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 compat_log;
 111 extern int latencytop_enabled;
 112 extern int sysctl_nr_open_min, sysctl_nr_open_max;
 113 #ifndef CONFIG_MMU
 114 extern int sysctl_nr_trim_pages;
 115 #endif
 116
 117 /* Constants used for minimum and  maximum */
 118 #ifdef CONFIG_LOCKUP_DETECTOR
 119 static int sixty = 60;
 120 #endif
 121
 122 static int __maybe_unused neg_one = -1;
 123
 124 static int zero;
 125 static int __maybe_unused one = 1;
 126 static int __maybe_unused two = 2;
 127 static int __maybe_unused four = 4;
 128 static unsigned long zero_ul;
 129 static unsigned long one_ul = 1;
 130 static unsigned long long_max = LONG_MAX;
 131 static int one_hundred = 100;
 132 #ifdef CONFIG_PRINTK
 133 static int ten_thousand = 10000;
 134 #endif
 135
 136 /* this is needed for the proc_doulongvec_minmax of vm_dirty_bytes */
 137 static unsigned long dirty_bytes_min = 2 * PAGE_SIZE;
 138
 139 /* this is needed for the proc_dointvec_minmax for [fs_]overflow UID and GID */
 140 static int maxolduid = 65535;
 141 static int minolduid;
 142
 143 static int ngroups_max = NGROUPS_MAX;
 144 static const int cap_last_cap = CAP_LAST_CAP;
 145
 146 /*this is needed for proc_doulongvec_minmax of sysctl_hung_task_timeout_secs */
 147 #ifdef CONFIG_DETECT_HUNG_TASK
 148 static unsigned long hung_task_timeout_max = (LONG_MAX/HZ);
 149 #endif
 150
 151 #ifdef CONFIG_INOTIFY_USER
 152 #include <linux/inotify.h>
 153 #endif
 154 #ifdef CONFIG_SPARC
 155 #endif
 156
 157 #ifdef __hppa__
 158 extern int pwrsw_enabled;
 159 #endif
 160
 161 #ifdef CONFIG_SYSCTL_ARCH_UNALIGN_ALLOW
 162 extern int unaligned_enabled;
 163 #endif
 164
 165 #ifdef CONFIG_IA64
 166 extern int unaligned_dump_stack;
 167 #endif
 168
 169 #ifdef CONFIG_SYSCTL_ARCH_UNALIGN_NO_WARN
 170 extern int no_unaligned_warning;
 171 #endif
 172
 173 #ifdef CONFIG_PROC_SYSCTL
 174
 175 #define SYSCTL_WRITES_LEGACY    -1
 176 #define SYSCTL_WRITES_WARN       0
 177 #define SYSCTL_WRITES_STRICT     1
 178
 179 static int sysctl_writes_strict = SYSCTL_WRITES_STRICT;
 180
 181 static int proc_do_cad_pid(struct ctl_table *table, int write,
 182                   void __user *buffer, size_t *lenp, loff_t *ppos);
 183 static int proc_taint(struct ctl_table *table, int write,
 184                                void __user *buffer, size_t *lenp, loff_t *ppos);
 185 #endif
 186
 187 #ifdef CONFIG_PRINTK
 188 static int proc_dointvec_minmax_sysadmin(struct ctl_table *table, int write,
 189                                 void __user *buffer, size_t *lenp, loff_t *ppos);
 190 #endif
 191
 192 static int proc_dointvec_minmax_coredump(struct ctl_table *table, int write,
 193                 void __user *buffer, size_t *lenp, loff_t *ppos);
 194 #ifdef CONFIG_COREDUMP
 195 static int proc_dostring_coredump(struct ctl_table *table, int write,
 196                 void __user *buffer, size_t *lenp, loff_t *ppos);
 197 #endif
 198
 199 #ifdef CONFIG_MAGIC_SYSRQ
 200 /* Note: sysrq code uses it's own private copy */
 201 static int __sysrq_enabled = CONFIG_MAGIC_SYSRQ_DEFAULT_ENABLE;
 202
 203 static int sysrq_sysctl_handler(struct ctl_table *table, int write,
 204                                 void __user *buffer, size_t *lenp,
 205                                 loff_t *ppos)
 206 {
 207         int error;
 208
 209         error = proc_dointvec(table, write, buffer, lenp, ppos);
 210         if (error)
 211                 return error;
 212
 213         if (write)
 214                 sysrq_toggle_support(__sysrq_enabled);
 215
 216         return 0;
 217 }
 218
 219 #endif
 220
 221 static struct ctl_table kern_table[];
 222 static struct ctl_table vm_table[];
 223 static struct ctl_table fs_table[];
 224 static struct ctl_table debug_table[];
 225 static struct ctl_table dev_table[];
 226 extern struct ctl_table random_table[];
 227 #ifdef CONFIG_EPOLL
 228 extern struct ctl_table epoll_table[];
 229 #endif
 230
 231 #ifdef HAVE_ARCH_PICK_MMAP_LAYOUT
 232 int sysctl_legacy_va_layout;
 233 #endif
 234
 235 /* The default sysctl tables: */
 236
 237 static struct ctl_table sysctl_base_table[] = {
 238         {
 239                 .procname       = "kernel",
 240                 .mode           = 0555,
 241                 .child          = kern_table,
 242         },
 243         {
 244                 .procname       = "vm",
 245                 .mode           = 0555,
 246                 .child          = vm_table,
 247         },
 248         {
 249                 .procname       = "fs",
 250                 .mode           = 0555,
 251                 .child          = fs_table,
 252         },
 253         {
 254                 .procname       = "debug",
 255                 .mode           = 0555,
 256                 .child          = debug_table,
 257         },
 258         {
 259                 .procname       = "dev",
 260                 .mode           = 0555,
 261                 .child          = dev_table,
 262         },
 263         { }
 264 };
 265
 266 #ifdef CONFIG_SCHED_DEBUG
 267 static int min_sched_granularity_ns = 100000;           /* 100 usecs */
 268 static int max_sched_granularity_ns = NSEC_PER_SEC;     /* 1 second */
 269 static int min_wakeup_granularity_ns;                   /* 0 usecs */
 270 static int max_wakeup_granularity_ns = NSEC_PER_SEC;    /* 1 second */
 271 #ifdef CONFIG_SMP
 272 static int min_sched_tunable_scaling = SCHED_TUNABLESCALING_NONE;
 273 static int max_sched_tunable_scaling = SCHED_TUNABLESCALING_END-1;
 274 #endif /* CONFIG_SMP */
 275 #endif /* CONFIG_SCHED_DEBUG */
 276
 277 #ifdef CONFIG_COMPACTION
 278 static int min_extfrag_threshold;
 279 static int max_extfrag_threshold = 1000;
 280 #endif
 281
 282 static struct ctl_table kern_table[] = {
 283         {
 284                 .procname       = "sched_child_runs_first",
 285                 .data           = &sysctl_sched_child_runs_first,
 286                 .maxlen         = sizeof(unsigned int),
 287                 .mode           = 0644,
 288                 .proc_handler   = proc_dointvec,
 289         },
 290 #ifdef CONFIG_SCHED_DEBUG
 291         {
 292                 .procname       = "sched_min_granularity_ns",
 293                 .data           = &sysctl_sched_min_granularity,
 294                 .maxlen         = sizeof(unsigned int),
 295                 .mode           = 0644,
 296                 .proc_handler   = sched_proc_update_handler,
 297                 .extra1         = &min_sched_granularity_ns,
 298                 .extra2         = &max_sched_granularity_ns,
 299         },
 300         {
 301                 .procname       = "sched_latency_ns",
 302                 .data           = &sysctl_sched_latency,
 303                 .maxlen         = sizeof(unsigned int),
 304                 .mode           = 0644,
 305                 .proc_handler   = sched_proc_update_handler,
 306                 .extra1         = &min_sched_granularity_ns,
 307                 .extra2         = &max_sched_granularity_ns,
 308         },
 309         {
 310                 .procname       = "sched_wakeup_granularity_ns",
 311                 .data           = &sysctl_sched_wakeup_granularity,
 312                 .maxlen         = sizeof(unsigned int),
 313                 .mode           = 0644,
 314                 .proc_handler   = sched_proc_update_handler,
 315                 .extra1         = &min_wakeup_granularity_ns,
 316                 .extra2         = &max_wakeup_granularity_ns,
 317         },
 318 #ifdef CONFIG_SMP
 319         {
 320                 .procname       = "sched_tunable_scaling",
 321                 .data           = &sysctl_sched_tunable_scaling,
 322                 .maxlen         = sizeof(enum sched_tunable_scaling),
 323                 .mode           = 0644,
 324                 .proc_handler   = sched_proc_update_handler,
 325                 .extra1         = &min_sched_tunable_scaling,
 326                 .extra2         = &max_sched_tunable_scaling,
 327         },
 328         {
 329                 .procname       = "sched_migration_cost_ns",
 330                 .data           = &sysctl_sched_migration_cost,
 331                 .maxlen         = sizeof(unsigned int),
 332                 .mode           = 0644,
 333                 .proc_handler   = proc_dointvec,
 334         },
 335         {
 336                 .procname       = "sched_nr_migrate",
 337                 .data           = &sysctl_sched_nr_migrate,
 338                 .maxlen         = sizeof(unsigned int),
 339                 .mode           = 0644,
 340                 .proc_handler   = proc_dointvec,
 341         },
 342         {
 343                 .procname       = "sched_time_avg_ms",
 344                 .data           = &sysctl_sched_time_avg,
 345                 .maxlen         = sizeof(unsigned int),
 346                 .mode           = 0644,
 347                 .proc_handler   = proc_dointvec_minmax,
 348                 .extra1         = &one,
 349         },
 350         {
 351                 .procname       = "sched_shares_window_ns",
 352                 .data           = &sysctl_sched_shares_window,
 353                 .maxlen         = sizeof(unsigned int),
 354                 .mode           = 0644,
 355                 .proc_handler   = proc_dointvec,
 356         },
 357 #endif /* CONFIG_SMP */
 358 #ifdef CONFIG_NUMA_BALANCING
 359         {
 360                 .procname       = "numa_balancing_scan_delay_ms",
 361                 .data           = &sysctl_numa_balancing_scan_delay,
 362                 .maxlen         = sizeof(unsigned int),
 363                 .mode           = 0644,
 364                 .proc_handler   = proc_dointvec,
 365         },
 366         {
 367                 .procname       = "numa_balancing_scan_period_min_ms",
 368                 .data           = &sysctl_numa_balancing_scan_period_min,
 369                 .maxlen         = sizeof(unsigned int),
 370                 .mode           = 0644,
 371                 .proc_handler   = proc_dointvec,
 372         },
 373         {
 374                 .procname       = "numa_balancing_scan_period_max_ms",
 375                 .data           = &sysctl_numa_balancing_scan_period_max,
 376                 .maxlen         = sizeof(unsigned int),
 377                 .mode           = 0644,
 378                 .proc_handler   = proc_dointvec,
 379         },
 380         {
 381                 .procname       = "numa_balancing_scan_size_mb",
 382                 .data           = &sysctl_numa_balancing_scan_size,
 383                 .maxlen         = sizeof(unsigned int),
 384                 .mode           = 0644,
 385                 .proc_handler   = proc_dointvec_minmax,
 386                 .extra1         = &one,
 387         },
 388         {
 389                 .procname       = "numa_balancing",
 390                 .data           = NULL, /* filled in by handler */
 391                 .maxlen         = sizeof(unsigned int),
 392                 .mode           = 0644,
 393                 .proc_handler   = sysctl_numa_balancing,
 394                 .extra1         = &zero,
 395                 .extra2         = &one,
 396         },
 397 #endif /* CONFIG_NUMA_BALANCING */
 398 #endif /* CONFIG_SCHED_DEBUG */
 399         {
 400                 .procname       = "sched_rt_period_us",
 401                 .data           = &sysctl_sched_rt_period,
 402                 .maxlen         = sizeof(unsigned int),
 403                 .mode           = 0644,
 404                 .proc_handler   = sched_rt_handler,
 405         },
 406         {
 407                 .procname       = "sched_rt_runtime_us",
 408                 .data           = &sysctl_sched_rt_runtime,
 409                 .maxlen         = sizeof(int),
 410                 .mode           = 0644,
 411                 .proc_handler   = sched_rt_handler,
 412         },
 413         {
 414                 .procname       = "sched_rr_timeslice_ms",
 415                 .data           = &sched_rr_timeslice,
 416                 .maxlen         = sizeof(int),
 417                 .mode           = 0644,
 418                 .proc_handler   = sched_rr_handler,
 419         },
 420 #ifdef CONFIG_SCHED_AUTOGROUP
 421         {
 422                 .procname       = "sched_autogroup_enabled",
 423                 .data           = &sysctl_sched_autogroup_enabled,
 424                 .maxlen         = sizeof(unsigned int),
 425                 .mode           = 0644,
 426                 .proc_handler   = proc_dointvec_minmax,
 427                 .extra1         = &zero,
 428                 .extra2         = &one,
 429         },
 430 #endif
 431 #ifdef CONFIG_CFS_BANDWIDTH
 432         {
 433                 .procname       = "sched_cfs_bandwidth_slice_us",
 434                 .data           = &sysctl_sched_cfs_bandwidth_slice,
 435                 .maxlen         = sizeof(unsigned int),
 436                 .mode           = 0644,
 437                 .proc_handler   = proc_dointvec_minmax,
 438                 .extra1         = &one,
 439         },
 440 #endif
 441 #ifdef CONFIG_PROVE_LOCKING
 442         {
 443                 .procname       = "prove_locking",
 444                 .data           = &prove_locking,
 445                 .maxlen         = sizeof(int),
 446                 .mode           = 0644,
 447                 .proc_handler   = proc_dointvec,
 448         },
 449 #endif
 450 #ifdef CONFIG_LOCK_STAT
 451         {
 452                 .procname       = "lock_stat",
 453                 .data           = &lock_stat,
 454                 .maxlen         = sizeof(int),
 455                 .mode           = 0644,
 456                 .proc_handler   = proc_dointvec,
 457         },
 458 #endif
 459         {
 460                 .procname       = "panic",
 461                 .data           = &panic_timeout,
 462                 .maxlen         = sizeof(int),
 463                 .mode           = 0644,
 464                 .proc_handler   = proc_dointvec,
 465         },
 466 #ifdef CONFIG_COREDUMP
 467         {
 468                 .procname       = "core_uses_pid",
 469                 .data           = &core_uses_pid,
 470                 .maxlen         = sizeof(int),
 471                 .mode           = 0644,
 472                 .proc_handler   = proc_dointvec,
 473         },
 474         {
 475                 .procname       = "core_pattern",
 476                 .data           = core_pattern,
 477                 .maxlen         = CORENAME_MAX_SIZE,
 478                 .mode           = 0644,
 479                 .proc_handler   = proc_dostring_coredump,
 480         },
 481         {
 482                 .procname       = "core_pipe_limit",
 483                 .data           = &core_pipe_limit,
 484                 .maxlen         = sizeof(unsigned int),
 485                 .mode           = 0644,
 486                 .proc_handler   = proc_dointvec,
 487         },
 488 #endif
 489 #ifdef CONFIG_PROC_SYSCTL
 490         {
 491                 .procname       = "tainted",
 492                 .maxlen         = sizeof(long),
 493                 .mode           = 0644,
 494                 .proc_handler   = proc_taint,
 495         },
 496         {
 497                 .procname       = "sysctl_writes_strict",
 498                 .data           = &sysctl_writes_strict,
 499                 .maxlen         = sizeof(int),
 500                 .mode           = 0644,
 501                 .proc_handler   = proc_dointvec_minmax,
 502                 .extra1         = &neg_one,
 503                 .extra2         = &one,
 504         },
 505 #endif
 506 #ifdef CONFIG_LATENCYTOP
 507         {
 508                 .procname       = "latencytop",
 509                 .data           = &latencytop_enabled,
 510                 .maxlen         = sizeof(int),
 511                 .mode           = 0644,
 512                 .proc_handler   = proc_dointvec,
 513         },
 514 #endif
 515 #ifdef CONFIG_BLK_DEV_INITRD
 516         {
 517                 .procname       = "real-root-dev",
 518                 .data           = &real_root_dev,
 519                 .maxlen         = sizeof(int),
 520                 .mode           = 0644,
 521                 .proc_handler   = proc_dointvec,
 522         },
 523 #endif
 524         {
 525                 .procname       = "print-fatal-signals",
 526                 .data           = &print_fatal_signals,
 527                 .maxlen         = sizeof(int),
 528                 .mode           = 0644,
 529                 .proc_handler   = proc_dointvec,
 530         },
 531 #ifdef CONFIG_SPARC
 532         {
 533                 .procname       = "reboot-cmd",
 534                 .data           = reboot_command,
 535                 .maxlen         = 256,
 536                 .mode           = 0644,
 537                 .proc_handler   = proc_dostring,
 538         },
 539         {
 540                 .procname       = "stop-a",
 541                 .data           = &stop_a_enabled,
 542                 .maxlen         = sizeof (int),
 543                 .mode           = 0644,
 544                 .proc_handler   = proc_dointvec,
 545         },
 546         {
 547                 .procname       = "scons-poweroff",
 548                 .data           = &scons_pwroff,
 549                 .maxlen         = sizeof (int),
 550                 .mode           = 0644,
 551                 .proc_handler   = proc_dointvec,
 552         },
 553 #endif
 554 #ifdef CONFIG_SPARC64
 555         {
 556                 .procname       = "tsb-ratio",
 557                 .data           = &sysctl_tsb_ratio,
 558                 .maxlen         = sizeof (int),
 559                 .mode           = 0644,
 560                 .proc_handler   = proc_dointvec,
 561         },
 562 #endif
 563 #ifdef __hppa__
 564         {
 565                 .procname       = "soft-power",
 566                 .data           = &pwrsw_enabled,
 567                 .maxlen         = sizeof (int),
 568                 .mode           = 0644,
 569                 .proc_handler   = proc_dointvec,
 570         },
 571 #endif
 572 #ifdef CONFIG_SYSCTL_ARCH_UNALIGN_ALLOW
 573         {
 574                 .procname       = "unaligned-trap",
 575                 .data           = &unaligned_enabled,
 576                 .maxlen         = sizeof (int),
 577                 .mode           = 0644,
 578                 .proc_handler   = proc_dointvec,
 579         },
 580 #endif
 581         {
 582                 .procname       = "ctrl-alt-del",
 583                 .data           = &C_A_D,
 584                 .maxlen         = sizeof(int),
 585                 .mode           = 0644,
 586                 .proc_handler   = proc_dointvec,
 587         },
 588 #ifdef CONFIG_FUNCTION_TRACER
 589         {
 590                 .procname       = "ftrace_enabled",
 591                 .data           = &ftrace_enabled,
 592                 .maxlen         = sizeof(int),
 593                 .mode           = 0644,
 594                 .proc_handler   = ftrace_enable_sysctl,
 595         },
 596 #endif
 597 #ifdef CONFIG_STACK_TRACER
 598         {
 599                 .procname       = "stack_tracer_enabled",
 600                 .data           = &stack_tracer_enabled,
 601                 .maxlen         = sizeof(int),
 602                 .mode           = 0644,
 603                 .proc_handler   = stack_trace_sysctl,
 604         },
 605 #endif
 606 #ifdef CONFIG_TRACING
 607         {
 608                 .procname       = "ftrace_dump_on_oops",
 609                 .data           = &ftrace_dump_on_oops,
 610                 .maxlen         = sizeof(int),
 611                 .mode           = 0644,
 612                 .proc_handler   = proc_dointvec,
 613         },
 614         {
 615                 .procname       = "traceoff_on_warning",
 616                 .data           = &__disable_trace_on_warning,
 617                 .maxlen         = sizeof(__disable_trace_on_warning),
 618                 .mode           = 0644,
 619                 .proc_handler   = proc_dointvec,
 620         },
 621         {
 622                 .procname       = "tracepoint_printk",
 623                 .data           = &tracepoint_printk,
 624                 .maxlen         = sizeof(tracepoint_printk),
 625                 .mode           = 0644,
 626                 .proc_handler   = proc_dointvec,
 627         },
 628 #endif
 629 #ifdef CONFIG_KEXEC_CORE
 630         {
 631                 .procname       = "kexec_load_disabled",
 632                 .data           = &kexec_load_disabled,
 633                 .maxlen         = sizeof(int),
 634                 .mode           = 0644,
 635                 /* only handle a transition from default "0" to "1" */
 636                 .proc_handler   = proc_dointvec_minmax,
 637                 .extra1         = &one,
 638                 .extra2         = &one,
 639         },
 640 #endif
 641 #ifdef CONFIG_MODULES
 642         {
 643                 .procname       = "modprobe",
 644                 .data           = &modprobe_path,
 645                 .maxlen         = KMOD_PATH_LEN,
 646                 .mode           = 0644,
 647                 .proc_handler   = proc_dostring,
 648         },
 649         {
 650                 .procname       = "modules_disabled",
 651                 .data           = &modules_disabled,
 652                 .maxlen         = sizeof(int),
 653                 .mode           = 0644,
 654                 /* only handle a transition from default "0" to "1" */
 655                 .proc_handler   = proc_dointvec_minmax,
 656                 .extra1         = &one,
 657                 .extra2         = &one,
 658         },
 659 #endif
 660 #ifdef CONFIG_UEVENT_HELPER
 661         {
 662                 .procname       = "hotplug",
 663                 .data           = &uevent_helper,
 664                 .maxlen         = UEVENT_HELPER_PATH_LEN,
 665                 .mode           = 0644,
 666                 .proc_handler   = proc_dostring,
 667         },
 668 #endif
 669 #ifdef CONFIG_CHR_DEV_SG
 670         {
 671                 .procname       = "sg-big-buff",
 672                 .data           = &sg_big_buff,
 673                 .maxlen         = sizeof (int),
 674                 .mode           = 0444,
 675                 .proc_handler   = proc_dointvec,
 676         },
 677 #endif
 678 #ifdef CONFIG_BSD_PROCESS_ACCT
 679         {
 680                 .procname       = "acct",
 681                 .data           = &acct_parm,
 682                 .maxlen         = 3*sizeof(int),
 683                 .mode           = 0644,
 684                 .proc_handler   = proc_dointvec,
 685         },
 686 #endif
 687 #ifdef CONFIG_MAGIC_SYSRQ
 688         {
 689                 .procname       = "sysrq",
 690                 .data           = &__sysrq_enabled,
 691                 .maxlen         = sizeof (int),
 692                 .mode           = 0644,
 693                 .proc_handler   = sysrq_sysctl_handler,
 694         },
 695 #endif
 696 #ifdef CONFIG_PROC_SYSCTL
 697         {
 698                 .procname       = "cad_pid",
 699                 .data           = NULL,
 700                 .maxlen         = sizeof (int),
 701                 .mode           = 0600,
 702                 .proc_handler   = proc_do_cad_pid,
 703         },
 704 #endif
 705         {
 706                 .procname       = "threads-max",
 707                 .data           = NULL,
 708                 .maxlen         = sizeof(int),
 709                 .mode           = 0644,
 710                 .proc_handler   = sysctl_max_threads,
 711         },
 712         {
 713                 .procname       = "random",
 714                 .mode           = 0555,
 715                 .child          = random_table,
 716         },
 717         {
 718                 .procname       = "usermodehelper",
 719                 .mode           = 0555,
 720                 .child          = usermodehelper_table,
 721         },
 722         {
 723                 .procname       = "overflowuid",
 724                 .data           = &overflowuid,
 725                 .maxlen         = sizeof(int),
 726                 .mode           = 0644,
 727                 .proc_handler   = proc_dointvec_minmax,
 728                 .extra1         = &minolduid,
 729                 .extra2         = &maxolduid,
 730         },
 731         {
 732                 .procname       = "overflowgid",
 733                 .data           = &overflowgid,
 734                 .maxlen         = sizeof(int),
 735                 .mode           = 0644,
 736                 .proc_handler   = proc_dointvec_minmax,
 737                 .extra1         = &minolduid,
 738                 .extra2         = &maxolduid,
 739         },
 740 #ifdef CONFIG_S390
 741 #ifdef CONFIG_MATHEMU
 742         {
 743                 .procname       = "ieee_emulation_warnings",
 744                 .data           = &sysctl_ieee_emulation_warnings,
 745                 .maxlen         = sizeof(int),
 746                 .mode           = 0644,
 747                 .proc_handler   = proc_dointvec,
 748         },
 749 #endif
 750         {
 751                 .procname       = "userprocess_debug",
 752                 .data           = &show_unhandled_signals,
 753                 .maxlen         = sizeof(int),
 754                 .mode           = 0644,
 755                 .proc_handler   = proc_dointvec,
 756         },
 757 #endif
 758         {
 759                 .procname       = "pid_max",
 760                 .data           = &pid_max,
 761                 .maxlen         = sizeof (int),
 762                 .mode           = 0644,
 763                 .proc_handler   = proc_dointvec_minmax,
 764                 .extra1         = &pid_max_min,
 765                 .extra2         = &pid_max_max,
 766         },
 767         {
 768                 .procname       = "panic_on_oops",
 769                 .data           = &panic_on_oops,
 770                 .maxlen         = sizeof(int),
 771                 .mode           = 0644,
 772                 .proc_handler   = proc_dointvec,
 773         },
 774 #if defined CONFIG_PRINTK
 775         {
 776                 .procname       = "printk",
 777                 .data           = &console_loglevel,
 778                 .maxlen         = 4*sizeof(int),
 779                 .mode           = 0644,
 780                 .proc_handler   = proc_dointvec,
 781         },
 782         {
 783                 .procname       = "printk_ratelimit",
 784                 .data           = &printk_ratelimit_state.interval,
 785                 .maxlen         = sizeof(int),
 786                 .mode           = 0644,
 787                 .proc_handler   = proc_dointvec_jiffies,
 788         },
 789         {
 790                 .procname       = "printk_ratelimit_burst",
 791                 .data           = &printk_ratelimit_state.burst,
 792                 .maxlen         = sizeof(int),
 793                 .mode           = 0644,
 794                 .proc_handler   = proc_dointvec,
 795         },
 796         {
 797                 .procname       = "printk_delay",
 798                 .data           = &printk_delay_msec,
 799                 .maxlen         = sizeof(int),
 800                 .mode           = 0644,
 801                 .proc_handler   = proc_dointvec_minmax,
 802                 .extra1         = &zero,
 803                 .extra2         = &ten_thousand,
 804         },
 805         {
 806                 .procname       = "dmesg_restrict",
 807                 .data           = &dmesg_restrict,
 808                 .maxlen         = sizeof(int),
 809                 .mode           = 0644,
 810                 .proc_handler   = proc_dointvec_minmax_sysadmin,
 811                 .extra1         = &zero,
 812                 .extra2         = &one,
 813         },
 814         {
 815                 .procname       = "kptr_restrict",
 816                 .data           = &kptr_restrict,
 817                 .maxlen         = sizeof(int),
 818                 .mode           = 0644,
 819                 .proc_handler   = proc_dointvec_minmax_sysadmin,
 820                 .extra1         = &zero,
 821                 .extra2         = &two,
 822         },
 823 #endif
 824         {
 825                 .procname       = "ngroups_max",
 826                 .data           = &ngroups_max,
 827                 .maxlen         = sizeof (int),
 828                 .mode           = 0444,
 829                 .proc_handler   = proc_dointvec,
 830         },
 831         {
 832                 .procname       = "cap_last_cap",
 833                 .data           = (void *)&cap_last_cap,
 834                 .maxlen         = sizeof(int),
 835                 .mode           = 0444,
 836                 .proc_handler   = proc_dointvec,
 837         },
 838 #if defined(CONFIG_LOCKUP_DETECTOR)
 839         {
 840                 .procname       = "watchdog",
 841                 .data           = &watchdog_user_enabled,
 842                 .maxlen         = sizeof (int),
 843                 .mode           = 0644,
 844                 .proc_handler   = proc_watchdog,
 845                 .extra1         = &zero,
 846                 .extra2         = &one,
 847         },
 848         {
 849                 .procname       = "watchdog_thresh",
 850                 .data           = &watchdog_thresh,
 851                 .maxlen         = sizeof(int),
 852                 .mode           = 0644,
 853                 .proc_handler   = proc_watchdog_thresh,
 854                 .extra1         = &zero,
 855                 .extra2         = &sixty,
 856         },
 857         {
 858                 .procname       = "nmi_watchdog",
 859                 .data           = &nmi_watchdog_enabled,
 860                 .maxlen         = sizeof (int),
 861                 .mode           = 0644,
 862                 .proc_handler   = proc_nmi_watchdog,
 863                 .extra1         = &zero,
 864 #if defined(CONFIG_HAVE_NMI_WATCHDOG) || defined(CONFIG_HARDLOCKUP_DETECTOR)
 865                 .extra2         = &one,
 866 #else
 867                 .extra2         = &zero,
 868 #endif
 869         },
 870         {
 871                 .procname       = "soft_watchdog",
 872                 .data           = &soft_watchdog_enabled,
 873                 .maxlen         = sizeof (int),
 874                 .mode           = 0644,
 875                 .proc_handler   = proc_soft_watchdog,
 876                 .extra1         = &zero,
 877                 .extra2         = &one,
 878         },
 879         {
 880                 .procname       = "watchdog_cpumask",
 881                 .data           = &watchdog_cpumask_bits,
 882                 .maxlen         = NR_CPUS,
 883                 .mode           = 0644,
 884                 .proc_handler   = proc_watchdog_cpumask,
 885         },
 886         {
 887                 .procname       = "softlockup_panic",
 888                 .data           = &softlockup_panic,
 889                 .maxlen         = sizeof(int),
 890                 .mode           = 0644,
 891                 .proc_handler   = proc_dointvec_minmax,
 892                 .extra1         = &zero,
 893                 .extra2         = &one,
 894         },
 895 #ifdef CONFIG_HARDLOCKUP_DETECTOR
 896         {
 897                 .procname       = "hardlockup_panic",
 898                 .data           = &hardlockup_panic,
 899                 .maxlen         = sizeof(int),
 900                 .mode           = 0644,
 901                 .proc_handler   = proc_dointvec_minmax,
 902                 .extra1         = &zero,
 903                 .extra2         = &one,
 904         },
 905 #endif
 906 #ifdef CONFIG_SMP
 907         {
 908                 .procname       = "softlockup_all_cpu_backtrace",
 909                 .data           = &sysctl_softlockup_all_cpu_backtrace,
 910                 .maxlen         = sizeof(int),
 911                 .mode           = 0644,
 912                 .proc_handler   = proc_dointvec_minmax,
 913                 .extra1         = &zero,
 914                 .extra2         = &one,
 915         },
 916         {
 917                 .procname       = "hardlockup_all_cpu_backtrace",
 918                 .data           = &sysctl_hardlockup_all_cpu_backtrace,
 919                 .maxlen         = sizeof(int),
 920                 .mode           = 0644,
 921                 .proc_handler   = proc_dointvec_minmax,
 922                 .extra1         = &zero,
 923                 .extra2         = &one,
 924         },
 925 #endif /* CONFIG_SMP */
 926 #endif
 927 #if defined(CONFIG_X86_LOCAL_APIC) && defined(CONFIG_X86)
 928         {
 929                 .procname       = "unknown_nmi_panic",
 930                 .data           = &unknown_nmi_panic,
 931                 .maxlen         = sizeof (int),
 932                 .mode           = 0644,
 933                 .proc_handler   = proc_dointvec,
 934         },
 935 #endif
 936 #if defined(CONFIG_X86)
 937         {
 938                 .procname       = "panic_on_unrecovered_nmi",
 939                 .data           = &panic_on_unrecovered_nmi,
 940                 .maxlen         = sizeof(int),
 941                 .mode           = 0644,
 942                 .proc_handler   = proc_dointvec,
 943         },
 944         {
 945                 .procname       = "panic_on_io_nmi",
 946                 .data           = &panic_on_io_nmi,
 947                 .maxlen         = sizeof(int),
 948                 .mode           = 0644,
 949                 .proc_handler   = proc_dointvec,
 950         },
 951 #ifdef CONFIG_DEBUG_STACKOVERFLOW
 952         {
 953                 .procname       = "panic_on_stackoverflow",
 954                 .data           = &sysctl_panic_on_stackoverflow,
 955                 .maxlen         = sizeof(int),
 956                 .mode           = 0644,
 957                 .proc_handler   = proc_dointvec,
 958         },
 959 #endif
 960         {
 961                 .procname       = "bootloader_type",
 962                 .data           = &bootloader_type,
 963                 .maxlen         = sizeof (int),
 964                 .mode           = 0444,
 965                 .proc_handler   = proc_dointvec,
 966         },
 967         {
 968                 .procname       = "bootloader_version",
 969                 .data           = &bootloader_version,
 970                 .maxlen         = sizeof (int),
 971                 .mode           = 0444,
 972                 .proc_handler   = proc_dointvec,
 973         },
 974         {
 975                 .procname       = "kstack_depth_to_print",
 976                 .data           = &kstack_depth_to_print,
 977                 .maxlen         = sizeof(int),
 978                 .mode           = 0644,
 979                 .proc_handler   = proc_dointvec,
 980         },
 981         {
 982                 .procname       = "io_delay_type",
 983                 .data           = &io_delay_type,
 984                 .maxlen         = sizeof(int),
 985                 .mode           = 0644,
 986                 .proc_handler   = proc_dointvec,
 987         },
 988 #endif
 989 #if defined(CONFIG_MMU)
 990         {
 991                 .procname       = "randomize_va_space",
 992                 .data           = &randomize_va_space,
 993                 .maxlen         = sizeof(int),
 994                 .mode           = 0644,
 995                 .proc_handler   = proc_dointvec,
 996         },
 997 #endif
 998 #if defined(CONFIG_S390) && defined(CONFIG_SMP)
 999         {
1000                 .procname       = "spin_retry",
1001                 .data           = &spin_retry,
1002                 .maxlen         = sizeof (int),
1003                 .mode           = 0644,
1004                 .proc_handler   = proc_dointvec,
1005         },
1006 #endif
1007 #if     defined(CONFIG_ACPI_SLEEP) && defined(CONFIG_X86)
1008         {
1009                 .procname       = "acpi_video_flags",
1010                 .data           = &acpi_realmode_flags,
1011                 .maxlen         = sizeof (unsigned long),
1012                 .mode           = 0644,
1013                 .proc_handler   = proc_doulongvec_minmax,
1014         },
1015 #endif
1016 #ifdef CONFIG_SYSCTL_ARCH_UNALIGN_NO_WARN
1017         {
1018                 .procname       = "ignore-unaligned-usertrap",
1019                 .data           = &no_unaligned_warning,
1020                 .maxlen         = sizeof (int),
1021                 .mode           = 0644,
1022                 .proc_handler   = proc_dointvec,
1023         },
1024 #endif
1025 #ifdef CONFIG_IA64
1026         {
1027                 .procname       = "unaligned-dump-stack",
1028                 .data           = &unaligned_dump_stack,
1029                 .maxlen         = sizeof (int),
1030                 .mode           = 0644,
1031                 .proc_handler   = proc_dointvec,
1032         },
1033 #endif
1034 #ifdef CONFIG_DETECT_HUNG_TASK
1035         {
1036                 .procname       = "hung_task_panic",
1037                 .data           = &sysctl_hung_task_panic,
1038                 .maxlen         = sizeof(int),
1039                 .mode           = 0644,
1040                 .proc_handler   = proc_dointvec_minmax,
1041                 .extra1         = &zero,
1042                 .extra2         = &one,
1043         },
1044         {
1045                 .procname       = "hung_task_check_count",
1046                 .data           = &sysctl_hung_task_check_count,
1047                 .maxlen         = sizeof(int),
1048                 .mode           = 0644,
1049                 .proc_handler   = proc_dointvec_minmax,
1050                 .extra1         = &zero,
1051         },
1052         {
1053                 .procname       = "hung_task_timeout_secs",
1054                 .data           = &sysctl_hung_task_timeout_secs,
1055                 .maxlen         = sizeof(unsigned long),
1056                 .mode           = 0644,
1057                 .proc_handler   = proc_dohung_task_timeout_secs,
1058                 .extra2         = &hung_task_timeout_max,
1059         },
1060         {
1061                 .procname       = "hung_task_warnings",
1062                 .data           = &sysctl_hung_task_warnings,
1063                 .maxlen         = sizeof(int),
1064                 .mode           = 0644,
1065                 .proc_handler   = proc_dointvec_minmax,
1066                 .extra1         = &neg_one,
1067         },
1068 #endif
1069 #ifdef CONFIG_COMPAT
1070         {
1071                 .procname       = "compat-log",
1072                 .data           = &compat_log,
1073                 .maxlen         = sizeof (int),
1074                 .mode           = 0644,
1075                 .proc_handler   = proc_dointvec,
1076         },
1077 #endif
1078 #ifdef CONFIG_RT_MUTEXES
1079         {
1080                 .procname       = "max_lock_depth",
1081                 .data           = &max_lock_depth,
1082                 .maxlen         = sizeof(int),
1083                 .mode           = 0644,
1084                 .proc_handler   = proc_dointvec,
1085         },
1086 #endif
1087         {
1088                 .procname       = "poweroff_cmd",
1089                 .data           = &poweroff_cmd,
1090                 .maxlen         = POWEROFF_CMD_PATH_LEN,
1091                 .mode           = 0644,
1092                 .proc_handler   = proc_dostring,
1093         },
1094 #ifdef CONFIG_KEYS
1095         {
1096                 .procname       = "keys",
1097                 .mode           = 0555,
1098                 .child          = key_sysctls,
1099         },
1100 #endif
1101 #ifdef CONFIG_PERF_EVENTS
1102         /*
1103          * User-space scripts rely on the existence of this file
1104          * as a feature check for perf_events being enabled.
1105          *
1106          * So it's an ABI, do not remove!
1107          */
1108         {
1109                 .procname       = "perf_event_paranoid",
1110                 .data           = &sysctl_perf_event_paranoid,
1111                 .maxlen         = sizeof(sysctl_perf_event_paranoid),
1112                 .mode           = 0644,
1113                 .proc_handler   = proc_dointvec,
1114         },
1115         {
1116                 .procname       = "perf_event_mlock_kb",
1117                 .data           = &sysctl_perf_event_mlock,
1118                 .maxlen         = sizeof(sysctl_perf_event_mlock),
1119                 .mode           = 0644,
1120                 .proc_handler   = proc_dointvec,
1121         },
1122         {
1123                 .procname       = "perf_event_max_sample_rate",
1124                 .data           = &sysctl_perf_event_sample_rate,
1125                 .maxlen         = sizeof(sysctl_perf_event_sample_rate),
1126                 .mode           = 0644,
1127                 .proc_handler   = perf_proc_update_handler,
1128                 .extra1         = &one,
1129         },
1130         {
1131                 .procname       = "perf_cpu_time_max_percent",
1132                 .data           = &sysctl_perf_cpu_time_max_percent,
1133                 .maxlen         = sizeof(sysctl_perf_cpu_time_max_percent),
1134                 .mode           = 0644,
1135                 .proc_handler   = perf_cpu_time_max_percent_handler,
1136                 .extra1         = &zero,
1137                 .extra2         = &one_hundred,
1138         },
1139 #endif
1140 #ifdef CONFIG_KMEMCHECK
1141         {
1142                 .procname       = "kmemcheck",
1143                 .data           = &kmemcheck_enabled,
1144                 .maxlen         = sizeof(int),
1145                 .mode           = 0644,
1146                 .proc_handler   = proc_dointvec,
1147         },
1148 #endif
1149         {
1150                 .procname       = "panic_on_warn",
1151                 .data           = &panic_on_warn,
1152                 .maxlen         = sizeof(int),
1153                 .mode           = 0644,
1154                 .proc_handler   = proc_dointvec_minmax,
1155                 .extra1         = &zero,
1156                 .extra2         = &one,
1157         },
1158 #if defined(CONFIG_SMP) && defined(CONFIG_NO_HZ_COMMON)
1159         {
1160                 .procname       = "timer_migration",
1161                 .data           = &sysctl_timer_migration,
1162                 .maxlen         = sizeof(unsigned int),
1163                 .mode           = 0644,
1164                 .proc_handler   = timer_migration_handler,
1165                 .extra1         = &zero,
1166                 .extra2         = &one,
1167         },
1168 #endif
1169 #ifdef CONFIG_BPF_SYSCALL
1170         {
1171                 .procname       = "unprivileged_bpf_disabled",
1172                 .data           = &sysctl_unprivileged_bpf_disabled,
1173                 .maxlen         = sizeof(sysctl_unprivileged_bpf_disabled),
1174                 .mode           = 0644,
1175                 /* only handle a transition from default "0" to "1" */
1176                 .proc_handler   = proc_dointvec_minmax,
1177                 .extra1         = &one,
1178                 .extra2         = &one,
1179         },
1180 #endif
1181         { }
1182 };
1183
1184 static struct ctl_table vm_table[] = {
1185         {
1186                 .procname       = "overcommit_memory",
1187                 .data           = &sysctl_overcommit_memory,
1188                 .maxlen         = sizeof(sysctl_overcommit_memory),
1189                 .mode           = 0644,
1190                 .proc_handler   = proc_dointvec_minmax,
1191                 .extra1         = &zero,
1192                 .extra2         = &two,
1193         },
1194         {
1195                 .procname       = "panic_on_oom",
1196                 .data           = &sysctl_panic_on_oom,
1197                 .maxlen         = sizeof(sysctl_panic_on_oom),
1198                 .mode           = 0644,
1199                 .proc_handler   = proc_dointvec_minmax,
1200                 .extra1         = &zero,
1201                 .extra2         = &two,
1202         },
1203         {
1204                 .procname       = "oom_kill_allocating_task",
1205                 .data           = &sysctl_oom_kill_allocating_task,
1206                 .maxlen         = sizeof(sysctl_oom_kill_allocating_task),
1207                 .mode           = 0644,
1208                 .proc_handler   = proc_dointvec,
1209         },
1210         {
1211                 .procname       = "oom_dump_tasks",
1212                 .data           = &sysctl_oom_dump_tasks,
1213                 .maxlen         = sizeof(sysctl_oom_dump_tasks),
1214                 .mode           = 0644,
1215                 .proc_handler   = proc_dointvec,
1216         },
1217         {
1218                 .procname       = "overcommit_ratio",
1219                 .data           = &sysctl_overcommit_ratio,
1220                 .maxlen         = sizeof(sysctl_overcommit_ratio),
1221                 .mode           = 0644,
1222                 .proc_handler   = overcommit_ratio_handler,
1223         },
1224         {
1225                 .procname       = "overcommit_kbytes",
1226                 .data           = &sysctl_overcommit_kbytes,
1227                 .maxlen         = sizeof(sysctl_overcommit_kbytes),
1228                 .mode           = 0644,
1229                 .proc_handler   = overcommit_kbytes_handler,
1230         },
1231         {
1232                 .procname       = "page-cluster",
1233                 .data           = &page_cluster,
1234                 .maxlen         = sizeof(int),
1235                 .mode           = 0644,
1236                 .proc_handler   = proc_dointvec_minmax,
1237                 .extra1         = &zero,
1238         },
1239         {
1240                 .procname       = "dirty_background_ratio",
1241                 .data           = &dirty_background_ratio,
1242                 .maxlen         = sizeof(dirty_background_ratio),
1243                 .mode           = 0644,
1244                 .proc_handler   = dirty_background_ratio_handler,
1245                 .extra1         = &zero,
1246                 .extra2         = &one_hundred,
1247         },
1248         {
1249                 .procname       = "dirty_background_bytes",
1250                 .data           = &dirty_background_bytes,
1251                 .maxlen         = sizeof(dirty_background_bytes),
1252                 .mode           = 0644,
1253                 .proc_handler   = dirty_background_bytes_handler,
1254                 .extra1         = &one_ul,
1255         },
1256         {
1257                 .procname       = "dirty_ratio",
1258                 .data           = &vm_dirty_ratio,
1259                 .maxlen         = sizeof(vm_dirty_ratio),
1260                 .mode           = 0644,
1261                 .proc_handler   = dirty_ratio_handler,
1262                 .extra1         = &zero,
1263                 .extra2         = &one_hundred,
1264         },
1265         {
1266                 .procname       = "dirty_bytes",
1267                 .data           = &vm_dirty_bytes,
1268                 .maxlen         = sizeof(vm_dirty_bytes),
1269                 .mode           = 0644,
1270                 .proc_handler   = dirty_bytes_handler,
1271                 .extra1         = &dirty_bytes_min,
1272         },
1273         {
1274                 .procname       = "dirty_writeback_centisecs",
1275                 .data           = &dirty_writeback_interval,
1276                 .maxlen         = sizeof(dirty_writeback_interval),
1277                 .mode           = 0644,
1278                 .proc_handler   = dirty_writeback_centisecs_handler,
1279         },
1280         {
1281                 .procname       = "dirty_expire_centisecs",
1282                 .data           = &dirty_expire_interval,
1283                 .maxlen         = sizeof(dirty_expire_interval),
1284                 .mode           = 0644,
1285                 .proc_handler   = proc_dointvec_minmax,
1286                 .extra1         = &zero,
1287         },
1288         {
1289                 .procname       = "dirtytime_expire_seconds",
1290                 .data           = &dirtytime_expire_interval,
1291                 .maxlen         = sizeof(dirty_expire_interval),
1292                 .mode           = 0644,
1293                 .proc_handler   = dirtytime_interval_handler,
1294                 .extra1         = &zero,
1295         },
1296         {
1297                 .procname       = "nr_pdflush_threads",
1298                 .mode           = 0444 /* read-only */,
1299                 .proc_handler   = pdflush_proc_obsolete,
1300         },
1301         {
1302                 .procname       = "swappiness",
1303                 .data           = &vm_swappiness,
1304                 .maxlen         = sizeof(vm_swappiness),
1305                 .mode           = 0644,
1306                 .proc_handler   = proc_dointvec_minmax,
1307                 .extra1         = &zero,
1308                 .extra2         = &one_hundred,
1309         },
1310 #ifdef CONFIG_HUGETLB_PAGE
1311         {
1312                 .procname       = "nr_hugepages",
1313                 .data           = NULL,
1314                 .maxlen         = sizeof(unsigned long),
1315                 .mode           = 0644,
1316                 .proc_handler   = hugetlb_sysctl_handler,
1317         },
1318 #ifdef CONFIG_NUMA
1319         {
1320                 .procname       = "nr_hugepages_mempolicy",
1321                 .data           = NULL,
1322                 .maxlen         = sizeof(unsigned long),
1323                 .mode           = 0644,
1324                 .proc_handler   = &hugetlb_mempolicy_sysctl_handler,
1325         },
1326 #endif
1327          {
1328                 .procname       = "hugetlb_shm_group",
1329                 .data           = &sysctl_hugetlb_shm_group,
1330                 .maxlen         = sizeof(gid_t),
1331                 .mode           = 0644,
1332                 .proc_handler   = proc_dointvec,
1333          },
1334          {
1335                 .procname       = "hugepages_treat_as_movable",
1336                 .data           = &hugepages_treat_as_movable,
1337                 .maxlen         = sizeof(int),
1338                 .mode           = 0644,
1339                 .proc_handler   = proc_dointvec,
1340         },
1341         {
1342                 .procname       = "nr_overcommit_hugepages",
1343                 .data           = NULL,
1344                 .maxlen         = sizeof(unsigned long),
1345                 .mode           = 0644,
1346                 .proc_handler   = hugetlb_overcommit_handler,
1347         },
1348 #endif
1349         {
1350                 .procname       = "lowmem_reserve_ratio",
1351                 .data           = &sysctl_lowmem_reserve_ratio,
1352                 .maxlen         = sizeof(sysctl_lowmem_reserve_ratio),
1353                 .mode           = 0644,
1354                 .proc_handler   = lowmem_reserve_ratio_sysctl_handler,
1355         },
1356         {
1357                 .procname       = "drop_caches",
1358                 .data           = &sysctl_drop_caches,
1359                 .maxlen         = sizeof(int),
1360                 .mode           = 0200,
1361                 .proc_handler   = drop_caches_sysctl_handler,
1362                 .extra1         = &one,
1363                 .extra2         = &four,
1364         },
1365 #ifdef CONFIG_COMPACTION
1366         {
1367                 .procname       = "compact_memory",
1368                 .data           = &sysctl_compact_memory,
1369                 .maxlen         = sizeof(int),
1370                 .mode           = 0200,
1371                 .proc_handler   = sysctl_compaction_handler,
1372         },
1373         {
1374                 .procname       = "extfrag_threshold",
1375                 .data           = &sysctl_extfrag_threshold,
1376                 .maxlen         = sizeof(int),
1377                 .mode           = 0644,
1378                 .proc_handler   = sysctl_extfrag_handler,
1379                 .extra1         = &min_extfrag_threshold,
1380                 .extra2         = &max_extfrag_threshold,
1381         },
1382         {
1383                 .procname       = "compact_unevictable_allowed",
1384                 .data           = &sysctl_compact_unevictable_allowed,
1385                 .maxlen         = sizeof(int),
1386                 .mode           = 0644,
1387                 .proc_handler   = proc_dointvec,
1388                 .extra1         = &zero,
1389                 .extra2         = &one,
1390         },
1391
1392 #endif /* CONFIG_COMPACTION */
1393         {
1394                 .procname       = "min_free_kbytes",
1395                 .data           = &min_free_kbytes,
1396                 .maxlen         = sizeof(min_free_kbytes),
1397                 .mode           = 0644,
1398                 .proc_handler   = min_free_kbytes_sysctl_handler,
1399                 .extra1         = &zero,
1400         },
1401         {
1402                 .procname       = "percpu_pagelist_fraction",
1403                 .data           = &percpu_pagelist_fraction,
1404                 .maxlen         = sizeof(percpu_pagelist_fraction),
1405                 .mode           = 0644,
1406                 .proc_handler   = percpu_pagelist_fraction_sysctl_handler,
1407                 .extra1         = &zero,
1408         },
1409 #ifdef CONFIG_MMU
1410         {
1411                 .procname       = "max_map_count",
1412                 .data           = &sysctl_max_map_count,
1413                 .maxlen         = sizeof(sysctl_max_map_count),
1414                 .mode           = 0644,
1415                 .proc_handler   = proc_dointvec_minmax,
1416                 .extra1         = &zero,
1417         },
1418 #else
1419         {
1420                 .procname       = "nr_trim_pages",
1421                 .data           = &sysctl_nr_trim_pages,
1422                 .maxlen         = sizeof(sysctl_nr_trim_pages),
1423                 .mode           = 0644,
1424                 .proc_handler   = proc_dointvec_minmax,
1425                 .extra1         = &zero,
1426         },
1427 #endif
1428         {
1429                 .procname       = "laptop_mode",
1430                 .data           = &laptop_mode,
1431                 .maxlen         = sizeof(laptop_mode),
1432                 .mode           = 0644,
1433                 .proc_handler   = proc_dointvec_jiffies,
1434         },
1435         {
1436                 .procname       = "block_dump",
1437                 .data           = &block_dump,
1438                 .maxlen         = sizeof(block_dump),
1439                 .mode           = 0644,
1440                 .proc_handler   = proc_dointvec,
1441                 .extra1         = &zero,
1442         },
1443         {
1444                 .procname       = "vfs_cache_pressure",
1445                 .data           = &sysctl_vfs_cache_pressure,
1446                 .maxlen         = sizeof(sysctl_vfs_cache_pressure),
1447                 .mode           = 0644,
1448                 .proc_handler   = proc_dointvec,
1449                 .extra1         = &zero,
1450         },
1451 #ifdef HAVE_ARCH_PICK_MMAP_LAYOUT
1452         {
1453                 .procname       = "legacy_va_layout",
1454                 .data           = &sysctl_legacy_va_layout,
1455                 .maxlen         = sizeof(sysctl_legacy_va_layout),
1456                 .mode           = 0644,
1457                 .proc_handler   = proc_dointvec,
1458                 .extra1         = &zero,
1459         },
1460 #endif
1461 #ifdef CONFIG_NUMA
1462         {
1463                 .procname       = "zone_reclaim_mode",
1464                 .data           = &zone_reclaim_mode,
1465                 .maxlen         = sizeof(zone_reclaim_mode),
1466                 .mode           = 0644,
1467                 .proc_handler   = proc_dointvec,
1468                 .extra1         = &zero,
1469         },
1470         {
1471                 .procname       = "min_unmapped_ratio",
1472                 .data           = &sysctl_min_unmapped_ratio,
1473                 .maxlen         = sizeof(sysctl_min_unmapped_ratio),
1474                 .mode           = 0644,
1475                 .proc_handler   = sysctl_min_unmapped_ratio_sysctl_handler,
1476                 .extra1         = &zero,
1477                 .extra2         = &one_hundred,
1478         },
1479         {
1480                 .procname       = "min_slab_ratio",
1481                 .data           = &sysctl_min_slab_ratio,
1482                 .maxlen         = sizeof(sysctl_min_slab_ratio),
1483                 .mode           = 0644,
1484                 .proc_handler   = sysctl_min_slab_ratio_sysctl_handler,
1485                 .extra1         = &zero,
1486                 .extra2         = &one_hundred,
1487         },
1488 #endif
1489 #ifdef CONFIG_SMP
1490         {
1491                 .procname       = "stat_interval",
1492                 .data           = &sysctl_stat_interval,
1493                 .maxlen         = sizeof(sysctl_stat_interval),
1494                 .mode           = 0644,
1495                 .proc_handler   = proc_dointvec_jiffies,
1496         },
1497 #endif
1498 #ifdef CONFIG_MMU
1499         {
1500                 .procname       = "mmap_min_addr",
1501                 .data           = &dac_mmap_min_addr,
1502                 .maxlen         = sizeof(unsigned long),
1503                 .mode           = 0644,
1504                 .proc_handler   = mmap_min_addr_handler,
1505         },
1506 #endif
1507 #ifdef CONFIG_NUMA
1508         {
1509                 .procname       = "numa_zonelist_order",
1510                 .data           = &numa_zonelist_order,
1511                 .maxlen         = NUMA_ZONELIST_ORDER_LEN,
1512                 .mode           = 0644,
1513                 .proc_handler   = numa_zonelist_order_handler,
1514         },
1515 #endif
1516 #if (defined(CONFIG_X86_32) && !defined(CONFIG_UML))|| \
1517    (defined(CONFIG_SUPERH) && defined(CONFIG_VSYSCALL))
1518         {
1519                 .procname       = "vdso_enabled",
1520 #ifdef CONFIG_X86_32
1521                 .data           = &vdso32_enabled,
1522                 .maxlen         = sizeof(vdso32_enabled),
1523 #else
1524                 .data           = &vdso_enabled,
1525                 .maxlen         = sizeof(vdso_enabled),
1526 #endif
1527                 .mode           = 0644,
1528                 .proc_handler   = proc_dointvec,
1529                 .extra1         = &zero,
1530         },
1531 #endif
1532 #ifdef CONFIG_HIGHMEM
1533         {
1534                 .procname       = "highmem_is_dirtyable",
1535                 .data           = &vm_highmem_is_dirtyable,
1536                 .maxlen         = sizeof(vm_highmem_is_dirtyable),
1537                 .mode           = 0644,
1538                 .proc_handler   = proc_dointvec_minmax,
1539                 .extra1         = &zero,
1540                 .extra2         = &one,
1541         },
1542 #endif
1543 #ifdef CONFIG_MEMORY_FAILURE
1544         {
1545                 .procname       = "memory_failure_early_kill",
1546                 .data           = &sysctl_memory_failure_early_kill,
1547                 .maxlen         = sizeof(sysctl_memory_failure_early_kill),
1548                 .mode           = 0644,
1549                 .proc_handler   = proc_dointvec_minmax,
1550                 .extra1         = &zero,
1551                 .extra2         = &one,
1552         },
1553         {
1554                 .procname       = "memory_failure_recovery",
1555                 .data           = &sysctl_memory_failure_recovery,
1556                 .maxlen         = sizeof(sysctl_memory_failure_recovery),
1557                 .mode           = 0644,
1558                 .proc_handler   = proc_dointvec_minmax,
1559                 .extra1         = &zero,
1560                 .extra2         = &one,
1561         },
1562 #endif
1563         {
1564                 .procname       = "user_reserve_kbytes",
1565                 .data           = &sysctl_user_reserve_kbytes,
1566                 .maxlen         = sizeof(sysctl_user_reserve_kbytes),
1567                 .mode           = 0644,
1568                 .proc_handler   = proc_doulongvec_minmax,
1569         },
1570         {
1571                 .procname       = "admin_reserve_kbytes",
1572                 .data           = &sysctl_admin_reserve_kbytes,
1573                 .maxlen         = sizeof(sysctl_admin_reserve_kbytes),
1574                 .mode           = 0644,
1575                 .proc_handler   = proc_doulongvec_minmax,
1576         },
1577         { }
1578 };
1579
1580 static struct ctl_table fs_table[] = {
1581         {
1582                 .procname       = "inode-nr",
1583                 .data           = &inodes_stat,
1584                 .maxlen         = 2*sizeof(long),
1585                 .mode           = 0444,
1586                 .proc_handler   = proc_nr_inodes,
1587         },
1588         {
1589                 .procname       = "inode-state",
1590                 .data           = &inodes_stat,
1591                 .maxlen         = 7*sizeof(long),
1592                 .mode           = 0444,
1593                 .proc_handler   = proc_nr_inodes,
1594         },
1595         {
1596                 .procname       = "file-nr",
1597                 .data           = &files_stat,
1598                 .maxlen         = sizeof(files_stat),
1599                 .mode           = 0444,
1600                 .proc_handler   = proc_nr_files,
1601         },
1602         {
1603                 .procname       = "file-max",
1604                 .data           = &files_stat.max_files,
1605                 .maxlen         = sizeof(files_stat.max_files),
1606                 .mode           = 0644,
1607                 .proc_handler   = proc_doulongvec_minmax,
1608                 .extra1         = &zero_ul,
1609                 .extra2         = &long_max,
1610         },
1611         {
1612                 .procname       = "nr_open",
1613                 .data           = &sysctl_nr_open,
1614                 .maxlen         = sizeof(int),
1615                 .mode           = 0644,
1616                 .proc_handler   = proc_dointvec_minmax,
1617                 .extra1         = &sysctl_nr_open_min,
1618                 .extra2         = &sysctl_nr_open_max,
1619         },
1620         {
1621                 .procname       = "dentry-state",
1622                 .data           = &dentry_stat,
1623                 .maxlen         = 6*sizeof(long),
1624                 .mode           = 0444,
1625                 .proc_handler   = proc_nr_dentry,
1626         },
1627         {
1628                 .procname       = "overflowuid",
1629                 .data           = &fs_overflowuid,
1630                 .maxlen         = sizeof(int),
1631                 .mode           = 0644,
1632                 .proc_handler   = proc_dointvec_minmax,
1633                 .extra1         = &minolduid,
1634                 .extra2         = &maxolduid,
1635         },
1636         {
1637                 .procname       = "overflowgid",
1638                 .data           = &fs_overflowgid,
1639                 .maxlen         = sizeof(int),
1640                 .mode           = 0644,
1641                 .proc_handler   = proc_dointvec_minmax,
1642                 .extra1         = &minolduid,
1643                 .extra2         = &maxolduid,
1644         },
1645 #ifdef CONFIG_FILE_LOCKING
1646         {
1647                 .procname       = "leases-enable",
1648                 .data           = &leases_enable,
1649                 .maxlen         = sizeof(int),
1650                 .mode           = 0644,
1651                 .proc_handler   = proc_dointvec,
1652         },
1653 #endif
1654 #ifdef CONFIG_DNOTIFY
1655         {
1656                 .procname       = "dir-notify-enable",
1657                 .data           = &dir_notify_enable,
1658                 .maxlen         = sizeof(int),
1659                 .mode           = 0644,
1660                 .proc_handler   = proc_dointvec,
1661         },
1662 #endif
1663 #ifdef CONFIG_MMU
1664 #ifdef CONFIG_FILE_LOCKING
1665         {
1666                 .procname       = "lease-break-time",
1667                 .data           = &lease_break_time,
1668                 .maxlen         = sizeof(int),
1669                 .mode           = 0644,
1670                 .proc_handler   = proc_dointvec,
1671         },
1672 #endif
1673 #ifdef CONFIG_AIO
1674         {
1675                 .procname       = "aio-nr",
1676                 .data           = &aio_nr,
1677                 .maxlen         = sizeof(aio_nr),
1678                 .mode           = 0444,
1679                 .proc_handler   = proc_doulongvec_minmax,
1680         },
1681         {
1682                 .procname       = "aio-max-nr",
1683                 .data           = &aio_max_nr,
1684                 .maxlen         = sizeof(aio_max_nr),
1685                 .mode           = 0644,
1686                 .proc_handler   = proc_doulongvec_minmax,
1687         },
1688 #endif /* CONFIG_AIO */
1689 #ifdef CONFIG_INOTIFY_USER
1690         {
1691                 .procname       = "inotify",
1692                 .mode           = 0555,
1693                 .child          = inotify_table,
1694         },
1695 #endif
1696 #ifdef CONFIG_EPOLL
1697         {
1698                 .procname       = "epoll",
1699                 .mode           = 0555,
1700                 .child          = epoll_table,
1701         },
1702 #endif
1703 #endif
1704         {
1705                 .procname       = "protected_symlinks",
1706                 .data           = &sysctl_protected_symlinks,
1707                 .maxlen         = sizeof(int),
1708                 .mode           = 0600,
1709                 .proc_handler   = proc_dointvec_minmax,
1710                 .extra1         = &zero,
1711                 .extra2         = &one,
1712         },
1713         {
1714                 .procname       = "protected_hardlinks",
1715                 .data           = &sysctl_protected_hardlinks,
1716                 .maxlen         = sizeof(int),
1717                 .mode           = 0600,
1718                 .proc_handler   = proc_dointvec_minmax,
1719                 .extra1         = &zero,
1720                 .extra2         = &one,
1721         },
1722         {
1723                 .procname       = "protected_fifos",
1724                 .data           = &sysctl_protected_fifos,
1725                 .maxlen         = sizeof(int),
1726                 .mode           = 0600,
1727                 .proc_handler   = proc_dointvec_minmax,
1728                 .extra1         = &zero,
1729                 .extra2         = &two,
1730         },
1731         {
1732                 .procname       = "protected_regular",
1733                 .data           = &sysctl_protected_regular,
1734                 .maxlen         = sizeof(int),
1735                 .mode           = 0600,
1736                 .proc_handler   = proc_dointvec_minmax,
1737                 .extra1         = &zero,
1738                 .extra2         = &two,
1739         },
1740         {
1741                 .procname       = "suid_dumpable",
1742                 .data           = &suid_dumpable,
1743                 .maxlen         = sizeof(int),
1744                 .mode           = 0644,
1745                 .proc_handler   = proc_dointvec_minmax_coredump,
1746                 .extra1         = &zero,
1747                 .extra2         = &two,
1748         },
1749 #if defined(CONFIG_BINFMT_MISC) || defined(CONFIG_BINFMT_MISC_MODULE)
1750         {
1751                 .procname       = "binfmt_misc",
1752                 .mode           = 0555,
1753                 .child          = sysctl_mount_point,
1754         },
1755 #endif
1756         {
1757                 .procname       = "pipe-max-size",
1758                 .data           = &pipe_max_size,
1759                 .maxlen         = sizeof(int),
1760                 .mode           = 0644,
1761                 .proc_handler   = &pipe_proc_fn,
1762                 .extra1         = &pipe_min_size,
1763         },
1764         {
1765                 .procname       = "pipe-user-pages-hard",
1766                 .data           = &pipe_user_pages_hard,
1767                 .maxlen         = sizeof(pipe_user_pages_hard),
1768                 .mode           = 0644,
1769                 .proc_handler   = proc_doulongvec_minmax,
1770         },
1771         {
1772                 .procname       = "pipe-user-pages-soft",
1773                 .data           = &pipe_user_pages_soft,
1774                 .maxlen         = sizeof(pipe_user_pages_soft),
1775                 .mode           = 0644,
1776                 .proc_handler   = proc_doulongvec_minmax,
1777         },
1778         {
1779                 .procname       = "mount-max",
1780                 .data           = &sysctl_mount_max,
1781                 .maxlen         = sizeof(unsigned int),
1782                 .mode           = 0644,
1783                 .proc_handler   = proc_dointvec_minmax,
1784                 .extra1         = &one,
1785         },
1786         { }
1787 };
1788
1789 static struct ctl_table debug_table[] = {
1790 #ifdef CONFIG_SYSCTL_EXCEPTION_TRACE
1791         {
1792                 .procname       = "exception-trace",
1793                 .data           = &show_unhandled_signals,
1794                 .maxlen         = sizeof(int),
1795                 .mode           = 0644,
1796                 .proc_handler   = proc_dointvec
1797         },
1798 #endif
1799 #if defined(CONFIG_OPTPROBES)
1800         {
1801                 .procname       = "kprobes-optimization",
1802                 .data           = &sysctl_kprobes_optimization,
1803                 .maxlen         = sizeof(int),
1804                 .mode           = 0644,
1805                 .proc_handler   = proc_kprobes_optimization_handler,
1806                 .extra1         = &zero,
1807                 .extra2         = &one,
1808         },
1809 #endif
1810         { }
1811 };
1812
1813 static struct ctl_table dev_table[] = {
1814         { }
1815 };
1816
1817 int __init sysctl_init(void)
1818 {
1819         struct ctl_table_header *hdr;
1820
1821         hdr = register_sysctl_table(sysctl_base_table);
1822         kmemleak_not_leak(hdr);
1823         return 0;
1824 }
1825
1826 #endif /* CONFIG_SYSCTL */
1827
1828 /*
1829  * /proc/sys support
1830  */
1831
1832 #ifdef CONFIG_PROC_SYSCTL
1833
1834 static int _proc_do_string(char *data, int maxlen, int write,
1835                            char __user *buffer,
1836                            size_t *lenp, loff_t *ppos)
1837 {
1838         size_t len;
1839         char __user *p;
1840         char c;
1841
1842         if (!data || !maxlen || !*lenp) {
1843                 *lenp = 0;
1844                 return 0;
1845         }
1846
1847         if (write) {
1848                 if (sysctl_writes_strict == SYSCTL_WRITES_STRICT) {
1849                         /* Only continue writes not past the end of buffer. */
1850                         len = strlen(data);
1851                         if (len > maxlen - 1)
1852                                 len = maxlen - 1;
1853
1854                         if (*ppos > len)
1855                                 return 0;
1856                         len = *ppos;
1857                 } else {
1858                         /* Start writing from beginning of buffer. */
1859                         len = 0;
1860                 }
1861
1862                 *ppos += *lenp;
1863                 p = buffer;
1864                 while ((p - buffer) < *lenp && len < maxlen - 1) {
1865                         if (get_user(c, p++))
1866                                 return -EFAULT;
1867                         if (c == 0 || c == '\n')
1868                                 break;
1869                         data[len++] = c;
1870                 }
1871                 data[len] = 0;
1872         } else {
1873                 len = strlen(data);
1874                 if (len > maxlen)
1875                         len = maxlen;
1876
1877                 if (*ppos > len) {
1878                         *lenp = 0;
1879                         return 0;
1880                 }
1881
1882                 data += *ppos;
1883                 len  -= *ppos;
1884
1885                 if (len > *lenp)
1886                         len = *lenp;
1887                 if (len)
1888                         if (copy_to_user(buffer, data, len))
1889                                 return -EFAULT;
1890                 if (len < *lenp) {
1891                         if (put_user('\n', buffer + len))
1892                                 return -EFAULT;
1893                         len++;
1894                 }
1895                 *lenp = len;
1896                 *ppos += len;
1897         }
1898         return 0;
1899 }
1900
1901 static void warn_sysctl_write(struct ctl_table *table)
1902 {
1903         pr_warn_once("%s wrote to %s when file position was not 0!\n"
1904                 "This will not be supported in the future. To silence this\n"
1905                 "warning, set kernel.sysctl_writes_strict = -1\n",
1906                 current->comm, table->procname);
1907 }
1908
1909 /**
1910  * proc_dostring - read a string sysctl
1911  * @table: the sysctl table
1912  * @write: %TRUE if this is a write to the sysctl file
1913  * @buffer: the user buffer
1914  * @lenp: the size of the user buffer
1915  * @ppos: file position
1916  *
1917  * Reads/writes a string from/to the user buffer. If the kernel
1918  * buffer provided is not large enough to hold the string, the
1919  * string is truncated. The copied string is %NULL-terminated.
1920  * If the string is being read by the user process, it is copied
1921  * and a newline '\n' is added. It is truncated if the buffer is
1922  * not large enough.
1923  *
1924  * Returns 0 on success.
1925  */
1926 int proc_dostring(struct ctl_table *table, int write,
1927                   void __user *buffer, size_t *lenp, loff_t *ppos)
1928 {
1929         if (write && *ppos && sysctl_writes_strict == SYSCTL_WRITES_WARN)
1930                 warn_sysctl_write(table);
1931
1932         return _proc_do_string((char *)(table->data), table->maxlen, write,
1933                                (char __user *)buffer, lenp, ppos);
1934 }
1935
1936 static size_t proc_skip_spaces(char **buf)
1937 {
1938         size_t ret;
1939         char *tmp = skip_spaces(*buf);
1940         ret = tmp - *buf;
1941         *buf = tmp;
1942         return ret;
1943 }
1944
1945 static void proc_skip_char(char **buf, size_t *size, const char v)
1946 {
1947         while (*size) {
1948                 if (**buf != v)
1949                         break;
1950                 (*size)--;
1951                 (*buf)++;
1952         }
1953 }
1954
1955 #define TMPBUFLEN 22
1956 /**
1957  * proc_get_long - reads an ASCII formatted integer from a user buffer
1958  *
1959  * @buf: a kernel buffer
1960  * @size: size of the kernel buffer
1961  * @val: this is where the number will be stored
1962  * @neg: set to %TRUE if number is negative
1963  * @perm_tr: a vector which contains the allowed trailers
1964  * @perm_tr_len: size of the perm_tr vector
1965  * @tr: pointer to store the trailer character
1966  *
1967  * In case of success %0 is returned and @buf and @size are updated with
1968  * the amount of bytes read. If @tr is non-NULL and a trailing
1969  * character exists (size is non-zero after returning from this
1970  * function), @tr is updated with the trailing character.
1971  */
1972 static int proc_get_long(char **buf, size_t *size,
1973                           unsigned long *val, bool *neg,
1974                           const char *perm_tr, unsigned perm_tr_len, char *tr)
1975 {
1976         int len;
1977         char *p, tmp[TMPBUFLEN];
1978
1979         if (!*size)
1980                 return -EINVAL;
1981
1982         len = *size;
1983         if (len > TMPBUFLEN - 1)
1984                 len = TMPBUFLEN - 1;
1985
1986         memcpy(tmp, *buf, len);
1987
1988         tmp[len] = 0;
1989         p = tmp;
1990         if (*p == '-' && *size > 1) {
1991                 *neg = true;
1992                 p++;
1993         } else
1994                 *neg = false;
1995         if (!isdigit(*p))
1996                 return -EINVAL;
1997
1998         *val = simple_strtoul(p, &p, 0);
1999
2000         len = p - tmp;
2001
2002         /* We don't know if the next char is whitespace thus we may accept
2003          * invalid integers (e.g. 1234...a) or two integers instead of one
2004          * (e.g. 123...1). So lets not allow such large numbers. */
2005         if (len == TMPBUFLEN - 1)
2006                 return -EINVAL;
2007
2008         if (len < *size && perm_tr_len && !memchr(perm_tr, *p, perm_tr_len))
2009                 return -EINVAL;
2010
2011         if (tr && (len < *size))
2012                 *tr = *p;
2013
2014         *buf += len;
2015         *size -= len;
2016
2017         return 0;
2018 }
2019
2020 /**
2021  * proc_put_long - converts an integer to a decimal ASCII formatted string
2022  *
2023  * @buf: the user buffer
2024  * @size: the size of the user buffer
2025  * @val: the integer to be converted
2026  * @neg: sign of the number, %TRUE for negative
2027  *
2028  * In case of success %0 is returned and @buf and @size are updated with
2029  * the amount of bytes written.
2030  */
2031 static int proc_put_long(void __user **buf, size_t *size, unsigned long val,
2032                           bool neg)
2033 {
2034         int len;
2035         char tmp[TMPBUFLEN], *p = tmp;
2036
2037         sprintf(p, "%s%lu", neg ? "-" : "", val);
2038         len = strlen(tmp);
2039         if (len > *size)
2040                 len = *size;
2041         if (copy_to_user(*buf, tmp, len))
2042                 return -EFAULT;
2043         *size -= len;
2044         *buf += len;
2045         return 0;
2046 }
2047 #undef TMPBUFLEN
2048
2049 static int proc_put_char(void __user **buf, size_t *size, char c)
2050 {
2051         if (*size) {
2052                 char __user **buffer = (char __user **)buf;
2053                 if (put_user(c, *buffer))
2054                         return -EFAULT;
2055                 (*size)--, (*buffer)++;
2056                 *buf = *buffer;
2057         }
2058         return 0;
2059 }
2060
2061 static int do_proc_dointvec_conv(bool *negp, unsigned long *lvalp,
2062                                  int *valp,
2063                                  int write, void *data)
2064 {
2065         if (write) {
2066                 if (*negp) {
2067                         if (*lvalp > (unsigned long) INT_MAX + 1)
2068                                 return -EINVAL;
2069                         *valp = -*lvalp;
2070                 } else {
2071                         if (*lvalp > (unsigned long) INT_MAX)
2072                                 return -EINVAL;
2073                         *valp = *lvalp;
2074                 }
2075         } else {
2076                 int val = *valp;
2077                 if (val < 0) {
2078                         *negp = true;
2079                         *lvalp = -(unsigned long)val;
2080                 } else {
2081                         *negp = false;
2082                         *lvalp = (unsigned long)val;
2083                 }
2084         }
2085         return 0;
2086 }
2087
2088 static int do_proc_douintvec_conv(bool *negp, unsigned long *lvalp,
2089                                  int *valp,
2090                                  int write, void *data)
2091 {
2092         if (write) {
2093                 if (*negp)
2094                         return -EINVAL;
2095                 if (*lvalp > UINT_MAX)
2096                         return -EINVAL;
2097                 *valp = *lvalp;
2098         } else {
2099                 unsigned int val = *valp;
2100                 *negp = false;
2101                 *lvalp = (unsigned long)val;
2102         }
2103         return 0;
2104 }
2105
2106 static const char proc_wspace_sep[] = { ' ', '\t', '\n' };
2107
2108 static int __do_proc_dointvec(void *tbl_data, struct ctl_table *table,
2109                   int write, void __user *buffer,
2110                   size_t *lenp, loff_t *ppos,
2111                   int (*conv)(bool *negp, unsigned long *lvalp, int *valp,
2112                               int write, void *data),
2113                   void *data)
2114 {
2115         int *i, vleft, first = 1, err = 0;
2116         unsigned long page = 0;
2117         size_t left;
2118         char *kbuf;
2119
2120         if (!tbl_data || !table->maxlen || !*lenp || (*ppos && !write)) {
2121                 *lenp = 0;
2122                 return 0;
2123         }
2124
2125         i = (int *) tbl_data;
2126         vleft = table->maxlen / sizeof(*i);
2127         left = *lenp;
2128
2129         if (!conv)
2130                 conv = do_proc_dointvec_conv;
2131
2132         if (write) {
2133                 if (*ppos) {
2134                         switch (sysctl_writes_strict) {
2135                         case SYSCTL_WRITES_STRICT:
2136                                 goto out;
2137                         case SYSCTL_WRITES_WARN:
2138                                 warn_sysctl_write(table);
2139                                 break;
2140                         default:
2141                                 break;
2142                         }
2143                 }
2144
2145                 if (left > PAGE_SIZE - 1)
2146                         left = PAGE_SIZE - 1;
2147                 page = __get_free_page(GFP_TEMPORARY);
2148                 kbuf = (char *) page;
2149                 if (!kbuf)
2150                         return -ENOMEM;
2151                 if (copy_from_user(kbuf, buffer, left)) {
2152                         err = -EFAULT;
2153                         goto free;
2154                 }
2155                 kbuf[left] = 0;
2156         }
2157
2158         for (; left && vleft--; i++, first=0) {
2159                 unsigned long lval;
2160                 bool neg;
2161
2162                 if (write) {
2163                         left -= proc_skip_spaces(&kbuf);
2164
2165                         if (!left)
2166                                 break;
2167                         err = proc_get_long(&kbuf, &left, &lval, &neg,
2168                                              proc_wspace_sep,
2169                                              sizeof(proc_wspace_sep), NULL);
2170                         if (err)
2171                                 break;
2172                         if (conv(&neg, &lval, i, 1, data)) {
2173                                 err = -EINVAL;
2174                                 break;
2175                         }
2176                 } else {
2177                         if (conv(&neg, &lval, i, 0, data)) {
2178                                 err = -EINVAL;
2179                                 break;
2180                         }
2181                         if (!first)
2182                                 err = proc_put_char(&buffer, &left, '\t');
2183                         if (err)
2184                                 break;
2185                         err = proc_put_long(&buffer, &left, lval, neg);
2186                         if (err)
2187                                 break;
2188                 }
2189         }
2190
2191         if (!write && !first && left && !err)
2192                 err = proc_put_char(&buffer, &left, '\n');
2193         if (write && !err && left)
2194                 left -= proc_skip_spaces(&kbuf);
2195 free:
2196         if (write) {
2197                 free_page(page);
2198                 if (first)
2199                         return err ? : -EINVAL;
2200         }
2201         *lenp -= left;
2202 out:
2203         *ppos += *lenp;
2204         return err;
2205 }
2206
2207 static int do_proc_dointvec(struct ctl_table *table, int write,
2208                   void __user *buffer, size_t *lenp, loff_t *ppos,
2209                   int (*conv)(bool *negp, unsigned long *lvalp, int *valp,
2210                               int write, void *data),
2211                   void *data)
2212 {
2213         return __do_proc_dointvec(table->data, table, write,
2214                         buffer, lenp, ppos, conv, data);
2215 }
2216
2217 /**
2218  * proc_dointvec - read a vector of integers
2219  * @table: the sysctl table
2220  * @write: %TRUE if this is a write to the sysctl file
2221  * @buffer: the user buffer
2222  * @lenp: the size of the user buffer
2223  * @ppos: file position
2224  *
2225  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2226  * values from/to the user buffer, treated as an ASCII string.
2227  *
2228  * Returns 0 on success.
2229  */
2230 int proc_dointvec(struct ctl_table *table, int write,
2231                      void __user *buffer, size_t *lenp, loff_t *ppos)
2232 {
2233         return do_proc_dointvec(table, write, buffer, lenp, ppos, NULL, NULL);
2234 }
2235
2236 /**
2237  * proc_douintvec - read a vector of unsigned integers
2238  * @table: the sysctl table
2239  * @write: %TRUE if this is a write to the sysctl file
2240  * @buffer: the user buffer
2241  * @lenp: the size of the user buffer
2242  * @ppos: file position
2243  *
2244  * Reads/writes up to table->maxlen/sizeof(unsigned int) unsigned integer
2245  * values from/to the user buffer, treated as an ASCII string.
2246  *
2247  * Returns 0 on success.
2248  */
2249 int proc_douintvec(struct ctl_table *table, int write,
2250                      void __user *buffer, size_t *lenp, loff_t *ppos)
2251 {
2252         return do_proc_dointvec(table, write, buffer, lenp, ppos,
2253                                 do_proc_douintvec_conv, NULL);
2254 }
2255
2256 /*
2257  * Taint values can only be increased
2258  * This means we can safely use a temporary.
2259  */
2260 static int proc_taint(struct ctl_table *table, int write,
2261                                void __user *buffer, size_t *lenp, loff_t *ppos)
2262 {
2263         struct ctl_table t;
2264         unsigned long tmptaint = get_taint();
2265         int err;
2266
2267         if (write && !capable(CAP_SYS_ADMIN))
2268                 return -EPERM;
2269
2270         t = *table;
2271         t.data = &tmptaint;
2272         err = proc_doulongvec_minmax(&t, write, buffer, lenp, ppos);
2273         if (err < 0)
2274                 return err;
2275
2276         if (write) {
2277                 /*
2278                  * Poor man's atomic or. Not worth adding a primitive
2279                  * to everyone's atomic.h for this
2280                  */
2281                 int i;
2282                 for (i = 0; i < BITS_PER_LONG && tmptaint >> i; i++) {
2283                         if ((tmptaint >> i) & 1)
2284                                 add_taint(i, LOCKDEP_STILL_OK);
2285                 }
2286         }
2287
2288         return err;
2289 }
2290
2291 #ifdef CONFIG_PRINTK
2292 static int proc_dointvec_minmax_sysadmin(struct ctl_table *table, int write,
2293                                 void __user *buffer, size_t *lenp, loff_t *ppos)
2294 {
2295         if (write && !capable(CAP_SYS_ADMIN))
2296                 return -EPERM;
2297
2298         return proc_dointvec_minmax(table, write, buffer, lenp, ppos);
2299 }
2300 #endif
2301
2302 struct do_proc_dointvec_minmax_conv_param {
2303         int *min;
2304         int *max;
2305 };
2306
2307 static int do_proc_dointvec_minmax_conv(bool *negp, unsigned long *lvalp,
2308                                         int *valp,
2309                                         int write, void *data)
2310 {
2311         struct do_proc_dointvec_minmax_conv_param *param = data;
2312         if (write) {
2313                 int val;
2314                 if (*negp) {
2315                         if (*lvalp > (unsigned long) INT_MAX + 1)
2316                                 return -EINVAL;
2317                         val = -*lvalp;
2318                 } else {
2319                         if (*lvalp > (unsigned long) INT_MAX)
2320                                 return -EINVAL;
2321                         val = *lvalp;
2322                 }
2323                 if ((param->min && *param->min > val) ||
2324                     (param->max && *param->max < val))
2325                         return -EINVAL;
2326                 *valp = val;
2327         } else {
2328                 int val = *valp;
2329                 if (val < 0) {
2330                         *negp = true;
2331                         *lvalp = -(unsigned long)val;
2332                 } else {
2333                         *negp = false;
2334                         *lvalp = (unsigned long)val;
2335                 }
2336         }
2337         return 0;
2338 }
2339
2340 /**
2341  * proc_dointvec_minmax - read a vector of integers with min/max values
2342  * @table: the sysctl table
2343  * @write: %TRUE if this is a write to the sysctl file
2344  * @buffer: the user buffer
2345  * @lenp: the size of the user buffer
2346  * @ppos: file position
2347  *
2348  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2349  * values from/to the user buffer, treated as an ASCII string.
2350  *
2351  * This routine will ensure the values are within the range specified by
2352  * table->extra1 (min) and table->extra2 (max).
2353  *
2354  * Returns 0 on success.
2355  */
2356 int proc_dointvec_minmax(struct ctl_table *table, int write,
2357                   void __user *buffer, size_t *lenp, loff_t *ppos)
2358 {
2359         struct do_proc_dointvec_minmax_conv_param param = {
2360                 .min = (int *) table->extra1,
2361                 .max = (int *) table->extra2,
2362         };
2363         return do_proc_dointvec(table, write, buffer, lenp, ppos,
2364                                 do_proc_dointvec_minmax_conv, &param);
2365 }
2366
2367 static void validate_coredump_safety(void)
2368 {
2369 #ifdef CONFIG_COREDUMP
2370         if (suid_dumpable == SUID_DUMP_ROOT &&
2371             core_pattern[0] != '/' && core_pattern[0] != '|') {
2372                 printk(KERN_WARNING "Unsafe core_pattern used with "\
2373                         "suid_dumpable=2. Pipe handler or fully qualified "\
2374                         "core dump path required.\n");
2375         }
2376 #endif
2377 }
2378
2379 static int proc_dointvec_minmax_coredump(struct ctl_table *table, int write,
2380                 void __user *buffer, size_t *lenp, loff_t *ppos)
2381 {
2382         int error = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
2383         if (!error)
2384                 validate_coredump_safety();
2385         return error;
2386 }
2387
2388 #ifdef CONFIG_COREDUMP
2389 static int proc_dostring_coredump(struct ctl_table *table, int write,
2390                   void __user *buffer, size_t *lenp, loff_t *ppos)
2391 {
2392         int error = proc_dostring(table, write, buffer, lenp, ppos);
2393         if (!error)
2394                 validate_coredump_safety();
2395         return error;
2396 }
2397 #endif
2398
2399 static int __do_proc_doulongvec_minmax(void *data, struct ctl_table *table, int write,
2400                                      void __user *buffer,
2401                                      size_t *lenp, loff_t *ppos,
2402                                      unsigned long convmul,
2403                                      unsigned long convdiv)
2404 {
2405         unsigned long *i, *min, *max;
2406         int vleft, first = 1, err = 0;
2407         unsigned long page = 0;
2408         size_t left;
2409         char *kbuf;
2410
2411         if (!data || !table->maxlen || !*lenp || (*ppos && !write)) {
2412                 *lenp = 0;
2413                 return 0;
2414         }
2415
2416         i = (unsigned long *) data;
2417         min = (unsigned long *) table->extra1;
2418         max = (unsigned long *) table->extra2;
2419         vleft = table->maxlen / sizeof(unsigned long);
2420         left = *lenp;
2421
2422         if (write) {
2423                 if (*ppos) {
2424                         switch (sysctl_writes_strict) {
2425                         case SYSCTL_WRITES_STRICT:
2426                                 goto out;
2427                         case SYSCTL_WRITES_WARN:
2428                                 warn_sysctl_write(table);
2429                                 break;
2430                         default:
2431                                 break;
2432                         }
2433                 }
2434
2435                 if (left > PAGE_SIZE - 1)
2436                         left = PAGE_SIZE - 1;
2437                 page = __get_free_page(GFP_TEMPORARY);
2438                 kbuf = (char *) page;
2439                 if (!kbuf)
2440                         return -ENOMEM;
2441                 if (copy_from_user(kbuf, buffer, left)) {
2442                         err = -EFAULT;
2443                         goto free;
2444                 }
2445                 kbuf[left] = 0;
2446         }
2447
2448         for (; left && vleft--; i++, first = 0) {
2449                 unsigned long val;
2450
2451                 if (write) {
2452                         bool neg;
2453
2454                         left -= proc_skip_spaces(&kbuf);
2455
2456                         err = proc_get_long(&kbuf, &left, &val, &neg,
2457                                              proc_wspace_sep,
2458                                              sizeof(proc_wspace_sep), NULL);
2459                         if (err)
2460                                 break;
2461                         if (neg)
2462                                 continue;
2463                         val = convmul * val / convdiv;
2464                         if ((min && val < *min) || (max && val > *max)) {
2465                                 err = -EINVAL;
2466                                 break;
2467                         }
2468                         *i = val;
2469                 } else {
2470                         val = convdiv * (*i) / convmul;
2471                         if (!first) {
2472                                 err = proc_put_char(&buffer, &left, '\t');
2473                                 if (err)
2474                                         break;
2475                         }
2476                         err = proc_put_long(&buffer, &left, val, false);
2477                         if (err)
2478                                 break;
2479                 }
2480         }
2481
2482         if (!write && !first && left && !err)
2483                 err = proc_put_char(&buffer, &left, '\n');
2484         if (write && !err)
2485                 left -= proc_skip_spaces(&kbuf);
2486 free:
2487         if (write) {
2488                 free_page(page);
2489                 if (first)
2490                         return err ? : -EINVAL;
2491         }
2492         *lenp -= left;
2493 out:
2494         *ppos += *lenp;
2495         return err;
2496 }
2497
2498 static int do_proc_doulongvec_minmax(struct ctl_table *table, int write,
2499                                      void __user *buffer,
2500                                      size_t *lenp, loff_t *ppos,
2501                                      unsigned long convmul,
2502                                      unsigned long convdiv)
2503 {
2504         return __do_proc_doulongvec_minmax(table->data, table, write,
2505                         buffer, lenp, ppos, convmul, convdiv);
2506 }
2507
2508 /**
2509  * proc_doulongvec_minmax - read a vector of long integers with min/max values
2510  * @table: the sysctl table
2511  * @write: %TRUE if this is a write to the sysctl file
2512  * @buffer: the user buffer
2513  * @lenp: the size of the user buffer
2514  * @ppos: file position
2515  *
2516  * Reads/writes up to table->maxlen/sizeof(unsigned long) unsigned long
2517  * values from/to the user buffer, treated as an ASCII string.
2518  *
2519  * This routine will ensure the values are within the range specified by
2520  * table->extra1 (min) and table->extra2 (max).
2521  *
2522  * Returns 0 on success.
2523  */
2524 int proc_doulongvec_minmax(struct ctl_table *table, int write,
2525                            void __user *buffer, size_t *lenp, loff_t *ppos)
2526 {
2527     return do_proc_doulongvec_minmax(table, write, buffer, lenp, ppos, 1l, 1l);
2528 }
2529
2530 /**
2531  * proc_doulongvec_ms_jiffies_minmax - read a vector of millisecond values with min/max values
2532  * @table: the sysctl table
2533  * @write: %TRUE if this is a write to the sysctl file
2534  * @buffer: the user buffer
2535  * @lenp: the size of the user buffer
2536  * @ppos: file position
2537  *
2538  * Reads/writes up to table->maxlen/sizeof(unsigned long) unsigned long
2539  * values from/to the user buffer, treated as an ASCII string. The values
2540  * are treated as milliseconds, and converted to jiffies when they are stored.
2541  *
2542  * This routine will ensure the values are within the range specified by
2543  * table->extra1 (min) and table->extra2 (max).
2544  *
2545  * Returns 0 on success.
2546  */
2547 int proc_doulongvec_ms_jiffies_minmax(struct ctl_table *table, int write,
2548                                       void __user *buffer,
2549                                       size_t *lenp, loff_t *ppos)
2550 {
2551     return do_proc_doulongvec_minmax(table, write, buffer,
2552                                      lenp, ppos, HZ, 1000l);
2553 }
2554
2555
2556 static int do_proc_dointvec_jiffies_conv(bool *negp, unsigned long *lvalp,
2557                                          int *valp,
2558                                          int write, void *data)
2559 {
2560         if (write) {
2561                 if (*lvalp > LONG_MAX / HZ)
2562                         return 1;
2563                 *valp = *negp ? -(*lvalp*HZ) : (*lvalp*HZ);
2564         } else {
2565                 int val = *valp;
2566                 unsigned long lval;
2567                 if (val < 0) {
2568                         *negp = true;
2569                         lval = -(unsigned long)val;
2570                 } else {
2571                         *negp = false;
2572                         lval = (unsigned long)val;
2573                 }
2574                 *lvalp = lval / HZ;
2575         }
2576         return 0;
2577 }
2578
2579 static int do_proc_dointvec_userhz_jiffies_conv(bool *negp, unsigned long *lvalp,
2580                                                 int *valp,
2581                                                 int write, void *data)
2582 {
2583         if (write) {
2584                 if (USER_HZ < HZ && *lvalp > (LONG_MAX / HZ) * USER_HZ)
2585                         return 1;
2586                 *valp = clock_t_to_jiffies(*negp ? -*lvalp : *lvalp);
2587         } else {
2588                 int val = *valp;
2589                 unsigned long lval;
2590                 if (val < 0) {
2591                         *negp = true;
2592                         lval = -(unsigned long)val;
2593                 } else {
2594                         *negp = false;
2595                         lval = (unsigned long)val;
2596                 }
2597                 *lvalp = jiffies_to_clock_t(lval);
2598         }
2599         return 0;
2600 }
2601
2602 static int do_proc_dointvec_ms_jiffies_conv(bool *negp, unsigned long *lvalp,
2603                                             int *valp,
2604                                             int write, void *data)
2605 {
2606         if (write) {
2607                 unsigned long jif = msecs_to_jiffies(*negp ? -*lvalp : *lvalp);
2608
2609                 if (jif > INT_MAX)
2610                         return 1;
2611                 *valp = (int)jif;
2612         } else {
2613                 int val = *valp;
2614                 unsigned long lval;
2615                 if (val < 0) {
2616                         *negp = true;
2617                         lval = -(unsigned long)val;
2618                 } else {
2619                         *negp = false;
2620                         lval = (unsigned long)val;
2621                 }
2622                 *lvalp = jiffies_to_msecs(lval);
2623         }
2624         return 0;
2625 }
2626
2627 /**
2628  * proc_dointvec_jiffies - read a vector of integers as seconds
2629  * @table: the sysctl table
2630  * @write: %TRUE if this is a write to the sysctl file
2631  * @buffer: the user buffer
2632  * @lenp: the size of the user buffer
2633  * @ppos: file position
2634  *
2635  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2636  * values from/to the user buffer, treated as an ASCII string.
2637  * The values read are assumed to be in seconds, and are converted into
2638  * jiffies.
2639  *
2640  * Returns 0 on success.
2641  */
2642 int proc_dointvec_jiffies(struct ctl_table *table, int write,
2643                           void __user *buffer, size_t *lenp, loff_t *ppos)
2644 {
2645     return do_proc_dointvec(table,write,buffer,lenp,ppos,
2646                             do_proc_dointvec_jiffies_conv,NULL);
2647 }
2648
2649 /**
2650  * proc_dointvec_userhz_jiffies - read a vector of integers as 1/USER_HZ seconds
2651  * @table: the sysctl table
2652  * @write: %TRUE if this is a write to the sysctl file
2653  * @buffer: the user buffer
2654  * @lenp: the size of the user buffer
2655  * @ppos: pointer to the file position
2656  *
2657  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2658  * values from/to the user buffer, treated as an ASCII string.
2659  * The values read are assumed to be in 1/USER_HZ seconds, and
2660  * are converted into jiffies.
2661  *
2662  * Returns 0 on success.
2663  */
2664 int proc_dointvec_userhz_jiffies(struct ctl_table *table, int write,
2665                                  void __user *buffer, size_t *lenp, loff_t *ppos)
2666 {
2667     return do_proc_dointvec(table,write,buffer,lenp,ppos,
2668                             do_proc_dointvec_userhz_jiffies_conv,NULL);
2669 }
2670
2671 /**
2672  * proc_dointvec_ms_jiffies - read a vector of integers as 1 milliseconds
2673  * @table: the sysctl table
2674  * @write: %TRUE if this is a write to the sysctl file
2675  * @buffer: the user buffer
2676  * @lenp: the size of the user buffer
2677  * @ppos: file position
2678  * @ppos: the current position in the file
2679  *
2680  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2681  * values from/to the user buffer, treated as an ASCII string.
2682  * The values read are assumed to be in 1/1000 seconds, and
2683  * are converted into jiffies.
2684  *
2685  * Returns 0 on success.
2686  */
2687 int proc_dointvec_ms_jiffies(struct ctl_table *table, int write,
2688                              void __user *buffer, size_t *lenp, loff_t *ppos)
2689 {
2690         return do_proc_dointvec(table, write, buffer, lenp, ppos,
2691                                 do_proc_dointvec_ms_jiffies_conv, NULL);
2692 }
2693
2694 static int proc_do_cad_pid(struct ctl_table *table, int write,
2695                            void __user *buffer, size_t *lenp, loff_t *ppos)
2696 {
2697         struct pid *new_pid;
2698         pid_t tmp;
2699         int r;
2700
2701         tmp = pid_vnr(cad_pid);
2702
2703         r = __do_proc_dointvec(&tmp, table, write, buffer,
2704                                lenp, ppos, NULL, NULL);
2705         if (r || !write)
2706                 return r;
2707
2708         new_pid = find_get_pid(tmp);
2709         if (!new_pid)
2710                 return -ESRCH;
2711
2712         put_pid(xchg(&cad_pid, new_pid));
2713         return 0;
2714 }
2715
2716 /**
2717  * proc_do_large_bitmap - read/write from/to a large bitmap
2718  * @table: the sysctl table
2719  * @write: %TRUE if this is a write to the sysctl file
2720  * @buffer: the user buffer
2721  * @lenp: the size of the user buffer
2722  * @ppos: file position
2723  *
2724  * The bitmap is stored at table->data and the bitmap length (in bits)
2725  * in table->maxlen.
2726  *
2727  * We use a range comma separated format (e.g. 1,3-4,10-10) so that
2728  * large bitmaps may be represented in a compact manner. Writing into
2729  * the file will clear the bitmap then update it with the given input.
2730  *
2731  * Returns 0 on success.
2732  */
2733 int proc_do_large_bitmap(struct ctl_table *table, int write,
2734                          void __user *buffer, size_t *lenp, loff_t *ppos)
2735 {
2736         int err = 0;
2737         bool first = 1;
2738         size_t left = *lenp;
2739         unsigned long bitmap_len = table->maxlen;
2740         unsigned long *bitmap = *(unsigned long **) table->data;
2741         unsigned long *tmp_bitmap = NULL;
2742         char tr_a[] = { '-', ',', '\n' }, tr_b[] = { ',', '\n', 0 }, c;
2743
2744         if (!bitmap || !bitmap_len || !left || (*ppos && !write)) {
2745                 *lenp = 0;
2746                 return 0;
2747         }
2748
2749         if (write) {
2750                 unsigned long page = 0;
2751                 char *kbuf;
2752
2753                 if (left > PAGE_SIZE - 1)
2754                         left = PAGE_SIZE - 1;
2755
2756                 page = __get_free_page(GFP_TEMPORARY);
2757                 kbuf = (char *) page;
2758                 if (!kbuf)
2759                         return -ENOMEM;
2760                 if (copy_from_user(kbuf, buffer, left)) {
2761                         free_page(page);
2762                         return -EFAULT;
2763                 }
2764                 kbuf[left] = 0;
2765
2766                 tmp_bitmap = kzalloc(BITS_TO_LONGS(bitmap_len) * sizeof(unsigned long),
2767                                      GFP_KERNEL);
2768                 if (!tmp_bitmap) {
2769                         free_page(page);
2770                         return -ENOMEM;
2771                 }
2772                 proc_skip_char(&kbuf, &left, '\n');
2773                 while (!err && left) {
2774                         unsigned long val_a, val_b;
2775                         bool neg;
2776
2777                         err = proc_get_long(&kbuf, &left, &val_a, &neg, tr_a,
2778                                              sizeof(tr_a), &c);
2779                         if (err)
2780                                 break;
2781                         if (val_a >= bitmap_len || neg) {
2782                                 err = -EINVAL;
2783                                 break;
2784                         }
2785
2786                         val_b = val_a;
2787                         if (left) {
2788                                 kbuf++;
2789                                 left--;
2790                         }
2791
2792                         if (c == '-') {
2793                                 err = proc_get_long(&kbuf, &left, &val_b,
2794                                                      &neg, tr_b, sizeof(tr_b),
2795                                                      &c);
2796                                 if (err)
2797                                         break;
2798                                 if (val_b >= bitmap_len || neg ||
2799                                     val_a > val_b) {
2800                                         err = -EINVAL;
2801                                         break;
2802                                 }
2803                                 if (left) {
2804                                         kbuf++;
2805                                         left--;
2806                                 }
2807                         }
2808
2809                         bitmap_set(tmp_bitmap, val_a, val_b - val_a + 1);
2810                         first = 0;
2811                         proc_skip_char(&kbuf, &left, '\n');
2812                 }
2813                 free_page(page);
2814         } else {
2815                 unsigned long bit_a, bit_b = 0;
2816
2817                 while (left) {
2818                         bit_a = find_next_bit(bitmap, bitmap_len, bit_b);
2819                         if (bit_a >= bitmap_len)
2820                                 break;
2821                         bit_b = find_next_zero_bit(bitmap, bitmap_len,
2822                                                    bit_a + 1) - 1;
2823
2824                         if (!first) {
2825                                 err = proc_put_char(&buffer, &left, ',');
2826                                 if (err)
2827                                         break;
2828                         }
2829                         err = proc_put_long(&buffer, &left, bit_a, false);
2830                         if (err)
2831                                 break;
2832                         if (bit_a != bit_b) {
2833                                 err = proc_put_char(&buffer, &left, '-');
2834                                 if (err)
2835                                         break;
2836                                 err = proc_put_long(&buffer, &left, bit_b, false);
2837                                 if (err)
2838                                         break;
2839                         }
2840
2841                         first = 0; bit_b++;
2842                 }
2843                 if (!err)
2844                         err = proc_put_char(&buffer, &left, '\n');
2845         }
2846
2847         if (!err) {
2848                 if (write) {
2849                         if (*ppos)
2850                                 bitmap_or(bitmap, bitmap, tmp_bitmap, bitmap_len);
2851                         else
2852                                 bitmap_copy(bitmap, tmp_bitmap, bitmap_len);
2853                 }
2854                 kfree(tmp_bitmap);
2855                 *lenp -= left;
2856                 *ppos += *lenp;
2857                 return 0;
2858         } else {
2859                 kfree(tmp_bitmap);
2860                 return err;
2861         }
2862 }
2863
2864 #else /* CONFIG_PROC_SYSCTL */
2865
2866 int proc_dostring(struct ctl_table *table, int write,
2867                   void __user *buffer, size_t *lenp, loff_t *ppos)
2868 {
2869         return -ENOSYS;
2870 }
2871
2872 int proc_dointvec(struct ctl_table *table, int write,
2873                   void __user *buffer, size_t *lenp, loff_t *ppos)
2874 {
2875         return -ENOSYS;
2876 }
2877
2878 int proc_douintvec(struct ctl_table *table, int write,
2879                   void __user *buffer, size_t *lenp, loff_t *ppos)
2880 {
2881         return -ENOSYS;
2882 }
2883
2884 int proc_dointvec_minmax(struct ctl_table *table, int write,
2885                     void __user *buffer, size_t *lenp, loff_t *ppos)
2886 {
2887         return -ENOSYS;
2888 }
2889
2890 int proc_dointvec_jiffies(struct ctl_table *table, int write,
2891                     void __user *buffer, size_t *lenp, loff_t *ppos)
2892 {
2893         return -ENOSYS;
2894 }
2895
2896 int proc_dointvec_userhz_jiffies(struct ctl_table *table, int write,
2897                     void __user *buffer, size_t *lenp, loff_t *ppos)
2898 {
2899         return -ENOSYS;
2900 }
2901
2902 int proc_dointvec_ms_jiffies(struct ctl_table *table, int write,
2903                              void __user *buffer, size_t *lenp, loff_t *ppos)
2904 {
2905         return -ENOSYS;
2906 }
2907
2908 int proc_doulongvec_minmax(struct ctl_table *table, int write,
2909                     void __user *buffer, size_t *lenp, loff_t *ppos)
2910 {
2911         return -ENOSYS;
2912 }
2913
2914 int proc_doulongvec_ms_jiffies_minmax(struct ctl_table *table, int write,
2915                                       void __user *buffer,
2916                                       size_t *lenp, loff_t *ppos)
2917 {
2918     return -ENOSYS;
2919 }
2920
2921
2922 #endif /* CONFIG_PROC_SYSCTL */
2923
2924 /*
2925  * No sense putting this after each symbol definition, twice,
2926  * exception granted :-)
2927  */
2928 EXPORT_SYMBOL(proc_dointvec);
2929 EXPORT_SYMBOL(proc_douintvec);
2930 EXPORT_SYMBOL(proc_dointvec_jiffies);
2931 EXPORT_SYMBOL(proc_dointvec_minmax);
2932 EXPORT_SYMBOL(proc_dointvec_userhz_jiffies);
2933 EXPORT_SYMBOL(proc_dointvec_ms_jiffies);
2934 EXPORT_SYMBOL(proc_dostring);
2935 EXPORT_SYMBOL(proc_doulongvec_minmax);
2936 EXPORT_SYMBOL(proc_doulongvec_ms_jiffies_minmax);