2149                 } else {
2150                         if (*lvalp > (unsigned long) INT_MAX)
2151                                 return -EINVAL;
2152                         *valp = *lvalp;
2153                 }
2154         } else {
2155                 int val = *valp;
2156                 if (val < 0) {
2157                         *negp = true;
2158                         *lvalp = -(unsigned long)val;
2159                 } else {
2160                         *negp = false;
2161                         *lvalp = (unsigned long)val;
2162                 }
2163         }
2164         return 0;
2165 }
2166
2167 static int do_proc_douintvec_conv(bool *negp, unsigned long *lvalp,
2168                                  int *valp,
2169                                  int write, void *data)
2170 {
2171         if (write) {
2172                 if (*negp)
2173                         return -EINVAL;
2174                 if (*lvalp > UINT_MAX)
2175                         return -EINVAL;
2176                 *valp = *lvalp;
2177         } else {
2178                 unsigned int val = *valp;
2179                 *negp = false;
2180                 *lvalp = (unsigned long)val;
2181         }
2182         return 0;
2183 }
2184
2185 static const char proc_wspace_sep[] = { ' ', '\t', '\n' };
2186
2187 static int __do_proc_dointvec(void *tbl_data, struct ctl_table *table,
2188                   int write, void __user *buffer,
2189                   size_t *lenp, loff_t *ppos,
2190                   int (*conv)(bool *negp, unsigned long *lvalp, int *valp,
2191                               int write, void *data),
2192                   void *data)
2193 {
2194         int *i, vleft, first = 1, err = 0;
2195         size_t left;
2196         char *kbuf = NULL, *p;
2197
2198         if (!tbl_data || !table->maxlen || !*lenp || (*ppos && !write)) {
2199                 *lenp = 0;
2200                 return 0;
2201         }
2202
2203         i = (int *) tbl_data;
2204         vleft = table->maxlen / sizeof(*i);
2205         left = *lenp;
2206
2207         if (!conv)
2208                 conv = do_proc_dointvec_conv;
2209
2210         if (write) {
2211                 if (*ppos) {
2212                         switch (sysctl_writes_strict) {
2213                         case SYSCTL_WRITES_STRICT:
2214                                 goto out;
2215                         case SYSCTL_WRITES_WARN:
2216                                 warn_sysctl_write(table);
2217                                 break;
2218                         default:
2219                                 break;
2220                         }
2221                 }
2222
2223                 if (left > PAGE_SIZE - 1)
2224                         left = PAGE_SIZE - 1;
2225                 p = kbuf = memdup_user_nul(buffer, left);
2226                 if (IS_ERR(kbuf))
2227                         return PTR_ERR(kbuf);
2228         }
2229
2230         for (; left && vleft--; i++, first=0) {
2231                 unsigned long lval;
2232                 bool neg;
2233
2234                 if (write) {
2235                         left -= proc_skip_spaces(&p);
2236
2237                         if (!left)
2238                                 break;
2239                         err = proc_get_long(&p, &left, &lval, &neg,
2240                                              proc_wspace_sep,
2241                                              sizeof(proc_wspace_sep), NULL);
2242                         if (err)
2243                                 break;
2244                         if (conv(&neg, &lval, i, 1, data)) {
2245                                 err = -EINVAL;
2246                                 break;
2247                         }
2248                 } else {
2249                         if (conv(&neg, &lval, i, 0, data)) {
2250                                 err = -EINVAL;
2251                                 break;
2252                         }
2253                         if (!first)
2254                                 err = proc_put_char(&buffer, &left, '\t');
2255                         if (err)
2256                                 break;
2257                         err = proc_put_long(&buffer, &left, lval, neg);
2258                         if (err)
2259                                 break;
2260                 }
2261         }
2262
2263         if (!write && !first && left && !err)
2264                 err = proc_put_char(&buffer, &left, '\n');
2265         if (write && !err && left)
2266                 left -= proc_skip_spaces(&p);
2267         if (write) {
2268                 kfree(kbuf);
2269                 if (first)
2270                         return err ? : -EINVAL;
2271         }
2272         *lenp -= left;
2273 out:
2274         *ppos += *lenp;
2275         return err;
2276 }
2277
2278 static int do_proc_dointvec(struct ctl_table *table, int write,
2279                   void __user *buffer, size_t *lenp, loff_t *ppos,
2280                   int (*conv)(bool *negp, unsigned long *lvalp, int *valp,
2281                               int write, void *data),
2282                   void *data)
2283 {
2284         return __do_proc_dointvec(table->data, table, write,
2285                         buffer, lenp, ppos, conv, data);
2286 }
2287
2288 /**
2289  * proc_dointvec - read a vector of integers
2290  * @table: the sysctl table
2291  * @write: %TRUE if this is a write to the sysctl file
2292  * @buffer: the user buffer
2293  * @lenp: the size of the user buffer
2294  * @ppos: file position
2295  *
2296  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2297  * values from/to the user buffer, treated as an ASCII string.
2298  *
2299  * Returns 0 on success.
2300  */
2301 int proc_dointvec(struct ctl_table *table, int write,
2302                      void __user *buffer, size_t *lenp, loff_t *ppos)
2303 {
2304         return do_proc_dointvec(table, write, buffer, lenp, ppos, NULL, NULL);
2305 }
2306
2307 /**
2308  * proc_douintvec - read a vector of unsigned integers
2309  * @table: the sysctl table
2310  * @write: %TRUE if this is a write to the sysctl file
2311  * @buffer: the user buffer
2312  * @lenp: the size of the user buffer
2313  * @ppos: file position
2314  *
2315  * Reads/writes up to table->maxlen/sizeof(unsigned int) unsigned integer
2316  * values from/to the user buffer, treated as an ASCII string.
2317  *
2318  * Returns 0 on success.
2319  */
2320 int proc_douintvec(struct ctl_table *table, int write,
2321                      void __user *buffer, size_t *lenp, loff_t *ppos)
2322 {
2323         return do_proc_dointvec(table, write, buffer, lenp, ppos,
2324                                 do_proc_douintvec_conv, NULL);
2325 }
2326
2327 /*
2328  * Taint values can only be increased
2329  * This means we can safely use a temporary.
2330  */
2331 static int proc_taint(struct ctl_table *table, int write,
2332                                void __user *buffer, size_t *lenp, loff_t *ppos)
2333 {
2334         struct ctl_table t;
2335         unsigned long tmptaint = get_taint();
2336         int err;
2337
2338         if (write && !capable(CAP_SYS_ADMIN))
2339                 return -EPERM;
2340
2341         t = *table;
2342         t.data = &tmptaint;
2343         err = proc_doulongvec_minmax(&t, write, buffer, lenp, ppos);
2344         if (err < 0)
2345                 return err;
2346
2347         if (write) {
2348                 /*
2349                  * Poor man's atomic or. Not worth adding a primitive
2350                  * to everyone's atomic.h for this
2351                  */
2352                 int i;
2353                 for (i = 0; i < BITS_PER_LONG && tmptaint >> i; i++) {
2354                         if ((tmptaint >> i) & 1)
2355                                 add_taint(i, LOCKDEP_STILL_OK);
2356                 }
2357         }
2358
2359         return err;
2360 }
2361
2362 #ifdef CONFIG_PRINTK
2363 static int proc_dointvec_minmax_sysadmin(struct ctl_table *table, int write,
2364                                 void __user *buffer, size_t *lenp, loff_t *ppos)
2365 {
2366         if (write && !capable(CAP_SYS_ADMIN))
2367                 return -EPERM;
2368
2369         return proc_dointvec_minmax(table, write, buffer, lenp, ppos);
2370 }
2371 #endif
2372
2373 struct do_proc_dointvec_minmax_conv_param {
2374         int *min;
2375         int *max;
2376 };
2377
2378 static int do_proc_dointvec_minmax_conv(bool *negp, unsigned long *lvalp,
2379                                         int *valp,
2380                                         int write, void *data)
2381 {
2382         struct do_proc_dointvec_minmax_conv_param *param = data;
2383         if (write) {
2384                 int val;
2385                 if (*negp) {
2386                         if (*lvalp > (unsigned long) INT_MAX + 1)
2387                                 return -EINVAL;
2388                         val = -*lvalp;
2389                 } else {
2390                         if (*lvalp > (unsigned long) INT_MAX)
2391                                 return -EINVAL;
2392                         val = *lvalp;
2393                 }
2394                 if ((param->min && *param->min > val) ||
2395                     (param->max && *param->max < val))
2396                         return -EINVAL;
2397                 *valp = val;
2398         } else {
2399                 int val = *valp;
2400                 if (val < 0) {
2401                         *negp = true;
2402                         *lvalp = -(unsigned long)val;
2403                 } else {
2404                         *negp = false;
2405                         *lvalp = (unsigned long)val;
2406                 }
2407         }
2408         return 0;
2409 }
2410
2411 /**
2412  * proc_dointvec_minmax - read a vector of integers with min/max values
2413  * @table: the sysctl table
2414  * @write: %TRUE if this is a write to the sysctl file
2415  * @buffer: the user buffer
2416  * @lenp: the size of the user buffer
2417  * @ppos: file position
2418  *
2419  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2420  * values from/to the user buffer, treated as an ASCII string.
2421  *
2422  * This routine will ensure the values are within the range specified by
2423  * table->extra1 (min) and table->extra2 (max).
2424  *
2425  * Returns 0 on success.
2426  */
2427 int proc_dointvec_minmax(struct ctl_table *table, int write,
2428                   void __user *buffer, size_t *lenp, loff_t *ppos)
2429 {
2430         struct do_proc_dointvec_minmax_conv_param param = {
2431                 .min = (int *) table->extra1,
2432                 .max = (int *) table->extra2,
2433         };
2434         return do_proc_dointvec(table, write, buffer, lenp, ppos,
2435                                 do_proc_dointvec_minmax_conv, &param);
2436 }
2437
2438 static void validate_coredump_safety(void)
2439 {
2440 #ifdef CONFIG_COREDUMP
2441         if (suid_dumpable == SUID_DUMP_ROOT &&
2442             core_pattern[0] != '/' && core_pattern[0] != '|') {
2443                 printk(KERN_WARNING "Unsafe core_pattern used with "\
2444                         "suid_dumpable=2. Pipe handler or fully qualified "\
2445                         "core dump path required.\n");
2446         }
2447 #endif
2448 }
2449
2450 static int proc_dointvec_minmax_coredump(struct ctl_table *table, int write,
2451                 void __user *buffer, size_t *lenp, loff_t *ppos)
2452 {
2453         int error = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
2454         if (!error)
2455                 validate_coredump_safety();
2456         return error;
2457 }
2458
2459 #ifdef CONFIG_COREDUMP
2460 static int proc_dostring_coredump(struct ctl_table *table, int write,
2461                   void __user *buffer, size_t *lenp, loff_t *ppos)
2462 {
2463         int error = proc_dostring(table, write, buffer, lenp, ppos);
2464         if (!error)
2465                 validate_coredump_safety();
2466         return error;
2467 }
2468 #endif
2469
2470 static int __do_proc_doulongvec_minmax(void *data, 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         unsigned long *i, *min, *max;
2477         int vleft, first = 1, err = 0;
2478         size_t left;
2479         char *kbuf = NULL, *p;
2480
2481         if (!data || !table->maxlen || !*lenp || (*ppos && !write)) {
2482                 *lenp = 0;
2483                 return 0;
2484         }
2485
2486         i = (unsigned long *) data;
2487         min = (unsigned long *) table->extra1;
2488         max = (unsigned long *) table->extra2;
2489         vleft = table->maxlen / sizeof(unsigned long);
2490         left = *lenp;
2491
2492         if (write) {
2493                 if (*ppos) {
2494                         switch (sysctl_writes_strict) {
2495                         case SYSCTL_WRITES_STRICT:
2496                                 goto out;
2497                         case SYSCTL_WRITES_WARN:
2498                                 warn_sysctl_write(table);
2499                                 break;
2500                         default:
2501                                 break;
2502                         }
2503                 }
2504
2505                 if (left > PAGE_SIZE - 1)
2506                         left = PAGE_SIZE - 1;
2507                 p = kbuf = memdup_user_nul(buffer, left);
2508                 if (IS_ERR(kbuf))
2509                         return PTR_ERR(kbuf);
2510         }
2511
2512         for (; left && vleft--; i++, first = 0) {
2513                 unsigned long val;
2514
2515                 if (write) {
2516                         bool neg;
2517
2518                         left -= proc_skip_spaces(&p);
2519                         if (!left)
2520                                 break;
2521
2522                         err = proc_get_long(&p, &left, &val, &neg,
2523                                              proc_wspace_sep,
2524                                              sizeof(proc_wspace_sep), NULL);
2525                         if (err)
2526                                 break;
2527                         if (neg)
2528                                 continue;
2529                         val = convmul * val / convdiv;
2530                         if ((min && val < *min) || (max && val > *max)) {
2531                                 err = -EINVAL;
2532                                 break;
2533                         }
2534                         *i = val;
2535                 } else {
2536                         val = convdiv * (*i) / convmul;
2537                         if (!first) {
2538                                 err = proc_put_char(&buffer, &left, '\t');
2539                                 if (err)
2540                                         break;
2541                         }
2542                         err = proc_put_long(&buffer, &left, val, false);
2543                         if (err)
2544                                 break;
2545                 }
2546         }
2547
2548         if (!write && !first && left && !err)
2549                 err = proc_put_char(&buffer, &left, '\n');
2550         if (write && !err)
2551                 left -= proc_skip_spaces(&p);
2552         if (write) {
2553                 kfree(kbuf);
2554                 if (first)
2555                         return err ? : -EINVAL;
2556         }
2557         *lenp -= left;
2558 out:
2559         *ppos += *lenp;
2560         return err;
2561 }
2562
2563 static int do_proc_doulongvec_minmax(struct ctl_table *table, int write,
2564                                      void __user *buffer,
2565                                      size_t *lenp, loff_t *ppos,
2566                                      unsigned long convmul,
2567                                      unsigned long convdiv)
2568 {
2569         return __do_proc_doulongvec_minmax(table->data, table, write,
2570                         buffer, lenp, ppos, convmul, convdiv);
2571 }
2572
2573 /**
2574  * proc_doulongvec_minmax - read a vector of long integers with min/max values
2575  * @table: the sysctl table
2576  * @write: %TRUE if this is a write to the sysctl file
2577  * @buffer: the user buffer
2578  * @lenp: the size of the user buffer
2579  * @ppos: file position
2580  *
2581  * Reads/writes up to table->maxlen/sizeof(unsigned long) unsigned long
2582  * values from/to the user buffer, treated as an ASCII string.
2583  *
2584  * This routine will ensure the values are within the range specified by
2585  * table->extra1 (min) and table->extra2 (max).
2586  *
2587  * Returns 0 on success.
2588  */
2589 int proc_doulongvec_minmax(struct ctl_table *table, int write,
2590                            void __user *buffer, size_t *lenp, loff_t *ppos)
2591 {
2592     return do_proc_doulongvec_minmax(table, write, buffer, lenp, ppos, 1l, 1l);
2593 }
2594
2595 /**
2596  * proc_doulongvec_ms_jiffies_minmax - read a vector of millisecond values with min/max values
2597  * @table: the sysctl table
2598  * @write: %TRUE if this is a write to the sysctl file
2599  * @buffer: the user buffer
2600  * @lenp: the size of the user buffer
2601  * @ppos: file position
2602  *
2603  * Reads/writes up to table->maxlen/sizeof(unsigned long) unsigned long
2604  * values from/to the user buffer, treated as an ASCII string. The values
2605  * are treated as milliseconds, and converted to jiffies when they are stored.
2606  *
2607  * This routine will ensure the values are within the range specified by
2608  * table->extra1 (min) and table->extra2 (max).
2609  *
2610  * Returns 0 on success.
2611  */
2612 int proc_doulongvec_ms_jiffies_minmax(struct ctl_table *table, int write,
2613                                       void __user *buffer,
2614                                       size_t *lenp, loff_t *ppos)
2615 {
2616     return do_proc_doulongvec_minmax(table, write, buffer,
2617                                      lenp, ppos, HZ, 1000l);
2618 }
2619
2620
2621 static int do_proc_dointvec_jiffies_conv(bool *negp, unsigned long *lvalp,
2622                                          int *valp,
2623                                          int write, void *data)
2624 {
2625         if (write) {
2626                 if (*lvalp > LONG_MAX / HZ)
2627                         return 1;
2628                 *valp = *negp ? -(*lvalp*HZ) : (*lvalp*HZ);
2629         } else {
2630                 int val = *valp;
2631                 unsigned long lval;
2632                 if (val < 0) {
2633                         *negp = true;
2634                         lval = -(unsigned long)val;
2635                 } else {
2636                         *negp = false;
2637                         lval = (unsigned long)val;
2638                 }
2639                 *lvalp = lval / HZ;
2640         }
2641         return 0;
2642 }
2643
2644 static int do_proc_dointvec_userhz_jiffies_conv(bool *negp, unsigned long *lvalp,
2645                                                 int *valp,
2646                                                 int write, void *data)
2647 {
2648         if (write) {
2649                 if (USER_HZ < HZ && *lvalp > (LONG_MAX / HZ) * USER_HZ)
2650                         return 1;
2651                 *valp = clock_t_to_jiffies(*negp ? -*lvalp : *lvalp);
2652         } else {
2653                 int val = *valp;
2654                 unsigned long lval;
2655                 if (val < 0) {
2656                         *negp = true;
2657                         lval = -(unsigned long)val;
2658                 } else {
2659                         *negp = false;
2660                         lval = (unsigned long)val;
2661                 }
2662                 *lvalp = jiffies_to_clock_t(lval);
2663         }
2664         return 0;
2665 }
2666
2667 static int do_proc_dointvec_ms_jiffies_conv(bool *negp, unsigned long *lvalp,
2668                                             int *valp,
2669                                             int write, void *data)
2670 {
2671         if (write) {
2672                 unsigned long jif = msecs_to_jiffies(*negp ? -*lvalp : *lvalp);
2673
2674                 if (jif > INT_MAX)
2675                         return 1;
2676                 *valp = (int)jif;
2677         } else {
2678                 int val = *valp;
2679                 unsigned long lval;
2680                 if (val < 0) {
2681                         *negp = true;
2682                         lval = -(unsigned long)val;
2683                 } else {
2684                         *negp = false;
2685                         lval = (unsigned long)val;
2686                 }
2687                 *lvalp = jiffies_to_msecs(lval);
2688         }
2689         return 0;
2690 }
2691
2692 /**
2693  * proc_dointvec_jiffies - read a vector of integers as seconds
2694  * @table: the sysctl table
2695  * @write: %TRUE if this is a write to the sysctl file
2696  * @buffer: the user buffer
2697  * @lenp: the size of the user buffer
2698  * @ppos: file position
2699  *
2700  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2701  * values from/to the user buffer, treated as an ASCII string.
2702  * The values read are assumed to be in seconds, and are converted into
2703  * jiffies.
2704  *
2705  * Returns 0 on success.
2706  */
2707 int proc_dointvec_jiffies(struct ctl_table *table, int write,
2708                           void __user *buffer, size_t *lenp, loff_t *ppos)
2709 {
2710     return do_proc_dointvec(table,write,buffer,lenp,ppos,
2711                             do_proc_dointvec_jiffies_conv,NULL);
2712 }
2713
2714 /**
2715  * proc_dointvec_userhz_jiffies - read a vector of integers as 1/USER_HZ seconds
2716  * @table: the sysctl table
2717  * @write: %TRUE if this is a write to the sysctl file
2718  * @buffer: the user buffer
2719  * @lenp: the size of the user buffer
2720  * @ppos: pointer to the file position
2721  *
2722  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2723  * values from/to the user buffer, treated as an ASCII string.
2724  * The values read are assumed to be in 1/USER_HZ seconds, and
2725  * are converted into jiffies.
2726  *
2727  * Returns 0 on success.
2728  */
2729 int proc_dointvec_userhz_jiffies(struct ctl_table *table, int write,
2730                                  void __user *buffer, size_t *lenp, loff_t *ppos)
2731 {
2732     return do_proc_dointvec(table,write,buffer,lenp,ppos,
2733                             do_proc_dointvec_userhz_jiffies_conv,NULL);
2734 }
2735
2736 /**
2737  * proc_dointvec_ms_jiffies - read a vector of integers as 1 milliseconds
2738  * @table: the sysctl table
2739  * @write: %TRUE if this is a write to the sysctl file
2740  * @buffer: the user buffer
2741  * @lenp: the size of the user buffer
2742  * @ppos: file position
2743  * @ppos: the current position in the file
2744  *
2745  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2746  * values from/to the user buffer, treated as an ASCII string.
2747  * The values read are assumed to be in 1/1000 seconds, and
2748  * are converted into jiffies.
2749  *
2750  * Returns 0 on success.
2751  */
2752 int proc_dointvec_ms_jiffies(struct ctl_table *table, int write,
2753                              void __user *buffer, size_t *lenp, loff_t *ppos)
2754 {
2755         return do_proc_dointvec(table, write, buffer, lenp, ppos,
2756                                 do_proc_dointvec_ms_jiffies_conv, NULL);
2757 }
2758
2759 static int proc_do_cad_pid(struct ctl_table *table, int write,
2760                            void __user *buffer, size_t *lenp, loff_t *ppos)
2761 {
2762         struct pid *new_pid;
2763         pid_t tmp;
2764         int r;
2765
2766         tmp = pid_vnr(cad_pid);
2767
2768         r = __do_proc_dointvec(&tmp, table, write, buffer,
2769                                lenp, ppos, NULL, NULL);
2770         if (r || !write)
2771                 return r;
2772
2773         new_pid = find_get_pid(tmp);
2774         if (!new_pid)
2775                 return -ESRCH;
2776
2777         put_pid(xchg(&cad_pid, new_pid));
2778         return 0;
2779 }
2780
2781 /**
2782  * proc_do_large_bitmap - read/write from/to a large bitmap
2783  * @table: the sysctl table
2784  * @write: %TRUE if this is a write to the sysctl file
2785  * @buffer: the user buffer
2786  * @lenp: the size of the user buffer
2787  * @ppos: file position
2788  *
2789  * The bitmap is stored at table->data and the bitmap length (in bits)
2790  * in table->maxlen.
2791  *
2792  * We use a range comma separated format (e.g. 1,3-4,10-10) so that
2793  * large bitmaps may be represented in a compact manner. Writing into
2794  * the file will clear the bitmap then update it with the given input.
2795  *
2796  * Returns 0 on success.
2797  */
2798 int proc_do_large_bitmap(struct ctl_table *table, int write,
2799                          void __user *buffer, size_t *lenp, loff_t *ppos)
2800 {
2801         int err = 0;
2802         bool first = 1;
2803         size_t left = *lenp;
2804         unsigned long bitmap_len = table->maxlen;
2805         unsigned long *bitmap = *(unsigned long **) table->data;
2806         unsigned long *tmp_bitmap = NULL;
2807         char tr_a[] = { '-', ',', '\n' }, tr_b[] = { ',', '\n', 0 }, c;
2808
2809         if (!bitmap || !bitmap_len || !left || (*ppos && !write)) {
2810                 *lenp = 0;
2811                 return 0;
2812         }
2813
2814         if (write) {
2815                 char *kbuf, *p;
2816
2817                 if (left > PAGE_SIZE - 1)
2818                         left = PAGE_SIZE - 1;
2819
2820                 p = kbuf = memdup_user_nul(buffer, left);
2821                 if (IS_ERR(kbuf))
2822                         return PTR_ERR(kbuf);
2823
2824                 tmp_bitmap = kzalloc(BITS_TO_LONGS(bitmap_len) * sizeof(unsigned long),
2825                                      GFP_KERNEL);
2826                 if (!tmp_bitmap) {
2827                         kfree(kbuf);
2828                         return -ENOMEM;
2829                 }
2830                 proc_skip_char(&p, &left, '\n');
2831                 while (!err && left) {
2832                         unsigned long val_a, val_b;
2833                         bool neg;
2834
2835                         err = proc_get_long(&p, &left, &val_a, &neg, tr_a,
2836                                              sizeof(tr_a), &c);
2837                         if (err)
2838                                 break;
2839                         if (val_a >= bitmap_len || neg) {
2840                                 err = -EINVAL;
2841                                 break;
2842                         }
2843
2844                         val_b = val_a;
2845                         if (left) {
2846                                 p++;
2847                                 left--;
2848                         }
2849
2850                         if (c == '-') {
2851                                 err = proc_get_long(&p, &left, &val_b,
2852                                                      &neg, tr_b, sizeof(tr_b),
2853                                                      &c);
2854                                 if (err)
2855                                         break;
2856                                 if (val_b >= bitmap_len || neg ||
2857                                     val_a > val_b) {
2858                                         err = -EINVAL;
2859                                         break;
2860                                 }
2861                                 if (left) {
2862                                         p++;
2863                                         left--;
2864                                 }
2865                         }
2866
2867                         bitmap_set(tmp_bitmap, val_a, val_b - val_a + 1);
2868                         first = 0;
2869                         proc_skip_char(&p, &left, '\n');
2870                 }
2871                 kfree(kbuf);
2872         } else {
2873                 unsigned long bit_a, bit_b = 0;
2874
2875                 while (left) {
2876                         bit_a = find_next_bit(bitmap, bitmap_len, bit_b);
2877                         if (bit_a >= bitmap_len)
2878                                 break;
2879                         bit_b = find_next_zero_bit(bitmap, bitmap_len,
2880                                                    bit_a + 1) - 1;
2881
2882                         if (!first) {
2883                                 err = proc_put_char(&buffer, &left, ',');
2884                                 if (err)
2885                                         break;
2886                         }
2887                         err = proc_put_long(&buffer, &left, bit_a, false);
2888                         if (err)
2889                                 break;
2890                         if (bit_a != bit_b) {
2891                                 err = proc_put_char(&buffer, &left, '-');
2892                                 if (err)
2893                                         break;
2894                                 err = proc_put_long(&buffer, &left, bit_b, false);
2895                                 if (err)
2896                                         break;
2897                         }
2898
2899                         first = 0; bit_b++;
2900                 }
2901                 if (!err)
2902                         err = proc_put_char(&buffer, &left, '\n');
2903         }
2904
2905         if (!err) {
2906                 if (write) {
2907                         if (*ppos)
2908                                 bitmap_or(bitmap, bitmap, tmp_bitmap, bitmap_len);
2909                         else
2910                                 bitmap_copy(bitmap, tmp_bitmap, bitmap_len);
2911                 }
2912                 kfree(tmp_bitmap);
2913                 *lenp -= left;
2914                 *ppos += *lenp;
2915                 return 0;
2916         } else {
2917                 kfree(tmp_bitmap);
2918                 return err;
2919         }
2920 }
2921
2922 #else /* CONFIG_PROC_SYSCTL */
2923
2924 int proc_dostring(struct ctl_table *table, int write,
2925                   void __user *buffer, size_t *lenp, loff_t *ppos)
2926 {
2927         return -ENOSYS;
2928 }
2929
2930 int proc_dointvec(struct ctl_table *table, int write,
2931                   void __user *buffer, size_t *lenp, loff_t *ppos)
2932 {
2933         return -ENOSYS;
2934 }
2935
2936 int proc_douintvec(struct ctl_table *table, int write,
2937                   void __user *buffer, size_t *lenp, loff_t *ppos)
2938 {
2939         return -ENOSYS;
2940 }
2941
2942 int proc_dointvec_minmax(struct ctl_table *table, int write,
2943                     void __user *buffer, size_t *lenp, loff_t *ppos)
2944 {
2945         return -ENOSYS;
2946 }
2947
2948 int proc_dointvec_jiffies(struct ctl_table *table, int write,
2949                     void __user *buffer, size_t *lenp, loff_t *ppos)
2950 {
2951         return -ENOSYS;
2952 }
2953
2954 int proc_dointvec_userhz_jiffies(struct ctl_table *table, int write,
2955                     void __user *buffer, size_t *lenp, loff_t *ppos)
2956 {
2957         return -ENOSYS;
2958 }
2959
2960 int proc_dointvec_ms_jiffies(struct ctl_table *table, int write,
2961                              void __user *buffer, size_t *lenp, loff_t *ppos)
2962 {
2963         return -ENOSYS;
2964 }
2965
2966 int proc_doulongvec_minmax(struct ctl_table *table, int write,
2967                     void __user *buffer, size_t *lenp, loff_t *ppos)
2968 {
2969         return -ENOSYS;
2970 }
2971
2972 int proc_doulongvec_ms_jiffies_minmax(struct ctl_table *table, int write,
2973                                       void __user *buffer,
2974                                       size_t *lenp, loff_t *ppos)
2975 {
2976     return -ENOSYS;
2977 }
2978
2979
2980 #endif /* CONFIG_PROC_SYSCTL */
2981
2982 /*
2983  * No sense putting this after each symbol definition, twice,
2984  * exception granted :-)
2985  */
2986 EXPORT_SYMBOL(proc_dointvec);
2987 EXPORT_SYMBOL(proc_douintvec);
2988 EXPORT_SYMBOL(proc_dointvec_jiffies);
2989 EXPORT_SYMBOL(proc_dointvec_minmax);
2990 EXPORT_SYMBOL(proc_dointvec_userhz_jiffies);
2991 EXPORT_SYMBOL(proc_dointvec_ms_jiffies);
2992 EXPORT_SYMBOL(proc_dostring);
2993 EXPORT_SYMBOL(proc_doulongvec_minmax);
2994 EXPORT_SYMBOL(proc_doulongvec_ms_jiffies_minmax);