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/mm.h>
  23 #include <linux/swap.h>
  24 #include <linux/slab.h>
  25 #include <linux/sysctl.h>
  26 #include <linux/bitmap.h>
  27 #include <linux/signal.h>
  28 #include <linux/printk.h>
  29 #include <linux/proc_fs.h>
  30 #include <linux/security.h>
  31 #include <linux/ctype.h>
  32 #include <linux/kmemcheck.h>
  33 #include <linux/kmemleak.h>
  34 #include <linux/fs.h>
  35 #include <linux/init.h>
  36 #include <linux/kernel.h>
  37 #include <linux/kobject.h>
  38 #include <linux/net.h>
  39 #include <linux/sysrq.h>
  40 #include <linux/highuid.h>
  41 #include <linux/writeback.h>
  42 #include <linux/ratelimit.h>
  43 #include <linux/compaction.h>
  44 #include <linux/hugetlb.h>
  45 #include <linux/initrd.h>
  46 #include <linux/key.h>
  47 #include <linux/times.h>
  48 #include <linux/limits.h>
  49 #include <linux/dcache.h>
  50 #include <linux/dnotify.h>
  51 #include <linux/syscalls.h>
  52 #include <linux/vmstat.h>
  53 #include <linux/nfs_fs.h>
  54 #include <linux/acpi.h>
  55 #include <linux/reboot.h>
  56 #include <linux/ftrace.h>
  57 #include <linux/perf_event.h>
  58 #include <linux/kprobes.h>
  59 #include <linux/pipe_fs_i.h>
  60 #include <linux/oom.h>
  61 #include <linux/kmod.h>
  62 #include <linux/capability.h>
  63 #include <linux/binfmts.h>
  64 #include <linux/sched/sysctl.h>
  65
  66 #include <asm/uaccess.h>
  67 #include <asm/processor.h>
  68
  69 #ifdef CONFIG_X86
  70 #include <asm/nmi.h>
  71 #include <asm/stacktrace.h>
  72 #include <asm/io.h>
  73 #endif
  74 #ifdef CONFIG_SPARC
  75 #include <asm/setup.h>
  76 #endif
  77 #ifdef CONFIG_BSD_PROCESS_ACCT
  78 #include <linux/acct.h>
  79 #endif
  80 #ifdef CONFIG_RT_MUTEXES
  81 #include <linux/rtmutex.h>
  82 #endif
  83 #if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_LOCK_STAT)
  84 #include <linux/lockdep.h>
  85 #endif
  86 #ifdef CONFIG_CHR_DEV_SG
  87 #include <scsi/sg.h>
  88 #endif
  89
  90 #ifdef CONFIG_LOCKUP_DETECTOR
  91 #include <linux/nmi.h>
  92 #endif
  93
  94
  95 #if defined(CONFIG_SYSCTL)
  96
  97 /* External variables not in a header file. */
  98 extern int sysctl_overcommit_memory;
  99 extern int sysctl_overcommit_ratio;
 100 extern int max_threads;
 101 extern int suid_dumpable;
 102 #ifdef CONFIG_COREDUMP
 103 extern int core_uses_pid;
 104 extern char core_pattern[];
 105 extern unsigned int core_pipe_limit;
 106 #endif
 107 extern int pid_max;
 108 extern int pid_max_min, pid_max_max;
 109 extern int percpu_pagelist_fraction;
 110 extern int compat_log;
 111 extern int latencytop_enabled;
 112 extern int sysctl_nr_open_min, sysctl_nr_open_max;
 113 #ifndef CONFIG_MMU
 114 extern int sysctl_nr_trim_pages;
 115 #endif
 116 #ifdef CONFIG_BLOCK
 117 extern int blk_iopoll_enabled;
 118 #endif
 119
 120 /* Constants used for minimum and  maximum */
 121 #ifdef CONFIG_LOCKUP_DETECTOR
 122 static int sixty = 60;
 123 #endif
 124
 125 static int zero;
 126 static int __maybe_unused one = 1;
 127 static int __maybe_unused two = 2;
 128 static int __maybe_unused three = 3;
 129 static unsigned long one_ul = 1;
 130 static int one_hundred = 100;
 131 #ifdef CONFIG_PRINTK
 132 static int ten_thousand = 10000;
 133 #endif
 134
 135 /* this is needed for the proc_doulongvec_minmax of vm_dirty_bytes */
 136 static unsigned long dirty_bytes_min = 2 * PAGE_SIZE;
 137
 138 /* this is needed for the proc_dointvec_minmax for [fs_]overflow UID and GID */
 139 static int maxolduid = 65535;
 140 static int minolduid;
 141 static int min_percpu_pagelist_fract = 8;
 142
 143 static int ngroups_max = NGROUPS_MAX;
 144 static const int cap_last_cap = CAP_LAST_CAP;
 145
 146 #ifdef CONFIG_INOTIFY_USER
 147 #include <linux/inotify.h>
 148 #endif
 149 #ifdef CONFIG_SPARC
 150 #endif
 151
 152 #ifdef CONFIG_SPARC64
 153 extern int sysctl_tsb_ratio;
 154 #endif
 155
 156 #ifdef __hppa__
 157 extern int pwrsw_enabled;
 158 #endif
 159
 160 #ifdef CONFIG_SYSCTL_ARCH_UNALIGN_ALLOW
 161 extern int unaligned_enabled;
 162 #endif
 163
 164 #ifdef CONFIG_IA64
 165 extern int unaligned_dump_stack;
 166 #endif
 167
 168 #ifdef CONFIG_SYSCTL_ARCH_UNALIGN_NO_WARN
 169 extern int no_unaligned_warning;
 170 #endif
 171
 172 #ifdef CONFIG_PROC_SYSCTL
 173 static int proc_do_cad_pid(struct ctl_table *table, int write,
 174                   void __user *buffer, size_t *lenp, loff_t *ppos);
 175 static int proc_taint(struct ctl_table *table, int write,
 176                                void __user *buffer, size_t *lenp, loff_t *ppos);
 177 #endif
 178
 179 #ifdef CONFIG_PRINTK
 180 static int proc_dointvec_minmax_sysadmin(struct ctl_table *table, int write,
 181                                 void __user *buffer, size_t *lenp, loff_t *ppos);
 182 #endif
 183
 184 static int proc_dointvec_minmax_coredump(struct ctl_table *table, int write,
 185                 void __user *buffer, size_t *lenp, loff_t *ppos);
 186 #ifdef CONFIG_COREDUMP
 187 static int proc_dostring_coredump(struct ctl_table *table, int write,
 188                 void __user *buffer, size_t *lenp, loff_t *ppos);
 189 #endif
 190
 191 #ifdef CONFIG_MAGIC_SYSRQ
 192 /* Note: sysrq code uses it's own private copy */
 193 static int __sysrq_enabled = CONFIG_MAGIC_SYSRQ_DEFAULT_ENABLE;
 194
 195 static int sysrq_sysctl_handler(ctl_table *table, int write,
 196                                 void __user *buffer, size_t *lenp,
 197                                 loff_t *ppos)
 198 {
 199         int error;
 200
 201         error = proc_dointvec(table, write, buffer, lenp, ppos);
 202         if (error)
 203                 return error;
 204
 205         if (write)
 206                 sysrq_toggle_support(__sysrq_enabled);
 207
 208         return 0;
 209 }
 210
 211 #endif
 212
 213 static struct ctl_table kern_table[];
 214 static struct ctl_table vm_table[];
 215 static struct ctl_table fs_table[];
 216 static struct ctl_table debug_table[];
 217 static struct ctl_table dev_table[];
 218 extern struct ctl_table random_table[];
 219 #ifdef CONFIG_EPOLL
 220 extern struct ctl_table epoll_table[];
 221 #endif
 222
 223 #ifdef HAVE_ARCH_PICK_MMAP_LAYOUT
 224 int sysctl_legacy_va_layout;
 225 #endif
 226
 227 /* The default sysctl tables: */
 228
 229 static struct ctl_table sysctl_base_table[] = {
 230         {
 231                 .procname       = "kernel",
 232                 .mode           = 0555,
 233                 .child          = kern_table,
 234         },
 235         {
 236                 .procname       = "vm",
 237                 .mode           = 0555,
 238                 .child          = vm_table,
 239         },
 240         {
 241                 .procname       = "fs",
 242                 .mode           = 0555,
 243                 .child          = fs_table,
 244         },
 245         {
 246                 .procname       = "debug",
 247                 .mode           = 0555,
 248                 .child          = debug_table,
 249         },
 250         {
 251                 .procname       = "dev",
 252                 .mode           = 0555,
 253                 .child          = dev_table,
 254         },
 255         { }
 256 };
 257
 258 #ifdef CONFIG_SCHED_DEBUG
 259 static int min_sched_granularity_ns = 100000;           /* 100 usecs */
 260 static int max_sched_granularity_ns = NSEC_PER_SEC;     /* 1 second */
 261 static int min_wakeup_granularity_ns;                   /* 0 usecs */
 262 static int max_wakeup_granularity_ns = NSEC_PER_SEC;    /* 1 second */
 263 #ifdef CONFIG_SMP
 264 static int min_sched_tunable_scaling = SCHED_TUNABLESCALING_NONE;
 265 static int max_sched_tunable_scaling = SCHED_TUNABLESCALING_END-1;
 266 #endif /* CONFIG_SMP */
 267 #endif /* CONFIG_SCHED_DEBUG */
 268
 269 #ifdef CONFIG_COMPACTION
 270 static int min_extfrag_threshold;
 271 static int max_extfrag_threshold = 1000;
 272 #endif
 273
 274 static struct ctl_table kern_table[] = {
 275         {
 276                 .procname       = "sched_child_runs_first",
 277                 .data           = &sysctl_sched_child_runs_first,
 278                 .maxlen         = sizeof(unsigned int),
 279                 .mode           = 0644,
 280                 .proc_handler   = proc_dointvec,
 281         },
 282 #ifdef CONFIG_SCHED_DEBUG
 283         {
 284                 .procname       = "sched_min_granularity_ns",
 285                 .data           = &sysctl_sched_min_granularity,
 286                 .maxlen         = sizeof(unsigned int),
 287                 .mode           = 0644,
 288                 .proc_handler   = sched_proc_update_handler,
 289                 .extra1         = &min_sched_granularity_ns,
 290                 .extra2         = &max_sched_granularity_ns,
 291         },
 292         {
 293                 .procname       = "sched_latency_ns",
 294                 .data           = &sysctl_sched_latency,
 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_wakeup_granularity_ns",
 303                 .data           = &sysctl_sched_wakeup_granularity,
 304                 .maxlen         = sizeof(unsigned int),
 305                 .mode           = 0644,
 306                 .proc_handler   = sched_proc_update_handler,
 307                 .extra1         = &min_wakeup_granularity_ns,
 308                 .extra2         = &max_wakeup_granularity_ns,
 309         },
 310 #ifdef CONFIG_SMP
 311         {
 312                 .procname       = "sched_tunable_scaling",
 313                 .data           = &sysctl_sched_tunable_scaling,
 314                 .maxlen         = sizeof(enum sched_tunable_scaling),
 315                 .mode           = 0644,
 316                 .proc_handler   = sched_proc_update_handler,
 317                 .extra1         = &min_sched_tunable_scaling,
 318                 .extra2         = &max_sched_tunable_scaling,
 319         },
 320         {
 321                 .procname       = "sched_migration_cost_ns",
 322                 .data           = &sysctl_sched_migration_cost,
 323                 .maxlen         = sizeof(unsigned int),
 324                 .mode           = 0644,
 325                 .proc_handler   = proc_dointvec,
 326         },
 327         {
 328                 .procname       = "sched_nr_migrate",
 329                 .data           = &sysctl_sched_nr_migrate,
 330                 .maxlen         = sizeof(unsigned int),
 331                 .mode           = 0644,
 332                 .proc_handler   = proc_dointvec,
 333         },
 334         {
 335                 .procname       = "sched_time_avg_ms",
 336                 .data           = &sysctl_sched_time_avg,
 337                 .maxlen         = sizeof(unsigned int),
 338                 .mode           = 0644,
 339                 .proc_handler   = proc_dointvec,
 340         },
 341         {
 342                 .procname       = "sched_shares_window_ns",
 343                 .data           = &sysctl_sched_shares_window,
 344                 .maxlen         = sizeof(unsigned int),
 345                 .mode           = 0644,
 346                 .proc_handler   = proc_dointvec,
 347         },
 348         {
 349                 .procname       = "timer_migration",
 350                 .data           = &sysctl_timer_migration,
 351                 .maxlen         = sizeof(unsigned int),
 352                 .mode           = 0644,
 353                 .proc_handler   = proc_dointvec_minmax,
 354                 .extra1         = &zero,
 355                 .extra2         = &one,
 356         },
 357 #endif /* CONFIG_SMP */
 358 #ifdef CONFIG_NUMA_BALANCING
 359         {
 360                 .procname       = "numa_balancing_scan_delay_ms",
 361                 .data           = &sysctl_numa_balancing_scan_delay,
 362                 .maxlen         = sizeof(unsigned int),
 363                 .mode           = 0644,
 364                 .proc_handler   = proc_dointvec,
 365         },
 366         {
 367                 .procname       = "numa_balancing_scan_period_min_ms",
 368                 .data           = &sysctl_numa_balancing_scan_period_min,
 369                 .maxlen         = sizeof(unsigned int),
 370                 .mode           = 0644,
 371                 .proc_handler   = proc_dointvec,
 372         },
 373         {
 374                 .procname       = "numa_balancing_scan_period_max_ms",
 375                 .data           = &sysctl_numa_balancing_scan_period_max,
 376                 .maxlen         = sizeof(unsigned int),
 377                 .mode           = 0644,
 378                 .proc_handler   = proc_dointvec,
 379         },
 380         {
 381                 .procname       = "numa_balancing_scan_size_mb",
 382                 .data           = &sysctl_numa_balancing_scan_size,
 383                 .maxlen         = sizeof(unsigned int),
 384                 .mode           = 0644,
 385                 .proc_handler   = proc_dointvec,
 386         },
 387 #endif /* CONFIG_NUMA_BALANCING */
 388 #endif /* CONFIG_SCHED_DEBUG */
 389         {
 390                 .procname       = "sched_rt_period_us",
 391                 .data           = &sysctl_sched_rt_period,
 392                 .maxlen         = sizeof(unsigned int),
 393                 .mode           = 0644,
 394                 .proc_handler   = sched_rt_handler,
 395         },
 396         {
 397                 .procname       = "sched_rt_runtime_us",
 398                 .data           = &sysctl_sched_rt_runtime,
 399                 .maxlen         = sizeof(int),
 400                 .mode           = 0644,
 401                 .proc_handler   = sched_rt_handler,
 402         },
 403         {
 404                 .procname       = "sched_rr_timeslice_ms",
 405                 .data           = &sched_rr_timeslice,
 406                 .maxlen         = sizeof(int),
 407                 .mode           = 0644,
 408                 .proc_handler   = sched_rr_handler,
 409         },
 410 #ifdef CONFIG_SCHED_AUTOGROUP
 411         {
 412                 .procname       = "sched_autogroup_enabled",
 413                 .data           = &sysctl_sched_autogroup_enabled,
 414                 .maxlen         = sizeof(unsigned int),
 415                 .mode           = 0644,
 416                 .proc_handler   = proc_dointvec_minmax,
 417                 .extra1         = &zero,
 418                 .extra2         = &one,
 419         },
 420 #endif
 421 #ifdef CONFIG_CFS_BANDWIDTH
 422         {
 423                 .procname       = "sched_cfs_bandwidth_slice_us",
 424                 .data           = &sysctl_sched_cfs_bandwidth_slice,
 425                 .maxlen         = sizeof(unsigned int),
 426                 .mode           = 0644,
 427                 .proc_handler   = proc_dointvec_minmax,
 428                 .extra1         = &one,
 429         },
 430 #endif
 431 #ifdef CONFIG_PROVE_LOCKING
 432         {
 433                 .procname       = "prove_locking",
 434                 .data           = &prove_locking,
 435                 .maxlen         = sizeof(int),
 436                 .mode           = 0644,
 437                 .proc_handler   = proc_dointvec,
 438         },
 439 #endif
 440 #ifdef CONFIG_LOCK_STAT
 441         {
 442                 .procname       = "lock_stat",
 443                 .data           = &lock_stat,
 444                 .maxlen         = sizeof(int),
 445                 .mode           = 0644,
 446                 .proc_handler   = proc_dointvec,
 447         },
 448 #endif
 449         {
 450                 .procname       = "panic",
 451                 .data           = &panic_timeout,
 452                 .maxlen         = sizeof(int),
 453                 .mode           = 0644,
 454                 .proc_handler   = proc_dointvec,
 455         },
 456 #ifdef CONFIG_COREDUMP
 457         {
 458                 .procname       = "core_uses_pid",
 459                 .data           = &core_uses_pid,
 460                 .maxlen         = sizeof(int),
 461                 .mode           = 0644,
 462                 .proc_handler   = proc_dointvec,
 463         },
 464         {
 465                 .procname       = "core_pattern",
 466                 .data           = core_pattern,
 467                 .maxlen         = CORENAME_MAX_SIZE,
 468                 .mode           = 0644,
 469                 .proc_handler   = proc_dostring_coredump,
 470         },
 471         {
 472                 .procname       = "core_pipe_limit",
 473                 .data           = &core_pipe_limit,
 474                 .maxlen         = sizeof(unsigned int),
 475                 .mode           = 0644,
 476                 .proc_handler   = proc_dointvec,
 477         },
 478 #endif
 479 #ifdef CONFIG_PROC_SYSCTL
 480         {
 481                 .procname       = "tainted",
 482                 .maxlen         = sizeof(long),
 483                 .mode           = 0644,
 484                 .proc_handler   = proc_taint,
 485         },
 486 #endif
 487 #ifdef CONFIG_LATENCYTOP
 488         {
 489                 .procname       = "latencytop",
 490                 .data           = &latencytop_enabled,
 491                 .maxlen         = sizeof(int),
 492                 .mode           = 0644,
 493                 .proc_handler   = proc_dointvec,
 494         },
 495 #endif
 496 #ifdef CONFIG_BLK_DEV_INITRD
 497         {
 498                 .procname       = "real-root-dev",
 499                 .data           = &real_root_dev,
 500                 .maxlen         = sizeof(int),
 501                 .mode           = 0644,
 502                 .proc_handler   = proc_dointvec,
 503         },
 504 #endif
 505         {
 506                 .procname       = "print-fatal-signals",
 507                 .data           = &print_fatal_signals,
 508                 .maxlen         = sizeof(int),
 509                 .mode           = 0644,
 510                 .proc_handler   = proc_dointvec,
 511         },
 512 #ifdef CONFIG_SPARC
 513         {
 514                 .procname       = "reboot-cmd",
 515                 .data           = reboot_command,
 516                 .maxlen         = 256,
 517                 .mode           = 0644,
 518                 .proc_handler   = proc_dostring,
 519         },
 520         {
 521                 .procname       = "stop-a",
 522                 .data           = &stop_a_enabled,
 523                 .maxlen         = sizeof (int),
 524                 .mode           = 0644,
 525                 .proc_handler   = proc_dointvec,
 526         },
 527         {
 528                 .procname       = "scons-poweroff",
 529                 .data           = &scons_pwroff,
 530                 .maxlen         = sizeof (int),
 531                 .mode           = 0644,
 532                 .proc_handler   = proc_dointvec,
 533         },
 534 #endif
 535 #ifdef CONFIG_SPARC64
 536         {
 537                 .procname       = "tsb-ratio",
 538                 .data           = &sysctl_tsb_ratio,
 539                 .maxlen         = sizeof (int),
 540                 .mode           = 0644,
 541                 .proc_handler   = proc_dointvec,
 542         },
 543 #endif
 544 #ifdef __hppa__
 545         {
 546                 .procname       = "soft-power",
 547                 .data           = &pwrsw_enabled,
 548                 .maxlen         = sizeof (int),
 549                 .mode           = 0644,
 550                 .proc_handler   = proc_dointvec,
 551         },
 552 #endif
 553 #ifdef CONFIG_SYSCTL_ARCH_UNALIGN_ALLOW
 554         {
 555                 .procname       = "unaligned-trap",
 556                 .data           = &unaligned_enabled,
 557                 .maxlen         = sizeof (int),
 558                 .mode           = 0644,
 559                 .proc_handler   = proc_dointvec,
 560         },
 561 #endif
 562         {
 563                 .procname       = "ctrl-alt-del",
 564                 .data           = &C_A_D,
 565                 .maxlen         = sizeof(int),
 566                 .mode           = 0644,
 567                 .proc_handler   = proc_dointvec,
 568         },
 569 #ifdef CONFIG_FUNCTION_TRACER
 570         {
 571                 .procname       = "ftrace_enabled",
 572                 .data           = &ftrace_enabled,
 573                 .maxlen         = sizeof(int),
 574                 .mode           = 0644,
 575                 .proc_handler   = ftrace_enable_sysctl,
 576         },
 577 #endif
 578 #ifdef CONFIG_STACK_TRACER
 579         {
 580                 .procname       = "stack_tracer_enabled",
 581                 .data           = &stack_tracer_enabled,
 582                 .maxlen         = sizeof(int),
 583                 .mode           = 0644,
 584                 .proc_handler   = stack_trace_sysctl,
 585         },
 586 #endif
 587 #ifdef CONFIG_TRACING
 588         {
 589                 .procname       = "ftrace_dump_on_oops",
 590                 .data           = &ftrace_dump_on_oops,
 591                 .maxlen         = sizeof(int),
 592                 .mode           = 0644,
 593                 .proc_handler   = proc_dointvec,
 594         },
 595         {
 596                 .procname       = "traceoff_on_warning",
 597                 .data           = &__disable_trace_on_warning,
 598                 .maxlen         = sizeof(__disable_trace_on_warning),
 599                 .mode           = 0644,
 600                 .proc_handler   = proc_dointvec,
 601         },
 602 #endif
 603 #ifdef CONFIG_MODULES
 604         {
 605                 .procname       = "modprobe",
 606                 .data           = &modprobe_path,
 607                 .maxlen         = KMOD_PATH_LEN,
 608                 .mode           = 0644,
 609                 .proc_handler   = proc_dostring,
 610         },
 611         {
 612                 .procname       = "modules_disabled",
 613                 .data           = &modules_disabled,
 614                 .maxlen         = sizeof(int),
 615                 .mode           = 0644,
 616                 /* only handle a transition from default "0" to "1" */
 617                 .proc_handler   = proc_dointvec_minmax,
 618                 .extra1         = &one,
 619                 .extra2         = &one,
 620         },
 621 #endif
 622
 623         {
 624                 .procname       = "hotplug",
 625                 .data           = &uevent_helper,
 626                 .maxlen         = UEVENT_HELPER_PATH_LEN,
 627                 .mode           = 0644,
 628                 .proc_handler   = proc_dostring,
 629         },
 630
 631 #ifdef CONFIG_CHR_DEV_SG
 632         {
 633                 .procname       = "sg-big-buff",
 634                 .data           = &sg_big_buff,
 635                 .maxlen         = sizeof (int),
 636                 .mode           = 0444,
 637                 .proc_handler   = proc_dointvec,
 638         },
 639 #endif
 640 #ifdef CONFIG_BSD_PROCESS_ACCT
 641         {
 642                 .procname       = "acct",
 643                 .data           = &acct_parm,
 644                 .maxlen         = 3*sizeof(int),
 645                 .mode           = 0644,
 646                 .proc_handler   = proc_dointvec,
 647         },
 648 #endif
 649 #ifdef CONFIG_MAGIC_SYSRQ
 650         {
 651                 .procname       = "sysrq",
 652                 .data           = &__sysrq_enabled,
 653                 .maxlen         = sizeof (int),
 654                 .mode           = 0644,
 655                 .proc_handler   = sysrq_sysctl_handler,
 656         },
 657 #endif
 658 #ifdef CONFIG_PROC_SYSCTL
 659         {
 660                 .procname       = "cad_pid",
 661                 .data           = NULL,
 662                 .maxlen         = sizeof (int),
 663                 .mode           = 0600,
 664                 .proc_handler   = proc_do_cad_pid,
 665         },
 666 #endif
 667         {
 668                 .procname       = "threads-max",
 669                 .data           = &max_threads,
 670                 .maxlen         = sizeof(int),
 671                 .mode           = 0644,
 672                 .proc_handler   = proc_dointvec,
 673         },
 674         {
 675                 .procname       = "random",
 676                 .mode           = 0555,
 677                 .child          = random_table,
 678         },
 679         {
 680                 .procname       = "usermodehelper",
 681                 .mode           = 0555,
 682                 .child          = usermodehelper_table,
 683         },
 684         {
 685                 .procname       = "overflowuid",
 686                 .data           = &overflowuid,
 687                 .maxlen         = sizeof(int),
 688                 .mode           = 0644,
 689                 .proc_handler   = proc_dointvec_minmax,
 690                 .extra1         = &minolduid,
 691                 .extra2         = &maxolduid,
 692         },
 693         {
 694                 .procname       = "overflowgid",
 695                 .data           = &overflowgid,
 696                 .maxlen         = sizeof(int),
 697                 .mode           = 0644,
 698                 .proc_handler   = proc_dointvec_minmax,
 699                 .extra1         = &minolduid,
 700                 .extra2         = &maxolduid,
 701         },
 702 #ifdef CONFIG_S390
 703 #ifdef CONFIG_MATHEMU
 704         {
 705                 .procname       = "ieee_emulation_warnings",
 706                 .data           = &sysctl_ieee_emulation_warnings,
 707                 .maxlen         = sizeof(int),
 708                 .mode           = 0644,
 709                 .proc_handler   = proc_dointvec,
 710         },
 711 #endif
 712         {
 713                 .procname       = "userprocess_debug",
 714                 .data           = &show_unhandled_signals,
 715                 .maxlen         = sizeof(int),
 716                 .mode           = 0644,
 717                 .proc_handler   = proc_dointvec,
 718         },
 719 #endif
 720         {
 721                 .procname       = "pid_max",
 722                 .data           = &pid_max,
 723                 .maxlen         = sizeof (int),
 724                 .mode           = 0644,
 725                 .proc_handler   = proc_dointvec_minmax,
 726                 .extra1         = &pid_max_min,
 727                 .extra2         = &pid_max_max,
 728         },
 729         {
 730                 .procname       = "panic_on_oops",
 731                 .data           = &panic_on_oops,
 732                 .maxlen         = sizeof(int),
 733                 .mode           = 0644,
 734                 .proc_handler   = proc_dointvec,
 735         },
 736 #if defined CONFIG_PRINTK
 737         {
 738                 .procname       = "printk",
 739                 .data           = &console_loglevel,
 740                 .maxlen         = 4*sizeof(int),
 741                 .mode           = 0644,
 742                 .proc_handler   = proc_dointvec,
 743         },
 744         {
 745                 .procname       = "printk_ratelimit",
 746                 .data           = &printk_ratelimit_state.interval,
 747                 .maxlen         = sizeof(int),
 748                 .mode           = 0644,
 749                 .proc_handler   = proc_dointvec_jiffies,
 750         },
 751         {
 752                 .procname       = "printk_ratelimit_burst",
 753                 .data           = &printk_ratelimit_state.burst,
 754                 .maxlen         = sizeof(int),
 755                 .mode           = 0644,
 756                 .proc_handler   = proc_dointvec,
 757         },
 758         {
 759                 .procname       = "printk_delay",
 760                 .data           = &printk_delay_msec,
 761                 .maxlen         = sizeof(int),
 762                 .mode           = 0644,
 763                 .proc_handler   = proc_dointvec_minmax,
 764                 .extra1         = &zero,
 765                 .extra2         = &ten_thousand,
 766         },
 767         {
 768                 .procname       = "dmesg_restrict",
 769                 .data           = &dmesg_restrict,
 770                 .maxlen         = sizeof(int),
 771                 .mode           = 0644,
 772                 .proc_handler   = proc_dointvec_minmax_sysadmin,
 773                 .extra1         = &zero,
 774                 .extra2         = &one,
 775         },
 776         {
 777                 .procname       = "kptr_restrict",
 778                 .data           = &kptr_restrict,
 779                 .maxlen         = sizeof(int),
 780                 .mode           = 0644,
 781                 .proc_handler   = proc_dointvec_minmax_sysadmin,
 782                 .extra1         = &zero,
 783                 .extra2         = &two,
 784         },
 785 #endif
 786         {
 787                 .procname       = "ngroups_max",
 788                 .data           = &ngroups_max,
 789                 .maxlen         = sizeof (int),
 790                 .mode           = 0444,
 791                 .proc_handler   = proc_dointvec,
 792         },
 793         {
 794                 .procname       = "cap_last_cap",
 795                 .data           = (void *)&cap_last_cap,
 796                 .maxlen         = sizeof(int),
 797                 .mode           = 0444,
 798                 .proc_handler   = proc_dointvec,
 799         },
 800 #if defined(CONFIG_LOCKUP_DETECTOR)
 801         {
 802                 .procname       = "watchdog",
 803                 .data           = &watchdog_user_enabled,
 804                 .maxlen         = sizeof (int),
 805                 .mode           = 0644,
 806                 .proc_handler   = proc_dowatchdog,
 807                 .extra1         = &zero,
 808                 .extra2         = &one,
 809         },
 810         {
 811                 .procname       = "watchdog_thresh",
 812                 .data           = &watchdog_thresh,
 813                 .maxlen         = sizeof(int),
 814                 .mode           = 0644,
 815                 .proc_handler   = proc_dowatchdog,
 816                 .extra1         = &zero,
 817                 .extra2         = &sixty,
 818         },
 819         {
 820                 .procname       = "softlockup_panic",
 821                 .data           = &softlockup_panic,
 822                 .maxlen         = sizeof(int),
 823                 .mode           = 0644,
 824                 .proc_handler   = proc_dointvec_minmax,
 825                 .extra1         = &zero,
 826                 .extra2         = &one,
 827         },
 828         {
 829                 .procname       = "nmi_watchdog",
 830                 .data           = &watchdog_user_enabled,
 831                 .maxlen         = sizeof (int),
 832                 .mode           = 0644,
 833                 .proc_handler   = proc_dowatchdog,
 834                 .extra1         = &zero,
 835                 .extra2         = &one,
 836         },
 837 #endif
 838 #if defined(CONFIG_X86_LOCAL_APIC) && defined(CONFIG_X86)
 839         {
 840                 .procname       = "unknown_nmi_panic",
 841                 .data           = &unknown_nmi_panic,
 842                 .maxlen         = sizeof (int),
 843                 .mode           = 0644,
 844                 .proc_handler   = proc_dointvec,
 845         },
 846 #endif
 847 #if defined(CONFIG_X86)
 848         {
 849                 .procname       = "panic_on_unrecovered_nmi",
 850                 .data           = &panic_on_unrecovered_nmi,
 851                 .maxlen         = sizeof(int),
 852                 .mode           = 0644,
 853                 .proc_handler   = proc_dointvec,
 854         },
 855         {
 856                 .procname       = "panic_on_io_nmi",
 857                 .data           = &panic_on_io_nmi,
 858                 .maxlen         = sizeof(int),
 859                 .mode           = 0644,
 860                 .proc_handler   = proc_dointvec,
 861         },
 862 #ifdef CONFIG_DEBUG_STACKOVERFLOW
 863         {
 864                 .procname       = "panic_on_stackoverflow",
 865                 .data           = &sysctl_panic_on_stackoverflow,
 866                 .maxlen         = sizeof(int),
 867                 .mode           = 0644,
 868                 .proc_handler   = proc_dointvec,
 869         },
 870 #endif
 871         {
 872                 .procname       = "bootloader_type",
 873                 .data           = &bootloader_type,
 874                 .maxlen         = sizeof (int),
 875                 .mode           = 0444,
 876                 .proc_handler   = proc_dointvec,
 877         },
 878         {
 879                 .procname       = "bootloader_version",
 880                 .data           = &bootloader_version,
 881                 .maxlen         = sizeof (int),
 882                 .mode           = 0444,
 883                 .proc_handler   = proc_dointvec,
 884         },
 885         {
 886                 .procname       = "kstack_depth_to_print",
 887                 .data           = &kstack_depth_to_print,
 888                 .maxlen         = sizeof(int),
 889                 .mode           = 0644,
 890                 .proc_handler   = proc_dointvec,
 891         },
 892         {
 893                 .procname       = "io_delay_type",
 894                 .data           = &io_delay_type,
 895                 .maxlen         = sizeof(int),
 896                 .mode           = 0644,
 897                 .proc_handler   = proc_dointvec,
 898         },
 899 #endif
 900 #if defined(CONFIG_MMU)
 901         {
 902                 .procname       = "randomize_va_space",
 903                 .data           = &randomize_va_space,
 904                 .maxlen         = sizeof(int),
 905                 .mode           = 0644,
 906                 .proc_handler   = proc_dointvec,
 907         },
 908 #endif
 909 #if defined(CONFIG_S390) && defined(CONFIG_SMP)
 910         {
 911                 .procname       = "spin_retry",
 912                 .data           = &spin_retry,
 913                 .maxlen         = sizeof (int),
 914                 .mode           = 0644,
 915                 .proc_handler   = proc_dointvec,
 916         },
 917 #endif
 918 #if     defined(CONFIG_ACPI_SLEEP) && defined(CONFIG_X86)
 919         {
 920                 .procname       = "acpi_video_flags",
 921                 .data           = &acpi_realmode_flags,
 922                 .maxlen         = sizeof (unsigned long),
 923                 .mode           = 0644,
 924                 .proc_handler   = proc_doulongvec_minmax,
 925         },
 926 #endif
 927 #ifdef CONFIG_SYSCTL_ARCH_UNALIGN_NO_WARN
 928         {
 929                 .procname       = "ignore-unaligned-usertrap",
 930                 .data           = &no_unaligned_warning,
 931                 .maxlen         = sizeof (int),
 932                 .mode           = 0644,
 933                 .proc_handler   = proc_dointvec,
 934         },
 935 #endif
 936 #ifdef CONFIG_IA64
 937         {
 938                 .procname       = "unaligned-dump-stack",
 939                 .data           = &unaligned_dump_stack,
 940                 .maxlen         = sizeof (int),
 941                 .mode           = 0644,
 942                 .proc_handler   = proc_dointvec,
 943         },
 944 #endif
 945 #ifdef CONFIG_DETECT_HUNG_TASK
 946         {
 947                 .procname       = "hung_task_panic",
 948                 .data           = &sysctl_hung_task_panic,
 949                 .maxlen         = sizeof(int),
 950                 .mode           = 0644,
 951                 .proc_handler   = proc_dointvec_minmax,
 952                 .extra1         = &zero,
 953                 .extra2         = &one,
 954         },
 955         {
 956                 .procname       = "hung_task_check_count",
 957                 .data           = &sysctl_hung_task_check_count,
 958                 .maxlen         = sizeof(int),
 959                 .mode           = 0644,
 960                 .proc_handler   = proc_dointvec_minmax,
 961                 .extra1         = &zero,
 962         },
 963         {
 964                 .procname       = "hung_task_timeout_secs",
 965                 .data           = &sysctl_hung_task_timeout_secs,
 966                 .maxlen         = sizeof(unsigned long),
 967                 .mode           = 0644,
 968                 .proc_handler   = proc_dohung_task_timeout_secs,
 969         },
 970         {
 971                 .procname       = "hung_task_warnings",
 972                 .data           = &sysctl_hung_task_warnings,
 973                 .maxlen         = sizeof(unsigned long),
 974                 .mode           = 0644,
 975                 .proc_handler   = proc_doulongvec_minmax,
 976         },
 977 #endif
 978 #ifdef CONFIG_COMPAT
 979         {
 980                 .procname       = "compat-log",
 981                 .data           = &compat_log,
 982                 .maxlen         = sizeof (int),
 983                 .mode           = 0644,
 984                 .proc_handler   = proc_dointvec,
 985         },
 986 #endif
 987 #ifdef CONFIG_RT_MUTEXES
 988         {
 989                 .procname       = "max_lock_depth",
 990                 .data           = &max_lock_depth,
 991                 .maxlen         = sizeof(int),
 992                 .mode           = 0644,
 993                 .proc_handler   = proc_dointvec,
 994         },
 995 #endif
 996         {
 997                 .procname       = "poweroff_cmd",
 998                 .data           = &poweroff_cmd,
 999                 .maxlen         = POWEROFF_CMD_PATH_LEN,
1000                 .mode           = 0644,
1001                 .proc_handler   = proc_dostring,
1002         },
1003 #ifdef CONFIG_KEYS
1004         {
1005                 .procname       = "keys",
1006                 .mode           = 0555,
1007                 .child          = key_sysctls,
1008         },
1009 #endif
1010 #ifdef CONFIG_RCU_TORTURE_TEST
1011         {
1012                 .procname       = "rcutorture_runnable",
1013                 .data           = &rcutorture_runnable,
1014                 .maxlen         = sizeof(int),
1015                 .mode           = 0644,
1016                 .proc_handler   = proc_dointvec,
1017         },
1018 #endif
1019 #ifdef CONFIG_PERF_EVENTS
1020         /*
1021          * User-space scripts rely on the existence of this file
1022          * as a feature check for perf_events being enabled.
1023          *
1024          * So it's an ABI, do not remove!
1025          */
1026         {
1027                 .procname       = "perf_event_paranoid",
1028                 .data           = &sysctl_perf_event_paranoid,
1029                 .maxlen         = sizeof(sysctl_perf_event_paranoid),
1030                 .mode           = 0644,
1031                 .proc_handler   = proc_dointvec,
1032         },
1033         {
1034                 .procname       = "perf_event_mlock_kb",
1035                 .data           = &sysctl_perf_event_mlock,
1036                 .maxlen         = sizeof(sysctl_perf_event_mlock),
1037                 .mode           = 0644,
1038                 .proc_handler   = proc_dointvec,
1039         },
1040         {
1041                 .procname       = "perf_event_max_sample_rate",
1042                 .data           = &sysctl_perf_event_sample_rate,
1043                 .maxlen         = sizeof(sysctl_perf_event_sample_rate),
1044                 .mode           = 0644,
1045                 .proc_handler   = perf_proc_update_handler,
1046                 .extra1         = &one,
1047         },
1048         {
1049                 .procname       = "perf_cpu_time_max_percent",
1050                 .data           = &sysctl_perf_cpu_time_max_percent,
1051                 .maxlen         = sizeof(sysctl_perf_cpu_time_max_percent),
1052                 .mode           = 0644,
1053                 .proc_handler   = perf_cpu_time_max_percent_handler,
1054                 .extra1         = &zero,
1055                 .extra2         = &one_hundred,
1056         },
1057 #endif
1058 #ifdef CONFIG_KMEMCHECK
1059         {
1060                 .procname       = "kmemcheck",
1061                 .data           = &kmemcheck_enabled,
1062                 .maxlen         = sizeof(int),
1063                 .mode           = 0644,
1064                 .proc_handler   = proc_dointvec,
1065         },
1066 #endif
1067 #ifdef CONFIG_BLOCK
1068         {
1069                 .procname       = "blk_iopoll",
1070                 .data           = &blk_iopoll_enabled,
1071                 .maxlen         = sizeof(int),
1072                 .mode           = 0644,
1073                 .proc_handler   = proc_dointvec,
1074         },
1075 #endif
1076         { }
1077 };
1078
1079 static struct ctl_table vm_table[] = {
1080         {
1081                 .procname       = "overcommit_memory",
1082                 .data           = &sysctl_overcommit_memory,
1083                 .maxlen         = sizeof(sysctl_overcommit_memory),
1084                 .mode           = 0644,
1085                 .proc_handler   = proc_dointvec_minmax,
1086                 .extra1         = &zero,
1087                 .extra2         = &two,
1088         },
1089         {
1090                 .procname       = "panic_on_oom",
1091                 .data           = &sysctl_panic_on_oom,
1092                 .maxlen         = sizeof(sysctl_panic_on_oom),
1093                 .mode           = 0644,
1094                 .proc_handler   = proc_dointvec_minmax,
1095                 .extra1         = &zero,
1096                 .extra2         = &two,
1097         },
1098         {
1099                 .procname       = "oom_kill_allocating_task",
1100                 .data           = &sysctl_oom_kill_allocating_task,
1101                 .maxlen         = sizeof(sysctl_oom_kill_allocating_task),
1102                 .mode           = 0644,
1103                 .proc_handler   = proc_dointvec,
1104         },
1105         {
1106                 .procname       = "oom_dump_tasks",
1107                 .data           = &sysctl_oom_dump_tasks,
1108                 .maxlen         = sizeof(sysctl_oom_dump_tasks),
1109                 .mode           = 0644,
1110                 .proc_handler   = proc_dointvec,
1111         },
1112         {
1113                 .procname       = "overcommit_ratio",
1114                 .data           = &sysctl_overcommit_ratio,
1115                 .maxlen         = sizeof(sysctl_overcommit_ratio),
1116                 .mode           = 0644,
1117                 .proc_handler   = proc_dointvec,
1118         },
1119         {
1120                 .procname       = "page-cluster",
1121                 .data           = &page_cluster,
1122                 .maxlen         = sizeof(int),
1123                 .mode           = 0644,
1124                 .proc_handler   = proc_dointvec_minmax,
1125                 .extra1         = &zero,
1126         },
1127         {
1128                 .procname       = "dirty_background_ratio",
1129                 .data           = &dirty_background_ratio,
1130                 .maxlen         = sizeof(dirty_background_ratio),
1131                 .mode           = 0644,
1132                 .proc_handler   = dirty_background_ratio_handler,
1133                 .extra1         = &zero,
1134                 .extra2         = &one_hundred,
1135         },
1136         {
1137                 .procname       = "dirty_background_bytes",
1138                 .data           = &dirty_background_bytes,
1139                 .maxlen         = sizeof(dirty_background_bytes),
1140                 .mode           = 0644,
1141                 .proc_handler   = dirty_background_bytes_handler,
1142                 .extra1         = &one_ul,
1143         },
1144         {
1145                 .procname       = "dirty_ratio",
1146                 .data           = &vm_dirty_ratio,
1147                 .maxlen         = sizeof(vm_dirty_ratio),
1148                 .mode           = 0644,
1149                 .proc_handler   = dirty_ratio_handler,
1150                 .extra1         = &zero,
1151                 .extra2         = &one_hundred,
1152         },
1153         {
1154                 .procname       = "dirty_bytes",
1155                 .data           = &vm_dirty_bytes,
1156                 .maxlen         = sizeof(vm_dirty_bytes),
1157                 .mode           = 0644,
1158                 .proc_handler   = dirty_bytes_handler,
1159                 .extra1         = &dirty_bytes_min,
1160         },
1161         {
1162                 .procname       = "dirty_writeback_centisecs",
1163                 .data           = &dirty_writeback_interval,
1164                 .maxlen         = sizeof(dirty_writeback_interval),
1165                 .mode           = 0644,
1166                 .proc_handler   = dirty_writeback_centisecs_handler,
1167         },
1168         {
1169                 .procname       = "dirty_expire_centisecs",
1170                 .data           = &dirty_expire_interval,
1171                 .maxlen         = sizeof(dirty_expire_interval),
1172                 .mode           = 0644,
1173                 .proc_handler   = proc_dointvec_minmax,
1174                 .extra1         = &zero,
1175         },
1176         {
1177                 .procname       = "nr_pdflush_threads",
1178                 .mode           = 0444 /* read-only */,
1179                 .proc_handler   = pdflush_proc_obsolete,
1180         },
1181         {
1182                 .procname       = "swappiness",
1183                 .data           = &vm_swappiness,
1184                 .maxlen         = sizeof(vm_swappiness),
1185                 .mode           = 0644,
1186                 .proc_handler   = proc_dointvec_minmax,
1187                 .extra1         = &zero,
1188                 .extra2         = &one_hundred,
1189         },
1190 #ifdef CONFIG_HUGETLB_PAGE
1191         {
1192                 .procname       = "nr_hugepages",
1193                 .data           = NULL,
1194                 .maxlen         = sizeof(unsigned long),
1195                 .mode           = 0644,
1196                 .proc_handler   = hugetlb_sysctl_handler,
1197                 .extra1         = (void *)&hugetlb_zero,
1198                 .extra2         = (void *)&hugetlb_infinity,
1199         },
1200 #ifdef CONFIG_NUMA
1201         {
1202                 .procname       = "nr_hugepages_mempolicy",
1203                 .data           = NULL,
1204                 .maxlen         = sizeof(unsigned long),
1205                 .mode           = 0644,
1206                 .proc_handler   = &hugetlb_mempolicy_sysctl_handler,
1207                 .extra1         = (void *)&hugetlb_zero,
1208                 .extra2         = (void *)&hugetlb_infinity,
1209         },
1210 #endif
1211          {
1212                 .procname       = "hugetlb_shm_group",
1213                 .data           = &sysctl_hugetlb_shm_group,
1214                 .maxlen         = sizeof(gid_t),
1215                 .mode           = 0644,
1216                 .proc_handler   = proc_dointvec,
1217          },
1218          {
1219                 .procname       = "hugepages_treat_as_movable",
1220                 .data           = &hugepages_treat_as_movable,
1221                 .maxlen         = sizeof(int),
1222                 .mode           = 0644,
1223                 .proc_handler   = proc_dointvec,
1224         },
1225         {
1226                 .procname       = "nr_overcommit_hugepages",
1227                 .data           = NULL,
1228                 .maxlen         = sizeof(unsigned long),
1229                 .mode           = 0644,
1230                 .proc_handler   = hugetlb_overcommit_handler,
1231                 .extra1         = (void *)&hugetlb_zero,
1232                 .extra2         = (void *)&hugetlb_infinity,
1233         },
1234 #endif
1235         {
1236                 .procname       = "lowmem_reserve_ratio",
1237                 .data           = &sysctl_lowmem_reserve_ratio,
1238                 .maxlen         = sizeof(sysctl_lowmem_reserve_ratio),
1239                 .mode           = 0644,
1240                 .proc_handler   = lowmem_reserve_ratio_sysctl_handler,
1241         },
1242         {
1243                 .procname       = "drop_caches",
1244                 .data           = &sysctl_drop_caches,
1245                 .maxlen         = sizeof(int),
1246                 .mode           = 0644,
1247                 .proc_handler   = drop_caches_sysctl_handler,
1248                 .extra1         = &one,
1249                 .extra2         = &three,
1250         },
1251 #ifdef CONFIG_COMPACTION
1252         {
1253                 .procname       = "compact_memory",
1254                 .data           = &sysctl_compact_memory,
1255                 .maxlen         = sizeof(int),
1256                 .mode           = 0200,
1257                 .proc_handler   = sysctl_compaction_handler,
1258         },
1259         {
1260                 .procname       = "extfrag_threshold",
1261                 .data           = &sysctl_extfrag_threshold,
1262                 .maxlen         = sizeof(int),
1263                 .mode           = 0644,
1264                 .proc_handler   = sysctl_extfrag_handler,
1265                 .extra1         = &min_extfrag_threshold,
1266                 .extra2         = &max_extfrag_threshold,
1267         },
1268
1269 #endif /* CONFIG_COMPACTION */
1270         {
1271                 .procname       = "min_free_kbytes",
1272                 .data           = &min_free_kbytes,
1273                 .maxlen         = sizeof(min_free_kbytes),
1274                 .mode           = 0644,
1275                 .proc_handler   = min_free_kbytes_sysctl_handler,
1276                 .extra1         = &zero,
1277         },
1278         {
1279                 .procname       = "percpu_pagelist_fraction",
1280                 .data           = &percpu_pagelist_fraction,
1281                 .maxlen         = sizeof(percpu_pagelist_fraction),
1282                 .mode           = 0644,
1283                 .proc_handler   = percpu_pagelist_fraction_sysctl_handler,
1284                 .extra1         = &min_percpu_pagelist_fract,
1285         },
1286 #ifdef CONFIG_MMU
1287         {
1288                 .procname       = "max_map_count",
1289                 .data           = &sysctl_max_map_count,
1290                 .maxlen         = sizeof(sysctl_max_map_count),
1291                 .mode           = 0644,
1292                 .proc_handler   = proc_dointvec_minmax,
1293                 .extra1         = &zero,
1294         },
1295 #else
1296         {
1297                 .procname       = "nr_trim_pages",
1298                 .data           = &sysctl_nr_trim_pages,
1299                 .maxlen         = sizeof(sysctl_nr_trim_pages),
1300                 .mode           = 0644,
1301                 .proc_handler   = proc_dointvec_minmax,
1302                 .extra1         = &zero,
1303         },
1304 #endif
1305         {
1306                 .procname       = "laptop_mode",
1307                 .data           = &laptop_mode,
1308                 .maxlen         = sizeof(laptop_mode),
1309                 .mode           = 0644,
1310                 .proc_handler   = proc_dointvec_jiffies,
1311         },
1312         {
1313                 .procname       = "block_dump",
1314                 .data           = &block_dump,
1315                 .maxlen         = sizeof(block_dump),
1316                 .mode           = 0644,
1317                 .proc_handler   = proc_dointvec,
1318                 .extra1         = &zero,
1319         },
1320         {
1321                 .procname       = "vfs_cache_pressure",
1322                 .data           = &sysctl_vfs_cache_pressure,
1323                 .maxlen         = sizeof(sysctl_vfs_cache_pressure),
1324                 .mode           = 0644,
1325                 .proc_handler   = proc_dointvec,
1326                 .extra1         = &zero,
1327         },
1328 #ifdef HAVE_ARCH_PICK_MMAP_LAYOUT
1329         {
1330                 .procname       = "legacy_va_layout",
1331                 .data           = &sysctl_legacy_va_layout,
1332                 .maxlen         = sizeof(sysctl_legacy_va_layout),
1333                 .mode           = 0644,
1334                 .proc_handler   = proc_dointvec,
1335                 .extra1         = &zero,
1336         },
1337 #endif
1338 #ifdef CONFIG_NUMA
1339         {
1340                 .procname       = "zone_reclaim_mode",
1341                 .data           = &zone_reclaim_mode,
1342                 .maxlen         = sizeof(zone_reclaim_mode),
1343                 .mode           = 0644,
1344                 .proc_handler   = proc_dointvec,
1345                 .extra1         = &zero,
1346         },
1347         {
1348                 .procname       = "min_unmapped_ratio",
1349                 .data           = &sysctl_min_unmapped_ratio,
1350                 .maxlen         = sizeof(sysctl_min_unmapped_ratio),
1351                 .mode           = 0644,
1352                 .proc_handler   = sysctl_min_unmapped_ratio_sysctl_handler,
1353                 .extra1         = &zero,
1354                 .extra2         = &one_hundred,
1355         },
1356         {
1357                 .procname       = "min_slab_ratio",
1358                 .data           = &sysctl_min_slab_ratio,
1359                 .maxlen         = sizeof(sysctl_min_slab_ratio),
1360                 .mode           = 0644,
1361                 .proc_handler   = sysctl_min_slab_ratio_sysctl_handler,
1362                 .extra1         = &zero,
1363                 .extra2         = &one_hundred,
1364         },
1365 #endif
1366 #ifdef CONFIG_SMP
1367         {
1368                 .procname       = "stat_interval",
1369                 .data           = &sysctl_stat_interval,
1370                 .maxlen         = sizeof(sysctl_stat_interval),
1371                 .mode           = 0644,
1372                 .proc_handler   = proc_dointvec_jiffies,
1373         },
1374 #endif
1375 #ifdef CONFIG_MMU
1376         {
1377                 .procname       = "mmap_min_addr",
1378                 .data           = &dac_mmap_min_addr,
1379                 .maxlen         = sizeof(unsigned long),
1380                 .mode           = 0644,
1381                 .proc_handler   = mmap_min_addr_handler,
1382         },
1383 #endif
1384 #ifdef CONFIG_NUMA
1385         {
1386                 .procname       = "numa_zonelist_order",
1387                 .data           = &numa_zonelist_order,
1388                 .maxlen         = NUMA_ZONELIST_ORDER_LEN,
1389                 .mode           = 0644,
1390                 .proc_handler   = numa_zonelist_order_handler,
1391         },
1392 #endif
1393 #if (defined(CONFIG_X86_32) && !defined(CONFIG_UML))|| \
1394    (defined(CONFIG_SUPERH) && defined(CONFIG_VSYSCALL))
1395         {
1396                 .procname       = "vdso_enabled",
1397                 .data           = &vdso_enabled,
1398                 .maxlen         = sizeof(vdso_enabled),
1399                 .mode           = 0644,
1400                 .proc_handler   = proc_dointvec,
1401                 .extra1         = &zero,
1402         },
1403 #endif
1404 #ifdef CONFIG_HIGHMEM
1405         {
1406                 .procname       = "highmem_is_dirtyable",
1407                 .data           = &vm_highmem_is_dirtyable,
1408                 .maxlen         = sizeof(vm_highmem_is_dirtyable),
1409                 .mode           = 0644,
1410                 .proc_handler   = proc_dointvec_minmax,
1411                 .extra1         = &zero,
1412                 .extra2         = &one,
1413         },
1414 #endif
1415         {
1416                 .procname       = "scan_unevictable_pages",
1417                 .data           = &scan_unevictable_pages,
1418                 .maxlen         = sizeof(scan_unevictable_pages),
1419                 .mode           = 0644,
1420                 .proc_handler   = scan_unevictable_handler,
1421         },
1422 #ifdef CONFIG_MEMORY_FAILURE
1423         {
1424                 .procname       = "memory_failure_early_kill",
1425                 .data           = &sysctl_memory_failure_early_kill,
1426                 .maxlen         = sizeof(sysctl_memory_failure_early_kill),
1427                 .mode           = 0644,
1428                 .proc_handler   = proc_dointvec_minmax,
1429                 .extra1         = &zero,
1430                 .extra2         = &one,
1431         },
1432         {
1433                 .procname       = "memory_failure_recovery",
1434                 .data           = &sysctl_memory_failure_recovery,
1435                 .maxlen         = sizeof(sysctl_memory_failure_recovery),
1436                 .mode           = 0644,
1437                 .proc_handler   = proc_dointvec_minmax,
1438                 .extra1         = &zero,
1439                 .extra2         = &one,
1440         },
1441 #endif
1442         {
1443                 .procname       = "user_reserve_kbytes",
1444                 .data           = &sysctl_user_reserve_kbytes,
1445                 .maxlen         = sizeof(sysctl_user_reserve_kbytes),
1446                 .mode           = 0644,
1447                 .proc_handler   = proc_doulongvec_minmax,
1448         },
1449         {
1450                 .procname       = "admin_reserve_kbytes",
1451                 .data           = &sysctl_admin_reserve_kbytes,
1452                 .maxlen         = sizeof(sysctl_admin_reserve_kbytes),
1453                 .mode           = 0644,
1454                 .proc_handler   = proc_doulongvec_minmax,
1455         },
1456         { }
1457 };
1458
1459 #if defined(CONFIG_BINFMT_MISC) || defined(CONFIG_BINFMT_MISC_MODULE)
1460 static struct ctl_table binfmt_misc_table[] = {
1461         { }
1462 };
1463 #endif
1464
1465 static struct ctl_table fs_table[] = {
1466         {
1467                 .procname       = "inode-nr",
1468                 .data           = &inodes_stat,
1469                 .maxlen         = 2*sizeof(long),
1470                 .mode           = 0444,
1471                 .proc_handler   = proc_nr_inodes,
1472         },
1473         {
1474                 .procname       = "inode-state",
1475                 .data           = &inodes_stat,
1476                 .maxlen         = 7*sizeof(long),
1477                 .mode           = 0444,
1478                 .proc_handler   = proc_nr_inodes,
1479         },
1480         {
1481                 .procname       = "file-nr",
1482                 .data           = &files_stat,
1483                 .maxlen         = sizeof(files_stat),
1484                 .mode           = 0444,
1485                 .proc_handler   = proc_nr_files,
1486         },
1487         {
1488                 .procname       = "file-max",
1489                 .data           = &files_stat.max_files,
1490                 .maxlen         = sizeof(files_stat.max_files),
1491                 .mode           = 0644,
1492                 .proc_handler   = proc_doulongvec_minmax,
1493         },
1494         {
1495                 .procname       = "nr_open",
1496                 .data           = &sysctl_nr_open,
1497                 .maxlen         = sizeof(int),
1498                 .mode           = 0644,
1499                 .proc_handler   = proc_dointvec_minmax,
1500                 .extra1         = &sysctl_nr_open_min,
1501                 .extra2         = &sysctl_nr_open_max,
1502         },
1503         {
1504                 .procname       = "dentry-state",
1505                 .data           = &dentry_stat,
1506                 .maxlen         = 6*sizeof(long),
1507                 .mode           = 0444,
1508                 .proc_handler   = proc_nr_dentry,
1509         },
1510         {
1511                 .procname       = "overflowuid",
1512                 .data           = &fs_overflowuid,
1513                 .maxlen         = sizeof(int),
1514                 .mode           = 0644,
1515                 .proc_handler   = proc_dointvec_minmax,
1516                 .extra1         = &minolduid,
1517                 .extra2         = &maxolduid,
1518         },
1519         {
1520                 .procname       = "overflowgid",
1521                 .data           = &fs_overflowgid,
1522                 .maxlen         = sizeof(int),
1523                 .mode           = 0644,
1524                 .proc_handler   = proc_dointvec_minmax,
1525                 .extra1         = &minolduid,
1526                 .extra2         = &maxolduid,
1527         },
1528 #ifdef CONFIG_FILE_LOCKING
1529         {
1530                 .procname       = "leases-enable",
1531                 .data           = &leases_enable,
1532                 .maxlen         = sizeof(int),
1533                 .mode           = 0644,
1534                 .proc_handler   = proc_dointvec,
1535         },
1536 #endif
1537 #ifdef CONFIG_DNOTIFY
1538         {
1539                 .procname       = "dir-notify-enable",
1540                 .data           = &dir_notify_enable,
1541                 .maxlen         = sizeof(int),
1542                 .mode           = 0644,
1543                 .proc_handler   = proc_dointvec,
1544         },
1545 #endif
1546 #ifdef CONFIG_MMU
1547 #ifdef CONFIG_FILE_LOCKING
1548         {
1549                 .procname       = "lease-break-time",
1550                 .data           = &lease_break_time,
1551                 .maxlen         = sizeof(int),
1552                 .mode           = 0644,
1553                 .proc_handler   = proc_dointvec,
1554         },
1555 #endif
1556 #ifdef CONFIG_AIO
1557         {
1558                 .procname       = "aio-nr",
1559                 .data           = &aio_nr,
1560                 .maxlen         = sizeof(aio_nr),
1561                 .mode           = 0444,
1562                 .proc_handler   = proc_doulongvec_minmax,
1563         },
1564         {
1565                 .procname       = "aio-max-nr",
1566                 .data           = &aio_max_nr,
1567                 .maxlen         = sizeof(aio_max_nr),
1568                 .mode           = 0644,
1569                 .proc_handler   = proc_doulongvec_minmax,
1570         },
1571 #endif /* CONFIG_AIO */
1572 #ifdef CONFIG_INOTIFY_USER
1573         {
1574                 .procname       = "inotify",
1575                 .mode           = 0555,
1576                 .child          = inotify_table,
1577         },
1578 #endif
1579 #ifdef CONFIG_EPOLL
1580         {
1581                 .procname       = "epoll",
1582                 .mode           = 0555,
1583                 .child          = epoll_table,
1584         },
1585 #endif
1586 #endif
1587         {
1588                 .procname       = "protected_symlinks",
1589                 .data           = &sysctl_protected_symlinks,
1590                 .maxlen         = sizeof(int),
1591                 .mode           = 0600,
1592                 .proc_handler   = proc_dointvec_minmax,
1593                 .extra1         = &zero,
1594                 .extra2         = &one,
1595         },
1596         {
1597                 .procname       = "protected_hardlinks",
1598                 .data           = &sysctl_protected_hardlinks,
1599                 .maxlen         = sizeof(int),
1600                 .mode           = 0600,
1601                 .proc_handler   = proc_dointvec_minmax,
1602                 .extra1         = &zero,
1603                 .extra2         = &one,
1604         },
1605         {
1606                 .procname       = "suid_dumpable",
1607                 .data           = &suid_dumpable,
1608                 .maxlen         = sizeof(int),
1609                 .mode           = 0644,
1610                 .proc_handler   = proc_dointvec_minmax_coredump,
1611                 .extra1         = &zero,
1612                 .extra2         = &two,
1613         },
1614 #if defined(CONFIG_BINFMT_MISC) || defined(CONFIG_BINFMT_MISC_MODULE)
1615         {
1616                 .procname       = "binfmt_misc",
1617                 .mode           = 0555,
1618                 .child          = binfmt_misc_table,
1619         },
1620 #endif
1621         {
1622                 .procname       = "pipe-max-size",
1623                 .data           = &pipe_max_size,
1624                 .maxlen         = sizeof(int),
1625                 .mode           = 0644,
1626                 .proc_handler   = &pipe_proc_fn,
1627                 .extra1         = &pipe_min_size,
1628         },
1629         { }
1630 };
1631
1632 static struct ctl_table debug_table[] = {
1633 #ifdef CONFIG_SYSCTL_EXCEPTION_TRACE
1634         {
1635                 .procname       = "exception-trace",
1636                 .data           = &show_unhandled_signals,
1637                 .maxlen         = sizeof(int),
1638                 .mode           = 0644,
1639                 .proc_handler   = proc_dointvec
1640         },
1641 #endif
1642 #if defined(CONFIG_OPTPROBES)
1643         {
1644                 .procname       = "kprobes-optimization",
1645                 .data           = &sysctl_kprobes_optimization,
1646                 .maxlen         = sizeof(int),
1647                 .mode           = 0644,
1648                 .proc_handler   = proc_kprobes_optimization_handler,
1649                 .extra1         = &zero,
1650                 .extra2         = &one,
1651         },
1652 #endif
1653         { }
1654 };
1655
1656 static struct ctl_table dev_table[] = {
1657         { }
1658 };
1659
1660 int __init sysctl_init(void)
1661 {
1662         struct ctl_table_header *hdr;
1663
1664         hdr = register_sysctl_table(sysctl_base_table);
1665         kmemleak_not_leak(hdr);
1666         return 0;
1667 }
1668
1669 #endif /* CONFIG_SYSCTL */
1670
1671 /*
1672  * /proc/sys support
1673  */
1674
1675 #ifdef CONFIG_PROC_SYSCTL
1676
1677 static int _proc_do_string(void* data, int maxlen, int write,
1678                            void __user *buffer,
1679                            size_t *lenp, loff_t *ppos)
1680 {
1681         size_t len;
1682         char __user *p;
1683         char c;
1684
1685         if (!data || !maxlen || !*lenp) {
1686                 *lenp = 0;
1687                 return 0;
1688         }
1689
1690         if (write) {
1691                 len = 0;
1692                 p = buffer;
1693                 while (len < *lenp) {
1694                         if (get_user(c, p++))
1695                                 return -EFAULT;
1696                         if (c == 0 || c == '\n')
1697                                 break;
1698                         len++;
1699                 }
1700                 if (len >= maxlen)
1701                         len = maxlen-1;
1702                 if(copy_from_user(data, buffer, len))
1703                         return -EFAULT;
1704                 ((char *) data)[len] = 0;
1705                 *ppos += *lenp;
1706         } else {
1707                 len = strlen(data);
1708                 if (len > maxlen)
1709                         len = maxlen;
1710
1711                 if (*ppos > len) {
1712                         *lenp = 0;
1713                         return 0;
1714                 }
1715
1716                 data += *ppos;
1717                 len  -= *ppos;
1718
1719                 if (len > *lenp)
1720                         len = *lenp;
1721                 if (len)
1722                         if(copy_to_user(buffer, data, len))
1723                                 return -EFAULT;
1724                 if (len < *lenp) {
1725                         if(put_user('\n', ((char __user *) buffer) + len))
1726                                 return -EFAULT;
1727                         len++;
1728                 }
1729                 *lenp = len;
1730                 *ppos += len;
1731         }
1732         return 0;
1733 }
1734
1735 /**
1736  * proc_dostring - read a string sysctl
1737  * @table: the sysctl table
1738  * @write: %TRUE if this is a write to the sysctl file
1739  * @buffer: the user buffer
1740  * @lenp: the size of the user buffer
1741  * @ppos: file position
1742  *
1743  * Reads/writes a string from/to the user buffer. If the kernel
1744  * buffer provided is not large enough to hold the string, the
1745  * string is truncated. The copied string is %NULL-terminated.
1746  * If the string is being read by the user process, it is copied
1747  * and a newline '\n' is added. It is truncated if the buffer is
1748  * not large enough.
1749  *
1750  * Returns 0 on success.
1751  */
1752 int proc_dostring(struct ctl_table *table, int write,
1753                   void __user *buffer, size_t *lenp, loff_t *ppos)
1754 {
1755         return _proc_do_string(table->data, table->maxlen, write,
1756                                buffer, lenp, ppos);
1757 }
1758
1759 static size_t proc_skip_spaces(char **buf)
1760 {
1761         size_t ret;
1762         char *tmp = skip_spaces(*buf);
1763         ret = tmp - *buf;
1764         *buf = tmp;
1765         return ret;
1766 }
1767
1768 static void proc_skip_char(char **buf, size_t *size, const char v)
1769 {
1770         while (*size) {
1771                 if (**buf != v)
1772                         break;
1773                 (*size)--;
1774                 (*buf)++;
1775         }
1776 }
1777
1778 #define TMPBUFLEN 22
1779 /**
1780  * proc_get_long - reads an ASCII formatted integer from a user buffer
1781  *
1782  * @buf: a kernel buffer
1783  * @size: size of the kernel buffer
1784  * @val: this is where the number will be stored
1785  * @neg: set to %TRUE if number is negative
1786  * @perm_tr: a vector which contains the allowed trailers
1787  * @perm_tr_len: size of the perm_tr vector
1788  * @tr: pointer to store the trailer character
1789  *
1790  * In case of success %0 is returned and @buf and @size are updated with
1791  * the amount of bytes read. If @tr is non-NULL and a trailing
1792  * character exists (size is non-zero after returning from this
1793  * function), @tr is updated with the trailing character.
1794  */
1795 static int proc_get_long(char **buf, size_t *size,
1796                           unsigned long *val, bool *neg,
1797                           const char *perm_tr, unsigned perm_tr_len, char *tr)
1798 {
1799         int len;
1800         char *p, tmp[TMPBUFLEN];
1801
1802         if (!*size)
1803                 return -EINVAL;
1804
1805         len = *size;
1806         if (len > TMPBUFLEN - 1)
1807                 len = TMPBUFLEN - 1;
1808
1809         memcpy(tmp, *buf, len);
1810
1811         tmp[len] = 0;
1812         p = tmp;
1813         if (*p == '-' && *size > 1) {
1814                 *neg = true;
1815                 p++;
1816         } else
1817                 *neg = false;
1818         if (!isdigit(*p))
1819                 return -EINVAL;
1820
1821         *val = simple_strtoul(p, &p, 0);
1822
1823         len = p - tmp;
1824
1825         /* We don't know if the next char is whitespace thus we may accept
1826          * invalid integers (e.g. 1234...a) or two integers instead of one
1827          * (e.g. 123...1). So lets not allow such large numbers. */
1828         if (len == TMPBUFLEN - 1)
1829                 return -EINVAL;
1830
1831         if (len < *size && perm_tr_len && !memchr(perm_tr, *p, perm_tr_len))
1832                 return -EINVAL;
1833
1834         if (tr && (len < *size))
1835                 *tr = *p;
1836
1837         *buf += len;
1838         *size -= len;
1839
1840         return 0;
1841 }
1842
1843 /**
1844  * proc_put_long - converts an integer to a decimal ASCII formatted string
1845  *
1846  * @buf: the user buffer
1847  * @size: the size of the user buffer
1848  * @val: the integer to be converted
1849  * @neg: sign of the number, %TRUE for negative
1850  *
1851  * In case of success %0 is returned and @buf and @size are updated with
1852  * the amount of bytes written.
1853  */
1854 static int proc_put_long(void __user **buf, size_t *size, unsigned long val,
1855                           bool neg)
1856 {
1857         int len;
1858         char tmp[TMPBUFLEN], *p = tmp;
1859
1860         sprintf(p, "%s%lu", neg ? "-" : "", val);
1861         len = strlen(tmp);
1862         if (len > *size)
1863                 len = *size;
1864         if (copy_to_user(*buf, tmp, len))
1865                 return -EFAULT;
1866         *size -= len;
1867         *buf += len;
1868         return 0;
1869 }
1870 #undef TMPBUFLEN
1871
1872 static int proc_put_char(void __user **buf, size_t *size, char c)
1873 {
1874         if (*size) {
1875                 char __user **buffer = (char __user **)buf;
1876                 if (put_user(c, *buffer))
1877                         return -EFAULT;
1878                 (*size)--, (*buffer)++;
1879                 *buf = *buffer;
1880         }
1881         return 0;
1882 }
1883
1884 static int do_proc_dointvec_conv(bool *negp, unsigned long *lvalp,
1885                                  int *valp,
1886                                  int write, void *data)
1887 {
1888         if (write) {
1889                 *valp = *negp ? -*lvalp : *lvalp;
1890         } else {
1891                 int val = *valp;
1892                 if (val < 0) {
1893                         *negp = true;
1894                         *lvalp = (unsigned long)-val;
1895                 } else {
1896                         *negp = false;
1897                         *lvalp = (unsigned long)val;
1898                 }
1899         }
1900         return 0;
1901 }
1902
1903 static const char proc_wspace_sep[] = { ' ', '\t', '\n' };
1904
1905 static int __do_proc_dointvec(void *tbl_data, struct ctl_table *table,
1906                   int write, void __user *buffer,
1907                   size_t *lenp, loff_t *ppos,
1908                   int (*conv)(bool *negp, unsigned long *lvalp, int *valp,
1909                               int write, void *data),
1910                   void *data)
1911 {
1912         int *i, vleft, first = 1, err = 0;
1913         unsigned long page = 0;
1914         size_t left;
1915         char *kbuf;
1916
1917         if (!tbl_data || !table->maxlen || !*lenp || (*ppos && !write)) {
1918                 *lenp = 0;
1919                 return 0;
1920         }
1921
1922         i = (int *) tbl_data;
1923         vleft = table->maxlen / sizeof(*i);
1924         left = *lenp;
1925
1926         if (!conv)
1927                 conv = do_proc_dointvec_conv;
1928
1929         if (write) {
1930                 if (left > PAGE_SIZE - 1)
1931                         left = PAGE_SIZE - 1;
1932                 page = __get_free_page(GFP_TEMPORARY);
1933                 kbuf = (char *) page;
1934                 if (!kbuf)
1935                         return -ENOMEM;
1936                 if (copy_from_user(kbuf, buffer, left)) {
1937                         err = -EFAULT;
1938                         goto free;
1939                 }
1940                 kbuf[left] = 0;
1941         }
1942
1943         for (; left && vleft--; i++, first=0) {
1944                 unsigned long lval;
1945                 bool neg;
1946
1947                 if (write) {
1948                         left -= proc_skip_spaces(&kbuf);
1949
1950                         if (!left)
1951                                 break;
1952                         err = proc_get_long(&kbuf, &left, &lval, &neg,
1953                                              proc_wspace_sep,
1954                                              sizeof(proc_wspace_sep), NULL);
1955                         if (err)
1956                                 break;
1957                         if (conv(&neg, &lval, i, 1, data)) {
1958                                 err = -EINVAL;
1959                                 break;
1960                         }
1961                 } else {
1962                         if (conv(&neg, &lval, i, 0, data)) {
1963                                 err = -EINVAL;
1964                                 break;
1965                         }
1966                         if (!first)
1967                                 err = proc_put_char(&buffer, &left, '\t');
1968                         if (err)
1969                                 break;
1970                         err = proc_put_long(&buffer, &left, lval, neg);
1971                         if (err)
1972                                 break;
1973                 }
1974         }
1975
1976         if (!write && !first && left && !err)
1977                 err = proc_put_char(&buffer, &left, '\n');
1978         if (write && !err && left)
1979                 left -= proc_skip_spaces(&kbuf);
1980 free:
1981         if (write) {
1982                 free_page(page);
1983                 if (first)
1984                         return err ? : -EINVAL;
1985         }
1986         *lenp -= left;
1987         *ppos += *lenp;
1988         return err;
1989 }
1990
1991 static int do_proc_dointvec(struct ctl_table *table, int write,
1992                   void __user *buffer, size_t *lenp, loff_t *ppos,
1993                   int (*conv)(bool *negp, unsigned long *lvalp, int *valp,
1994                               int write, void *data),
1995                   void *data)
1996 {
1997         return __do_proc_dointvec(table->data, table, write,
1998                         buffer, lenp, ppos, conv, data);
1999 }
2000
2001 /**
2002  * proc_dointvec - read a vector of integers
2003  * @table: the sysctl table
2004  * @write: %TRUE if this is a write to the sysctl file
2005  * @buffer: the user buffer
2006  * @lenp: the size of the user buffer
2007  * @ppos: file position
2008  *
2009  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2010  * values from/to the user buffer, treated as an ASCII string.
2011  *
2012  * Returns 0 on success.
2013  */
2014 int proc_dointvec(struct ctl_table *table, int write,
2015                      void __user *buffer, size_t *lenp, loff_t *ppos)
2016 {
2017     return do_proc_dointvec(table,write,buffer,lenp,ppos,
2018                             NULL,NULL);
2019 }
2020
2021 /*
2022  * Taint values can only be increased
2023  * This means we can safely use a temporary.
2024  */
2025 static int proc_taint(struct ctl_table *table, int write,
2026                                void __user *buffer, size_t *lenp, loff_t *ppos)
2027 {
2028         struct ctl_table t;
2029         unsigned long tmptaint = get_taint();
2030         int err;
2031
2032         if (write && !capable(CAP_SYS_ADMIN))
2033                 return -EPERM;
2034
2035         t = *table;
2036         t.data = &tmptaint;
2037         err = proc_doulongvec_minmax(&t, write, buffer, lenp, ppos);
2038         if (err < 0)
2039                 return err;
2040
2041         if (write) {
2042                 /*
2043                  * Poor man's atomic or. Not worth adding a primitive
2044                  * to everyone's atomic.h for this
2045                  */
2046                 int i;
2047                 for (i = 0; i < BITS_PER_LONG && tmptaint >> i; i++) {
2048                         if ((tmptaint >> i) & 1)
2049                                 add_taint(i, LOCKDEP_STILL_OK);
2050                 }
2051         }
2052
2053         return err;
2054 }
2055
2056 #ifdef CONFIG_PRINTK
2057 static int proc_dointvec_minmax_sysadmin(struct ctl_table *table, int write,
2058                                 void __user *buffer, size_t *lenp, loff_t *ppos)
2059 {
2060         if (write && !capable(CAP_SYS_ADMIN))
2061                 return -EPERM;
2062
2063         return proc_dointvec_minmax(table, write, buffer, lenp, ppos);
2064 }
2065 #endif
2066
2067 struct do_proc_dointvec_minmax_conv_param {
2068         int *min;
2069         int *max;
2070 };
2071
2072 static int do_proc_dointvec_minmax_conv(bool *negp, unsigned long *lvalp,
2073                                         int *valp,
2074                                         int write, void *data)
2075 {
2076         struct do_proc_dointvec_minmax_conv_param *param = data;
2077         if (write) {
2078                 int val = *negp ? -*lvalp : *lvalp;
2079                 if ((param->min && *param->min > val) ||
2080                     (param->max && *param->max < val))
2081                         return -EINVAL;
2082                 *valp = val;
2083         } else {
2084                 int val = *valp;
2085                 if (val < 0) {
2086                         *negp = true;
2087                         *lvalp = (unsigned long)-val;
2088                 } else {
2089                         *negp = false;
2090                         *lvalp = (unsigned long)val;
2091                 }
2092         }
2093         return 0;
2094 }
2095
2096 /**
2097  * proc_dointvec_minmax - read a vector of integers with min/max values
2098  * @table: the sysctl table
2099  * @write: %TRUE if this is a write to the sysctl file
2100  * @buffer: the user buffer
2101  * @lenp: the size of the user buffer
2102  * @ppos: file position
2103  *
2104  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2105  * values from/to the user buffer, treated as an ASCII string.
2106  *
2107  * This routine will ensure the values are within the range specified by
2108  * table->extra1 (min) and table->extra2 (max).
2109  *
2110  * Returns 0 on success.
2111  */
2112 int proc_dointvec_minmax(struct ctl_table *table, int write,
2113                   void __user *buffer, size_t *lenp, loff_t *ppos)
2114 {
2115         struct do_proc_dointvec_minmax_conv_param param = {
2116                 .min = (int *) table->extra1,
2117                 .max = (int *) table->extra2,
2118         };
2119         return do_proc_dointvec(table, write, buffer, lenp, ppos,
2120                                 do_proc_dointvec_minmax_conv, &param);
2121 }
2122
2123 static void validate_coredump_safety(void)
2124 {
2125 #ifdef CONFIG_COREDUMP
2126         if (suid_dumpable == SUID_DUMP_ROOT &&
2127             core_pattern[0] != '/' && core_pattern[0] != '|') {
2128                 printk(KERN_WARNING "Unsafe core_pattern used with "\
2129                         "suid_dumpable=2. Pipe handler or fully qualified "\
2130                         "core dump path required.\n");
2131         }
2132 #endif
2133 }
2134
2135 static int proc_dointvec_minmax_coredump(struct ctl_table *table, int write,
2136                 void __user *buffer, size_t *lenp, loff_t *ppos)
2137 {
2138         int error = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
2139         if (!error)
2140                 validate_coredump_safety();
2141         return error;
2142 }
2143
2144 #ifdef CONFIG_COREDUMP
2145 static int proc_dostring_coredump(struct ctl_table *table, int write,
2146                   void __user *buffer, size_t *lenp, loff_t *ppos)
2147 {
2148         int error = proc_dostring(table, write, buffer, lenp, ppos);
2149         if (!error)
2150                 validate_coredump_safety();
2151         return error;
2152 }
2153 #endif
2154
2155 static int __do_proc_doulongvec_minmax(void *data, struct ctl_table *table, int write,
2156                                      void __user *buffer,
2157                                      size_t *lenp, loff_t *ppos,
2158                                      unsigned long convmul,
2159                                      unsigned long convdiv)
2160 {
2161         unsigned long *i, *min, *max;
2162         int vleft, first = 1, err = 0;
2163         unsigned long page = 0;
2164         size_t left;
2165         char *kbuf;
2166
2167         if (!data || !table->maxlen || !*lenp || (*ppos && !write)) {
2168                 *lenp = 0;
2169                 return 0;
2170         }
2171
2172         i = (unsigned long *) data;
2173         min = (unsigned long *) table->extra1;
2174         max = (unsigned long *) table->extra2;
2175         vleft = table->maxlen / sizeof(unsigned long);
2176         left = *lenp;
2177
2178         if (write) {
2179                 if (left > PAGE_SIZE - 1)
2180                         left = PAGE_SIZE - 1;
2181                 page = __get_free_page(GFP_TEMPORARY);
2182                 kbuf = (char *) page;
2183                 if (!kbuf)
2184                         return -ENOMEM;
2185                 if (copy_from_user(kbuf, buffer, left)) {
2186                         err = -EFAULT;
2187                         goto free;
2188                 }
2189                 kbuf[left] = 0;
2190         }
2191
2192         for (; left && vleft--; i++, first = 0) {
2193                 unsigned long val;
2194
2195                 if (write) {
2196                         bool neg;
2197
2198                         left -= proc_skip_spaces(&kbuf);
2199
2200                         err = proc_get_long(&kbuf, &left, &val, &neg,
2201                                              proc_wspace_sep,
2202                                              sizeof(proc_wspace_sep), NULL);
2203                         if (err)
2204                                 break;
2205                         if (neg)
2206                                 continue;
2207                         if ((min && val < *min) || (max && val > *max))
2208                                 continue;
2209                         *i = val;
2210                 } else {
2211                         val = convdiv * (*i) / convmul;
2212                         if (!first) {
2213                                 err = proc_put_char(&buffer, &left, '\t');
2214                                 if (err)
2215                                         break;
2216                         }
2217                         err = proc_put_long(&buffer, &left, val, false);
2218                         if (err)
2219                                 break;
2220                 }
2221         }
2222
2223         if (!write && !first && left && !err)
2224                 err = proc_put_char(&buffer, &left, '\n');
2225         if (write && !err)
2226                 left -= proc_skip_spaces(&kbuf);
2227 free:
2228         if (write) {
2229                 free_page(page);
2230                 if (first)
2231                         return err ? : -EINVAL;
2232         }
2233         *lenp -= left;
2234         *ppos += *lenp;
2235         return err;
2236 }
2237
2238 static int do_proc_doulongvec_minmax(struct ctl_table *table, int write,
2239                                      void __user *buffer,
2240                                      size_t *lenp, loff_t *ppos,
2241                                      unsigned long convmul,
2242                                      unsigned long convdiv)
2243 {
2244         return __do_proc_doulongvec_minmax(table->data, table, write,
2245                         buffer, lenp, ppos, convmul, convdiv);
2246 }
2247
2248 /**
2249  * proc_doulongvec_minmax - read a vector of long integers with min/max values
2250  * @table: the sysctl table
2251  * @write: %TRUE if this is a write to the sysctl file
2252  * @buffer: the user buffer
2253  * @lenp: the size of the user buffer
2254  * @ppos: file position
2255  *
2256  * Reads/writes up to table->maxlen/sizeof(unsigned long) unsigned long
2257  * values from/to the user buffer, treated as an ASCII string.
2258  *
2259  * This routine will ensure the values are within the range specified by
2260  * table->extra1 (min) and table->extra2 (max).
2261  *
2262  * Returns 0 on success.
2263  */
2264 int proc_doulongvec_minmax(struct ctl_table *table, int write,
2265                            void __user *buffer, size_t *lenp, loff_t *ppos)
2266 {
2267     return do_proc_doulongvec_minmax(table, write, buffer, lenp, ppos, 1l, 1l);
2268 }
2269
2270 /**
2271  * proc_doulongvec_ms_jiffies_minmax - read a vector of millisecond values with min/max values
2272  * @table: the sysctl table
2273  * @write: %TRUE if this is a write to the sysctl file
2274  * @buffer: the user buffer
2275  * @lenp: the size of the user buffer
2276  * @ppos: file position
2277  *
2278  * Reads/writes up to table->maxlen/sizeof(unsigned long) unsigned long
2279  * values from/to the user buffer, treated as an ASCII string. The values
2280  * are treated as milliseconds, and converted to jiffies when they are stored.
2281  *
2282  * This routine will ensure the values are within the range specified by
2283  * table->extra1 (min) and table->extra2 (max).
2284  *
2285  * Returns 0 on success.
2286  */
2287 int proc_doulongvec_ms_jiffies_minmax(struct ctl_table *table, int write,
2288                                       void __user *buffer,
2289                                       size_t *lenp, loff_t *ppos)
2290 {
2291     return do_proc_doulongvec_minmax(table, write, buffer,
2292                                      lenp, ppos, HZ, 1000l);
2293 }
2294
2295
2296 static int do_proc_dointvec_jiffies_conv(bool *negp, unsigned long *lvalp,
2297                                          int *valp,
2298                                          int write, void *data)
2299 {
2300         if (write) {
2301                 if (*lvalp > LONG_MAX / HZ)
2302                         return 1;
2303                 *valp = *negp ? -(*lvalp*HZ) : (*lvalp*HZ);
2304         } else {
2305                 int val = *valp;
2306                 unsigned long lval;
2307                 if (val < 0) {
2308                         *negp = true;
2309                         lval = (unsigned long)-val;
2310                 } else {
2311                         *negp = false;
2312                         lval = (unsigned long)val;
2313                 }
2314                 *lvalp = lval / HZ;
2315         }
2316         return 0;
2317 }
2318
2319 static int do_proc_dointvec_userhz_jiffies_conv(bool *negp, unsigned long *lvalp,
2320                                                 int *valp,
2321                                                 int write, void *data)
2322 {
2323         if (write) {
2324                 if (USER_HZ < HZ && *lvalp > (LONG_MAX / HZ) * USER_HZ)
2325                         return 1;
2326                 *valp = clock_t_to_jiffies(*negp ? -*lvalp : *lvalp);
2327         } else {
2328                 int val = *valp;
2329                 unsigned long lval;
2330                 if (val < 0) {
2331                         *negp = true;
2332                         lval = (unsigned long)-val;
2333                 } else {
2334                         *negp = false;
2335                         lval = (unsigned long)val;
2336                 }
2337                 *lvalp = jiffies_to_clock_t(lval);
2338         }
2339         return 0;
2340 }
2341
2342 static int do_proc_dointvec_ms_jiffies_conv(bool *negp, unsigned long *lvalp,
2343                                             int *valp,
2344                                             int write, void *data)
2345 {
2346         if (write) {
2347                 unsigned long jif = msecs_to_jiffies(*negp ? -*lvalp : *lvalp);
2348
2349                 if (jif > INT_MAX)
2350                         return 1;
2351                 *valp = (int)jif;
2352         } else {
2353                 int val = *valp;
2354                 unsigned long lval;
2355                 if (val < 0) {
2356                         *negp = true;
2357                         lval = (unsigned long)-val;
2358                 } else {
2359                         *negp = false;
2360                         lval = (unsigned long)val;
2361                 }
2362                 *lvalp = jiffies_to_msecs(lval);
2363         }
2364         return 0;
2365 }
2366
2367 /**
2368  * proc_dointvec_jiffies - read a vector of integers as seconds
2369  * @table: the sysctl table
2370  * @write: %TRUE if this is a write to the sysctl file
2371  * @buffer: the user buffer
2372  * @lenp: the size of the user buffer
2373  * @ppos: file position
2374  *
2375  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2376  * values from/to the user buffer, treated as an ASCII string.
2377  * The values read are assumed to be in seconds, and are converted into
2378  * jiffies.
2379  *
2380  * Returns 0 on success.
2381  */
2382 int proc_dointvec_jiffies(struct ctl_table *table, int write,
2383                           void __user *buffer, size_t *lenp, loff_t *ppos)
2384 {
2385     return do_proc_dointvec(table,write,buffer,lenp,ppos,
2386                             do_proc_dointvec_jiffies_conv,NULL);
2387 }
2388
2389 /**
2390  * proc_dointvec_userhz_jiffies - read a vector of integers as 1/USER_HZ seconds
2391  * @table: the sysctl table
2392  * @write: %TRUE if this is a write to the sysctl file
2393  * @buffer: the user buffer
2394  * @lenp: the size of the user buffer
2395  * @ppos: pointer to the file position
2396  *
2397  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2398  * values from/to the user buffer, treated as an ASCII string.
2399  * The values read are assumed to be in 1/USER_HZ seconds, and
2400  * are converted into jiffies.
2401  *
2402  * Returns 0 on success.
2403  */
2404 int proc_dointvec_userhz_jiffies(struct ctl_table *table, int write,
2405                                  void __user *buffer, size_t *lenp, loff_t *ppos)
2406 {
2407     return do_proc_dointvec(table,write,buffer,lenp,ppos,
2408                             do_proc_dointvec_userhz_jiffies_conv,NULL);
2409 }
2410
2411 /**
2412  * proc_dointvec_ms_jiffies - read a vector of integers as 1 milliseconds
2413  * @table: the sysctl table
2414  * @write: %TRUE if this is a write to the sysctl file
2415  * @buffer: the user buffer
2416  * @lenp: the size of the user buffer
2417  * @ppos: file position
2418  * @ppos: the current position in the file
2419  *
2420  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2421  * values from/to the user buffer, treated as an ASCII string.
2422  * The values read are assumed to be in 1/1000 seconds, and
2423  * are converted into jiffies.
2424  *
2425  * Returns 0 on success.
2426  */
2427 int proc_dointvec_ms_jiffies(struct ctl_table *table, int write,
2428                              void __user *buffer, size_t *lenp, loff_t *ppos)
2429 {
2430         return do_proc_dointvec(table, write, buffer, lenp, ppos,
2431                                 do_proc_dointvec_ms_jiffies_conv, NULL);
2432 }
2433
2434 static int proc_do_cad_pid(struct ctl_table *table, int write,
2435                            void __user *buffer, size_t *lenp, loff_t *ppos)
2436 {
2437         struct pid *new_pid;
2438         pid_t tmp;
2439         int r;
2440
2441         tmp = pid_vnr(cad_pid);
2442
2443         r = __do_proc_dointvec(&tmp, table, write, buffer,
2444                                lenp, ppos, NULL, NULL);
2445         if (r || !write)
2446                 return r;
2447
2448         new_pid = find_get_pid(tmp);
2449         if (!new_pid)
2450                 return -ESRCH;
2451
2452         put_pid(xchg(&cad_pid, new_pid));
2453         return 0;
2454 }
2455
2456 /**
2457  * proc_do_large_bitmap - read/write from/to a large bitmap
2458  * @table: the sysctl table
2459  * @write: %TRUE if this is a write to the sysctl file
2460  * @buffer: the user buffer
2461  * @lenp: the size of the user buffer
2462  * @ppos: file position
2463  *
2464  * The bitmap is stored at table->data and the bitmap length (in bits)
2465  * in table->maxlen.
2466  *
2467  * We use a range comma separated format (e.g. 1,3-4,10-10) so that
2468  * large bitmaps may be represented in a compact manner. Writing into
2469  * the file will clear the bitmap then update it with the given input.
2470  *
2471  * Returns 0 on success.
2472  */
2473 int proc_do_large_bitmap(struct ctl_table *table, int write,
2474                          void __user *buffer, size_t *lenp, loff_t *ppos)
2475 {
2476         int err = 0;
2477         bool first = 1;
2478         size_t left = *lenp;
2479         unsigned long bitmap_len = table->maxlen;
2480         unsigned long *bitmap = (unsigned long *) table->data;
2481         unsigned long *tmp_bitmap = NULL;
2482         char tr_a[] = { '-', ',', '\n' }, tr_b[] = { ',', '\n', 0 }, c;
2483
2484         if (!bitmap_len || !left || (*ppos && !write)) {
2485                 *lenp = 0;
2486                 return 0;
2487         }
2488
2489         if (write) {
2490                 unsigned long page = 0;
2491                 char *kbuf;
2492
2493                 if (left > PAGE_SIZE - 1)
2494                         left = PAGE_SIZE - 1;
2495
2496                 page = __get_free_page(GFP_TEMPORARY);
2497                 kbuf = (char *) page;
2498                 if (!kbuf)
2499                         return -ENOMEM;
2500                 if (copy_from_user(kbuf, buffer, left)) {
2501                         free_page(page);
2502                         return -EFAULT;
2503                 }
2504                 kbuf[left] = 0;
2505
2506                 tmp_bitmap = kzalloc(BITS_TO_LONGS(bitmap_len) * sizeof(unsigned long),
2507                                      GFP_KERNEL);
2508                 if (!tmp_bitmap) {
2509                         free_page(page);
2510                         return -ENOMEM;
2511                 }
2512                 proc_skip_char(&kbuf, &left, '\n');
2513                 while (!err && left) {
2514                         unsigned long val_a, val_b;
2515                         bool neg;
2516
2517                         err = proc_get_long(&kbuf, &left, &val_a, &neg, tr_a,
2518                                              sizeof(tr_a), &c);
2519                         if (err)
2520                                 break;
2521                         if (val_a >= bitmap_len || neg) {
2522                                 err = -EINVAL;
2523                                 break;
2524                         }
2525
2526                         val_b = val_a;
2527                         if (left) {
2528                                 kbuf++;
2529                                 left--;
2530                         }
2531
2532                         if (c == '-') {
2533                                 err = proc_get_long(&kbuf, &left, &val_b,
2534                                                      &neg, tr_b, sizeof(tr_b),
2535                                                      &c);
2536                                 if (err)
2537                                         break;
2538                                 if (val_b >= bitmap_len || neg ||
2539                                     val_a > val_b) {
2540                                         err = -EINVAL;
2541                                         break;
2542                                 }
2543                                 if (left) {
2544                                         kbuf++;
2545                                         left--;
2546                                 }
2547                         }
2548
2549                         bitmap_set(tmp_bitmap, val_a, val_b - val_a + 1);
2550                         first = 0;
2551                         proc_skip_char(&kbuf, &left, '\n');
2552                 }
2553                 free_page(page);
2554         } else {
2555                 unsigned long bit_a, bit_b = 0;
2556
2557                 while (left) {
2558                         bit_a = find_next_bit(bitmap, bitmap_len, bit_b);
2559                         if (bit_a >= bitmap_len)
2560                                 break;
2561                         bit_b = find_next_zero_bit(bitmap, bitmap_len,
2562                                                    bit_a + 1) - 1;
2563
2564                         if (!first) {
2565                                 err = proc_put_char(&buffer, &left, ',');
2566                                 if (err)
2567                                         break;
2568                         }
2569                         err = proc_put_long(&buffer, &left, bit_a, false);
2570                         if (err)
2571                                 break;
2572                         if (bit_a != bit_b) {
2573                                 err = proc_put_char(&buffer, &left, '-');
2574                                 if (err)
2575                                         break;
2576                                 err = proc_put_long(&buffer, &left, bit_b, false);
2577                                 if (err)
2578                                         break;
2579                         }
2580
2581                         first = 0; bit_b++;
2582                 }
2583                 if (!err)
2584                         err = proc_put_char(&buffer, &left, '\n');
2585         }
2586
2587         if (!err) {
2588                 if (write) {
2589                         if (*ppos)
2590                                 bitmap_or(bitmap, bitmap, tmp_bitmap, bitmap_len);
2591                         else
2592                                 bitmap_copy(bitmap, tmp_bitmap, bitmap_len);
2593                 }
2594                 kfree(tmp_bitmap);
2595                 *lenp -= left;
2596                 *ppos += *lenp;
2597                 return 0;
2598         } else {
2599                 kfree(tmp_bitmap);
2600                 return err;
2601         }
2602 }
2603
2604 #else /* CONFIG_PROC_SYSCTL */
2605
2606 int proc_dostring(struct ctl_table *table, int write,
2607                   void __user *buffer, size_t *lenp, loff_t *ppos)
2608 {
2609         return -ENOSYS;
2610 }
2611
2612 int proc_dointvec(struct ctl_table *table, int write,
2613                   void __user *buffer, size_t *lenp, loff_t *ppos)
2614 {
2615         return -ENOSYS;
2616 }
2617
2618 int proc_dointvec_minmax(struct ctl_table *table, int write,
2619                     void __user *buffer, size_t *lenp, loff_t *ppos)
2620 {
2621         return -ENOSYS;
2622 }
2623
2624 int proc_dointvec_jiffies(struct ctl_table *table, int write,
2625                     void __user *buffer, size_t *lenp, loff_t *ppos)
2626 {
2627         return -ENOSYS;
2628 }
2629
2630 int proc_dointvec_userhz_jiffies(struct ctl_table *table, int write,
2631                     void __user *buffer, size_t *lenp, loff_t *ppos)
2632 {
2633         return -ENOSYS;
2634 }
2635
2636 int proc_dointvec_ms_jiffies(struct ctl_table *table, int write,
2637                              void __user *buffer, size_t *lenp, loff_t *ppos)
2638 {
2639         return -ENOSYS;
2640 }
2641
2642 int proc_doulongvec_minmax(struct ctl_table *table, int write,
2643                     void __user *buffer, size_t *lenp, loff_t *ppos)
2644 {
2645         return -ENOSYS;
2646 }
2647
2648 int proc_doulongvec_ms_jiffies_minmax(struct ctl_table *table, int write,
2649                                       void __user *buffer,
2650                                       size_t *lenp, loff_t *ppos)
2651 {
2652     return -ENOSYS;
2653 }
2654
2655
2656 #endif /* CONFIG_PROC_SYSCTL */
2657
2658 /*
2659  * No sense putting this after each symbol definition, twice,
2660  * exception granted :-)
2661  */
2662 EXPORT_SYMBOL(proc_dointvec);
2663 EXPORT_SYMBOL(proc_dointvec_jiffies);
2664 EXPORT_SYMBOL(proc_dointvec_minmax);
2665 EXPORT_SYMBOL(proc_dointvec_userhz_jiffies);
2666 EXPORT_SYMBOL(proc_dointvec_ms_jiffies);
2667 EXPORT_SYMBOL(proc_dostring);
2668 EXPORT_SYMBOL(proc_doulongvec_minmax);
2669 EXPORT_SYMBOL(proc_doulongvec_ms_jiffies_minmax);