kernel/sysctl.c

   1 /*
   2  * sysctl.c: General linux system control interface
   3  *
   4  * Begun 24 March 1995, Stephen Tweedie
   5  * Added /proc support, Dec 1995
   6  * Added bdflush entry and intvec min/max checking, 2/23/96, Tom Dyas.
   7  * Added hooks for /proc/sys/net (minor, minor patch), 96/4/1, Mike Shaver.
   8  * Added kernel/java-{interpreter,appletviewer}, 96/5/10, Mike Shaver.
   9  * Dynamic registration fixes, Stephen Tweedie.
  10  * Added kswapd-interval, ctrl-alt-del, printk stuff, 1/8/97, Chris Horn.
  11  * Made sysctl support optional via CONFIG_SYSCTL, 1/10/97, Chris
  12  *  Horn.
  13  * Added proc_doulongvec_ms_jiffies_minmax, 09/08/99, Carlos H. Bauer.
  14  * Added proc_doulongvec_minmax, 09/08/99, Carlos H. Bauer.
  15  * Changed linked lists to use list.h instead of lists.h, 02/24/00, Bill
  16  *  Wendling.
  17  * The list_for_each() macro wasn't appropriate for the sysctl loop.
  18  *  Removed it and replaced it with older style, 03/23/00, Bill Wendling
  19  */
  20
  21 #include <linux/module.h>
  22 #include <linux/mm.h>
  23 #include <linux/swap.h>
  24 #include <linux/slab.h>
  25 #include <linux/sysctl.h>
  26 #include <linux/bitmap.h>
  27 #include <linux/signal.h>
  28 #include <linux/printk.h>
  29 #include <linux/proc_fs.h>
  30 #include <linux/security.h>
  31 #include <linux/ctype.h>
  32 #include <linux/kmemcheck.h>
  33 #include <linux/kmemleak.h>
  34 #include <linux/fs.h>
  35 #include <linux/init.h>
  36 #include <linux/kernel.h>
  37 #include <linux/kobject.h>
  38 #include <linux/net.h>
  39 #include <linux/sysrq.h>
  40 #include <linux/highuid.h>
  41 #include <linux/writeback.h>
  42 #include <linux/ratelimit.h>
  43 #include <linux/compaction.h>
  44 #include <linux/hugetlb.h>
  45 #include <linux/initrd.h>
  46 #include <linux/key.h>
  47 #include <linux/times.h>
  48 #include <linux/limits.h>
  49 #include <linux/dcache.h>
  50 #include <linux/dnotify.h>
  51 #include <linux/syscalls.h>
  52 #include <linux/vmstat.h>
  53 #include <linux/nfs_fs.h>
  54 #include <linux/acpi.h>
  55 #include <linux/reboot.h>
  56 #include <linux/ftrace.h>
  57 #include <linux/perf_event.h>
  58 #include <linux/kprobes.h>
  59 #include <linux/pipe_fs_i.h>
  60 #include <linux/oom.h>
  61 #include <linux/kmod.h>
  62 #include <linux/capability.h>
  63 #include <linux/binfmts.h>
  64 #include <linux/sched/sysctl.h>
  65
  66 #include <asm/uaccess.h>
  67 #include <asm/processor.h>
  68
  69 #ifdef CONFIG_X86
  70 #include <asm/nmi.h>
  71 #include <asm/stacktrace.h>
  72 #include <asm/io.h>
  73 #endif
  74 #ifdef CONFIG_SPARC
  75 #include <asm/setup.h>
  76 #endif
  77 #ifdef CONFIG_BSD_PROCESS_ACCT
  78 #include <linux/acct.h>
  79 #endif
  80 #ifdef CONFIG_RT_MUTEXES
  81 #include <linux/rtmutex.h>
  82 #endif
  83 #if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_LOCK_STAT)
  84 #include <linux/lockdep.h>
  85 #endif
  86 #ifdef CONFIG_CHR_DEV_SG
  87 #include <scsi/sg.h>
  88 #endif
  89
  90 #ifdef CONFIG_LOCKUP_DETECTOR
  91 #include <linux/nmi.h>
  92 #endif
  93
  94
  95 #if defined(CONFIG_SYSCTL)
  96
  97 /* External variables not in a header file. */
  98 extern int sysctl_overcommit_memory;
  99 extern int sysctl_overcommit_ratio;
 100 extern int max_threads;
 101 extern int suid_dumpable;
 102 #ifdef CONFIG_COREDUMP
 103 extern int core_uses_pid;
 104 extern char core_pattern[];
 105 extern unsigned int core_pipe_limit;
 106 #endif
 107 extern int pid_max;
 108 extern int pid_max_min, pid_max_max;
 109 extern int percpu_pagelist_fraction;
 110 extern int compat_log;
 111 extern int latencytop_enabled;
 112 extern int sysctl_nr_open_min, sysctl_nr_open_max;
 113 #ifndef CONFIG_MMU
 114 extern int sysctl_nr_trim_pages;
 115 #endif
 116 #ifdef CONFIG_BLOCK
 117 extern int blk_iopoll_enabled;
 118 #endif
 119
 120 /* Constants used for minimum and  maximum */
 121 #ifdef CONFIG_LOCKUP_DETECTOR
 122 static int sixty = 60;
 123 #endif
 124
 125 static int zero;
 126 static int __maybe_unused one = 1;
 127 static int __maybe_unused two = 2;
 128 static int __maybe_unused three = 3;
 129 static unsigned long one_ul = 1;
 130 static int one_hundred = 100;
 131 #ifdef CONFIG_PRINTK
 132 static int ten_thousand = 10000;
 133 #endif
 134
 135 /* this is needed for the proc_doulongvec_minmax of vm_dirty_bytes */
 136 static unsigned long dirty_bytes_min = 2 * PAGE_SIZE;
 137
 138 /* this is needed for the proc_dointvec_minmax for [fs_]overflow UID and GID */
 139 static int maxolduid = 65535;
 140 static int minolduid;
 141 static int min_percpu_pagelist_fract = 8;
 142
 143 static int ngroups_max = NGROUPS_MAX;
 144 static const int cap_last_cap = CAP_LAST_CAP;
 145
 146 #ifdef CONFIG_INOTIFY_USER
 147 #include <linux/inotify.h>
 148 #endif
 149 #ifdef CONFIG_SPARC
 150 #endif
 151
 152 #ifdef CONFIG_SPARC64
 153 extern int sysctl_tsb_ratio;
 154 #endif
 155
 156 #ifdef __hppa__
 157 extern int pwrsw_enabled;
 158 #endif
 159
 160 #ifdef CONFIG_SYSCTL_ARCH_UNALIGN_ALLOW
 161 extern int unaligned_enabled;
 162 #endif
 163
 164 #ifdef CONFIG_IA64
 165 extern int unaligned_dump_stack;
 166 #endif
 167
 168 #ifdef CONFIG_SYSCTL_ARCH_UNALIGN_NO_WARN
 169 extern int no_unaligned_warning;
 170 #endif
 171
 172 #ifdef CONFIG_PROC_SYSCTL
 173 static int proc_do_cad_pid(struct ctl_table *table, int write,
 174                   void __user *buffer, size_t *lenp, loff_t *ppos);
 175 static int proc_taint(struct ctl_table *table, int write,
 176                                void __user *buffer, size_t *lenp, loff_t *ppos);
 177 #endif
 178
 179 #ifdef CONFIG_PRINTK
 180 static int proc_dointvec_minmax_sysadmin(struct ctl_table *table, int write,
 181                                 void __user *buffer, size_t *lenp, loff_t *ppos);
 182 #endif
 183
 184 static int proc_dointvec_minmax_coredump(struct ctl_table *table, int write,
 185                 void __user *buffer, size_t *lenp, loff_t *ppos);
 186 #ifdef CONFIG_COREDUMP
 187 static int proc_dostring_coredump(struct ctl_table *table, int write,
 188                 void __user *buffer, size_t *lenp, loff_t *ppos);
 189 #endif
 190
 191 #ifdef CONFIG_MAGIC_SYSRQ
 192 /* Note: sysrq code uses it's own private copy */
 193 static int __sysrq_enabled = SYSRQ_DEFAULT_ENABLE;
 194
 195 static int sysrq_sysctl_handler(ctl_table *table, int write,
 196                                 void __user *buffer, size_t *lenp,
 197                                 loff_t *ppos)
 198 {
 199         int error;
 200
 201         error = proc_dointvec(table, write, buffer, lenp, ppos);
 202         if (error)
 203                 return error;
 204
 205         if (write)
 206                 sysrq_toggle_support(__sysrq_enabled);
 207
 208         return 0;
 209 }
 210
 211 #endif
 212
 213 static struct ctl_table kern_table[];
 214 static struct ctl_table vm_table[];
 215 static struct ctl_table fs_table[];
 216 static struct ctl_table debug_table[];
 217 static struct ctl_table dev_table[];
 218 extern struct ctl_table random_table[];
 219 #ifdef CONFIG_EPOLL
 220 extern struct ctl_table epoll_table[];
 221 #endif
 222
 223 #ifdef HAVE_ARCH_PICK_MMAP_LAYOUT
 224 int sysctl_legacy_va_layout;
 225 #endif
 226
 227 /* The default sysctl tables: */
 228
 229 static struct ctl_table sysctl_base_table[] = {
 230         {
 231                 .procname       = "kernel",
 232                 .mode           = 0555,
 233                 .child          = kern_table,
 234         },
 235         {
 236                 .procname       = "vm",
 237                 .mode           = 0555,
 238                 .child          = vm_table,
 239         },
 240         {
 241                 .procname       = "fs",
 242                 .mode           = 0555,
 243                 .child          = fs_table,
 244         },
 245         {
 246                 .procname       = "debug",
 247                 .mode           = 0555,
 248                 .child          = debug_table,
 249         },
 250         {
 251                 .procname       = "dev",
 252                 .mode           = 0555,
 253                 .child          = dev_table,
 254         },
 255         { }
 256 };
 257
 258 #ifdef CONFIG_SCHED_DEBUG
 259 static int min_sched_granularity_ns = 100000;           /* 100 usecs */
 260 static int max_sched_granularity_ns = NSEC_PER_SEC;     /* 1 second */
 261 static int min_wakeup_granularity_ns;                   /* 0 usecs */
 262 static int max_wakeup_granularity_ns = NSEC_PER_SEC;    /* 1 second */
 263 #ifdef CONFIG_SMP
 264 static int min_sched_tunable_scaling = SCHED_TUNABLESCALING_NONE;
 265 static int max_sched_tunable_scaling = SCHED_TUNABLESCALING_END-1;
 266 #endif /* CONFIG_SMP */
 267 #endif /* CONFIG_SCHED_DEBUG */
 268
 269 #ifdef CONFIG_COMPACTION
 270 static int min_extfrag_threshold;
 271 static int max_extfrag_threshold = 1000;
 272 #endif
 273
 274 static struct ctl_table kern_table[] = {
 275         {
 276                 .procname       = "sched_child_runs_first",
 277                 .data           = &sysctl_sched_child_runs_first,
 278                 .maxlen         = sizeof(unsigned int),
 279                 .mode           = 0644,
 280                 .proc_handler   = proc_dointvec,
 281         },
 282 #ifdef CONFIG_SCHED_DEBUG
 283         {
 284                 .procname       = "sched_min_granularity_ns",
 285                 .data           = &sysctl_sched_min_granularity,
 286                 .maxlen         = sizeof(unsigned int),
 287                 .mode           = 0644,
 288                 .proc_handler   = sched_proc_update_handler,
 289                 .extra1         = &min_sched_granularity_ns,
 290                 .extra2         = &max_sched_granularity_ns,
 291         },
 292         {
 293                 .procname       = "sched_latency_ns",
 294                 .data           = &sysctl_sched_latency,
 295                 .maxlen         = sizeof(unsigned int),
 296                 .mode           = 0644,
 297                 .proc_handler   = sched_proc_update_handler,
 298                 .extra1         = &min_sched_granularity_ns,
 299                 .extra2         = &max_sched_granularity_ns,
 300         },
 301         {
 302                 .procname       = "sched_wakeup_granularity_ns",
 303                 .data           = &sysctl_sched_wakeup_granularity,
 304                 .maxlen         = sizeof(unsigned int),
 305                 .mode           = 0644,
 306                 .proc_handler   = sched_proc_update_handler,
 307                 .extra1         = &min_wakeup_granularity_ns,
 308                 .extra2         = &max_wakeup_granularity_ns,
 309         },
 310 #ifdef CONFIG_SMP
 311         {
 312                 .procname       = "sched_tunable_scaling",
 313                 .data           = &sysctl_sched_tunable_scaling,
 314                 .maxlen         = sizeof(enum sched_tunable_scaling),
 315                 .mode           = 0644,
 316                 .proc_handler   = sched_proc_update_handler,
 317                 .extra1         = &min_sched_tunable_scaling,
 318                 .extra2         = &max_sched_tunable_scaling,
 319         },
 320         {
 321                 .procname       = "sched_migration_cost_ns",
 322                 .data           = &sysctl_sched_migration_cost,
 323                 .maxlen         = sizeof(unsigned int),
 324                 .mode           = 0644,
 325                 .proc_handler   = proc_dointvec,
 326         },
 327         {
 328                 .procname       = "sched_nr_migrate",
 329                 .data           = &sysctl_sched_nr_migrate,
 330                 .maxlen         = sizeof(unsigned int),
 331                 .mode           = 0644,
 332                 .proc_handler   = proc_dointvec,
 333         },
 334         {
 335                 .procname       = "sched_time_avg_ms",
 336                 .data           = &sysctl_sched_time_avg,
 337                 .maxlen         = sizeof(unsigned int),
 338                 .mode           = 0644,
 339                 .proc_handler   = proc_dointvec,
 340         },
 341         {
 342                 .procname       = "sched_shares_window_ns",
 343                 .data           = &sysctl_sched_shares_window,
 344                 .maxlen         = sizeof(unsigned int),
 345                 .mode           = 0644,
 346                 .proc_handler   = proc_dointvec,
 347         },
 348         {
 349                 .procname       = "timer_migration",
 350                 .data           = &sysctl_timer_migration,
 351                 .maxlen         = sizeof(unsigned int),
 352                 .mode           = 0644,
 353                 .proc_handler   = proc_dointvec_minmax,
 354                 .extra1         = &zero,
 355                 .extra2         = &one,
 356         },
 357 #endif /* CONFIG_SMP */
 358 #ifdef CONFIG_NUMA_BALANCING
 359         {
 360                 .procname       = "numa_balancing_scan_delay_ms",
 361                 .data           = &sysctl_numa_balancing_scan_delay,
 362                 .maxlen         = sizeof(unsigned int),
 363                 .mode           = 0644,
 364                 .proc_handler   = proc_dointvec,
 365         },
 366         {
 367                 .procname       = "numa_balancing_scan_period_min_ms",
 368                 .data           = &sysctl_numa_balancing_scan_period_min,
 369                 .maxlen         = sizeof(unsigned int),
 370                 .mode           = 0644,
 371                 .proc_handler   = proc_dointvec,
 372         },
 373         {
 374                 .procname       = "numa_balancing_scan_period_reset",
 375                 .data           = &sysctl_numa_balancing_scan_period_reset,
 376                 .maxlen         = sizeof(unsigned int),
 377                 .mode           = 0644,
 378                 .proc_handler   = proc_dointvec,
 379         },
 380         {
 381                 .procname       = "numa_balancing_scan_period_max_ms",
 382                 .data           = &sysctl_numa_balancing_scan_period_max,
 383                 .maxlen         = sizeof(unsigned int),
 384                 .mode           = 0644,
 385                 .proc_handler   = proc_dointvec,
 386         },
 387         {
 388                 .procname       = "numa_balancing_scan_size_mb",
 389                 .data           = &sysctl_numa_balancing_scan_size,
 390                 .maxlen         = sizeof(unsigned int),
 391                 .mode           = 0644,
 392                 .proc_handler   = proc_dointvec,
 393         },
 394 #endif /* CONFIG_NUMA_BALANCING */
 395 #endif /* CONFIG_SCHED_DEBUG */
 396         {
 397                 .procname       = "sched_rt_period_us",
 398                 .data           = &sysctl_sched_rt_period,
 399                 .maxlen         = sizeof(unsigned int),
 400                 .mode           = 0644,
 401                 .proc_handler   = sched_rt_handler,
 402         },
 403         {
 404                 .procname       = "sched_rt_runtime_us",
 405                 .data           = &sysctl_sched_rt_runtime,
 406                 .maxlen         = sizeof(int),
 407                 .mode           = 0644,
 408                 .proc_handler   = sched_rt_handler,
 409         },
 410         {
 411                 .procname       = "sched_rr_timeslice_ms",
 412                 .data           = &sched_rr_timeslice,
 413                 .maxlen         = sizeof(int),
 414                 .mode           = 0644,
 415                 .proc_handler   = sched_rr_handler,
 416         },
 417 #ifdef CONFIG_SCHED_AUTOGROUP
 418         {
 419                 .procname       = "sched_autogroup_enabled",
 420                 .data           = &sysctl_sched_autogroup_enabled,
 421                 .maxlen         = sizeof(unsigned int),
 422                 .mode           = 0644,
 423                 .proc_handler   = proc_dointvec_minmax,
 424                 .extra1         = &zero,
 425                 .extra2         = &one,
 426         },
 427 #endif
 428 #ifdef CONFIG_CFS_BANDWIDTH
 429         {
 430                 .procname       = "sched_cfs_bandwidth_slice_us",
 431                 .data           = &sysctl_sched_cfs_bandwidth_slice,
 432                 .maxlen         = sizeof(unsigned int),
 433                 .mode           = 0644,
 434                 .proc_handler   = proc_dointvec_minmax,
 435                 .extra1         = &one,
 436         },
 437 #endif
 438 #ifdef CONFIG_PROVE_LOCKING
 439         {
 440                 .procname       = "prove_locking",
 441                 .data           = &prove_locking,
 442                 .maxlen         = sizeof(int),
 443                 .mode           = 0644,
 444                 .proc_handler   = proc_dointvec,
 445         },
 446 #endif
 447 #ifdef CONFIG_LOCK_STAT
 448         {
 449                 .procname       = "lock_stat",
 450                 .data           = &lock_stat,
 451                 .maxlen         = sizeof(int),
 452                 .mode           = 0644,
 453                 .proc_handler   = proc_dointvec,
 454         },
 455 #endif
 456         {
 457                 .procname       = "panic",
 458                 .data           = &panic_timeout,
 459                 .maxlen         = sizeof(int),
 460                 .mode           = 0644,
 461                 .proc_handler   = proc_dointvec,
 462         },
 463 #ifdef CONFIG_COREDUMP
 464         {
 465                 .procname       = "core_uses_pid",
 466                 .data           = &core_uses_pid,
 467                 .maxlen         = sizeof(int),
 468                 .mode           = 0644,
 469                 .proc_handler   = proc_dointvec,
 470         },
 471         {
 472                 .procname       = "core_pattern",
 473                 .data           = core_pattern,
 474                 .maxlen         = CORENAME_MAX_SIZE,
 475                 .mode           = 0644,
 476                 .proc_handler   = proc_dostring_coredump,
 477         },
 478         {
 479                 .procname       = "core_pipe_limit",
 480                 .data           = &core_pipe_limit,
 481                 .maxlen         = sizeof(unsigned int),
 482                 .mode           = 0644,
 483                 .proc_handler   = proc_dointvec,
 484         },
 485 #endif
 486 #ifdef CONFIG_PROC_SYSCTL
 487         {
 488                 .procname       = "tainted",
 489                 .maxlen         = sizeof(long),
 490                 .mode           = 0644,
 491                 .proc_handler   = proc_taint,
 492         },
 493 #endif
 494 #ifdef CONFIG_LATENCYTOP
 495         {
 496                 .procname       = "latencytop",
 497                 .data           = &latencytop_enabled,
 498                 .maxlen         = sizeof(int),
 499                 .mode           = 0644,
 500                 .proc_handler   = proc_dointvec,
 501         },
 502 #endif
 503 #ifdef CONFIG_BLK_DEV_INITRD
 504         {
 505                 .procname       = "real-root-dev",
 506                 .data           = &real_root_dev,
 507                 .maxlen         = sizeof(int),
 508                 .mode           = 0644,
 509                 .proc_handler   = proc_dointvec,
 510         },
 511 #endif
 512         {
 513                 .procname       = "print-fatal-signals",
 514                 .data           = &print_fatal_signals,
 515                 .maxlen         = sizeof(int),
 516                 .mode           = 0644,
 517                 .proc_handler   = proc_dointvec,
 518         },
 519 #ifdef CONFIG_SPARC
 520         {
 521                 .procname       = "reboot-cmd",
 522                 .data           = reboot_command,
 523                 .maxlen         = 256,
 524                 .mode           = 0644,
 525                 .proc_handler   = proc_dostring,
 526         },
 527         {
 528                 .procname       = "stop-a",
 529                 .data           = &stop_a_enabled,
 530                 .maxlen         = sizeof (int),
 531                 .mode           = 0644,
 532                 .proc_handler   = proc_dointvec,
 533         },
 534         {
 535                 .procname       = "scons-poweroff",
 536                 .data           = &scons_pwroff,
 537                 .maxlen         = sizeof (int),
 538                 .mode           = 0644,
 539                 .proc_handler   = proc_dointvec,
 540         },
 541 #endif
 542 #ifdef CONFIG_SPARC64
 543         {
 544                 .procname       = "tsb-ratio",
 545                 .data           = &sysctl_tsb_ratio,
 546                 .maxlen         = sizeof (int),
 547                 .mode           = 0644,
 548                 .proc_handler   = proc_dointvec,
 549         },
 550 #endif
 551 #ifdef __hppa__
 552         {
 553                 .procname       = "soft-power",
 554                 .data           = &pwrsw_enabled,
 555                 .maxlen         = sizeof (int),
 556                 .mode           = 0644,
 557                 .proc_handler   = proc_dointvec,
 558         },
 559 #endif
 560 #ifdef CONFIG_SYSCTL_ARCH_UNALIGN_ALLOW
 561         {
 562                 .procname       = "unaligned-trap",
 563                 .data           = &unaligned_enabled,
 564                 .maxlen         = sizeof (int),
 565                 .mode           = 0644,
 566                 .proc_handler   = proc_dointvec,
 567         },
 568 #endif
 569         {
 570                 .procname       = "ctrl-alt-del",
 571                 .data           = &C_A_D,
 572                 .maxlen         = sizeof(int),
 573                 .mode           = 0644,
 574                 .proc_handler   = proc_dointvec,
 575         },
 576 #ifdef CONFIG_FUNCTION_TRACER
 577         {
 578                 .procname       = "ftrace_enabled",
 579                 .data           = &ftrace_enabled,
 580                 .maxlen         = sizeof(int),
 581                 .mode           = 0644,
 582                 .proc_handler   = ftrace_enable_sysctl,
 583         },
 584 #endif
 585 #ifdef CONFIG_STACK_TRACER
 586         {
 587                 .procname       = "stack_tracer_enabled",
 588                 .data           = &stack_tracer_enabled,
 589                 .maxlen         = sizeof(int),
 590                 .mode           = 0644,
 591                 .proc_handler   = stack_trace_sysctl,
 592         },
 593 #endif
 594 #ifdef CONFIG_TRACING
 595         {
 596                 .procname       = "ftrace_dump_on_oops",
 597                 .data           = &ftrace_dump_on_oops,
 598                 .maxlen         = sizeof(int),
 599                 .mode           = 0644,
 600                 .proc_handler   = proc_dointvec,
 601         },
 602         {
 603                 .procname       = "traceoff_on_warning",
 604                 .data           = &__disable_trace_on_warning,
 605                 .maxlen         = sizeof(__disable_trace_on_warning),
 606                 .mode           = 0644,
 607                 .proc_handler   = proc_dointvec,
 608         },
 609 #endif
 610 #ifdef CONFIG_MODULES
 611         {
 612                 .procname       = "modprobe",
 613                 .data           = &modprobe_path,
 614                 .maxlen         = KMOD_PATH_LEN,
 615                 .mode           = 0644,
 616                 .proc_handler   = proc_dostring,
 617         },
 618         {
 619                 .procname       = "modules_disabled",
 620                 .data           = &modules_disabled,
 621                 .maxlen         = sizeof(int),
 622                 .mode           = 0644,
 623                 /* only handle a transition from default "0" to "1" */
 624                 .proc_handler   = proc_dointvec_minmax,
 625                 .extra1         = &one,
 626                 .extra2         = &one,
 627         },
 628 #endif
 629
 630         {
 631                 .procname       = "hotplug",
 632                 .data           = &uevent_helper,
 633                 .maxlen         = UEVENT_HELPER_PATH_LEN,
 634                 .mode           = 0644,
 635                 .proc_handler   = proc_dostring,
 636         },
 637
 638 #ifdef CONFIG_CHR_DEV_SG
 639         {
 640                 .procname       = "sg-big-buff",
 641                 .data           = &sg_big_buff,
 642                 .maxlen         = sizeof (int),
 643                 .mode           = 0444,
 644                 .proc_handler   = proc_dointvec,
 645         },
 646 #endif
 647 #ifdef CONFIG_BSD_PROCESS_ACCT
 648         {
 649                 .procname       = "acct",
 650                 .data           = &acct_parm,
 651                 .maxlen         = 3*sizeof(int),
 652                 .mode           = 0644,
 653                 .proc_handler   = proc_dointvec,
 654         },
 655 #endif
 656 #ifdef CONFIG_MAGIC_SYSRQ
 657         {
 658                 .procname       = "sysrq",
 659                 .data           = &__sysrq_enabled,
 660                 .maxlen         = sizeof (int),
 661                 .mode           = 0644,
 662                 .proc_handler   = sysrq_sysctl_handler,
 663         },
 664 #endif
 665 #ifdef CONFIG_PROC_SYSCTL
 666         {
 667                 .procname       = "cad_pid",
 668                 .data           = NULL,
 669                 .maxlen         = sizeof (int),
 670                 .mode           = 0600,
 671                 .proc_handler   = proc_do_cad_pid,
 672         },
 673 #endif
 674         {
 675                 .procname       = "threads-max",
 676                 .data           = &max_threads,
 677                 .maxlen         = sizeof(int),
 678                 .mode           = 0644,
 679                 .proc_handler   = proc_dointvec,
 680         },
 681         {
 682                 .procname       = "random",
 683                 .mode           = 0555,
 684                 .child          = random_table,
 685         },
 686         {
 687                 .procname       = "usermodehelper",
 688                 .mode           = 0555,
 689                 .child          = usermodehelper_table,
 690         },
 691         {
 692                 .procname       = "overflowuid",
 693                 .data           = &overflowuid,
 694                 .maxlen         = sizeof(int),
 695                 .mode           = 0644,
 696                 .proc_handler   = proc_dointvec_minmax,
 697                 .extra1         = &minolduid,
 698                 .extra2         = &maxolduid,
 699         },
 700         {
 701                 .procname       = "overflowgid",
 702                 .data           = &overflowgid,
 703                 .maxlen         = sizeof(int),
 704                 .mode           = 0644,
 705                 .proc_handler   = proc_dointvec_minmax,
 706                 .extra1         = &minolduid,
 707                 .extra2         = &maxolduid,
 708         },
 709 #ifdef CONFIG_S390
 710 #ifdef CONFIG_MATHEMU
 711         {
 712                 .procname       = "ieee_emulation_warnings",
 713                 .data           = &sysctl_ieee_emulation_warnings,
 714                 .maxlen         = sizeof(int),
 715                 .mode           = 0644,
 716                 .proc_handler   = proc_dointvec,
 717         },
 718 #endif
 719         {
 720                 .procname       = "userprocess_debug",
 721                 .data           = &show_unhandled_signals,
 722                 .maxlen         = sizeof(int),
 723                 .mode           = 0644,
 724                 .proc_handler   = proc_dointvec,
 725         },
 726 #endif
 727         {
 728                 .procname       = "pid_max",
 729                 .data           = &pid_max,
 730                 .maxlen         = sizeof (int),
 731                 .mode           = 0644,
 732                 .proc_handler   = proc_dointvec_minmax,
 733                 .extra1         = &pid_max_min,
 734                 .extra2         = &pid_max_max,
 735         },
 736         {
 737                 .procname       = "panic_on_oops",
 738                 .data           = &panic_on_oops,
 739                 .maxlen         = sizeof(int),
 740                 .mode           = 0644,
 741                 .proc_handler   = proc_dointvec,
 742         },
 743 #if defined CONFIG_PRINTK
 744         {
 745                 .procname       = "printk",
 746                 .data           = &console_loglevel,
 747                 .maxlen         = 4*sizeof(int),
 748                 .mode           = 0644,
 749                 .proc_handler   = proc_dointvec,
 750         },
 751         {
 752                 .procname       = "printk_ratelimit",
 753                 .data           = &printk_ratelimit_state.interval,
 754                 .maxlen         = sizeof(int),
 755                 .mode           = 0644,
 756                 .proc_handler   = proc_dointvec_jiffies,
 757         },
 758         {
 759                 .procname       = "printk_ratelimit_burst",
 760                 .data           = &printk_ratelimit_state.burst,
 761                 .maxlen         = sizeof(int),
 762                 .mode           = 0644,
 763                 .proc_handler   = proc_dointvec,
 764         },
 765         {
 766                 .procname       = "printk_delay",
 767                 .data           = &printk_delay_msec,
 768                 .maxlen         = sizeof(int),
 769                 .mode           = 0644,
 770                 .proc_handler   = proc_dointvec_minmax,
 771                 .extra1         = &zero,
 772                 .extra2         = &ten_thousand,
 773         },
 774         {
 775                 .procname       = "dmesg_restrict",
 776                 .data           = &dmesg_restrict,
 777                 .maxlen         = sizeof(int),
 778                 .mode           = 0644,
 779                 .proc_handler   = proc_dointvec_minmax_sysadmin,
 780                 .extra1         = &zero,
 781                 .extra2         = &one,
 782         },
 783         {
 784                 .procname       = "kptr_restrict",
 785                 .data           = &kptr_restrict,
 786                 .maxlen         = sizeof(int),
 787                 .mode           = 0644,
 788                 .proc_handler   = proc_dointvec_minmax_sysadmin,
 789                 .extra1         = &zero,
 790                 .extra2         = &two,
 791         },
 792 #endif
 793         {
 794                 .procname       = "ngroups_max",
 795                 .data           = &ngroups_max,
 796                 .maxlen         = sizeof (int),
 797                 .mode           = 0444,
 798                 .proc_handler   = proc_dointvec,
 799         },
 800         {
 801                 .procname       = "cap_last_cap",
 802                 .data           = (void *)&cap_last_cap,
 803                 .maxlen         = sizeof(int),
 804                 .mode           = 0444,
 805                 .proc_handler   = proc_dointvec,
 806         },
 807 #if defined(CONFIG_LOCKUP_DETECTOR)
 808         {
 809                 .procname       = "watchdog",
 810                 .data           = &watchdog_user_enabled,
 811                 .maxlen         = sizeof (int),
 812                 .mode           = 0644,
 813                 .proc_handler   = proc_dowatchdog,
 814                 .extra1         = &zero,
 815                 .extra2         = &one,
 816         },
 817         {
 818                 .procname       = "watchdog_thresh",
 819                 .data           = &watchdog_thresh,
 820                 .maxlen         = sizeof(int),
 821                 .mode           = 0644,
 822                 .proc_handler   = proc_dowatchdog,
 823                 .extra1         = &zero,
 824                 .extra2         = &sixty,
 825         },
 826         {
 827                 .procname       = "softlockup_panic",
 828                 .data           = &softlockup_panic,
 829                 .maxlen         = sizeof(int),
 830                 .mode           = 0644,
 831                 .proc_handler   = proc_dointvec_minmax,
 832                 .extra1         = &zero,
 833                 .extra2         = &one,
 834         },
 835         {
 836                 .procname       = "nmi_watchdog",
 837                 .data           = &watchdog_user_enabled,
 838                 .maxlen         = sizeof (int),
 839                 .mode           = 0644,
 840                 .proc_handler   = proc_dowatchdog,
 841                 .extra1         = &zero,
 842                 .extra2         = &one,
 843         },
 844 #endif
 845 #if defined(CONFIG_X86_LOCAL_APIC) && defined(CONFIG_X86)
 846         {
 847                 .procname       = "unknown_nmi_panic",
 848                 .data           = &unknown_nmi_panic,
 849                 .maxlen         = sizeof (int),
 850                 .mode           = 0644,
 851                 .proc_handler   = proc_dointvec,
 852         },
 853 #endif
 854 #if defined(CONFIG_X86)
 855         {
 856                 .procname       = "panic_on_unrecovered_nmi",
 857                 .data           = &panic_on_unrecovered_nmi,
 858                 .maxlen         = sizeof(int),
 859                 .mode           = 0644,
 860                 .proc_handler   = proc_dointvec,
 861         },
 862         {
 863                 .procname       = "panic_on_io_nmi",
 864                 .data           = &panic_on_io_nmi,
 865                 .maxlen         = sizeof(int),
 866                 .mode           = 0644,
 867                 .proc_handler   = proc_dointvec,
 868         },
 869 #ifdef CONFIG_DEBUG_STACKOVERFLOW
 870         {
 871                 .procname       = "panic_on_stackoverflow",
 872                 .data           = &sysctl_panic_on_stackoverflow,
 873                 .maxlen         = sizeof(int),
 874                 .mode           = 0644,
 875                 .proc_handler   = proc_dointvec,
 876         },
 877 #endif
 878         {
 879                 .procname       = "bootloader_type",
 880                 .data           = &bootloader_type,
 881                 .maxlen         = sizeof (int),
 882                 .mode           = 0444,
 883                 .proc_handler   = proc_dointvec,
 884         },
 885         {
 886                 .procname       = "bootloader_version",
 887                 .data           = &bootloader_version,
 888                 .maxlen         = sizeof (int),
 889                 .mode           = 0444,
 890                 .proc_handler   = proc_dointvec,
 891         },
 892         {
 893                 .procname       = "kstack_depth_to_print",
 894                 .data           = &kstack_depth_to_print,
 895                 .maxlen         = sizeof(int),
 896                 .mode           = 0644,
 897                 .proc_handler   = proc_dointvec,
 898         },
 899         {
 900                 .procname       = "io_delay_type",
 901                 .data           = &io_delay_type,
 902                 .maxlen         = sizeof(int),
 903                 .mode           = 0644,
 904                 .proc_handler   = proc_dointvec,
 905         },
 906 #endif
 907 #if defined(CONFIG_MMU)
 908         {
 909                 .procname       = "randomize_va_space",
 910                 .data           = &randomize_va_space,
 911                 .maxlen         = sizeof(int),
 912                 .mode           = 0644,
 913                 .proc_handler   = proc_dointvec,
 914         },
 915 #endif
 916 #if defined(CONFIG_S390) && defined(CONFIG_SMP)
 917         {
 918                 .procname       = "spin_retry",
 919                 .data           = &spin_retry,
 920                 .maxlen         = sizeof (int),
 921                 .mode           = 0644,
 922                 .proc_handler   = proc_dointvec,
 923         },
 924 #endif
 925 #if     defined(CONFIG_ACPI_SLEEP) && defined(CONFIG_X86)
 926         {
 927                 .procname       = "acpi_video_flags",
 928                 .data           = &acpi_realmode_flags,
 929                 .maxlen         = sizeof (unsigned long),
 930                 .mode           = 0644,
 931                 .proc_handler   = proc_doulongvec_minmax,
 932         },
 933 #endif
 934 #ifdef CONFIG_SYSCTL_ARCH_UNALIGN_NO_WARN
 935         {
 936                 .procname       = "ignore-unaligned-usertrap",
 937                 .data           = &no_unaligned_warning,
 938                 .maxlen         = sizeof (int),
 939                 .mode           = 0644,
 940                 .proc_handler   = proc_dointvec,
 941         },
 942 #endif
 943 #ifdef CONFIG_IA64
 944         {
 945                 .procname       = "unaligned-dump-stack",
 946                 .data           = &unaligned_dump_stack,
 947                 .maxlen         = sizeof (int),
 948                 .mode           = 0644,
 949                 .proc_handler   = proc_dointvec,
 950         },
 951 #endif
 952 #ifdef CONFIG_DETECT_HUNG_TASK
 953         {
 954                 .procname       = "hung_task_panic",
 955                 .data           = &sysctl_hung_task_panic,
 956                 .maxlen         = sizeof(int),
 957                 .mode           = 0644,
 958                 .proc_handler   = proc_dointvec_minmax,
 959                 .extra1         = &zero,
 960                 .extra2         = &one,
 961         },
 962         {
 963                 .procname       = "hung_task_check_count",
 964                 .data           = &sysctl_hung_task_check_count,
 965                 .maxlen         = sizeof(unsigned long),
 966                 .mode           = 0644,
 967                 .proc_handler   = proc_doulongvec_minmax,
 968         },
 969         {
 970                 .procname       = "hung_task_timeout_secs",
 971                 .data           = &sysctl_hung_task_timeout_secs,
 972                 .maxlen         = sizeof(unsigned long),
 973                 .mode           = 0644,
 974                 .proc_handler   = proc_dohung_task_timeout_secs,
 975         },
 976         {
 977                 .procname       = "hung_task_warnings",
 978                 .data           = &sysctl_hung_task_warnings,
 979                 .maxlen         = sizeof(unsigned long),
 980                 .mode           = 0644,
 981                 .proc_handler   = proc_doulongvec_minmax,
 982         },
 983 #endif
 984 #ifdef CONFIG_COMPAT
 985         {
 986                 .procname       = "compat-log",
 987                 .data           = &compat_log,
 988                 .maxlen         = sizeof (int),
 989                 .mode           = 0644,
 990                 .proc_handler   = proc_dointvec,
 991         },
 992 #endif
 993 #ifdef CONFIG_RT_MUTEXES
 994         {
 995                 .procname       = "max_lock_depth",
 996                 .data           = &max_lock_depth,
 997                 .maxlen         = sizeof(int),
 998                 .mode           = 0644,
 999                 .proc_handler   = proc_dointvec,
1000         },
1001 #endif
1002         {
1003                 .procname       = "poweroff_cmd",
1004                 .data           = &poweroff_cmd,
1005                 .maxlen         = POWEROFF_CMD_PATH_LEN,
1006                 .mode           = 0644,
1007                 .proc_handler   = proc_dostring,
1008         },
1009 #ifdef CONFIG_KEYS
1010         {
1011                 .procname       = "keys",
1012                 .mode           = 0555,
1013                 .child          = key_sysctls,
1014         },
1015 #endif
1016 #ifdef CONFIG_RCU_TORTURE_TEST
1017         {
1018                 .procname       = "rcutorture_runnable",
1019                 .data           = &rcutorture_runnable,
1020                 .maxlen         = sizeof(int),
1021                 .mode           = 0644,
1022                 .proc_handler   = proc_dointvec,
1023         },
1024 #endif
1025 #ifdef CONFIG_PERF_EVENTS
1026         /*
1027          * User-space scripts rely on the existence of this file
1028          * as a feature check for perf_events being enabled.
1029          *
1030          * So it's an ABI, do not remove!
1031          */
1032         {
1033                 .procname       = "perf_event_paranoid",
1034                 .data           = &sysctl_perf_event_paranoid,
1035                 .maxlen         = sizeof(sysctl_perf_event_paranoid),
1036                 .mode           = 0644,
1037                 .proc_handler   = proc_dointvec,
1038         },
1039         {
1040                 .procname       = "perf_event_mlock_kb",
1041                 .data           = &sysctl_perf_event_mlock,
1042                 .maxlen         = sizeof(sysctl_perf_event_mlock),
1043                 .mode           = 0644,
1044                 .proc_handler   = proc_dointvec,
1045         },
1046         {
1047                 .procname       = "perf_event_max_sample_rate",
1048                 .data           = &sysctl_perf_event_sample_rate,
1049                 .maxlen         = sizeof(sysctl_perf_event_sample_rate),
1050                 .mode           = 0644,
1051                 .proc_handler   = perf_proc_update_handler,
1052                 .extra1         = &one,
1053         },
1054         {
1055                 .procname       = "perf_cpu_time_max_percent",
1056                 .data           = &sysctl_perf_cpu_time_max_percent,
1057                 .maxlen         = sizeof(sysctl_perf_cpu_time_max_percent),
1058                 .mode           = 0644,
1059                 .proc_handler   = perf_cpu_time_max_percent_handler,
1060                 .extra1         = &zero,
1061                 .extra2         = &one_hundred,
1062         },
1063 #endif
1064 #ifdef CONFIG_KMEMCHECK
1065         {
1066                 .procname       = "kmemcheck",
1067                 .data           = &kmemcheck_enabled,
1068                 .maxlen         = sizeof(int),
1069                 .mode           = 0644,
1070                 .proc_handler   = proc_dointvec,
1071         },
1072 #endif
1073 #ifdef CONFIG_BLOCK
1074         {
1075                 .procname       = "blk_iopoll",
1076                 .data           = &blk_iopoll_enabled,
1077                 .maxlen         = sizeof(int),
1078                 .mode           = 0644,
1079                 .proc_handler   = proc_dointvec,
1080         },
1081 #endif
1082         { }
1083 };
1084
1085 static struct ctl_table vm_table[] = {
1086         {
1087                 .procname       = "overcommit_memory",
1088                 .data           = &sysctl_overcommit_memory,
1089                 .maxlen         = sizeof(sysctl_overcommit_memory),
1090                 .mode           = 0644,
1091                 .proc_handler   = proc_dointvec_minmax,
1092                 .extra1         = &zero,
1093                 .extra2         = &two,
1094         },
1095         {
1096                 .procname       = "panic_on_oom",
1097                 .data           = &sysctl_panic_on_oom,
1098                 .maxlen         = sizeof(sysctl_panic_on_oom),
1099                 .mode           = 0644,
1100                 .proc_handler   = proc_dointvec_minmax,
1101                 .extra1         = &zero,
1102                 .extra2         = &two,
1103         },
1104         {
1105                 .procname       = "oom_kill_allocating_task",
1106                 .data           = &sysctl_oom_kill_allocating_task,
1107                 .maxlen         = sizeof(sysctl_oom_kill_allocating_task),
1108                 .mode           = 0644,
1109                 .proc_handler   = proc_dointvec,
1110         },
1111         {
1112                 .procname       = "oom_dump_tasks",
1113                 .data           = &sysctl_oom_dump_tasks,
1114                 .maxlen         = sizeof(sysctl_oom_dump_tasks),
1115                 .mode           = 0644,
1116                 .proc_handler   = proc_dointvec,
1117         },
1118         {
1119                 .procname       = "overcommit_ratio",
1120                 .data           = &sysctl_overcommit_ratio,
1121                 .maxlen         = sizeof(sysctl_overcommit_ratio),
1122                 .mode           = 0644,
1123                 .proc_handler   = proc_dointvec,
1124         },
1125         {
1126                 .procname       = "page-cluster",
1127                 .data           = &page_cluster,
1128                 .maxlen         = sizeof(int),
1129                 .mode           = 0644,
1130                 .proc_handler   = proc_dointvec_minmax,
1131                 .extra1         = &zero,
1132         },
1133         {
1134                 .procname       = "dirty_background_ratio",
1135                 .data           = &dirty_background_ratio,
1136                 .maxlen         = sizeof(dirty_background_ratio),
1137                 .mode           = 0644,
1138                 .proc_handler   = dirty_background_ratio_handler,
1139                 .extra1         = &zero,
1140                 .extra2         = &one_hundred,
1141         },
1142         {
1143                 .procname       = "dirty_background_bytes",
1144                 .data           = &dirty_background_bytes,
1145                 .maxlen         = sizeof(dirty_background_bytes),
1146                 .mode           = 0644,
1147                 .proc_handler   = dirty_background_bytes_handler,
1148                 .extra1         = &one_ul,
1149         },
1150         {
1151                 .procname       = "dirty_ratio",
1152                 .data           = &vm_dirty_ratio,
1153                 .maxlen         = sizeof(vm_dirty_ratio),
1154                 .mode           = 0644,
1155                 .proc_handler   = dirty_ratio_handler,
1156                 .extra1         = &zero,
1157                 .extra2         = &one_hundred,
1158         },
1159         {
1160                 .procname       = "dirty_bytes",
1161                 .data           = &vm_dirty_bytes,
1162                 .maxlen         = sizeof(vm_dirty_bytes),
1163                 .mode           = 0644,
1164                 .proc_handler   = dirty_bytes_handler,
1165                 .extra1         = &dirty_bytes_min,
1166         },
1167         {
1168                 .procname       = "dirty_writeback_centisecs",
1169                 .data           = &dirty_writeback_interval,
1170                 .maxlen         = sizeof(dirty_writeback_interval),
1171                 .mode           = 0644,
1172                 .proc_handler   = dirty_writeback_centisecs_handler,
1173         },
1174         {
1175                 .procname       = "dirty_expire_centisecs",
1176                 .data           = &dirty_expire_interval,
1177                 .maxlen         = sizeof(dirty_expire_interval),
1178                 .mode           = 0644,
1179                 .proc_handler   = proc_dointvec_minmax,
1180                 .extra1         = &zero,
1181         },
1182         {
1183                 .procname       = "nr_pdflush_threads",
1184                 .mode           = 0444 /* read-only */,
1185                 .proc_handler   = pdflush_proc_obsolete,
1186         },
1187         {
1188                 .procname       = "swappiness",
1189                 .data           = &vm_swappiness,
1190                 .maxlen         = sizeof(vm_swappiness),
1191                 .mode           = 0644,
1192                 .proc_handler   = proc_dointvec_minmax,
1193                 .extra1         = &zero,
1194                 .extra2         = &one_hundred,
1195         },
1196 #ifdef CONFIG_HUGETLB_PAGE
1197         {
1198                 .procname       = "nr_hugepages",
1199                 .data           = NULL,
1200                 .maxlen         = sizeof(unsigned long),
1201                 .mode           = 0644,
1202                 .proc_handler   = hugetlb_sysctl_handler,
1203                 .extra1         = (void *)&hugetlb_zero,
1204                 .extra2         = (void *)&hugetlb_infinity,
1205         },
1206 #ifdef CONFIG_NUMA
1207         {
1208                 .procname       = "nr_hugepages_mempolicy",
1209                 .data           = NULL,
1210                 .maxlen         = sizeof(unsigned long),
1211                 .mode           = 0644,
1212                 .proc_handler   = &hugetlb_mempolicy_sysctl_handler,
1213                 .extra1         = (void *)&hugetlb_zero,
1214                 .extra2         = (void *)&hugetlb_infinity,
1215         },
1216 #endif
1217          {
1218                 .procname       = "hugetlb_shm_group",
1219                 .data           = &sysctl_hugetlb_shm_group,
1220                 .maxlen         = sizeof(gid_t),
1221                 .mode           = 0644,
1222                 .proc_handler   = proc_dointvec,
1223          },
1224          {
1225                 .procname       = "hugepages_treat_as_movable",
1226                 .data           = &hugepages_treat_as_movable,
1227                 .maxlen         = sizeof(int),
1228                 .mode           = 0644,
1229                 .proc_handler   = proc_dointvec,
1230         },
1231         {
1232                 .procname       = "nr_overcommit_hugepages",
1233                 .data           = NULL,
1234                 .maxlen         = sizeof(unsigned long),
1235                 .mode           = 0644,
1236                 .proc_handler   = hugetlb_overcommit_handler,
1237                 .extra1         = (void *)&hugetlb_zero,
1238                 .extra2         = (void *)&hugetlb_infinity,
1239         },
1240 #endif
1241         {
1242                 .procname       = "lowmem_reserve_ratio",
1243                 .data           = &sysctl_lowmem_reserve_ratio,
1244                 .maxlen         = sizeof(sysctl_lowmem_reserve_ratio),
1245                 .mode           = 0644,
1246                 .proc_handler   = lowmem_reserve_ratio_sysctl_handler,
1247         },
1248         {
1249                 .procname       = "drop_caches",
1250                 .data           = &sysctl_drop_caches,
1251                 .maxlen         = sizeof(int),
1252                 .mode           = 0644,
1253                 .proc_handler   = drop_caches_sysctl_handler,
1254                 .extra1         = &one,
1255                 .extra2         = &three,
1256         },
1257 #ifdef CONFIG_COMPACTION
1258         {
1259                 .procname       = "compact_memory",
1260                 .data           = &sysctl_compact_memory,
1261                 .maxlen         = sizeof(int),
1262                 .mode           = 0200,
1263                 .proc_handler   = sysctl_compaction_handler,
1264         },
1265         {
1266                 .procname       = "extfrag_threshold",
1267                 .data           = &sysctl_extfrag_threshold,
1268                 .maxlen         = sizeof(int),
1269                 .mode           = 0644,
1270                 .proc_handler   = sysctl_extfrag_handler,
1271                 .extra1         = &min_extfrag_threshold,
1272                 .extra2         = &max_extfrag_threshold,
1273         },
1274
1275 #endif /* CONFIG_COMPACTION */
1276         {
1277                 .procname       = "min_free_kbytes",
1278                 .data           = &min_free_kbytes,
1279                 .maxlen         = sizeof(min_free_kbytes),
1280                 .mode           = 0644,
1281                 .proc_handler   = min_free_kbytes_sysctl_handler,
1282                 .extra1         = &zero,
1283         },
1284         {
1285                 .procname       = "percpu_pagelist_fraction",
1286                 .data           = &percpu_pagelist_fraction,
1287                 .maxlen         = sizeof(percpu_pagelist_fraction),
1288                 .mode           = 0644,
1289                 .proc_handler   = percpu_pagelist_fraction_sysctl_handler,
1290                 .extra1         = &min_percpu_pagelist_fract,
1291         },
1292 #ifdef CONFIG_MMU
1293         {
1294                 .procname       = "max_map_count",
1295                 .data           = &sysctl_max_map_count,
1296                 .maxlen         = sizeof(sysctl_max_map_count),
1297                 .mode           = 0644,
1298                 .proc_handler   = proc_dointvec_minmax,
1299                 .extra1         = &zero,
1300         },
1301 #else
1302         {
1303                 .procname       = "nr_trim_pages",
1304                 .data           = &sysctl_nr_trim_pages,
1305                 .maxlen         = sizeof(sysctl_nr_trim_pages),
1306                 .mode           = 0644,
1307                 .proc_handler   = proc_dointvec_minmax,
1308                 .extra1         = &zero,
1309         },
1310 #endif
1311         {
1312                 .procname       = "laptop_mode",
1313                 .data           = &laptop_mode,
1314                 .maxlen         = sizeof(laptop_mode),
1315                 .mode           = 0644,
1316                 .proc_handler   = proc_dointvec_jiffies,
1317         },
1318         {
1319                 .procname       = "block_dump",
1320                 .data           = &block_dump,
1321                 .maxlen         = sizeof(block_dump),
1322                 .mode           = 0644,
1323                 .proc_handler   = proc_dointvec,
1324                 .extra1         = &zero,
1325         },
1326         {
1327                 .procname       = "vfs_cache_pressure",
1328                 .data           = &sysctl_vfs_cache_pressure,
1329                 .maxlen         = sizeof(sysctl_vfs_cache_pressure),
1330                 .mode           = 0644,
1331                 .proc_handler   = proc_dointvec,
1332                 .extra1         = &zero,
1333         },
1334 #ifdef HAVE_ARCH_PICK_MMAP_LAYOUT
1335         {
1336                 .procname       = "legacy_va_layout",
1337                 .data           = &sysctl_legacy_va_layout,
1338                 .maxlen         = sizeof(sysctl_legacy_va_layout),
1339                 .mode           = 0644,
1340                 .proc_handler   = proc_dointvec,
1341                 .extra1         = &zero,
1342         },
1343 #endif
1344 #ifdef CONFIG_NUMA
1345         {
1346                 .procname       = "zone_reclaim_mode",
1347                 .data           = &zone_reclaim_mode,
1348                 .maxlen         = sizeof(zone_reclaim_mode),
1349                 .mode           = 0644,
1350                 .proc_handler   = proc_dointvec,
1351                 .extra1         = &zero,
1352         },
1353         {
1354                 .procname       = "min_unmapped_ratio",
1355                 .data           = &sysctl_min_unmapped_ratio,
1356                 .maxlen         = sizeof(sysctl_min_unmapped_ratio),
1357                 .mode           = 0644,
1358                 .proc_handler   = sysctl_min_unmapped_ratio_sysctl_handler,
1359                 .extra1         = &zero,
1360                 .extra2         = &one_hundred,
1361         },
1362         {
1363                 .procname       = "min_slab_ratio",
1364                 .data           = &sysctl_min_slab_ratio,
1365                 .maxlen         = sizeof(sysctl_min_slab_ratio),
1366                 .mode           = 0644,
1367                 .proc_handler   = sysctl_min_slab_ratio_sysctl_handler,
1368                 .extra1         = &zero,
1369                 .extra2         = &one_hundred,
1370         },
1371 #endif
1372 #ifdef CONFIG_SMP
1373         {
1374                 .procname       = "stat_interval",
1375                 .data           = &sysctl_stat_interval,
1376                 .maxlen         = sizeof(sysctl_stat_interval),
1377                 .mode           = 0644,
1378                 .proc_handler   = proc_dointvec_jiffies,
1379         },
1380 #endif
1381 #ifdef CONFIG_MMU
1382         {
1383                 .procname       = "mmap_min_addr",
1384                 .data           = &dac_mmap_min_addr,
1385                 .maxlen         = sizeof(unsigned long),
1386                 .mode           = 0644,
1387                 .proc_handler   = mmap_min_addr_handler,
1388         },
1389 #endif
1390 #ifdef CONFIG_NUMA
1391         {
1392                 .procname       = "numa_zonelist_order",
1393                 .data           = &numa_zonelist_order,
1394                 .maxlen         = NUMA_ZONELIST_ORDER_LEN,
1395                 .mode           = 0644,
1396                 .proc_handler   = numa_zonelist_order_handler,
1397         },
1398 #endif
1399 #if (defined(CONFIG_X86_32) && !defined(CONFIG_UML))|| \
1400    (defined(CONFIG_SUPERH) && defined(CONFIG_VSYSCALL))
1401         {
1402                 .procname       = "vdso_enabled",
1403                 .data           = &vdso_enabled,
1404                 .maxlen         = sizeof(vdso_enabled),
1405                 .mode           = 0644,
1406                 .proc_handler   = proc_dointvec,
1407                 .extra1         = &zero,
1408         },
1409 #endif
1410 #ifdef CONFIG_HIGHMEM
1411         {
1412                 .procname       = "highmem_is_dirtyable",
1413                 .data           = &vm_highmem_is_dirtyable,
1414                 .maxlen         = sizeof(vm_highmem_is_dirtyable),
1415                 .mode           = 0644,
1416                 .proc_handler   = proc_dointvec_minmax,
1417                 .extra1         = &zero,
1418                 .extra2         = &one,
1419         },
1420 #endif
1421         {
1422                 .procname       = "scan_unevictable_pages",
1423                 .data           = &scan_unevictable_pages,
1424                 .maxlen         = sizeof(scan_unevictable_pages),
1425                 .mode           = 0644,
1426                 .proc_handler   = scan_unevictable_handler,
1427         },
1428 #ifdef CONFIG_MEMORY_FAILURE
1429         {
1430                 .procname       = "memory_failure_early_kill",
1431                 .data           = &sysctl_memory_failure_early_kill,
1432                 .maxlen         = sizeof(sysctl_memory_failure_early_kill),
1433                 .mode           = 0644,
1434                 .proc_handler   = proc_dointvec_minmax,
1435                 .extra1         = &zero,
1436                 .extra2         = &one,
1437         },
1438         {
1439                 .procname       = "memory_failure_recovery",
1440                 .data           = &sysctl_memory_failure_recovery,
1441                 .maxlen         = sizeof(sysctl_memory_failure_recovery),
1442                 .mode           = 0644,
1443                 .proc_handler   = proc_dointvec_minmax,
1444                 .extra1         = &zero,
1445                 .extra2         = &one,
1446         },
1447 #endif
1448         {
1449                 .procname       = "user_reserve_kbytes",
1450                 .data           = &sysctl_user_reserve_kbytes,
1451                 .maxlen         = sizeof(sysctl_user_reserve_kbytes),
1452                 .mode           = 0644,
1453                 .proc_handler   = proc_doulongvec_minmax,
1454         },
1455         {
1456                 .procname       = "admin_reserve_kbytes",
1457                 .data           = &sysctl_admin_reserve_kbytes,
1458                 .maxlen         = sizeof(sysctl_admin_reserve_kbytes),
1459                 .mode           = 0644,
1460                 .proc_handler   = proc_doulongvec_minmax,
1461         },
1462         { }
1463 };
1464
1465 #if defined(CONFIG_BINFMT_MISC) || defined(CONFIG_BINFMT_MISC_MODULE)
1466 static struct ctl_table binfmt_misc_table[] = {
1467         { }
1468 };
1469 #endif
1470
1471 static struct ctl_table fs_table[] = {
1472         {
1473                 .procname       = "inode-nr",
1474                 .data           = &inodes_stat,
1475                 .maxlen         = 2*sizeof(long),
1476                 .mode           = 0444,
1477                 .proc_handler   = proc_nr_inodes,
1478         },
1479         {
1480                 .procname       = "inode-state",
1481                 .data           = &inodes_stat,
1482                 .maxlen         = 7*sizeof(long),
1483                 .mode           = 0444,
1484                 .proc_handler   = proc_nr_inodes,
1485         },
1486         {
1487                 .procname       = "file-nr",
1488                 .data           = &files_stat,
1489                 .maxlen         = sizeof(files_stat),
1490                 .mode           = 0444,
1491                 .proc_handler   = proc_nr_files,
1492         },
1493         {
1494                 .procname       = "file-max",
1495                 .data           = &files_stat.max_files,
1496                 .maxlen         = sizeof(files_stat.max_files),
1497                 .mode           = 0644,
1498                 .proc_handler   = proc_doulongvec_minmax,
1499         },
1500         {
1501                 .procname       = "nr_open",
1502                 .data           = &sysctl_nr_open,
1503                 .maxlen         = sizeof(int),
1504                 .mode           = 0644,
1505                 .proc_handler   = proc_dointvec_minmax,
1506                 .extra1         = &sysctl_nr_open_min,
1507                 .extra2         = &sysctl_nr_open_max,
1508         },
1509         {
1510                 .procname       = "dentry-state",
1511                 .data           = &dentry_stat,
1512                 .maxlen         = 6*sizeof(long),
1513                 .mode           = 0444,
1514                 .proc_handler   = proc_nr_dentry,
1515         },
1516         {
1517                 .procname       = "overflowuid",
1518                 .data           = &fs_overflowuid,
1519                 .maxlen         = sizeof(int),
1520                 .mode           = 0644,
1521                 .proc_handler   = proc_dointvec_minmax,
1522                 .extra1         = &minolduid,
1523                 .extra2         = &maxolduid,
1524         },
1525         {
1526                 .procname       = "overflowgid",
1527                 .data           = &fs_overflowgid,
1528                 .maxlen         = sizeof(int),
1529                 .mode           = 0644,
1530                 .proc_handler   = proc_dointvec_minmax,
1531                 .extra1         = &minolduid,
1532                 .extra2         = &maxolduid,
1533         },
1534 #ifdef CONFIG_FILE_LOCKING
1535         {
1536                 .procname       = "leases-enable",
1537                 .data           = &leases_enable,
1538                 .maxlen         = sizeof(int),
1539                 .mode           = 0644,
1540                 .proc_handler   = proc_dointvec,
1541         },
1542 #endif
1543 #ifdef CONFIG_DNOTIFY
1544         {
1545                 .procname       = "dir-notify-enable",
1546                 .data           = &dir_notify_enable,
1547                 .maxlen         = sizeof(int),
1548                 .mode           = 0644,
1549                 .proc_handler   = proc_dointvec,
1550         },
1551 #endif
1552 #ifdef CONFIG_MMU
1553 #ifdef CONFIG_FILE_LOCKING
1554         {
1555                 .procname       = "lease-break-time",
1556                 .data           = &lease_break_time,
1557                 .maxlen         = sizeof(int),
1558                 .mode           = 0644,
1559                 .proc_handler   = proc_dointvec,
1560         },
1561 #endif
1562 #ifdef CONFIG_AIO
1563         {
1564                 .procname       = "aio-nr",
1565                 .data           = &aio_nr,
1566                 .maxlen         = sizeof(aio_nr),
1567                 .mode           = 0444,
1568                 .proc_handler   = proc_doulongvec_minmax,
1569         },
1570         {
1571                 .procname       = "aio-max-nr",
1572                 .data           = &aio_max_nr,
1573                 .maxlen         = sizeof(aio_max_nr),
1574                 .mode           = 0644,
1575                 .proc_handler   = proc_doulongvec_minmax,
1576         },
1577 #endif /* CONFIG_AIO */
1578 #ifdef CONFIG_INOTIFY_USER
1579         {
1580                 .procname       = "inotify",
1581                 .mode           = 0555,
1582                 .child          = inotify_table,
1583         },
1584 #endif
1585 #ifdef CONFIG_EPOLL
1586         {
1587                 .procname       = "epoll",
1588                 .mode           = 0555,
1589                 .child          = epoll_table,
1590         },
1591 #endif
1592 #endif
1593         {
1594                 .procname       = "protected_symlinks",
1595                 .data           = &sysctl_protected_symlinks,
1596                 .maxlen         = sizeof(int),
1597                 .mode           = 0600,
1598                 .proc_handler   = proc_dointvec_minmax,
1599                 .extra1         = &zero,
1600                 .extra2         = &one,
1601         },
1602         {
1603                 .procname       = "protected_hardlinks",
1604                 .data           = &sysctl_protected_hardlinks,
1605                 .maxlen         = sizeof(int),
1606                 .mode           = 0600,
1607                 .proc_handler   = proc_dointvec_minmax,
1608                 .extra1         = &zero,
1609                 .extra2         = &one,
1610         },
1611         {
1612                 .procname       = "suid_dumpable",
1613                 .data           = &suid_dumpable,
1614                 .maxlen         = sizeof(int),
1615                 .mode           = 0644,
1616                 .proc_handler   = proc_dointvec_minmax_coredump,
1617                 .extra1         = &zero,
1618                 .extra2         = &two,
1619         },
1620 #if defined(CONFIG_BINFMT_MISC) || defined(CONFIG_BINFMT_MISC_MODULE)
1621         {
1622                 .procname       = "binfmt_misc",
1623                 .mode           = 0555,
1624                 .child          = binfmt_misc_table,
1625         },
1626 #endif
1627         {
1628                 .procname       = "pipe-max-size",
1629                 .data           = &pipe_max_size,
1630                 .maxlen         = sizeof(int),
1631                 .mode           = 0644,
1632                 .proc_handler   = &pipe_proc_fn,
1633                 .extra1         = &pipe_min_size,
1634         },
1635         { }
1636 };
1637
1638 static struct ctl_table debug_table[] = {
1639 #ifdef CONFIG_SYSCTL_EXCEPTION_TRACE
1640         {
1641                 .procname       = "exception-trace",
1642                 .data           = &show_unhandled_signals,
1643                 .maxlen         = sizeof(int),
1644                 .mode           = 0644,
1645                 .proc_handler   = proc_dointvec
1646         },
1647 #endif
1648 #if defined(CONFIG_OPTPROBES)
1649         {
1650                 .procname       = "kprobes-optimization",
1651                 .data           = &sysctl_kprobes_optimization,
1652                 .maxlen         = sizeof(int),
1653                 .mode           = 0644,
1654                 .proc_handler   = proc_kprobes_optimization_handler,
1655                 .extra1         = &zero,
1656                 .extra2         = &one,
1657         },
1658 #endif
1659         { }
1660 };
1661
1662 static struct ctl_table dev_table[] = {
1663         { }
1664 };
1665
1666 int __init sysctl_init(void)
1667 {
1668         struct ctl_table_header *hdr;
1669
1670         hdr = register_sysctl_table(sysctl_base_table);
1671         kmemleak_not_leak(hdr);
1672         return 0;
1673 }
1674
1675 #endif /* CONFIG_SYSCTL */
1676
1677 /*
1678  * /proc/sys support
1679  */
1680
1681 #ifdef CONFIG_PROC_SYSCTL
1682
1683 static int _proc_do_string(void* data, int maxlen, int write,
1684                            void __user *buffer,
1685                            size_t *lenp, loff_t *ppos)
1686 {
1687         size_t len;
1688         char __user *p;
1689         char c;
1690
1691         if (!data || !maxlen || !*lenp) {
1692                 *lenp = 0;
1693                 return 0;
1694         }
1695
1696         if (write) {
1697                 len = 0;
1698                 p = buffer;
1699                 while (len < *lenp) {
1700                         if (get_user(c, p++))
1701                                 return -EFAULT;
1702                         if (c == 0 || c == '\n')
1703                                 break;
1704                         len++;
1705                 }
1706                 if (len >= maxlen)
1707                         len = maxlen-1;
1708                 if(copy_from_user(data, buffer, len))
1709                         return -EFAULT;
1710                 ((char *) data)[len] = 0;
1711                 *ppos += *lenp;
1712         } else {
1713                 len = strlen(data);
1714                 if (len > maxlen)
1715                         len = maxlen;
1716
1717                 if (*ppos > len) {
1718                         *lenp = 0;
1719                         return 0;
1720                 }
1721
1722                 data += *ppos;
1723                 len  -= *ppos;
1724
1725                 if (len > *lenp)
1726                         len = *lenp;
1727                 if (len)
1728                         if(copy_to_user(buffer, data, len))
1729                                 return -EFAULT;
1730                 if (len < *lenp) {
1731                         if(put_user('\n', ((char __user *) buffer) + len))
1732                                 return -EFAULT;
1733                         len++;
1734                 }
1735                 *lenp = len;
1736                 *ppos += len;
1737         }
1738         return 0;
1739 }
1740
1741 /**
1742  * proc_dostring - read a string sysctl
1743  * @table: the sysctl table
1744  * @write: %TRUE if this is a write to the sysctl file
1745  * @buffer: the user buffer
1746  * @lenp: the size of the user buffer
1747  * @ppos: file position
1748  *
1749  * Reads/writes a string from/to the user buffer. If the kernel
1750  * buffer provided is not large enough to hold the string, the
1751  * string is truncated. The copied string is %NULL-terminated.
1752  * If the string is being read by the user process, it is copied
1753  * and a newline '\n' is added. It is truncated if the buffer is
1754  * not large enough.
1755  *
1756  * Returns 0 on success.
1757  */
1758 int proc_dostring(struct ctl_table *table, int write,
1759                   void __user *buffer, size_t *lenp, loff_t *ppos)
1760 {
1761         return _proc_do_string(table->data, table->maxlen, write,
1762                                buffer, lenp, ppos);
1763 }
1764
1765 static size_t proc_skip_spaces(char **buf)
1766 {
1767         size_t ret;
1768         char *tmp = skip_spaces(*buf);
1769         ret = tmp - *buf;
1770         *buf = tmp;
1771         return ret;
1772 }
1773
1774 static void proc_skip_char(char **buf, size_t *size, const char v)
1775 {
1776         while (*size) {
1777                 if (**buf != v)
1778                         break;
1779                 (*size)--;
1780                 (*buf)++;
1781         }
1782 }
1783
1784 #define TMPBUFLEN 22
1785 /**
1786  * proc_get_long - reads an ASCII formatted integer from a user buffer
1787  *
1788  * @buf: a kernel buffer
1789  * @size: size of the kernel buffer
1790  * @val: this is where the number will be stored
1791  * @neg: set to %TRUE if number is negative
1792  * @perm_tr: a vector which contains the allowed trailers
1793  * @perm_tr_len: size of the perm_tr vector
1794  * @tr: pointer to store the trailer character
1795  *
1796  * In case of success %0 is returned and @buf and @size are updated with
1797  * the amount of bytes read. If @tr is non-NULL and a trailing
1798  * character exists (size is non-zero after returning from this
1799  * function), @tr is updated with the trailing character.
1800  */
1801 static int proc_get_long(char **buf, size_t *size,
1802                           unsigned long *val, bool *neg,
1803                           const char *perm_tr, unsigned perm_tr_len, char *tr)
1804 {
1805         int len;
1806         char *p, tmp[TMPBUFLEN];
1807
1808         if (!*size)
1809                 return -EINVAL;
1810
1811         len = *size;
1812         if (len > TMPBUFLEN - 1)
1813                 len = TMPBUFLEN - 1;
1814
1815         memcpy(tmp, *buf, len);
1816
1817         tmp[len] = 0;
1818         p = tmp;
1819         if (*p == '-' && *size > 1) {
1820                 *neg = true;
1821                 p++;
1822         } else
1823                 *neg = false;
1824         if (!isdigit(*p))
1825                 return -EINVAL;
1826
1827         *val = simple_strtoul(p, &p, 0);
1828
1829         len = p - tmp;
1830
1831         /* We don't know if the next char is whitespace thus we may accept
1832          * invalid integers (e.g. 1234...a) or two integers instead of one
1833          * (e.g. 123...1). So lets not allow such large numbers. */
1834         if (len == TMPBUFLEN - 1)
1835                 return -EINVAL;
1836
1837         if (len < *size && perm_tr_len && !memchr(perm_tr, *p, perm_tr_len))
1838                 return -EINVAL;
1839
1840         if (tr && (len < *size))
1841                 *tr = *p;
1842
1843         *buf += len;
1844         *size -= len;
1845
1846         return 0;
1847 }
1848
1849 /**
1850  * proc_put_long - converts an integer to a decimal ASCII formatted string
1851  *
1852  * @buf: the user buffer
1853  * @size: the size of the user buffer
1854  * @val: the integer to be converted
1855  * @neg: sign of the number, %TRUE for negative
1856  *
1857  * In case of success %0 is returned and @buf and @size are updated with
1858  * the amount of bytes written.
1859  */
1860 static int proc_put_long(void __user **buf, size_t *size, unsigned long val,
1861                           bool neg)
1862 {
1863         int len;
1864         char tmp[TMPBUFLEN], *p = tmp;
1865
1866         sprintf(p, "%s%lu", neg ? "-" : "", val);
1867         len = strlen(tmp);
1868         if (len > *size)
1869                 len = *size;
1870         if (copy_to_user(*buf, tmp, len))
1871                 return -EFAULT;
1872         *size -= len;
1873         *buf += len;
1874         return 0;
1875 }
1876 #undef TMPBUFLEN
1877
1878 static int proc_put_char(void __user **buf, size_t *size, char c)
1879 {
1880         if (*size) {
1881                 char __user **buffer = (char __user **)buf;
1882                 if (put_user(c, *buffer))
1883                         return -EFAULT;
1884                 (*size)--, (*buffer)++;
1885                 *buf = *buffer;
1886         }
1887         return 0;
1888 }
1889
1890 static int do_proc_dointvec_conv(bool *negp, unsigned long *lvalp,
1891                                  int *valp,
1892                                  int write, void *data)
1893 {
1894         if (write) {
1895                 *valp = *negp ? -*lvalp : *lvalp;
1896         } else {
1897                 int val = *valp;
1898                 if (val < 0) {
1899                         *negp = true;
1900                         *lvalp = (unsigned long)-val;
1901                 } else {
1902                         *negp = false;
1903                         *lvalp = (unsigned long)val;
1904                 }
1905         }
1906         return 0;
1907 }
1908
1909 static const char proc_wspace_sep[] = { ' ', '\t', '\n' };
1910
1911 static int __do_proc_dointvec(void *tbl_data, struct ctl_table *table,
1912                   int write, void __user *buffer,
1913                   size_t *lenp, loff_t *ppos,
1914                   int (*conv)(bool *negp, unsigned long *lvalp, int *valp,
1915                               int write, void *data),
1916                   void *data)
1917 {
1918         int *i, vleft, first = 1, err = 0;
1919         unsigned long page = 0;
1920         size_t left;
1921         char *kbuf;
1922
1923         if (!tbl_data || !table->maxlen || !*lenp || (*ppos && !write)) {
1924                 *lenp = 0;
1925                 return 0;
1926         }
1927
1928         i = (int *) tbl_data;
1929         vleft = table->maxlen / sizeof(*i);
1930         left = *lenp;
1931
1932         if (!conv)
1933                 conv = do_proc_dointvec_conv;
1934
1935         if (write) {
1936                 if (left > PAGE_SIZE - 1)
1937                         left = PAGE_SIZE - 1;
1938                 page = __get_free_page(GFP_TEMPORARY);
1939                 kbuf = (char *) page;
1940                 if (!kbuf)
1941                         return -ENOMEM;
1942                 if (copy_from_user(kbuf, buffer, left)) {
1943                         err = -EFAULT;
1944                         goto free;
1945                 }
1946                 kbuf[left] = 0;
1947         }
1948
1949         for (; left && vleft--; i++, first=0) {
1950                 unsigned long lval;
1951                 bool neg;
1952
1953                 if (write) {
1954                         left -= proc_skip_spaces(&kbuf);
1955
1956                         if (!left)
1957                                 break;
1958                         err = proc_get_long(&kbuf, &left, &lval, &neg,
1959                                              proc_wspace_sep,
1960                                              sizeof(proc_wspace_sep), NULL);
1961                         if (err)
1962                                 break;
1963                         if (conv(&neg, &lval, i, 1, data)) {
1964                                 err = -EINVAL;
1965                                 break;
1966                         }
1967                 } else {
1968                         if (conv(&neg, &lval, i, 0, data)) {
1969                                 err = -EINVAL;
1970                                 break;
1971                         }
1972                         if (!first)
1973                                 err = proc_put_char(&buffer, &left, '\t');
1974                         if (err)
1975                                 break;
1976                         err = proc_put_long(&buffer, &left, lval, neg);
1977                         if (err)
1978                                 break;
1979                 }
1980         }
1981
1982         if (!write && !first && left && !err)
1983                 err = proc_put_char(&buffer, &left, '\n');
1984         if (write && !err && left)
1985                 left -= proc_skip_spaces(&kbuf);
1986 free:
1987         if (write) {
1988                 free_page(page);
1989                 if (first)
1990                         return err ? : -EINVAL;
1991         }
1992         *lenp -= left;
1993         *ppos += *lenp;
1994         return err;
1995 }
1996
1997 static int do_proc_dointvec(struct ctl_table *table, int write,
1998                   void __user *buffer, size_t *lenp, loff_t *ppos,
1999                   int (*conv)(bool *negp, unsigned long *lvalp, int *valp,
2000                               int write, void *data),
2001                   void *data)
2002 {
2003         return __do_proc_dointvec(table->data, table, write,
2004                         buffer, lenp, ppos, conv, data);
2005 }
2006
2007 /**
2008  * proc_dointvec - read a vector of integers
2009  * @table: the sysctl table
2010  * @write: %TRUE if this is a write to the sysctl file
2011  * @buffer: the user buffer
2012  * @lenp: the size of the user buffer
2013  * @ppos: file position
2014  *
2015  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2016  * values from/to the user buffer, treated as an ASCII string.
2017  *
2018  * Returns 0 on success.
2019  */
2020 int proc_dointvec(struct ctl_table *table, int write,
2021                      void __user *buffer, size_t *lenp, loff_t *ppos)
2022 {
2023     return do_proc_dointvec(table,write,buffer,lenp,ppos,
2024                             NULL,NULL);
2025 }
2026
2027 /*
2028  * Taint values can only be increased
2029  * This means we can safely use a temporary.
2030  */
2031 static int proc_taint(struct ctl_table *table, int write,
2032                                void __user *buffer, size_t *lenp, loff_t *ppos)
2033 {
2034         struct ctl_table t;
2035         unsigned long tmptaint = get_taint();
2036         int err;
2037
2038         if (write && !capable(CAP_SYS_ADMIN))
2039                 return -EPERM;
2040
2041         t = *table;
2042         t.data = &tmptaint;
2043         err = proc_doulongvec_minmax(&t, write, buffer, lenp, ppos);
2044         if (err < 0)
2045                 return err;
2046
2047         if (write) {
2048                 /*
2049                  * Poor man's atomic or. Not worth adding a primitive
2050                  * to everyone's atomic.h for this
2051                  */
2052                 int i;
2053                 for (i = 0; i < BITS_PER_LONG && tmptaint >> i; i++) {
2054                         if ((tmptaint >> i) & 1)
2055                                 add_taint(i, LOCKDEP_STILL_OK);
2056                 }
2057         }
2058
2059         return err;
2060 }
2061
2062 #ifdef CONFIG_PRINTK
2063 static int proc_dointvec_minmax_sysadmin(struct ctl_table *table, int write,
2064                                 void __user *buffer, size_t *lenp, loff_t *ppos)
2065 {
2066         if (write && !capable(CAP_SYS_ADMIN))
2067                 return -EPERM;
2068
2069         return proc_dointvec_minmax(table, write, buffer, lenp, ppos);
2070 }
2071 #endif
2072
2073 struct do_proc_dointvec_minmax_conv_param {
2074         int *min;
2075         int *max;
2076 };
2077
2078 static int do_proc_dointvec_minmax_conv(bool *negp, unsigned long *lvalp,
2079                                         int *valp,
2080                                         int write, void *data)
2081 {
2082         struct do_proc_dointvec_minmax_conv_param *param = data;
2083         if (write) {
2084                 int val = *negp ? -*lvalp : *lvalp;
2085                 if ((param->min && *param->min > val) ||
2086                     (param->max && *param->max < val))
2087                         return -EINVAL;
2088                 *valp = val;
2089         } else {
2090                 int val = *valp;
2091                 if (val < 0) {
2092                         *negp = true;
2093                         *lvalp = (unsigned long)-val;
2094                 } else {
2095                         *negp = false;
2096                         *lvalp = (unsigned long)val;
2097                 }
2098         }
2099         return 0;
2100 }
2101
2102 /**
2103  * proc_dointvec_minmax - read a vector of integers with min/max values
2104  * @table: the sysctl table
2105  * @write: %TRUE if this is a write to the sysctl file
2106  * @buffer: the user buffer
2107  * @lenp: the size of the user buffer
2108  * @ppos: file position
2109  *
2110  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2111  * values from/to the user buffer, treated as an ASCII string.
2112  *
2113  * This routine will ensure the values are within the range specified by
2114  * table->extra1 (min) and table->extra2 (max).
2115  *
2116  * Returns 0 on success.
2117  */
2118 int proc_dointvec_minmax(struct ctl_table *table, int write,
2119                   void __user *buffer, size_t *lenp, loff_t *ppos)
2120 {
2121         struct do_proc_dointvec_minmax_conv_param param = {
2122                 .min = (int *) table->extra1,
2123                 .max = (int *) table->extra2,
2124         };
2125         return do_proc_dointvec(table, write, buffer, lenp, ppos,
2126                                 do_proc_dointvec_minmax_conv, &param);
2127 }
2128
2129 static void validate_coredump_safety(void)
2130 {
2131 #ifdef CONFIG_COREDUMP
2132         if (suid_dumpable == SUID_DUMP_ROOT &&
2133             core_pattern[0] != '/' && core_pattern[0] != '|') {
2134                 printk(KERN_WARNING "Unsafe core_pattern used with "\
2135                         "suid_dumpable=2. Pipe handler or fully qualified "\
2136                         "core dump path required.\n");
2137         }
2138 #endif
2139 }
2140
2141 static int proc_dointvec_minmax_coredump(struct ctl_table *table, int write,
2142                 void __user *buffer, size_t *lenp, loff_t *ppos)
2143 {
2144         int error = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
2145         if (!error)
2146                 validate_coredump_safety();
2147         return error;
2148 }
2149
2150 #ifdef CONFIG_COREDUMP
2151 static int proc_dostring_coredump(struct ctl_table *table, int write,
2152                   void __user *buffer, size_t *lenp, loff_t *ppos)
2153 {
2154         int error = proc_dostring(table, write, buffer, lenp, ppos);
2155         if (!error)
2156                 validate_coredump_safety();
2157         return error;
2158 }
2159 #endif
2160
2161 static int __do_proc_doulongvec_minmax(void *data, struct ctl_table *table, int write,
2162                                      void __user *buffer,
2163                                      size_t *lenp, loff_t *ppos,
2164                                      unsigned long convmul,
2165                                      unsigned long convdiv)
2166 {
2167         unsigned long *i, *min, *max;
2168         int vleft, first = 1, err = 0;
2169         unsigned long page = 0;
2170         size_t left;
2171         char *kbuf;
2172
2173         if (!data || !table->maxlen || !*lenp || (*ppos && !write)) {
2174                 *lenp = 0;
2175                 return 0;
2176         }
2177
2178         i = (unsigned long *) data;
2179         min = (unsigned long *) table->extra1;
2180         max = (unsigned long *) table->extra2;
2181         vleft = table->maxlen / sizeof(unsigned long);
2182         left = *lenp;
2183
2184         if (write) {
2185                 if (left > PAGE_SIZE - 1)
2186                         left = PAGE_SIZE - 1;
2187                 page = __get_free_page(GFP_TEMPORARY);
2188                 kbuf = (char *) page;
2189                 if (!kbuf)
2190                         return -ENOMEM;
2191                 if (copy_from_user(kbuf, buffer, left)) {
2192                         err = -EFAULT;
2193                         goto free;
2194                 }
2195                 kbuf[left] = 0;
2196         }
2197
2198         for (; left && vleft--; i++, first = 0) {
2199                 unsigned long val;
2200
2201                 if (write) {
2202                         bool neg;
2203
2204                         left -= proc_skip_spaces(&kbuf);
2205
2206                         err = proc_get_long(&kbuf, &left, &val, &neg,
2207                                              proc_wspace_sep,
2208                                              sizeof(proc_wspace_sep), NULL);
2209                         if (err)
2210                                 break;
2211                         if (neg)
2212                                 continue;
2213                         if ((min && val < *min) || (max && val > *max))
2214                                 continue;
2215                         *i = val;
2216                 } else {
2217                         val = convdiv * (*i) / convmul;
2218                         if (!first)
2219                                 err = proc_put_char(&buffer, &left, '\t');
2220                         err = proc_put_long(&buffer, &left, val, false);
2221                         if (err)
2222                                 break;
2223                 }
2224         }
2225
2226         if (!write && !first && left && !err)
2227                 err = proc_put_char(&buffer, &left, '\n');
2228         if (write && !err)
2229                 left -= proc_skip_spaces(&kbuf);
2230 free:
2231         if (write) {
2232                 free_page(page);
2233                 if (first)
2234                         return err ? : -EINVAL;
2235         }
2236         *lenp -= left;
2237         *ppos += *lenp;
2238         return err;
2239 }
2240
2241 static int do_proc_doulongvec_minmax(struct ctl_table *table, int write,
2242                                      void __user *buffer,
2243                                      size_t *lenp, loff_t *ppos,
2244                                      unsigned long convmul,
2245                                      unsigned long convdiv)
2246 {
2247         return __do_proc_doulongvec_minmax(table->data, table, write,
2248                         buffer, lenp, ppos, convmul, convdiv);
2249 }
2250
2251 /**
2252  * proc_doulongvec_minmax - read a vector of long integers with min/max values
2253  * @table: the sysctl table
2254  * @write: %TRUE if this is a write to the sysctl file
2255  * @buffer: the user buffer
2256  * @lenp: the size of the user buffer
2257  * @ppos: file position
2258  *
2259  * Reads/writes up to table->maxlen/sizeof(unsigned long) unsigned long
2260  * values from/to the user buffer, treated as an ASCII string.
2261  *
2262  * This routine will ensure the values are within the range specified by
2263  * table->extra1 (min) and table->extra2 (max).
2264  *
2265  * Returns 0 on success.
2266  */
2267 int proc_doulongvec_minmax(struct ctl_table *table, int write,
2268                            void __user *buffer, size_t *lenp, loff_t *ppos)
2269 {
2270     return do_proc_doulongvec_minmax(table, write, buffer, lenp, ppos, 1l, 1l);
2271 }
2272
2273 /**
2274  * proc_doulongvec_ms_jiffies_minmax - read a vector of millisecond values with min/max values
2275  * @table: the sysctl table
2276  * @write: %TRUE if this is a write to the sysctl file
2277  * @buffer: the user buffer
2278  * @lenp: the size of the user buffer
2279  * @ppos: file position
2280  *
2281  * Reads/writes up to table->maxlen/sizeof(unsigned long) unsigned long
2282  * values from/to the user buffer, treated as an ASCII string. The values
2283  * are treated as milliseconds, and converted to jiffies when they are stored.
2284  *
2285  * This routine will ensure the values are within the range specified by
2286  * table->extra1 (min) and table->extra2 (max).
2287  *
2288  * Returns 0 on success.
2289  */
2290 int proc_doulongvec_ms_jiffies_minmax(struct ctl_table *table, int write,
2291                                       void __user *buffer,
2292                                       size_t *lenp, loff_t *ppos)
2293 {
2294     return do_proc_doulongvec_minmax(table, write, buffer,
2295                                      lenp, ppos, HZ, 1000l);
2296 }
2297
2298
2299 static int do_proc_dointvec_jiffies_conv(bool *negp, unsigned long *lvalp,
2300                                          int *valp,
2301                                          int write, void *data)
2302 {
2303         if (write) {
2304                 if (*lvalp > LONG_MAX / HZ)
2305                         return 1;
2306                 *valp = *negp ? -(*lvalp*HZ) : (*lvalp*HZ);
2307         } else {
2308                 int val = *valp;
2309                 unsigned long lval;
2310                 if (val < 0) {
2311                         *negp = true;
2312                         lval = (unsigned long)-val;
2313                 } else {
2314                         *negp = false;
2315                         lval = (unsigned long)val;
2316                 }
2317                 *lvalp = lval / HZ;
2318         }
2319         return 0;
2320 }
2321
2322 static int do_proc_dointvec_userhz_jiffies_conv(bool *negp, unsigned long *lvalp,
2323                                                 int *valp,
2324                                                 int write, void *data)
2325 {
2326         if (write) {
2327                 if (USER_HZ < HZ && *lvalp > (LONG_MAX / HZ) * USER_HZ)
2328                         return 1;
2329                 *valp = clock_t_to_jiffies(*negp ? -*lvalp : *lvalp);
2330         } else {
2331                 int val = *valp;
2332                 unsigned long lval;
2333                 if (val < 0) {
2334                         *negp = true;
2335                         lval = (unsigned long)-val;
2336                 } else {
2337                         *negp = false;
2338                         lval = (unsigned long)val;
2339                 }
2340                 *lvalp = jiffies_to_clock_t(lval);
2341         }
2342         return 0;
2343 }
2344
2345 static int do_proc_dointvec_ms_jiffies_conv(bool *negp, unsigned long *lvalp,
2346                                             int *valp,
2347                                             int write, void *data)
2348 {
2349         if (write) {
2350                 unsigned long jif = msecs_to_jiffies(*negp ? -*lvalp : *lvalp);
2351
2352                 if (jif > INT_MAX)
2353                         return 1;
2354                 *valp = (int)jif;
2355         } else {
2356                 int val = *valp;
2357                 unsigned long lval;
2358                 if (val < 0) {
2359                         *negp = true;
2360                         lval = (unsigned long)-val;
2361                 } else {
2362                         *negp = false;
2363                         lval = (unsigned long)val;
2364                 }
2365                 *lvalp = jiffies_to_msecs(lval);
2366         }
2367         return 0;
2368 }
2369
2370 /**
2371  * proc_dointvec_jiffies - read a vector of integers as seconds
2372  * @table: the sysctl table
2373  * @write: %TRUE if this is a write to the sysctl file
2374  * @buffer: the user buffer
2375  * @lenp: the size of the user buffer
2376  * @ppos: file position
2377  *
2378  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2379  * values from/to the user buffer, treated as an ASCII string.
2380  * The values read are assumed to be in seconds, and are converted into
2381  * jiffies.
2382  *
2383  * Returns 0 on success.
2384  */
2385 int proc_dointvec_jiffies(struct ctl_table *table, int write,
2386                           void __user *buffer, size_t *lenp, loff_t *ppos)
2387 {
2388     return do_proc_dointvec(table,write,buffer,lenp,ppos,
2389                             do_proc_dointvec_jiffies_conv,NULL);
2390 }
2391
2392 /**
2393  * proc_dointvec_userhz_jiffies - read a vector of integers as 1/USER_HZ seconds
2394  * @table: the sysctl table
2395  * @write: %TRUE if this is a write to the sysctl file
2396  * @buffer: the user buffer
2397  * @lenp: the size of the user buffer
2398  * @ppos: pointer to the file position
2399  *
2400  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2401  * values from/to the user buffer, treated as an ASCII string.
2402  * The values read are assumed to be in 1/USER_HZ seconds, and
2403  * are converted into jiffies.
2404  *
2405  * Returns 0 on success.
2406  */
2407 int proc_dointvec_userhz_jiffies(struct ctl_table *table, int write,
2408                                  void __user *buffer, size_t *lenp, loff_t *ppos)
2409 {
2410     return do_proc_dointvec(table,write,buffer,lenp,ppos,
2411                             do_proc_dointvec_userhz_jiffies_conv,NULL);
2412 }
2413
2414 /**
2415  * proc_dointvec_ms_jiffies - read a vector of integers as 1 milliseconds
2416  * @table: the sysctl table
2417  * @write: %TRUE if this is a write to the sysctl file
2418  * @buffer: the user buffer
2419  * @lenp: the size of the user buffer
2420  * @ppos: file position
2421  * @ppos: the current position in the file
2422  *
2423  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2424  * values from/to the user buffer, treated as an ASCII string.
2425  * The values read are assumed to be in 1/1000 seconds, and
2426  * are converted into jiffies.
2427  *
2428  * Returns 0 on success.
2429  */
2430 int proc_dointvec_ms_jiffies(struct ctl_table *table, int write,
2431                              void __user *buffer, size_t *lenp, loff_t *ppos)
2432 {
2433         return do_proc_dointvec(table, write, buffer, lenp, ppos,
2434                                 do_proc_dointvec_ms_jiffies_conv, NULL);
2435 }
2436
2437 static int proc_do_cad_pid(struct ctl_table *table, int write,
2438                            void __user *buffer, size_t *lenp, loff_t *ppos)
2439 {
2440         struct pid *new_pid;
2441         pid_t tmp;
2442         int r;
2443
2444         tmp = pid_vnr(cad_pid);
2445
2446         r = __do_proc_dointvec(&tmp, table, write, buffer,
2447                                lenp, ppos, NULL, NULL);
2448         if (r || !write)
2449                 return r;
2450
2451         new_pid = find_get_pid(tmp);
2452         if (!new_pid)
2453                 return -ESRCH;
2454
2455         put_pid(xchg(&cad_pid, new_pid));
2456         return 0;
2457 }
2458
2459 /**
2460  * proc_do_large_bitmap - read/write from/to a large bitmap
2461  * @table: the sysctl table
2462  * @write: %TRUE if this is a write to the sysctl file
2463  * @buffer: the user buffer
2464  * @lenp: the size of the user buffer
2465  * @ppos: file position
2466  *
2467  * The bitmap is stored at table->data and the bitmap length (in bits)
2468  * in table->maxlen.
2469  *
2470  * We use a range comma separated format (e.g. 1,3-4,10-10) so that
2471  * large bitmaps may be represented in a compact manner. Writing into
2472  * the file will clear the bitmap then update it with the given input.
2473  *
2474  * Returns 0 on success.
2475  */
2476 int proc_do_large_bitmap(struct ctl_table *table, int write,
2477                          void __user *buffer, size_t *lenp, loff_t *ppos)
2478 {
2479         int err = 0;
2480         bool first = 1;
2481         size_t left = *lenp;
2482         unsigned long bitmap_len = table->maxlen;
2483         unsigned long *bitmap = (unsigned long *) table->data;
2484         unsigned long *tmp_bitmap = NULL;
2485         char tr_a[] = { '-', ',', '\n' }, tr_b[] = { ',', '\n', 0 }, c;
2486
2487         if (!bitmap_len || !left || (*ppos && !write)) {
2488                 *lenp = 0;
2489                 return 0;
2490         }
2491
2492         if (write) {
2493                 unsigned long page = 0;
2494                 char *kbuf;
2495
2496                 if (left > PAGE_SIZE - 1)
2497                         left = PAGE_SIZE - 1;
2498
2499                 page = __get_free_page(GFP_TEMPORARY);
2500                 kbuf = (char *) page;
2501                 if (!kbuf)
2502                         return -ENOMEM;
2503                 if (copy_from_user(kbuf, buffer, left)) {
2504                         free_page(page);
2505                         return -EFAULT;
2506                 }
2507                 kbuf[left] = 0;
2508
2509                 tmp_bitmap = kzalloc(BITS_TO_LONGS(bitmap_len) * sizeof(unsigned long),
2510                                      GFP_KERNEL);
2511                 if (!tmp_bitmap) {
2512                         free_page(page);
2513                         return -ENOMEM;
2514                 }
2515                 proc_skip_char(&kbuf, &left, '\n');
2516                 while (!err && left) {
2517                         unsigned long val_a, val_b;
2518                         bool neg;
2519
2520                         err = proc_get_long(&kbuf, &left, &val_a, &neg, tr_a,
2521                                              sizeof(tr_a), &c);
2522                         if (err)
2523                                 break;
2524                         if (val_a >= bitmap_len || neg) {
2525                                 err = -EINVAL;
2526                                 break;
2527                         }
2528
2529                         val_b = val_a;
2530                         if (left) {
2531                                 kbuf++;
2532                                 left--;
2533                         }
2534
2535                         if (c == '-') {
2536                                 err = proc_get_long(&kbuf, &left, &val_b,
2537                                                      &neg, tr_b, sizeof(tr_b),
2538                                                      &c);
2539                                 if (err)
2540                                         break;
2541                                 if (val_b >= bitmap_len || neg ||
2542                                     val_a > val_b) {
2543                                         err = -EINVAL;
2544                                         break;
2545                                 }
2546                                 if (left) {
2547                                         kbuf++;
2548                                         left--;
2549                                 }
2550                         }
2551
2552                         bitmap_set(tmp_bitmap, val_a, val_b - val_a + 1);
2553                         first = 0;
2554                         proc_skip_char(&kbuf, &left, '\n');
2555                 }
2556                 free_page(page);
2557         } else {
2558                 unsigned long bit_a, bit_b = 0;
2559
2560                 while (left) {
2561                         bit_a = find_next_bit(bitmap, bitmap_len, bit_b);
2562                         if (bit_a >= bitmap_len)
2563                                 break;
2564                         bit_b = find_next_zero_bit(bitmap, bitmap_len,
2565                                                    bit_a + 1) - 1;
2566
2567                         if (!first) {
2568                                 err = proc_put_char(&buffer, &left, ',');
2569                                 if (err)
2570                                         break;
2571                         }
2572                         err = proc_put_long(&buffer, &left, bit_a, false);
2573                         if (err)
2574                                 break;
2575                         if (bit_a != bit_b) {
2576                                 err = proc_put_char(&buffer, &left, '-');
2577                                 if (err)
2578                                         break;
2579                                 err = proc_put_long(&buffer, &left, bit_b, false);
2580                                 if (err)
2581                                         break;
2582                         }
2583
2584                         first = 0; bit_b++;
2585                 }
2586                 if (!err)
2587                         err = proc_put_char(&buffer, &left, '\n');
2588         }
2589
2590         if (!err) {
2591                 if (write) {
2592                         if (*ppos)
2593                                 bitmap_or(bitmap, bitmap, tmp_bitmap, bitmap_len);
2594                         else
2595                                 bitmap_copy(bitmap, tmp_bitmap, bitmap_len);
2596                 }
2597                 kfree(tmp_bitmap);
2598                 *lenp -= left;
2599                 *ppos += *lenp;
2600                 return 0;
2601         } else {
2602                 kfree(tmp_bitmap);
2603                 return err;
2604         }
2605 }
2606
2607 #else /* CONFIG_PROC_SYSCTL */
2608
2609 int proc_dostring(struct ctl_table *table, int write,
2610                   void __user *buffer, size_t *lenp, loff_t *ppos)
2611 {
2612         return -ENOSYS;
2613 }
2614
2615 int proc_dointvec(struct ctl_table *table, int write,
2616                   void __user *buffer, size_t *lenp, loff_t *ppos)
2617 {
2618         return -ENOSYS;
2619 }
2620
2621 int proc_dointvec_minmax(struct ctl_table *table, int write,
2622                     void __user *buffer, size_t *lenp, loff_t *ppos)
2623 {
2624         return -ENOSYS;
2625 }
2626
2627 int proc_dointvec_jiffies(struct ctl_table *table, int write,
2628                     void __user *buffer, size_t *lenp, loff_t *ppos)
2629 {
2630         return -ENOSYS;
2631 }
2632
2633 int proc_dointvec_userhz_jiffies(struct ctl_table *table, int write,
2634                     void __user *buffer, size_t *lenp, loff_t *ppos)
2635 {
2636         return -ENOSYS;
2637 }
2638
2639 int proc_dointvec_ms_jiffies(struct ctl_table *table, int write,
2640                              void __user *buffer, size_t *lenp, loff_t *ppos)
2641 {
2642         return -ENOSYS;
2643 }
2644
2645 int proc_doulongvec_minmax(struct ctl_table *table, int write,
2646                     void __user *buffer, size_t *lenp, loff_t *ppos)
2647 {
2648         return -ENOSYS;
2649 }
2650
2651 int proc_doulongvec_ms_jiffies_minmax(struct ctl_table *table, int write,
2652                                       void __user *buffer,
2653                                       size_t *lenp, loff_t *ppos)
2654 {
2655     return -ENOSYS;
2656 }
2657
2658
2659 #endif /* CONFIG_PROC_SYSCTL */
2660
2661 /*
2662  * No sense putting this after each symbol definition, twice,
2663  * exception granted :-)
2664  */
2665 EXPORT_SYMBOL(proc_dointvec);
2666 EXPORT_SYMBOL(proc_dointvec_jiffies);
2667 EXPORT_SYMBOL(proc_dointvec_minmax);
2668 EXPORT_SYMBOL(proc_dointvec_userhz_jiffies);
2669 EXPORT_SYMBOL(proc_dointvec_ms_jiffies);
2670 EXPORT_SYMBOL(proc_dostring);
2671 EXPORT_SYMBOL(proc_doulongvec_minmax);
2672 EXPORT_SYMBOL(proc_doulongvec_ms_jiffies_minmax);