kernel/sysctl.c

   1 /*
   2  * sysctl.c: General linux system control interface
   3  *
   4  * Begun 24 March 1995, Stephen Tweedie
   5  * Added /proc support, Dec 1995
   6  * Added bdflush entry and intvec min/max checking, 2/23/96, Tom Dyas.
   7  * Added hooks for /proc/sys/net (minor, minor patch), 96/4/1, Mike Shaver.
   8  * Added kernel/java-{interpreter,appletviewer}, 96/5/10, Mike Shaver.
   9  * Dynamic registration fixes, Stephen Tweedie.
  10  * Added kswapd-interval, ctrl-alt-del, printk stuff, 1/8/97, Chris Horn.
  11  * Made sysctl support optional via CONFIG_SYSCTL, 1/10/97, Chris
  12  *  Horn.
  13  * Added proc_doulongvec_ms_jiffies_minmax, 09/08/99, Carlos H. Bauer.
  14  * Added proc_doulongvec_minmax, 09/08/99, Carlos H. Bauer.
  15  * Changed linked lists to use list.h instead of lists.h, 02/24/00, Bill
  16  *  Wendling.
  17  * The list_for_each() macro wasn't appropriate for the sysctl loop.
  18  *  Removed it and replaced it with older style, 03/23/00, Bill Wendling
  19  */
  20
  21 #include <linux/module.h>
  22 #include <linux/mm.h>
  23 #include <linux/swap.h>
  24 #include <linux/slab.h>
  25 #include <linux/sysctl.h>
  26 #include <linux/bitmap.h>
  27 #include <linux/signal.h>
  28 #include <linux/printk.h>
  29 #include <linux/proc_fs.h>
  30 #include <linux/security.h>
  31 #include <linux/ctype.h>
  32 #include <linux/kmemcheck.h>
  33 #include <linux/kmemleak.h>
  34 #include <linux/fs.h>
  35 #include <linux/init.h>
  36 #include <linux/kernel.h>
  37 #include <linux/kobject.h>
  38 #include <linux/net.h>
  39 #include <linux/sysrq.h>
  40 #include <linux/highuid.h>
  41 #include <linux/writeback.h>
  42 #include <linux/ratelimit.h>
  43 #include <linux/compaction.h>
  44 #include <linux/hugetlb.h>
  45 #include <linux/initrd.h>
  46 #include <linux/key.h>
  47 #include <linux/times.h>
  48 #include <linux/limits.h>
  49 #include <linux/dcache.h>
  50 #include <linux/dnotify.h>
  51 #include <linux/syscalls.h>
  52 #include <linux/vmstat.h>
  53 #include <linux/nfs_fs.h>
  54 #include <linux/acpi.h>
  55 #include <linux/reboot.h>
  56 #include <linux/ftrace.h>
  57 #include <linux/perf_event.h>
  58 #include <linux/kprobes.h>
  59 #include <linux/pipe_fs_i.h>
  60 #include <linux/oom.h>
  61 #include <linux/kmod.h>
  62 #include <linux/capability.h>
  63 #include <linux/binfmts.h>
  64 #include <linux/sched/sysctl.h>
  65
  66 #include <asm/uaccess.h>
  67 #include <asm/processor.h>
  68
  69 #ifdef CONFIG_X86
  70 #include <asm/nmi.h>
  71 #include <asm/stacktrace.h>
  72 #include <asm/io.h>
  73 #endif
  74 #ifdef CONFIG_SPARC
  75 #include <asm/setup.h>
  76 #endif
  77 #ifdef CONFIG_BSD_PROCESS_ACCT
  78 #include <linux/acct.h>
  79 #endif
  80 #ifdef CONFIG_RT_MUTEXES
  81 #include <linux/rtmutex.h>
  82 #endif
  83 #if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_LOCK_STAT)
  84 #include <linux/lockdep.h>
  85 #endif
  86 #ifdef CONFIG_CHR_DEV_SG
  87 #include <scsi/sg.h>
  88 #endif
  89
  90 #ifdef CONFIG_LOCKUP_DETECTOR
  91 #include <linux/nmi.h>
  92 #endif
  93
  94
  95 #if defined(CONFIG_SYSCTL)
  96
  97 /* External variables not in a header file. */
  98 extern int sysctl_overcommit_memory;
  99 extern int sysctl_overcommit_ratio;
 100 extern int max_threads;
 101 extern int suid_dumpable;
 102 #ifdef CONFIG_COREDUMP
 103 extern int core_uses_pid;
 104 extern char core_pattern[];
 105 extern unsigned int core_pipe_limit;
 106 #endif
 107 extern int pid_max;
 108 extern int min_free_kbytes;
 109 extern int pid_max_min, pid_max_max;
 110 extern int sysctl_drop_caches;
 111 extern int percpu_pagelist_fraction;
 112 extern int compat_log;
 113 extern int latencytop_enabled;
 114 extern int sysctl_nr_open_min, sysctl_nr_open_max;
 115 #ifndef CONFIG_MMU
 116 extern int sysctl_nr_trim_pages;
 117 #endif
 118 #ifdef CONFIG_BLOCK
 119 extern int blk_iopoll_enabled;
 120 #endif
 121
 122 /* Constants used for minimum and  maximum */
 123 #ifdef CONFIG_LOCKUP_DETECTOR
 124 static int sixty = 60;
 125 static int neg_one = -1;
 126 #endif
 127
 128 static int zero;
 129 static int __maybe_unused one = 1;
 130 static int __maybe_unused two = 2;
 131 static int __maybe_unused three = 3;
 132 static unsigned long one_ul = 1;
 133 static int one_hundred = 100;
 134 #ifdef CONFIG_PRINTK
 135 static int ten_thousand = 10000;
 136 #endif
 137
 138 /* this is needed for the proc_doulongvec_minmax of vm_dirty_bytes */
 139 static unsigned long dirty_bytes_min = 2 * PAGE_SIZE;
 140
 141 /* this is needed for the proc_dointvec_minmax for [fs_]overflow UID and GID */
 142 static int maxolduid = 65535;
 143 static int minolduid;
 144 static int min_percpu_pagelist_fract = 8;
 145
 146 static int ngroups_max = NGROUPS_MAX;
 147 static const int cap_last_cap = CAP_LAST_CAP;
 148
 149 #ifdef CONFIG_INOTIFY_USER
 150 #include <linux/inotify.h>
 151 #endif
 152 #ifdef CONFIG_SPARC
 153 #endif
 154
 155 #ifdef CONFIG_SPARC64
 156 extern int sysctl_tsb_ratio;
 157 #endif
 158
 159 #ifdef __hppa__
 160 extern int pwrsw_enabled;
 161 extern int unaligned_enabled;
 162 #endif
 163
 164 #ifdef CONFIG_IA64
 165 extern int no_unaligned_warning;
 166 extern int unaligned_dump_stack;
 167 #endif
 168
 169 #ifdef CONFIG_PROC_SYSCTL
 170 static int proc_do_cad_pid(struct ctl_table *table, int write,
 171                   void __user *buffer, size_t *lenp, loff_t *ppos);
 172 static int proc_taint(struct ctl_table *table, int write,
 173                                void __user *buffer, size_t *lenp, loff_t *ppos);
 174 #endif
 175
 176 #ifdef CONFIG_PRINTK
 177 static int proc_dointvec_minmax_sysadmin(struct ctl_table *table, int write,
 178                                 void __user *buffer, size_t *lenp, loff_t *ppos);
 179 #endif
 180
 181 static int proc_dointvec_minmax_coredump(struct ctl_table *table, int write,
 182                 void __user *buffer, size_t *lenp, loff_t *ppos);
 183 #ifdef CONFIG_COREDUMP
 184 static int proc_dostring_coredump(struct ctl_table *table, int write,
 185                 void __user *buffer, size_t *lenp, loff_t *ppos);
 186 #endif
 187
 188 #ifdef CONFIG_MAGIC_SYSRQ
 189 /* Note: sysrq code uses it's own private copy */
 190 static int __sysrq_enabled = SYSRQ_DEFAULT_ENABLE;
 191
 192 static int sysrq_sysctl_handler(ctl_table *table, int write,
 193                                 void __user *buffer, size_t *lenp,
 194                                 loff_t *ppos)
 195 {
 196         int error;
 197
 198         error = proc_dointvec(table, write, buffer, lenp, ppos);
 199         if (error)
 200                 return error;
 201
 202         if (write)
 203                 sysrq_toggle_support(__sysrq_enabled);
 204
 205         return 0;
 206 }
 207
 208 #endif
 209
 210 static struct ctl_table kern_table[];
 211 static struct ctl_table vm_table[];
 212 static struct ctl_table fs_table[];
 213 static struct ctl_table debug_table[];
 214 static struct ctl_table dev_table[];
 215 extern struct ctl_table random_table[];
 216 #ifdef CONFIG_EPOLL
 217 extern struct ctl_table epoll_table[];
 218 #endif
 219
 220 #ifdef HAVE_ARCH_PICK_MMAP_LAYOUT
 221 int sysctl_legacy_va_layout;
 222 #endif
 223
 224 /* The default sysctl tables: */
 225
 226 static struct ctl_table sysctl_base_table[] = {
 227         {
 228                 .procname       = "kernel",
 229                 .mode           = 0555,
 230                 .child          = kern_table,
 231         },
 232         {
 233                 .procname       = "vm",
 234                 .mode           = 0555,
 235                 .child          = vm_table,
 236         },
 237         {
 238                 .procname       = "fs",
 239                 .mode           = 0555,
 240                 .child          = fs_table,
 241         },
 242         {
 243                 .procname       = "debug",
 244                 .mode           = 0555,
 245                 .child          = debug_table,
 246         },
 247         {
 248                 .procname       = "dev",
 249                 .mode           = 0555,
 250                 .child          = dev_table,
 251         },
 252         { }
 253 };
 254
 255 #ifdef CONFIG_SCHED_DEBUG
 256 static int min_sched_granularity_ns = 100000;           /* 100 usecs */
 257 static int max_sched_granularity_ns = NSEC_PER_SEC;     /* 1 second */
 258 static int min_wakeup_granularity_ns;                   /* 0 usecs */
 259 static int max_wakeup_granularity_ns = NSEC_PER_SEC;    /* 1 second */
 260 #ifdef CONFIG_SMP
 261 static int min_sched_tunable_scaling = SCHED_TUNABLESCALING_NONE;
 262 static int max_sched_tunable_scaling = SCHED_TUNABLESCALING_END-1;
 263 #endif /* CONFIG_SMP */
 264 #endif /* CONFIG_SCHED_DEBUG */
 265
 266 #ifdef CONFIG_COMPACTION
 267 static int min_extfrag_threshold;
 268 static int max_extfrag_threshold = 1000;
 269 #endif
 270
 271 static struct ctl_table kern_table[] = {
 272         {
 273                 .procname       = "sched_child_runs_first",
 274                 .data           = &sysctl_sched_child_runs_first,
 275                 .maxlen         = sizeof(unsigned int),
 276                 .mode           = 0644,
 277                 .proc_handler   = proc_dointvec,
 278         },
 279 #ifdef CONFIG_SCHED_DEBUG
 280         {
 281                 .procname       = "sched_min_granularity_ns",
 282                 .data           = &sysctl_sched_min_granularity,
 283                 .maxlen         = sizeof(unsigned int),
 284                 .mode           = 0644,
 285                 .proc_handler   = sched_proc_update_handler,
 286                 .extra1         = &min_sched_granularity_ns,
 287                 .extra2         = &max_sched_granularity_ns,
 288         },
 289         {
 290                 .procname       = "sched_latency_ns",
 291                 .data           = &sysctl_sched_latency,
 292                 .maxlen         = sizeof(unsigned int),
 293                 .mode           = 0644,
 294                 .proc_handler   = sched_proc_update_handler,
 295                 .extra1         = &min_sched_granularity_ns,
 296                 .extra2         = &max_sched_granularity_ns,
 297         },
 298         {
 299                 .procname       = "sched_wakeup_granularity_ns",
 300                 .data           = &sysctl_sched_wakeup_granularity,
 301                 .maxlen         = sizeof(unsigned int),
 302                 .mode           = 0644,
 303                 .proc_handler   = sched_proc_update_handler,
 304                 .extra1         = &min_wakeup_granularity_ns,
 305                 .extra2         = &max_wakeup_granularity_ns,
 306         },
 307 #ifdef CONFIG_SMP
 308         {
 309                 .procname       = "sched_tunable_scaling",
 310                 .data           = &sysctl_sched_tunable_scaling,
 311                 .maxlen         = sizeof(enum sched_tunable_scaling),
 312                 .mode           = 0644,
 313                 .proc_handler   = sched_proc_update_handler,
 314                 .extra1         = &min_sched_tunable_scaling,
 315                 .extra2         = &max_sched_tunable_scaling,
 316         },
 317         {
 318                 .procname       = "sched_migration_cost_ns",
 319                 .data           = &sysctl_sched_migration_cost,
 320                 .maxlen         = sizeof(unsigned int),
 321                 .mode           = 0644,
 322                 .proc_handler   = proc_dointvec,
 323         },
 324         {
 325                 .procname       = "sched_nr_migrate",
 326                 .data           = &sysctl_sched_nr_migrate,
 327                 .maxlen         = sizeof(unsigned int),
 328                 .mode           = 0644,
 329                 .proc_handler   = proc_dointvec,
 330         },
 331         {
 332                 .procname       = "sched_time_avg_ms",
 333                 .data           = &sysctl_sched_time_avg,
 334                 .maxlen         = sizeof(unsigned int),
 335                 .mode           = 0644,
 336                 .proc_handler   = proc_dointvec,
 337         },
 338         {
 339                 .procname       = "sched_shares_window_ns",
 340                 .data           = &sysctl_sched_shares_window,
 341                 .maxlen         = sizeof(unsigned int),
 342                 .mode           = 0644,
 343                 .proc_handler   = proc_dointvec,
 344         },
 345         {
 346                 .procname       = "timer_migration",
 347                 .data           = &sysctl_timer_migration,
 348                 .maxlen         = sizeof(unsigned int),
 349                 .mode           = 0644,
 350                 .proc_handler   = proc_dointvec_minmax,
 351                 .extra1         = &zero,
 352                 .extra2         = &one,
 353         },
 354 #endif /* CONFIG_SMP */
 355 #ifdef CONFIG_NUMA_BALANCING
 356         {
 357                 .procname       = "numa_balancing_scan_delay_ms",
 358                 .data           = &sysctl_numa_balancing_scan_delay,
 359                 .maxlen         = sizeof(unsigned int),
 360                 .mode           = 0644,
 361                 .proc_handler   = proc_dointvec,
 362         },
 363         {
 364                 .procname       = "numa_balancing_scan_period_min_ms",
 365                 .data           = &sysctl_numa_balancing_scan_period_min,
 366                 .maxlen         = sizeof(unsigned int),
 367                 .mode           = 0644,
 368                 .proc_handler   = proc_dointvec,
 369         },
 370         {
 371                 .procname       = "numa_balancing_scan_period_reset",
 372                 .data           = &sysctl_numa_balancing_scan_period_reset,
 373                 .maxlen         = sizeof(unsigned int),
 374                 .mode           = 0644,
 375                 .proc_handler   = proc_dointvec,
 376         },
 377         {
 378                 .procname       = "numa_balancing_scan_period_max_ms",
 379                 .data           = &sysctl_numa_balancing_scan_period_max,
 380                 .maxlen         = sizeof(unsigned int),
 381                 .mode           = 0644,
 382                 .proc_handler   = proc_dointvec,
 383         },
 384         {
 385                 .procname       = "numa_balancing_scan_size_mb",
 386                 .data           = &sysctl_numa_balancing_scan_size,
 387                 .maxlen         = sizeof(unsigned int),
 388                 .mode           = 0644,
 389                 .proc_handler   = proc_dointvec,
 390         },
 391 #endif /* CONFIG_NUMA_BALANCING */
 392 #endif /* CONFIG_SCHED_DEBUG */
 393         {
 394                 .procname       = "sched_rt_period_us",
 395                 .data           = &sysctl_sched_rt_period,
 396                 .maxlen         = sizeof(unsigned int),
 397                 .mode           = 0644,
 398                 .proc_handler   = sched_rt_handler,
 399         },
 400         {
 401                 .procname       = "sched_rt_runtime_us",
 402                 .data           = &sysctl_sched_rt_runtime,
 403                 .maxlen         = sizeof(int),
 404                 .mode           = 0644,
 405                 .proc_handler   = sched_rt_handler,
 406         },
 407 #ifdef CONFIG_SCHED_AUTOGROUP
 408         {
 409                 .procname       = "sched_autogroup_enabled",
 410                 .data           = &sysctl_sched_autogroup_enabled,
 411                 .maxlen         = sizeof(unsigned int),
 412                 .mode           = 0644,
 413                 .proc_handler   = proc_dointvec_minmax,
 414                 .extra1         = &zero,
 415                 .extra2         = &one,
 416         },
 417 #endif
 418 #ifdef CONFIG_CFS_BANDWIDTH
 419         {
 420                 .procname       = "sched_cfs_bandwidth_slice_us",
 421                 .data           = &sysctl_sched_cfs_bandwidth_slice,
 422                 .maxlen         = sizeof(unsigned int),
 423                 .mode           = 0644,
 424                 .proc_handler   = proc_dointvec_minmax,
 425                 .extra1         = &one,
 426         },
 427 #endif
 428 #ifdef CONFIG_PROVE_LOCKING
 429         {
 430                 .procname       = "prove_locking",
 431                 .data           = &prove_locking,
 432                 .maxlen         = sizeof(int),
 433                 .mode           = 0644,
 434                 .proc_handler   = proc_dointvec,
 435         },
 436 #endif
 437 #ifdef CONFIG_LOCK_STAT
 438         {
 439                 .procname       = "lock_stat",
 440                 .data           = &lock_stat,
 441                 .maxlen         = sizeof(int),
 442                 .mode           = 0644,
 443                 .proc_handler   = proc_dointvec,
 444         },
 445 #endif
 446         {
 447                 .procname       = "panic",
 448                 .data           = &panic_timeout,
 449                 .maxlen         = sizeof(int),
 450                 .mode           = 0644,
 451                 .proc_handler   = proc_dointvec,
 452         },
 453 #ifdef CONFIG_COREDUMP
 454         {
 455                 .procname       = "core_uses_pid",
 456                 .data           = &core_uses_pid,
 457                 .maxlen         = sizeof(int),
 458                 .mode           = 0644,
 459                 .proc_handler   = proc_dointvec,
 460         },
 461         {
 462                 .procname       = "core_pattern",
 463                 .data           = core_pattern,
 464                 .maxlen         = CORENAME_MAX_SIZE,
 465                 .mode           = 0644,
 466                 .proc_handler   = proc_dostring_coredump,
 467         },
 468         {
 469                 .procname       = "core_pipe_limit",
 470                 .data           = &core_pipe_limit,
 471                 .maxlen         = sizeof(unsigned int),
 472                 .mode           = 0644,
 473                 .proc_handler   = proc_dointvec,
 474         },
 475 #endif
 476 #ifdef CONFIG_PROC_SYSCTL
 477         {
 478                 .procname       = "tainted",
 479                 .maxlen         = sizeof(long),
 480                 .mode           = 0644,
 481                 .proc_handler   = proc_taint,
 482         },
 483 #endif
 484 #ifdef CONFIG_LATENCYTOP
 485         {
 486                 .procname       = "latencytop",
 487                 .data           = &latencytop_enabled,
 488                 .maxlen         = sizeof(int),
 489                 .mode           = 0644,
 490                 .proc_handler   = proc_dointvec,
 491         },
 492 #endif
 493 #ifdef CONFIG_BLK_DEV_INITRD
 494         {
 495                 .procname       = "real-root-dev",
 496                 .data           = &real_root_dev,
 497                 .maxlen         = sizeof(int),
 498                 .mode           = 0644,
 499                 .proc_handler   = proc_dointvec,
 500         },
 501 #endif
 502         {
 503                 .procname       = "print-fatal-signals",
 504                 .data           = &print_fatal_signals,
 505                 .maxlen         = sizeof(int),
 506                 .mode           = 0644,
 507                 .proc_handler   = proc_dointvec,
 508         },
 509 #ifdef CONFIG_SPARC
 510         {
 511                 .procname       = "reboot-cmd",
 512                 .data           = reboot_command,
 513                 .maxlen         = 256,
 514                 .mode           = 0644,
 515                 .proc_handler   = proc_dostring,
 516         },
 517         {
 518                 .procname       = "stop-a",
 519                 .data           = &stop_a_enabled,
 520                 .maxlen         = sizeof (int),
 521                 .mode           = 0644,
 522                 .proc_handler   = proc_dointvec,
 523         },
 524         {
 525                 .procname       = "scons-poweroff",
 526                 .data           = &scons_pwroff,
 527                 .maxlen         = sizeof (int),
 528                 .mode           = 0644,
 529                 .proc_handler   = proc_dointvec,
 530         },
 531 #endif
 532 #ifdef CONFIG_SPARC64
 533         {
 534                 .procname       = "tsb-ratio",
 535                 .data           = &sysctl_tsb_ratio,
 536                 .maxlen         = sizeof (int),
 537                 .mode           = 0644,
 538                 .proc_handler   = proc_dointvec,
 539         },
 540 #endif
 541 #ifdef __hppa__
 542         {
 543                 .procname       = "soft-power",
 544                 .data           = &pwrsw_enabled,
 545                 .maxlen         = sizeof (int),
 546                 .mode           = 0644,
 547                 .proc_handler   = proc_dointvec,
 548         },
 549         {
 550                 .procname       = "unaligned-trap",
 551                 .data           = &unaligned_enabled,
 552                 .maxlen         = sizeof (int),
 553                 .mode           = 0644,
 554                 .proc_handler   = proc_dointvec,
 555         },
 556 #endif
 557         {
 558                 .procname       = "ctrl-alt-del",
 559                 .data           = &C_A_D,
 560                 .maxlen         = sizeof(int),
 561                 .mode           = 0644,
 562                 .proc_handler   = proc_dointvec,
 563         },
 564 #ifdef CONFIG_FUNCTION_TRACER
 565         {
 566                 .procname       = "ftrace_enabled",
 567                 .data           = &ftrace_enabled,
 568                 .maxlen         = sizeof(int),
 569                 .mode           = 0644,
 570                 .proc_handler   = ftrace_enable_sysctl,
 571         },
 572 #endif
 573 #ifdef CONFIG_STACK_TRACER
 574         {
 575                 .procname       = "stack_tracer_enabled",
 576                 .data           = &stack_tracer_enabled,
 577                 .maxlen         = sizeof(int),
 578                 .mode           = 0644,
 579                 .proc_handler   = stack_trace_sysctl,
 580         },
 581 #endif
 582 #ifdef CONFIG_TRACING
 583         {
 584                 .procname       = "ftrace_dump_on_oops",
 585                 .data           = &ftrace_dump_on_oops,
 586                 .maxlen         = sizeof(int),
 587                 .mode           = 0644,
 588                 .proc_handler   = proc_dointvec,
 589         },
 590 #endif
 591 #ifdef CONFIG_MODULES
 592         {
 593                 .procname       = "modprobe",
 594                 .data           = &modprobe_path,
 595                 .maxlen         = KMOD_PATH_LEN,
 596                 .mode           = 0644,
 597                 .proc_handler   = proc_dostring,
 598         },
 599         {
 600                 .procname       = "modules_disabled",
 601                 .data           = &modules_disabled,
 602                 .maxlen         = sizeof(int),
 603                 .mode           = 0644,
 604                 /* only handle a transition from default "0" to "1" */
 605                 .proc_handler   = proc_dointvec_minmax,
 606                 .extra1         = &one,
 607                 .extra2         = &one,
 608         },
 609 #endif
 610
 611         {
 612                 .procname       = "hotplug",
 613                 .data           = &uevent_helper,
 614                 .maxlen         = UEVENT_HELPER_PATH_LEN,
 615                 .mode           = 0644,
 616                 .proc_handler   = proc_dostring,
 617         },
 618
 619 #ifdef CONFIG_CHR_DEV_SG
 620         {
 621                 .procname       = "sg-big-buff",
 622                 .data           = &sg_big_buff,
 623                 .maxlen         = sizeof (int),
 624                 .mode           = 0444,
 625                 .proc_handler   = proc_dointvec,
 626         },
 627 #endif
 628 #ifdef CONFIG_BSD_PROCESS_ACCT
 629         {
 630                 .procname       = "acct",
 631                 .data           = &acct_parm,
 632                 .maxlen         = 3*sizeof(int),
 633                 .mode           = 0644,
 634                 .proc_handler   = proc_dointvec,
 635         },
 636 #endif
 637 #ifdef CONFIG_MAGIC_SYSRQ
 638         {
 639                 .procname       = "sysrq",
 640                 .data           = &__sysrq_enabled,
 641                 .maxlen         = sizeof (int),
 642                 .mode           = 0644,
 643                 .proc_handler   = sysrq_sysctl_handler,
 644         },
 645 #endif
 646 #ifdef CONFIG_PROC_SYSCTL
 647         {
 648                 .procname       = "cad_pid",
 649                 .data           = NULL,
 650                 .maxlen         = sizeof (int),
 651                 .mode           = 0600,
 652                 .proc_handler   = proc_do_cad_pid,
 653         },
 654 #endif
 655         {
 656                 .procname       = "threads-max",
 657                 .data           = &max_threads,
 658                 .maxlen         = sizeof(int),
 659                 .mode           = 0644,
 660                 .proc_handler   = proc_dointvec,
 661         },
 662         {
 663                 .procname       = "random",
 664                 .mode           = 0555,
 665                 .child          = random_table,
 666         },
 667         {
 668                 .procname       = "usermodehelper",
 669                 .mode           = 0555,
 670                 .child          = usermodehelper_table,
 671         },
 672         {
 673                 .procname       = "overflowuid",
 674                 .data           = &overflowuid,
 675                 .maxlen         = sizeof(int),
 676                 .mode           = 0644,
 677                 .proc_handler   = proc_dointvec_minmax,
 678                 .extra1         = &minolduid,
 679                 .extra2         = &maxolduid,
 680         },
 681         {
 682                 .procname       = "overflowgid",
 683                 .data           = &overflowgid,
 684                 .maxlen         = sizeof(int),
 685                 .mode           = 0644,
 686                 .proc_handler   = proc_dointvec_minmax,
 687                 .extra1         = &minolduid,
 688                 .extra2         = &maxolduid,
 689         },
 690 #ifdef CONFIG_S390
 691 #ifdef CONFIG_MATHEMU
 692         {
 693                 .procname       = "ieee_emulation_warnings",
 694                 .data           = &sysctl_ieee_emulation_warnings,
 695                 .maxlen         = sizeof(int),
 696                 .mode           = 0644,
 697                 .proc_handler   = proc_dointvec,
 698         },
 699 #endif
 700         {
 701                 .procname       = "userprocess_debug",
 702                 .data           = &show_unhandled_signals,
 703                 .maxlen         = sizeof(int),
 704                 .mode           = 0644,
 705                 .proc_handler   = proc_dointvec,
 706         },
 707 #endif
 708         {
 709                 .procname       = "pid_max",
 710                 .data           = &pid_max,
 711                 .maxlen         = sizeof (int),
 712                 .mode           = 0644,
 713                 .proc_handler   = proc_dointvec_minmax,
 714                 .extra1         = &pid_max_min,
 715                 .extra2         = &pid_max_max,
 716         },
 717         {
 718                 .procname       = "panic_on_oops",
 719                 .data           = &panic_on_oops,
 720                 .maxlen         = sizeof(int),
 721                 .mode           = 0644,
 722                 .proc_handler   = proc_dointvec,
 723         },
 724 #if defined CONFIG_PRINTK
 725         {
 726                 .procname       = "printk",
 727                 .data           = &console_loglevel,
 728                 .maxlen         = 4*sizeof(int),
 729                 .mode           = 0644,
 730                 .proc_handler   = proc_dointvec,
 731         },
 732         {
 733                 .procname       = "printk_ratelimit",
 734                 .data           = &printk_ratelimit_state.interval,
 735                 .maxlen         = sizeof(int),
 736                 .mode           = 0644,
 737                 .proc_handler   = proc_dointvec_jiffies,
 738         },
 739         {
 740                 .procname       = "printk_ratelimit_burst",
 741                 .data           = &printk_ratelimit_state.burst,
 742                 .maxlen         = sizeof(int),
 743                 .mode           = 0644,
 744                 .proc_handler   = proc_dointvec,
 745         },
 746         {
 747                 .procname       = "printk_delay",
 748                 .data           = &printk_delay_msec,
 749                 .maxlen         = sizeof(int),
 750                 .mode           = 0644,
 751                 .proc_handler   = proc_dointvec_minmax,
 752                 .extra1         = &zero,
 753                 .extra2         = &ten_thousand,
 754         },
 755         {
 756                 .procname       = "dmesg_restrict",
 757                 .data           = &dmesg_restrict,
 758                 .maxlen         = sizeof(int),
 759                 .mode           = 0644,
 760                 .proc_handler   = proc_dointvec_minmax_sysadmin,
 761                 .extra1         = &zero,
 762                 .extra2         = &one,
 763         },
 764         {
 765                 .procname       = "kptr_restrict",
 766                 .data           = &kptr_restrict,
 767                 .maxlen         = sizeof(int),
 768                 .mode           = 0644,
 769                 .proc_handler   = proc_dointvec_minmax_sysadmin,
 770                 .extra1         = &zero,
 771                 .extra2         = &two,
 772         },
 773 #endif
 774         {
 775                 .procname       = "ngroups_max",
 776                 .data           = &ngroups_max,
 777                 .maxlen         = sizeof (int),
 778                 .mode           = 0444,
 779                 .proc_handler   = proc_dointvec,
 780         },
 781         {
 782                 .procname       = "cap_last_cap",
 783                 .data           = (void *)&cap_last_cap,
 784                 .maxlen         = sizeof(int),
 785                 .mode           = 0444,
 786                 .proc_handler   = proc_dointvec,
 787         },
 788 #if defined(CONFIG_LOCKUP_DETECTOR)
 789         {
 790                 .procname       = "watchdog",
 791                 .data           = &watchdog_enabled,
 792                 .maxlen         = sizeof (int),
 793                 .mode           = 0644,
 794                 .proc_handler   = proc_dowatchdog,
 795                 .extra1         = &zero,
 796                 .extra2         = &one,
 797         },
 798         {
 799                 .procname       = "watchdog_thresh",
 800                 .data           = &watchdog_thresh,
 801                 .maxlen         = sizeof(int),
 802                 .mode           = 0644,
 803                 .proc_handler   = proc_dowatchdog,
 804                 .extra1         = &neg_one,
 805                 .extra2         = &sixty,
 806         },
 807         {
 808                 .procname       = "softlockup_panic",
 809                 .data           = &softlockup_panic,
 810                 .maxlen         = sizeof(int),
 811                 .mode           = 0644,
 812                 .proc_handler   = proc_dointvec_minmax,
 813                 .extra1         = &zero,
 814                 .extra2         = &one,
 815         },
 816         {
 817                 .procname       = "nmi_watchdog",
 818                 .data           = &watchdog_enabled,
 819                 .maxlen         = sizeof (int),
 820                 .mode           = 0644,
 821                 .proc_handler   = proc_dowatchdog,
 822                 .extra1         = &zero,
 823                 .extra2         = &one,
 824         },
 825 #endif
 826 #if defined(CONFIG_X86_LOCAL_APIC) && defined(CONFIG_X86)
 827         {
 828                 .procname       = "unknown_nmi_panic",
 829                 .data           = &unknown_nmi_panic,
 830                 .maxlen         = sizeof (int),
 831                 .mode           = 0644,
 832                 .proc_handler   = proc_dointvec,
 833         },
 834 #endif
 835 #if defined(CONFIG_X86)
 836         {
 837                 .procname       = "panic_on_unrecovered_nmi",
 838                 .data           = &panic_on_unrecovered_nmi,
 839                 .maxlen         = sizeof(int),
 840                 .mode           = 0644,
 841                 .proc_handler   = proc_dointvec,
 842         },
 843         {
 844                 .procname       = "panic_on_io_nmi",
 845                 .data           = &panic_on_io_nmi,
 846                 .maxlen         = sizeof(int),
 847                 .mode           = 0644,
 848                 .proc_handler   = proc_dointvec,
 849         },
 850 #ifdef CONFIG_DEBUG_STACKOVERFLOW
 851         {
 852                 .procname       = "panic_on_stackoverflow",
 853                 .data           = &sysctl_panic_on_stackoverflow,
 854                 .maxlen         = sizeof(int),
 855                 .mode           = 0644,
 856                 .proc_handler   = proc_dointvec,
 857         },
 858 #endif
 859         {
 860                 .procname       = "bootloader_type",
 861                 .data           = &bootloader_type,
 862                 .maxlen         = sizeof (int),
 863                 .mode           = 0444,
 864                 .proc_handler   = proc_dointvec,
 865         },
 866         {
 867                 .procname       = "bootloader_version",
 868                 .data           = &bootloader_version,
 869                 .maxlen         = sizeof (int),
 870                 .mode           = 0444,
 871                 .proc_handler   = proc_dointvec,
 872         },
 873         {
 874                 .procname       = "kstack_depth_to_print",
 875                 .data           = &kstack_depth_to_print,
 876                 .maxlen         = sizeof(int),
 877                 .mode           = 0644,
 878                 .proc_handler   = proc_dointvec,
 879         },
 880         {
 881                 .procname       = "io_delay_type",
 882                 .data           = &io_delay_type,
 883                 .maxlen         = sizeof(int),
 884                 .mode           = 0644,
 885                 .proc_handler   = proc_dointvec,
 886         },
 887 #endif
 888 #if defined(CONFIG_MMU)
 889         {
 890                 .procname       = "randomize_va_space",
 891                 .data           = &randomize_va_space,
 892                 .maxlen         = sizeof(int),
 893                 .mode           = 0644,
 894                 .proc_handler   = proc_dointvec,
 895         },
 896 #endif
 897 #if defined(CONFIG_S390) && defined(CONFIG_SMP)
 898         {
 899                 .procname       = "spin_retry",
 900                 .data           = &spin_retry,
 901                 .maxlen         = sizeof (int),
 902                 .mode           = 0644,
 903                 .proc_handler   = proc_dointvec,
 904         },
 905 #endif
 906 #if     defined(CONFIG_ACPI_SLEEP) && defined(CONFIG_X86)
 907         {
 908                 .procname       = "acpi_video_flags",
 909                 .data           = &acpi_realmode_flags,
 910                 .maxlen         = sizeof (unsigned long),
 911                 .mode           = 0644,
 912                 .proc_handler   = proc_doulongvec_minmax,
 913         },
 914 #endif
 915 #ifdef CONFIG_IA64
 916         {
 917                 .procname       = "ignore-unaligned-usertrap",
 918                 .data           = &no_unaligned_warning,
 919                 .maxlen         = sizeof (int),
 920                 .mode           = 0644,
 921                 .proc_handler   = proc_dointvec,
 922         },
 923         {
 924                 .procname       = "unaligned-dump-stack",
 925                 .data           = &unaligned_dump_stack,
 926                 .maxlen         = sizeof (int),
 927                 .mode           = 0644,
 928                 .proc_handler   = proc_dointvec,
 929         },
 930 #endif
 931 #ifdef CONFIG_DETECT_HUNG_TASK
 932         {
 933                 .procname       = "hung_task_panic",
 934                 .data           = &sysctl_hung_task_panic,
 935                 .maxlen         = sizeof(int),
 936                 .mode           = 0644,
 937                 .proc_handler   = proc_dointvec_minmax,
 938                 .extra1         = &zero,
 939                 .extra2         = &one,
 940         },
 941         {
 942                 .procname       = "hung_task_check_count",
 943                 .data           = &sysctl_hung_task_check_count,
 944                 .maxlen         = sizeof(unsigned long),
 945                 .mode           = 0644,
 946                 .proc_handler   = proc_doulongvec_minmax,
 947         },
 948         {
 949                 .procname       = "hung_task_timeout_secs",
 950                 .data           = &sysctl_hung_task_timeout_secs,
 951                 .maxlen         = sizeof(unsigned long),
 952                 .mode           = 0644,
 953                 .proc_handler   = proc_dohung_task_timeout_secs,
 954         },
 955         {
 956                 .procname       = "hung_task_warnings",
 957                 .data           = &sysctl_hung_task_warnings,
 958                 .maxlen         = sizeof(unsigned long),
 959                 .mode           = 0644,
 960                 .proc_handler   = proc_doulongvec_minmax,
 961         },
 962 #endif
 963 #ifdef CONFIG_COMPAT
 964         {
 965                 .procname       = "compat-log",
 966                 .data           = &compat_log,
 967                 .maxlen         = sizeof (int),
 968                 .mode           = 0644,
 969                 .proc_handler   = proc_dointvec,
 970         },
 971 #endif
 972 #ifdef CONFIG_RT_MUTEXES
 973         {
 974                 .procname       = "max_lock_depth",
 975                 .data           = &max_lock_depth,
 976                 .maxlen         = sizeof(int),
 977                 .mode           = 0644,
 978                 .proc_handler   = proc_dointvec,
 979         },
 980 #endif
 981         {
 982                 .procname       = "poweroff_cmd",
 983                 .data           = &poweroff_cmd,
 984                 .maxlen         = POWEROFF_CMD_PATH_LEN,
 985                 .mode           = 0644,
 986                 .proc_handler   = proc_dostring,
 987         },
 988 #ifdef CONFIG_KEYS
 989         {
 990                 .procname       = "keys",
 991                 .mode           = 0555,
 992                 .child          = key_sysctls,
 993         },
 994 #endif
 995 #ifdef CONFIG_RCU_TORTURE_TEST
 996         {
 997                 .procname       = "rcutorture_runnable",
 998                 .data           = &rcutorture_runnable,
 999                 .maxlen         = sizeof(int),
1000                 .mode           = 0644,
1001                 .proc_handler   = proc_dointvec,
1002         },
1003 #endif
1004 #ifdef CONFIG_PERF_EVENTS
1005         /*
1006          * User-space scripts rely on the existence of this file
1007          * as a feature check for perf_events being enabled.
1008          *
1009          * So it's an ABI, do not remove!
1010          */
1011         {
1012                 .procname       = "perf_event_paranoid",
1013                 .data           = &sysctl_perf_event_paranoid,
1014                 .maxlen         = sizeof(sysctl_perf_event_paranoid),
1015                 .mode           = 0644,
1016                 .proc_handler   = proc_dointvec,
1017         },
1018         {
1019                 .procname       = "perf_event_mlock_kb",
1020                 .data           = &sysctl_perf_event_mlock,
1021                 .maxlen         = sizeof(sysctl_perf_event_mlock),
1022                 .mode           = 0644,
1023                 .proc_handler   = proc_dointvec,
1024         },
1025         {
1026                 .procname       = "perf_event_max_sample_rate",
1027                 .data           = &sysctl_perf_event_sample_rate,
1028                 .maxlen         = sizeof(sysctl_perf_event_sample_rate),
1029                 .mode           = 0644,
1030                 .proc_handler   = perf_proc_update_handler,
1031         },
1032 #endif
1033 #ifdef CONFIG_KMEMCHECK
1034         {
1035                 .procname       = "kmemcheck",
1036                 .data           = &kmemcheck_enabled,
1037                 .maxlen         = sizeof(int),
1038                 .mode           = 0644,
1039                 .proc_handler   = proc_dointvec,
1040         },
1041 #endif
1042 #ifdef CONFIG_BLOCK
1043         {
1044                 .procname       = "blk_iopoll",
1045                 .data           = &blk_iopoll_enabled,
1046                 .maxlen         = sizeof(int),
1047                 .mode           = 0644,
1048                 .proc_handler   = proc_dointvec,
1049         },
1050 #endif
1051         { }
1052 };
1053
1054 static struct ctl_table vm_table[] = {
1055         {
1056                 .procname       = "overcommit_memory",
1057                 .data           = &sysctl_overcommit_memory,
1058                 .maxlen         = sizeof(sysctl_overcommit_memory),
1059                 .mode           = 0644,
1060                 .proc_handler   = proc_dointvec_minmax,
1061                 .extra1         = &zero,
1062                 .extra2         = &two,
1063         },
1064         {
1065                 .procname       = "panic_on_oom",
1066                 .data           = &sysctl_panic_on_oom,
1067                 .maxlen         = sizeof(sysctl_panic_on_oom),
1068                 .mode           = 0644,
1069                 .proc_handler   = proc_dointvec_minmax,
1070                 .extra1         = &zero,
1071                 .extra2         = &two,
1072         },
1073         {
1074                 .procname       = "oom_kill_allocating_task",
1075                 .data           = &sysctl_oom_kill_allocating_task,
1076                 .maxlen         = sizeof(sysctl_oom_kill_allocating_task),
1077                 .mode           = 0644,
1078                 .proc_handler   = proc_dointvec,
1079         },
1080         {
1081                 .procname       = "oom_dump_tasks",
1082                 .data           = &sysctl_oom_dump_tasks,
1083                 .maxlen         = sizeof(sysctl_oom_dump_tasks),
1084                 .mode           = 0644,
1085                 .proc_handler   = proc_dointvec,
1086         },
1087         {
1088                 .procname       = "overcommit_ratio",
1089                 .data           = &sysctl_overcommit_ratio,
1090                 .maxlen         = sizeof(sysctl_overcommit_ratio),
1091                 .mode           = 0644,
1092                 .proc_handler   = proc_dointvec,
1093         },
1094         {
1095                 .procname       = "page-cluster",
1096                 .data           = &page_cluster,
1097                 .maxlen         = sizeof(int),
1098                 .mode           = 0644,
1099                 .proc_handler   = proc_dointvec_minmax,
1100                 .extra1         = &zero,
1101         },
1102         {
1103                 .procname       = "dirty_background_ratio",
1104                 .data           = &dirty_background_ratio,
1105                 .maxlen         = sizeof(dirty_background_ratio),
1106                 .mode           = 0644,
1107                 .proc_handler   = dirty_background_ratio_handler,
1108                 .extra1         = &zero,
1109                 .extra2         = &one_hundred,
1110         },
1111         {
1112                 .procname       = "dirty_background_bytes",
1113                 .data           = &dirty_background_bytes,
1114                 .maxlen         = sizeof(dirty_background_bytes),
1115                 .mode           = 0644,
1116                 .proc_handler   = dirty_background_bytes_handler,
1117                 .extra1         = &one_ul,
1118         },
1119         {
1120                 .procname       = "dirty_ratio",
1121                 .data           = &vm_dirty_ratio,
1122                 .maxlen         = sizeof(vm_dirty_ratio),
1123                 .mode           = 0644,
1124                 .proc_handler   = dirty_ratio_handler,
1125                 .extra1         = &zero,
1126                 .extra2         = &one_hundred,
1127         },
1128         {
1129                 .procname       = "dirty_bytes",
1130                 .data           = &vm_dirty_bytes,
1131                 .maxlen         = sizeof(vm_dirty_bytes),
1132                 .mode           = 0644,
1133                 .proc_handler   = dirty_bytes_handler,
1134                 .extra1         = &dirty_bytes_min,
1135         },
1136         {
1137                 .procname       = "dirty_writeback_centisecs",
1138                 .data           = &dirty_writeback_interval,
1139                 .maxlen         = sizeof(dirty_writeback_interval),
1140                 .mode           = 0644,
1141                 .proc_handler   = dirty_writeback_centisecs_handler,
1142         },
1143         {
1144                 .procname       = "dirty_expire_centisecs",
1145                 .data           = &dirty_expire_interval,
1146                 .maxlen         = sizeof(dirty_expire_interval),
1147                 .mode           = 0644,
1148                 .proc_handler   = proc_dointvec_minmax,
1149                 .extra1         = &zero,
1150         },
1151         {
1152                 .procname       = "nr_pdflush_threads",
1153                 .mode           = 0444 /* read-only */,
1154                 .proc_handler   = pdflush_proc_obsolete,
1155         },
1156         {
1157                 .procname       = "swappiness",
1158                 .data           = &vm_swappiness,
1159                 .maxlen         = sizeof(vm_swappiness),
1160                 .mode           = 0644,
1161                 .proc_handler   = proc_dointvec_minmax,
1162                 .extra1         = &zero,
1163                 .extra2         = &one_hundred,
1164         },
1165 #ifdef CONFIG_HUGETLB_PAGE
1166         {
1167                 .procname       = "nr_hugepages",
1168                 .data           = NULL,
1169                 .maxlen         = sizeof(unsigned long),
1170                 .mode           = 0644,
1171                 .proc_handler   = hugetlb_sysctl_handler,
1172                 .extra1         = (void *)&hugetlb_zero,
1173                 .extra2         = (void *)&hugetlb_infinity,
1174         },
1175 #ifdef CONFIG_NUMA
1176         {
1177                 .procname       = "nr_hugepages_mempolicy",
1178                 .data           = NULL,
1179                 .maxlen         = sizeof(unsigned long),
1180                 .mode           = 0644,
1181                 .proc_handler   = &hugetlb_mempolicy_sysctl_handler,
1182                 .extra1         = (void *)&hugetlb_zero,
1183                 .extra2         = (void *)&hugetlb_infinity,
1184         },
1185 #endif
1186          {
1187                 .procname       = "hugetlb_shm_group",
1188                 .data           = &sysctl_hugetlb_shm_group,
1189                 .maxlen         = sizeof(gid_t),
1190                 .mode           = 0644,
1191                 .proc_handler   = proc_dointvec,
1192          },
1193          {
1194                 .procname       = "hugepages_treat_as_movable",
1195                 .data           = &hugepages_treat_as_movable,
1196                 .maxlen         = sizeof(int),
1197                 .mode           = 0644,
1198                 .proc_handler   = hugetlb_treat_movable_handler,
1199         },
1200         {
1201                 .procname       = "nr_overcommit_hugepages",
1202                 .data           = NULL,
1203                 .maxlen         = sizeof(unsigned long),
1204                 .mode           = 0644,
1205                 .proc_handler   = hugetlb_overcommit_handler,
1206                 .extra1         = (void *)&hugetlb_zero,
1207                 .extra2         = (void *)&hugetlb_infinity,
1208         },
1209 #endif
1210         {
1211                 .procname       = "lowmem_reserve_ratio",
1212                 .data           = &sysctl_lowmem_reserve_ratio,
1213                 .maxlen         = sizeof(sysctl_lowmem_reserve_ratio),
1214                 .mode           = 0644,
1215                 .proc_handler   = lowmem_reserve_ratio_sysctl_handler,
1216         },
1217         {
1218                 .procname       = "drop_caches",
1219                 .data           = &sysctl_drop_caches,
1220                 .maxlen         = sizeof(int),
1221                 .mode           = 0644,
1222                 .proc_handler   = drop_caches_sysctl_handler,
1223                 .extra1         = &one,
1224                 .extra2         = &three,
1225         },
1226 #ifdef CONFIG_COMPACTION
1227         {
1228                 .procname       = "compact_memory",
1229                 .data           = &sysctl_compact_memory,
1230                 .maxlen         = sizeof(int),
1231                 .mode           = 0200,
1232                 .proc_handler   = sysctl_compaction_handler,
1233         },
1234         {
1235                 .procname       = "extfrag_threshold",
1236                 .data           = &sysctl_extfrag_threshold,
1237                 .maxlen         = sizeof(int),
1238                 .mode           = 0644,
1239                 .proc_handler   = sysctl_extfrag_handler,
1240                 .extra1         = &min_extfrag_threshold,
1241                 .extra2         = &max_extfrag_threshold,
1242         },
1243
1244 #endif /* CONFIG_COMPACTION */
1245         {
1246                 .procname       = "min_free_kbytes",
1247                 .data           = &min_free_kbytes,
1248                 .maxlen         = sizeof(min_free_kbytes),
1249                 .mode           = 0644,
1250                 .proc_handler   = min_free_kbytes_sysctl_handler,
1251                 .extra1         = &zero,
1252         },
1253         {
1254                 .procname       = "percpu_pagelist_fraction",
1255                 .data           = &percpu_pagelist_fraction,
1256                 .maxlen         = sizeof(percpu_pagelist_fraction),
1257                 .mode           = 0644,
1258                 .proc_handler   = percpu_pagelist_fraction_sysctl_handler,
1259                 .extra1         = &min_percpu_pagelist_fract,
1260         },
1261 #ifdef CONFIG_MMU
1262         {
1263                 .procname       = "max_map_count",
1264                 .data           = &sysctl_max_map_count,
1265                 .maxlen         = sizeof(sysctl_max_map_count),
1266                 .mode           = 0644,
1267                 .proc_handler   = proc_dointvec_minmax,
1268                 .extra1         = &zero,
1269         },
1270 #else
1271         {
1272                 .procname       = "nr_trim_pages",
1273                 .data           = &sysctl_nr_trim_pages,
1274                 .maxlen         = sizeof(sysctl_nr_trim_pages),
1275                 .mode           = 0644,
1276                 .proc_handler   = proc_dointvec_minmax,
1277                 .extra1         = &zero,
1278         },
1279 #endif
1280         {
1281                 .procname       = "laptop_mode",
1282                 .data           = &laptop_mode,
1283                 .maxlen         = sizeof(laptop_mode),
1284                 .mode           = 0644,
1285                 .proc_handler   = proc_dointvec_jiffies,
1286         },
1287         {
1288                 .procname       = "block_dump",
1289                 .data           = &block_dump,
1290                 .maxlen         = sizeof(block_dump),
1291                 .mode           = 0644,
1292                 .proc_handler   = proc_dointvec,
1293                 .extra1         = &zero,
1294         },
1295         {
1296                 .procname       = "vfs_cache_pressure",
1297                 .data           = &sysctl_vfs_cache_pressure,
1298                 .maxlen         = sizeof(sysctl_vfs_cache_pressure),
1299                 .mode           = 0644,
1300                 .proc_handler   = proc_dointvec,
1301                 .extra1         = &zero,
1302         },
1303 #ifdef HAVE_ARCH_PICK_MMAP_LAYOUT
1304         {
1305                 .procname       = "legacy_va_layout",
1306                 .data           = &sysctl_legacy_va_layout,
1307                 .maxlen         = sizeof(sysctl_legacy_va_layout),
1308                 .mode           = 0644,
1309                 .proc_handler   = proc_dointvec,
1310                 .extra1         = &zero,
1311         },
1312 #endif
1313 #ifdef CONFIG_NUMA
1314         {
1315                 .procname       = "zone_reclaim_mode",
1316                 .data           = &zone_reclaim_mode,
1317                 .maxlen         = sizeof(zone_reclaim_mode),
1318                 .mode           = 0644,
1319                 .proc_handler   = proc_dointvec,
1320                 .extra1         = &zero,
1321         },
1322         {
1323                 .procname       = "min_unmapped_ratio",
1324                 .data           = &sysctl_min_unmapped_ratio,
1325                 .maxlen         = sizeof(sysctl_min_unmapped_ratio),
1326                 .mode           = 0644,
1327                 .proc_handler   = sysctl_min_unmapped_ratio_sysctl_handler,
1328                 .extra1         = &zero,
1329                 .extra2         = &one_hundred,
1330         },
1331         {
1332                 .procname       = "min_slab_ratio",
1333                 .data           = &sysctl_min_slab_ratio,
1334                 .maxlen         = sizeof(sysctl_min_slab_ratio),
1335                 .mode           = 0644,
1336                 .proc_handler   = sysctl_min_slab_ratio_sysctl_handler,
1337                 .extra1         = &zero,
1338                 .extra2         = &one_hundred,
1339         },
1340 #endif
1341 #ifdef CONFIG_SMP
1342         {
1343                 .procname       = "stat_interval",
1344                 .data           = &sysctl_stat_interval,
1345                 .maxlen         = sizeof(sysctl_stat_interval),
1346                 .mode           = 0644,
1347                 .proc_handler   = proc_dointvec_jiffies,
1348         },
1349 #endif
1350 #ifdef CONFIG_MMU
1351         {
1352                 .procname       = "mmap_min_addr",
1353                 .data           = &dac_mmap_min_addr,
1354                 .maxlen         = sizeof(unsigned long),
1355                 .mode           = 0644,
1356                 .proc_handler   = mmap_min_addr_handler,
1357         },
1358 #endif
1359 #ifdef CONFIG_NUMA
1360         {
1361                 .procname       = "numa_zonelist_order",
1362                 .data           = &numa_zonelist_order,
1363                 .maxlen         = NUMA_ZONELIST_ORDER_LEN,
1364                 .mode           = 0644,
1365                 .proc_handler   = numa_zonelist_order_handler,
1366         },
1367 #endif
1368 #if (defined(CONFIG_X86_32) && !defined(CONFIG_UML))|| \
1369    (defined(CONFIG_SUPERH) && defined(CONFIG_VSYSCALL))
1370         {
1371                 .procname       = "vdso_enabled",
1372                 .data           = &vdso_enabled,
1373                 .maxlen         = sizeof(vdso_enabled),
1374                 .mode           = 0644,
1375                 .proc_handler   = proc_dointvec,
1376                 .extra1         = &zero,
1377         },
1378 #endif
1379 #ifdef CONFIG_HIGHMEM
1380         {
1381                 .procname       = "highmem_is_dirtyable",
1382                 .data           = &vm_highmem_is_dirtyable,
1383                 .maxlen         = sizeof(vm_highmem_is_dirtyable),
1384                 .mode           = 0644,
1385                 .proc_handler   = proc_dointvec_minmax,
1386                 .extra1         = &zero,
1387                 .extra2         = &one,
1388         },
1389 #endif
1390         {
1391                 .procname       = "scan_unevictable_pages",
1392                 .data           = &scan_unevictable_pages,
1393                 .maxlen         = sizeof(scan_unevictable_pages),
1394                 .mode           = 0644,
1395                 .proc_handler   = scan_unevictable_handler,
1396         },
1397 #ifdef CONFIG_MEMORY_FAILURE
1398         {
1399                 .procname       = "memory_failure_early_kill",
1400                 .data           = &sysctl_memory_failure_early_kill,
1401                 .maxlen         = sizeof(sysctl_memory_failure_early_kill),
1402                 .mode           = 0644,
1403                 .proc_handler   = proc_dointvec_minmax,
1404                 .extra1         = &zero,
1405                 .extra2         = &one,
1406         },
1407         {
1408                 .procname       = "memory_failure_recovery",
1409                 .data           = &sysctl_memory_failure_recovery,
1410                 .maxlen         = sizeof(sysctl_memory_failure_recovery),
1411                 .mode           = 0644,
1412                 .proc_handler   = proc_dointvec_minmax,
1413                 .extra1         = &zero,
1414                 .extra2         = &one,
1415         },
1416 #endif
1417         { }
1418 };
1419
1420 #if defined(CONFIG_BINFMT_MISC) || defined(CONFIG_BINFMT_MISC_MODULE)
1421 static struct ctl_table binfmt_misc_table[] = {
1422         { }
1423 };
1424 #endif
1425
1426 static struct ctl_table fs_table[] = {
1427         {
1428                 .procname       = "inode-nr",
1429                 .data           = &inodes_stat,
1430                 .maxlen         = 2*sizeof(int),
1431                 .mode           = 0444,
1432                 .proc_handler   = proc_nr_inodes,
1433         },
1434         {
1435                 .procname       = "inode-state",
1436                 .data           = &inodes_stat,
1437                 .maxlen         = 7*sizeof(int),
1438                 .mode           = 0444,
1439                 .proc_handler   = proc_nr_inodes,
1440         },
1441         {
1442                 .procname       = "file-nr",
1443                 .data           = &files_stat,
1444                 .maxlen         = sizeof(files_stat),
1445                 .mode           = 0444,
1446                 .proc_handler   = proc_nr_files,
1447         },
1448         {
1449                 .procname       = "file-max",
1450                 .data           = &files_stat.max_files,
1451                 .maxlen         = sizeof(files_stat.max_files),
1452                 .mode           = 0644,
1453                 .proc_handler   = proc_doulongvec_minmax,
1454         },
1455         {
1456                 .procname       = "nr_open",
1457                 .data           = &sysctl_nr_open,
1458                 .maxlen         = sizeof(int),
1459                 .mode           = 0644,
1460                 .proc_handler   = proc_dointvec_minmax,
1461                 .extra1         = &sysctl_nr_open_min,
1462                 .extra2         = &sysctl_nr_open_max,
1463         },
1464         {
1465                 .procname       = "dentry-state",
1466                 .data           = &dentry_stat,
1467                 .maxlen         = 6*sizeof(int),
1468                 .mode           = 0444,
1469                 .proc_handler   = proc_nr_dentry,
1470         },
1471         {
1472                 .procname       = "overflowuid",
1473                 .data           = &fs_overflowuid,
1474                 .maxlen         = sizeof(int),
1475                 .mode           = 0644,
1476                 .proc_handler   = proc_dointvec_minmax,
1477                 .extra1         = &minolduid,
1478                 .extra2         = &maxolduid,
1479         },
1480         {
1481                 .procname       = "overflowgid",
1482                 .data           = &fs_overflowgid,
1483                 .maxlen         = sizeof(int),
1484                 .mode           = 0644,
1485                 .proc_handler   = proc_dointvec_minmax,
1486                 .extra1         = &minolduid,
1487                 .extra2         = &maxolduid,
1488         },
1489 #ifdef CONFIG_FILE_LOCKING
1490         {
1491                 .procname       = "leases-enable",
1492                 .data           = &leases_enable,
1493                 .maxlen         = sizeof(int),
1494                 .mode           = 0644,
1495                 .proc_handler   = proc_dointvec,
1496         },
1497 #endif
1498 #ifdef CONFIG_DNOTIFY
1499         {
1500                 .procname       = "dir-notify-enable",
1501                 .data           = &dir_notify_enable,
1502                 .maxlen         = sizeof(int),
1503                 .mode           = 0644,
1504                 .proc_handler   = proc_dointvec,
1505         },
1506 #endif
1507 #ifdef CONFIG_MMU
1508 #ifdef CONFIG_FILE_LOCKING
1509         {
1510                 .procname       = "lease-break-time",
1511                 .data           = &lease_break_time,
1512                 .maxlen         = sizeof(int),
1513                 .mode           = 0644,
1514                 .proc_handler   = proc_dointvec,
1515         },
1516 #endif
1517 #ifdef CONFIG_AIO
1518         {
1519                 .procname       = "aio-nr",
1520                 .data           = &aio_nr,
1521                 .maxlen         = sizeof(aio_nr),
1522                 .mode           = 0444,
1523                 .proc_handler   = proc_doulongvec_minmax,
1524         },
1525         {
1526                 .procname       = "aio-max-nr",
1527                 .data           = &aio_max_nr,
1528                 .maxlen         = sizeof(aio_max_nr),
1529                 .mode           = 0644,
1530                 .proc_handler   = proc_doulongvec_minmax,
1531         },
1532 #endif /* CONFIG_AIO */
1533 #ifdef CONFIG_INOTIFY_USER
1534         {
1535                 .procname       = "inotify",
1536                 .mode           = 0555,
1537                 .child          = inotify_table,
1538         },
1539 #endif
1540 #ifdef CONFIG_EPOLL
1541         {
1542                 .procname       = "epoll",
1543                 .mode           = 0555,
1544                 .child          = epoll_table,
1545         },
1546 #endif
1547 #endif
1548         {
1549                 .procname       = "protected_symlinks",
1550                 .data           = &sysctl_protected_symlinks,
1551                 .maxlen         = sizeof(int),
1552                 .mode           = 0600,
1553                 .proc_handler   = proc_dointvec_minmax,
1554                 .extra1         = &zero,
1555                 .extra2         = &one,
1556         },
1557         {
1558                 .procname       = "protected_hardlinks",
1559                 .data           = &sysctl_protected_hardlinks,
1560                 .maxlen         = sizeof(int),
1561                 .mode           = 0600,
1562                 .proc_handler   = proc_dointvec_minmax,
1563                 .extra1         = &zero,
1564                 .extra2         = &one,
1565         },
1566         {
1567                 .procname       = "suid_dumpable",
1568                 .data           = &suid_dumpable,
1569                 .maxlen         = sizeof(int),
1570                 .mode           = 0644,
1571                 .proc_handler   = proc_dointvec_minmax_coredump,
1572                 .extra1         = &zero,
1573                 .extra2         = &two,
1574         },
1575 #if defined(CONFIG_BINFMT_MISC) || defined(CONFIG_BINFMT_MISC_MODULE)
1576         {
1577                 .procname       = "binfmt_misc",
1578                 .mode           = 0555,
1579                 .child          = binfmt_misc_table,
1580         },
1581 #endif
1582         {
1583                 .procname       = "pipe-max-size",
1584                 .data           = &pipe_max_size,
1585                 .maxlen         = sizeof(int),
1586                 .mode           = 0644,
1587                 .proc_handler   = &pipe_proc_fn,
1588                 .extra1         = &pipe_min_size,
1589         },
1590         { }
1591 };
1592
1593 static struct ctl_table debug_table[] = {
1594 #ifdef CONFIG_SYSCTL_EXCEPTION_TRACE
1595         {
1596                 .procname       = "exception-trace",
1597                 .data           = &show_unhandled_signals,
1598                 .maxlen         = sizeof(int),
1599                 .mode           = 0644,
1600                 .proc_handler   = proc_dointvec
1601         },
1602 #endif
1603 #if defined(CONFIG_OPTPROBES)
1604         {
1605                 .procname       = "kprobes-optimization",
1606                 .data           = &sysctl_kprobes_optimization,
1607                 .maxlen         = sizeof(int),
1608                 .mode           = 0644,
1609                 .proc_handler   = proc_kprobes_optimization_handler,
1610                 .extra1         = &zero,
1611                 .extra2         = &one,
1612         },
1613 #endif
1614         { }
1615 };
1616
1617 static struct ctl_table dev_table[] = {
1618         { }
1619 };
1620
1621 int __init sysctl_init(void)
1622 {
1623         struct ctl_table_header *hdr;
1624
1625         hdr = register_sysctl_table(sysctl_base_table);
1626         kmemleak_not_leak(hdr);
1627         return 0;
1628 }
1629
1630 #endif /* CONFIG_SYSCTL */
1631
1632 /*
1633  * /proc/sys support
1634  */
1635
1636 #ifdef CONFIG_PROC_SYSCTL
1637
1638 static int _proc_do_string(void* data, int maxlen, int write,
1639                            void __user *buffer,
1640                            size_t *lenp, loff_t *ppos)
1641 {
1642         size_t len;
1643         char __user *p;
1644         char c;
1645
1646         if (!data || !maxlen || !*lenp) {
1647                 *lenp = 0;
1648                 return 0;
1649         }
1650
1651         if (write) {
1652                 len = 0;
1653                 p = buffer;
1654                 while (len < *lenp) {
1655                         if (get_user(c, p++))
1656                                 return -EFAULT;
1657                         if (c == 0 || c == '\n')
1658                                 break;
1659                         len++;
1660                 }
1661                 if (len >= maxlen)
1662                         len = maxlen-1;
1663                 if(copy_from_user(data, buffer, len))
1664                         return -EFAULT;
1665                 ((char *) data)[len] = 0;
1666                 *ppos += *lenp;
1667         } else {
1668                 len = strlen(data);
1669                 if (len > maxlen)
1670                         len = maxlen;
1671
1672                 if (*ppos > len) {
1673                         *lenp = 0;
1674                         return 0;
1675                 }
1676
1677                 data += *ppos;
1678                 len  -= *ppos;
1679
1680                 if (len > *lenp)
1681                         len = *lenp;
1682                 if (len)
1683                         if(copy_to_user(buffer, data, len))
1684                                 return -EFAULT;
1685                 if (len < *lenp) {
1686                         if(put_user('\n', ((char __user *) buffer) + len))
1687                                 return -EFAULT;
1688                         len++;
1689                 }
1690                 *lenp = len;
1691                 *ppos += len;
1692         }
1693         return 0;
1694 }
1695
1696 /**
1697  * proc_dostring - read a string sysctl
1698  * @table: the sysctl table
1699  * @write: %TRUE if this is a write to the sysctl file
1700  * @buffer: the user buffer
1701  * @lenp: the size of the user buffer
1702  * @ppos: file position
1703  *
1704  * Reads/writes a string from/to the user buffer. If the kernel
1705  * buffer provided is not large enough to hold the string, the
1706  * string is truncated. The copied string is %NULL-terminated.
1707  * If the string is being read by the user process, it is copied
1708  * and a newline '\n' is added. It is truncated if the buffer is
1709  * not large enough.
1710  *
1711  * Returns 0 on success.
1712  */
1713 int proc_dostring(struct ctl_table *table, int write,
1714                   void __user *buffer, size_t *lenp, loff_t *ppos)
1715 {
1716         return _proc_do_string(table->data, table->maxlen, write,
1717                                buffer, lenp, ppos);
1718 }
1719
1720 static size_t proc_skip_spaces(char **buf)
1721 {
1722         size_t ret;
1723         char *tmp = skip_spaces(*buf);
1724         ret = tmp - *buf;
1725         *buf = tmp;
1726         return ret;
1727 }
1728
1729 static void proc_skip_char(char **buf, size_t *size, const char v)
1730 {
1731         while (*size) {
1732                 if (**buf != v)
1733                         break;
1734                 (*size)--;
1735                 (*buf)++;
1736         }
1737 }
1738
1739 #define TMPBUFLEN 22
1740 /**
1741  * proc_get_long - reads an ASCII formatted integer from a user buffer
1742  *
1743  * @buf: a kernel buffer
1744  * @size: size of the kernel buffer
1745  * @val: this is where the number will be stored
1746  * @neg: set to %TRUE if number is negative
1747  * @perm_tr: a vector which contains the allowed trailers
1748  * @perm_tr_len: size of the perm_tr vector
1749  * @tr: pointer to store the trailer character
1750  *
1751  * In case of success %0 is returned and @buf and @size are updated with
1752  * the amount of bytes read. If @tr is non-NULL and a trailing
1753  * character exists (size is non-zero after returning from this
1754  * function), @tr is updated with the trailing character.
1755  */
1756 static int proc_get_long(char **buf, size_t *size,
1757                           unsigned long *val, bool *neg,
1758                           const char *perm_tr, unsigned perm_tr_len, char *tr)
1759 {
1760         int len;
1761         char *p, tmp[TMPBUFLEN];
1762
1763         if (!*size)
1764                 return -EINVAL;
1765
1766         len = *size;
1767         if (len > TMPBUFLEN - 1)
1768                 len = TMPBUFLEN - 1;
1769
1770         memcpy(tmp, *buf, len);
1771
1772         tmp[len] = 0;
1773         p = tmp;
1774         if (*p == '-' && *size > 1) {
1775                 *neg = true;
1776                 p++;
1777         } else
1778                 *neg = false;
1779         if (!isdigit(*p))
1780                 return -EINVAL;
1781
1782         *val = simple_strtoul(p, &p, 0);
1783
1784         len = p - tmp;
1785
1786         /* We don't know if the next char is whitespace thus we may accept
1787          * invalid integers (e.g. 1234...a) or two integers instead of one
1788          * (e.g. 123...1). So lets not allow such large numbers. */
1789         if (len == TMPBUFLEN - 1)
1790                 return -EINVAL;
1791
1792         if (len < *size && perm_tr_len && !memchr(perm_tr, *p, perm_tr_len))
1793                 return -EINVAL;
1794
1795         if (tr && (len < *size))
1796                 *tr = *p;
1797
1798         *buf += len;
1799         *size -= len;
1800
1801         return 0;
1802 }
1803
1804 /**
1805  * proc_put_long - converts an integer to a decimal ASCII formatted string
1806  *
1807  * @buf: the user buffer
1808  * @size: the size of the user buffer
1809  * @val: the integer to be converted
1810  * @neg: sign of the number, %TRUE for negative
1811  *
1812  * In case of success %0 is returned and @buf and @size are updated with
1813  * the amount of bytes written.
1814  */
1815 static int proc_put_long(void __user **buf, size_t *size, unsigned long val,
1816                           bool neg)
1817 {
1818         int len;
1819         char tmp[TMPBUFLEN], *p = tmp;
1820
1821         sprintf(p, "%s%lu", neg ? "-" : "", val);
1822         len = strlen(tmp);
1823         if (len > *size)
1824                 len = *size;
1825         if (copy_to_user(*buf, tmp, len))
1826                 return -EFAULT;
1827         *size -= len;
1828         *buf += len;
1829         return 0;
1830 }
1831 #undef TMPBUFLEN
1832
1833 static int proc_put_char(void __user **buf, size_t *size, char c)
1834 {
1835         if (*size) {
1836                 char __user **buffer = (char __user **)buf;
1837                 if (put_user(c, *buffer))
1838                         return -EFAULT;
1839                 (*size)--, (*buffer)++;
1840                 *buf = *buffer;
1841         }
1842         return 0;
1843 }
1844
1845 static int do_proc_dointvec_conv(bool *negp, unsigned long *lvalp,
1846                                  int *valp,
1847                                  int write, void *data)
1848 {
1849         if (write) {
1850                 *valp = *negp ? -*lvalp : *lvalp;
1851         } else {
1852                 int val = *valp;
1853                 if (val < 0) {
1854                         *negp = true;
1855                         *lvalp = (unsigned long)-val;
1856                 } else {
1857                         *negp = false;
1858                         *lvalp = (unsigned long)val;
1859                 }
1860         }
1861         return 0;
1862 }
1863
1864 static const char proc_wspace_sep[] = { ' ', '\t', '\n' };
1865
1866 static int __do_proc_dointvec(void *tbl_data, struct ctl_table *table,
1867                   int write, void __user *buffer,
1868                   size_t *lenp, loff_t *ppos,
1869                   int (*conv)(bool *negp, unsigned long *lvalp, int *valp,
1870                               int write, void *data),
1871                   void *data)
1872 {
1873         int *i, vleft, first = 1, err = 0;
1874         unsigned long page = 0;
1875         size_t left;
1876         char *kbuf;
1877
1878         if (!tbl_data || !table->maxlen || !*lenp || (*ppos && !write)) {
1879                 *lenp = 0;
1880                 return 0;
1881         }
1882
1883         i = (int *) tbl_data;
1884         vleft = table->maxlen / sizeof(*i);
1885         left = *lenp;
1886
1887         if (!conv)
1888                 conv = do_proc_dointvec_conv;
1889
1890         if (write) {
1891                 if (left > PAGE_SIZE - 1)
1892                         left = PAGE_SIZE - 1;
1893                 page = __get_free_page(GFP_TEMPORARY);
1894                 kbuf = (char *) page;
1895                 if (!kbuf)
1896                         return -ENOMEM;
1897                 if (copy_from_user(kbuf, buffer, left)) {
1898                         err = -EFAULT;
1899                         goto free;
1900                 }
1901                 kbuf[left] = 0;
1902         }
1903
1904         for (; left && vleft--; i++, first=0) {
1905                 unsigned long lval;
1906                 bool neg;
1907
1908                 if (write) {
1909                         left -= proc_skip_spaces(&kbuf);
1910
1911                         if (!left)
1912                                 break;
1913                         err = proc_get_long(&kbuf, &left, &lval, &neg,
1914                                              proc_wspace_sep,
1915                                              sizeof(proc_wspace_sep), NULL);
1916                         if (err)
1917                                 break;
1918                         if (conv(&neg, &lval, i, 1, data)) {
1919                                 err = -EINVAL;
1920                                 break;
1921                         }
1922                 } else {
1923                         if (conv(&neg, &lval, i, 0, data)) {
1924                                 err = -EINVAL;
1925                                 break;
1926                         }
1927                         if (!first)
1928                                 err = proc_put_char(&buffer, &left, '\t');
1929                         if (err)
1930                                 break;
1931                         err = proc_put_long(&buffer, &left, lval, neg);
1932                         if (err)
1933                                 break;
1934                 }
1935         }
1936
1937         if (!write && !first && left && !err)
1938                 err = proc_put_char(&buffer, &left, '\n');
1939         if (write && !err && left)
1940                 left -= proc_skip_spaces(&kbuf);
1941 free:
1942         if (write) {
1943                 free_page(page);
1944                 if (first)
1945                         return err ? : -EINVAL;
1946         }
1947         *lenp -= left;
1948         *ppos += *lenp;
1949         return err;
1950 }
1951
1952 static int do_proc_dointvec(struct ctl_table *table, int write,
1953                   void __user *buffer, size_t *lenp, loff_t *ppos,
1954                   int (*conv)(bool *negp, unsigned long *lvalp, int *valp,
1955                               int write, void *data),
1956                   void *data)
1957 {
1958         return __do_proc_dointvec(table->data, table, write,
1959                         buffer, lenp, ppos, conv, data);
1960 }
1961
1962 /**
1963  * proc_dointvec - read a vector of integers
1964  * @table: the sysctl table
1965  * @write: %TRUE if this is a write to the sysctl file
1966  * @buffer: the user buffer
1967  * @lenp: the size of the user buffer
1968  * @ppos: file position
1969  *
1970  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
1971  * values from/to the user buffer, treated as an ASCII string.
1972  *
1973  * Returns 0 on success.
1974  */
1975 int proc_dointvec(struct ctl_table *table, int write,
1976                      void __user *buffer, size_t *lenp, loff_t *ppos)
1977 {
1978     return do_proc_dointvec(table,write,buffer,lenp,ppos,
1979                             NULL,NULL);
1980 }
1981
1982 /*
1983  * Taint values can only be increased
1984  * This means we can safely use a temporary.
1985  */
1986 static int proc_taint(struct ctl_table *table, int write,
1987                                void __user *buffer, size_t *lenp, loff_t *ppos)
1988 {
1989         struct ctl_table t;
1990         unsigned long tmptaint = get_taint();
1991         int err;
1992
1993         if (write && !capable(CAP_SYS_ADMIN))
1994                 return -EPERM;
1995
1996         t = *table;
1997         t.data = &tmptaint;
1998         err = proc_doulongvec_minmax(&t, write, buffer, lenp, ppos);
1999         if (err < 0)
2000                 return err;
2001
2002         if (write) {
2003                 /*
2004                  * Poor man's atomic or. Not worth adding a primitive
2005                  * to everyone's atomic.h for this
2006                  */
2007                 int i;
2008                 for (i = 0; i < BITS_PER_LONG && tmptaint >> i; i++) {
2009                         if ((tmptaint >> i) & 1)
2010                                 add_taint(i);
2011                 }
2012         }
2013
2014         return err;
2015 }
2016
2017 #ifdef CONFIG_PRINTK
2018 static int proc_dointvec_minmax_sysadmin(struct ctl_table *table, int write,
2019                                 void __user *buffer, size_t *lenp, loff_t *ppos)
2020 {
2021         if (write && !capable(CAP_SYS_ADMIN))
2022                 return -EPERM;
2023
2024         return proc_dointvec_minmax(table, write, buffer, lenp, ppos);
2025 }
2026 #endif
2027
2028 struct do_proc_dointvec_minmax_conv_param {
2029         int *min;
2030         int *max;
2031 };
2032
2033 static int do_proc_dointvec_minmax_conv(bool *negp, unsigned long *lvalp,
2034                                         int *valp,
2035                                         int write, void *data)
2036 {
2037         struct do_proc_dointvec_minmax_conv_param *param = data;
2038         if (write) {
2039                 int val = *negp ? -*lvalp : *lvalp;
2040                 if ((param->min && *param->min > val) ||
2041                     (param->max && *param->max < val))
2042                         return -EINVAL;
2043                 *valp = val;
2044         } else {
2045                 int val = *valp;
2046                 if (val < 0) {
2047                         *negp = true;
2048                         *lvalp = (unsigned long)-val;
2049                 } else {
2050                         *negp = false;
2051                         *lvalp = (unsigned long)val;
2052                 }
2053         }
2054         return 0;
2055 }
2056
2057 /**
2058  * proc_dointvec_minmax - read a vector of integers with min/max values
2059  * @table: the sysctl table
2060  * @write: %TRUE if this is a write to the sysctl file
2061  * @buffer: the user buffer
2062  * @lenp: the size of the user buffer
2063  * @ppos: file position
2064  *
2065  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2066  * values from/to the user buffer, treated as an ASCII string.
2067  *
2068  * This routine will ensure the values are within the range specified by
2069  * table->extra1 (min) and table->extra2 (max).
2070  *
2071  * Returns 0 on success.
2072  */
2073 int proc_dointvec_minmax(struct ctl_table *table, int write,
2074                   void __user *buffer, size_t *lenp, loff_t *ppos)
2075 {
2076         struct do_proc_dointvec_minmax_conv_param param = {
2077                 .min = (int *) table->extra1,
2078                 .max = (int *) table->extra2,
2079         };
2080         return do_proc_dointvec(table, write, buffer, lenp, ppos,
2081                                 do_proc_dointvec_minmax_conv, &param);
2082 }
2083
2084 static void validate_coredump_safety(void)
2085 {
2086 #ifdef CONFIG_COREDUMP
2087         if (suid_dumpable == SUID_DUMPABLE_SAFE &&
2088             core_pattern[0] != '/' && core_pattern[0] != '|') {
2089                 printk(KERN_WARNING "Unsafe core_pattern used with "\
2090                         "suid_dumpable=2. Pipe handler or fully qualified "\
2091                         "core dump path required.\n");
2092         }
2093 #endif
2094 }
2095
2096 static int proc_dointvec_minmax_coredump(struct ctl_table *table, int write,
2097                 void __user *buffer, size_t *lenp, loff_t *ppos)
2098 {
2099         int error = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
2100         if (!error)
2101                 validate_coredump_safety();
2102         return error;
2103 }
2104
2105 #ifdef CONFIG_COREDUMP
2106 static int proc_dostring_coredump(struct ctl_table *table, int write,
2107                   void __user *buffer, size_t *lenp, loff_t *ppos)
2108 {
2109         int error = proc_dostring(table, write, buffer, lenp, ppos);
2110         if (!error)
2111                 validate_coredump_safety();
2112         return error;
2113 }
2114 #endif
2115
2116 static int __do_proc_doulongvec_minmax(void *data, struct ctl_table *table, int write,
2117                                      void __user *buffer,
2118                                      size_t *lenp, loff_t *ppos,
2119                                      unsigned long convmul,
2120                                      unsigned long convdiv)
2121 {
2122         unsigned long *i, *min, *max;
2123         int vleft, first = 1, err = 0;
2124         unsigned long page = 0;
2125         size_t left;
2126         char *kbuf;
2127
2128         if (!data || !table->maxlen || !*lenp || (*ppos && !write)) {
2129                 *lenp = 0;
2130                 return 0;
2131         }
2132
2133         i = (unsigned long *) data;
2134         min = (unsigned long *) table->extra1;
2135         max = (unsigned long *) table->extra2;
2136         vleft = table->maxlen / sizeof(unsigned long);
2137         left = *lenp;
2138
2139         if (write) {
2140                 if (left > PAGE_SIZE - 1)
2141                         left = PAGE_SIZE - 1;
2142                 page = __get_free_page(GFP_TEMPORARY);
2143                 kbuf = (char *) page;
2144                 if (!kbuf)
2145                         return -ENOMEM;
2146                 if (copy_from_user(kbuf, buffer, left)) {
2147                         err = -EFAULT;
2148                         goto free;
2149                 }
2150                 kbuf[left] = 0;
2151         }
2152
2153         for (; left && vleft--; i++, first = 0) {
2154                 unsigned long val;
2155
2156                 if (write) {
2157                         bool neg;
2158
2159                         left -= proc_skip_spaces(&kbuf);
2160
2161                         err = proc_get_long(&kbuf, &left, &val, &neg,
2162                                              proc_wspace_sep,
2163                                              sizeof(proc_wspace_sep), NULL);
2164                         if (err)
2165                                 break;
2166                         if (neg)
2167                                 continue;
2168                         if ((min && val < *min) || (max && val > *max))
2169                                 continue;
2170                         *i = val;
2171                 } else {
2172                         val = convdiv * (*i) / convmul;
2173                         if (!first)
2174                                 err = proc_put_char(&buffer, &left, '\t');
2175                         err = proc_put_long(&buffer, &left, val, false);
2176                         if (err)
2177                                 break;
2178                 }
2179         }
2180
2181         if (!write && !first && left && !err)
2182                 err = proc_put_char(&buffer, &left, '\n');
2183         if (write && !err)
2184                 left -= proc_skip_spaces(&kbuf);
2185 free:
2186         if (write) {
2187                 free_page(page);
2188                 if (first)
2189                         return err ? : -EINVAL;
2190         }
2191         *lenp -= left;
2192         *ppos += *lenp;
2193         return err;
2194 }
2195
2196 static int do_proc_doulongvec_minmax(struct ctl_table *table, int write,
2197                                      void __user *buffer,
2198                                      size_t *lenp, loff_t *ppos,
2199                                      unsigned long convmul,
2200                                      unsigned long convdiv)
2201 {
2202         return __do_proc_doulongvec_minmax(table->data, table, write,
2203                         buffer, lenp, ppos, convmul, convdiv);
2204 }
2205
2206 /**
2207  * proc_doulongvec_minmax - read a vector of long integers with min/max values
2208  * @table: the sysctl table
2209  * @write: %TRUE if this is a write to the sysctl file
2210  * @buffer: the user buffer
2211  * @lenp: the size of the user buffer
2212  * @ppos: file position
2213  *
2214  * Reads/writes up to table->maxlen/sizeof(unsigned long) unsigned long
2215  * values from/to the user buffer, treated as an ASCII string.
2216  *
2217  * This routine will ensure the values are within the range specified by
2218  * table->extra1 (min) and table->extra2 (max).
2219  *
2220  * Returns 0 on success.
2221  */
2222 int proc_doulongvec_minmax(struct ctl_table *table, int write,
2223                            void __user *buffer, size_t *lenp, loff_t *ppos)
2224 {
2225     return do_proc_doulongvec_minmax(table, write, buffer, lenp, ppos, 1l, 1l);
2226 }
2227
2228 /**
2229  * proc_doulongvec_ms_jiffies_minmax - read a vector of millisecond values with min/max values
2230  * @table: the sysctl table
2231  * @write: %TRUE if this is a write to the sysctl file
2232  * @buffer: the user buffer
2233  * @lenp: the size of the user buffer
2234  * @ppos: file position
2235  *
2236  * Reads/writes up to table->maxlen/sizeof(unsigned long) unsigned long
2237  * values from/to the user buffer, treated as an ASCII string. The values
2238  * are treated as milliseconds, and converted to jiffies when they are stored.
2239  *
2240  * This routine will ensure the values are within the range specified by
2241  * table->extra1 (min) and table->extra2 (max).
2242  *
2243  * Returns 0 on success.
2244  */
2245 int proc_doulongvec_ms_jiffies_minmax(struct ctl_table *table, int write,
2246                                       void __user *buffer,
2247                                       size_t *lenp, loff_t *ppos)
2248 {
2249     return do_proc_doulongvec_minmax(table, write, buffer,
2250                                      lenp, ppos, HZ, 1000l);
2251 }
2252
2253
2254 static int do_proc_dointvec_jiffies_conv(bool *negp, unsigned long *lvalp,
2255                                          int *valp,
2256                                          int write, void *data)
2257 {
2258         if (write) {
2259                 if (*lvalp > LONG_MAX / HZ)
2260                         return 1;
2261                 *valp = *negp ? -(*lvalp*HZ) : (*lvalp*HZ);
2262         } else {
2263                 int val = *valp;
2264                 unsigned long lval;
2265                 if (val < 0) {
2266                         *negp = true;
2267                         lval = (unsigned long)-val;
2268                 } else {
2269                         *negp = false;
2270                         lval = (unsigned long)val;
2271                 }
2272                 *lvalp = lval / HZ;
2273         }
2274         return 0;
2275 }
2276
2277 static int do_proc_dointvec_userhz_jiffies_conv(bool *negp, unsigned long *lvalp,
2278                                                 int *valp,
2279                                                 int write, void *data)
2280 {
2281         if (write) {
2282                 if (USER_HZ < HZ && *lvalp > (LONG_MAX / HZ) * USER_HZ)
2283                         return 1;
2284                 *valp = clock_t_to_jiffies(*negp ? -*lvalp : *lvalp);
2285         } else {
2286                 int val = *valp;
2287                 unsigned long lval;
2288                 if (val < 0) {
2289                         *negp = true;
2290                         lval = (unsigned long)-val;
2291                 } else {
2292                         *negp = false;
2293                         lval = (unsigned long)val;
2294                 }
2295                 *lvalp = jiffies_to_clock_t(lval);
2296         }
2297         return 0;
2298 }
2299
2300 static int do_proc_dointvec_ms_jiffies_conv(bool *negp, unsigned long *lvalp,
2301                                             int *valp,
2302                                             int write, void *data)
2303 {
2304         if (write) {
2305                 *valp = msecs_to_jiffies(*negp ? -*lvalp : *lvalp);
2306         } else {
2307                 int val = *valp;
2308                 unsigned long lval;
2309                 if (val < 0) {
2310                         *negp = true;
2311                         lval = (unsigned long)-val;
2312                 } else {
2313                         *negp = false;
2314                         lval = (unsigned long)val;
2315                 }
2316                 *lvalp = jiffies_to_msecs(lval);
2317         }
2318         return 0;
2319 }
2320
2321 /**
2322  * proc_dointvec_jiffies - read a vector of integers as seconds
2323  * @table: the sysctl table
2324  * @write: %TRUE if this is a write to the sysctl file
2325  * @buffer: the user buffer
2326  * @lenp: the size of the user buffer
2327  * @ppos: file position
2328  *
2329  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2330  * values from/to the user buffer, treated as an ASCII string.
2331  * The values read are assumed to be in seconds, and are converted into
2332  * jiffies.
2333  *
2334  * Returns 0 on success.
2335  */
2336 int proc_dointvec_jiffies(struct ctl_table *table, int write,
2337                           void __user *buffer, size_t *lenp, loff_t *ppos)
2338 {
2339     return do_proc_dointvec(table,write,buffer,lenp,ppos,
2340                             do_proc_dointvec_jiffies_conv,NULL);
2341 }
2342
2343 /**
2344  * proc_dointvec_userhz_jiffies - read a vector of integers as 1/USER_HZ seconds
2345  * @table: the sysctl table
2346  * @write: %TRUE if this is a write to the sysctl file
2347  * @buffer: the user buffer
2348  * @lenp: the size of the user buffer
2349  * @ppos: pointer to the file position
2350  *
2351  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2352  * values from/to the user buffer, treated as an ASCII string.
2353  * The values read are assumed to be in 1/USER_HZ seconds, and
2354  * are converted into jiffies.
2355  *
2356  * Returns 0 on success.
2357  */
2358 int proc_dointvec_userhz_jiffies(struct ctl_table *table, int write,
2359                                  void __user *buffer, size_t *lenp, loff_t *ppos)
2360 {
2361     return do_proc_dointvec(table,write,buffer,lenp,ppos,
2362                             do_proc_dointvec_userhz_jiffies_conv,NULL);
2363 }
2364
2365 /**
2366  * proc_dointvec_ms_jiffies - read a vector of integers as 1 milliseconds
2367  * @table: the sysctl table
2368  * @write: %TRUE if this is a write to the sysctl file
2369  * @buffer: the user buffer
2370  * @lenp: the size of the user buffer
2371  * @ppos: file position
2372  * @ppos: the current position in the file
2373  *
2374  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2375  * values from/to the user buffer, treated as an ASCII string.
2376  * The values read are assumed to be in 1/1000 seconds, and
2377  * are converted into jiffies.
2378  *
2379  * Returns 0 on success.
2380  */
2381 int proc_dointvec_ms_jiffies(struct ctl_table *table, int write,
2382                              void __user *buffer, size_t *lenp, loff_t *ppos)
2383 {
2384         return do_proc_dointvec(table, write, buffer, lenp, ppos,
2385                                 do_proc_dointvec_ms_jiffies_conv, NULL);
2386 }
2387
2388 static int proc_do_cad_pid(struct ctl_table *table, int write,
2389                            void __user *buffer, size_t *lenp, loff_t *ppos)
2390 {
2391         struct pid *new_pid;
2392         pid_t tmp;
2393         int r;
2394
2395         tmp = pid_vnr(cad_pid);
2396
2397         r = __do_proc_dointvec(&tmp, table, write, buffer,
2398                                lenp, ppos, NULL, NULL);
2399         if (r || !write)
2400                 return r;
2401
2402         new_pid = find_get_pid(tmp);
2403         if (!new_pid)
2404                 return -ESRCH;
2405
2406         put_pid(xchg(&cad_pid, new_pid));
2407         return 0;
2408 }
2409
2410 /**
2411  * proc_do_large_bitmap - read/write from/to a large bitmap
2412  * @table: the sysctl table
2413  * @write: %TRUE if this is a write to the sysctl file
2414  * @buffer: the user buffer
2415  * @lenp: the size of the user buffer
2416  * @ppos: file position
2417  *
2418  * The bitmap is stored at table->data and the bitmap length (in bits)
2419  * in table->maxlen.
2420  *
2421  * We use a range comma separated format (e.g. 1,3-4,10-10) so that
2422  * large bitmaps may be represented in a compact manner. Writing into
2423  * the file will clear the bitmap then update it with the given input.
2424  *
2425  * Returns 0 on success.
2426  */
2427 int proc_do_large_bitmap(struct ctl_table *table, int write,
2428                          void __user *buffer, size_t *lenp, loff_t *ppos)
2429 {
2430         int err = 0;
2431         bool first = 1;
2432         size_t left = *lenp;
2433         unsigned long bitmap_len = table->maxlen;
2434         unsigned long *bitmap = (unsigned long *) table->data;
2435         unsigned long *tmp_bitmap = NULL;
2436         char tr_a[] = { '-', ',', '\n' }, tr_b[] = { ',', '\n', 0 }, c;
2437
2438         if (!bitmap_len || !left || (*ppos && !write)) {
2439                 *lenp = 0;
2440                 return 0;
2441         }
2442
2443         if (write) {
2444                 unsigned long page = 0;
2445                 char *kbuf;
2446
2447                 if (left > PAGE_SIZE - 1)
2448                         left = PAGE_SIZE - 1;
2449
2450                 page = __get_free_page(GFP_TEMPORARY);
2451                 kbuf = (char *) page;
2452                 if (!kbuf)
2453                         return -ENOMEM;
2454                 if (copy_from_user(kbuf, buffer, left)) {
2455                         free_page(page);
2456                         return -EFAULT;
2457                 }
2458                 kbuf[left] = 0;
2459
2460                 tmp_bitmap = kzalloc(BITS_TO_LONGS(bitmap_len) * sizeof(unsigned long),
2461                                      GFP_KERNEL);
2462                 if (!tmp_bitmap) {
2463                         free_page(page);
2464                         return -ENOMEM;
2465                 }
2466                 proc_skip_char(&kbuf, &left, '\n');
2467                 while (!err && left) {
2468                         unsigned long val_a, val_b;
2469                         bool neg;
2470
2471                         err = proc_get_long(&kbuf, &left, &val_a, &neg, tr_a,
2472                                              sizeof(tr_a), &c);
2473                         if (err)
2474                                 break;
2475                         if (val_a >= bitmap_len || neg) {
2476                                 err = -EINVAL;
2477                                 break;
2478                         }
2479
2480                         val_b = val_a;
2481                         if (left) {
2482                                 kbuf++;
2483                                 left--;
2484                         }
2485
2486                         if (c == '-') {
2487                                 err = proc_get_long(&kbuf, &left, &val_b,
2488                                                      &neg, tr_b, sizeof(tr_b),
2489                                                      &c);
2490                                 if (err)
2491                                         break;
2492                                 if (val_b >= bitmap_len || neg ||
2493                                     val_a > val_b) {
2494                                         err = -EINVAL;
2495                                         break;
2496                                 }
2497                                 if (left) {
2498                                         kbuf++;
2499                                         left--;
2500                                 }
2501                         }
2502
2503                         bitmap_set(tmp_bitmap, val_a, val_b - val_a + 1);
2504                         first = 0;
2505                         proc_skip_char(&kbuf, &left, '\n');
2506                 }
2507                 free_page(page);
2508         } else {
2509                 unsigned long bit_a, bit_b = 0;
2510
2511                 while (left) {
2512                         bit_a = find_next_bit(bitmap, bitmap_len, bit_b);
2513                         if (bit_a >= bitmap_len)
2514                                 break;
2515                         bit_b = find_next_zero_bit(bitmap, bitmap_len,
2516                                                    bit_a + 1) - 1;
2517
2518                         if (!first) {
2519                                 err = proc_put_char(&buffer, &left, ',');
2520                                 if (err)
2521                                         break;
2522                         }
2523                         err = proc_put_long(&buffer, &left, bit_a, false);
2524                         if (err)
2525                                 break;
2526                         if (bit_a != bit_b) {
2527                                 err = proc_put_char(&buffer, &left, '-');
2528                                 if (err)
2529                                         break;
2530                                 err = proc_put_long(&buffer, &left, bit_b, false);
2531                                 if (err)
2532                                         break;
2533                         }
2534
2535                         first = 0; bit_b++;
2536                 }
2537                 if (!err)
2538                         err = proc_put_char(&buffer, &left, '\n');
2539         }
2540
2541         if (!err) {
2542                 if (write) {
2543                         if (*ppos)
2544                                 bitmap_or(bitmap, bitmap, tmp_bitmap, bitmap_len);
2545                         else
2546                                 bitmap_copy(bitmap, tmp_bitmap, bitmap_len);
2547                 }
2548                 kfree(tmp_bitmap);
2549                 *lenp -= left;
2550                 *ppos += *lenp;
2551                 return 0;
2552         } else {
2553                 kfree(tmp_bitmap);
2554                 return err;
2555         }
2556 }
2557
2558 #else /* CONFIG_PROC_SYSCTL */
2559
2560 int proc_dostring(struct ctl_table *table, int write,
2561                   void __user *buffer, size_t *lenp, loff_t *ppos)
2562 {
2563         return -ENOSYS;
2564 }
2565
2566 int proc_dointvec(struct ctl_table *table, int write,
2567                   void __user *buffer, size_t *lenp, loff_t *ppos)
2568 {
2569         return -ENOSYS;
2570 }
2571
2572 int proc_dointvec_minmax(struct ctl_table *table, int write,
2573                     void __user *buffer, size_t *lenp, loff_t *ppos)
2574 {
2575         return -ENOSYS;
2576 }
2577
2578 int proc_dointvec_jiffies(struct ctl_table *table, int write,
2579                     void __user *buffer, size_t *lenp, loff_t *ppos)
2580 {
2581         return -ENOSYS;
2582 }
2583
2584 int proc_dointvec_userhz_jiffies(struct ctl_table *table, int write,
2585                     void __user *buffer, size_t *lenp, loff_t *ppos)
2586 {
2587         return -ENOSYS;
2588 }
2589
2590 int proc_dointvec_ms_jiffies(struct ctl_table *table, int write,
2591                              void __user *buffer, size_t *lenp, loff_t *ppos)
2592 {
2593         return -ENOSYS;
2594 }
2595
2596 int proc_doulongvec_minmax(struct ctl_table *table, int write,
2597                     void __user *buffer, size_t *lenp, loff_t *ppos)
2598 {
2599         return -ENOSYS;
2600 }
2601
2602 int proc_doulongvec_ms_jiffies_minmax(struct ctl_table *table, int write,
2603                                       void __user *buffer,
2604                                       size_t *lenp, loff_t *ppos)
2605 {
2606     return -ENOSYS;
2607 }
2608
2609
2610 #endif /* CONFIG_PROC_SYSCTL */
2611
2612 /*
2613  * No sense putting this after each symbol definition, twice,
2614  * exception granted :-)
2615  */
2616 EXPORT_SYMBOL(proc_dointvec);
2617 EXPORT_SYMBOL(proc_dointvec_jiffies);
2618 EXPORT_SYMBOL(proc_dointvec_minmax);
2619 EXPORT_SYMBOL(proc_dointvec_userhz_jiffies);
2620 EXPORT_SYMBOL(proc_dointvec_ms_jiffies);
2621 EXPORT_SYMBOL(proc_dostring);
2622 EXPORT_SYMBOL(proc_doulongvec_minmax);
2623 EXPORT_SYMBOL(proc_doulongvec_ms_jiffies_minmax);