kernel/sysctl.c

   1 /*
   2  * sysctl.c: General linux system control interface
   3  *
   4  * Begun 24 March 1995, Stephen Tweedie
   5  * Added /proc support, Dec 1995
   6  * Added bdflush entry and intvec min/max checking, 2/23/96, Tom Dyas.
   7  * Added hooks for /proc/sys/net (minor, minor patch), 96/4/1, Mike Shaver.
   8  * Added kernel/java-{interpreter,appletviewer}, 96/5/10, Mike Shaver.
   9  * Dynamic registration fixes, Stephen Tweedie.
  10  * Added kswapd-interval, ctrl-alt-del, printk stuff, 1/8/97, Chris Horn.
  11  * Made sysctl support optional via CONFIG_SYSCTL, 1/10/97, Chris
  12  *  Horn.
  13  * Added proc_doulongvec_ms_jiffies_minmax, 09/08/99, Carlos H. Bauer.
  14  * Added proc_doulongvec_minmax, 09/08/99, Carlos H. Bauer.
  15  * Changed linked lists to use list.h instead of lists.h, 02/24/00, Bill
  16  *  Wendling.
  17  * The list_for_each() macro wasn't appropriate for the sysctl loop.
  18  *  Removed it and replaced it with older style, 03/23/00, Bill Wendling
  19  */
  20
  21 #include <linux/module.h>
  22 #include <linux/aio.h>
  23 #include <linux/mm.h>
  24 #include <linux/swap.h>
  25 #include <linux/slab.h>
  26 #include <linux/sysctl.h>
  27 #include <linux/bitmap.h>
  28 #include <linux/signal.h>
  29 #include <linux/printk.h>
  30 #include <linux/proc_fs.h>
  31 #include <linux/security.h>
  32 #include <linux/ctype.h>
  33 #include <linux/kmemcheck.h>
  34 #include <linux/kmemleak.h>
  35 #include <linux/fs.h>
  36 #include <linux/init.h>
  37 #include <linux/kernel.h>
  38 #include <linux/kobject.h>
  39 #include <linux/net.h>
  40 #include <linux/sysrq.h>
  41 #include <linux/highuid.h>
  42 #include <linux/writeback.h>
  43 #include <linux/ratelimit.h>
  44 #include <linux/compaction.h>
  45 #include <linux/hugetlb.h>
  46 #include <linux/initrd.h>
  47 #include <linux/key.h>
  48 #include <linux/times.h>
  49 #include <linux/limits.h>
  50 #include <linux/dcache.h>
  51 #include <linux/dnotify.h>
  52 #include <linux/syscalls.h>
  53 #include <linux/vmstat.h>
  54 #include <linux/nfs_fs.h>
  55 #include <linux/acpi.h>
  56 #include <linux/reboot.h>
  57 #include <linux/ftrace.h>
  58 #include <linux/perf_event.h>
  59 #include <linux/kprobes.h>
  60 #include <linux/pipe_fs_i.h>
  61 #include <linux/oom.h>
  62 #include <linux/kmod.h>
  63 #include <linux/capability.h>
  64 #include <linux/binfmts.h>
  65 #include <linux/sched/sysctl.h>
  66 #include <linux/kexec.h>
  67 #include <linux/bpf.h>
  68 #include <linux/mount.h>
  69
  70 #include <asm/uaccess.h>
  71 #include <asm/processor.h>
  72
  73 #ifdef CONFIG_X86
  74 #include <asm/nmi.h>
  75 #include <asm/stacktrace.h>
  76 #include <asm/io.h>
  77 #endif
  78 #ifdef CONFIG_SPARC
  79 #include <asm/setup.h>
  80 #endif
  81 #ifdef CONFIG_BSD_PROCESS_ACCT
  82 #include <linux/acct.h>
  83 #endif
  84 #ifdef CONFIG_RT_MUTEXES
  85 #include <linux/rtmutex.h>
  86 #endif
  87 #if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_LOCK_STAT)
  88 #include <linux/lockdep.h>
  89 #endif
  90 #ifdef CONFIG_CHR_DEV_SG
  91 #include <scsi/sg.h>
  92 #endif
  93
  94 #ifdef CONFIG_LOCKUP_DETECTOR
  95 #include <linux/nmi.h>
  96 #endif
  97
  98 #if defined(CONFIG_SYSCTL)
  99
 100 /* External variables not in a header file. */
 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 latencytop_enabled;
 111 extern unsigned int sysctl_nr_open_min, sysctl_nr_open_max;
 112 #ifndef CONFIG_MMU
 113 extern int sysctl_nr_trim_pages;
 114 #endif
 115
 116 /* Constants used for minimum and  maximum */
 117 #ifdef CONFIG_LOCKUP_DETECTOR
 118 static int sixty = 60;
 119 #endif
 120
 121 static int __maybe_unused neg_one = -1;
 122
 123 static int zero;
 124 static int __maybe_unused one = 1;
 125 static int __maybe_unused two = 2;
 126 static int __maybe_unused four = 4;
 127 static unsigned long one_ul = 1;
 128 static int one_hundred = 100;
 129 static int one_thousand = 1000;
 130 #ifdef CONFIG_PRINTK
 131 static int ten_thousand = 10000;
 132 #endif
 133 #ifdef CONFIG_PERF_EVENTS
 134 static int six_hundred_forty_kb = 640 * 1024;
 135 #endif
 136
 137 /* this is needed for the proc_doulongvec_minmax of vm_dirty_bytes */
 138 static unsigned long dirty_bytes_min = 2 * PAGE_SIZE;
 139
 140 /* this is needed for the proc_dointvec_minmax for [fs_]overflow UID and GID */
 141 static int maxolduid = 65535;
 142 static int minolduid;
 143
 144 static int ngroups_max = NGROUPS_MAX;
 145 static const int cap_last_cap = CAP_LAST_CAP;
 146
 147 /*this is needed for proc_doulongvec_minmax of sysctl_hung_task_timeout_secs */
 148 #ifdef CONFIG_DETECT_HUNG_TASK
 149 static unsigned long hung_task_timeout_max = (LONG_MAX/HZ);
 150 #endif
 151
 152 #ifdef CONFIG_INOTIFY_USER
 153 #include <linux/inotify.h>
 154 #endif
 155 #ifdef CONFIG_SPARC
 156 #endif
 157
 158 #ifdef __hppa__
 159 extern int pwrsw_enabled;
 160 #endif
 161
 162 #ifdef CONFIG_SYSCTL_ARCH_UNALIGN_ALLOW
 163 extern int unaligned_enabled;
 164 #endif
 165
 166 #ifdef CONFIG_IA64
 167 extern int unaligned_dump_stack;
 168 #endif
 169
 170 #ifdef CONFIG_SYSCTL_ARCH_UNALIGN_NO_WARN
 171 extern int no_unaligned_warning;
 172 #endif
 173
 174 #ifdef CONFIG_PROC_SYSCTL
 175
 176 #define SYSCTL_WRITES_LEGACY    -1
 177 #define SYSCTL_WRITES_WARN       0
 178 #define SYSCTL_WRITES_STRICT     1
 179
 180 static int sysctl_writes_strict = SYSCTL_WRITES_STRICT;
 181
 182 static int proc_do_cad_pid(struct ctl_table *table, int write,
 183                   void __user *buffer, size_t *lenp, loff_t *ppos);
 184 static int proc_taint(struct ctl_table *table, int write,
 185                                void __user *buffer, size_t *lenp, loff_t *ppos);
 186 #endif
 187
 188 #ifdef CONFIG_PRINTK
 189 static int proc_dointvec_minmax_sysadmin(struct ctl_table *table, int write,
 190                                 void __user *buffer, size_t *lenp, loff_t *ppos);
 191 #endif
 192
 193 static int proc_dointvec_minmax_coredump(struct ctl_table *table, int write,
 194                 void __user *buffer, size_t *lenp, loff_t *ppos);
 195 #ifdef CONFIG_COREDUMP
 196 static int proc_dostring_coredump(struct ctl_table *table, int write,
 197                 void __user *buffer, size_t *lenp, loff_t *ppos);
 198 #endif
 199
 200 #ifdef CONFIG_MAGIC_SYSRQ
 201 /* Note: sysrq code uses it's own private copy */
 202 static int __sysrq_enabled = CONFIG_MAGIC_SYSRQ_DEFAULT_ENABLE;
 203
 204 static int sysrq_sysctl_handler(struct ctl_table *table, int write,
 205                                 void __user *buffer, size_t *lenp,
 206                                 loff_t *ppos)
 207 {
 208         int error;
 209
 210         error = proc_dointvec(table, write, buffer, lenp, ppos);
 211         if (error)
 212                 return error;
 213
 214         if (write)
 215                 sysrq_toggle_support(__sysrq_enabled);
 216
 217         return 0;
 218 }
 219
 220 #endif
 221
 222 static struct ctl_table kern_table[];
 223 static struct ctl_table vm_table[];
 224 static struct ctl_table fs_table[];
 225 static struct ctl_table debug_table[];
 226 static struct ctl_table dev_table[];
 227 extern struct ctl_table random_table[];
 228 #ifdef CONFIG_EPOLL
 229 extern struct ctl_table epoll_table[];
 230 #endif
 231
 232 #ifdef HAVE_ARCH_PICK_MMAP_LAYOUT
 233 int sysctl_legacy_va_layout;
 234 #endif
 235
 236 /* The default sysctl tables: */
 237
 238 static struct ctl_table sysctl_base_table[] = {
 239         {
 240                 .procname       = "kernel",
 241                 .mode           = 0555,
 242                 .child          = kern_table,
 243         },
 244         {
 245                 .procname       = "vm",
 246                 .mode           = 0555,
 247                 .child          = vm_table,
 248         },
 249         {
 250                 .procname       = "fs",
 251                 .mode           = 0555,
 252                 .child          = fs_table,
 253         },
 254         {
 255                 .procname       = "debug",
 256                 .mode           = 0555,
 257                 .child          = debug_table,
 258         },
 259         {
 260                 .procname       = "dev",
 261                 .mode           = 0555,
 262                 .child          = dev_table,
 263         },
 264         { }
 265 };
 266
 267 #ifdef CONFIG_SCHED_DEBUG
 268 static int min_sched_granularity_ns = 100000;           /* 100 usecs */
 269 static int max_sched_granularity_ns = NSEC_PER_SEC;     /* 1 second */
 270 static int min_wakeup_granularity_ns;                   /* 0 usecs */
 271 static int max_wakeup_granularity_ns = NSEC_PER_SEC;    /* 1 second */
 272 #ifdef CONFIG_SMP
 273 static int min_sched_tunable_scaling = SCHED_TUNABLESCALING_NONE;
 274 static int max_sched_tunable_scaling = SCHED_TUNABLESCALING_END-1;
 275 #endif /* CONFIG_SMP */
 276 #endif /* CONFIG_SCHED_DEBUG */
 277
 278 #ifdef CONFIG_COMPACTION
 279 static int min_extfrag_threshold;
 280 static int max_extfrag_threshold = 1000;
 281 #endif
 282
 283 static struct ctl_table kern_table[] = {
 284         {
 285                 .procname       = "sched_child_runs_first",
 286                 .data           = &sysctl_sched_child_runs_first,
 287                 .maxlen         = sizeof(unsigned int),
 288                 .mode           = 0644,
 289                 .proc_handler   = proc_dointvec,
 290         },
 291 #ifdef CONFIG_SCHED_DEBUG
 292         {
 293                 .procname       = "sched_min_granularity_ns",
 294                 .data           = &sysctl_sched_min_granularity,
 295                 .maxlen         = sizeof(unsigned int),
 296                 .mode           = 0644,
 297                 .proc_handler   = sched_proc_update_handler,
 298                 .extra1         = &min_sched_granularity_ns,
 299                 .extra2         = &max_sched_granularity_ns,
 300         },
 301         {
 302                 .procname       = "sched_latency_ns",
 303                 .data           = &sysctl_sched_latency,
 304                 .maxlen         = sizeof(unsigned int),
 305                 .mode           = 0644,
 306                 .proc_handler   = sched_proc_update_handler,
 307                 .extra1         = &min_sched_granularity_ns,
 308                 .extra2         = &max_sched_granularity_ns,
 309         },
 310         {
 311                 .procname       = "sched_wakeup_granularity_ns",
 312                 .data           = &sysctl_sched_wakeup_granularity,
 313                 .maxlen         = sizeof(unsigned int),
 314                 .mode           = 0644,
 315                 .proc_handler   = sched_proc_update_handler,
 316                 .extra1         = &min_wakeup_granularity_ns,
 317                 .extra2         = &max_wakeup_granularity_ns,
 318         },
 319 #ifdef CONFIG_SMP
 320         {
 321                 .procname       = "sched_tunable_scaling",
 322                 .data           = &sysctl_sched_tunable_scaling,
 323                 .maxlen         = sizeof(enum sched_tunable_scaling),
 324                 .mode           = 0644,
 325                 .proc_handler   = sched_proc_update_handler,
 326                 .extra1         = &min_sched_tunable_scaling,
 327                 .extra2         = &max_sched_tunable_scaling,
 328         },
 329         {
 330                 .procname       = "sched_migration_cost_ns",
 331                 .data           = &sysctl_sched_migration_cost,
 332                 .maxlen         = sizeof(unsigned int),
 333                 .mode           = 0644,
 334                 .proc_handler   = proc_dointvec,
 335         },
 336         {
 337                 .procname       = "sched_nr_migrate",
 338                 .data           = &sysctl_sched_nr_migrate,
 339                 .maxlen         = sizeof(unsigned int),
 340                 .mode           = 0644,
 341                 .proc_handler   = proc_dointvec,
 342         },
 343         {
 344                 .procname       = "sched_time_avg_ms",
 345                 .data           = &sysctl_sched_time_avg,
 346                 .maxlen         = sizeof(unsigned int),
 347                 .mode           = 0644,
 348                 .proc_handler   = proc_dointvec_minmax,
 349                 .extra1         = &one,
 350         },
 351         {
 352                 .procname       = "sched_shares_window_ns",
 353                 .data           = &sysctl_sched_shares_window,
 354                 .maxlen         = sizeof(unsigned int),
 355                 .mode           = 0644,
 356                 .proc_handler   = proc_dointvec,
 357         },
 358 #ifdef CONFIG_SCHEDSTATS
 359         {
 360                 .procname       = "sched_schedstats",
 361                 .data           = NULL,
 362                 .maxlen         = sizeof(unsigned int),
 363                 .mode           = 0644,
 364                 .proc_handler   = sysctl_schedstats,
 365                 .extra1         = &zero,
 366                 .extra2         = &one,
 367         },
 368 #endif /* CONFIG_SCHEDSTATS */
 369 #endif /* CONFIG_SMP */
 370 #ifdef CONFIG_NUMA_BALANCING
 371         {
 372                 .procname       = "numa_balancing_scan_delay_ms",
 373                 .data           = &sysctl_numa_balancing_scan_delay,
 374                 .maxlen         = sizeof(unsigned int),
 375                 .mode           = 0644,
 376                 .proc_handler   = proc_dointvec,
 377         },
 378         {
 379                 .procname       = "numa_balancing_scan_period_min_ms",
 380                 .data           = &sysctl_numa_balancing_scan_period_min,
 381                 .maxlen         = sizeof(unsigned int),
 382                 .mode           = 0644,
 383                 .proc_handler   = proc_dointvec,
 384         },
 385         {
 386                 .procname       = "numa_balancing_scan_period_max_ms",
 387                 .data           = &sysctl_numa_balancing_scan_period_max,
 388                 .maxlen         = sizeof(unsigned int),
 389                 .mode           = 0644,
 390                 .proc_handler   = proc_dointvec,
 391         },
 392         {
 393                 .procname       = "numa_balancing_scan_size_mb",
 394                 .data           = &sysctl_numa_balancing_scan_size,
 395                 .maxlen         = sizeof(unsigned int),
 396                 .mode           = 0644,
 397                 .proc_handler   = proc_dointvec_minmax,
 398                 .extra1         = &one,
 399         },
 400         {
 401                 .procname       = "numa_balancing",
 402                 .data           = NULL, /* filled in by handler */
 403                 .maxlen         = sizeof(unsigned int),
 404                 .mode           = 0644,
 405                 .proc_handler   = sysctl_numa_balancing,
 406                 .extra1         = &zero,
 407                 .extra2         = &one,
 408         },
 409 #endif /* CONFIG_NUMA_BALANCING */
 410 #endif /* CONFIG_SCHED_DEBUG */
 411         {
 412                 .procname       = "sched_rt_period_us",
 413                 .data           = &sysctl_sched_rt_period,
 414                 .maxlen         = sizeof(unsigned int),
 415                 .mode           = 0644,
 416                 .proc_handler   = sched_rt_handler,
 417         },
 418         {
 419                 .procname       = "sched_rt_runtime_us",
 420                 .data           = &sysctl_sched_rt_runtime,
 421                 .maxlen         = sizeof(int),
 422                 .mode           = 0644,
 423                 .proc_handler   = sched_rt_handler,
 424         },
 425         {
 426                 .procname       = "sched_rr_timeslice_ms",
 427                 .data           = &sched_rr_timeslice,
 428                 .maxlen         = sizeof(int),
 429                 .mode           = 0644,
 430                 .proc_handler   = sched_rr_handler,
 431         },
 432 #ifdef CONFIG_SCHED_AUTOGROUP
 433         {
 434                 .procname       = "sched_autogroup_enabled",
 435                 .data           = &sysctl_sched_autogroup_enabled,
 436                 .maxlen         = sizeof(unsigned int),
 437                 .mode           = 0644,
 438                 .proc_handler   = proc_dointvec_minmax,
 439                 .extra1         = &zero,
 440                 .extra2         = &one,
 441         },
 442 #endif
 443 #ifdef CONFIG_CFS_BANDWIDTH
 444         {
 445                 .procname       = "sched_cfs_bandwidth_slice_us",
 446                 .data           = &sysctl_sched_cfs_bandwidth_slice,
 447                 .maxlen         = sizeof(unsigned int),
 448                 .mode           = 0644,
 449                 .proc_handler   = proc_dointvec_minmax,
 450                 .extra1         = &one,
 451         },
 452 #endif
 453 #ifdef CONFIG_PROVE_LOCKING
 454         {
 455                 .procname       = "prove_locking",
 456                 .data           = &prove_locking,
 457                 .maxlen         = sizeof(int),
 458                 .mode           = 0644,
 459                 .proc_handler   = proc_dointvec,
 460         },
 461 #endif
 462 #ifdef CONFIG_LOCK_STAT
 463         {
 464                 .procname       = "lock_stat",
 465                 .data           = &lock_stat,
 466                 .maxlen         = sizeof(int),
 467                 .mode           = 0644,
 468                 .proc_handler   = proc_dointvec,
 469         },
 470 #endif
 471         {
 472                 .procname       = "panic",
 473                 .data           = &panic_timeout,
 474                 .maxlen         = sizeof(int),
 475                 .mode           = 0644,
 476                 .proc_handler   = proc_dointvec,
 477         },
 478 #ifdef CONFIG_COREDUMP
 479         {
 480                 .procname       = "core_uses_pid",
 481                 .data           = &core_uses_pid,
 482                 .maxlen         = sizeof(int),
 483                 .mode           = 0644,
 484                 .proc_handler   = proc_dointvec,
 485         },
 486         {
 487                 .procname       = "core_pattern",
 488                 .data           = core_pattern,
 489                 .maxlen         = CORENAME_MAX_SIZE,
 490                 .mode           = 0644,
 491                 .proc_handler   = proc_dostring_coredump,
 492         },
 493         {
 494                 .procname       = "core_pipe_limit",
 495                 .data           = &core_pipe_limit,
 496                 .maxlen         = sizeof(unsigned int),
 497                 .mode           = 0644,
 498                 .proc_handler   = proc_dointvec,
 499         },
 500 #endif
 501 #ifdef CONFIG_PROC_SYSCTL
 502         {
 503                 .procname       = "tainted",
 504                 .maxlen         = sizeof(long),
 505                 .mode           = 0644,
 506                 .proc_handler   = proc_taint,
 507         },
 508         {
 509                 .procname       = "sysctl_writes_strict",
 510                 .data           = &sysctl_writes_strict,
 511                 .maxlen         = sizeof(int),
 512                 .mode           = 0644,
 513                 .proc_handler   = proc_dointvec_minmax,
 514                 .extra1         = &neg_one,
 515                 .extra2         = &one,
 516         },
 517 #endif
 518 #ifdef CONFIG_LATENCYTOP
 519         {
 520                 .procname       = "latencytop",
 521                 .data           = &latencytop_enabled,
 522                 .maxlen         = sizeof(int),
 523                 .mode           = 0644,
 524                 .proc_handler   = sysctl_latencytop,
 525         },
 526 #endif
 527 #ifdef CONFIG_BLK_DEV_INITRD
 528         {
 529                 .procname       = "real-root-dev",
 530                 .data           = &real_root_dev,
 531                 .maxlen         = sizeof(int),
 532                 .mode           = 0644,
 533                 .proc_handler   = proc_dointvec,
 534         },
 535 #endif
 536         {
 537                 .procname       = "print-fatal-signals",
 538                 .data           = &print_fatal_signals,
 539                 .maxlen         = sizeof(int),
 540                 .mode           = 0644,
 541                 .proc_handler   = proc_dointvec,
 542         },
 543 #ifdef CONFIG_SPARC
 544         {
 545                 .procname       = "reboot-cmd",
 546                 .data           = reboot_command,
 547                 .maxlen         = 256,
 548                 .mode           = 0644,
 549                 .proc_handler   = proc_dostring,
 550         },
 551         {
 552                 .procname       = "stop-a",
 553                 .data           = &stop_a_enabled,
 554                 .maxlen         = sizeof (int),
 555                 .mode           = 0644,
 556                 .proc_handler   = proc_dointvec,
 557         },
 558         {
 559                 .procname       = "scons-poweroff",
 560                 .data           = &scons_pwroff,
 561                 .maxlen         = sizeof (int),
 562                 .mode           = 0644,
 563                 .proc_handler   = proc_dointvec,
 564         },
 565 #endif
 566 #ifdef CONFIG_SPARC64
 567         {
 568                 .procname       = "tsb-ratio",
 569                 .data           = &sysctl_tsb_ratio,
 570                 .maxlen         = sizeof (int),
 571                 .mode           = 0644,
 572                 .proc_handler   = proc_dointvec,
 573         },
 574 #endif
 575 #ifdef __hppa__
 576         {
 577                 .procname       = "soft-power",
 578                 .data           = &pwrsw_enabled,
 579                 .maxlen         = sizeof (int),
 580                 .mode           = 0644,
 581                 .proc_handler   = proc_dointvec,
 582         },
 583 #endif
 584 #ifdef CONFIG_SYSCTL_ARCH_UNALIGN_ALLOW
 585         {
 586                 .procname       = "unaligned-trap",
 587                 .data           = &unaligned_enabled,
 588                 .maxlen         = sizeof (int),
 589                 .mode           = 0644,
 590                 .proc_handler   = proc_dointvec,
 591         },
 592 #endif
 593         {
 594                 .procname       = "ctrl-alt-del",
 595                 .data           = &C_A_D,
 596                 .maxlen         = sizeof(int),
 597                 .mode           = 0644,
 598                 .proc_handler   = proc_dointvec,
 599         },
 600 #ifdef CONFIG_FUNCTION_TRACER
 601         {
 602                 .procname       = "ftrace_enabled",
 603                 .data           = &ftrace_enabled,
 604                 .maxlen         = sizeof(int),
 605                 .mode           = 0644,
 606                 .proc_handler   = ftrace_enable_sysctl,
 607         },
 608 #endif
 609 #ifdef CONFIG_STACK_TRACER
 610         {
 611                 .procname       = "stack_tracer_enabled",
 612                 .data           = &stack_tracer_enabled,
 613                 .maxlen         = sizeof(int),
 614                 .mode           = 0644,
 615                 .proc_handler   = stack_trace_sysctl,
 616         },
 617 #endif
 618 #ifdef CONFIG_TRACING
 619         {
 620                 .procname       = "ftrace_dump_on_oops",
 621                 .data           = &ftrace_dump_on_oops,
 622                 .maxlen         = sizeof(int),
 623                 .mode           = 0644,
 624                 .proc_handler   = proc_dointvec,
 625         },
 626         {
 627                 .procname       = "traceoff_on_warning",
 628                 .data           = &__disable_trace_on_warning,
 629                 .maxlen         = sizeof(__disable_trace_on_warning),
 630                 .mode           = 0644,
 631                 .proc_handler   = proc_dointvec,
 632         },
 633         {
 634                 .procname       = "tracepoint_printk",
 635                 .data           = &tracepoint_printk,
 636                 .maxlen         = sizeof(tracepoint_printk),
 637                 .mode           = 0644,
 638                 .proc_handler   = proc_dointvec,
 639         },
 640 #endif
 641 #ifdef CONFIG_KEXEC_CORE
 642         {
 643                 .procname       = "kexec_load_disabled",
 644                 .data           = &kexec_load_disabled,
 645                 .maxlen         = sizeof(int),
 646                 .mode           = 0644,
 647                 /* only handle a transition from default "0" to "1" */
 648                 .proc_handler   = proc_dointvec_minmax,
 649                 .extra1         = &one,
 650                 .extra2         = &one,
 651         },
 652 #endif
 653 #ifdef CONFIG_MODULES
 654         {
 655                 .procname       = "modprobe",
 656                 .data           = &modprobe_path,
 657                 .maxlen         = KMOD_PATH_LEN,
 658                 .mode           = 0644,
 659                 .proc_handler   = proc_dostring,
 660         },
 661         {
 662                 .procname       = "modules_disabled",
 663                 .data           = &modules_disabled,
 664                 .maxlen         = sizeof(int),
 665                 .mode           = 0644,
 666                 /* only handle a transition from default "0" to "1" */
 667                 .proc_handler   = proc_dointvec_minmax,
 668                 .extra1         = &one,
 669                 .extra2         = &one,
 670         },
 671 #endif
 672 #ifdef CONFIG_UEVENT_HELPER
 673         {
 674                 .procname       = "hotplug",
 675                 .data           = &uevent_helper,
 676                 .maxlen         = UEVENT_HELPER_PATH_LEN,
 677                 .mode           = 0644,
 678                 .proc_handler   = proc_dostring,
 679         },
 680 #endif
 681 #ifdef CONFIG_CHR_DEV_SG
 682         {
 683                 .procname       = "sg-big-buff",
 684                 .data           = &sg_big_buff,
 685                 .maxlen         = sizeof (int),
 686                 .mode           = 0444,
 687                 .proc_handler   = proc_dointvec,
 688         },
 689 #endif
 690 #ifdef CONFIG_BSD_PROCESS_ACCT
 691         {
 692                 .procname       = "acct",
 693                 .data           = &acct_parm,
 694                 .maxlen         = 3*sizeof(int),
 695                 .mode           = 0644,
 696                 .proc_handler   = proc_dointvec,
 697         },
 698 #endif
 699 #ifdef CONFIG_MAGIC_SYSRQ
 700         {
 701                 .procname       = "sysrq",
 702                 .data           = &__sysrq_enabled,
 703                 .maxlen         = sizeof (int),
 704                 .mode           = 0644,
 705                 .proc_handler   = sysrq_sysctl_handler,
 706         },
 707 #endif
 708 #ifdef CONFIG_PROC_SYSCTL
 709         {
 710                 .procname       = "cad_pid",
 711                 .data           = NULL,
 712                 .maxlen         = sizeof (int),
 713                 .mode           = 0600,
 714                 .proc_handler   = proc_do_cad_pid,
 715         },
 716 #endif
 717         {
 718                 .procname       = "threads-max",
 719                 .data           = NULL,
 720                 .maxlen         = sizeof(int),
 721                 .mode           = 0644,
 722                 .proc_handler   = sysctl_max_threads,
 723         },
 724         {
 725                 .procname       = "random",
 726                 .mode           = 0555,
 727                 .child          = random_table,
 728         },
 729         {
 730                 .procname       = "usermodehelper",
 731                 .mode           = 0555,
 732                 .child          = usermodehelper_table,
 733         },
 734         {
 735                 .procname       = "overflowuid",
 736                 .data           = &overflowuid,
 737                 .maxlen         = sizeof(int),
 738                 .mode           = 0644,
 739                 .proc_handler   = proc_dointvec_minmax,
 740                 .extra1         = &minolduid,
 741                 .extra2         = &maxolduid,
 742         },
 743         {
 744                 .procname       = "overflowgid",
 745                 .data           = &overflowgid,
 746                 .maxlen         = sizeof(int),
 747                 .mode           = 0644,
 748                 .proc_handler   = proc_dointvec_minmax,
 749                 .extra1         = &minolduid,
 750                 .extra2         = &maxolduid,
 751         },
 752 #ifdef CONFIG_S390
 753 #ifdef CONFIG_MATHEMU
 754         {
 755                 .procname       = "ieee_emulation_warnings",
 756                 .data           = &sysctl_ieee_emulation_warnings,
 757                 .maxlen         = sizeof(int),
 758                 .mode           = 0644,
 759                 .proc_handler   = proc_dointvec,
 760         },
 761 #endif
 762         {
 763                 .procname       = "userprocess_debug",
 764                 .data           = &show_unhandled_signals,
 765                 .maxlen         = sizeof(int),
 766                 .mode           = 0644,
 767                 .proc_handler   = proc_dointvec,
 768         },
 769 #endif
 770         {
 771                 .procname       = "pid_max",
 772                 .data           = &pid_max,
 773                 .maxlen         = sizeof (int),
 774                 .mode           = 0644,
 775                 .proc_handler   = proc_dointvec_minmax,
 776                 .extra1         = &pid_max_min,
 777                 .extra2         = &pid_max_max,
 778         },
 779         {
 780                 .procname       = "panic_on_oops",
 781                 .data           = &panic_on_oops,
 782                 .maxlen         = sizeof(int),
 783                 .mode           = 0644,
 784                 .proc_handler   = proc_dointvec,
 785         },
 786 #if defined CONFIG_PRINTK
 787         {
 788                 .procname       = "printk",
 789                 .data           = &console_loglevel,
 790                 .maxlen         = 4*sizeof(int),
 791                 .mode           = 0644,
 792                 .proc_handler   = proc_dointvec,
 793         },
 794         {
 795                 .procname       = "printk_ratelimit",
 796                 .data           = &printk_ratelimit_state.interval,
 797                 .maxlen         = sizeof(int),
 798                 .mode           = 0644,
 799                 .proc_handler   = proc_dointvec_jiffies,
 800         },
 801         {
 802                 .procname       = "printk_ratelimit_burst",
 803                 .data           = &printk_ratelimit_state.burst,
 804                 .maxlen         = sizeof(int),
 805                 .mode           = 0644,
 806                 .proc_handler   = proc_dointvec,
 807         },
 808         {
 809                 .procname       = "printk_delay",
 810                 .data           = &printk_delay_msec,
 811                 .maxlen         = sizeof(int),
 812                 .mode           = 0644,
 813                 .proc_handler   = proc_dointvec_minmax,
 814                 .extra1         = &zero,
 815                 .extra2         = &ten_thousand,
 816         },
 817         {
 818                 .procname       = "printk_devkmsg",
 819                 .data           = devkmsg_log_str,
 820                 .maxlen         = DEVKMSG_STR_MAX_SIZE,
 821                 .mode           = 0644,
 822                 .proc_handler   = devkmsg_sysctl_set_loglvl,
 823         },
 824         {
 825                 .procname       = "dmesg_restrict",
 826                 .data           = &dmesg_restrict,
 827                 .maxlen         = sizeof(int),
 828                 .mode           = 0644,
 829                 .proc_handler   = proc_dointvec_minmax_sysadmin,
 830                 .extra1         = &zero,
 831                 .extra2         = &one,
 832         },
 833         {
 834                 .procname       = "kptr_restrict",
 835                 .data           = &kptr_restrict,
 836                 .maxlen         = sizeof(int),
 837                 .mode           = 0644,
 838                 .proc_handler   = proc_dointvec_minmax_sysadmin,
 839                 .extra1         = &zero,
 840                 .extra2         = &two,
 841         },
 842 #endif
 843         {
 844                 .procname       = "ngroups_max",
 845                 .data           = &ngroups_max,
 846                 .maxlen         = sizeof (int),
 847                 .mode           = 0444,
 848                 .proc_handler   = proc_dointvec,
 849         },
 850         {
 851                 .procname       = "cap_last_cap",
 852                 .data           = (void *)&cap_last_cap,
 853                 .maxlen         = sizeof(int),
 854                 .mode           = 0444,
 855                 .proc_handler   = proc_dointvec,
 856         },
 857 #if defined(CONFIG_LOCKUP_DETECTOR)
 858         {
 859                 .procname       = "watchdog",
 860                 .data           = &watchdog_user_enabled,
 861                 .maxlen         = sizeof (int),
 862                 .mode           = 0644,
 863                 .proc_handler   = proc_watchdog,
 864                 .extra1         = &zero,
 865                 .extra2         = &one,
 866         },
 867         {
 868                 .procname       = "watchdog_thresh",
 869                 .data           = &watchdog_thresh,
 870                 .maxlen         = sizeof(int),
 871                 .mode           = 0644,
 872                 .proc_handler   = proc_watchdog_thresh,
 873                 .extra1         = &zero,
 874                 .extra2         = &sixty,
 875         },
 876         {
 877                 .procname       = "nmi_watchdog",
 878                 .data           = &nmi_watchdog_enabled,
 879                 .maxlen         = sizeof (int),
 880                 .mode           = 0644,
 881                 .proc_handler   = proc_nmi_watchdog,
 882                 .extra1         = &zero,
 883 #if defined(CONFIG_HAVE_NMI_WATCHDOG) || defined(CONFIG_HARDLOCKUP_DETECTOR)
 884                 .extra2         = &one,
 885 #else
 886                 .extra2         = &zero,
 887 #endif
 888         },
 889         {
 890                 .procname       = "soft_watchdog",
 891                 .data           = &soft_watchdog_enabled,
 892                 .maxlen         = sizeof (int),
 893                 .mode           = 0644,
 894                 .proc_handler   = proc_soft_watchdog,
 895                 .extra1         = &zero,
 896                 .extra2         = &one,
 897         },
 898         {
 899                 .procname       = "watchdog_cpumask",
 900                 .data           = &watchdog_cpumask_bits,
 901                 .maxlen         = NR_CPUS,
 902                 .mode           = 0644,
 903                 .proc_handler   = proc_watchdog_cpumask,
 904         },
 905         {
 906                 .procname       = "softlockup_panic",
 907                 .data           = &softlockup_panic,
 908                 .maxlen         = sizeof(int),
 909                 .mode           = 0644,
 910                 .proc_handler   = proc_dointvec_minmax,
 911                 .extra1         = &zero,
 912                 .extra2         = &one,
 913         },
 914 #ifdef CONFIG_HARDLOCKUP_DETECTOR
 915         {
 916                 .procname       = "hardlockup_panic",
 917                 .data           = &hardlockup_panic,
 918                 .maxlen         = sizeof(int),
 919                 .mode           = 0644,
 920                 .proc_handler   = proc_dointvec_minmax,
 921                 .extra1         = &zero,
 922                 .extra2         = &one,
 923         },
 924 #endif
 925 #ifdef CONFIG_SMP
 926         {
 927                 .procname       = "softlockup_all_cpu_backtrace",
 928                 .data           = &sysctl_softlockup_all_cpu_backtrace,
 929                 .maxlen         = sizeof(int),
 930                 .mode           = 0644,
 931                 .proc_handler   = proc_dointvec_minmax,
 932                 .extra1         = &zero,
 933                 .extra2         = &one,
 934         },
 935         {
 936                 .procname       = "hardlockup_all_cpu_backtrace",
 937                 .data           = &sysctl_hardlockup_all_cpu_backtrace,
 938                 .maxlen         = sizeof(int),
 939                 .mode           = 0644,
 940                 .proc_handler   = proc_dointvec_minmax,
 941                 .extra1         = &zero,
 942                 .extra2         = &one,
 943         },
 944 #endif /* CONFIG_SMP */
 945 #endif
 946 #if defined(CONFIG_X86_LOCAL_APIC) && defined(CONFIG_X86)
 947         {
 948                 .procname       = "unknown_nmi_panic",
 949                 .data           = &unknown_nmi_panic,
 950                 .maxlen         = sizeof (int),
 951                 .mode           = 0644,
 952                 .proc_handler   = proc_dointvec,
 953         },
 954 #endif
 955 #if defined(CONFIG_X86)
 956         {
 957                 .procname       = "panic_on_unrecovered_nmi",
 958                 .data           = &panic_on_unrecovered_nmi,
 959                 .maxlen         = sizeof(int),
 960                 .mode           = 0644,
 961                 .proc_handler   = proc_dointvec,
 962         },
 963         {
 964                 .procname       = "panic_on_io_nmi",
 965                 .data           = &panic_on_io_nmi,
 966                 .maxlen         = sizeof(int),
 967                 .mode           = 0644,
 968                 .proc_handler   = proc_dointvec,
 969         },
 970 #ifdef CONFIG_DEBUG_STACKOVERFLOW
 971         {
 972                 .procname       = "panic_on_stackoverflow",
 973                 .data           = &sysctl_panic_on_stackoverflow,
 974                 .maxlen         = sizeof(int),
 975                 .mode           = 0644,
 976                 .proc_handler   = proc_dointvec,
 977         },
 978 #endif
 979         {
 980                 .procname       = "bootloader_type",
 981                 .data           = &bootloader_type,
 982                 .maxlen         = sizeof (int),
 983                 .mode           = 0444,
 984                 .proc_handler   = proc_dointvec,
 985         },
 986         {
 987                 .procname       = "bootloader_version",
 988                 .data           = &bootloader_version,
 989                 .maxlen         = sizeof (int),
 990                 .mode           = 0444,
 991                 .proc_handler   = proc_dointvec,
 992         },
 993         {
 994                 .procname       = "kstack_depth_to_print",
 995                 .data           = &kstack_depth_to_print,
 996                 .maxlen         = sizeof(int),
 997                 .mode           = 0644,
 998                 .proc_handler   = proc_dointvec,
 999         },
1000         {
1001                 .procname       = "io_delay_type",
1002                 .data           = &io_delay_type,
1003                 .maxlen         = sizeof(int),
1004                 .mode           = 0644,
1005                 .proc_handler   = proc_dointvec,
1006         },
1007 #endif
1008 #if defined(CONFIG_MMU)
1009         {
1010                 .procname       = "randomize_va_space",
1011                 .data           = &randomize_va_space,
1012                 .maxlen         = sizeof(int),
1013                 .mode           = 0644,
1014                 .proc_handler   = proc_dointvec,
1015         },
1016 #endif
1017 #if defined(CONFIG_S390) && defined(CONFIG_SMP)
1018         {
1019                 .procname       = "spin_retry",
1020                 .data           = &spin_retry,
1021                 .maxlen         = sizeof (int),
1022                 .mode           = 0644,
1023                 .proc_handler   = proc_dointvec,
1024         },
1025 #endif
1026 #if     defined(CONFIG_ACPI_SLEEP) && defined(CONFIG_X86)
1027         {
1028                 .procname       = "acpi_video_flags",
1029                 .data           = &acpi_realmode_flags,
1030                 .maxlen         = sizeof (unsigned long),
1031                 .mode           = 0644,
1032                 .proc_handler   = proc_doulongvec_minmax,
1033         },
1034 #endif
1035 #ifdef CONFIG_SYSCTL_ARCH_UNALIGN_NO_WARN
1036         {
1037                 .procname       = "ignore-unaligned-usertrap",
1038                 .data           = &no_unaligned_warning,
1039                 .maxlen         = sizeof (int),
1040                 .mode           = 0644,
1041                 .proc_handler   = proc_dointvec,
1042         },
1043 #endif
1044 #ifdef CONFIG_IA64
1045         {
1046                 .procname       = "unaligned-dump-stack",
1047                 .data           = &unaligned_dump_stack,
1048                 .maxlen         = sizeof (int),
1049                 .mode           = 0644,
1050                 .proc_handler   = proc_dointvec,
1051         },
1052 #endif
1053 #ifdef CONFIG_DETECT_HUNG_TASK
1054         {
1055                 .procname       = "hung_task_panic",
1056                 .data           = &sysctl_hung_task_panic,
1057                 .maxlen         = sizeof(int),
1058                 .mode           = 0644,
1059                 .proc_handler   = proc_dointvec_minmax,
1060                 .extra1         = &zero,
1061                 .extra2         = &one,
1062         },
1063         {
1064                 .procname       = "hung_task_check_count",
1065                 .data           = &sysctl_hung_task_check_count,
1066                 .maxlen         = sizeof(int),
1067                 .mode           = 0644,
1068                 .proc_handler   = proc_dointvec_minmax,
1069                 .extra1         = &zero,
1070         },
1071         {
1072                 .procname       = "hung_task_timeout_secs",
1073                 .data           = &sysctl_hung_task_timeout_secs,
1074                 .maxlen         = sizeof(unsigned long),
1075                 .mode           = 0644,
1076                 .proc_handler   = proc_dohung_task_timeout_secs,
1077                 .extra2         = &hung_task_timeout_max,
1078         },
1079         {
1080                 .procname       = "hung_task_warnings",
1081                 .data           = &sysctl_hung_task_warnings,
1082                 .maxlen         = sizeof(int),
1083                 .mode           = 0644,
1084                 .proc_handler   = proc_dointvec_minmax,
1085                 .extra1         = &neg_one,
1086         },
1087 #endif
1088 #ifdef CONFIG_RT_MUTEXES
1089         {
1090                 .procname       = "max_lock_depth",
1091                 .data           = &max_lock_depth,
1092                 .maxlen         = sizeof(int),
1093                 .mode           = 0644,
1094                 .proc_handler   = proc_dointvec,
1095         },
1096 #endif
1097         {
1098                 .procname       = "poweroff_cmd",
1099                 .data           = &poweroff_cmd,
1100                 .maxlen         = POWEROFF_CMD_PATH_LEN,
1101                 .mode           = 0644,
1102                 .proc_handler   = proc_dostring,
1103         },
1104 #ifdef CONFIG_KEYS
1105         {
1106                 .procname       = "keys",
1107                 .mode           = 0555,
1108                 .child          = key_sysctls,
1109         },
1110 #endif
1111 #ifdef CONFIG_PERF_EVENTS
1112         /*
1113          * User-space scripts rely on the existence of this file
1114          * as a feature check for perf_events being enabled.
1115          *
1116          * So it's an ABI, do not remove!
1117          */
1118         {
1119                 .procname       = "perf_event_paranoid",
1120                 .data           = &sysctl_perf_event_paranoid,
1121                 .maxlen         = sizeof(sysctl_perf_event_paranoid),
1122                 .mode           = 0644,
1123                 .proc_handler   = proc_dointvec,
1124         },
1125         {
1126                 .procname       = "perf_event_mlock_kb",
1127                 .data           = &sysctl_perf_event_mlock,
1128                 .maxlen         = sizeof(sysctl_perf_event_mlock),
1129                 .mode           = 0644,
1130                 .proc_handler   = proc_dointvec,
1131         },
1132         {
1133                 .procname       = "perf_event_max_sample_rate",
1134                 .data           = &sysctl_perf_event_sample_rate,
1135                 .maxlen         = sizeof(sysctl_perf_event_sample_rate),
1136                 .mode           = 0644,
1137                 .proc_handler   = perf_proc_update_handler,
1138                 .extra1         = &one,
1139         },
1140         {
1141                 .procname       = "perf_cpu_time_max_percent",
1142                 .data           = &sysctl_perf_cpu_time_max_percent,
1143                 .maxlen         = sizeof(sysctl_perf_cpu_time_max_percent),
1144                 .mode           = 0644,
1145                 .proc_handler   = perf_cpu_time_max_percent_handler,
1146                 .extra1         = &zero,
1147                 .extra2         = &one_hundred,
1148         },
1149         {
1150                 .procname       = "perf_event_max_stack",
1151                 .data           = &sysctl_perf_event_max_stack,
1152                 .maxlen         = sizeof(sysctl_perf_event_max_stack),
1153                 .mode           = 0644,
1154                 .proc_handler   = perf_event_max_stack_handler,
1155                 .extra1         = &zero,
1156                 .extra2         = &six_hundred_forty_kb,
1157         },
1158         {
1159                 .procname       = "perf_event_max_contexts_per_stack",
1160                 .data           = &sysctl_perf_event_max_contexts_per_stack,
1161                 .maxlen         = sizeof(sysctl_perf_event_max_contexts_per_stack),
1162                 .mode           = 0644,
1163                 .proc_handler   = perf_event_max_stack_handler,
1164                 .extra1         = &zero,
1165                 .extra2         = &one_thousand,
1166         },
1167 #endif
1168 #ifdef CONFIG_KMEMCHECK
1169         {
1170                 .procname       = "kmemcheck",
1171                 .data           = &kmemcheck_enabled,
1172                 .maxlen         = sizeof(int),
1173                 .mode           = 0644,
1174                 .proc_handler   = proc_dointvec,
1175         },
1176 #endif
1177         {
1178                 .procname       = "panic_on_warn",
1179                 .data           = &panic_on_warn,
1180                 .maxlen         = sizeof(int),
1181                 .mode           = 0644,
1182                 .proc_handler   = proc_dointvec_minmax,
1183                 .extra1         = &zero,
1184                 .extra2         = &one,
1185         },
1186 #if defined(CONFIG_SMP) && defined(CONFIG_NO_HZ_COMMON)
1187         {
1188                 .procname       = "timer_migration",
1189                 .data           = &sysctl_timer_migration,
1190                 .maxlen         = sizeof(unsigned int),
1191                 .mode           = 0644,
1192                 .proc_handler   = timer_migration_handler,
1193                 .extra1         = &zero,
1194                 .extra2         = &one,
1195         },
1196 #endif
1197 #ifdef CONFIG_BPF_SYSCALL
1198         {
1199                 .procname       = "unprivileged_bpf_disabled",
1200                 .data           = &sysctl_unprivileged_bpf_disabled,
1201                 .maxlen         = sizeof(sysctl_unprivileged_bpf_disabled),
1202                 .mode           = 0644,
1203                 /* only handle a transition from default "0" to "1" */
1204                 .proc_handler   = proc_dointvec_minmax,
1205                 .extra1         = &one,
1206                 .extra2         = &one,
1207         },
1208 #endif
1209 #if defined(CONFIG_TREE_RCU) || defined(CONFIG_PREEMPT_RCU)
1210         {
1211                 .procname       = "panic_on_rcu_stall",
1212                 .data           = &sysctl_panic_on_rcu_stall,
1213                 .maxlen         = sizeof(sysctl_panic_on_rcu_stall),
1214                 .mode           = 0644,
1215                 .proc_handler   = proc_dointvec_minmax,
1216                 .extra1         = &zero,
1217                 .extra2         = &one,
1218         },
1219 #endif
1220         { }
1221 };
1222
1223 static struct ctl_table vm_table[] = {
1224         {
1225                 .procname       = "overcommit_memory",
1226                 .data           = &sysctl_overcommit_memory,
1227                 .maxlen         = sizeof(sysctl_overcommit_memory),
1228                 .mode           = 0644,
1229                 .proc_handler   = proc_dointvec_minmax,
1230                 .extra1         = &zero,
1231                 .extra2         = &two,
1232         },
1233         {
1234                 .procname       = "panic_on_oom",
1235                 .data           = &sysctl_panic_on_oom,
1236                 .maxlen         = sizeof(sysctl_panic_on_oom),
1237                 .mode           = 0644,
1238                 .proc_handler   = proc_dointvec_minmax,
1239                 .extra1         = &zero,
1240                 .extra2         = &two,
1241         },
1242         {
1243                 .procname       = "oom_kill_allocating_task",
1244                 .data           = &sysctl_oom_kill_allocating_task,
1245                 .maxlen         = sizeof(sysctl_oom_kill_allocating_task),
1246                 .mode           = 0644,
1247                 .proc_handler   = proc_dointvec,
1248         },
1249         {
1250                 .procname       = "oom_dump_tasks",
1251                 .data           = &sysctl_oom_dump_tasks,
1252                 .maxlen         = sizeof(sysctl_oom_dump_tasks),
1253                 .mode           = 0644,
1254                 .proc_handler   = proc_dointvec,
1255         },
1256         {
1257                 .procname       = "overcommit_ratio",
1258                 .data           = &sysctl_overcommit_ratio,
1259                 .maxlen         = sizeof(sysctl_overcommit_ratio),
1260                 .mode           = 0644,
1261                 .proc_handler   = overcommit_ratio_handler,
1262         },
1263         {
1264                 .procname       = "overcommit_kbytes",
1265                 .data           = &sysctl_overcommit_kbytes,
1266                 .maxlen         = sizeof(sysctl_overcommit_kbytes),
1267                 .mode           = 0644,
1268                 .proc_handler   = overcommit_kbytes_handler,
1269         },
1270         {
1271                 .procname       = "page-cluster",
1272                 .data           = &page_cluster,
1273                 .maxlen         = sizeof(int),
1274                 .mode           = 0644,
1275                 .proc_handler   = proc_dointvec_minmax,
1276                 .extra1         = &zero,
1277         },
1278         {
1279                 .procname       = "dirty_background_ratio",
1280                 .data           = &dirty_background_ratio,
1281                 .maxlen         = sizeof(dirty_background_ratio),
1282                 .mode           = 0644,
1283                 .proc_handler   = dirty_background_ratio_handler,
1284                 .extra1         = &zero,
1285                 .extra2         = &one_hundred,
1286         },
1287         {
1288                 .procname       = "dirty_background_bytes",
1289                 .data           = &dirty_background_bytes,
1290                 .maxlen         = sizeof(dirty_background_bytes),
1291                 .mode           = 0644,
1292                 .proc_handler   = dirty_background_bytes_handler,
1293                 .extra1         = &one_ul,
1294         },
1295         {
1296                 .procname       = "dirty_ratio",
1297                 .data           = &vm_dirty_ratio,
1298                 .maxlen         = sizeof(vm_dirty_ratio),
1299                 .mode           = 0644,
1300                 .proc_handler   = dirty_ratio_handler,
1301                 .extra1         = &zero,
1302                 .extra2         = &one_hundred,
1303         },
1304         {
1305                 .procname       = "dirty_bytes",
1306                 .data           = &vm_dirty_bytes,
1307                 .maxlen         = sizeof(vm_dirty_bytes),
1308                 .mode           = 0644,
1309                 .proc_handler   = dirty_bytes_handler,
1310                 .extra1         = &dirty_bytes_min,
1311         },
1312         {
1313                 .procname       = "dirty_writeback_centisecs",
1314                 .data           = &dirty_writeback_interval,
1315                 .maxlen         = sizeof(dirty_writeback_interval),
1316                 .mode           = 0644,
1317                 .proc_handler   = dirty_writeback_centisecs_handler,
1318         },
1319         {
1320                 .procname       = "dirty_expire_centisecs",
1321                 .data           = &dirty_expire_interval,
1322                 .maxlen         = sizeof(dirty_expire_interval),
1323                 .mode           = 0644,
1324                 .proc_handler   = proc_dointvec_minmax,
1325                 .extra1         = &zero,
1326         },
1327         {
1328                 .procname       = "dirtytime_expire_seconds",
1329                 .data           = &dirtytime_expire_interval,
1330                 .maxlen         = sizeof(dirty_expire_interval),
1331                 .mode           = 0644,
1332                 .proc_handler   = dirtytime_interval_handler,
1333                 .extra1         = &zero,
1334         },
1335         {
1336                 .procname       = "nr_pdflush_threads",
1337                 .mode           = 0444 /* read-only */,
1338                 .proc_handler   = pdflush_proc_obsolete,
1339         },
1340         {
1341                 .procname       = "swappiness",
1342                 .data           = &vm_swappiness,
1343                 .maxlen         = sizeof(vm_swappiness),
1344                 .mode           = 0644,
1345                 .proc_handler   = proc_dointvec_minmax,
1346                 .extra1         = &zero,
1347                 .extra2         = &one_hundred,
1348         },
1349 #ifdef CONFIG_HUGETLB_PAGE
1350         {
1351                 .procname       = "nr_hugepages",
1352                 .data           = NULL,
1353                 .maxlen         = sizeof(unsigned long),
1354                 .mode           = 0644,
1355                 .proc_handler   = hugetlb_sysctl_handler,
1356         },
1357 #ifdef CONFIG_NUMA
1358         {
1359                 .procname       = "nr_hugepages_mempolicy",
1360                 .data           = NULL,
1361                 .maxlen         = sizeof(unsigned long),
1362                 .mode           = 0644,
1363                 .proc_handler   = &hugetlb_mempolicy_sysctl_handler,
1364         },
1365 #endif
1366          {
1367                 .procname       = "hugetlb_shm_group",
1368                 .data           = &sysctl_hugetlb_shm_group,
1369                 .maxlen         = sizeof(gid_t),
1370                 .mode           = 0644,
1371                 .proc_handler   = proc_dointvec,
1372          },
1373          {
1374                 .procname       = "hugepages_treat_as_movable",
1375                 .data           = &hugepages_treat_as_movable,
1376                 .maxlen         = sizeof(int),
1377                 .mode           = 0644,
1378                 .proc_handler   = proc_dointvec,
1379         },
1380         {
1381                 .procname       = "nr_overcommit_hugepages",
1382                 .data           = NULL,
1383                 .maxlen         = sizeof(unsigned long),
1384                 .mode           = 0644,
1385                 .proc_handler   = hugetlb_overcommit_handler,
1386         },
1387 #endif
1388         {
1389                 .procname       = "lowmem_reserve_ratio",
1390                 .data           = &sysctl_lowmem_reserve_ratio,
1391                 .maxlen         = sizeof(sysctl_lowmem_reserve_ratio),
1392                 .mode           = 0644,
1393                 .proc_handler   = lowmem_reserve_ratio_sysctl_handler,
1394         },
1395         {
1396                 .procname       = "drop_caches",
1397                 .data           = &sysctl_drop_caches,
1398                 .maxlen         = sizeof(int),
1399                 .mode           = 0644,
1400                 .proc_handler   = drop_caches_sysctl_handler,
1401                 .extra1         = &one,
1402                 .extra2         = &four,
1403         },
1404 #ifdef CONFIG_COMPACTION
1405         {
1406                 .procname       = "compact_memory",
1407                 .data           = &sysctl_compact_memory,
1408                 .maxlen         = sizeof(int),
1409                 .mode           = 0200,
1410                 .proc_handler   = sysctl_compaction_handler,
1411         },
1412         {
1413                 .procname       = "extfrag_threshold",
1414                 .data           = &sysctl_extfrag_threshold,
1415                 .maxlen         = sizeof(int),
1416                 .mode           = 0644,
1417                 .proc_handler   = sysctl_extfrag_handler,
1418                 .extra1         = &min_extfrag_threshold,
1419                 .extra2         = &max_extfrag_threshold,
1420         },
1421         {
1422                 .procname       = "compact_unevictable_allowed",
1423                 .data           = &sysctl_compact_unevictable_allowed,
1424                 .maxlen         = sizeof(int),
1425                 .mode           = 0644,
1426                 .proc_handler   = proc_dointvec,
1427                 .extra1         = &zero,
1428                 .extra2         = &one,
1429         },
1430
1431 #endif /* CONFIG_COMPACTION */
1432         {
1433                 .procname       = "min_free_kbytes",
1434                 .data           = &min_free_kbytes,
1435                 .maxlen         = sizeof(min_free_kbytes),
1436                 .mode           = 0644,
1437                 .proc_handler   = min_free_kbytes_sysctl_handler,
1438                 .extra1         = &zero,
1439         },
1440         {
1441                 .procname       = "watermark_scale_factor",
1442                 .data           = &watermark_scale_factor,
1443                 .maxlen         = sizeof(watermark_scale_factor),
1444                 .mode           = 0644,
1445                 .proc_handler   = watermark_scale_factor_sysctl_handler,
1446                 .extra1         = &one,
1447                 .extra2         = &one_thousand,
1448         },
1449         {
1450                 .procname       = "percpu_pagelist_fraction",
1451                 .data           = &percpu_pagelist_fraction,
1452                 .maxlen         = sizeof(percpu_pagelist_fraction),
1453                 .mode           = 0644,
1454                 .proc_handler   = percpu_pagelist_fraction_sysctl_handler,
1455                 .extra1         = &zero,
1456         },
1457 #ifdef CONFIG_MMU
1458         {
1459                 .procname       = "max_map_count",
1460                 .data           = &sysctl_max_map_count,
1461                 .maxlen         = sizeof(sysctl_max_map_count),
1462                 .mode           = 0644,
1463                 .proc_handler   = proc_dointvec_minmax,
1464                 .extra1         = &zero,
1465         },
1466 #else
1467         {
1468                 .procname       = "nr_trim_pages",
1469                 .data           = &sysctl_nr_trim_pages,
1470                 .maxlen         = sizeof(sysctl_nr_trim_pages),
1471                 .mode           = 0644,
1472                 .proc_handler   = proc_dointvec_minmax,
1473                 .extra1         = &zero,
1474         },
1475 #endif
1476         {
1477                 .procname       = "laptop_mode",
1478                 .data           = &laptop_mode,
1479                 .maxlen         = sizeof(laptop_mode),
1480                 .mode           = 0644,
1481                 .proc_handler   = proc_dointvec_jiffies,
1482         },
1483         {
1484                 .procname       = "block_dump",
1485                 .data           = &block_dump,
1486                 .maxlen         = sizeof(block_dump),
1487                 .mode           = 0644,
1488                 .proc_handler   = proc_dointvec,
1489                 .extra1         = &zero,
1490         },
1491         {
1492                 .procname       = "vfs_cache_pressure",
1493                 .data           = &sysctl_vfs_cache_pressure,
1494                 .maxlen         = sizeof(sysctl_vfs_cache_pressure),
1495                 .mode           = 0644,
1496                 .proc_handler   = proc_dointvec,
1497                 .extra1         = &zero,
1498         },
1499 #ifdef HAVE_ARCH_PICK_MMAP_LAYOUT
1500         {
1501                 .procname       = "legacy_va_layout",
1502                 .data           = &sysctl_legacy_va_layout,
1503                 .maxlen         = sizeof(sysctl_legacy_va_layout),
1504                 .mode           = 0644,
1505                 .proc_handler   = proc_dointvec,
1506                 .extra1         = &zero,
1507         },
1508 #endif
1509 #ifdef CONFIG_NUMA
1510         {
1511                 .procname       = "zone_reclaim_mode",
1512                 .data           = &node_reclaim_mode,
1513                 .maxlen         = sizeof(node_reclaim_mode),
1514                 .mode           = 0644,
1515                 .proc_handler   = proc_dointvec,
1516                 .extra1         = &zero,
1517         },
1518         {
1519                 .procname       = "min_unmapped_ratio",
1520                 .data           = &sysctl_min_unmapped_ratio,
1521                 .maxlen         = sizeof(sysctl_min_unmapped_ratio),
1522                 .mode           = 0644,
1523                 .proc_handler   = sysctl_min_unmapped_ratio_sysctl_handler,
1524                 .extra1         = &zero,
1525                 .extra2         = &one_hundred,
1526         },
1527         {
1528                 .procname       = "min_slab_ratio",
1529                 .data           = &sysctl_min_slab_ratio,
1530                 .maxlen         = sizeof(sysctl_min_slab_ratio),
1531                 .mode           = 0644,
1532                 .proc_handler   = sysctl_min_slab_ratio_sysctl_handler,
1533                 .extra1         = &zero,
1534                 .extra2         = &one_hundred,
1535         },
1536 #endif
1537 #ifdef CONFIG_SMP
1538         {
1539                 .procname       = "stat_interval",
1540                 .data           = &sysctl_stat_interval,
1541                 .maxlen         = sizeof(sysctl_stat_interval),
1542                 .mode           = 0644,
1543                 .proc_handler   = proc_dointvec_jiffies,
1544         },
1545         {
1546                 .procname       = "stat_refresh",
1547                 .data           = NULL,
1548                 .maxlen         = 0,
1549                 .mode           = 0600,
1550                 .proc_handler   = vmstat_refresh,
1551         },
1552 #endif
1553 #ifdef CONFIG_MMU
1554         {
1555                 .procname       = "mmap_min_addr",
1556                 .data           = &dac_mmap_min_addr,
1557                 .maxlen         = sizeof(unsigned long),
1558                 .mode           = 0644,
1559                 .proc_handler   = mmap_min_addr_handler,
1560         },
1561 #endif
1562 #ifdef CONFIG_NUMA
1563         {
1564                 .procname       = "numa_zonelist_order",
1565                 .data           = &numa_zonelist_order,
1566                 .maxlen         = NUMA_ZONELIST_ORDER_LEN,
1567                 .mode           = 0644,
1568                 .proc_handler   = numa_zonelist_order_handler,
1569         },
1570 #endif
1571 #if (defined(CONFIG_X86_32) && !defined(CONFIG_UML))|| \
1572    (defined(CONFIG_SUPERH) && defined(CONFIG_VSYSCALL))
1573         {
1574                 .procname       = "vdso_enabled",
1575 #ifdef CONFIG_X86_32
1576                 .data           = &vdso32_enabled,
1577                 .maxlen         = sizeof(vdso32_enabled),
1578 #else
1579                 .data           = &vdso_enabled,
1580                 .maxlen         = sizeof(vdso_enabled),
1581 #endif
1582                 .mode           = 0644,
1583                 .proc_handler   = proc_dointvec,
1584                 .extra1         = &zero,
1585         },
1586 #endif
1587 #ifdef CONFIG_HIGHMEM
1588         {
1589                 .procname       = "highmem_is_dirtyable",
1590                 .data           = &vm_highmem_is_dirtyable,
1591                 .maxlen         = sizeof(vm_highmem_is_dirtyable),
1592                 .mode           = 0644,
1593                 .proc_handler   = proc_dointvec_minmax,
1594                 .extra1         = &zero,
1595                 .extra2         = &one,
1596         },
1597 #endif
1598 #ifdef CONFIG_MEMORY_FAILURE
1599         {
1600                 .procname       = "memory_failure_early_kill",
1601                 .data           = &sysctl_memory_failure_early_kill,
1602                 .maxlen         = sizeof(sysctl_memory_failure_early_kill),
1603                 .mode           = 0644,
1604                 .proc_handler   = proc_dointvec_minmax,
1605                 .extra1         = &zero,
1606                 .extra2         = &one,
1607         },
1608         {
1609                 .procname       = "memory_failure_recovery",
1610                 .data           = &sysctl_memory_failure_recovery,
1611                 .maxlen         = sizeof(sysctl_memory_failure_recovery),
1612                 .mode           = 0644,
1613                 .proc_handler   = proc_dointvec_minmax,
1614                 .extra1         = &zero,
1615                 .extra2         = &one,
1616         },
1617 #endif
1618         {
1619                 .procname       = "user_reserve_kbytes",
1620                 .data           = &sysctl_user_reserve_kbytes,
1621                 .maxlen         = sizeof(sysctl_user_reserve_kbytes),
1622                 .mode           = 0644,
1623                 .proc_handler   = proc_doulongvec_minmax,
1624         },
1625         {
1626                 .procname       = "admin_reserve_kbytes",
1627                 .data           = &sysctl_admin_reserve_kbytes,
1628                 .maxlen         = sizeof(sysctl_admin_reserve_kbytes),
1629                 .mode           = 0644,
1630                 .proc_handler   = proc_doulongvec_minmax,
1631         },
1632 #ifdef CONFIG_HAVE_ARCH_MMAP_RND_BITS
1633         {
1634                 .procname       = "mmap_rnd_bits",
1635                 .data           = &mmap_rnd_bits,
1636                 .maxlen         = sizeof(mmap_rnd_bits),
1637                 .mode           = 0600,
1638                 .proc_handler   = proc_dointvec_minmax,
1639                 .extra1         = (void *)&mmap_rnd_bits_min,
1640                 .extra2         = (void *)&mmap_rnd_bits_max,
1641         },
1642 #endif
1643 #ifdef CONFIG_HAVE_ARCH_MMAP_RND_COMPAT_BITS
1644         {
1645                 .procname       = "mmap_rnd_compat_bits",
1646                 .data           = &mmap_rnd_compat_bits,
1647                 .maxlen         = sizeof(mmap_rnd_compat_bits),
1648                 .mode           = 0600,
1649                 .proc_handler   = proc_dointvec_minmax,
1650                 .extra1         = (void *)&mmap_rnd_compat_bits_min,
1651                 .extra2         = (void *)&mmap_rnd_compat_bits_max,
1652         },
1653 #endif
1654         { }
1655 };
1656
1657 static struct ctl_table fs_table[] = {
1658         {
1659                 .procname       = "inode-nr",
1660                 .data           = &inodes_stat,
1661                 .maxlen         = 2*sizeof(long),
1662                 .mode           = 0444,
1663                 .proc_handler   = proc_nr_inodes,
1664         },
1665         {
1666                 .procname       = "inode-state",
1667                 .data           = &inodes_stat,
1668                 .maxlen         = 7*sizeof(long),
1669                 .mode           = 0444,
1670                 .proc_handler   = proc_nr_inodes,
1671         },
1672         {
1673                 .procname       = "file-nr",
1674                 .data           = &files_stat,
1675                 .maxlen         = sizeof(files_stat),
1676                 .mode           = 0444,
1677                 .proc_handler   = proc_nr_files,
1678         },
1679         {
1680                 .procname       = "file-max",
1681                 .data           = &files_stat.max_files,
1682                 .maxlen         = sizeof(files_stat.max_files),
1683                 .mode           = 0644,
1684                 .proc_handler   = proc_doulongvec_minmax,
1685         },
1686         {
1687                 .procname       = "nr_open",
1688                 .data           = &sysctl_nr_open,
1689                 .maxlen         = sizeof(unsigned int),
1690                 .mode           = 0644,
1691                 .proc_handler   = proc_dointvec_minmax,
1692                 .extra1         = &sysctl_nr_open_min,
1693                 .extra2         = &sysctl_nr_open_max,
1694         },
1695         {
1696                 .procname       = "dentry-state",
1697                 .data           = &dentry_stat,
1698                 .maxlen         = 6*sizeof(long),
1699                 .mode           = 0444,
1700                 .proc_handler   = proc_nr_dentry,
1701         },
1702         {
1703                 .procname       = "overflowuid",
1704                 .data           = &fs_overflowuid,
1705                 .maxlen         = sizeof(int),
1706                 .mode           = 0644,
1707                 .proc_handler   = proc_dointvec_minmax,
1708                 .extra1         = &minolduid,
1709                 .extra2         = &maxolduid,
1710         },
1711         {
1712                 .procname       = "overflowgid",
1713                 .data           = &fs_overflowgid,
1714                 .maxlen         = sizeof(int),
1715                 .mode           = 0644,
1716                 .proc_handler   = proc_dointvec_minmax,
1717                 .extra1         = &minolduid,
1718                 .extra2         = &maxolduid,
1719         },
1720 #ifdef CONFIG_FILE_LOCKING
1721         {
1722                 .procname       = "leases-enable",
1723                 .data           = &leases_enable,
1724                 .maxlen         = sizeof(int),
1725                 .mode           = 0644,
1726                 .proc_handler   = proc_dointvec,
1727         },
1728 #endif
1729 #ifdef CONFIG_DNOTIFY
1730         {
1731                 .procname       = "dir-notify-enable",
1732                 .data           = &dir_notify_enable,
1733                 .maxlen         = sizeof(int),
1734                 .mode           = 0644,
1735                 .proc_handler   = proc_dointvec,
1736         },
1737 #endif
1738 #ifdef CONFIG_MMU
1739 #ifdef CONFIG_FILE_LOCKING
1740         {
1741                 .procname       = "lease-break-time",
1742                 .data           = &lease_break_time,
1743                 .maxlen         = sizeof(int),
1744                 .mode           = 0644,
1745                 .proc_handler   = proc_dointvec,
1746         },
1747 #endif
1748 #ifdef CONFIG_AIO
1749         {
1750                 .procname       = "aio-nr",
1751                 .data           = &aio_nr,
1752                 .maxlen         = sizeof(aio_nr),
1753                 .mode           = 0444,
1754                 .proc_handler   = proc_doulongvec_minmax,
1755         },
1756         {
1757                 .procname       = "aio-max-nr",
1758                 .data           = &aio_max_nr,
1759                 .maxlen         = sizeof(aio_max_nr),
1760                 .mode           = 0644,
1761                 .proc_handler   = proc_doulongvec_minmax,
1762         },
1763 #endif /* CONFIG_AIO */
1764 #ifdef CONFIG_INOTIFY_USER
1765         {
1766                 .procname       = "inotify",
1767                 .mode           = 0555,
1768                 .child          = inotify_table,
1769         },
1770 #endif
1771 #ifdef CONFIG_EPOLL
1772         {
1773                 .procname       = "epoll",
1774                 .mode           = 0555,
1775                 .child          = epoll_table,
1776         },
1777 #endif
1778 #endif
1779         {
1780                 .procname       = "protected_symlinks",
1781                 .data           = &sysctl_protected_symlinks,
1782                 .maxlen         = sizeof(int),
1783                 .mode           = 0600,
1784                 .proc_handler   = proc_dointvec_minmax,
1785                 .extra1         = &zero,
1786                 .extra2         = &one,
1787         },
1788         {
1789                 .procname       = "protected_hardlinks",
1790                 .data           = &sysctl_protected_hardlinks,
1791                 .maxlen         = sizeof(int),
1792                 .mode           = 0600,
1793                 .proc_handler   = proc_dointvec_minmax,
1794                 .extra1         = &zero,
1795                 .extra2         = &one,
1796         },
1797         {
1798                 .procname       = "protected_fifos",
1799                 .data           = &sysctl_protected_fifos,
1800                 .maxlen         = sizeof(int),
1801                 .mode           = 0600,
1802                 .proc_handler   = proc_dointvec_minmax,
1803                 .extra1         = &zero,
1804                 .extra2         = &two,
1805         },
1806         {
1807                 .procname       = "protected_regular",
1808                 .data           = &sysctl_protected_regular,
1809                 .maxlen         = sizeof(int),
1810                 .mode           = 0600,
1811                 .proc_handler   = proc_dointvec_minmax,
1812                 .extra1         = &zero,
1813                 .extra2         = &two,
1814         },
1815         {
1816                 .procname       = "suid_dumpable",
1817                 .data           = &suid_dumpable,
1818                 .maxlen         = sizeof(int),
1819                 .mode           = 0644,
1820                 .proc_handler   = proc_dointvec_minmax_coredump,
1821                 .extra1         = &zero,
1822                 .extra2         = &two,
1823         },
1824 #if defined(CONFIG_BINFMT_MISC) || defined(CONFIG_BINFMT_MISC_MODULE)
1825         {
1826                 .procname       = "binfmt_misc",
1827                 .mode           = 0555,
1828                 .child          = sysctl_mount_point,
1829         },
1830 #endif
1831         {
1832                 .procname       = "pipe-max-size",
1833                 .data           = &pipe_max_size,
1834                 .maxlen         = sizeof(int),
1835                 .mode           = 0644,
1836                 .proc_handler   = &pipe_proc_fn,
1837                 .extra1         = &pipe_min_size,
1838         },
1839         {
1840                 .procname       = "pipe-user-pages-hard",
1841                 .data           = &pipe_user_pages_hard,
1842                 .maxlen         = sizeof(pipe_user_pages_hard),
1843                 .mode           = 0644,
1844                 .proc_handler   = proc_doulongvec_minmax,
1845         },
1846         {
1847                 .procname       = "pipe-user-pages-soft",
1848                 .data           = &pipe_user_pages_soft,
1849                 .maxlen         = sizeof(pipe_user_pages_soft),
1850                 .mode           = 0644,
1851                 .proc_handler   = proc_doulongvec_minmax,
1852         },
1853         {
1854                 .procname       = "mount-max",
1855                 .data           = &sysctl_mount_max,
1856                 .maxlen         = sizeof(unsigned int),
1857                 .mode           = 0644,
1858                 .proc_handler   = proc_dointvec_minmax,
1859                 .extra1         = &one,
1860         },
1861         { }
1862 };
1863
1864 static struct ctl_table debug_table[] = {
1865 #ifdef CONFIG_SYSCTL_EXCEPTION_TRACE
1866         {
1867                 .procname       = "exception-trace",
1868                 .data           = &show_unhandled_signals,
1869                 .maxlen         = sizeof(int),
1870                 .mode           = 0644,
1871                 .proc_handler   = proc_dointvec
1872         },
1873 #endif
1874 #if defined(CONFIG_OPTPROBES)
1875         {
1876                 .procname       = "kprobes-optimization",
1877                 .data           = &sysctl_kprobes_optimization,
1878                 .maxlen         = sizeof(int),
1879                 .mode           = 0644,
1880                 .proc_handler   = proc_kprobes_optimization_handler,
1881                 .extra1         = &zero,
1882                 .extra2         = &one,
1883         },
1884 #endif
1885         { }
1886 };
1887
1888 static struct ctl_table dev_table[] = {
1889         { }
1890 };
1891
1892 int __init sysctl_init(void)
1893 {
1894         struct ctl_table_header *hdr;
1895
1896         hdr = register_sysctl_table(sysctl_base_table);
1897         kmemleak_not_leak(hdr);
1898         return 0;
1899 }
1900
1901 #endif /* CONFIG_SYSCTL */
1902
1903 /*
1904  * /proc/sys support
1905  */
1906
1907 #ifdef CONFIG_PROC_SYSCTL
1908
1909 static int _proc_do_string(char *data, int maxlen, int write,
1910                            char __user *buffer,
1911                            size_t *lenp, loff_t *ppos)
1912 {
1913         size_t len;
1914         char __user *p;
1915         char c;
1916
1917         if (!data || !maxlen || !*lenp) {
1918                 *lenp = 0;
1919                 return 0;
1920         }
1921
1922         if (write) {
1923                 if (sysctl_writes_strict == SYSCTL_WRITES_STRICT) {
1924                         /* Only continue writes not past the end of buffer. */
1925                         len = strlen(data);
1926                         if (len > maxlen - 1)
1927                                 len = maxlen - 1;
1928
1929                         if (*ppos > len)
1930                                 return 0;
1931                         len = *ppos;
1932                 } else {
1933                         /* Start writing from beginning of buffer. */
1934                         len = 0;
1935                 }
1936
1937                 *ppos += *lenp;
1938                 p = buffer;
1939                 while ((p - buffer) < *lenp && len < maxlen - 1) {
1940                         if (get_user(c, p++))
1941                                 return -EFAULT;
1942                         if (c == 0 || c == '\n')
1943                                 break;
1944                         data[len++] = c;
1945                 }
1946                 data[len] = 0;
1947         } else {
1948                 len = strlen(data);
1949                 if (len > maxlen)
1950                         len = maxlen;
1951
1952                 if (*ppos > len) {
1953                         *lenp = 0;
1954                         return 0;
1955                 }
1956
1957                 data += *ppos;
1958                 len  -= *ppos;
1959
1960                 if (len > *lenp)
1961                         len = *lenp;
1962                 if (len)
1963                         if (copy_to_user(buffer, data, len))
1964                                 return -EFAULT;
1965                 if (len < *lenp) {
1966                         if (put_user('\n', buffer + len))
1967                                 return -EFAULT;
1968                         len++;
1969                 }
1970                 *lenp = len;
1971                 *ppos += len;
1972         }
1973         return 0;
1974 }
1975
1976 static void warn_sysctl_write(struct ctl_table *table)
1977 {
1978         pr_warn_once("%s wrote to %s when file position was not 0!\n"
1979                 "This will not be supported in the future. To silence this\n"
1980                 "warning, set kernel.sysctl_writes_strict = -1\n",
1981                 current->comm, table->procname);
1982 }
1983
1984 /**
1985  * proc_dostring - read a string sysctl
1986  * @table: the sysctl table
1987  * @write: %TRUE if this is a write to the sysctl file
1988  * @buffer: the user buffer
1989  * @lenp: the size of the user buffer
1990  * @ppos: file position
1991  *
1992  * Reads/writes a string from/to the user buffer. If the kernel
1993  * buffer provided is not large enough to hold the string, the
1994  * string is truncated. The copied string is %NULL-terminated.
1995  * If the string is being read by the user process, it is copied
1996  * and a newline '\n' is added. It is truncated if the buffer is
1997  * not large enough.
1998  *
1999  * Returns 0 on success.
2000  */
2001 int proc_dostring(struct ctl_table *table, int write,
2002                   void __user *buffer, size_t *lenp, loff_t *ppos)
2003 {
2004         if (write && *ppos && sysctl_writes_strict == SYSCTL_WRITES_WARN)
2005                 warn_sysctl_write(table);
2006
2007         return _proc_do_string((char *)(table->data), table->maxlen, write,
2008                                (char __user *)buffer, lenp, ppos);
2009 }
2010
2011 static size_t proc_skip_spaces(char **buf)
2012 {
2013         size_t ret;
2014         char *tmp = skip_spaces(*buf);
2015         ret = tmp - *buf;
2016         *buf = tmp;
2017         return ret;
2018 }
2019
2020 static void proc_skip_char(char **buf, size_t *size, const char v)
2021 {
2022         while (*size) {
2023                 if (**buf != v)
2024                         break;
2025                 (*size)--;
2026                 (*buf)++;
2027         }
2028 }
2029
2030 #define TMPBUFLEN 22
2031 /**
2032  * proc_get_long - reads an ASCII formatted integer from a user buffer
2033  *
2034  * @buf: a kernel buffer
2035  * @size: size of the kernel buffer
2036  * @val: this is where the number will be stored
2037  * @neg: set to %TRUE if number is negative
2038  * @perm_tr: a vector which contains the allowed trailers
2039  * @perm_tr_len: size of the perm_tr vector
2040  * @tr: pointer to store the trailer character
2041  *
2042  * In case of success %0 is returned and @buf and @size are updated with
2043  * the amount of bytes read. If @tr is non-NULL and a trailing
2044  * character exists (size is non-zero after returning from this
2045  * function), @tr is updated with the trailing character.
2046  */
2047 static int proc_get_long(char **buf, size_t *size,
2048                           unsigned long *val, bool *neg,
2049                           const char *perm_tr, unsigned perm_tr_len, char *tr)
2050 {
2051         int len;
2052         char *p, tmp[TMPBUFLEN];
2053
2054         if (!*size)
2055                 return -EINVAL;
2056
2057         len = *size;
2058         if (len > TMPBUFLEN - 1)
2059                 len = TMPBUFLEN - 1;
2060
2061         memcpy(tmp, *buf, len);
2062
2063         tmp[len] = 0;
2064         p = tmp;
2065         if (*p == '-' && *size > 1) {
2066                 *neg = true;
2067                 p++;
2068         } else
2069                 *neg = false;
2070         if (!isdigit(*p))
2071                 return -EINVAL;
2072
2073         *val = simple_strtoul(p, &p, 0);
2074
2075         len = p - tmp;
2076
2077         /* We don't know if the next char is whitespace thus we may accept
2078          * invalid integers (e.g. 1234...a) or two integers instead of one
2079          * (e.g. 123...1). So lets not allow such large numbers. */
2080         if (len == TMPBUFLEN - 1)
2081                 return -EINVAL;
2082
2083         if (len < *size && perm_tr_len && !memchr(perm_tr, *p, perm_tr_len))
2084                 return -EINVAL;
2085
2086         if (tr && (len < *size))
2087                 *tr = *p;
2088
2089         *buf += len;
2090         *size -= len;
2091
2092         return 0;
2093 }
2094
2095 /**
2096  * proc_put_long - converts an integer to a decimal ASCII formatted string
2097  *
2098  * @buf: the user buffer
2099  * @size: the size of the user buffer
2100  * @val: the integer to be converted
2101  * @neg: sign of the number, %TRUE for negative
2102  *
2103  * In case of success %0 is returned and @buf and @size are updated with
2104  * the amount of bytes written.
2105  */
2106 static int proc_put_long(void __user **buf, size_t *size, unsigned long val,
2107                           bool neg)
2108 {
2109         int len;
2110         char tmp[TMPBUFLEN], *p = tmp;
2111
2112         sprintf(p, "%s%lu", neg ? "-" : "", val);
2113         len = strlen(tmp);
2114         if (len > *size)
2115                 len = *size;
2116         if (copy_to_user(*buf, tmp, len))
2117                 return -EFAULT;
2118         *size -= len;
2119         *buf += len;
2120         return 0;
2121 }
2122 #undef TMPBUFLEN
2123
2124 static int proc_put_char(void __user **buf, size_t *size, char c)
2125 {
2126         if (*size) {
2127                 char __user **buffer = (char __user **)buf;
2128                 if (put_user(c, *buffer))
2129                         return -EFAULT;
2130                 (*size)--, (*buffer)++;
2131                 *buf = *buffer;
2132         }
2133         return 0;
2134 }
2135
2136 static int do_proc_dointvec_conv(bool *negp, unsigned long *lvalp,
2137                                  int *valp,
2138                                  int write, void *data)
2139 {
2140         if (write) {
2141                 if (*negp) {
2142                         if (*lvalp > (unsigned long) INT_MAX + 1)
2143                                 return -EINVAL;
2144                         *valp = -*lvalp;
2145                 } else {
2146                         if (*lvalp > (unsigned long) INT_MAX)
2147                                 return -EINVAL;
2148                         *valp = *lvalp;
2149                 }
2150         } else {
2151                 int val = *valp;
2152                 if (val < 0) {
2153                         *negp = true;
2154                         *lvalp = -(unsigned long)val;
2155                 } else {
2156                         *negp = false;
2157                         *lvalp = (unsigned long)val;
2158                 }
2159         }
2160         return 0;
2161 }
2162
2163 static int do_proc_douintvec_conv(bool *negp, unsigned long *lvalp,
2164                                  int *valp,
2165                                  int write, void *data)
2166 {
2167         if (write) {
2168                 if (*negp)
2169                         return -EINVAL;
2170                 if (*lvalp > UINT_MAX)
2171                         return -EINVAL;
2172                 *valp = *lvalp;
2173         } else {
2174                 unsigned int val = *valp;
2175                 *negp = false;
2176                 *lvalp = (unsigned long)val;
2177         }
2178         return 0;
2179 }
2180
2181 static const char proc_wspace_sep[] = { ' ', '\t', '\n' };
2182
2183 static int __do_proc_dointvec(void *tbl_data, struct ctl_table *table,
2184                   int write, void __user *buffer,
2185                   size_t *lenp, loff_t *ppos,
2186                   int (*conv)(bool *negp, unsigned long *lvalp, int *valp,
2187                               int write, void *data),
2188                   void *data)
2189 {
2190         int *i, vleft, first = 1, err = 0;
2191         size_t left;
2192         char *kbuf = NULL, *p;
2193
2194         if (!tbl_data || !table->maxlen || !*lenp || (*ppos && !write)) {
2195                 *lenp = 0;
2196                 return 0;
2197         }
2198
2199         i = (int *) tbl_data;
2200         vleft = table->maxlen / sizeof(*i);
2201         left = *lenp;
2202
2203         if (!conv)
2204                 conv = do_proc_dointvec_conv;
2205
2206         if (write) {
2207                 if (*ppos) {
2208                         switch (sysctl_writes_strict) {
2209                         case SYSCTL_WRITES_STRICT:
2210                                 goto out;
2211                         case SYSCTL_WRITES_WARN:
2212                                 warn_sysctl_write(table);
2213                                 break;
2214                         default:
2215                                 break;
2216                         }
2217                 }
2218
2219                 if (left > PAGE_SIZE - 1)
2220                         left = PAGE_SIZE - 1;
2221                 p = kbuf = memdup_user_nul(buffer, left);
2222                 if (IS_ERR(kbuf))
2223                         return PTR_ERR(kbuf);
2224         }
2225
2226         for (; left && vleft--; i++, first=0) {
2227                 unsigned long lval;
2228                 bool neg;
2229
2230                 if (write) {
2231                         left -= proc_skip_spaces(&p);
2232
2233                         if (!left)
2234                                 break;
2235                         err = proc_get_long(&p, &left, &lval, &neg,
2236                                              proc_wspace_sep,
2237                                              sizeof(proc_wspace_sep), NULL);
2238                         if (err)
2239                                 break;
2240                         if (conv(&neg, &lval, i, 1, data)) {
2241                                 err = -EINVAL;
2242                                 break;
2243                         }
2244                 } else {
2245                         if (conv(&neg, &lval, i, 0, data)) {
2246                                 err = -EINVAL;
2247                                 break;
2248                         }
2249                         if (!first)
2250                                 err = proc_put_char(&buffer, &left, '\t');
2251                         if (err)
2252                                 break;
2253                         err = proc_put_long(&buffer, &left, lval, neg);
2254                         if (err)
2255                                 break;
2256                 }
2257         }
2258
2259         if (!write && !first && left && !err)
2260                 err = proc_put_char(&buffer, &left, '\n');
2261         if (write && !err && left)
2262                 left -= proc_skip_spaces(&p);
2263         if (write) {
2264                 kfree(kbuf);
2265                 if (first)
2266                         return err ? : -EINVAL;
2267         }
2268         *lenp -= left;
2269 out:
2270         *ppos += *lenp;
2271         return err;
2272 }
2273
2274 static int do_proc_dointvec(struct ctl_table *table, int write,
2275                   void __user *buffer, size_t *lenp, loff_t *ppos,
2276                   int (*conv)(bool *negp, unsigned long *lvalp, int *valp,
2277                               int write, void *data),
2278                   void *data)
2279 {
2280         return __do_proc_dointvec(table->data, table, write,
2281                         buffer, lenp, ppos, conv, data);
2282 }
2283
2284 /**
2285  * proc_dointvec - read a vector of integers
2286  * @table: the sysctl table
2287  * @write: %TRUE if this is a write to the sysctl file
2288  * @buffer: the user buffer
2289  * @lenp: the size of the user buffer
2290  * @ppos: file position
2291  *
2292  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2293  * values from/to the user buffer, treated as an ASCII string.
2294  *
2295  * Returns 0 on success.
2296  */
2297 int proc_dointvec(struct ctl_table *table, int write,
2298                      void __user *buffer, size_t *lenp, loff_t *ppos)
2299 {
2300         return do_proc_dointvec(table, write, buffer, lenp, ppos, NULL, NULL);
2301 }
2302
2303 /**
2304  * proc_douintvec - read a vector of unsigned integers
2305  * @table: the sysctl table
2306  * @write: %TRUE if this is a write to the sysctl file
2307  * @buffer: the user buffer
2308  * @lenp: the size of the user buffer
2309  * @ppos: file position
2310  *
2311  * Reads/writes up to table->maxlen/sizeof(unsigned int) unsigned integer
2312  * values from/to the user buffer, treated as an ASCII string.
2313  *
2314  * Returns 0 on success.
2315  */
2316 int proc_douintvec(struct ctl_table *table, int write,
2317                      void __user *buffer, size_t *lenp, loff_t *ppos)
2318 {
2319         return do_proc_dointvec(table, write, buffer, lenp, ppos,
2320                                 do_proc_douintvec_conv, NULL);
2321 }
2322
2323 /*
2324  * Taint values can only be increased
2325  * This means we can safely use a temporary.
2326  */
2327 static int proc_taint(struct ctl_table *table, int write,
2328                                void __user *buffer, size_t *lenp, loff_t *ppos)
2329 {
2330         struct ctl_table t;
2331         unsigned long tmptaint = get_taint();
2332         int err;
2333
2334         if (write && !capable(CAP_SYS_ADMIN))
2335                 return -EPERM;
2336
2337         t = *table;
2338         t.data = &tmptaint;
2339         err = proc_doulongvec_minmax(&t, write, buffer, lenp, ppos);
2340         if (err < 0)
2341                 return err;
2342
2343         if (write) {
2344                 /*
2345                  * Poor man's atomic or. Not worth adding a primitive
2346                  * to everyone's atomic.h for this
2347                  */
2348                 int i;
2349                 for (i = 0; i < BITS_PER_LONG && tmptaint >> i; i++) {
2350                         if ((tmptaint >> i) & 1)
2351                                 add_taint(i, LOCKDEP_STILL_OK);
2352                 }
2353         }
2354
2355         return err;
2356 }
2357
2358 #ifdef CONFIG_PRINTK
2359 static int proc_dointvec_minmax_sysadmin(struct ctl_table *table, int write,
2360                                 void __user *buffer, size_t *lenp, loff_t *ppos)
2361 {
2362         if (write && !capable(CAP_SYS_ADMIN))
2363                 return -EPERM;
2364
2365         return proc_dointvec_minmax(table, write, buffer, lenp, ppos);
2366 }
2367 #endif
2368
2369 struct do_proc_dointvec_minmax_conv_param {
2370         int *min;
2371         int *max;
2372 };
2373
2374 static int do_proc_dointvec_minmax_conv(bool *negp, unsigned long *lvalp,
2375                                         int *valp,
2376                                         int write, void *data)
2377 {
2378         struct do_proc_dointvec_minmax_conv_param *param = data;
2379         if (write) {
2380                 int val = *negp ? -*lvalp : *lvalp;
2381                 if ((param->min && *param->min > val) ||
2382                     (param->max && *param->max < val))
2383                         return -EINVAL;
2384                 *valp = val;
2385         } else {
2386                 int val = *valp;
2387                 if (val < 0) {
2388                         *negp = true;
2389                         *lvalp = -(unsigned long)val;
2390                 } else {
2391                         *negp = false;
2392                         *lvalp = (unsigned long)val;
2393                 }
2394         }
2395         return 0;
2396 }
2397
2398 /**
2399  * proc_dointvec_minmax - read a vector of integers with min/max values
2400  * @table: the sysctl table
2401  * @write: %TRUE if this is a write to the sysctl file
2402  * @buffer: the user buffer
2403  * @lenp: the size of the user buffer
2404  * @ppos: file position
2405  *
2406  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2407  * values from/to the user buffer, treated as an ASCII string.
2408  *
2409  * This routine will ensure the values are within the range specified by
2410  * table->extra1 (min) and table->extra2 (max).
2411  *
2412  * Returns 0 on success.
2413  */
2414 int proc_dointvec_minmax(struct ctl_table *table, int write,
2415                   void __user *buffer, size_t *lenp, loff_t *ppos)
2416 {
2417         struct do_proc_dointvec_minmax_conv_param param = {
2418                 .min = (int *) table->extra1,
2419                 .max = (int *) table->extra2,
2420         };
2421         return do_proc_dointvec(table, write, buffer, lenp, ppos,
2422                                 do_proc_dointvec_minmax_conv, &param);
2423 }
2424
2425 static void validate_coredump_safety(void)
2426 {
2427 #ifdef CONFIG_COREDUMP
2428         if (suid_dumpable == SUID_DUMP_ROOT &&
2429             core_pattern[0] != '/' && core_pattern[0] != '|') {
2430                 printk(KERN_WARNING "Unsafe core_pattern used with "\
2431                         "suid_dumpable=2. Pipe handler or fully qualified "\
2432                         "core dump path required.\n");
2433         }
2434 #endif
2435 }
2436
2437 static int proc_dointvec_minmax_coredump(struct ctl_table *table, int write,
2438                 void __user *buffer, size_t *lenp, loff_t *ppos)
2439 {
2440         int error = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
2441         if (!error)
2442                 validate_coredump_safety();
2443         return error;
2444 }
2445
2446 #ifdef CONFIG_COREDUMP
2447 static int proc_dostring_coredump(struct ctl_table *table, int write,
2448                   void __user *buffer, size_t *lenp, loff_t *ppos)
2449 {
2450         int error = proc_dostring(table, write, buffer, lenp, ppos);
2451         if (!error)
2452                 validate_coredump_safety();
2453         return error;
2454 }
2455 #endif
2456
2457 static int __do_proc_doulongvec_minmax(void *data, struct ctl_table *table, int write,
2458                                      void __user *buffer,
2459                                      size_t *lenp, loff_t *ppos,
2460                                      unsigned long convmul,
2461                                      unsigned long convdiv)
2462 {
2463         unsigned long *i, *min, *max;
2464         int vleft, first = 1, err = 0;
2465         size_t left;
2466         char *kbuf = NULL, *p;
2467
2468         if (!data || !table->maxlen || !*lenp || (*ppos && !write)) {
2469                 *lenp = 0;
2470                 return 0;
2471         }
2472
2473         i = (unsigned long *) data;
2474         min = (unsigned long *) table->extra1;
2475         max = (unsigned long *) table->extra2;
2476         vleft = table->maxlen / sizeof(unsigned long);
2477         left = *lenp;
2478
2479         if (write) {
2480                 if (*ppos) {
2481                         switch (sysctl_writes_strict) {
2482                         case SYSCTL_WRITES_STRICT:
2483                                 goto out;
2484                         case SYSCTL_WRITES_WARN:
2485                                 warn_sysctl_write(table);
2486                                 break;
2487                         default:
2488                                 break;
2489                         }
2490                 }
2491
2492                 if (left > PAGE_SIZE - 1)
2493                         left = PAGE_SIZE - 1;
2494                 p = kbuf = memdup_user_nul(buffer, left);
2495                 if (IS_ERR(kbuf))
2496                         return PTR_ERR(kbuf);
2497         }
2498
2499         for (; left && vleft--; i++, first = 0) {
2500                 unsigned long val;
2501
2502                 if (write) {
2503                         bool neg;
2504
2505                         left -= proc_skip_spaces(&p);
2506
2507                         err = proc_get_long(&p, &left, &val, &neg,
2508                                              proc_wspace_sep,
2509                                              sizeof(proc_wspace_sep), NULL);
2510                         if (err)
2511                                 break;
2512                         if (neg)
2513                                 continue;
2514                         val = convmul * val / convdiv;
2515                         if ((min && val < *min) || (max && val > *max))
2516                                 continue;
2517                         *i = val;
2518                 } else {
2519                         val = convdiv * (*i) / convmul;
2520                         if (!first) {
2521                                 err = proc_put_char(&buffer, &left, '\t');
2522                                 if (err)
2523                                         break;
2524                         }
2525                         err = proc_put_long(&buffer, &left, val, false);
2526                         if (err)
2527                                 break;
2528                 }
2529         }
2530
2531         if (!write && !first && left && !err)
2532                 err = proc_put_char(&buffer, &left, '\n');
2533         if (write && !err)
2534                 left -= proc_skip_spaces(&p);
2535         if (write) {
2536                 kfree(kbuf);
2537                 if (first)
2538                         return err ? : -EINVAL;
2539         }
2540         *lenp -= left;
2541 out:
2542         *ppos += *lenp;
2543         return err;
2544 }
2545
2546 static int do_proc_doulongvec_minmax(struct ctl_table *table, int write,
2547                                      void __user *buffer,
2548                                      size_t *lenp, loff_t *ppos,
2549                                      unsigned long convmul,
2550                                      unsigned long convdiv)
2551 {
2552         return __do_proc_doulongvec_minmax(table->data, table, write,
2553                         buffer, lenp, ppos, convmul, convdiv);
2554 }
2555
2556 /**
2557  * proc_doulongvec_minmax - read a vector of long integers with min/max values
2558  * @table: the sysctl table
2559  * @write: %TRUE if this is a write to the sysctl file
2560  * @buffer: the user buffer
2561  * @lenp: the size of the user buffer
2562  * @ppos: file position
2563  *
2564  * Reads/writes up to table->maxlen/sizeof(unsigned long) unsigned long
2565  * values from/to the user buffer, treated as an ASCII string.
2566  *
2567  * This routine will ensure the values are within the range specified by
2568  * table->extra1 (min) and table->extra2 (max).
2569  *
2570  * Returns 0 on success.
2571  */
2572 int proc_doulongvec_minmax(struct ctl_table *table, int write,
2573                            void __user *buffer, size_t *lenp, loff_t *ppos)
2574 {
2575     return do_proc_doulongvec_minmax(table, write, buffer, lenp, ppos, 1l, 1l);
2576 }
2577
2578 /**
2579  * proc_doulongvec_ms_jiffies_minmax - read a vector of millisecond values with min/max values
2580  * @table: the sysctl table
2581  * @write: %TRUE if this is a write to the sysctl file
2582  * @buffer: the user buffer
2583  * @lenp: the size of the user buffer
2584  * @ppos: file position
2585  *
2586  * Reads/writes up to table->maxlen/sizeof(unsigned long) unsigned long
2587  * values from/to the user buffer, treated as an ASCII string. The values
2588  * are treated as milliseconds, and converted to jiffies when they are stored.
2589  *
2590  * This routine will ensure the values are within the range specified by
2591  * table->extra1 (min) and table->extra2 (max).
2592  *
2593  * Returns 0 on success.
2594  */
2595 int proc_doulongvec_ms_jiffies_minmax(struct ctl_table *table, int write,
2596                                       void __user *buffer,
2597                                       size_t *lenp, loff_t *ppos)
2598 {
2599     return do_proc_doulongvec_minmax(table, write, buffer,
2600                                      lenp, ppos, HZ, 1000l);
2601 }
2602
2603
2604 static int do_proc_dointvec_jiffies_conv(bool *negp, unsigned long *lvalp,
2605                                          int *valp,
2606                                          int write, void *data)
2607 {
2608         if (write) {
2609                 if (*lvalp > LONG_MAX / HZ)
2610                         return 1;
2611                 *valp = *negp ? -(*lvalp*HZ) : (*lvalp*HZ);
2612         } else {
2613                 int val = *valp;
2614                 unsigned long lval;
2615                 if (val < 0) {
2616                         *negp = true;
2617                         lval = -(unsigned long)val;
2618                 } else {
2619                         *negp = false;
2620                         lval = (unsigned long)val;
2621                 }
2622                 *lvalp = lval / HZ;
2623         }
2624         return 0;
2625 }
2626
2627 static int do_proc_dointvec_userhz_jiffies_conv(bool *negp, unsigned long *lvalp,
2628                                                 int *valp,
2629                                                 int write, void *data)
2630 {
2631         if (write) {
2632                 if (USER_HZ < HZ && *lvalp > (LONG_MAX / HZ) * USER_HZ)
2633                         return 1;
2634                 *valp = clock_t_to_jiffies(*negp ? -*lvalp : *lvalp);
2635         } else {
2636                 int val = *valp;
2637                 unsigned long lval;
2638                 if (val < 0) {
2639                         *negp = true;
2640                         lval = -(unsigned long)val;
2641                 } else {
2642                         *negp = false;
2643                         lval = (unsigned long)val;
2644                 }
2645                 *lvalp = jiffies_to_clock_t(lval);
2646         }
2647         return 0;
2648 }
2649
2650 static int do_proc_dointvec_ms_jiffies_conv(bool *negp, unsigned long *lvalp,
2651                                             int *valp,
2652                                             int write, void *data)
2653 {
2654         if (write) {
2655                 unsigned long jif = msecs_to_jiffies(*negp ? -*lvalp : *lvalp);
2656
2657                 if (jif > INT_MAX)
2658                         return 1;
2659                 *valp = (int)jif;
2660         } else {
2661                 int val = *valp;
2662                 unsigned long lval;
2663                 if (val < 0) {
2664                         *negp = true;
2665                         lval = -(unsigned long)val;
2666                 } else {
2667                         *negp = false;
2668                         lval = (unsigned long)val;
2669                 }
2670                 *lvalp = jiffies_to_msecs(lval);
2671         }
2672         return 0;
2673 }
2674
2675 /**
2676  * proc_dointvec_jiffies - read a vector of integers as seconds
2677  * @table: the sysctl table
2678  * @write: %TRUE if this is a write to the sysctl file
2679  * @buffer: the user buffer
2680  * @lenp: the size of the user buffer
2681  * @ppos: file position
2682  *
2683  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2684  * values from/to the user buffer, treated as an ASCII string.
2685  * The values read are assumed to be in seconds, and are converted into
2686  * jiffies.
2687  *
2688  * Returns 0 on success.
2689  */
2690 int proc_dointvec_jiffies(struct ctl_table *table, int write,
2691                           void __user *buffer, size_t *lenp, loff_t *ppos)
2692 {
2693     return do_proc_dointvec(table,write,buffer,lenp,ppos,
2694                             do_proc_dointvec_jiffies_conv,NULL);
2695 }
2696
2697 /**
2698  * proc_dointvec_userhz_jiffies - read a vector of integers as 1/USER_HZ seconds
2699  * @table: the sysctl table
2700  * @write: %TRUE if this is a write to the sysctl file
2701  * @buffer: the user buffer
2702  * @lenp: the size of the user buffer
2703  * @ppos: pointer to the file position
2704  *
2705  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2706  * values from/to the user buffer, treated as an ASCII string.
2707  * The values read are assumed to be in 1/USER_HZ seconds, and
2708  * are converted into jiffies.
2709  *
2710  * Returns 0 on success.
2711  */
2712 int proc_dointvec_userhz_jiffies(struct ctl_table *table, int write,
2713                                  void __user *buffer, size_t *lenp, loff_t *ppos)
2714 {
2715     return do_proc_dointvec(table,write,buffer,lenp,ppos,
2716                             do_proc_dointvec_userhz_jiffies_conv,NULL);
2717 }
2718
2719 /**
2720  * proc_dointvec_ms_jiffies - read a vector of integers as 1 milliseconds
2721  * @table: the sysctl table
2722  * @write: %TRUE if this is a write to the sysctl file
2723  * @buffer: the user buffer
2724  * @lenp: the size of the user buffer
2725  * @ppos: file position
2726  * @ppos: the current position in the file
2727  *
2728  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2729  * values from/to the user buffer, treated as an ASCII string.
2730  * The values read are assumed to be in 1/1000 seconds, and
2731  * are converted into jiffies.
2732  *
2733  * Returns 0 on success.
2734  */
2735 int proc_dointvec_ms_jiffies(struct ctl_table *table, int write,
2736                              void __user *buffer, size_t *lenp, loff_t *ppos)
2737 {
2738         return do_proc_dointvec(table, write, buffer, lenp, ppos,
2739                                 do_proc_dointvec_ms_jiffies_conv, NULL);
2740 }
2741
2742 static int proc_do_cad_pid(struct ctl_table *table, int write,
2743                            void __user *buffer, size_t *lenp, loff_t *ppos)
2744 {
2745         struct pid *new_pid;
2746         pid_t tmp;
2747         int r;
2748
2749         tmp = pid_vnr(cad_pid);
2750
2751         r = __do_proc_dointvec(&tmp, table, write, buffer,
2752                                lenp, ppos, NULL, NULL);
2753         if (r || !write)
2754                 return r;
2755
2756         new_pid = find_get_pid(tmp);
2757         if (!new_pid)
2758                 return -ESRCH;
2759
2760         put_pid(xchg(&cad_pid, new_pid));
2761         return 0;
2762 }
2763
2764 /**
2765  * proc_do_large_bitmap - read/write from/to a large bitmap
2766  * @table: the sysctl table
2767  * @write: %TRUE if this is a write to the sysctl file
2768  * @buffer: the user buffer
2769  * @lenp: the size of the user buffer
2770  * @ppos: file position
2771  *
2772  * The bitmap is stored at table->data and the bitmap length (in bits)
2773  * in table->maxlen.
2774  *
2775  * We use a range comma separated format (e.g. 1,3-4,10-10) so that
2776  * large bitmaps may be represented in a compact manner. Writing into
2777  * the file will clear the bitmap then update it with the given input.
2778  *
2779  * Returns 0 on success.
2780  */
2781 int proc_do_large_bitmap(struct ctl_table *table, int write,
2782                          void __user *buffer, size_t *lenp, loff_t *ppos)
2783 {
2784         int err = 0;
2785         bool first = 1;
2786         size_t left = *lenp;
2787         unsigned long bitmap_len = table->maxlen;
2788         unsigned long *bitmap = *(unsigned long **) table->data;
2789         unsigned long *tmp_bitmap = NULL;
2790         char tr_a[] = { '-', ',', '\n' }, tr_b[] = { ',', '\n', 0 }, c;
2791
2792         if (!bitmap || !bitmap_len || !left || (*ppos && !write)) {
2793                 *lenp = 0;
2794                 return 0;
2795         }
2796
2797         if (write) {
2798                 char *kbuf, *p;
2799
2800                 if (left > PAGE_SIZE - 1)
2801                         left = PAGE_SIZE - 1;
2802
2803                 p = kbuf = memdup_user_nul(buffer, left);
2804                 if (IS_ERR(kbuf))
2805                         return PTR_ERR(kbuf);
2806
2807                 tmp_bitmap = kzalloc(BITS_TO_LONGS(bitmap_len) * sizeof(unsigned long),
2808                                      GFP_KERNEL);
2809                 if (!tmp_bitmap) {
2810                         kfree(kbuf);
2811                         return -ENOMEM;
2812                 }
2813                 proc_skip_char(&p, &left, '\n');
2814                 while (!err && left) {
2815                         unsigned long val_a, val_b;
2816                         bool neg;
2817
2818                         err = proc_get_long(&p, &left, &val_a, &neg, tr_a,
2819                                              sizeof(tr_a), &c);
2820                         if (err)
2821                                 break;
2822                         if (val_a >= bitmap_len || neg) {
2823                                 err = -EINVAL;
2824                                 break;
2825                         }
2826
2827                         val_b = val_a;
2828                         if (left) {
2829                                 p++;
2830                                 left--;
2831                         }
2832
2833                         if (c == '-') {
2834                                 err = proc_get_long(&p, &left, &val_b,
2835                                                      &neg, tr_b, sizeof(tr_b),
2836                                                      &c);
2837                                 if (err)
2838                                         break;
2839                                 if (val_b >= bitmap_len || neg ||
2840                                     val_a > val_b) {
2841                                         err = -EINVAL;
2842                                         break;
2843                                 }
2844                                 if (left) {
2845                                         p++;
2846                                         left--;
2847                                 }
2848                         }
2849
2850                         bitmap_set(tmp_bitmap, val_a, val_b - val_a + 1);
2851                         first = 0;
2852                         proc_skip_char(&p, &left, '\n');
2853                 }
2854                 kfree(kbuf);
2855         } else {
2856                 unsigned long bit_a, bit_b = 0;
2857
2858                 while (left) {
2859                         bit_a = find_next_bit(bitmap, bitmap_len, bit_b);
2860                         if (bit_a >= bitmap_len)
2861                                 break;
2862                         bit_b = find_next_zero_bit(bitmap, bitmap_len,
2863                                                    bit_a + 1) - 1;
2864
2865                         if (!first) {
2866                                 err = proc_put_char(&buffer, &left, ',');
2867                                 if (err)
2868                                         break;
2869                         }
2870                         err = proc_put_long(&buffer, &left, bit_a, false);
2871                         if (err)
2872                                 break;
2873                         if (bit_a != bit_b) {
2874                                 err = proc_put_char(&buffer, &left, '-');
2875                                 if (err)
2876                                         break;
2877                                 err = proc_put_long(&buffer, &left, bit_b, false);
2878                                 if (err)
2879                                         break;
2880                         }
2881
2882                         first = 0; bit_b++;
2883                 }
2884                 if (!err)
2885                         err = proc_put_char(&buffer, &left, '\n');
2886         }
2887
2888         if (!err) {
2889                 if (write) {
2890                         if (*ppos)
2891                                 bitmap_or(bitmap, bitmap, tmp_bitmap, bitmap_len);
2892                         else
2893                                 bitmap_copy(bitmap, tmp_bitmap, bitmap_len);
2894                 }
2895                 kfree(tmp_bitmap);
2896                 *lenp -= left;
2897                 *ppos += *lenp;
2898                 return 0;
2899         } else {
2900                 kfree(tmp_bitmap);
2901                 return err;
2902         }
2903 }
2904
2905 #else /* CONFIG_PROC_SYSCTL */
2906
2907 int proc_dostring(struct ctl_table *table, int write,
2908                   void __user *buffer, size_t *lenp, loff_t *ppos)
2909 {
2910         return -ENOSYS;
2911 }
2912
2913 int proc_dointvec(struct ctl_table *table, int write,
2914                   void __user *buffer, size_t *lenp, loff_t *ppos)
2915 {
2916         return -ENOSYS;
2917 }
2918
2919 int proc_douintvec(struct ctl_table *table, int write,
2920                   void __user *buffer, size_t *lenp, loff_t *ppos)
2921 {
2922         return -ENOSYS;
2923 }
2924
2925 int proc_dointvec_minmax(struct ctl_table *table, int write,
2926                     void __user *buffer, size_t *lenp, loff_t *ppos)
2927 {
2928         return -ENOSYS;
2929 }
2930
2931 int proc_dointvec_jiffies(struct ctl_table *table, int write,
2932                     void __user *buffer, size_t *lenp, loff_t *ppos)
2933 {
2934         return -ENOSYS;
2935 }
2936
2937 int proc_dointvec_userhz_jiffies(struct ctl_table *table, int write,
2938                     void __user *buffer, size_t *lenp, loff_t *ppos)
2939 {
2940         return -ENOSYS;
2941 }
2942
2943 int proc_dointvec_ms_jiffies(struct ctl_table *table, int write,
2944                              void __user *buffer, size_t *lenp, loff_t *ppos)
2945 {
2946         return -ENOSYS;
2947 }
2948
2949 int proc_doulongvec_minmax(struct ctl_table *table, int write,
2950                     void __user *buffer, size_t *lenp, loff_t *ppos)
2951 {
2952         return -ENOSYS;
2953 }
2954
2955 int proc_doulongvec_ms_jiffies_minmax(struct ctl_table *table, int write,
2956                                       void __user *buffer,
2957                                       size_t *lenp, loff_t *ppos)
2958 {
2959     return -ENOSYS;
2960 }
2961
2962
2963 #endif /* CONFIG_PROC_SYSCTL */
2964
2965 /*
2966  * No sense putting this after each symbol definition, twice,
2967  * exception granted :-)
2968  */
2969 EXPORT_SYMBOL(proc_dointvec);
2970 EXPORT_SYMBOL(proc_douintvec);
2971 EXPORT_SYMBOL(proc_dointvec_jiffies);
2972 EXPORT_SYMBOL(proc_dointvec_minmax);
2973 EXPORT_SYMBOL(proc_dointvec_userhz_jiffies);
2974 EXPORT_SYMBOL(proc_dointvec_ms_jiffies);
2975 EXPORT_SYMBOL(proc_dostring);
2976 EXPORT_SYMBOL(proc_doulongvec_minmax);
2977 EXPORT_SYMBOL(proc_doulongvec_ms_jiffies_minmax);