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
  65 #include <asm/uaccess.h>
  66 #include <asm/processor.h>
  67
  68 #ifdef CONFIG_X86
  69 #include <asm/nmi.h>
  70 #include <asm/stacktrace.h>
  71 #include <asm/io.h>
  72 #endif
  73 #ifdef CONFIG_SPARC
  74 #include <asm/setup.h>
  75 #endif
  76 #ifdef CONFIG_BSD_PROCESS_ACCT
  77 #include <linux/acct.h>
  78 #endif
  79 #ifdef CONFIG_RT_MUTEXES
  80 #include <linux/rtmutex.h>
  81 #endif
  82 #if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_LOCK_STAT)
  83 #include <linux/lockdep.h>
  84 #endif
  85 #ifdef CONFIG_CHR_DEV_SG
  86 #include <scsi/sg.h>
  87 #endif
  88
  89 #ifdef CONFIG_LOCKUP_DETECTOR
  90 #include <linux/nmi.h>
  91 #endif
  92
  93
  94 #if defined(CONFIG_SYSCTL)
  95
  96 /* External variables not in a header file. */
  97 extern int sysctl_overcommit_memory;
  98 extern int sysctl_overcommit_ratio;
  99 extern int max_threads;
 100 extern int suid_dumpable;
 101 #ifdef CONFIG_COREDUMP
 102 extern int core_uses_pid;
 103 extern char core_pattern[];
 104 extern unsigned int core_pipe_limit;
 105 #endif
 106 extern int pid_max;
 107 extern int min_free_kbytes;
 108 extern int pid_max_min, pid_max_max;
 109 extern int sysctl_drop_caches;
 110 extern int percpu_pagelist_fraction;
 111 extern int compat_log;
 112 extern int latencytop_enabled;
 113 extern int sysctl_nr_open_min, sysctl_nr_open_max;
 114 #ifndef CONFIG_MMU
 115 extern int sysctl_nr_trim_pages;
 116 #endif
 117 #ifdef CONFIG_BLOCK
 118 extern int blk_iopoll_enabled;
 119 #endif
 120
 121 /* Constants used for minimum and  maximum */
 122 #ifdef CONFIG_LOCKUP_DETECTOR
 123 static int sixty = 60;
 124 static int neg_one = -1;
 125 #endif
 126
 127 static int zero;
 128 static int __maybe_unused one = 1;
 129 static int __maybe_unused two = 2;
 130 static int __maybe_unused three = 3;
 131 static unsigned long one_ul = 1;
 132 static int one_hundred = 100;
 133 #ifdef CONFIG_PRINTK
 134 static int ten_thousand = 10000;
 135 #endif
 136
 137 /* this is needed for the proc_doulongvec_minmax of vm_dirty_bytes */
 138 static unsigned long dirty_bytes_min = 2 * PAGE_SIZE;
 139
 140 /* this is needed for the proc_dointvec_minmax for [fs_]overflow UID and GID */
 141 static int maxolduid = 65535;
 142 static int minolduid;
 143 static int min_percpu_pagelist_fract = 8;
 144
 145 static int ngroups_max = NGROUPS_MAX;
 146 static const int cap_last_cap = CAP_LAST_CAP;
 147
 148 #ifdef CONFIG_INOTIFY_USER
 149 #include <linux/inotify.h>
 150 #endif
 151 #ifdef CONFIG_SPARC
 152 #endif
 153
 154 #ifdef CONFIG_SPARC64
 155 extern int sysctl_tsb_ratio;
 156 #endif
 157
 158 #ifdef __hppa__
 159 extern int pwrsw_enabled;
 160 extern int unaligned_enabled;
 161 #endif
 162
 163 #ifdef CONFIG_IA64
 164 extern int no_unaligned_warning;
 165 extern int unaligned_dump_stack;
 166 #endif
 167
 168 #ifdef CONFIG_PROC_SYSCTL
 169 static int proc_do_cad_pid(struct ctl_table *table, int write,
 170                   void __user *buffer, size_t *lenp, loff_t *ppos);
 171 static int proc_taint(struct ctl_table *table, int write,
 172                                void __user *buffer, size_t *lenp, loff_t *ppos);
 173 #endif
 174
 175 #ifdef CONFIG_PRINTK
 176 static int proc_dointvec_minmax_sysadmin(struct ctl_table *table, int write,
 177                                 void __user *buffer, size_t *lenp, loff_t *ppos);
 178 #endif
 179
 180 static int proc_dointvec_minmax_coredump(struct ctl_table *table, int write,
 181                 void __user *buffer, size_t *lenp, loff_t *ppos);
 182 #ifdef CONFIG_COREDUMP
 183 static int proc_dostring_coredump(struct ctl_table *table, int write,
 184                 void __user *buffer, size_t *lenp, loff_t *ppos);
 185 #endif
 186
 187 #ifdef CONFIG_MAGIC_SYSRQ
 188 /* Note: sysrq code uses it's own private copy */
 189 static int __sysrq_enabled = SYSRQ_DEFAULT_ENABLE;
 190
 191 static int sysrq_sysctl_handler(ctl_table *table, int write,
 192                                 void __user *buffer, size_t *lenp,
 193                                 loff_t *ppos)
 194 {
 195         int error;
 196
 197         error = proc_dointvec(table, write, buffer, lenp, ppos);
 198         if (error)
 199                 return error;
 200
 201         if (write)
 202                 sysrq_toggle_support(__sysrq_enabled);
 203
 204         return 0;
 205 }
 206
 207 #endif
 208
 209 static struct ctl_table kern_table[];
 210 static struct ctl_table vm_table[];
 211 static struct ctl_table fs_table[];
 212 static struct ctl_table debug_table[];
 213 static struct ctl_table dev_table[];
 214 extern struct ctl_table random_table[];
 215 #ifdef CONFIG_EPOLL
 216 extern struct ctl_table epoll_table[];
 217 #endif
 218
 219 #ifdef HAVE_ARCH_PICK_MMAP_LAYOUT
 220 int sysctl_legacy_va_layout;
 221 #endif
 222
 223 /* The default sysctl tables: */
 224
 225 static struct ctl_table sysctl_base_table[] = {
 226         {
 227                 .procname       = "kernel",
 228                 .mode           = 0555,
 229                 .child          = kern_table,
 230         },
 231         {
 232                 .procname       = "vm",
 233                 .mode           = 0555,
 234                 .child          = vm_table,
 235         },
 236         {
 237                 .procname       = "fs",
 238                 .mode           = 0555,
 239                 .child          = fs_table,
 240         },
 241         {
 242                 .procname       = "debug",
 243                 .mode           = 0555,
 244                 .child          = debug_table,
 245         },
 246         {
 247                 .procname       = "dev",
 248                 .mode           = 0555,
 249                 .child          = dev_table,
 250         },
 251         { }
 252 };
 253
 254 #ifdef CONFIG_SCHED_DEBUG
 255 static int min_sched_granularity_ns = 100000;           /* 100 usecs */
 256 static int max_sched_granularity_ns = NSEC_PER_SEC;     /* 1 second */
 257 static int min_wakeup_granularity_ns;                   /* 0 usecs */
 258 static int max_wakeup_granularity_ns = NSEC_PER_SEC;    /* 1 second */
 259 #ifdef CONFIG_SMP
 260 static int min_sched_tunable_scaling = SCHED_TUNABLESCALING_NONE;
 261 static int max_sched_tunable_scaling = SCHED_TUNABLESCALING_END-1;
 262 #endif /* CONFIG_SMP */
 263 #endif /* CONFIG_SCHED_DEBUG */
 264
 265 #ifdef CONFIG_COMPACTION
 266 static int min_extfrag_threshold;
 267 static int max_extfrag_threshold = 1000;
 268 #endif
 269
 270 static struct ctl_table kern_table[] = {
 271         {
 272                 .procname       = "sched_child_runs_first",
 273                 .data           = &sysctl_sched_child_runs_first,
 274                 .maxlen         = sizeof(unsigned int),
 275                 .mode           = 0644,
 276                 .proc_handler   = proc_dointvec,
 277         },
 278 #ifdef CONFIG_SCHED_DEBUG
 279         {
 280                 .procname       = "sched_min_granularity_ns",
 281                 .data           = &sysctl_sched_min_granularity,
 282                 .maxlen         = sizeof(unsigned int),
 283                 .mode           = 0644,
 284                 .proc_handler   = sched_proc_update_handler,
 285                 .extra1         = &min_sched_granularity_ns,
 286                 .extra2         = &max_sched_granularity_ns,
 287         },
 288         {
 289                 .procname       = "sched_latency_ns",
 290                 .data           = &sysctl_sched_latency,
 291                 .maxlen         = sizeof(unsigned int),
 292                 .mode           = 0644,
 293                 .proc_handler   = sched_proc_update_handler,
 294                 .extra1         = &min_sched_granularity_ns,
 295                 .extra2         = &max_sched_granularity_ns,
 296         },
 297         {
 298                 .procname       = "sched_wakeup_granularity_ns",
 299                 .data           = &sysctl_sched_wakeup_granularity,
 300                 .maxlen         = sizeof(unsigned int),
 301                 .mode           = 0644,
 302                 .proc_handler   = sched_proc_update_handler,
 303                 .extra1         = &min_wakeup_granularity_ns,
 304                 .extra2         = &max_wakeup_granularity_ns,
 305         },
 306 #ifdef CONFIG_SMP
 307         {
 308                 .procname       = "sched_tunable_scaling",
 309                 .data           = &sysctl_sched_tunable_scaling,
 310                 .maxlen         = sizeof(enum sched_tunable_scaling),
 311                 .mode           = 0644,
 312                 .proc_handler   = sched_proc_update_handler,
 313                 .extra1         = &min_sched_tunable_scaling,
 314                 .extra2         = &max_sched_tunable_scaling,
 315         },
 316         {
 317                 .procname       = "sched_migration_cost_ns",
 318                 .data           = &sysctl_sched_migration_cost,
 319                 .maxlen         = sizeof(unsigned int),
 320                 .mode           = 0644,
 321                 .proc_handler   = proc_dointvec,
 322         },
 323         {
 324                 .procname       = "sched_nr_migrate",
 325                 .data           = &sysctl_sched_nr_migrate,
 326                 .maxlen         = sizeof(unsigned int),
 327                 .mode           = 0644,
 328                 .proc_handler   = proc_dointvec,
 329         },
 330         {
 331                 .procname       = "sched_time_avg_ms",
 332                 .data           = &sysctl_sched_time_avg,
 333                 .maxlen         = sizeof(unsigned int),
 334                 .mode           = 0644,
 335                 .proc_handler   = proc_dointvec,
 336         },
 337         {
 338                 .procname       = "sched_shares_window_ns",
 339                 .data           = &sysctl_sched_shares_window,
 340                 .maxlen         = sizeof(unsigned int),
 341                 .mode           = 0644,
 342                 .proc_handler   = proc_dointvec,
 343         },
 344         {
 345                 .procname       = "timer_migration",
 346                 .data           = &sysctl_timer_migration,
 347                 .maxlen         = sizeof(unsigned int),
 348                 .mode           = 0644,
 349                 .proc_handler   = proc_dointvec_minmax,
 350                 .extra1         = &zero,
 351                 .extra2         = &one,
 352         },
 353 #endif /* CONFIG_SMP */
 354 #ifdef CONFIG_NUMA_BALANCING
 355         {
 356                 .procname       = "numa_balancing_scan_delay_ms",
 357                 .data           = &sysctl_numa_balancing_scan_delay,
 358                 .maxlen         = sizeof(unsigned int),
 359                 .mode           = 0644,
 360                 .proc_handler   = proc_dointvec,
 361         },
 362         {
 363                 .procname       = "numa_balancing_scan_period_min_ms",
 364                 .data           = &sysctl_numa_balancing_scan_period_min,
 365                 .maxlen         = sizeof(unsigned int),
 366                 .mode           = 0644,
 367                 .proc_handler   = proc_dointvec,
 368         },
 369         {
 370                 .procname       = "numa_balancing_scan_period_max_ms",
 371                 .data           = &sysctl_numa_balancing_scan_period_max,
 372                 .maxlen         = sizeof(unsigned int),
 373                 .mode           = 0644,
 374                 .proc_handler   = proc_dointvec,
 375         },
 376         {
 377                 .procname       = "numa_balancing_scan_size_mb",
 378                 .data           = &sysctl_numa_balancing_scan_size,
 379                 .maxlen         = sizeof(unsigned int),
 380                 .mode           = 0644,
 381                 .proc_handler   = proc_dointvec,
 382         },
 383 #endif /* CONFIG_NUMA_BALANCING */
 384 #endif /* CONFIG_SCHED_DEBUG */
 385         {
 386                 .procname       = "sched_rt_period_us",
 387                 .data           = &sysctl_sched_rt_period,
 388                 .maxlen         = sizeof(unsigned int),
 389                 .mode           = 0644,
 390                 .proc_handler   = sched_rt_handler,
 391         },
 392         {
 393                 .procname       = "sched_rt_runtime_us",
 394                 .data           = &sysctl_sched_rt_runtime,
 395                 .maxlen         = sizeof(int),
 396                 .mode           = 0644,
 397                 .proc_handler   = sched_rt_handler,
 398         },
 399 #ifdef CONFIG_SCHED_AUTOGROUP
 400         {
 401                 .procname       = "sched_autogroup_enabled",
 402                 .data           = &sysctl_sched_autogroup_enabled,
 403                 .maxlen         = sizeof(unsigned int),
 404                 .mode           = 0644,
 405                 .proc_handler   = proc_dointvec_minmax,
 406                 .extra1         = &zero,
 407                 .extra2         = &one,
 408         },
 409 #endif
 410 #ifdef CONFIG_CFS_BANDWIDTH
 411         {
 412                 .procname       = "sched_cfs_bandwidth_slice_us",
 413                 .data           = &sysctl_sched_cfs_bandwidth_slice,
 414                 .maxlen         = sizeof(unsigned int),
 415                 .mode           = 0644,
 416                 .proc_handler   = proc_dointvec_minmax,
 417                 .extra1         = &one,
 418         },
 419 #endif
 420 #ifdef CONFIG_PROVE_LOCKING
 421         {
 422                 .procname       = "prove_locking",
 423                 .data           = &prove_locking,
 424                 .maxlen         = sizeof(int),
 425                 .mode           = 0644,
 426                 .proc_handler   = proc_dointvec,
 427         },
 428 #endif
 429 #ifdef CONFIG_LOCK_STAT
 430         {
 431                 .procname       = "lock_stat",
 432                 .data           = &lock_stat,
 433                 .maxlen         = sizeof(int),
 434                 .mode           = 0644,
 435                 .proc_handler   = proc_dointvec,
 436         },
 437 #endif
 438         {
 439                 .procname       = "panic",
 440                 .data           = &panic_timeout,
 441                 .maxlen         = sizeof(int),
 442                 .mode           = 0644,
 443                 .proc_handler   = proc_dointvec,
 444         },
 445 #ifdef CONFIG_COREDUMP
 446         {
 447                 .procname       = "core_uses_pid",
 448                 .data           = &core_uses_pid,
 449                 .maxlen         = sizeof(int),
 450                 .mode           = 0644,
 451                 .proc_handler   = proc_dointvec,
 452         },
 453         {
 454                 .procname       = "core_pattern",
 455                 .data           = core_pattern,
 456                 .maxlen         = CORENAME_MAX_SIZE,
 457                 .mode           = 0644,
 458                 .proc_handler   = proc_dostring_coredump,
 459         },
 460         {
 461                 .procname       = "core_pipe_limit",
 462                 .data           = &core_pipe_limit,
 463                 .maxlen         = sizeof(unsigned int),
 464                 .mode           = 0644,
 465                 .proc_handler   = proc_dointvec,
 466         },
 467 #endif
 468 #ifdef CONFIG_PROC_SYSCTL
 469         {
 470                 .procname       = "tainted",
 471                 .maxlen         = sizeof(long),
 472                 .mode           = 0644,
 473                 .proc_handler   = proc_taint,
 474         },
 475 #endif
 476 #ifdef CONFIG_LATENCYTOP
 477         {
 478                 .procname       = "latencytop",
 479                 .data           = &latencytop_enabled,
 480                 .maxlen         = sizeof(int),
 481                 .mode           = 0644,
 482                 .proc_handler   = proc_dointvec,
 483         },
 484 #endif
 485 #ifdef CONFIG_BLK_DEV_INITRD
 486         {
 487                 .procname       = "real-root-dev",
 488                 .data           = &real_root_dev,
 489                 .maxlen         = sizeof(int),
 490                 .mode           = 0644,
 491                 .proc_handler   = proc_dointvec,
 492         },
 493 #endif
 494         {
 495                 .procname       = "print-fatal-signals",
 496                 .data           = &print_fatal_signals,
 497                 .maxlen         = sizeof(int),
 498                 .mode           = 0644,
 499                 .proc_handler   = proc_dointvec,
 500         },
 501 #ifdef CONFIG_SPARC
 502         {
 503                 .procname       = "reboot-cmd",
 504                 .data           = reboot_command,
 505                 .maxlen         = 256,
 506                 .mode           = 0644,
 507                 .proc_handler   = proc_dostring,
 508         },
 509         {
 510                 .procname       = "stop-a",
 511                 .data           = &stop_a_enabled,
 512                 .maxlen         = sizeof (int),
 513                 .mode           = 0644,
 514                 .proc_handler   = proc_dointvec,
 515         },
 516         {
 517                 .procname       = "scons-poweroff",
 518                 .data           = &scons_pwroff,
 519                 .maxlen         = sizeof (int),
 520                 .mode           = 0644,
 521                 .proc_handler   = proc_dointvec,
 522         },
 523 #endif
 524 #ifdef CONFIG_SPARC64
 525         {
 526                 .procname       = "tsb-ratio",
 527                 .data           = &sysctl_tsb_ratio,
 528                 .maxlen         = sizeof (int),
 529                 .mode           = 0644,
 530                 .proc_handler   = proc_dointvec,
 531         },
 532 #endif
 533 #ifdef __hppa__
 534         {
 535                 .procname       = "soft-power",
 536                 .data           = &pwrsw_enabled,
 537                 .maxlen         = sizeof (int),
 538                 .mode           = 0644,
 539                 .proc_handler   = proc_dointvec,
 540         },
 541         {
 542                 .procname       = "unaligned-trap",
 543                 .data           = &unaligned_enabled,
 544                 .maxlen         = sizeof (int),
 545                 .mode           = 0644,
 546                 .proc_handler   = proc_dointvec,
 547         },
 548 #endif
 549         {
 550                 .procname       = "ctrl-alt-del",
 551                 .data           = &C_A_D,
 552                 .maxlen         = sizeof(int),
 553                 .mode           = 0644,
 554                 .proc_handler   = proc_dointvec,
 555         },
 556 #ifdef CONFIG_FUNCTION_TRACER
 557         {
 558                 .procname       = "ftrace_enabled",
 559                 .data           = &ftrace_enabled,
 560                 .maxlen         = sizeof(int),
 561                 .mode           = 0644,
 562                 .proc_handler   = ftrace_enable_sysctl,
 563         },
 564 #endif
 565 #ifdef CONFIG_STACK_TRACER
 566         {
 567                 .procname       = "stack_tracer_enabled",
 568                 .data           = &stack_tracer_enabled,
 569                 .maxlen         = sizeof(int),
 570                 .mode           = 0644,
 571                 .proc_handler   = stack_trace_sysctl,
 572         },
 573 #endif
 574 #ifdef CONFIG_TRACING
 575         {
 576                 .procname       = "ftrace_dump_on_oops",
 577                 .data           = &ftrace_dump_on_oops,
 578                 .maxlen         = sizeof(int),
 579                 .mode           = 0644,
 580                 .proc_handler   = proc_dointvec,
 581         },
 582 #endif
 583 #ifdef CONFIG_MODULES
 584         {
 585                 .procname       = "modprobe",
 586                 .data           = &modprobe_path,
 587                 .maxlen         = KMOD_PATH_LEN,
 588                 .mode           = 0644,
 589                 .proc_handler   = proc_dostring,
 590         },
 591         {
 592                 .procname       = "modules_disabled",
 593                 .data           = &modules_disabled,
 594                 .maxlen         = sizeof(int),
 595                 .mode           = 0644,
 596                 /* only handle a transition from default "0" to "1" */
 597                 .proc_handler   = proc_dointvec_minmax,
 598                 .extra1         = &one,
 599                 .extra2         = &one,
 600         },
 601 #endif
 602 #ifdef CONFIG_HOTPLUG
 603         {
 604                 .procname       = "hotplug",
 605                 .data           = &uevent_helper,
 606                 .maxlen         = UEVENT_HELPER_PATH_LEN,
 607                 .mode           = 0644,
 608                 .proc_handler   = proc_dostring,
 609         },
 610 #endif
 611 #ifdef CONFIG_CHR_DEV_SG
 612         {
 613                 .procname       = "sg-big-buff",
 614                 .data           = &sg_big_buff,
 615                 .maxlen         = sizeof (int),
 616                 .mode           = 0444,
 617                 .proc_handler   = proc_dointvec,
 618         },
 619 #endif
 620 #ifdef CONFIG_BSD_PROCESS_ACCT
 621         {
 622                 .procname       = "acct",
 623                 .data           = &acct_parm,
 624                 .maxlen         = 3*sizeof(int),
 625                 .mode           = 0644,
 626                 .proc_handler   = proc_dointvec,
 627         },
 628 #endif
 629 #ifdef CONFIG_MAGIC_SYSRQ
 630         {
 631                 .procname       = "sysrq",
 632                 .data           = &__sysrq_enabled,
 633                 .maxlen         = sizeof (int),
 634                 .mode           = 0644,
 635                 .proc_handler   = sysrq_sysctl_handler,
 636         },
 637 #endif
 638 #ifdef CONFIG_PROC_SYSCTL
 639         {
 640                 .procname       = "cad_pid",
 641                 .data           = NULL,
 642                 .maxlen         = sizeof (int),
 643                 .mode           = 0600,
 644                 .proc_handler   = proc_do_cad_pid,
 645         },
 646 #endif
 647         {
 648                 .procname       = "threads-max",
 649                 .data           = &max_threads,
 650                 .maxlen         = sizeof(int),
 651                 .mode           = 0644,
 652                 .proc_handler   = proc_dointvec,
 653         },
 654         {
 655                 .procname       = "random",
 656                 .mode           = 0555,
 657                 .child          = random_table,
 658         },
 659         {
 660                 .procname       = "usermodehelper",
 661                 .mode           = 0555,
 662                 .child          = usermodehelper_table,
 663         },
 664         {
 665                 .procname       = "overflowuid",
 666                 .data           = &overflowuid,
 667                 .maxlen         = sizeof(int),
 668                 .mode           = 0644,
 669                 .proc_handler   = proc_dointvec_minmax,
 670                 .extra1         = &minolduid,
 671                 .extra2         = &maxolduid,
 672         },
 673         {
 674                 .procname       = "overflowgid",
 675                 .data           = &overflowgid,
 676                 .maxlen         = sizeof(int),
 677                 .mode           = 0644,
 678                 .proc_handler   = proc_dointvec_minmax,
 679                 .extra1         = &minolduid,
 680                 .extra2         = &maxolduid,
 681         },
 682 #ifdef CONFIG_S390
 683 #ifdef CONFIG_MATHEMU
 684         {
 685                 .procname       = "ieee_emulation_warnings",
 686                 .data           = &sysctl_ieee_emulation_warnings,
 687                 .maxlen         = sizeof(int),
 688                 .mode           = 0644,
 689                 .proc_handler   = proc_dointvec,
 690         },
 691 #endif
 692         {
 693                 .procname       = "userprocess_debug",
 694                 .data           = &show_unhandled_signals,
 695                 .maxlen         = sizeof(int),
 696                 .mode           = 0644,
 697                 .proc_handler   = proc_dointvec,
 698         },
 699 #endif
 700         {
 701                 .procname       = "pid_max",
 702                 .data           = &pid_max,
 703                 .maxlen         = sizeof (int),
 704                 .mode           = 0644,
 705                 .proc_handler   = proc_dointvec_minmax,
 706                 .extra1         = &pid_max_min,
 707                 .extra2         = &pid_max_max,
 708         },
 709         {
 710                 .procname       = "panic_on_oops",
 711                 .data           = &panic_on_oops,
 712                 .maxlen         = sizeof(int),
 713                 .mode           = 0644,
 714                 .proc_handler   = proc_dointvec,
 715         },
 716 #if defined CONFIG_PRINTK
 717         {
 718                 .procname       = "printk",
 719                 .data           = &console_loglevel,
 720                 .maxlen         = 4*sizeof(int),
 721                 .mode           = 0644,
 722                 .proc_handler   = proc_dointvec,
 723         },
 724         {
 725                 .procname       = "printk_ratelimit",
 726                 .data           = &printk_ratelimit_state.interval,
 727                 .maxlen         = sizeof(int),
 728                 .mode           = 0644,
 729                 .proc_handler   = proc_dointvec_jiffies,
 730         },
 731         {
 732                 .procname       = "printk_ratelimit_burst",
 733                 .data           = &printk_ratelimit_state.burst,
 734                 .maxlen         = sizeof(int),
 735                 .mode           = 0644,
 736                 .proc_handler   = proc_dointvec,
 737         },
 738         {
 739                 .procname       = "printk_delay",
 740                 .data           = &printk_delay_msec,
 741                 .maxlen         = sizeof(int),
 742                 .mode           = 0644,
 743                 .proc_handler   = proc_dointvec_minmax,
 744                 .extra1         = &zero,
 745                 .extra2         = &ten_thousand,
 746         },
 747         {
 748                 .procname       = "dmesg_restrict",
 749                 .data           = &dmesg_restrict,
 750                 .maxlen         = sizeof(int),
 751                 .mode           = 0644,
 752                 .proc_handler   = proc_dointvec_minmax_sysadmin,
 753                 .extra1         = &zero,
 754                 .extra2         = &one,
 755         },
 756         {
 757                 .procname       = "kptr_restrict",
 758                 .data           = &kptr_restrict,
 759                 .maxlen         = sizeof(int),
 760                 .mode           = 0644,
 761                 .proc_handler   = proc_dointvec_minmax_sysadmin,
 762                 .extra1         = &zero,
 763                 .extra2         = &two,
 764         },
 765 #endif
 766         {
 767                 .procname       = "ngroups_max",
 768                 .data           = &ngroups_max,
 769                 .maxlen         = sizeof (int),
 770                 .mode           = 0444,
 771                 .proc_handler   = proc_dointvec,
 772         },
 773         {
 774                 .procname       = "cap_last_cap",
 775                 .data           = (void *)&cap_last_cap,
 776                 .maxlen         = sizeof(int),
 777                 .mode           = 0444,
 778                 .proc_handler   = proc_dointvec,
 779         },
 780 #if defined(CONFIG_LOCKUP_DETECTOR)
 781         {
 782                 .procname       = "watchdog",
 783                 .data           = &watchdog_enabled,
 784                 .maxlen         = sizeof (int),
 785                 .mode           = 0644,
 786                 .proc_handler   = proc_dowatchdog,
 787                 .extra1         = &zero,
 788                 .extra2         = &one,
 789         },
 790         {
 791                 .procname       = "watchdog_thresh",
 792                 .data           = &watchdog_thresh,
 793                 .maxlen         = sizeof(int),
 794                 .mode           = 0644,
 795                 .proc_handler   = proc_dowatchdog,
 796                 .extra1         = &neg_one,
 797                 .extra2         = &sixty,
 798         },
 799         {
 800                 .procname       = "softlockup_panic",
 801                 .data           = &softlockup_panic,
 802                 .maxlen         = sizeof(int),
 803                 .mode           = 0644,
 804                 .proc_handler   = proc_dointvec_minmax,
 805                 .extra1         = &zero,
 806                 .extra2         = &one,
 807         },
 808         {
 809                 .procname       = "nmi_watchdog",
 810                 .data           = &watchdog_enabled,
 811                 .maxlen         = sizeof (int),
 812                 .mode           = 0644,
 813                 .proc_handler   = proc_dowatchdog,
 814                 .extra1         = &zero,
 815                 .extra2         = &one,
 816         },
 817 #endif
 818 #if defined(CONFIG_X86_LOCAL_APIC) && defined(CONFIG_X86)
 819         {
 820                 .procname       = "unknown_nmi_panic",
 821                 .data           = &unknown_nmi_panic,
 822                 .maxlen         = sizeof (int),
 823                 .mode           = 0644,
 824                 .proc_handler   = proc_dointvec,
 825         },
 826 #endif
 827 #if defined(CONFIG_X86)
 828         {
 829                 .procname       = "panic_on_unrecovered_nmi",
 830                 .data           = &panic_on_unrecovered_nmi,
 831                 .maxlen         = sizeof(int),
 832                 .mode           = 0644,
 833                 .proc_handler   = proc_dointvec,
 834         },
 835         {
 836                 .procname       = "panic_on_io_nmi",
 837                 .data           = &panic_on_io_nmi,
 838                 .maxlen         = sizeof(int),
 839                 .mode           = 0644,
 840                 .proc_handler   = proc_dointvec,
 841         },
 842 #ifdef CONFIG_DEBUG_STACKOVERFLOW
 843         {
 844                 .procname       = "panic_on_stackoverflow",
 845                 .data           = &sysctl_panic_on_stackoverflow,
 846                 .maxlen         = sizeof(int),
 847                 .mode           = 0644,
 848                 .proc_handler   = proc_dointvec,
 849         },
 850 #endif
 851         {
 852                 .procname       = "bootloader_type",
 853                 .data           = &bootloader_type,
 854                 .maxlen         = sizeof (int),
 855                 .mode           = 0444,
 856                 .proc_handler   = proc_dointvec,
 857         },
 858         {
 859                 .procname       = "bootloader_version",
 860                 .data           = &bootloader_version,
 861                 .maxlen         = sizeof (int),
 862                 .mode           = 0444,
 863                 .proc_handler   = proc_dointvec,
 864         },
 865         {
 866                 .procname       = "kstack_depth_to_print",
 867                 .data           = &kstack_depth_to_print,
 868                 .maxlen         = sizeof(int),
 869                 .mode           = 0644,
 870                 .proc_handler   = proc_dointvec,
 871         },
 872         {
 873                 .procname       = "io_delay_type",
 874                 .data           = &io_delay_type,
 875                 .maxlen         = sizeof(int),
 876                 .mode           = 0644,
 877                 .proc_handler   = proc_dointvec,
 878         },
 879 #endif
 880 #if defined(CONFIG_MMU)
 881         {
 882                 .procname       = "randomize_va_space",
 883                 .data           = &randomize_va_space,
 884                 .maxlen         = sizeof(int),
 885                 .mode           = 0644,
 886                 .proc_handler   = proc_dointvec,
 887         },
 888 #endif
 889 #if defined(CONFIG_S390) && defined(CONFIG_SMP)
 890         {
 891                 .procname       = "spin_retry",
 892                 .data           = &spin_retry,
 893                 .maxlen         = sizeof (int),
 894                 .mode           = 0644,
 895                 .proc_handler   = proc_dointvec,
 896         },
 897 #endif
 898 #if     defined(CONFIG_ACPI_SLEEP) && defined(CONFIG_X86)
 899         {
 900                 .procname       = "acpi_video_flags",
 901                 .data           = &acpi_realmode_flags,
 902                 .maxlen         = sizeof (unsigned long),
 903                 .mode           = 0644,
 904                 .proc_handler   = proc_doulongvec_minmax,
 905         },
 906 #endif
 907 #ifdef CONFIG_IA64
 908         {
 909                 .procname       = "ignore-unaligned-usertrap",
 910                 .data           = &no_unaligned_warning,
 911                 .maxlen         = sizeof (int),
 912                 .mode           = 0644,
 913                 .proc_handler   = proc_dointvec,
 914         },
 915         {
 916                 .procname       = "unaligned-dump-stack",
 917                 .data           = &unaligned_dump_stack,
 918                 .maxlen         = sizeof (int),
 919                 .mode           = 0644,
 920                 .proc_handler   = proc_dointvec,
 921         },
 922 #endif
 923 #ifdef CONFIG_DETECT_HUNG_TASK
 924         {
 925                 .procname       = "hung_task_panic",
 926                 .data           = &sysctl_hung_task_panic,
 927                 .maxlen         = sizeof(int),
 928                 .mode           = 0644,
 929                 .proc_handler   = proc_dointvec_minmax,
 930                 .extra1         = &zero,
 931                 .extra2         = &one,
 932         },
 933         {
 934                 .procname       = "hung_task_check_count",
 935                 .data           = &sysctl_hung_task_check_count,
 936                 .maxlen         = sizeof(unsigned long),
 937                 .mode           = 0644,
 938                 .proc_handler   = proc_doulongvec_minmax,
 939         },
 940         {
 941                 .procname       = "hung_task_timeout_secs",
 942                 .data           = &sysctl_hung_task_timeout_secs,
 943                 .maxlen         = sizeof(unsigned long),
 944                 .mode           = 0644,
 945                 .proc_handler   = proc_dohung_task_timeout_secs,
 946         },
 947         {
 948                 .procname       = "hung_task_warnings",
 949                 .data           = &sysctl_hung_task_warnings,
 950                 .maxlen         = sizeof(unsigned long),
 951                 .mode           = 0644,
 952                 .proc_handler   = proc_doulongvec_minmax,
 953         },
 954 #endif
 955 #ifdef CONFIG_COMPAT
 956         {
 957                 .procname       = "compat-log",
 958                 .data           = &compat_log,
 959                 .maxlen         = sizeof (int),
 960                 .mode           = 0644,
 961                 .proc_handler   = proc_dointvec,
 962         },
 963 #endif
 964 #ifdef CONFIG_RT_MUTEXES
 965         {
 966                 .procname       = "max_lock_depth",
 967                 .data           = &max_lock_depth,
 968                 .maxlen         = sizeof(int),
 969                 .mode           = 0644,
 970                 .proc_handler   = proc_dointvec,
 971         },
 972 #endif
 973         {
 974                 .procname       = "poweroff_cmd",
 975                 .data           = &poweroff_cmd,
 976                 .maxlen         = POWEROFF_CMD_PATH_LEN,
 977                 .mode           = 0644,
 978                 .proc_handler   = proc_dostring,
 979         },
 980 #ifdef CONFIG_KEYS
 981         {
 982                 .procname       = "keys",
 983                 .mode           = 0555,
 984                 .child          = key_sysctls,
 985         },
 986 #endif
 987 #ifdef CONFIG_RCU_TORTURE_TEST
 988         {
 989                 .procname       = "rcutorture_runnable",
 990                 .data           = &rcutorture_runnable,
 991                 .maxlen         = sizeof(int),
 992                 .mode           = 0644,
 993                 .proc_handler   = proc_dointvec,
 994         },
 995 #endif
 996 #ifdef CONFIG_PERF_EVENTS
 997         /*
 998          * User-space scripts rely on the existence of this file
 999          * as a feature check for perf_events being enabled.
1000          *
1001          * So it's an ABI, do not remove!
1002          */
1003         {
1004                 .procname       = "perf_event_paranoid",
1005                 .data           = &sysctl_perf_event_paranoid,
1006                 .maxlen         = sizeof(sysctl_perf_event_paranoid),
1007                 .mode           = 0644,
1008                 .proc_handler   = proc_dointvec,
1009         },
1010         {
1011                 .procname       = "perf_event_mlock_kb",
1012                 .data           = &sysctl_perf_event_mlock,
1013                 .maxlen         = sizeof(sysctl_perf_event_mlock),
1014                 .mode           = 0644,
1015                 .proc_handler   = proc_dointvec,
1016         },
1017         {
1018                 .procname       = "perf_event_max_sample_rate",
1019                 .data           = &sysctl_perf_event_sample_rate,
1020                 .maxlen         = sizeof(sysctl_perf_event_sample_rate),
1021                 .mode           = 0644,
1022                 .proc_handler   = perf_proc_update_handler,
1023         },
1024 #endif
1025 #ifdef CONFIG_KMEMCHECK
1026         {
1027                 .procname       = "kmemcheck",
1028                 .data           = &kmemcheck_enabled,
1029                 .maxlen         = sizeof(int),
1030                 .mode           = 0644,
1031                 .proc_handler   = proc_dointvec,
1032         },
1033 #endif
1034 #ifdef CONFIG_BLOCK
1035         {
1036                 .procname       = "blk_iopoll",
1037                 .data           = &blk_iopoll_enabled,
1038                 .maxlen         = sizeof(int),
1039                 .mode           = 0644,
1040                 .proc_handler   = proc_dointvec,
1041         },
1042 #endif
1043         { }
1044 };
1045
1046 static struct ctl_table vm_table[] = {
1047         {
1048                 .procname       = "overcommit_memory",
1049                 .data           = &sysctl_overcommit_memory,
1050                 .maxlen         = sizeof(sysctl_overcommit_memory),
1051                 .mode           = 0644,
1052                 .proc_handler   = proc_dointvec_minmax,
1053                 .extra1         = &zero,
1054                 .extra2         = &two,
1055         },
1056         {
1057                 .procname       = "panic_on_oom",
1058                 .data           = &sysctl_panic_on_oom,
1059                 .maxlen         = sizeof(sysctl_panic_on_oom),
1060                 .mode           = 0644,
1061                 .proc_handler   = proc_dointvec_minmax,
1062                 .extra1         = &zero,
1063                 .extra2         = &two,
1064         },
1065         {
1066                 .procname       = "oom_kill_allocating_task",
1067                 .data           = &sysctl_oom_kill_allocating_task,
1068                 .maxlen         = sizeof(sysctl_oom_kill_allocating_task),
1069                 .mode           = 0644,
1070                 .proc_handler   = proc_dointvec,
1071         },
1072         {
1073                 .procname       = "oom_dump_tasks",
1074                 .data           = &sysctl_oom_dump_tasks,
1075                 .maxlen         = sizeof(sysctl_oom_dump_tasks),
1076                 .mode           = 0644,
1077                 .proc_handler   = proc_dointvec,
1078         },
1079         {
1080                 .procname       = "overcommit_ratio",
1081                 .data           = &sysctl_overcommit_ratio,
1082                 .maxlen         = sizeof(sysctl_overcommit_ratio),
1083                 .mode           = 0644,
1084                 .proc_handler   = proc_dointvec,
1085         },
1086         {
1087                 .procname       = "page-cluster",
1088                 .data           = &page_cluster,
1089                 .maxlen         = sizeof(int),
1090                 .mode           = 0644,
1091                 .proc_handler   = proc_dointvec_minmax,
1092                 .extra1         = &zero,
1093         },
1094         {
1095                 .procname       = "dirty_background_ratio",
1096                 .data           = &dirty_background_ratio,
1097                 .maxlen         = sizeof(dirty_background_ratio),
1098                 .mode           = 0644,
1099                 .proc_handler   = dirty_background_ratio_handler,
1100                 .extra1         = &zero,
1101                 .extra2         = &one_hundred,
1102         },
1103         {
1104                 .procname       = "dirty_background_bytes",
1105                 .data           = &dirty_background_bytes,
1106                 .maxlen         = sizeof(dirty_background_bytes),
1107                 .mode           = 0644,
1108                 .proc_handler   = dirty_background_bytes_handler,
1109                 .extra1         = &one_ul,
1110         },
1111         {
1112                 .procname       = "dirty_ratio",
1113                 .data           = &vm_dirty_ratio,
1114                 .maxlen         = sizeof(vm_dirty_ratio),
1115                 .mode           = 0644,
1116                 .proc_handler   = dirty_ratio_handler,
1117                 .extra1         = &zero,
1118                 .extra2         = &one_hundred,
1119         },
1120         {
1121                 .procname       = "dirty_bytes",
1122                 .data           = &vm_dirty_bytes,
1123                 .maxlen         = sizeof(vm_dirty_bytes),
1124                 .mode           = 0644,
1125                 .proc_handler   = dirty_bytes_handler,
1126                 .extra1         = &dirty_bytes_min,
1127         },
1128         {
1129                 .procname       = "dirty_writeback_centisecs",
1130                 .data           = &dirty_writeback_interval,
1131                 .maxlen         = sizeof(dirty_writeback_interval),
1132                 .mode           = 0644,
1133                 .proc_handler   = dirty_writeback_centisecs_handler,
1134         },
1135         {
1136                 .procname       = "dirty_expire_centisecs",
1137                 .data           = &dirty_expire_interval,
1138                 .maxlen         = sizeof(dirty_expire_interval),
1139                 .mode           = 0644,
1140                 .proc_handler   = proc_dointvec_minmax,
1141                 .extra1         = &zero,
1142         },
1143         {
1144                 .procname       = "nr_pdflush_threads",
1145                 .mode           = 0444 /* read-only */,
1146                 .proc_handler   = pdflush_proc_obsolete,
1147         },
1148         {
1149                 .procname       = "swappiness",
1150                 .data           = &vm_swappiness,
1151                 .maxlen         = sizeof(vm_swappiness),
1152                 .mode           = 0644,
1153                 .proc_handler   = proc_dointvec_minmax,
1154                 .extra1         = &zero,
1155                 .extra2         = &one_hundred,
1156         },
1157 #ifdef CONFIG_HUGETLB_PAGE
1158         {
1159                 .procname       = "nr_hugepages",
1160                 .data           = NULL,
1161                 .maxlen         = sizeof(unsigned long),
1162                 .mode           = 0644,
1163                 .proc_handler   = hugetlb_sysctl_handler,
1164                 .extra1         = (void *)&hugetlb_zero,
1165                 .extra2         = (void *)&hugetlb_infinity,
1166         },
1167 #ifdef CONFIG_NUMA
1168         {
1169                 .procname       = "nr_hugepages_mempolicy",
1170                 .data           = NULL,
1171                 .maxlen         = sizeof(unsigned long),
1172                 .mode           = 0644,
1173                 .proc_handler   = &hugetlb_mempolicy_sysctl_handler,
1174                 .extra1         = (void *)&hugetlb_zero,
1175                 .extra2         = (void *)&hugetlb_infinity,
1176         },
1177 #endif
1178          {
1179                 .procname       = "hugetlb_shm_group",
1180                 .data           = &sysctl_hugetlb_shm_group,
1181                 .maxlen         = sizeof(gid_t),
1182                 .mode           = 0644,
1183                 .proc_handler   = proc_dointvec,
1184          },
1185          {
1186                 .procname       = "hugepages_treat_as_movable",
1187                 .data           = &hugepages_treat_as_movable,
1188                 .maxlen         = sizeof(int),
1189                 .mode           = 0644,
1190                 .proc_handler   = hugetlb_treat_movable_handler,
1191         },
1192         {
1193                 .procname       = "nr_overcommit_hugepages",
1194                 .data           = NULL,
1195                 .maxlen         = sizeof(unsigned long),
1196                 .mode           = 0644,
1197                 .proc_handler   = hugetlb_overcommit_handler,
1198                 .extra1         = (void *)&hugetlb_zero,
1199                 .extra2         = (void *)&hugetlb_infinity,
1200         },
1201 #endif
1202         {
1203                 .procname       = "lowmem_reserve_ratio",
1204                 .data           = &sysctl_lowmem_reserve_ratio,
1205                 .maxlen         = sizeof(sysctl_lowmem_reserve_ratio),
1206                 .mode           = 0644,
1207                 .proc_handler   = lowmem_reserve_ratio_sysctl_handler,
1208         },
1209         {
1210                 .procname       = "drop_caches",
1211                 .data           = &sysctl_drop_caches,
1212                 .maxlen         = sizeof(int),
1213                 .mode           = 0644,
1214                 .proc_handler   = drop_caches_sysctl_handler,
1215                 .extra1         = &one,
1216                 .extra2         = &three,
1217         },
1218 #ifdef CONFIG_COMPACTION
1219         {
1220                 .procname       = "compact_memory",
1221                 .data           = &sysctl_compact_memory,
1222                 .maxlen         = sizeof(int),
1223                 .mode           = 0200,
1224                 .proc_handler   = sysctl_compaction_handler,
1225         },
1226         {
1227                 .procname       = "extfrag_threshold",
1228                 .data           = &sysctl_extfrag_threshold,
1229                 .maxlen         = sizeof(int),
1230                 .mode           = 0644,
1231                 .proc_handler   = sysctl_extfrag_handler,
1232                 .extra1         = &min_extfrag_threshold,
1233                 .extra2         = &max_extfrag_threshold,
1234         },
1235
1236 #endif /* CONFIG_COMPACTION */
1237         {
1238                 .procname       = "min_free_kbytes",
1239                 .data           = &min_free_kbytes,
1240                 .maxlen         = sizeof(min_free_kbytes),
1241                 .mode           = 0644,
1242                 .proc_handler   = min_free_kbytes_sysctl_handler,
1243                 .extra1         = &zero,
1244         },
1245         {
1246                 .procname       = "percpu_pagelist_fraction",
1247                 .data           = &percpu_pagelist_fraction,
1248                 .maxlen         = sizeof(percpu_pagelist_fraction),
1249                 .mode           = 0644,
1250                 .proc_handler   = percpu_pagelist_fraction_sysctl_handler,
1251                 .extra1         = &min_percpu_pagelist_fract,
1252         },
1253 #ifdef CONFIG_MMU
1254         {
1255                 .procname       = "max_map_count",
1256                 .data           = &sysctl_max_map_count,
1257                 .maxlen         = sizeof(sysctl_max_map_count),
1258                 .mode           = 0644,
1259                 .proc_handler   = proc_dointvec_minmax,
1260                 .extra1         = &zero,
1261         },
1262 #else
1263         {
1264                 .procname       = "nr_trim_pages",
1265                 .data           = &sysctl_nr_trim_pages,
1266                 .maxlen         = sizeof(sysctl_nr_trim_pages),
1267                 .mode           = 0644,
1268                 .proc_handler   = proc_dointvec_minmax,
1269                 .extra1         = &zero,
1270         },
1271 #endif
1272         {
1273                 .procname       = "laptop_mode",
1274                 .data           = &laptop_mode,
1275                 .maxlen         = sizeof(laptop_mode),
1276                 .mode           = 0644,
1277                 .proc_handler   = proc_dointvec_jiffies,
1278         },
1279         {
1280                 .procname       = "block_dump",
1281                 .data           = &block_dump,
1282                 .maxlen         = sizeof(block_dump),
1283                 .mode           = 0644,
1284                 .proc_handler   = proc_dointvec,
1285                 .extra1         = &zero,
1286         },
1287         {
1288                 .procname       = "vfs_cache_pressure",
1289                 .data           = &sysctl_vfs_cache_pressure,
1290                 .maxlen         = sizeof(sysctl_vfs_cache_pressure),
1291                 .mode           = 0644,
1292                 .proc_handler   = proc_dointvec,
1293                 .extra1         = &zero,
1294         },
1295 #ifdef HAVE_ARCH_PICK_MMAP_LAYOUT
1296         {
1297                 .procname       = "legacy_va_layout",
1298                 .data           = &sysctl_legacy_va_layout,
1299                 .maxlen         = sizeof(sysctl_legacy_va_layout),
1300                 .mode           = 0644,
1301                 .proc_handler   = proc_dointvec,
1302                 .extra1         = &zero,
1303         },
1304 #endif
1305 #ifdef CONFIG_NUMA
1306         {
1307                 .procname       = "zone_reclaim_mode",
1308                 .data           = &zone_reclaim_mode,
1309                 .maxlen         = sizeof(zone_reclaim_mode),
1310                 .mode           = 0644,
1311                 .proc_handler   = proc_dointvec,
1312                 .extra1         = &zero,
1313         },
1314         {
1315                 .procname       = "min_unmapped_ratio",
1316                 .data           = &sysctl_min_unmapped_ratio,
1317                 .maxlen         = sizeof(sysctl_min_unmapped_ratio),
1318                 .mode           = 0644,
1319                 .proc_handler   = sysctl_min_unmapped_ratio_sysctl_handler,
1320                 .extra1         = &zero,
1321                 .extra2         = &one_hundred,
1322         },
1323         {
1324                 .procname       = "min_slab_ratio",
1325                 .data           = &sysctl_min_slab_ratio,
1326                 .maxlen         = sizeof(sysctl_min_slab_ratio),
1327                 .mode           = 0644,
1328                 .proc_handler   = sysctl_min_slab_ratio_sysctl_handler,
1329                 .extra1         = &zero,
1330                 .extra2         = &one_hundred,
1331         },
1332 #endif
1333 #ifdef CONFIG_SMP
1334         {
1335                 .procname       = "stat_interval",
1336                 .data           = &sysctl_stat_interval,
1337                 .maxlen         = sizeof(sysctl_stat_interval),
1338                 .mode           = 0644,
1339                 .proc_handler   = proc_dointvec_jiffies,
1340         },
1341 #endif
1342 #ifdef CONFIG_MMU
1343         {
1344                 .procname       = "mmap_min_addr",
1345                 .data           = &dac_mmap_min_addr,
1346                 .maxlen         = sizeof(unsigned long),
1347                 .mode           = 0644,
1348                 .proc_handler   = mmap_min_addr_handler,
1349         },
1350 #endif
1351 #ifdef CONFIG_NUMA
1352         {
1353                 .procname       = "numa_zonelist_order",
1354                 .data           = &numa_zonelist_order,
1355                 .maxlen         = NUMA_ZONELIST_ORDER_LEN,
1356                 .mode           = 0644,
1357                 .proc_handler   = numa_zonelist_order_handler,
1358         },
1359 #endif
1360 #if (defined(CONFIG_X86_32) && !defined(CONFIG_UML))|| \
1361    (defined(CONFIG_SUPERH) && defined(CONFIG_VSYSCALL))
1362         {
1363                 .procname       = "vdso_enabled",
1364                 .data           = &vdso_enabled,
1365                 .maxlen         = sizeof(vdso_enabled),
1366                 .mode           = 0644,
1367                 .proc_handler   = proc_dointvec,
1368                 .extra1         = &zero,
1369         },
1370 #endif
1371 #ifdef CONFIG_HIGHMEM
1372         {
1373                 .procname       = "highmem_is_dirtyable",
1374                 .data           = &vm_highmem_is_dirtyable,
1375                 .maxlen         = sizeof(vm_highmem_is_dirtyable),
1376                 .mode           = 0644,
1377                 .proc_handler   = proc_dointvec_minmax,
1378                 .extra1         = &zero,
1379                 .extra2         = &one,
1380         },
1381 #endif
1382         {
1383                 .procname       = "scan_unevictable_pages",
1384                 .data           = &scan_unevictable_pages,
1385                 .maxlen         = sizeof(scan_unevictable_pages),
1386                 .mode           = 0644,
1387                 .proc_handler   = scan_unevictable_handler,
1388         },
1389 #ifdef CONFIG_MEMORY_FAILURE
1390         {
1391                 .procname       = "memory_failure_early_kill",
1392                 .data           = &sysctl_memory_failure_early_kill,
1393                 .maxlen         = sizeof(sysctl_memory_failure_early_kill),
1394                 .mode           = 0644,
1395                 .proc_handler   = proc_dointvec_minmax,
1396                 .extra1         = &zero,
1397                 .extra2         = &one,
1398         },
1399         {
1400                 .procname       = "memory_failure_recovery",
1401                 .data           = &sysctl_memory_failure_recovery,
1402                 .maxlen         = sizeof(sysctl_memory_failure_recovery),
1403                 .mode           = 0644,
1404                 .proc_handler   = proc_dointvec_minmax,
1405                 .extra1         = &zero,
1406                 .extra2         = &one,
1407         },
1408 #endif
1409         { }
1410 };
1411
1412 #if defined(CONFIG_BINFMT_MISC) || defined(CONFIG_BINFMT_MISC_MODULE)
1413 static struct ctl_table binfmt_misc_table[] = {
1414         { }
1415 };
1416 #endif
1417
1418 static struct ctl_table fs_table[] = {
1419         {
1420                 .procname       = "inode-nr",
1421                 .data           = &inodes_stat,
1422                 .maxlen         = 2*sizeof(int),
1423                 .mode           = 0444,
1424                 .proc_handler   = proc_nr_inodes,
1425         },
1426         {
1427                 .procname       = "inode-state",
1428                 .data           = &inodes_stat,
1429                 .maxlen         = 7*sizeof(int),
1430                 .mode           = 0444,
1431                 .proc_handler   = proc_nr_inodes,
1432         },
1433         {
1434                 .procname       = "file-nr",
1435                 .data           = &files_stat,
1436                 .maxlen         = sizeof(files_stat),
1437                 .mode           = 0444,
1438                 .proc_handler   = proc_nr_files,
1439         },
1440         {
1441                 .procname       = "file-max",
1442                 .data           = &files_stat.max_files,
1443                 .maxlen         = sizeof(files_stat.max_files),
1444                 .mode           = 0644,
1445                 .proc_handler   = proc_doulongvec_minmax,
1446         },
1447         {
1448                 .procname       = "nr_open",
1449                 .data           = &sysctl_nr_open,
1450                 .maxlen         = sizeof(int),
1451                 .mode           = 0644,
1452                 .proc_handler   = proc_dointvec_minmax,
1453                 .extra1         = &sysctl_nr_open_min,
1454                 .extra2         = &sysctl_nr_open_max,
1455         },
1456         {
1457                 .procname       = "dentry-state",
1458                 .data           = &dentry_stat,
1459                 .maxlen         = 6*sizeof(int),
1460                 .mode           = 0444,
1461                 .proc_handler   = proc_nr_dentry,
1462         },
1463         {
1464                 .procname       = "overflowuid",
1465                 .data           = &fs_overflowuid,
1466                 .maxlen         = sizeof(int),
1467                 .mode           = 0644,
1468                 .proc_handler   = proc_dointvec_minmax,
1469                 .extra1         = &minolduid,
1470                 .extra2         = &maxolduid,
1471         },
1472         {
1473                 .procname       = "overflowgid",
1474                 .data           = &fs_overflowgid,
1475                 .maxlen         = sizeof(int),
1476                 .mode           = 0644,
1477                 .proc_handler   = proc_dointvec_minmax,
1478                 .extra1         = &minolduid,
1479                 .extra2         = &maxolduid,
1480         },
1481 #ifdef CONFIG_FILE_LOCKING
1482         {
1483                 .procname       = "leases-enable",
1484                 .data           = &leases_enable,
1485                 .maxlen         = sizeof(int),
1486                 .mode           = 0644,
1487                 .proc_handler   = proc_dointvec,
1488         },
1489 #endif
1490 #ifdef CONFIG_DNOTIFY
1491         {
1492                 .procname       = "dir-notify-enable",
1493                 .data           = &dir_notify_enable,
1494                 .maxlen         = sizeof(int),
1495                 .mode           = 0644,
1496                 .proc_handler   = proc_dointvec,
1497         },
1498 #endif
1499 #ifdef CONFIG_MMU
1500 #ifdef CONFIG_FILE_LOCKING
1501         {
1502                 .procname       = "lease-break-time",
1503                 .data           = &lease_break_time,
1504                 .maxlen         = sizeof(int),
1505                 .mode           = 0644,
1506                 .proc_handler   = proc_dointvec,
1507         },
1508 #endif
1509 #ifdef CONFIG_AIO
1510         {
1511                 .procname       = "aio-nr",
1512                 .data           = &aio_nr,
1513                 .maxlen         = sizeof(aio_nr),
1514                 .mode           = 0444,
1515                 .proc_handler   = proc_doulongvec_minmax,
1516         },
1517         {
1518                 .procname       = "aio-max-nr",
1519                 .data           = &aio_max_nr,
1520                 .maxlen         = sizeof(aio_max_nr),
1521                 .mode           = 0644,
1522                 .proc_handler   = proc_doulongvec_minmax,
1523         },
1524 #endif /* CONFIG_AIO */
1525 #ifdef CONFIG_INOTIFY_USER
1526         {
1527                 .procname       = "inotify",
1528                 .mode           = 0555,
1529                 .child          = inotify_table,
1530         },
1531 #endif
1532 #ifdef CONFIG_EPOLL
1533         {
1534                 .procname       = "epoll",
1535                 .mode           = 0555,
1536                 .child          = epoll_table,
1537         },
1538 #endif
1539 #endif
1540         {
1541                 .procname       = "protected_symlinks",
1542                 .data           = &sysctl_protected_symlinks,
1543                 .maxlen         = sizeof(int),
1544                 .mode           = 0600,
1545                 .proc_handler   = proc_dointvec_minmax,
1546                 .extra1         = &zero,
1547                 .extra2         = &one,
1548         },
1549         {
1550                 .procname       = "protected_hardlinks",
1551                 .data           = &sysctl_protected_hardlinks,
1552                 .maxlen         = sizeof(int),
1553                 .mode           = 0600,
1554                 .proc_handler   = proc_dointvec_minmax,
1555                 .extra1         = &zero,
1556                 .extra2         = &one,
1557         },
1558         {
1559                 .procname       = "suid_dumpable",
1560                 .data           = &suid_dumpable,
1561                 .maxlen         = sizeof(int),
1562                 .mode           = 0644,
1563                 .proc_handler   = proc_dointvec_minmax_coredump,
1564                 .extra1         = &zero,
1565                 .extra2         = &two,
1566         },
1567 #if defined(CONFIG_BINFMT_MISC) || defined(CONFIG_BINFMT_MISC_MODULE)
1568         {
1569                 .procname       = "binfmt_misc",
1570                 .mode           = 0555,
1571                 .child          = binfmt_misc_table,
1572         },
1573 #endif
1574         {
1575                 .procname       = "pipe-max-size",
1576                 .data           = &pipe_max_size,
1577                 .maxlen         = sizeof(int),
1578                 .mode           = 0644,
1579                 .proc_handler   = &pipe_proc_fn,
1580                 .extra1         = &pipe_min_size,
1581         },
1582         { }
1583 };
1584
1585 static struct ctl_table debug_table[] = {
1586 #ifdef CONFIG_SYSCTL_EXCEPTION_TRACE
1587         {
1588                 .procname       = "exception-trace",
1589                 .data           = &show_unhandled_signals,
1590                 .maxlen         = sizeof(int),
1591                 .mode           = 0644,
1592                 .proc_handler   = proc_dointvec
1593         },
1594 #endif
1595 #if defined(CONFIG_OPTPROBES)
1596         {
1597                 .procname       = "kprobes-optimization",
1598                 .data           = &sysctl_kprobes_optimization,
1599                 .maxlen         = sizeof(int),
1600                 .mode           = 0644,
1601                 .proc_handler   = proc_kprobes_optimization_handler,
1602                 .extra1         = &zero,
1603                 .extra2         = &one,
1604         },
1605 #endif
1606         { }
1607 };
1608
1609 static struct ctl_table dev_table[] = {
1610         { }
1611 };
1612
1613 int __init sysctl_init(void)
1614 {
1615         struct ctl_table_header *hdr;
1616
1617         hdr = register_sysctl_table(sysctl_base_table);
1618         kmemleak_not_leak(hdr);
1619         return 0;
1620 }
1621
1622 #endif /* CONFIG_SYSCTL */
1623
1624 /*
1625  * /proc/sys support
1626  */
1627
1628 #ifdef CONFIG_PROC_SYSCTL
1629
1630 static int _proc_do_string(void* data, int maxlen, int write,
1631                            void __user *buffer,
1632                            size_t *lenp, loff_t *ppos)
1633 {
1634         size_t len;
1635         char __user *p;
1636         char c;
1637
1638         if (!data || !maxlen || !*lenp) {
1639                 *lenp = 0;
1640                 return 0;
1641         }
1642
1643         if (write) {
1644                 len = 0;
1645                 p = buffer;
1646                 while (len < *lenp) {
1647                         if (get_user(c, p++))
1648                                 return -EFAULT;
1649                         if (c == 0 || c == '\n')
1650                                 break;
1651                         len++;
1652                 }
1653                 if (len >= maxlen)
1654                         len = maxlen-1;
1655                 if(copy_from_user(data, buffer, len))
1656                         return -EFAULT;
1657                 ((char *) data)[len] = 0;
1658                 *ppos += *lenp;
1659         } else {
1660                 len = strlen(data);
1661                 if (len > maxlen)
1662                         len = maxlen;
1663
1664                 if (*ppos > len) {
1665                         *lenp = 0;
1666                         return 0;
1667                 }
1668
1669                 data += *ppos;
1670                 len  -= *ppos;
1671
1672                 if (len > *lenp)
1673                         len = *lenp;
1674                 if (len)
1675                         if(copy_to_user(buffer, data, len))
1676                                 return -EFAULT;
1677                 if (len < *lenp) {
1678                         if(put_user('\n', ((char __user *) buffer) + len))
1679                                 return -EFAULT;
1680                         len++;
1681                 }
1682                 *lenp = len;
1683                 *ppos += len;
1684         }
1685         return 0;
1686 }
1687
1688 /**
1689  * proc_dostring - read a string sysctl
1690  * @table: the sysctl table
1691  * @write: %TRUE if this is a write to the sysctl file
1692  * @buffer: the user buffer
1693  * @lenp: the size of the user buffer
1694  * @ppos: file position
1695  *
1696  * Reads/writes a string from/to the user buffer. If the kernel
1697  * buffer provided is not large enough to hold the string, the
1698  * string is truncated. The copied string is %NULL-terminated.
1699  * If the string is being read by the user process, it is copied
1700  * and a newline '\n' is added. It is truncated if the buffer is
1701  * not large enough.
1702  *
1703  * Returns 0 on success.
1704  */
1705 int proc_dostring(struct ctl_table *table, int write,
1706                   void __user *buffer, size_t *lenp, loff_t *ppos)
1707 {
1708         return _proc_do_string(table->data, table->maxlen, write,
1709                                buffer, lenp, ppos);
1710 }
1711
1712 static size_t proc_skip_spaces(char **buf)
1713 {
1714         size_t ret;
1715         char *tmp = skip_spaces(*buf);
1716         ret = tmp - *buf;
1717         *buf = tmp;
1718         return ret;
1719 }
1720
1721 static void proc_skip_char(char **buf, size_t *size, const char v)
1722 {
1723         while (*size) {
1724                 if (**buf != v)
1725                         break;
1726                 (*size)--;
1727                 (*buf)++;
1728         }
1729 }
1730
1731 #define TMPBUFLEN 22
1732 /**
1733  * proc_get_long - reads an ASCII formatted integer from a user buffer
1734  *
1735  * @buf: a kernel buffer
1736  * @size: size of the kernel buffer
1737  * @val: this is where the number will be stored
1738  * @neg: set to %TRUE if number is negative
1739  * @perm_tr: a vector which contains the allowed trailers
1740  * @perm_tr_len: size of the perm_tr vector
1741  * @tr: pointer to store the trailer character
1742  *
1743  * In case of success %0 is returned and @buf and @size are updated with
1744  * the amount of bytes read. If @tr is non-NULL and a trailing
1745  * character exists (size is non-zero after returning from this
1746  * function), @tr is updated with the trailing character.
1747  */
1748 static int proc_get_long(char **buf, size_t *size,
1749                           unsigned long *val, bool *neg,
1750                           const char *perm_tr, unsigned perm_tr_len, char *tr)
1751 {
1752         int len;
1753         char *p, tmp[TMPBUFLEN];
1754
1755         if (!*size)
1756                 return -EINVAL;
1757
1758         len = *size;
1759         if (len > TMPBUFLEN - 1)
1760                 len = TMPBUFLEN - 1;
1761
1762         memcpy(tmp, *buf, len);
1763
1764         tmp[len] = 0;
1765         p = tmp;
1766         if (*p == '-' && *size > 1) {
1767                 *neg = true;
1768                 p++;
1769         } else
1770                 *neg = false;
1771         if (!isdigit(*p))
1772                 return -EINVAL;
1773
1774         *val = simple_strtoul(p, &p, 0);
1775
1776         len = p - tmp;
1777
1778         /* We don't know if the next char is whitespace thus we may accept
1779          * invalid integers (e.g. 1234...a) or two integers instead of one
1780          * (e.g. 123...1). So lets not allow such large numbers. */
1781         if (len == TMPBUFLEN - 1)
1782                 return -EINVAL;
1783
1784         if (len < *size && perm_tr_len && !memchr(perm_tr, *p, perm_tr_len))
1785                 return -EINVAL;
1786
1787         if (tr && (len < *size))
1788                 *tr = *p;
1789
1790         *buf += len;
1791         *size -= len;
1792
1793         return 0;
1794 }
1795
1796 /**
1797  * proc_put_long - converts an integer to a decimal ASCII formatted string
1798  *
1799  * @buf: the user buffer
1800  * @size: the size of the user buffer
1801  * @val: the integer to be converted
1802  * @neg: sign of the number, %TRUE for negative
1803  *
1804  * In case of success %0 is returned and @buf and @size are updated with
1805  * the amount of bytes written.
1806  */
1807 static int proc_put_long(void __user **buf, size_t *size, unsigned long val,
1808                           bool neg)
1809 {
1810         int len;
1811         char tmp[TMPBUFLEN], *p = tmp;
1812
1813         sprintf(p, "%s%lu", neg ? "-" : "", val);
1814         len = strlen(tmp);
1815         if (len > *size)
1816                 len = *size;
1817         if (copy_to_user(*buf, tmp, len))
1818                 return -EFAULT;
1819         *size -= len;
1820         *buf += len;
1821         return 0;
1822 }
1823 #undef TMPBUFLEN
1824
1825 static int proc_put_char(void __user **buf, size_t *size, char c)
1826 {
1827         if (*size) {
1828                 char __user **buffer = (char __user **)buf;
1829                 if (put_user(c, *buffer))
1830                         return -EFAULT;
1831                 (*size)--, (*buffer)++;
1832                 *buf = *buffer;
1833         }
1834         return 0;
1835 }
1836
1837 static int do_proc_dointvec_conv(bool *negp, unsigned long *lvalp,
1838                                  int *valp,
1839                                  int write, void *data)
1840 {
1841         if (write) {
1842                 *valp = *negp ? -*lvalp : *lvalp;
1843         } else {
1844                 int val = *valp;
1845                 if (val < 0) {
1846                         *negp = true;
1847                         *lvalp = (unsigned long)-val;
1848                 } else {
1849                         *negp = false;
1850                         *lvalp = (unsigned long)val;
1851                 }
1852         }
1853         return 0;
1854 }
1855
1856 static const char proc_wspace_sep[] = { ' ', '\t', '\n' };
1857
1858 static int __do_proc_dointvec(void *tbl_data, struct ctl_table *table,
1859                   int write, void __user *buffer,
1860                   size_t *lenp, loff_t *ppos,
1861                   int (*conv)(bool *negp, unsigned long *lvalp, int *valp,
1862                               int write, void *data),
1863                   void *data)
1864 {
1865         int *i, vleft, first = 1, err = 0;
1866         unsigned long page = 0;
1867         size_t left;
1868         char *kbuf;
1869
1870         if (!tbl_data || !table->maxlen || !*lenp || (*ppos && !write)) {
1871                 *lenp = 0;
1872                 return 0;
1873         }
1874
1875         i = (int *) tbl_data;
1876         vleft = table->maxlen / sizeof(*i);
1877         left = *lenp;
1878
1879         if (!conv)
1880                 conv = do_proc_dointvec_conv;
1881
1882         if (write) {
1883                 if (left > PAGE_SIZE - 1)
1884                         left = PAGE_SIZE - 1;
1885                 page = __get_free_page(GFP_TEMPORARY);
1886                 kbuf = (char *) page;
1887                 if (!kbuf)
1888                         return -ENOMEM;
1889                 if (copy_from_user(kbuf, buffer, left)) {
1890                         err = -EFAULT;
1891                         goto free;
1892                 }
1893                 kbuf[left] = 0;
1894         }
1895
1896         for (; left && vleft--; i++, first=0) {
1897                 unsigned long lval;
1898                 bool neg;
1899
1900                 if (write) {
1901                         left -= proc_skip_spaces(&kbuf);
1902
1903                         if (!left)
1904                                 break;
1905                         err = proc_get_long(&kbuf, &left, &lval, &neg,
1906                                              proc_wspace_sep,
1907                                              sizeof(proc_wspace_sep), NULL);
1908                         if (err)
1909                                 break;
1910                         if (conv(&neg, &lval, i, 1, data)) {
1911                                 err = -EINVAL;
1912                                 break;
1913                         }
1914                 } else {
1915                         if (conv(&neg, &lval, i, 0, data)) {
1916                                 err = -EINVAL;
1917                                 break;
1918                         }
1919                         if (!first)
1920                                 err = proc_put_char(&buffer, &left, '\t');
1921                         if (err)
1922                                 break;
1923                         err = proc_put_long(&buffer, &left, lval, neg);
1924                         if (err)
1925                                 break;
1926                 }
1927         }
1928
1929         if (!write && !first && left && !err)
1930                 err = proc_put_char(&buffer, &left, '\n');
1931         if (write && !err && left)
1932                 left -= proc_skip_spaces(&kbuf);
1933 free:
1934         if (write) {
1935                 free_page(page);
1936                 if (first)
1937                         return err ? : -EINVAL;
1938         }
1939         *lenp -= left;
1940         *ppos += *lenp;
1941         return err;
1942 }
1943
1944 static int do_proc_dointvec(struct ctl_table *table, int write,
1945                   void __user *buffer, size_t *lenp, loff_t *ppos,
1946                   int (*conv)(bool *negp, unsigned long *lvalp, int *valp,
1947                               int write, void *data),
1948                   void *data)
1949 {
1950         return __do_proc_dointvec(table->data, table, write,
1951                         buffer, lenp, ppos, conv, data);
1952 }
1953
1954 /**
1955  * proc_dointvec - read a vector of integers
1956  * @table: the sysctl table
1957  * @write: %TRUE if this is a write to the sysctl file
1958  * @buffer: the user buffer
1959  * @lenp: the size of the user buffer
1960  * @ppos: file position
1961  *
1962  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
1963  * values from/to the user buffer, treated as an ASCII string.
1964  *
1965  * Returns 0 on success.
1966  */
1967 int proc_dointvec(struct ctl_table *table, int write,
1968                      void __user *buffer, size_t *lenp, loff_t *ppos)
1969 {
1970     return do_proc_dointvec(table,write,buffer,lenp,ppos,
1971                             NULL,NULL);
1972 }
1973
1974 /*
1975  * Taint values can only be increased
1976  * This means we can safely use a temporary.
1977  */
1978 static int proc_taint(struct ctl_table *table, int write,
1979                                void __user *buffer, size_t *lenp, loff_t *ppos)
1980 {
1981         struct ctl_table t;
1982         unsigned long tmptaint = get_taint();
1983         int err;
1984
1985         if (write && !capable(CAP_SYS_ADMIN))
1986                 return -EPERM;
1987
1988         t = *table;
1989         t.data = &tmptaint;
1990         err = proc_doulongvec_minmax(&t, write, buffer, lenp, ppos);
1991         if (err < 0)
1992                 return err;
1993
1994         if (write) {
1995                 /*
1996                  * Poor man's atomic or. Not worth adding a primitive
1997                  * to everyone's atomic.h for this
1998                  */
1999                 int i;
2000                 for (i = 0; i < BITS_PER_LONG && tmptaint >> i; i++) {
2001                         if ((tmptaint >> i) & 1)
2002                                 add_taint(i);
2003                 }
2004         }
2005
2006         return err;
2007 }
2008
2009 #ifdef CONFIG_PRINTK
2010 static int proc_dointvec_minmax_sysadmin(struct ctl_table *table, int write,
2011                                 void __user *buffer, size_t *lenp, loff_t *ppos)
2012 {
2013         if (write && !capable(CAP_SYS_ADMIN))
2014                 return -EPERM;
2015
2016         return proc_dointvec_minmax(table, write, buffer, lenp, ppos);
2017 }
2018 #endif
2019
2020 struct do_proc_dointvec_minmax_conv_param {
2021         int *min;
2022         int *max;
2023 };
2024
2025 static int do_proc_dointvec_minmax_conv(bool *negp, unsigned long *lvalp,
2026                                         int *valp,
2027                                         int write, void *data)
2028 {
2029         struct do_proc_dointvec_minmax_conv_param *param = data;
2030         if (write) {
2031                 int val = *negp ? -*lvalp : *lvalp;
2032                 if ((param->min && *param->min > val) ||
2033                     (param->max && *param->max < val))
2034                         return -EINVAL;
2035                 *valp = val;
2036         } else {
2037                 int val = *valp;
2038                 if (val < 0) {
2039                         *negp = true;
2040                         *lvalp = (unsigned long)-val;
2041                 } else {
2042                         *negp = false;
2043                         *lvalp = (unsigned long)val;
2044                 }
2045         }
2046         return 0;
2047 }
2048
2049 /**
2050  * proc_dointvec_minmax - read a vector of integers with min/max values
2051  * @table: the sysctl table
2052  * @write: %TRUE if this is a write to the sysctl file
2053  * @buffer: the user buffer
2054  * @lenp: the size of the user buffer
2055  * @ppos: file position
2056  *
2057  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2058  * values from/to the user buffer, treated as an ASCII string.
2059  *
2060  * This routine will ensure the values are within the range specified by
2061  * table->extra1 (min) and table->extra2 (max).
2062  *
2063  * Returns 0 on success.
2064  */
2065 int proc_dointvec_minmax(struct ctl_table *table, int write,
2066                   void __user *buffer, size_t *lenp, loff_t *ppos)
2067 {
2068         struct do_proc_dointvec_minmax_conv_param param = {
2069                 .min = (int *) table->extra1,
2070                 .max = (int *) table->extra2,
2071         };
2072         return do_proc_dointvec(table, write, buffer, lenp, ppos,
2073                                 do_proc_dointvec_minmax_conv, &param);
2074 }
2075
2076 static void validate_coredump_safety(void)
2077 {
2078 #ifdef CONFIG_COREDUMP
2079         if (suid_dumpable == SUID_DUMPABLE_SAFE &&
2080             core_pattern[0] != '/' && core_pattern[0] != '|') {
2081                 printk(KERN_WARNING "Unsafe core_pattern used with "\
2082                         "suid_dumpable=2. Pipe handler or fully qualified "\
2083                         "core dump path required.\n");
2084         }
2085 #endif
2086 }
2087
2088 static int proc_dointvec_minmax_coredump(struct ctl_table *table, int write,
2089                 void __user *buffer, size_t *lenp, loff_t *ppos)
2090 {
2091         int error = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
2092         if (!error)
2093                 validate_coredump_safety();
2094         return error;
2095 }
2096
2097 #ifdef CONFIG_COREDUMP
2098 static int proc_dostring_coredump(struct ctl_table *table, int write,
2099                   void __user *buffer, size_t *lenp, loff_t *ppos)
2100 {
2101         int error = proc_dostring(table, write, buffer, lenp, ppos);
2102         if (!error)
2103                 validate_coredump_safety();
2104         return error;
2105 }
2106 #endif
2107
2108 static int __do_proc_doulongvec_minmax(void *data, struct ctl_table *table, int write,
2109                                      void __user *buffer,
2110                                      size_t *lenp, loff_t *ppos,
2111                                      unsigned long convmul,
2112                                      unsigned long convdiv)
2113 {
2114         unsigned long *i, *min, *max;
2115         int vleft, first = 1, err = 0;
2116         unsigned long page = 0;
2117         size_t left;
2118         char *kbuf;
2119
2120         if (!data || !table->maxlen || !*lenp || (*ppos && !write)) {
2121                 *lenp = 0;
2122                 return 0;
2123         }
2124
2125         i = (unsigned long *) data;
2126         min = (unsigned long *) table->extra1;
2127         max = (unsigned long *) table->extra2;
2128         vleft = table->maxlen / sizeof(unsigned long);
2129         left = *lenp;
2130
2131         if (write) {
2132                 if (left > PAGE_SIZE - 1)
2133                         left = PAGE_SIZE - 1;
2134                 page = __get_free_page(GFP_TEMPORARY);
2135                 kbuf = (char *) page;
2136                 if (!kbuf)
2137                         return -ENOMEM;
2138                 if (copy_from_user(kbuf, buffer, left)) {
2139                         err = -EFAULT;
2140                         goto free;
2141                 }
2142                 kbuf[left] = 0;
2143         }
2144
2145         for (; left && vleft--; i++, first = 0) {
2146                 unsigned long val;
2147
2148                 if (write) {
2149                         bool neg;
2150
2151                         left -= proc_skip_spaces(&kbuf);
2152
2153                         err = proc_get_long(&kbuf, &left, &val, &neg,
2154                                              proc_wspace_sep,
2155                                              sizeof(proc_wspace_sep), NULL);
2156                         if (err)
2157                                 break;
2158                         if (neg)
2159                                 continue;
2160                         if ((min && val < *min) || (max && val > *max))
2161                                 continue;
2162                         *i = val;
2163                 } else {
2164                         val = convdiv * (*i) / convmul;
2165                         if (!first)
2166                                 err = proc_put_char(&buffer, &left, '\t');
2167                         err = proc_put_long(&buffer, &left, val, false);
2168                         if (err)
2169                                 break;
2170                 }
2171         }
2172
2173         if (!write && !first && left && !err)
2174                 err = proc_put_char(&buffer, &left, '\n');
2175         if (write && !err)
2176                 left -= proc_skip_spaces(&kbuf);
2177 free:
2178         if (write) {
2179                 free_page(page);
2180                 if (first)
2181                         return err ? : -EINVAL;
2182         }
2183         *lenp -= left;
2184         *ppos += *lenp;
2185         return err;
2186 }
2187
2188 static int do_proc_doulongvec_minmax(struct ctl_table *table, int write,
2189                                      void __user *buffer,
2190                                      size_t *lenp, loff_t *ppos,
2191                                      unsigned long convmul,
2192                                      unsigned long convdiv)
2193 {
2194         return __do_proc_doulongvec_minmax(table->data, table, write,
2195                         buffer, lenp, ppos, convmul, convdiv);
2196 }
2197
2198 /**
2199  * proc_doulongvec_minmax - read a vector of long integers with min/max values
2200  * @table: the sysctl table
2201  * @write: %TRUE if this is a write to the sysctl file
2202  * @buffer: the user buffer
2203  * @lenp: the size of the user buffer
2204  * @ppos: file position
2205  *
2206  * Reads/writes up to table->maxlen/sizeof(unsigned long) unsigned long
2207  * values from/to the user buffer, treated as an ASCII string.
2208  *
2209  * This routine will ensure the values are within the range specified by
2210  * table->extra1 (min) and table->extra2 (max).
2211  *
2212  * Returns 0 on success.
2213  */
2214 int proc_doulongvec_minmax(struct ctl_table *table, int write,
2215                            void __user *buffer, size_t *lenp, loff_t *ppos)
2216 {
2217     return do_proc_doulongvec_minmax(table, write, buffer, lenp, ppos, 1l, 1l);
2218 }
2219
2220 /**
2221  * proc_doulongvec_ms_jiffies_minmax - read a vector of millisecond values with min/max values
2222  * @table: the sysctl table
2223  * @write: %TRUE if this is a write to the sysctl file
2224  * @buffer: the user buffer
2225  * @lenp: the size of the user buffer
2226  * @ppos: file position
2227  *
2228  * Reads/writes up to table->maxlen/sizeof(unsigned long) unsigned long
2229  * values from/to the user buffer, treated as an ASCII string. The values
2230  * are treated as milliseconds, and converted to jiffies when they are stored.
2231  *
2232  * This routine will ensure the values are within the range specified by
2233  * table->extra1 (min) and table->extra2 (max).
2234  *
2235  * Returns 0 on success.
2236  */
2237 int proc_doulongvec_ms_jiffies_minmax(struct ctl_table *table, int write,
2238                                       void __user *buffer,
2239                                       size_t *lenp, loff_t *ppos)
2240 {
2241     return do_proc_doulongvec_minmax(table, write, buffer,
2242                                      lenp, ppos, HZ, 1000l);
2243 }
2244
2245
2246 static int do_proc_dointvec_jiffies_conv(bool *negp, unsigned long *lvalp,
2247                                          int *valp,
2248                                          int write, void *data)
2249 {
2250         if (write) {
2251                 if (*lvalp > LONG_MAX / HZ)
2252                         return 1;
2253                 *valp = *negp ? -(*lvalp*HZ) : (*lvalp*HZ);
2254         } else {
2255                 int val = *valp;
2256                 unsigned long lval;
2257                 if (val < 0) {
2258                         *negp = true;
2259                         lval = (unsigned long)-val;
2260                 } else {
2261                         *negp = false;
2262                         lval = (unsigned long)val;
2263                 }
2264                 *lvalp = lval / HZ;
2265         }
2266         return 0;
2267 }
2268
2269 static int do_proc_dointvec_userhz_jiffies_conv(bool *negp, unsigned long *lvalp,
2270                                                 int *valp,
2271                                                 int write, void *data)
2272 {
2273         if (write) {
2274                 if (USER_HZ < HZ && *lvalp > (LONG_MAX / HZ) * USER_HZ)
2275                         return 1;
2276                 *valp = clock_t_to_jiffies(*negp ? -*lvalp : *lvalp);
2277         } else {
2278                 int val = *valp;
2279                 unsigned long lval;
2280                 if (val < 0) {
2281                         *negp = true;
2282                         lval = (unsigned long)-val;
2283                 } else {
2284                         *negp = false;
2285                         lval = (unsigned long)val;
2286                 }
2287                 *lvalp = jiffies_to_clock_t(lval);
2288         }
2289         return 0;
2290 }
2291
2292 static int do_proc_dointvec_ms_jiffies_conv(bool *negp, unsigned long *lvalp,
2293                                             int *valp,
2294                                             int write, void *data)
2295 {
2296         if (write) {
2297                 *valp = msecs_to_jiffies(*negp ? -*lvalp : *lvalp);
2298         } else {
2299                 int val = *valp;
2300                 unsigned long lval;
2301                 if (val < 0) {
2302                         *negp = true;
2303                         lval = (unsigned long)-val;
2304                 } else {
2305                         *negp = false;
2306                         lval = (unsigned long)val;
2307                 }
2308                 *lvalp = jiffies_to_msecs(lval);
2309         }
2310         return 0;
2311 }
2312
2313 /**
2314  * proc_dointvec_jiffies - read a vector of integers as seconds
2315  * @table: the sysctl table
2316  * @write: %TRUE if this is a write to the sysctl file
2317  * @buffer: the user buffer
2318  * @lenp: the size of the user buffer
2319  * @ppos: file position
2320  *
2321  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2322  * values from/to the user buffer, treated as an ASCII string.
2323  * The values read are assumed to be in seconds, and are converted into
2324  * jiffies.
2325  *
2326  * Returns 0 on success.
2327  */
2328 int proc_dointvec_jiffies(struct ctl_table *table, int write,
2329                           void __user *buffer, size_t *lenp, loff_t *ppos)
2330 {
2331     return do_proc_dointvec(table,write,buffer,lenp,ppos,
2332                             do_proc_dointvec_jiffies_conv,NULL);
2333 }
2334
2335 /**
2336  * proc_dointvec_userhz_jiffies - read a vector of integers as 1/USER_HZ seconds
2337  * @table: the sysctl table
2338  * @write: %TRUE if this is a write to the sysctl file
2339  * @buffer: the user buffer
2340  * @lenp: the size of the user buffer
2341  * @ppos: pointer to the file position
2342  *
2343  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2344  * values from/to the user buffer, treated as an ASCII string.
2345  * The values read are assumed to be in 1/USER_HZ seconds, and
2346  * are converted into jiffies.
2347  *
2348  * Returns 0 on success.
2349  */
2350 int proc_dointvec_userhz_jiffies(struct ctl_table *table, int write,
2351                                  void __user *buffer, size_t *lenp, loff_t *ppos)
2352 {
2353     return do_proc_dointvec(table,write,buffer,lenp,ppos,
2354                             do_proc_dointvec_userhz_jiffies_conv,NULL);
2355 }
2356
2357 /**
2358  * proc_dointvec_ms_jiffies - read a vector of integers as 1 milliseconds
2359  * @table: the sysctl table
2360  * @write: %TRUE if this is a write to the sysctl file
2361  * @buffer: the user buffer
2362  * @lenp: the size of the user buffer
2363  * @ppos: file position
2364  * @ppos: the current position in the file
2365  *
2366  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2367  * values from/to the user buffer, treated as an ASCII string.
2368  * The values read are assumed to be in 1/1000 seconds, and
2369  * are converted into jiffies.
2370  *
2371  * Returns 0 on success.
2372  */
2373 int proc_dointvec_ms_jiffies(struct ctl_table *table, int write,
2374                              void __user *buffer, size_t *lenp, loff_t *ppos)
2375 {
2376         return do_proc_dointvec(table, write, buffer, lenp, ppos,
2377                                 do_proc_dointvec_ms_jiffies_conv, NULL);
2378 }
2379
2380 static int proc_do_cad_pid(struct ctl_table *table, int write,
2381                            void __user *buffer, size_t *lenp, loff_t *ppos)
2382 {
2383         struct pid *new_pid;
2384         pid_t tmp;
2385         int r;
2386
2387         tmp = pid_vnr(cad_pid);
2388
2389         r = __do_proc_dointvec(&tmp, table, write, buffer,
2390                                lenp, ppos, NULL, NULL);
2391         if (r || !write)
2392                 return r;
2393
2394         new_pid = find_get_pid(tmp);
2395         if (!new_pid)
2396                 return -ESRCH;
2397
2398         put_pid(xchg(&cad_pid, new_pid));
2399         return 0;
2400 }
2401
2402 /**
2403  * proc_do_large_bitmap - read/write from/to a large bitmap
2404  * @table: the sysctl table
2405  * @write: %TRUE if this is a write to the sysctl file
2406  * @buffer: the user buffer
2407  * @lenp: the size of the user buffer
2408  * @ppos: file position
2409  *
2410  * The bitmap is stored at table->data and the bitmap length (in bits)
2411  * in table->maxlen.
2412  *
2413  * We use a range comma separated format (e.g. 1,3-4,10-10) so that
2414  * large bitmaps may be represented in a compact manner. Writing into
2415  * the file will clear the bitmap then update it with the given input.
2416  *
2417  * Returns 0 on success.
2418  */
2419 int proc_do_large_bitmap(struct ctl_table *table, int write,
2420                          void __user *buffer, size_t *lenp, loff_t *ppos)
2421 {
2422         int err = 0;
2423         bool first = 1;
2424         size_t left = *lenp;
2425         unsigned long bitmap_len = table->maxlen;
2426         unsigned long *bitmap = (unsigned long *) table->data;
2427         unsigned long *tmp_bitmap = NULL;
2428         char tr_a[] = { '-', ',', '\n' }, tr_b[] = { ',', '\n', 0 }, c;
2429
2430         if (!bitmap_len || !left || (*ppos && !write)) {
2431                 *lenp = 0;
2432                 return 0;
2433         }
2434
2435         if (write) {
2436                 unsigned long page = 0;
2437                 char *kbuf;
2438
2439                 if (left > PAGE_SIZE - 1)
2440                         left = PAGE_SIZE - 1;
2441
2442                 page = __get_free_page(GFP_TEMPORARY);
2443                 kbuf = (char *) page;
2444                 if (!kbuf)
2445                         return -ENOMEM;
2446                 if (copy_from_user(kbuf, buffer, left)) {
2447                         free_page(page);
2448                         return -EFAULT;
2449                 }
2450                 kbuf[left] = 0;
2451
2452                 tmp_bitmap = kzalloc(BITS_TO_LONGS(bitmap_len) * sizeof(unsigned long),
2453                                      GFP_KERNEL);
2454                 if (!tmp_bitmap) {
2455                         free_page(page);
2456                         return -ENOMEM;
2457                 }
2458                 proc_skip_char(&kbuf, &left, '\n');
2459                 while (!err && left) {
2460                         unsigned long val_a, val_b;
2461                         bool neg;
2462
2463                         err = proc_get_long(&kbuf, &left, &val_a, &neg, tr_a,
2464                                              sizeof(tr_a), &c);
2465                         if (err)
2466                                 break;
2467                         if (val_a >= bitmap_len || neg) {
2468                                 err = -EINVAL;
2469                                 break;
2470                         }
2471
2472                         val_b = val_a;
2473                         if (left) {
2474                                 kbuf++;
2475                                 left--;
2476                         }
2477
2478                         if (c == '-') {
2479                                 err = proc_get_long(&kbuf, &left, &val_b,
2480                                                      &neg, tr_b, sizeof(tr_b),
2481                                                      &c);
2482                                 if (err)
2483                                         break;
2484                                 if (val_b >= bitmap_len || neg ||
2485                                     val_a > val_b) {
2486                                         err = -EINVAL;
2487                                         break;
2488                                 }
2489                                 if (left) {
2490                                         kbuf++;
2491                                         left--;
2492                                 }
2493                         }
2494
2495                         bitmap_set(tmp_bitmap, val_a, val_b - val_a + 1);
2496                         first = 0;
2497                         proc_skip_char(&kbuf, &left, '\n');
2498                 }
2499                 free_page(page);
2500         } else {
2501                 unsigned long bit_a, bit_b = 0;
2502
2503                 while (left) {
2504                         bit_a = find_next_bit(bitmap, bitmap_len, bit_b);
2505                         if (bit_a >= bitmap_len)
2506                                 break;
2507                         bit_b = find_next_zero_bit(bitmap, bitmap_len,
2508                                                    bit_a + 1) - 1;
2509
2510                         if (!first) {
2511                                 err = proc_put_char(&buffer, &left, ',');
2512                                 if (err)
2513                                         break;
2514                         }
2515                         err = proc_put_long(&buffer, &left, bit_a, false);
2516                         if (err)
2517                                 break;
2518                         if (bit_a != bit_b) {
2519                                 err = proc_put_char(&buffer, &left, '-');
2520                                 if (err)
2521                                         break;
2522                                 err = proc_put_long(&buffer, &left, bit_b, false);
2523                                 if (err)
2524                                         break;
2525                         }
2526
2527                         first = 0; bit_b++;
2528                 }
2529                 if (!err)
2530                         err = proc_put_char(&buffer, &left, '\n');
2531         }
2532
2533         if (!err) {
2534                 if (write) {
2535                         if (*ppos)
2536                                 bitmap_or(bitmap, bitmap, tmp_bitmap, bitmap_len);
2537                         else
2538                                 bitmap_copy(bitmap, tmp_bitmap, bitmap_len);
2539                 }
2540                 kfree(tmp_bitmap);
2541                 *lenp -= left;
2542                 *ppos += *lenp;
2543                 return 0;
2544         } else {
2545                 kfree(tmp_bitmap);
2546                 return err;
2547         }
2548 }
2549
2550 #else /* CONFIG_PROC_SYSCTL */
2551
2552 int proc_dostring(struct ctl_table *table, int write,
2553                   void __user *buffer, size_t *lenp, loff_t *ppos)
2554 {
2555         return -ENOSYS;
2556 }
2557
2558 int proc_dointvec(struct ctl_table *table, int write,
2559                   void __user *buffer, size_t *lenp, loff_t *ppos)
2560 {
2561         return -ENOSYS;
2562 }
2563
2564 int proc_dointvec_minmax(struct ctl_table *table, int write,
2565                     void __user *buffer, size_t *lenp, loff_t *ppos)
2566 {
2567         return -ENOSYS;
2568 }
2569
2570 int proc_dointvec_jiffies(struct ctl_table *table, int write,
2571                     void __user *buffer, size_t *lenp, loff_t *ppos)
2572 {
2573         return -ENOSYS;
2574 }
2575
2576 int proc_dointvec_userhz_jiffies(struct ctl_table *table, int write,
2577                     void __user *buffer, size_t *lenp, loff_t *ppos)
2578 {
2579         return -ENOSYS;
2580 }
2581
2582 int proc_dointvec_ms_jiffies(struct ctl_table *table, int write,
2583                              void __user *buffer, size_t *lenp, loff_t *ppos)
2584 {
2585         return -ENOSYS;
2586 }
2587
2588 int proc_doulongvec_minmax(struct ctl_table *table, int write,
2589                     void __user *buffer, size_t *lenp, loff_t *ppos)
2590 {
2591         return -ENOSYS;
2592 }
2593
2594 int proc_doulongvec_ms_jiffies_minmax(struct ctl_table *table, int write,
2595                                       void __user *buffer,
2596                                       size_t *lenp, loff_t *ppos)
2597 {
2598     return -ENOSYS;
2599 }
2600
2601
2602 #endif /* CONFIG_PROC_SYSCTL */
2603
2604 /*
2605  * No sense putting this after each symbol definition, twice,
2606  * exception granted :-)
2607  */
2608 EXPORT_SYMBOL(proc_dointvec);
2609 EXPORT_SYMBOL(proc_dointvec_jiffies);
2610 EXPORT_SYMBOL(proc_dointvec_minmax);
2611 EXPORT_SYMBOL(proc_dointvec_userhz_jiffies);
2612 EXPORT_SYMBOL(proc_dointvec_ms_jiffies);
2613 EXPORT_SYMBOL(proc_dostring);
2614 EXPORT_SYMBOL(proc_doulongvec_minmax);
2615 EXPORT_SYMBOL(proc_doulongvec_ms_jiffies_minmax);