kernel/sysctl.c

   1 /*
   2  * sysctl.c: General linux system control interface
   3  *
   4  * Begun 24 March 1995, Stephen Tweedie
   5  * Added /proc support, Dec 1995
   6  * Added bdflush entry and intvec min/max checking, 2/23/96, Tom Dyas.
   7  * Added hooks for /proc/sys/net (minor, minor patch), 96/4/1, Mike Shaver.
   8  * Added kernel/java-{interpreter,appletviewer}, 96/5/10, Mike Shaver.
   9  * Dynamic registration fixes, Stephen Tweedie.
  10  * Added kswapd-interval, ctrl-alt-del, printk stuff, 1/8/97, Chris Horn.
  11  * Made sysctl support optional via CONFIG_SYSCTL, 1/10/97, Chris
  12  *  Horn.
  13  * Added proc_doulongvec_ms_jiffies_minmax, 09/08/99, Carlos H. Bauer.
  14  * Added proc_doulongvec_minmax, 09/08/99, Carlos H. Bauer.
  15  * Changed linked lists to use list.h instead of lists.h, 02/24/00, Bill
  16  *  Wendling.
  17  * The list_for_each() macro wasn't appropriate for the sysctl loop.
  18  *  Removed it and replaced it with older style, 03/23/00, Bill Wendling
  19  */
  20
  21 #include <linux/module.h>
  22 #include <linux/aio.h>
  23 #include <linux/mm.h>
  24 #include <linux/swap.h>
  25 #include <linux/slab.h>
  26 #include <linux/sysctl.h>
  27 #include <linux/bitmap.h>
  28 #include <linux/signal.h>
  29 #include <linux/printk.h>
  30 #include <linux/proc_fs.h>
  31 #include <linux/security.h>
  32 #include <linux/ctype.h>
  33 #include <linux/kmemcheck.h>
  34 #include <linux/kmemleak.h>
  35 #include <linux/fs.h>
  36 #include <linux/init.h>
  37 #include <linux/kernel.h>
  38 #include <linux/kobject.h>
  39 #include <linux/net.h>
  40 #include <linux/sysrq.h>
  41 #include <linux/highuid.h>
  42 #include <linux/writeback.h>
  43 #include <linux/ratelimit.h>
  44 #include <linux/compaction.h>
  45 #include <linux/hugetlb.h>
  46 #include <linux/initrd.h>
  47 #include <linux/key.h>
  48 #include <linux/times.h>
  49 #include <linux/limits.h>
  50 #include <linux/dcache.h>
  51 #include <linux/dnotify.h>
  52 #include <linux/syscalls.h>
  53 #include <linux/vmstat.h>
  54 #include <linux/nfs_fs.h>
  55 #include <linux/acpi.h>
  56 #include <linux/reboot.h>
  57 #include <linux/ftrace.h>
  58 #include <linux/perf_event.h>
  59 #include <linux/kprobes.h>
  60 #include <linux/pipe_fs_i.h>
  61 #include <linux/oom.h>
  62 #include <linux/kmod.h>
  63 #include <linux/capability.h>
  64 #include <linux/binfmts.h>
  65 #include <linux/sched/sysctl.h>
  66 #include <linux/kexec.h>
  67 #include <linux/bpf.h>
  68
  69 #include <asm/uaccess.h>
  70 #include <asm/processor.h>
  71
  72 #ifdef CONFIG_X86
  73 #include <asm/nmi.h>
  74 #include <asm/stacktrace.h>
  75 #include <asm/io.h>
  76 #endif
  77 #ifdef CONFIG_SPARC
  78 #include <asm/setup.h>
  79 #endif
  80 #ifdef CONFIG_BSD_PROCESS_ACCT
  81 #include <linux/acct.h>
  82 #endif
  83 #ifdef CONFIG_RT_MUTEXES
  84 #include <linux/rtmutex.h>
  85 #endif
  86 #if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_LOCK_STAT)
  87 #include <linux/lockdep.h>
  88 #endif
  89 #ifdef CONFIG_CHR_DEV_SG
  90 #include <scsi/sg.h>
  91 #endif
  92
  93 #ifdef CONFIG_LOCKUP_DETECTOR
  94 #include <linux/nmi.h>
  95 #endif
  96
  97 #if defined(CONFIG_SYSCTL)
  98
  99 /* External variables not in a header file. */
 100 extern int suid_dumpable;
 101 #ifdef CONFIG_COREDUMP
 102 extern int core_uses_pid;
 103 extern char core_pattern[];
 104 extern unsigned int core_pipe_limit;
 105 #endif
 106 extern int pid_max;
 107 extern int pid_max_min, pid_max_max;
 108 extern int percpu_pagelist_fraction;
 109 extern int compat_log;
 110 extern int latencytop_enabled;
 111 extern int sysctl_nr_open_min, sysctl_nr_open_max;
 112 #ifndef CONFIG_MMU
 113 extern int sysctl_nr_trim_pages;
 114 #endif
 115
 116 /* Constants used for minimum and  maximum */
 117 #ifdef CONFIG_LOCKUP_DETECTOR
 118 static int sixty = 60;
 119 #endif
 120
 121 static int __maybe_unused neg_one = -1;
 122
 123 static int zero;
 124 static int __maybe_unused one = 1;
 125 static int __maybe_unused two = 2;
 126 static int __maybe_unused four = 4;
 127 static unsigned long one_ul = 1;
 128 static int one_hundred = 100;
 129 static int one_thousand = 1000;
 130 #ifdef CONFIG_PRINTK
 131 static int ten_thousand = 10000;
 132 #endif
 133 #ifdef CONFIG_PERF_EVENTS
 134 static int six_hundred_forty_kb = 640 * 1024;
 135 #endif
 136
 137 /* this is needed for the proc_doulongvec_minmax of vm_dirty_bytes */
 138 static unsigned long dirty_bytes_min = 2 * PAGE_SIZE;
 139
 140 /* this is needed for the proc_dointvec_minmax for [fs_]overflow UID and GID */
 141 static int maxolduid = 65535;
 142 static int minolduid;
 143
 144 static int ngroups_max = NGROUPS_MAX;
 145 static const int cap_last_cap = CAP_LAST_CAP;
 146
 147 /*this is needed for proc_doulongvec_minmax of sysctl_hung_task_timeout_secs */
 148 #ifdef CONFIG_DETECT_HUNG_TASK
 149 static unsigned long hung_task_timeout_max = (LONG_MAX/HZ);
 150 #endif
 151
 152 #ifdef CONFIG_INOTIFY_USER
 153 #include <linux/inotify.h>
 154 #endif
 155 #ifdef CONFIG_SPARC
 156 #endif
 157
 158 #ifdef __hppa__
 159 extern int pwrsw_enabled;
 160 #endif
 161
 162 #ifdef CONFIG_SYSCTL_ARCH_UNALIGN_ALLOW
 163 extern int unaligned_enabled;
 164 #endif
 165
 166 #ifdef CONFIG_IA64
 167 extern int unaligned_dump_stack;
 168 #endif
 169
 170 #ifdef CONFIG_SYSCTL_ARCH_UNALIGN_NO_WARN
 171 extern int no_unaligned_warning;
 172 #endif
 173
 174 #ifdef CONFIG_PROC_SYSCTL
 175
 176 #define SYSCTL_WRITES_LEGACY    -1
 177 #define SYSCTL_WRITES_WARN       0
 178 #define SYSCTL_WRITES_STRICT     1
 179
 180 static int sysctl_writes_strict = SYSCTL_WRITES_STRICT;
 181
 182 static int proc_do_cad_pid(struct ctl_table *table, int write,
 183                   void __user *buffer, size_t *lenp, loff_t *ppos);
 184 static int proc_taint(struct ctl_table *table, int write,
 185                                void __user *buffer, size_t *lenp, loff_t *ppos);
 186 #endif
 187
 188 #ifdef CONFIG_PRINTK
 189 static int proc_dointvec_minmax_sysadmin(struct ctl_table *table, int write,
 190                                 void __user *buffer, size_t *lenp, loff_t *ppos);
 191 #endif
 192
 193 static int proc_dointvec_minmax_coredump(struct ctl_table *table, int write,
 194                 void __user *buffer, size_t *lenp, loff_t *ppos);
 195 #ifdef CONFIG_COREDUMP
 196 static int proc_dostring_coredump(struct ctl_table *table, int write,
 197                 void __user *buffer, size_t *lenp, loff_t *ppos);
 198 #endif
 199
 200 #ifdef CONFIG_MAGIC_SYSRQ
 201 /* Note: sysrq code uses it's own private copy */
 202 static int __sysrq_enabled = CONFIG_MAGIC_SYSRQ_DEFAULT_ENABLE;
 203
 204 static int sysrq_sysctl_handler(struct ctl_table *table, int write,
 205                                 void __user *buffer, size_t *lenp,
 206                                 loff_t *ppos)
 207 {
 208         int error;
 209
 210         error = proc_dointvec(table, write, buffer, lenp, ppos);
 211         if (error)
 212                 return error;
 213
 214         if (write)
 215                 sysrq_toggle_support(__sysrq_enabled);
 216
 217         return 0;
 218 }
 219
 220 #endif
 221
 222 static struct ctl_table kern_table[];
 223 static struct ctl_table vm_table[];
 224 static struct ctl_table fs_table[];
 225 static struct ctl_table debug_table[];
 226 static struct ctl_table dev_table[];
 227 extern struct ctl_table random_table[];
 228 #ifdef CONFIG_EPOLL
 229 extern struct ctl_table epoll_table[];
 230 #endif
 231
 232 #ifdef HAVE_ARCH_PICK_MMAP_LAYOUT
 233 int sysctl_legacy_va_layout;
 234 #endif
 235
 236 /* The default sysctl tables: */
 237
 238 static struct ctl_table sysctl_base_table[] = {
 239         {
 240                 .procname       = "kernel",
 241                 .mode           = 0555,
 242                 .child          = kern_table,
 243         },
 244         {
 245                 .procname       = "vm",
 246                 .mode           = 0555,
 247                 .child          = vm_table,
 248         },
 249         {
 250                 .procname       = "fs",
 251                 .mode           = 0555,
 252                 .child          = fs_table,
 253         },
 254         {
 255                 .procname       = "debug",
 256                 .mode           = 0555,
 257                 .child          = debug_table,
 258         },
 259         {
 260                 .procname       = "dev",
 261                 .mode           = 0555,
 262                 .child          = dev_table,
 263         },
 264         { }
 265 };
 266
 267 #ifdef CONFIG_SCHED_DEBUG
 268 static int min_sched_granularity_ns = 100000;           /* 100 usecs */
 269 static int max_sched_granularity_ns = NSEC_PER_SEC;     /* 1 second */
 270 static int min_wakeup_granularity_ns;                   /* 0 usecs */
 271 static int max_wakeup_granularity_ns = NSEC_PER_SEC;    /* 1 second */
 272 #ifdef CONFIG_SMP
 273 static int min_sched_tunable_scaling = SCHED_TUNABLESCALING_NONE;
 274 static int max_sched_tunable_scaling = SCHED_TUNABLESCALING_END-1;
 275 #endif /* CONFIG_SMP */
 276 #endif /* CONFIG_SCHED_DEBUG */
 277
 278 #ifdef CONFIG_COMPACTION
 279 static int min_extfrag_threshold;
 280 static int max_extfrag_threshold = 1000;
 281 #endif
 282
 283 static struct ctl_table kern_table[] = {
 284         {
 285                 .procname       = "sched_child_runs_first",
 286                 .data           = &sysctl_sched_child_runs_first,
 287                 .maxlen         = sizeof(unsigned int),
 288                 .mode           = 0644,
 289                 .proc_handler   = proc_dointvec,
 290         },
 291 #ifdef CONFIG_SCHED_DEBUG
 292         {
 293                 .procname       = "sched_min_granularity_ns",
 294                 .data           = &sysctl_sched_min_granularity,
 295                 .maxlen         = sizeof(unsigned int),
 296                 .mode           = 0644,
 297                 .proc_handler   = sched_proc_update_handler,
 298                 .extra1         = &min_sched_granularity_ns,
 299                 .extra2         = &max_sched_granularity_ns,
 300         },
 301         {
 302                 .procname       = "sched_latency_ns",
 303                 .data           = &sysctl_sched_latency,
 304                 .maxlen         = sizeof(unsigned int),
 305                 .mode           = 0644,
 306                 .proc_handler   = sched_proc_update_handler,
 307                 .extra1         = &min_sched_granularity_ns,
 308                 .extra2         = &max_sched_granularity_ns,
 309         },
 310         {
 311                 .procname       = "sched_wakeup_granularity_ns",
 312                 .data           = &sysctl_sched_wakeup_granularity,
 313                 .maxlen         = sizeof(unsigned int),
 314                 .mode           = 0644,
 315                 .proc_handler   = sched_proc_update_handler,
 316                 .extra1         = &min_wakeup_granularity_ns,
 317                 .extra2         = &max_wakeup_granularity_ns,
 318         },
 319 #ifdef CONFIG_SMP
 320         {
 321                 .procname       = "sched_tunable_scaling",
 322                 .data           = &sysctl_sched_tunable_scaling,
 323                 .maxlen         = sizeof(enum sched_tunable_scaling),
 324                 .mode           = 0644,
 325                 .proc_handler   = sched_proc_update_handler,
 326                 .extra1         = &min_sched_tunable_scaling,
 327                 .extra2         = &max_sched_tunable_scaling,
 328         },
 329         {
 330                 .procname       = "sched_migration_cost_ns",
 331                 .data           = &sysctl_sched_migration_cost,
 332                 .maxlen         = sizeof(unsigned int),
 333                 .mode           = 0644,
 334                 .proc_handler   = proc_dointvec,
 335         },
 336         {
 337                 .procname       = "sched_nr_migrate",
 338                 .data           = &sysctl_sched_nr_migrate,
 339                 .maxlen         = sizeof(unsigned int),
 340                 .mode           = 0644,
 341                 .proc_handler   = proc_dointvec,
 342         },
 343         {
 344                 .procname       = "sched_time_avg_ms",
 345                 .data           = &sysctl_sched_time_avg,
 346                 .maxlen         = sizeof(unsigned int),
 347                 .mode           = 0644,
 348                 .proc_handler   = proc_dointvec,
 349         },
 350         {
 351                 .procname       = "sched_shares_window_ns",
 352                 .data           = &sysctl_sched_shares_window,
 353                 .maxlen         = sizeof(unsigned int),
 354                 .mode           = 0644,
 355                 .proc_handler   = proc_dointvec,
 356         },
 357 #ifdef CONFIG_SCHEDSTATS
 358         {
 359                 .procname       = "sched_schedstats",
 360                 .data           = NULL,
 361                 .maxlen         = sizeof(unsigned int),
 362                 .mode           = 0644,
 363                 .proc_handler   = sysctl_schedstats,
 364                 .extra1         = &zero,
 365                 .extra2         = &one,
 366         },
 367 #endif /* CONFIG_SCHEDSTATS */
 368 #endif /* CONFIG_SMP */
 369 #ifdef CONFIG_NUMA_BALANCING
 370         {
 371                 .procname       = "numa_balancing_scan_delay_ms",
 372                 .data           = &sysctl_numa_balancing_scan_delay,
 373                 .maxlen         = sizeof(unsigned int),
 374                 .mode           = 0644,
 375                 .proc_handler   = proc_dointvec,
 376         },
 377         {
 378                 .procname       = "numa_balancing_scan_period_min_ms",
 379                 .data           = &sysctl_numa_balancing_scan_period_min,
 380                 .maxlen         = sizeof(unsigned int),
 381                 .mode           = 0644,
 382                 .proc_handler   = proc_dointvec,
 383         },
 384         {
 385                 .procname       = "numa_balancing_scan_period_max_ms",
 386                 .data           = &sysctl_numa_balancing_scan_period_max,
 387                 .maxlen         = sizeof(unsigned int),
 388                 .mode           = 0644,
 389                 .proc_handler   = proc_dointvec,
 390         },
 391         {
 392                 .procname       = "numa_balancing_scan_size_mb",
 393                 .data           = &sysctl_numa_balancing_scan_size,
 394                 .maxlen         = sizeof(unsigned int),
 395                 .mode           = 0644,
 396                 .proc_handler   = proc_dointvec_minmax,
 397                 .extra1         = &one,
 398         },
 399         {
 400                 .procname       = "numa_balancing",
 401                 .data           = NULL, /* filled in by handler */
 402                 .maxlen         = sizeof(unsigned int),
 403                 .mode           = 0644,
 404                 .proc_handler   = sysctl_numa_balancing,
 405                 .extra1         = &zero,
 406                 .extra2         = &one,
 407         },
 408 #endif /* CONFIG_NUMA_BALANCING */
 409 #endif /* CONFIG_SCHED_DEBUG */
 410         {
 411                 .procname       = "sched_rt_period_us",
 412                 .data           = &sysctl_sched_rt_period,
 413                 .maxlen         = sizeof(unsigned int),
 414                 .mode           = 0644,
 415                 .proc_handler   = sched_rt_handler,
 416         },
 417         {
 418                 .procname       = "sched_rt_runtime_us",
 419                 .data           = &sysctl_sched_rt_runtime,
 420                 .maxlen         = sizeof(int),
 421                 .mode           = 0644,
 422                 .proc_handler   = sched_rt_handler,
 423         },
 424         {
 425                 .procname       = "sched_rr_timeslice_ms",
 426                 .data           = &sched_rr_timeslice,
 427                 .maxlen         = sizeof(int),
 428                 .mode           = 0644,
 429                 .proc_handler   = sched_rr_handler,
 430         },
 431 #ifdef CONFIG_SCHED_AUTOGROUP
 432         {
 433                 .procname       = "sched_autogroup_enabled",
 434                 .data           = &sysctl_sched_autogroup_enabled,
 435                 .maxlen         = sizeof(unsigned int),
 436                 .mode           = 0644,
 437                 .proc_handler   = proc_dointvec_minmax,
 438                 .extra1         = &zero,
 439                 .extra2         = &one,
 440         },
 441 #endif
 442 #ifdef CONFIG_CFS_BANDWIDTH
 443         {
 444                 .procname       = "sched_cfs_bandwidth_slice_us",
 445                 .data           = &sysctl_sched_cfs_bandwidth_slice,
 446                 .maxlen         = sizeof(unsigned int),
 447                 .mode           = 0644,
 448                 .proc_handler   = proc_dointvec_minmax,
 449                 .extra1         = &one,
 450         },
 451 #endif
 452 #ifdef CONFIG_PROVE_LOCKING
 453         {
 454                 .procname       = "prove_locking",
 455                 .data           = &prove_locking,
 456                 .maxlen         = sizeof(int),
 457                 .mode           = 0644,
 458                 .proc_handler   = proc_dointvec,
 459         },
 460 #endif
 461 #ifdef CONFIG_LOCK_STAT
 462         {
 463                 .procname       = "lock_stat",
 464                 .data           = &lock_stat,
 465                 .maxlen         = sizeof(int),
 466                 .mode           = 0644,
 467                 .proc_handler   = proc_dointvec,
 468         },
 469 #endif
 470         {
 471                 .procname       = "panic",
 472                 .data           = &panic_timeout,
 473                 .maxlen         = sizeof(int),
 474                 .mode           = 0644,
 475                 .proc_handler   = proc_dointvec,
 476         },
 477 #ifdef CONFIG_COREDUMP
 478         {
 479                 .procname       = "core_uses_pid",
 480                 .data           = &core_uses_pid,
 481                 .maxlen         = sizeof(int),
 482                 .mode           = 0644,
 483                 .proc_handler   = proc_dointvec,
 484         },
 485         {
 486                 .procname       = "core_pattern",
 487                 .data           = core_pattern,
 488                 .maxlen         = CORENAME_MAX_SIZE,
 489                 .mode           = 0644,
 490                 .proc_handler   = proc_dostring_coredump,
 491         },
 492         {
 493                 .procname       = "core_pipe_limit",
 494                 .data           = &core_pipe_limit,
 495                 .maxlen         = sizeof(unsigned int),
 496                 .mode           = 0644,
 497                 .proc_handler   = proc_dointvec,
 498         },
 499 #endif
 500 #ifdef CONFIG_PROC_SYSCTL
 501         {
 502                 .procname       = "tainted",
 503                 .maxlen         = sizeof(long),
 504                 .mode           = 0644,
 505                 .proc_handler   = proc_taint,
 506         },
 507         {
 508                 .procname       = "sysctl_writes_strict",
 509                 .data           = &sysctl_writes_strict,
 510                 .maxlen         = sizeof(int),
 511                 .mode           = 0644,
 512                 .proc_handler   = proc_dointvec_minmax,
 513                 .extra1         = &neg_one,
 514                 .extra2         = &one,
 515         },
 516 #endif
 517 #ifdef CONFIG_LATENCYTOP
 518         {
 519                 .procname       = "latencytop",
 520                 .data           = &latencytop_enabled,
 521                 .maxlen         = sizeof(int),
 522                 .mode           = 0644,
 523                 .proc_handler   = sysctl_latencytop,
 524         },
 525 #endif
 526 #ifdef CONFIG_BLK_DEV_INITRD
 527         {
 528                 .procname       = "real-root-dev",
 529                 .data           = &real_root_dev,
 530                 .maxlen         = sizeof(int),
 531                 .mode           = 0644,
 532                 .proc_handler   = proc_dointvec,
 533         },
 534 #endif
 535         {
 536                 .procname       = "print-fatal-signals",
 537                 .data           = &print_fatal_signals,
 538                 .maxlen         = sizeof(int),
 539                 .mode           = 0644,
 540                 .proc_handler   = proc_dointvec,
 541         },
 542 #ifdef CONFIG_SPARC
 543         {
 544                 .procname       = "reboot-cmd",
 545                 .data           = reboot_command,
 546                 .maxlen         = 256,
 547                 .mode           = 0644,
 548                 .proc_handler   = proc_dostring,
 549         },
 550         {
 551                 .procname       = "stop-a",
 552                 .data           = &stop_a_enabled,
 553                 .maxlen         = sizeof (int),
 554                 .mode           = 0644,
 555                 .proc_handler   = proc_dointvec,
 556         },
 557         {
 558                 .procname       = "scons-poweroff",
 559                 .data           = &scons_pwroff,
 560                 .maxlen         = sizeof (int),
 561                 .mode           = 0644,
 562                 .proc_handler   = proc_dointvec,
 563         },
 564 #endif
 565 #ifdef CONFIG_SPARC64
 566         {
 567                 .procname       = "tsb-ratio",
 568                 .data           = &sysctl_tsb_ratio,
 569                 .maxlen         = sizeof (int),
 570                 .mode           = 0644,
 571                 .proc_handler   = proc_dointvec,
 572         },
 573 #endif
 574 #ifdef __hppa__
 575         {
 576                 .procname       = "soft-power",
 577                 .data           = &pwrsw_enabled,
 578                 .maxlen         = sizeof (int),
 579                 .mode           = 0644,
 580                 .proc_handler   = proc_dointvec,
 581         },
 582 #endif
 583 #ifdef CONFIG_SYSCTL_ARCH_UNALIGN_ALLOW
 584         {
 585                 .procname       = "unaligned-trap",
 586                 .data           = &unaligned_enabled,
 587                 .maxlen         = sizeof (int),
 588                 .mode           = 0644,
 589                 .proc_handler   = proc_dointvec,
 590         },
 591 #endif
 592         {
 593                 .procname       = "ctrl-alt-del",
 594                 .data           = &C_A_D,
 595                 .maxlen         = sizeof(int),
 596                 .mode           = 0644,
 597                 .proc_handler   = proc_dointvec,
 598         },
 599 #ifdef CONFIG_FUNCTION_TRACER
 600         {
 601                 .procname       = "ftrace_enabled",
 602                 .data           = &ftrace_enabled,
 603                 .maxlen         = sizeof(int),
 604                 .mode           = 0644,
 605                 .proc_handler   = ftrace_enable_sysctl,
 606         },
 607 #endif
 608 #ifdef CONFIG_STACK_TRACER
 609         {
 610                 .procname       = "stack_tracer_enabled",
 611                 .data           = &stack_tracer_enabled,
 612                 .maxlen         = sizeof(int),
 613                 .mode           = 0644,
 614                 .proc_handler   = stack_trace_sysctl,
 615         },
 616 #endif
 617 #ifdef CONFIG_TRACING
 618         {
 619                 .procname       = "ftrace_dump_on_oops",
 620                 .data           = &ftrace_dump_on_oops,
 621                 .maxlen         = sizeof(int),
 622                 .mode           = 0644,
 623                 .proc_handler   = proc_dointvec,
 624         },
 625         {
 626                 .procname       = "traceoff_on_warning",
 627                 .data           = &__disable_trace_on_warning,
 628                 .maxlen         = sizeof(__disable_trace_on_warning),
 629                 .mode           = 0644,
 630                 .proc_handler   = proc_dointvec,
 631         },
 632         {
 633                 .procname       = "tracepoint_printk",
 634                 .data           = &tracepoint_printk,
 635                 .maxlen         = sizeof(tracepoint_printk),
 636                 .mode           = 0644,
 637                 .proc_handler   = proc_dointvec,
 638         },
 639 #endif
 640 #ifdef CONFIG_KEXEC_CORE
 641         {
 642                 .procname       = "kexec_load_disabled",
 643                 .data           = &kexec_load_disabled,
 644                 .maxlen         = sizeof(int),
 645                 .mode           = 0644,
 646                 /* only handle a transition from default "0" to "1" */
 647                 .proc_handler   = proc_dointvec_minmax,
 648                 .extra1         = &one,
 649                 .extra2         = &one,
 650         },
 651 #endif
 652 #ifdef CONFIG_MODULES
 653         {
 654                 .procname       = "modprobe",
 655                 .data           = &modprobe_path,
 656                 .maxlen         = KMOD_PATH_LEN,
 657                 .mode           = 0644,
 658                 .proc_handler   = proc_dostring,
 659         },
 660         {
 661                 .procname       = "modules_disabled",
 662                 .data           = &modules_disabled,
 663                 .maxlen         = sizeof(int),
 664                 .mode           = 0644,
 665                 /* only handle a transition from default "0" to "1" */
 666                 .proc_handler   = proc_dointvec_minmax,
 667                 .extra1         = &one,
 668                 .extra2         = &one,
 669         },
 670 #endif
 671 #ifdef CONFIG_UEVENT_HELPER
 672         {
 673                 .procname       = "hotplug",
 674                 .data           = &uevent_helper,
 675                 .maxlen         = UEVENT_HELPER_PATH_LEN,
 676                 .mode           = 0644,
 677                 .proc_handler   = proc_dostring,
 678         },
 679 #endif
 680 #ifdef CONFIG_CHR_DEV_SG
 681         {
 682                 .procname       = "sg-big-buff",
 683                 .data           = &sg_big_buff,
 684                 .maxlen         = sizeof (int),
 685                 .mode           = 0444,
 686                 .proc_handler   = proc_dointvec,
 687         },
 688 #endif
 689 #ifdef CONFIG_BSD_PROCESS_ACCT
 690         {
 691                 .procname       = "acct",
 692                 .data           = &acct_parm,
 693                 .maxlen         = 3*sizeof(int),
 694                 .mode           = 0644,
 695                 .proc_handler   = proc_dointvec,
 696         },
 697 #endif
 698 #ifdef CONFIG_MAGIC_SYSRQ
 699         {
 700                 .procname       = "sysrq",
 701                 .data           = &__sysrq_enabled,
 702                 .maxlen         = sizeof (int),
 703                 .mode           = 0644,
 704                 .proc_handler   = sysrq_sysctl_handler,
 705         },
 706 #endif
 707 #ifdef CONFIG_PROC_SYSCTL
 708         {
 709                 .procname       = "cad_pid",
 710                 .data           = NULL,
 711                 .maxlen         = sizeof (int),
 712                 .mode           = 0600,
 713                 .proc_handler   = proc_do_cad_pid,
 714         },
 715 #endif
 716         {
 717                 .procname       = "threads-max",
 718                 .data           = NULL,
 719                 .maxlen         = sizeof(int),
 720                 .mode           = 0644,
 721                 .proc_handler   = sysctl_max_threads,
 722         },
 723         {
 724                 .procname       = "random",
 725                 .mode           = 0555,
 726                 .child          = random_table,
 727         },
 728         {
 729                 .procname       = "usermodehelper",
 730                 .mode           = 0555,
 731                 .child          = usermodehelper_table,
 732         },
 733         {
 734                 .procname       = "overflowuid",
 735                 .data           = &overflowuid,
 736                 .maxlen         = sizeof(int),
 737                 .mode           = 0644,
 738                 .proc_handler   = proc_dointvec_minmax,
 739                 .extra1         = &minolduid,
 740                 .extra2         = &maxolduid,
 741         },
 742         {
 743                 .procname       = "overflowgid",
 744                 .data           = &overflowgid,
 745                 .maxlen         = sizeof(int),
 746                 .mode           = 0644,
 747                 .proc_handler   = proc_dointvec_minmax,
 748                 .extra1         = &minolduid,
 749                 .extra2         = &maxolduid,
 750         },
 751 #ifdef CONFIG_S390
 752 #ifdef CONFIG_MATHEMU
 753         {
 754                 .procname       = "ieee_emulation_warnings",
 755                 .data           = &sysctl_ieee_emulation_warnings,
 756                 .maxlen         = sizeof(int),
 757                 .mode           = 0644,
 758                 .proc_handler   = proc_dointvec,
 759         },
 760 #endif
 761         {
 762                 .procname       = "userprocess_debug",
 763                 .data           = &show_unhandled_signals,
 764                 .maxlen         = sizeof(int),
 765                 .mode           = 0644,
 766                 .proc_handler   = proc_dointvec,
 767         },
 768 #endif
 769         {
 770                 .procname       = "pid_max",
 771                 .data           = &pid_max,
 772                 .maxlen         = sizeof (int),
 773                 .mode           = 0644,
 774                 .proc_handler   = proc_dointvec_minmax,
 775                 .extra1         = &pid_max_min,
 776                 .extra2         = &pid_max_max,
 777         },
 778         {
 779                 .procname       = "panic_on_oops",
 780                 .data           = &panic_on_oops,
 781                 .maxlen         = sizeof(int),
 782                 .mode           = 0644,
 783                 .proc_handler   = proc_dointvec,
 784         },
 785 #if defined CONFIG_PRINTK
 786         {
 787                 .procname       = "printk",
 788                 .data           = &console_loglevel,
 789                 .maxlen         = 4*sizeof(int),
 790                 .mode           = 0644,
 791                 .proc_handler   = proc_dointvec,
 792         },
 793         {
 794                 .procname       = "printk_ratelimit",
 795                 .data           = &printk_ratelimit_state.interval,
 796                 .maxlen         = sizeof(int),
 797                 .mode           = 0644,
 798                 .proc_handler   = proc_dointvec_jiffies,
 799         },
 800         {
 801                 .procname       = "printk_ratelimit_burst",
 802                 .data           = &printk_ratelimit_state.burst,
 803                 .maxlen         = sizeof(int),
 804                 .mode           = 0644,
 805                 .proc_handler   = proc_dointvec,
 806         },
 807         {
 808                 .procname       = "printk_delay",
 809                 .data           = &printk_delay_msec,
 810                 .maxlen         = sizeof(int),
 811                 .mode           = 0644,
 812                 .proc_handler   = proc_dointvec_minmax,
 813                 .extra1         = &zero,
 814                 .extra2         = &ten_thousand,
 815         },
 816         {
 817                 .procname       = "printk_devkmsg",
 818                 .data           = devkmsg_log_str,
 819                 .maxlen         = DEVKMSG_STR_MAX_SIZE,
 820                 .mode           = 0644,
 821                 .proc_handler   = devkmsg_sysctl_set_loglvl,
 822         },
 823         {
 824                 .procname       = "dmesg_restrict",
 825                 .data           = &dmesg_restrict,
 826                 .maxlen         = sizeof(int),
 827                 .mode           = 0644,
 828                 .proc_handler   = proc_dointvec_minmax_sysadmin,
 829                 .extra1         = &zero,
 830                 .extra2         = &one,
 831         },
 832         {
 833                 .procname       = "kptr_restrict",
 834                 .data           = &kptr_restrict,
 835                 .maxlen         = sizeof(int),
 836                 .mode           = 0644,
 837                 .proc_handler   = proc_dointvec_minmax_sysadmin,
 838                 .extra1         = &zero,
 839                 .extra2         = &two,
 840         },
 841 #endif
 842         {
 843                 .procname       = "ngroups_max",
 844                 .data           = &ngroups_max,
 845                 .maxlen         = sizeof (int),
 846                 .mode           = 0444,
 847                 .proc_handler   = proc_dointvec,
 848         },
 849         {
 850                 .procname       = "cap_last_cap",
 851                 .data           = (void *)&cap_last_cap,
 852                 .maxlen         = sizeof(int),
 853                 .mode           = 0444,
 854                 .proc_handler   = proc_dointvec,
 855         },
 856 #if defined(CONFIG_LOCKUP_DETECTOR)
 857         {
 858                 .procname       = "watchdog",
 859                 .data           = &watchdog_user_enabled,
 860                 .maxlen         = sizeof (int),
 861                 .mode           = 0644,
 862                 .proc_handler   = proc_watchdog,
 863                 .extra1         = &zero,
 864                 .extra2         = &one,
 865         },
 866         {
 867                 .procname       = "watchdog_thresh",
 868                 .data           = &watchdog_thresh,
 869                 .maxlen         = sizeof(int),
 870                 .mode           = 0644,
 871                 .proc_handler   = proc_watchdog_thresh,
 872                 .extra1         = &zero,
 873                 .extra2         = &sixty,
 874         },
 875         {
 876                 .procname       = "nmi_watchdog",
 877                 .data           = &nmi_watchdog_enabled,
 878                 .maxlen         = sizeof (int),
 879                 .mode           = 0644,
 880                 .proc_handler   = proc_nmi_watchdog,
 881                 .extra1         = &zero,
 882 #if defined(CONFIG_HAVE_NMI_WATCHDOG) || defined(CONFIG_HARDLOCKUP_DETECTOR)
 883                 .extra2         = &one,
 884 #else
 885                 .extra2         = &zero,
 886 #endif
 887         },
 888         {
 889                 .procname       = "soft_watchdog",
 890                 .data           = &soft_watchdog_enabled,
 891                 .maxlen         = sizeof (int),
 892                 .mode           = 0644,
 893                 .proc_handler   = proc_soft_watchdog,
 894                 .extra1         = &zero,
 895                 .extra2         = &one,
 896         },
 897         {
 898                 .procname       = "watchdog_cpumask",
 899                 .data           = &watchdog_cpumask_bits,
 900                 .maxlen         = NR_CPUS,
 901                 .mode           = 0644,
 902                 .proc_handler   = proc_watchdog_cpumask,
 903         },
 904         {
 905                 .procname       = "softlockup_panic",
 906                 .data           = &softlockup_panic,
 907                 .maxlen         = sizeof(int),
 908                 .mode           = 0644,
 909                 .proc_handler   = proc_dointvec_minmax,
 910                 .extra1         = &zero,
 911                 .extra2         = &one,
 912         },
 913 #ifdef CONFIG_HARDLOCKUP_DETECTOR
 914         {
 915                 .procname       = "hardlockup_panic",
 916                 .data           = &hardlockup_panic,
 917                 .maxlen         = sizeof(int),
 918                 .mode           = 0644,
 919                 .proc_handler   = proc_dointvec_minmax,
 920                 .extra1         = &zero,
 921                 .extra2         = &one,
 922         },
 923 #endif
 924 #ifdef CONFIG_SMP
 925         {
 926                 .procname       = "softlockup_all_cpu_backtrace",
 927                 .data           = &sysctl_softlockup_all_cpu_backtrace,
 928                 .maxlen         = sizeof(int),
 929                 .mode           = 0644,
 930                 .proc_handler   = proc_dointvec_minmax,
 931                 .extra1         = &zero,
 932                 .extra2         = &one,
 933         },
 934         {
 935                 .procname       = "hardlockup_all_cpu_backtrace",
 936                 .data           = &sysctl_hardlockup_all_cpu_backtrace,
 937                 .maxlen         = sizeof(int),
 938                 .mode           = 0644,
 939                 .proc_handler   = proc_dointvec_minmax,
 940                 .extra1         = &zero,
 941                 .extra2         = &one,
 942         },
 943 #endif /* CONFIG_SMP */
 944 #endif
 945 #if defined(CONFIG_X86_LOCAL_APIC) && defined(CONFIG_X86)
 946         {
 947                 .procname       = "unknown_nmi_panic",
 948                 .data           = &unknown_nmi_panic,
 949                 .maxlen         = sizeof (int),
 950                 .mode           = 0644,
 951                 .proc_handler   = proc_dointvec,
 952         },
 953 #endif
 954 #if defined(CONFIG_X86)
 955         {
 956                 .procname       = "panic_on_unrecovered_nmi",
 957                 .data           = &panic_on_unrecovered_nmi,
 958                 .maxlen         = sizeof(int),
 959                 .mode           = 0644,
 960                 .proc_handler   = proc_dointvec,
 961         },
 962         {
 963                 .procname       = "panic_on_io_nmi",
 964                 .data           = &panic_on_io_nmi,
 965                 .maxlen         = sizeof(int),
 966                 .mode           = 0644,
 967                 .proc_handler   = proc_dointvec,
 968         },
 969 #ifdef CONFIG_DEBUG_STACKOVERFLOW
 970         {
 971                 .procname       = "panic_on_stackoverflow",
 972                 .data           = &sysctl_panic_on_stackoverflow,
 973                 .maxlen         = sizeof(int),
 974                 .mode           = 0644,
 975                 .proc_handler   = proc_dointvec,
 976         },
 977 #endif
 978         {
 979                 .procname       = "bootloader_type",
 980                 .data           = &bootloader_type,
 981                 .maxlen         = sizeof (int),
 982                 .mode           = 0444,
 983                 .proc_handler   = proc_dointvec,
 984         },
 985         {
 986                 .procname       = "bootloader_version",
 987                 .data           = &bootloader_version,
 988                 .maxlen         = sizeof (int),
 989                 .mode           = 0444,
 990                 .proc_handler   = proc_dointvec,
 991         },
 992         {
 993                 .procname       = "kstack_depth_to_print",
 994                 .data           = &kstack_depth_to_print,
 995                 .maxlen         = sizeof(int),
 996                 .mode           = 0644,
 997                 .proc_handler   = proc_dointvec,
 998         },
 999         {
1000                 .procname       = "io_delay_type",
1001                 .data           = &io_delay_type,
1002                 .maxlen         = sizeof(int),
1003                 .mode           = 0644,
1004                 .proc_handler   = proc_dointvec,
1005         },
1006 #endif
1007 #if defined(CONFIG_MMU)
1008         {
1009                 .procname       = "randomize_va_space",
1010                 .data           = &randomize_va_space,
1011                 .maxlen         = sizeof(int),
1012                 .mode           = 0644,
1013                 .proc_handler   = proc_dointvec,
1014         },
1015 #endif
1016 #if defined(CONFIG_S390) && defined(CONFIG_SMP)
1017         {
1018                 .procname       = "spin_retry",
1019                 .data           = &spin_retry,
1020                 .maxlen         = sizeof (int),
1021                 .mode           = 0644,
1022                 .proc_handler   = proc_dointvec,
1023         },
1024 #endif
1025 #if     defined(CONFIG_ACPI_SLEEP) && defined(CONFIG_X86)
1026         {
1027                 .procname       = "acpi_video_flags",
1028                 .data           = &acpi_realmode_flags,
1029                 .maxlen         = sizeof (unsigned long),
1030                 .mode           = 0644,
1031                 .proc_handler   = proc_doulongvec_minmax,
1032         },
1033 #endif
1034 #ifdef CONFIG_SYSCTL_ARCH_UNALIGN_NO_WARN
1035         {
1036                 .procname       = "ignore-unaligned-usertrap",
1037                 .data           = &no_unaligned_warning,
1038                 .maxlen         = sizeof (int),
1039                 .mode           = 0644,
1040                 .proc_handler   = proc_dointvec,
1041         },
1042 #endif
1043 #ifdef CONFIG_IA64
1044         {
1045                 .procname       = "unaligned-dump-stack",
1046                 .data           = &unaligned_dump_stack,
1047                 .maxlen         = sizeof (int),
1048                 .mode           = 0644,
1049                 .proc_handler   = proc_dointvec,
1050         },
1051 #endif
1052 #ifdef CONFIG_DETECT_HUNG_TASK
1053         {
1054                 .procname       = "hung_task_panic",
1055                 .data           = &sysctl_hung_task_panic,
1056                 .maxlen         = sizeof(int),
1057                 .mode           = 0644,
1058                 .proc_handler   = proc_dointvec_minmax,
1059                 .extra1         = &zero,
1060                 .extra2         = &one,
1061         },
1062         {
1063                 .procname       = "hung_task_check_count",
1064                 .data           = &sysctl_hung_task_check_count,
1065                 .maxlen         = sizeof(int),
1066                 .mode           = 0644,
1067                 .proc_handler   = proc_dointvec_minmax,
1068                 .extra1         = &zero,
1069         },
1070         {
1071                 .procname       = "hung_task_timeout_secs",
1072                 .data           = &sysctl_hung_task_timeout_secs,
1073                 .maxlen         = sizeof(unsigned long),
1074                 .mode           = 0644,
1075                 .proc_handler   = proc_dohung_task_timeout_secs,
1076                 .extra2         = &hung_task_timeout_max,
1077         },
1078         {
1079                 .procname       = "hung_task_warnings",
1080                 .data           = &sysctl_hung_task_warnings,
1081                 .maxlen         = sizeof(int),
1082                 .mode           = 0644,
1083                 .proc_handler   = proc_dointvec_minmax,
1084                 .extra1         = &neg_one,
1085         },
1086 #endif
1087 #ifdef CONFIG_COMPAT
1088         {
1089                 .procname       = "compat-log",
1090                 .data           = &compat_log,
1091                 .maxlen         = sizeof (int),
1092                 .mode           = 0644,
1093                 .proc_handler   = proc_dointvec,
1094         },
1095 #endif
1096 #ifdef CONFIG_RT_MUTEXES
1097         {
1098                 .procname       = "max_lock_depth",
1099                 .data           = &max_lock_depth,
1100                 .maxlen         = sizeof(int),
1101                 .mode           = 0644,
1102                 .proc_handler   = proc_dointvec,
1103         },
1104 #endif
1105         {
1106                 .procname       = "poweroff_cmd",
1107                 .data           = &poweroff_cmd,
1108                 .maxlen         = POWEROFF_CMD_PATH_LEN,
1109                 .mode           = 0644,
1110                 .proc_handler   = proc_dostring,
1111         },
1112 #ifdef CONFIG_KEYS
1113         {
1114                 .procname       = "keys",
1115                 .mode           = 0555,
1116                 .child          = key_sysctls,
1117         },
1118 #endif
1119 #ifdef CONFIG_PERF_EVENTS
1120         /*
1121          * User-space scripts rely on the existence of this file
1122          * as a feature check for perf_events being enabled.
1123          *
1124          * So it's an ABI, do not remove!
1125          */
1126         {
1127                 .procname       = "perf_event_paranoid",
1128                 .data           = &sysctl_perf_event_paranoid,
1129                 .maxlen         = sizeof(sysctl_perf_event_paranoid),
1130                 .mode           = 0644,
1131                 .proc_handler   = proc_dointvec,
1132         },
1133         {
1134                 .procname       = "perf_event_mlock_kb",
1135                 .data           = &sysctl_perf_event_mlock,
1136                 .maxlen         = sizeof(sysctl_perf_event_mlock),
1137                 .mode           = 0644,
1138                 .proc_handler   = proc_dointvec,
1139         },
1140         {
1141                 .procname       = "perf_event_max_sample_rate",
1142                 .data           = &sysctl_perf_event_sample_rate,
1143                 .maxlen         = sizeof(sysctl_perf_event_sample_rate),
1144                 .mode           = 0644,
1145                 .proc_handler   = perf_proc_update_handler,
1146                 .extra1         = &one,
1147         },
1148         {
1149                 .procname       = "perf_cpu_time_max_percent",
1150                 .data           = &sysctl_perf_cpu_time_max_percent,
1151                 .maxlen         = sizeof(sysctl_perf_cpu_time_max_percent),
1152                 .mode           = 0644,
1153                 .proc_handler   = perf_cpu_time_max_percent_handler,
1154                 .extra1         = &zero,
1155                 .extra2         = &one_hundred,
1156         },
1157         {
1158                 .procname       = "perf_event_max_stack",
1159                 .data           = &sysctl_perf_event_max_stack,
1160                 .maxlen         = sizeof(sysctl_perf_event_max_stack),
1161                 .mode           = 0644,
1162                 .proc_handler   = perf_event_max_stack_handler,
1163                 .extra1         = &zero,
1164                 .extra2         = &six_hundred_forty_kb,
1165         },
1166         {
1167                 .procname       = "perf_event_max_contexts_per_stack",
1168                 .data           = &sysctl_perf_event_max_contexts_per_stack,
1169                 .maxlen         = sizeof(sysctl_perf_event_max_contexts_per_stack),
1170                 .mode           = 0644,
1171                 .proc_handler   = perf_event_max_stack_handler,
1172                 .extra1         = &zero,
1173                 .extra2         = &one_thousand,
1174         },
1175 #endif
1176 #ifdef CONFIG_KMEMCHECK
1177         {
1178                 .procname       = "kmemcheck",
1179                 .data           = &kmemcheck_enabled,
1180                 .maxlen         = sizeof(int),
1181                 .mode           = 0644,
1182                 .proc_handler   = proc_dointvec,
1183         },
1184 #endif
1185         {
1186                 .procname       = "panic_on_warn",
1187                 .data           = &panic_on_warn,
1188                 .maxlen         = sizeof(int),
1189                 .mode           = 0644,
1190                 .proc_handler   = proc_dointvec_minmax,
1191                 .extra1         = &zero,
1192                 .extra2         = &one,
1193         },
1194 #if defined(CONFIG_SMP) && defined(CONFIG_NO_HZ_COMMON)
1195         {
1196                 .procname       = "timer_migration",
1197                 .data           = &sysctl_timer_migration,
1198                 .maxlen         = sizeof(unsigned int),
1199                 .mode           = 0644,
1200                 .proc_handler   = timer_migration_handler,
1201         },
1202 #endif
1203 #ifdef CONFIG_BPF_SYSCALL
1204         {
1205                 .procname       = "unprivileged_bpf_disabled",
1206                 .data           = &sysctl_unprivileged_bpf_disabled,
1207                 .maxlen         = sizeof(sysctl_unprivileged_bpf_disabled),
1208                 .mode           = 0644,
1209                 /* only handle a transition from default "0" to "1" */
1210                 .proc_handler   = proc_dointvec_minmax,
1211                 .extra1         = &one,
1212                 .extra2         = &one,
1213         },
1214 #endif
1215 #if defined(CONFIG_TREE_RCU) || defined(CONFIG_PREEMPT_RCU)
1216         {
1217                 .procname       = "panic_on_rcu_stall",
1218                 .data           = &sysctl_panic_on_rcu_stall,
1219                 .maxlen         = sizeof(sysctl_panic_on_rcu_stall),
1220                 .mode           = 0644,
1221                 .proc_handler   = proc_dointvec_minmax,
1222                 .extra1         = &zero,
1223                 .extra2         = &one,
1224         },
1225 #endif
1226         { }
1227 };
1228
1229 static struct ctl_table vm_table[] = {
1230         {
1231                 .procname       = "overcommit_memory",
1232                 .data           = &sysctl_overcommit_memory,
1233                 .maxlen         = sizeof(sysctl_overcommit_memory),
1234                 .mode           = 0644,
1235                 .proc_handler   = proc_dointvec_minmax,
1236                 .extra1         = &zero,
1237                 .extra2         = &two,
1238         },
1239         {
1240                 .procname       = "panic_on_oom",
1241                 .data           = &sysctl_panic_on_oom,
1242                 .maxlen         = sizeof(sysctl_panic_on_oom),
1243                 .mode           = 0644,
1244                 .proc_handler   = proc_dointvec_minmax,
1245                 .extra1         = &zero,
1246                 .extra2         = &two,
1247         },
1248         {
1249                 .procname       = "oom_kill_allocating_task",
1250                 .data           = &sysctl_oom_kill_allocating_task,
1251                 .maxlen         = sizeof(sysctl_oom_kill_allocating_task),
1252                 .mode           = 0644,
1253                 .proc_handler   = proc_dointvec,
1254         },
1255         {
1256                 .procname       = "oom_dump_tasks",
1257                 .data           = &sysctl_oom_dump_tasks,
1258                 .maxlen         = sizeof(sysctl_oom_dump_tasks),
1259                 .mode           = 0644,
1260                 .proc_handler   = proc_dointvec,
1261         },
1262         {
1263                 .procname       = "overcommit_ratio",
1264                 .data           = &sysctl_overcommit_ratio,
1265                 .maxlen         = sizeof(sysctl_overcommit_ratio),
1266                 .mode           = 0644,
1267                 .proc_handler   = overcommit_ratio_handler,
1268         },
1269         {
1270                 .procname       = "overcommit_kbytes",
1271                 .data           = &sysctl_overcommit_kbytes,
1272                 .maxlen         = sizeof(sysctl_overcommit_kbytes),
1273                 .mode           = 0644,
1274                 .proc_handler   = overcommit_kbytes_handler,
1275         },
1276         {
1277                 .procname       = "page-cluster",
1278                 .data           = &page_cluster,
1279                 .maxlen         = sizeof(int),
1280                 .mode           = 0644,
1281                 .proc_handler   = proc_dointvec_minmax,
1282                 .extra1         = &zero,
1283         },
1284         {
1285                 .procname       = "dirty_background_ratio",
1286                 .data           = &dirty_background_ratio,
1287                 .maxlen         = sizeof(dirty_background_ratio),
1288                 .mode           = 0644,
1289                 .proc_handler   = dirty_background_ratio_handler,
1290                 .extra1         = &zero,
1291                 .extra2         = &one_hundred,
1292         },
1293         {
1294                 .procname       = "dirty_background_bytes",
1295                 .data           = &dirty_background_bytes,
1296                 .maxlen         = sizeof(dirty_background_bytes),
1297                 .mode           = 0644,
1298                 .proc_handler   = dirty_background_bytes_handler,
1299                 .extra1         = &one_ul,
1300         },
1301         {
1302                 .procname       = "dirty_ratio",
1303                 .data           = &vm_dirty_ratio,
1304                 .maxlen         = sizeof(vm_dirty_ratio),
1305                 .mode           = 0644,
1306                 .proc_handler   = dirty_ratio_handler,
1307                 .extra1         = &zero,
1308                 .extra2         = &one_hundred,
1309         },
1310         {
1311                 .procname       = "dirty_bytes",
1312                 .data           = &vm_dirty_bytes,
1313                 .maxlen         = sizeof(vm_dirty_bytes),
1314                 .mode           = 0644,
1315                 .proc_handler   = dirty_bytes_handler,
1316                 .extra1         = &dirty_bytes_min,
1317         },
1318         {
1319                 .procname       = "dirty_writeback_centisecs",
1320                 .data           = &dirty_writeback_interval,
1321                 .maxlen         = sizeof(dirty_writeback_interval),
1322                 .mode           = 0644,
1323                 .proc_handler   = dirty_writeback_centisecs_handler,
1324         },
1325         {
1326                 .procname       = "dirty_expire_centisecs",
1327                 .data           = &dirty_expire_interval,
1328                 .maxlen         = sizeof(dirty_expire_interval),
1329                 .mode           = 0644,
1330                 .proc_handler   = proc_dointvec_minmax,
1331                 .extra1         = &zero,
1332         },
1333         {
1334                 .procname       = "dirtytime_expire_seconds",
1335                 .data           = &dirtytime_expire_interval,
1336                 .maxlen         = sizeof(dirty_expire_interval),
1337                 .mode           = 0644,
1338                 .proc_handler   = dirtytime_interval_handler,
1339                 .extra1         = &zero,
1340         },
1341         {
1342                 .procname       = "nr_pdflush_threads",
1343                 .mode           = 0444 /* read-only */,
1344                 .proc_handler   = pdflush_proc_obsolete,
1345         },
1346         {
1347                 .procname       = "swappiness",
1348                 .data           = &vm_swappiness,
1349                 .maxlen         = sizeof(vm_swappiness),
1350                 .mode           = 0644,
1351                 .proc_handler   = proc_dointvec_minmax,
1352                 .extra1         = &zero,
1353                 .extra2         = &one_hundred,
1354         },
1355 #ifdef CONFIG_HUGETLB_PAGE
1356         {
1357                 .procname       = "nr_hugepages",
1358                 .data           = NULL,
1359                 .maxlen         = sizeof(unsigned long),
1360                 .mode           = 0644,
1361                 .proc_handler   = hugetlb_sysctl_handler,
1362         },
1363 #ifdef CONFIG_NUMA
1364         {
1365                 .procname       = "nr_hugepages_mempolicy",
1366                 .data           = NULL,
1367                 .maxlen         = sizeof(unsigned long),
1368                 .mode           = 0644,
1369                 .proc_handler   = &hugetlb_mempolicy_sysctl_handler,
1370         },
1371 #endif
1372          {
1373                 .procname       = "hugetlb_shm_group",
1374                 .data           = &sysctl_hugetlb_shm_group,
1375                 .maxlen         = sizeof(gid_t),
1376                 .mode           = 0644,
1377                 .proc_handler   = proc_dointvec,
1378          },
1379          {
1380                 .procname       = "hugepages_treat_as_movable",
1381                 .data           = &hugepages_treat_as_movable,
1382                 .maxlen         = sizeof(int),
1383                 .mode           = 0644,
1384                 .proc_handler   = proc_dointvec,
1385         },
1386         {
1387                 .procname       = "nr_overcommit_hugepages",
1388                 .data           = NULL,
1389                 .maxlen         = sizeof(unsigned long),
1390                 .mode           = 0644,
1391                 .proc_handler   = hugetlb_overcommit_handler,
1392         },
1393 #endif
1394         {
1395                 .procname       = "lowmem_reserve_ratio",
1396                 .data           = &sysctl_lowmem_reserve_ratio,
1397                 .maxlen         = sizeof(sysctl_lowmem_reserve_ratio),
1398                 .mode           = 0644,
1399                 .proc_handler   = lowmem_reserve_ratio_sysctl_handler,
1400         },
1401         {
1402                 .procname       = "drop_caches",
1403                 .data           = &sysctl_drop_caches,
1404                 .maxlen         = sizeof(int),
1405                 .mode           = 0644,
1406                 .proc_handler   = drop_caches_sysctl_handler,
1407                 .extra1         = &one,
1408                 .extra2         = &four,
1409         },
1410 #ifdef CONFIG_COMPACTION
1411         {
1412                 .procname       = "compact_memory",
1413                 .data           = &sysctl_compact_memory,
1414                 .maxlen         = sizeof(int),
1415                 .mode           = 0200,
1416                 .proc_handler   = sysctl_compaction_handler,
1417         },
1418         {
1419                 .procname       = "extfrag_threshold",
1420                 .data           = &sysctl_extfrag_threshold,
1421                 .maxlen         = sizeof(int),
1422                 .mode           = 0644,
1423                 .proc_handler   = sysctl_extfrag_handler,
1424                 .extra1         = &min_extfrag_threshold,
1425                 .extra2         = &max_extfrag_threshold,
1426         },
1427         {
1428                 .procname       = "compact_unevictable_allowed",
1429                 .data           = &sysctl_compact_unevictable_allowed,
1430                 .maxlen         = sizeof(int),
1431                 .mode           = 0644,
1432                 .proc_handler   = proc_dointvec,
1433                 .extra1         = &zero,
1434                 .extra2         = &one,
1435         },
1436
1437 #endif /* CONFIG_COMPACTION */
1438         {
1439                 .procname       = "min_free_kbytes",
1440                 .data           = &min_free_kbytes,
1441                 .maxlen         = sizeof(min_free_kbytes),
1442                 .mode           = 0644,
1443                 .proc_handler   = min_free_kbytes_sysctl_handler,
1444                 .extra1         = &zero,
1445         },
1446         {
1447                 .procname       = "watermark_scale_factor",
1448                 .data           = &watermark_scale_factor,
1449                 .maxlen         = sizeof(watermark_scale_factor),
1450                 .mode           = 0644,
1451                 .proc_handler   = watermark_scale_factor_sysctl_handler,
1452                 .extra1         = &one,
1453                 .extra2         = &one_thousand,
1454         },
1455         {
1456                 .procname       = "percpu_pagelist_fraction",
1457                 .data           = &percpu_pagelist_fraction,
1458                 .maxlen         = sizeof(percpu_pagelist_fraction),
1459                 .mode           = 0644,
1460                 .proc_handler   = percpu_pagelist_fraction_sysctl_handler,
1461                 .extra1         = &zero,
1462         },
1463 #ifdef CONFIG_MMU
1464         {
1465                 .procname       = "max_map_count",
1466                 .data           = &sysctl_max_map_count,
1467                 .maxlen         = sizeof(sysctl_max_map_count),
1468                 .mode           = 0644,
1469                 .proc_handler   = proc_dointvec_minmax,
1470                 .extra1         = &zero,
1471         },
1472 #else
1473         {
1474                 .procname       = "nr_trim_pages",
1475                 .data           = &sysctl_nr_trim_pages,
1476                 .maxlen         = sizeof(sysctl_nr_trim_pages),
1477                 .mode           = 0644,
1478                 .proc_handler   = proc_dointvec_minmax,
1479                 .extra1         = &zero,
1480         },
1481 #endif
1482         {
1483                 .procname       = "laptop_mode",
1484                 .data           = &laptop_mode,
1485                 .maxlen         = sizeof(laptop_mode),
1486                 .mode           = 0644,
1487                 .proc_handler   = proc_dointvec_jiffies,
1488         },
1489         {
1490                 .procname       = "block_dump",
1491                 .data           = &block_dump,
1492                 .maxlen         = sizeof(block_dump),
1493                 .mode           = 0644,
1494                 .proc_handler   = proc_dointvec,
1495                 .extra1         = &zero,
1496         },
1497         {
1498                 .procname       = "vfs_cache_pressure",
1499                 .data           = &sysctl_vfs_cache_pressure,
1500                 .maxlen         = sizeof(sysctl_vfs_cache_pressure),
1501                 .mode           = 0644,
1502                 .proc_handler   = proc_dointvec,
1503                 .extra1         = &zero,
1504         },
1505 #ifdef HAVE_ARCH_PICK_MMAP_LAYOUT
1506         {
1507                 .procname       = "legacy_va_layout",
1508                 .data           = &sysctl_legacy_va_layout,
1509                 .maxlen         = sizeof(sysctl_legacy_va_layout),
1510                 .mode           = 0644,
1511                 .proc_handler   = proc_dointvec,
1512                 .extra1         = &zero,
1513         },
1514 #endif
1515 #ifdef CONFIG_NUMA
1516         {
1517                 .procname       = "zone_reclaim_mode",
1518                 .data           = &node_reclaim_mode,
1519                 .maxlen         = sizeof(node_reclaim_mode),
1520                 .mode           = 0644,
1521                 .proc_handler   = proc_dointvec,
1522                 .extra1         = &zero,
1523         },
1524         {
1525                 .procname       = "min_unmapped_ratio",
1526                 .data           = &sysctl_min_unmapped_ratio,
1527                 .maxlen         = sizeof(sysctl_min_unmapped_ratio),
1528                 .mode           = 0644,
1529                 .proc_handler   = sysctl_min_unmapped_ratio_sysctl_handler,
1530                 .extra1         = &zero,
1531                 .extra2         = &one_hundred,
1532         },
1533         {
1534                 .procname       = "min_slab_ratio",
1535                 .data           = &sysctl_min_slab_ratio,
1536                 .maxlen         = sizeof(sysctl_min_slab_ratio),
1537                 .mode           = 0644,
1538                 .proc_handler   = sysctl_min_slab_ratio_sysctl_handler,
1539                 .extra1         = &zero,
1540                 .extra2         = &one_hundred,
1541         },
1542 #endif
1543 #ifdef CONFIG_SMP
1544         {
1545                 .procname       = "stat_interval",
1546                 .data           = &sysctl_stat_interval,
1547                 .maxlen         = sizeof(sysctl_stat_interval),
1548                 .mode           = 0644,
1549                 .proc_handler   = proc_dointvec_jiffies,
1550         },
1551         {
1552                 .procname       = "stat_refresh",
1553                 .data           = NULL,
1554                 .maxlen         = 0,
1555                 .mode           = 0600,
1556                 .proc_handler   = vmstat_refresh,
1557         },
1558 #endif
1559 #ifdef CONFIG_MMU
1560         {
1561                 .procname       = "mmap_min_addr",
1562                 .data           = &dac_mmap_min_addr,
1563                 .maxlen         = sizeof(unsigned long),
1564                 .mode           = 0644,
1565                 .proc_handler   = mmap_min_addr_handler,
1566         },
1567 #endif
1568 #ifdef CONFIG_NUMA
1569         {
1570                 .procname       = "numa_zonelist_order",
1571                 .data           = &numa_zonelist_order,
1572                 .maxlen         = NUMA_ZONELIST_ORDER_LEN,
1573                 .mode           = 0644,
1574                 .proc_handler   = numa_zonelist_order_handler,
1575         },
1576 #endif
1577 #if (defined(CONFIG_X86_32) && !defined(CONFIG_UML))|| \
1578    (defined(CONFIG_SUPERH) && defined(CONFIG_VSYSCALL))
1579         {
1580                 .procname       = "vdso_enabled",
1581 #ifdef CONFIG_X86_32
1582                 .data           = &vdso32_enabled,
1583                 .maxlen         = sizeof(vdso32_enabled),
1584 #else
1585                 .data           = &vdso_enabled,
1586                 .maxlen         = sizeof(vdso_enabled),
1587 #endif
1588                 .mode           = 0644,
1589                 .proc_handler   = proc_dointvec,
1590                 .extra1         = &zero,
1591         },
1592 #endif
1593 #ifdef CONFIG_HIGHMEM
1594         {
1595                 .procname       = "highmem_is_dirtyable",
1596                 .data           = &vm_highmem_is_dirtyable,
1597                 .maxlen         = sizeof(vm_highmem_is_dirtyable),
1598                 .mode           = 0644,
1599                 .proc_handler   = proc_dointvec_minmax,
1600                 .extra1         = &zero,
1601                 .extra2         = &one,
1602         },
1603 #endif
1604 #ifdef CONFIG_MEMORY_FAILURE
1605         {
1606                 .procname       = "memory_failure_early_kill",
1607                 .data           = &sysctl_memory_failure_early_kill,
1608                 .maxlen         = sizeof(sysctl_memory_failure_early_kill),
1609                 .mode           = 0644,
1610                 .proc_handler   = proc_dointvec_minmax,
1611                 .extra1         = &zero,
1612                 .extra2         = &one,
1613         },
1614         {
1615                 .procname       = "memory_failure_recovery",
1616                 .data           = &sysctl_memory_failure_recovery,
1617                 .maxlen         = sizeof(sysctl_memory_failure_recovery),
1618                 .mode           = 0644,
1619                 .proc_handler   = proc_dointvec_minmax,
1620                 .extra1         = &zero,
1621                 .extra2         = &one,
1622         },
1623 #endif
1624         {
1625                 .procname       = "user_reserve_kbytes",
1626                 .data           = &sysctl_user_reserve_kbytes,
1627                 .maxlen         = sizeof(sysctl_user_reserve_kbytes),
1628                 .mode           = 0644,
1629                 .proc_handler   = proc_doulongvec_minmax,
1630         },
1631         {
1632                 .procname       = "admin_reserve_kbytes",
1633                 .data           = &sysctl_admin_reserve_kbytes,
1634                 .maxlen         = sizeof(sysctl_admin_reserve_kbytes),
1635                 .mode           = 0644,
1636                 .proc_handler   = proc_doulongvec_minmax,
1637         },
1638 #ifdef CONFIG_HAVE_ARCH_MMAP_RND_BITS
1639         {
1640                 .procname       = "mmap_rnd_bits",
1641                 .data           = &mmap_rnd_bits,
1642                 .maxlen         = sizeof(mmap_rnd_bits),
1643                 .mode           = 0600,
1644                 .proc_handler   = proc_dointvec_minmax,
1645                 .extra1         = (void *)&mmap_rnd_bits_min,
1646                 .extra2         = (void *)&mmap_rnd_bits_max,
1647         },
1648 #endif
1649 #ifdef CONFIG_HAVE_ARCH_MMAP_RND_COMPAT_BITS
1650         {
1651                 .procname       = "mmap_rnd_compat_bits",
1652                 .data           = &mmap_rnd_compat_bits,
1653                 .maxlen         = sizeof(mmap_rnd_compat_bits),
1654                 .mode           = 0600,
1655                 .proc_handler   = proc_dointvec_minmax,
1656                 .extra1         = (void *)&mmap_rnd_compat_bits_min,
1657                 .extra2         = (void *)&mmap_rnd_compat_bits_max,
1658         },
1659 #endif
1660         { }
1661 };
1662
1663 static struct ctl_table fs_table[] = {
1664         {
1665                 .procname       = "inode-nr",
1666                 .data           = &inodes_stat,
1667                 .maxlen         = 2*sizeof(long),
1668                 .mode           = 0444,
1669                 .proc_handler   = proc_nr_inodes,
1670         },
1671         {
1672                 .procname       = "inode-state",
1673                 .data           = &inodes_stat,
1674                 .maxlen         = 7*sizeof(long),
1675                 .mode           = 0444,
1676                 .proc_handler   = proc_nr_inodes,
1677         },
1678         {
1679                 .procname       = "file-nr",
1680                 .data           = &files_stat,
1681                 .maxlen         = sizeof(files_stat),
1682                 .mode           = 0444,
1683                 .proc_handler   = proc_nr_files,
1684         },
1685         {
1686                 .procname       = "file-max",
1687                 .data           = &files_stat.max_files,
1688                 .maxlen         = sizeof(files_stat.max_files),
1689                 .mode           = 0644,
1690                 .proc_handler   = proc_doulongvec_minmax,
1691         },
1692         {
1693                 .procname       = "nr_open",
1694                 .data           = &sysctl_nr_open,
1695                 .maxlen         = sizeof(int),
1696                 .mode           = 0644,
1697                 .proc_handler   = proc_dointvec_minmax,
1698                 .extra1         = &sysctl_nr_open_min,
1699                 .extra2         = &sysctl_nr_open_max,
1700         },
1701         {
1702                 .procname       = "dentry-state",
1703                 .data           = &dentry_stat,
1704                 .maxlen         = 6*sizeof(long),
1705                 .mode           = 0444,
1706                 .proc_handler   = proc_nr_dentry,
1707         },
1708         {
1709                 .procname       = "overflowuid",
1710                 .data           = &fs_overflowuid,
1711                 .maxlen         = sizeof(int),
1712                 .mode           = 0644,
1713                 .proc_handler   = proc_dointvec_minmax,
1714                 .extra1         = &minolduid,
1715                 .extra2         = &maxolduid,
1716         },
1717         {
1718                 .procname       = "overflowgid",
1719                 .data           = &fs_overflowgid,
1720                 .maxlen         = sizeof(int),
1721                 .mode           = 0644,
1722                 .proc_handler   = proc_dointvec_minmax,
1723                 .extra1         = &minolduid,
1724                 .extra2         = &maxolduid,
1725         },
1726 #ifdef CONFIG_FILE_LOCKING
1727         {
1728                 .procname       = "leases-enable",
1729                 .data           = &leases_enable,
1730                 .maxlen         = sizeof(int),
1731                 .mode           = 0644,
1732                 .proc_handler   = proc_dointvec,
1733         },
1734 #endif
1735 #ifdef CONFIG_DNOTIFY
1736         {
1737                 .procname       = "dir-notify-enable",
1738                 .data           = &dir_notify_enable,
1739                 .maxlen         = sizeof(int),
1740                 .mode           = 0644,
1741                 .proc_handler   = proc_dointvec,
1742         },
1743 #endif
1744 #ifdef CONFIG_MMU
1745 #ifdef CONFIG_FILE_LOCKING
1746         {
1747                 .procname       = "lease-break-time",
1748                 .data           = &lease_break_time,
1749                 .maxlen         = sizeof(int),
1750                 .mode           = 0644,
1751                 .proc_handler   = proc_dointvec,
1752         },
1753 #endif
1754 #ifdef CONFIG_AIO
1755         {
1756                 .procname       = "aio-nr",
1757                 .data           = &aio_nr,
1758                 .maxlen         = sizeof(aio_nr),
1759                 .mode           = 0444,
1760                 .proc_handler   = proc_doulongvec_minmax,
1761         },
1762         {
1763                 .procname       = "aio-max-nr",
1764                 .data           = &aio_max_nr,
1765                 .maxlen         = sizeof(aio_max_nr),
1766                 .mode           = 0644,
1767                 .proc_handler   = proc_doulongvec_minmax,
1768         },
1769 #endif /* CONFIG_AIO */
1770 #ifdef CONFIG_INOTIFY_USER
1771         {
1772                 .procname       = "inotify",
1773                 .mode           = 0555,
1774                 .child          = inotify_table,
1775         },
1776 #endif
1777 #ifdef CONFIG_EPOLL
1778         {
1779                 .procname       = "epoll",
1780                 .mode           = 0555,
1781                 .child          = epoll_table,
1782         },
1783 #endif
1784 #endif
1785         {
1786                 .procname       = "protected_symlinks",
1787                 .data           = &sysctl_protected_symlinks,
1788                 .maxlen         = sizeof(int),
1789                 .mode           = 0600,
1790                 .proc_handler   = proc_dointvec_minmax,
1791                 .extra1         = &zero,
1792                 .extra2         = &one,
1793         },
1794         {
1795                 .procname       = "protected_hardlinks",
1796                 .data           = &sysctl_protected_hardlinks,
1797                 .maxlen         = sizeof(int),
1798                 .mode           = 0600,
1799                 .proc_handler   = proc_dointvec_minmax,
1800                 .extra1         = &zero,
1801                 .extra2         = &one,
1802         },
1803         {
1804                 .procname       = "suid_dumpable",
1805                 .data           = &suid_dumpable,
1806                 .maxlen         = sizeof(int),
1807                 .mode           = 0644,
1808                 .proc_handler   = proc_dointvec_minmax_coredump,
1809                 .extra1         = &zero,
1810                 .extra2         = &two,
1811         },
1812 #if defined(CONFIG_BINFMT_MISC) || defined(CONFIG_BINFMT_MISC_MODULE)
1813         {
1814                 .procname       = "binfmt_misc",
1815                 .mode           = 0555,
1816                 .child          = sysctl_mount_point,
1817         },
1818 #endif
1819         {
1820                 .procname       = "pipe-max-size",
1821                 .data           = &pipe_max_size,
1822                 .maxlen         = sizeof(int),
1823                 .mode           = 0644,
1824                 .proc_handler   = &pipe_proc_fn,
1825                 .extra1         = &pipe_min_size,
1826         },
1827         {
1828                 .procname       = "pipe-user-pages-hard",
1829                 .data           = &pipe_user_pages_hard,
1830                 .maxlen         = sizeof(pipe_user_pages_hard),
1831                 .mode           = 0644,
1832                 .proc_handler   = proc_doulongvec_minmax,
1833         },
1834         {
1835                 .procname       = "pipe-user-pages-soft",
1836                 .data           = &pipe_user_pages_soft,
1837                 .maxlen         = sizeof(pipe_user_pages_soft),
1838                 .mode           = 0644,
1839                 .proc_handler   = proc_doulongvec_minmax,
1840         },
1841         { }
1842 };
1843
1844 static struct ctl_table debug_table[] = {
1845 #ifdef CONFIG_SYSCTL_EXCEPTION_TRACE
1846         {
1847                 .procname       = "exception-trace",
1848                 .data           = &show_unhandled_signals,
1849                 .maxlen         = sizeof(int),
1850                 .mode           = 0644,
1851                 .proc_handler   = proc_dointvec
1852         },
1853 #endif
1854 #if defined(CONFIG_OPTPROBES)
1855         {
1856                 .procname       = "kprobes-optimization",
1857                 .data           = &sysctl_kprobes_optimization,
1858                 .maxlen         = sizeof(int),
1859                 .mode           = 0644,
1860                 .proc_handler   = proc_kprobes_optimization_handler,
1861                 .extra1         = &zero,
1862                 .extra2         = &one,
1863         },
1864 #endif
1865         { }
1866 };
1867
1868 static struct ctl_table dev_table[] = {
1869         { }
1870 };
1871
1872 int __init sysctl_init(void)
1873 {
1874         struct ctl_table_header *hdr;
1875
1876         hdr = register_sysctl_table(sysctl_base_table);
1877         kmemleak_not_leak(hdr);
1878         return 0;
1879 }
1880
1881 #endif /* CONFIG_SYSCTL */
1882
1883 /*
1884  * /proc/sys support
1885  */
1886
1887 #ifdef CONFIG_PROC_SYSCTL
1888
1889 static int _proc_do_string(char *data, int maxlen, int write,
1890                            char __user *buffer,
1891                            size_t *lenp, loff_t *ppos)
1892 {
1893         size_t len;
1894         char __user *p;
1895         char c;
1896
1897         if (!data || !maxlen || !*lenp) {
1898                 *lenp = 0;
1899                 return 0;
1900         }
1901
1902         if (write) {
1903                 if (sysctl_writes_strict == SYSCTL_WRITES_STRICT) {
1904                         /* Only continue writes not past the end of buffer. */
1905                         len = strlen(data);
1906                         if (len > maxlen - 1)
1907                                 len = maxlen - 1;
1908
1909                         if (*ppos > len)
1910                                 return 0;
1911                         len = *ppos;
1912                 } else {
1913                         /* Start writing from beginning of buffer. */
1914                         len = 0;
1915                 }
1916
1917                 *ppos += *lenp;
1918                 p = buffer;
1919                 while ((p - buffer) < *lenp && len < maxlen - 1) {
1920                         if (get_user(c, p++))
1921                                 return -EFAULT;
1922                         if (c == 0 || c == '\n')
1923                                 break;
1924                         data[len++] = c;
1925                 }
1926                 data[len] = 0;
1927         } else {
1928                 len = strlen(data);
1929                 if (len > maxlen)
1930                         len = maxlen;
1931
1932                 if (*ppos > len) {
1933                         *lenp = 0;
1934                         return 0;
1935                 }
1936
1937                 data += *ppos;
1938                 len  -= *ppos;
1939
1940                 if (len > *lenp)
1941                         len = *lenp;
1942                 if (len)
1943                         if (copy_to_user(buffer, data, len))
1944                                 return -EFAULT;
1945                 if (len < *lenp) {
1946                         if (put_user('\n', buffer + len))
1947                                 return -EFAULT;
1948                         len++;
1949                 }
1950                 *lenp = len;
1951                 *ppos += len;
1952         }
1953         return 0;
1954 }
1955
1956 static void warn_sysctl_write(struct ctl_table *table)
1957 {
1958         pr_warn_once("%s wrote to %s when file position was not 0!\n"
1959                 "This will not be supported in the future. To silence this\n"
1960                 "warning, set kernel.sysctl_writes_strict = -1\n",
1961                 current->comm, table->procname);
1962 }
1963
1964 /**
1965  * proc_dostring - read a string sysctl
1966  * @table: the sysctl table
1967  * @write: %TRUE if this is a write to the sysctl file
1968  * @buffer: the user buffer
1969  * @lenp: the size of the user buffer
1970  * @ppos: file position
1971  *
1972  * Reads/writes a string from/to the user buffer. If the kernel
1973  * buffer provided is not large enough to hold the string, the
1974  * string is truncated. The copied string is %NULL-terminated.
1975  * If the string is being read by the user process, it is copied
1976  * and a newline '\n' is added. It is truncated if the buffer is
1977  * not large enough.
1978  *
1979  * Returns 0 on success.
1980  */
1981 int proc_dostring(struct ctl_table *table, int write,
1982                   void __user *buffer, size_t *lenp, loff_t *ppos)
1983 {
1984         if (write && *ppos && sysctl_writes_strict == SYSCTL_WRITES_WARN)
1985                 warn_sysctl_write(table);
1986
1987         return _proc_do_string((char *)(table->data), table->maxlen, write,
1988                                (char __user *)buffer, lenp, ppos);
1989 }
1990
1991 static size_t proc_skip_spaces(char **buf)
1992 {
1993         size_t ret;
1994         char *tmp = skip_spaces(*buf);
1995         ret = tmp - *buf;
1996         *buf = tmp;
1997         return ret;
1998 }
1999
2000 static void proc_skip_char(char **buf, size_t *size, const char v)
2001 {
2002         while (*size) {
2003                 if (**buf != v)
2004                         break;
2005                 (*size)--;
2006                 (*buf)++;
2007         }
2008 }
2009
2010 #define TMPBUFLEN 22
2011 /**
2012  * proc_get_long - reads an ASCII formatted integer from a user buffer
2013  *
2014  * @buf: a kernel buffer
2015  * @size: size of the kernel buffer
2016  * @val: this is where the number will be stored
2017  * @neg: set to %TRUE if number is negative
2018  * @perm_tr: a vector which contains the allowed trailers
2019  * @perm_tr_len: size of the perm_tr vector
2020  * @tr: pointer to store the trailer character
2021  *
2022  * In case of success %0 is returned and @buf and @size are updated with
2023  * the amount of bytes read. If @tr is non-NULL and a trailing
2024  * character exists (size is non-zero after returning from this
2025  * function), @tr is updated with the trailing character.
2026  */
2027 static int proc_get_long(char **buf, size_t *size,
2028                           unsigned long *val, bool *neg,
2029                           const char *perm_tr, unsigned perm_tr_len, char *tr)
2030 {
2031         int len;
2032         char *p, tmp[TMPBUFLEN];
2033
2034         if (!*size)
2035                 return -EINVAL;
2036
2037         len = *size;
2038         if (len > TMPBUFLEN - 1)
2039                 len = TMPBUFLEN - 1;
2040
2041         memcpy(tmp, *buf, len);
2042
2043         tmp[len] = 0;
2044         p = tmp;
2045         if (*p == '-' && *size > 1) {
2046                 *neg = true;
2047                 p++;
2048         } else
2049                 *neg = false;
2050         if (!isdigit(*p))
2051                 return -EINVAL;
2052
2053         *val = simple_strtoul(p, &p, 0);
2054
2055         len = p - tmp;
2056
2057         /* We don't know if the next char is whitespace thus we may accept
2058          * invalid integers (e.g. 1234...a) or two integers instead of one
2059          * (e.g. 123...1). So lets not allow such large numbers. */
2060         if (len == TMPBUFLEN - 1)
2061                 return -EINVAL;
2062
2063         if (len < *size && perm_tr_len && !memchr(perm_tr, *p, perm_tr_len))
2064                 return -EINVAL;
2065
2066         if (tr && (len < *size))
2067                 *tr = *p;
2068
2069         *buf += len;
2070         *size -= len;
2071
2072         return 0;
2073 }
2074
2075 /**
2076  * proc_put_long - converts an integer to a decimal ASCII formatted string
2077  *
2078  * @buf: the user buffer
2079  * @size: the size of the user buffer
2080  * @val: the integer to be converted
2081  * @neg: sign of the number, %TRUE for negative
2082  *
2083  * In case of success %0 is returned and @buf and @size are updated with
2084  * the amount of bytes written.
2085  */
2086 static int proc_put_long(void __user **buf, size_t *size, unsigned long val,
2087                           bool neg)
2088 {
2089         int len;
2090         char tmp[TMPBUFLEN], *p = tmp;
2091
2092         sprintf(p, "%s%lu", neg ? "-" : "", val);
2093         len = strlen(tmp);
2094         if (len > *size)
2095                 len = *size;
2096         if (copy_to_user(*buf, tmp, len))
2097                 return -EFAULT;
2098         *size -= len;
2099         *buf += len;
2100         return 0;
2101 }
2102 #undef TMPBUFLEN
2103
2104 static int proc_put_char(void __user **buf, size_t *size, char c)
2105 {
2106         if (*size) {
2107                 char __user **buffer = (char __user **)buf;
2108                 if (put_user(c, *buffer))
2109                         return -EFAULT;
2110                 (*size)--, (*buffer)++;
2111                 *buf = *buffer;
2112         }
2113         return 0;
2114 }
2115
2116 static int do_proc_dointvec_conv(bool *negp, unsigned long *lvalp,
2117                                  int *valp,
2118                                  int write, void *data)
2119 {
2120         if (write) {
2121                 if (*negp) {
2122                         if (*lvalp > (unsigned long) INT_MAX + 1)
2123                                 return -EINVAL;
2124                         *valp = -*lvalp;
2125                 } else {
2126                         if (*lvalp > (unsigned long) INT_MAX)
2127                                 return -EINVAL;
2128                         *valp = *lvalp;
2129                 }
2130         } else {
2131                 int val = *valp;
2132                 if (val < 0) {
2133                         *negp = true;
2134                         *lvalp = -(unsigned long)val;
2135                 } else {
2136                         *negp = false;
2137                         *lvalp = (unsigned long)val;
2138                 }
2139         }
2140         return 0;
2141 }
2142
2143 static const char proc_wspace_sep[] = { ' ', '\t', '\n' };
2144
2145 static int __do_proc_dointvec(void *tbl_data, struct ctl_table *table,
2146                   int write, void __user *buffer,
2147                   size_t *lenp, loff_t *ppos,
2148                   int (*conv)(bool *negp, unsigned long *lvalp, int *valp,
2149                               int write, void *data),
2150                   void *data)
2151 {
2152         int *i, vleft, first = 1, err = 0;
2153         size_t left;
2154         char *kbuf = NULL, *p;
2155
2156         if (!tbl_data || !table->maxlen || !*lenp || (*ppos && !write)) {
2157                 *lenp = 0;
2158                 return 0;
2159         }
2160
2161         i = (int *) tbl_data;
2162         vleft = table->maxlen / sizeof(*i);
2163         left = *lenp;
2164
2165         if (!conv)
2166                 conv = do_proc_dointvec_conv;
2167
2168         if (write) {
2169                 if (*ppos) {
2170                         switch (sysctl_writes_strict) {
2171                         case SYSCTL_WRITES_STRICT:
2172                                 goto out;
2173                         case SYSCTL_WRITES_WARN:
2174                                 warn_sysctl_write(table);
2175                                 break;
2176                         default:
2177                                 break;
2178                         }
2179                 }
2180
2181                 if (left > PAGE_SIZE - 1)
2182                         left = PAGE_SIZE - 1;
2183                 p = kbuf = memdup_user_nul(buffer, left);
2184                 if (IS_ERR(kbuf))
2185                         return PTR_ERR(kbuf);
2186         }
2187
2188         for (; left && vleft--; i++, first=0) {
2189                 unsigned long lval;
2190                 bool neg;
2191
2192                 if (write) {
2193                         left -= proc_skip_spaces(&p);
2194
2195                         if (!left)
2196                                 break;
2197                         err = proc_get_long(&p, &left, &lval, &neg,
2198                                              proc_wspace_sep,
2199                                              sizeof(proc_wspace_sep), NULL);
2200                         if (err)
2201                                 break;
2202                         if (conv(&neg, &lval, i, 1, data)) {
2203                                 err = -EINVAL;
2204                                 break;
2205                         }
2206                 } else {
2207                         if (conv(&neg, &lval, i, 0, data)) {
2208                                 err = -EINVAL;
2209                                 break;
2210                         }
2211                         if (!first)
2212                                 err = proc_put_char(&buffer, &left, '\t');
2213                         if (err)
2214                                 break;
2215                         err = proc_put_long(&buffer, &left, lval, neg);
2216                         if (err)
2217                                 break;
2218                 }
2219         }
2220
2221         if (!write && !first && left && !err)
2222                 err = proc_put_char(&buffer, &left, '\n');
2223         if (write && !err && left)
2224                 left -= proc_skip_spaces(&p);
2225         if (write) {
2226                 kfree(kbuf);
2227                 if (first)
2228                         return err ? : -EINVAL;
2229         }
2230         *lenp -= left;
2231 out:
2232         *ppos += *lenp;
2233         return err;
2234 }
2235
2236 static int do_proc_dointvec(struct ctl_table *table, int write,
2237                   void __user *buffer, size_t *lenp, loff_t *ppos,
2238                   int (*conv)(bool *negp, unsigned long *lvalp, int *valp,
2239                               int write, void *data),
2240                   void *data)
2241 {
2242         return __do_proc_dointvec(table->data, table, write,
2243                         buffer, lenp, ppos, conv, data);
2244 }
2245
2246 /**
2247  * proc_dointvec - read a vector of integers
2248  * @table: the sysctl table
2249  * @write: %TRUE if this is a write to the sysctl file
2250  * @buffer: the user buffer
2251  * @lenp: the size of the user buffer
2252  * @ppos: file position
2253  *
2254  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2255  * values from/to the user buffer, treated as an ASCII string.
2256  *
2257  * Returns 0 on success.
2258  */
2259 int proc_dointvec(struct ctl_table *table, int write,
2260                      void __user *buffer, size_t *lenp, loff_t *ppos)
2261 {
2262     return do_proc_dointvec(table,write,buffer,lenp,ppos,
2263                             NULL,NULL);
2264 }
2265
2266 /*
2267  * Taint values can only be increased
2268  * This means we can safely use a temporary.
2269  */
2270 static int proc_taint(struct ctl_table *table, int write,
2271                                void __user *buffer, size_t *lenp, loff_t *ppos)
2272 {
2273         struct ctl_table t;
2274         unsigned long tmptaint = get_taint();
2275         int err;
2276
2277         if (write && !capable(CAP_SYS_ADMIN))
2278                 return -EPERM;
2279
2280         t = *table;
2281         t.data = &tmptaint;
2282         err = proc_doulongvec_minmax(&t, write, buffer, lenp, ppos);
2283         if (err < 0)
2284                 return err;
2285
2286         if (write) {
2287                 /*
2288                  * Poor man's atomic or. Not worth adding a primitive
2289                  * to everyone's atomic.h for this
2290                  */
2291                 int i;
2292                 for (i = 0; i < BITS_PER_LONG && tmptaint >> i; i++) {
2293                         if ((tmptaint >> i) & 1)
2294                                 add_taint(i, LOCKDEP_STILL_OK);
2295                 }
2296         }
2297
2298         return err;
2299 }
2300
2301 #ifdef CONFIG_PRINTK
2302 static int proc_dointvec_minmax_sysadmin(struct ctl_table *table, int write,
2303                                 void __user *buffer, size_t *lenp, loff_t *ppos)
2304 {
2305         if (write && !capable(CAP_SYS_ADMIN))
2306                 return -EPERM;
2307
2308         return proc_dointvec_minmax(table, write, buffer, lenp, ppos);
2309 }
2310 #endif
2311
2312 struct do_proc_dointvec_minmax_conv_param {
2313         int *min;
2314         int *max;
2315 };
2316
2317 static int do_proc_dointvec_minmax_conv(bool *negp, unsigned long *lvalp,
2318                                         int *valp,
2319                                         int write, void *data)
2320 {
2321         struct do_proc_dointvec_minmax_conv_param *param = data;
2322         if (write) {
2323                 int val = *negp ? -*lvalp : *lvalp;
2324                 if ((param->min && *param->min > val) ||
2325                     (param->max && *param->max < val))
2326                         return -EINVAL;
2327                 *valp = val;
2328         } else {
2329                 int val = *valp;
2330                 if (val < 0) {
2331                         *negp = true;
2332                         *lvalp = -(unsigned long)val;
2333                 } else {
2334                         *negp = false;
2335                         *lvalp = (unsigned long)val;
2336                 }
2337         }
2338         return 0;
2339 }
2340
2341 /**
2342  * proc_dointvec_minmax - read a vector of integers with min/max values
2343  * @table: the sysctl table
2344  * @write: %TRUE if this is a write to the sysctl file
2345  * @buffer: the user buffer
2346  * @lenp: the size of the user buffer
2347  * @ppos: file position
2348  *
2349  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2350  * values from/to the user buffer, treated as an ASCII string.
2351  *
2352  * This routine will ensure the values are within the range specified by
2353  * table->extra1 (min) and table->extra2 (max).
2354  *
2355  * Returns 0 on success.
2356  */
2357 int proc_dointvec_minmax(struct ctl_table *table, int write,
2358                   void __user *buffer, size_t *lenp, loff_t *ppos)
2359 {
2360         struct do_proc_dointvec_minmax_conv_param param = {
2361                 .min = (int *) table->extra1,
2362                 .max = (int *) table->extra2,
2363         };
2364         return do_proc_dointvec(table, write, buffer, lenp, ppos,
2365                                 do_proc_dointvec_minmax_conv, &param);
2366 }
2367
2368 static void validate_coredump_safety(void)
2369 {
2370 #ifdef CONFIG_COREDUMP
2371         if (suid_dumpable == SUID_DUMP_ROOT &&
2372             core_pattern[0] != '/' && core_pattern[0] != '|') {
2373                 printk(KERN_WARNING "Unsafe core_pattern used with "\
2374                         "suid_dumpable=2. Pipe handler or fully qualified "\
2375                         "core dump path required.\n");
2376         }
2377 #endif
2378 }
2379
2380 static int proc_dointvec_minmax_coredump(struct ctl_table *table, int write,
2381                 void __user *buffer, size_t *lenp, loff_t *ppos)
2382 {
2383         int error = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
2384         if (!error)
2385                 validate_coredump_safety();
2386         return error;
2387 }
2388
2389 #ifdef CONFIG_COREDUMP
2390 static int proc_dostring_coredump(struct ctl_table *table, int write,
2391                   void __user *buffer, size_t *lenp, loff_t *ppos)
2392 {
2393         int error = proc_dostring(table, write, buffer, lenp, ppos);
2394         if (!error)
2395                 validate_coredump_safety();
2396         return error;
2397 }
2398 #endif
2399
2400 static int __do_proc_doulongvec_minmax(void *data, struct ctl_table *table, int write,
2401                                      void __user *buffer,
2402                                      size_t *lenp, loff_t *ppos,
2403                                      unsigned long convmul,
2404                                      unsigned long convdiv)
2405 {
2406         unsigned long *i, *min, *max;
2407         int vleft, first = 1, err = 0;
2408         size_t left;
2409         char *kbuf = NULL, *p;
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                 p = kbuf = memdup_user_nul(buffer, left);
2438                 if (IS_ERR(kbuf))
2439                         return PTR_ERR(kbuf);
2440         }
2441
2442         for (; left && vleft--; i++, first = 0) {
2443                 unsigned long val;
2444
2445                 if (write) {
2446                         bool neg;
2447
2448                         left -= proc_skip_spaces(&p);
2449
2450                         err = proc_get_long(&p, &left, &val, &neg,
2451                                              proc_wspace_sep,
2452                                              sizeof(proc_wspace_sep), NULL);
2453                         if (err)
2454                                 break;
2455                         if (neg)
2456                                 continue;
2457                         if ((min && val < *min) || (max && val > *max))
2458                                 continue;
2459                         *i = val;
2460                 } else {
2461                         val = convdiv * (*i) / convmul;
2462                         if (!first) {
2463                                 err = proc_put_char(&buffer, &left, '\t');
2464                                 if (err)
2465                                         break;
2466                         }
2467                         err = proc_put_long(&buffer, &left, val, false);
2468                         if (err)
2469                                 break;
2470                 }
2471         }
2472
2473         if (!write && !first && left && !err)
2474                 err = proc_put_char(&buffer, &left, '\n');
2475         if (write && !err)
2476                 left -= proc_skip_spaces(&p);
2477         if (write) {
2478                 kfree(kbuf);
2479                 if (first)
2480                         return err ? : -EINVAL;
2481         }
2482         *lenp -= left;
2483 out:
2484         *ppos += *lenp;
2485         return err;
2486 }
2487
2488 static int do_proc_doulongvec_minmax(struct ctl_table *table, int write,
2489                                      void __user *buffer,
2490                                      size_t *lenp, loff_t *ppos,
2491                                      unsigned long convmul,
2492                                      unsigned long convdiv)
2493 {
2494         return __do_proc_doulongvec_minmax(table->data, table, write,
2495                         buffer, lenp, ppos, convmul, convdiv);
2496 }
2497
2498 /**
2499  * proc_doulongvec_minmax - read a vector of long integers with min/max values
2500  * @table: the sysctl table
2501  * @write: %TRUE if this is a write to the sysctl file
2502  * @buffer: the user buffer
2503  * @lenp: the size of the user buffer
2504  * @ppos: file position
2505  *
2506  * Reads/writes up to table->maxlen/sizeof(unsigned long) unsigned long
2507  * values from/to the user buffer, treated as an ASCII string.
2508  *
2509  * This routine will ensure the values are within the range specified by
2510  * table->extra1 (min) and table->extra2 (max).
2511  *
2512  * Returns 0 on success.
2513  */
2514 int proc_doulongvec_minmax(struct ctl_table *table, int write,
2515                            void __user *buffer, size_t *lenp, loff_t *ppos)
2516 {
2517     return do_proc_doulongvec_minmax(table, write, buffer, lenp, ppos, 1l, 1l);
2518 }
2519
2520 /**
2521  * proc_doulongvec_ms_jiffies_minmax - read a vector of millisecond values with min/max values
2522  * @table: the sysctl table
2523  * @write: %TRUE if this is a write to the sysctl file
2524  * @buffer: the user buffer
2525  * @lenp: the size of the user buffer
2526  * @ppos: file position
2527  *
2528  * Reads/writes up to table->maxlen/sizeof(unsigned long) unsigned long
2529  * values from/to the user buffer, treated as an ASCII string. The values
2530  * are treated as milliseconds, and converted to jiffies when they are stored.
2531  *
2532  * This routine will ensure the values are within the range specified by
2533  * table->extra1 (min) and table->extra2 (max).
2534  *
2535  * Returns 0 on success.
2536  */
2537 int proc_doulongvec_ms_jiffies_minmax(struct ctl_table *table, int write,
2538                                       void __user *buffer,
2539                                       size_t *lenp, loff_t *ppos)
2540 {
2541     return do_proc_doulongvec_minmax(table, write, buffer,
2542                                      lenp, ppos, HZ, 1000l);
2543 }
2544
2545
2546 static int do_proc_dointvec_jiffies_conv(bool *negp, unsigned long *lvalp,
2547                                          int *valp,
2548                                          int write, void *data)
2549 {
2550         if (write) {
2551                 if (*lvalp > LONG_MAX / HZ)
2552                         return 1;
2553                 *valp = *negp ? -(*lvalp*HZ) : (*lvalp*HZ);
2554         } else {
2555                 int val = *valp;
2556                 unsigned long lval;
2557                 if (val < 0) {
2558                         *negp = true;
2559                         lval = -(unsigned long)val;
2560                 } else {
2561                         *negp = false;
2562                         lval = (unsigned long)val;
2563                 }
2564                 *lvalp = lval / HZ;
2565         }
2566         return 0;
2567 }
2568
2569 static int do_proc_dointvec_userhz_jiffies_conv(bool *negp, unsigned long *lvalp,
2570                                                 int *valp,
2571                                                 int write, void *data)
2572 {
2573         if (write) {
2574                 if (USER_HZ < HZ && *lvalp > (LONG_MAX / HZ) * USER_HZ)
2575                         return 1;
2576                 *valp = clock_t_to_jiffies(*negp ? -*lvalp : *lvalp);
2577         } else {
2578                 int val = *valp;
2579                 unsigned long lval;
2580                 if (val < 0) {
2581                         *negp = true;
2582                         lval = -(unsigned long)val;
2583                 } else {
2584                         *negp = false;
2585                         lval = (unsigned long)val;
2586                 }
2587                 *lvalp = jiffies_to_clock_t(lval);
2588         }
2589         return 0;
2590 }
2591
2592 static int do_proc_dointvec_ms_jiffies_conv(bool *negp, unsigned long *lvalp,
2593                                             int *valp,
2594                                             int write, void *data)
2595 {
2596         if (write) {
2597                 unsigned long jif = msecs_to_jiffies(*negp ? -*lvalp : *lvalp);
2598
2599                 if (jif > INT_MAX)
2600                         return 1;
2601                 *valp = (int)jif;
2602         } else {
2603                 int val = *valp;
2604                 unsigned long lval;
2605                 if (val < 0) {
2606                         *negp = true;
2607                         lval = -(unsigned long)val;
2608                 } else {
2609                         *negp = false;
2610                         lval = (unsigned long)val;
2611                 }
2612                 *lvalp = jiffies_to_msecs(lval);
2613         }
2614         return 0;
2615 }
2616
2617 /**
2618  * proc_dointvec_jiffies - read a vector of integers as seconds
2619  * @table: the sysctl table
2620  * @write: %TRUE if this is a write to the sysctl file
2621  * @buffer: the user buffer
2622  * @lenp: the size of the user buffer
2623  * @ppos: file position
2624  *
2625  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2626  * values from/to the user buffer, treated as an ASCII string.
2627  * The values read are assumed to be in seconds, and are converted into
2628  * jiffies.
2629  *
2630  * Returns 0 on success.
2631  */
2632 int proc_dointvec_jiffies(struct ctl_table *table, int write,
2633                           void __user *buffer, size_t *lenp, loff_t *ppos)
2634 {
2635     return do_proc_dointvec(table,write,buffer,lenp,ppos,
2636                             do_proc_dointvec_jiffies_conv,NULL);
2637 }
2638
2639 /**
2640  * proc_dointvec_userhz_jiffies - read a vector of integers as 1/USER_HZ seconds
2641  * @table: the sysctl table
2642  * @write: %TRUE if this is a write to the sysctl file
2643  * @buffer: the user buffer
2644  * @lenp: the size of the user buffer
2645  * @ppos: pointer to the file position
2646  *
2647  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2648  * values from/to the user buffer, treated as an ASCII string.
2649  * The values read are assumed to be in 1/USER_HZ seconds, and
2650  * are converted into jiffies.
2651  *
2652  * Returns 0 on success.
2653  */
2654 int proc_dointvec_userhz_jiffies(struct ctl_table *table, int write,
2655                                  void __user *buffer, size_t *lenp, loff_t *ppos)
2656 {
2657     return do_proc_dointvec(table,write,buffer,lenp,ppos,
2658                             do_proc_dointvec_userhz_jiffies_conv,NULL);
2659 }
2660
2661 /**
2662  * proc_dointvec_ms_jiffies - read a vector of integers as 1 milliseconds
2663  * @table: the sysctl table
2664  * @write: %TRUE if this is a write to the sysctl file
2665  * @buffer: the user buffer
2666  * @lenp: the size of the user buffer
2667  * @ppos: file position
2668  * @ppos: the current position in the file
2669  *
2670  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2671  * values from/to the user buffer, treated as an ASCII string.
2672  * The values read are assumed to be in 1/1000 seconds, and
2673  * are converted into jiffies.
2674  *
2675  * Returns 0 on success.
2676  */
2677 int proc_dointvec_ms_jiffies(struct ctl_table *table, int write,
2678                              void __user *buffer, size_t *lenp, loff_t *ppos)
2679 {
2680         return do_proc_dointvec(table, write, buffer, lenp, ppos,
2681                                 do_proc_dointvec_ms_jiffies_conv, NULL);
2682 }
2683
2684 static int proc_do_cad_pid(struct ctl_table *table, int write,
2685                            void __user *buffer, size_t *lenp, loff_t *ppos)
2686 {
2687         struct pid *new_pid;
2688         pid_t tmp;
2689         int r;
2690
2691         tmp = pid_vnr(cad_pid);
2692
2693         r = __do_proc_dointvec(&tmp, table, write, buffer,
2694                                lenp, ppos, NULL, NULL);
2695         if (r || !write)
2696                 return r;
2697
2698         new_pid = find_get_pid(tmp);
2699         if (!new_pid)
2700                 return -ESRCH;
2701
2702         put_pid(xchg(&cad_pid, new_pid));
2703         return 0;
2704 }
2705
2706 /**
2707  * proc_do_large_bitmap - read/write from/to a large bitmap
2708  * @table: the sysctl table
2709  * @write: %TRUE if this is a write to the sysctl file
2710  * @buffer: the user buffer
2711  * @lenp: the size of the user buffer
2712  * @ppos: file position
2713  *
2714  * The bitmap is stored at table->data and the bitmap length (in bits)
2715  * in table->maxlen.
2716  *
2717  * We use a range comma separated format (e.g. 1,3-4,10-10) so that
2718  * large bitmaps may be represented in a compact manner. Writing into
2719  * the file will clear the bitmap then update it with the given input.
2720  *
2721  * Returns 0 on success.
2722  */
2723 int proc_do_large_bitmap(struct ctl_table *table, int write,
2724                          void __user *buffer, size_t *lenp, loff_t *ppos)
2725 {
2726         int err = 0;
2727         bool first = 1;
2728         size_t left = *lenp;
2729         unsigned long bitmap_len = table->maxlen;
2730         unsigned long *bitmap = *(unsigned long **) table->data;
2731         unsigned long *tmp_bitmap = NULL;
2732         char tr_a[] = { '-', ',', '\n' }, tr_b[] = { ',', '\n', 0 }, c;
2733
2734         if (!bitmap || !bitmap_len || !left || (*ppos && !write)) {
2735                 *lenp = 0;
2736                 return 0;
2737         }
2738
2739         if (write) {
2740                 char *kbuf, *p;
2741
2742                 if (left > PAGE_SIZE - 1)
2743                         left = PAGE_SIZE - 1;
2744
2745                 p = kbuf = memdup_user_nul(buffer, left);
2746                 if (IS_ERR(kbuf))
2747                         return PTR_ERR(kbuf);
2748
2749                 tmp_bitmap = kzalloc(BITS_TO_LONGS(bitmap_len) * sizeof(unsigned long),
2750                                      GFP_KERNEL);
2751                 if (!tmp_bitmap) {
2752                         kfree(kbuf);
2753                         return -ENOMEM;
2754                 }
2755                 proc_skip_char(&p, &left, '\n');
2756                 while (!err && left) {
2757                         unsigned long val_a, val_b;
2758                         bool neg;
2759
2760                         err = proc_get_long(&p, &left, &val_a, &neg, tr_a,
2761                                              sizeof(tr_a), &c);
2762                         if (err)
2763                                 break;
2764                         if (val_a >= bitmap_len || neg) {
2765                                 err = -EINVAL;
2766                                 break;
2767                         }
2768
2769                         val_b = val_a;
2770                         if (left) {
2771                                 p++;
2772                                 left--;
2773                         }
2774
2775                         if (c == '-') {
2776                                 err = proc_get_long(&p, &left, &val_b,
2777                                                      &neg, tr_b, sizeof(tr_b),
2778                                                      &c);
2779                                 if (err)
2780                                         break;
2781                                 if (val_b >= bitmap_len || neg ||
2782                                     val_a > val_b) {
2783                                         err = -EINVAL;
2784                                         break;
2785                                 }
2786                                 if (left) {
2787                                         p++;
2788                                         left--;
2789                                 }
2790                         }
2791
2792                         bitmap_set(tmp_bitmap, val_a, val_b - val_a + 1);
2793                         first = 0;
2794                         proc_skip_char(&p, &left, '\n');
2795                 }
2796                 kfree(kbuf);
2797         } else {
2798                 unsigned long bit_a, bit_b = 0;
2799
2800                 while (left) {
2801                         bit_a = find_next_bit(bitmap, bitmap_len, bit_b);
2802                         if (bit_a >= bitmap_len)
2803                                 break;
2804                         bit_b = find_next_zero_bit(bitmap, bitmap_len,
2805                                                    bit_a + 1) - 1;
2806
2807                         if (!first) {
2808                                 err = proc_put_char(&buffer, &left, ',');
2809                                 if (err)
2810                                         break;
2811                         }
2812                         err = proc_put_long(&buffer, &left, bit_a, false);
2813                         if (err)
2814                                 break;
2815                         if (bit_a != bit_b) {
2816                                 err = proc_put_char(&buffer, &left, '-');
2817                                 if (err)
2818                                         break;
2819                                 err = proc_put_long(&buffer, &left, bit_b, false);
2820                                 if (err)
2821                                         break;
2822                         }
2823
2824                         first = 0; bit_b++;
2825                 }
2826                 if (!err)
2827                         err = proc_put_char(&buffer, &left, '\n');
2828         }
2829
2830         if (!err) {
2831                 if (write) {
2832                         if (*ppos)
2833                                 bitmap_or(bitmap, bitmap, tmp_bitmap, bitmap_len);
2834                         else
2835                                 bitmap_copy(bitmap, tmp_bitmap, bitmap_len);
2836                 }
2837                 kfree(tmp_bitmap);
2838                 *lenp -= left;
2839                 *ppos += *lenp;
2840                 return 0;
2841         } else {
2842                 kfree(tmp_bitmap);
2843                 return err;
2844         }
2845 }
2846
2847 #else /* CONFIG_PROC_SYSCTL */
2848
2849 int proc_dostring(struct ctl_table *table, int write,
2850                   void __user *buffer, size_t *lenp, loff_t *ppos)
2851 {
2852         return -ENOSYS;
2853 }
2854
2855 int proc_dointvec(struct ctl_table *table, int write,
2856                   void __user *buffer, size_t *lenp, loff_t *ppos)
2857 {
2858         return -ENOSYS;
2859 }
2860
2861 int proc_dointvec_minmax(struct ctl_table *table, int write,
2862                     void __user *buffer, size_t *lenp, loff_t *ppos)
2863 {
2864         return -ENOSYS;
2865 }
2866
2867 int proc_dointvec_jiffies(struct ctl_table *table, int write,
2868                     void __user *buffer, size_t *lenp, loff_t *ppos)
2869 {
2870         return -ENOSYS;
2871 }
2872
2873 int proc_dointvec_userhz_jiffies(struct ctl_table *table, int write,
2874                     void __user *buffer, size_t *lenp, loff_t *ppos)
2875 {
2876         return -ENOSYS;
2877 }
2878
2879 int proc_dointvec_ms_jiffies(struct ctl_table *table, int write,
2880                              void __user *buffer, size_t *lenp, loff_t *ppos)
2881 {
2882         return -ENOSYS;
2883 }
2884
2885 int proc_doulongvec_minmax(struct ctl_table *table, int write,
2886                     void __user *buffer, size_t *lenp, loff_t *ppos)
2887 {
2888         return -ENOSYS;
2889 }
2890
2891 int proc_doulongvec_ms_jiffies_minmax(struct ctl_table *table, int write,
2892                                       void __user *buffer,
2893                                       size_t *lenp, loff_t *ppos)
2894 {
2895     return -ENOSYS;
2896 }
2897
2898
2899 #endif /* CONFIG_PROC_SYSCTL */
2900
2901 /*
2902  * No sense putting this after each symbol definition, twice,
2903  * exception granted :-)
2904  */
2905 EXPORT_SYMBOL(proc_dointvec);
2906 EXPORT_SYMBOL(proc_dointvec_jiffies);
2907 EXPORT_SYMBOL(proc_dointvec_minmax);
2908 EXPORT_SYMBOL(proc_dointvec_userhz_jiffies);
2909 EXPORT_SYMBOL(proc_dointvec_ms_jiffies);
2910 EXPORT_SYMBOL(proc_dostring);
2911 EXPORT_SYMBOL(proc_doulongvec_minmax);
2912 EXPORT_SYMBOL(proc_doulongvec_ms_jiffies_minmax);