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 extra_free_kbytes;
 109 extern int pid_max_min, pid_max_max;
 110 extern int percpu_pagelist_fraction;
 111 extern int compat_log;
 112 extern int latencytop_enabled;
 113 extern int sysctl_nr_open_min, sysctl_nr_open_max;
 114 #ifndef CONFIG_MMU
 115 extern int sysctl_nr_trim_pages;
 116 #endif
 117
 118 /* Constants used for minimum and  maximum */
 119 #ifdef CONFIG_LOCKUP_DETECTOR
 120 static int sixty = 60;
 121 #endif
 122
 123 static int __maybe_unused neg_one = -1;
 124
 125 static int zero;
 126 static int __maybe_unused one = 1;
 127 static int __maybe_unused two = 2;
 128 static int __maybe_unused four = 4;
 129 static unsigned long one_ul = 1;
 130 static int one_hundred = 100;
 131 #ifdef CONFIG_PRINTK
 132 static int ten_thousand = 10000;
 133 #endif
 134
 135 /* this is needed for the proc_doulongvec_minmax of vm_dirty_bytes */
 136 static unsigned long dirty_bytes_min = 2 * PAGE_SIZE;
 137
 138 /* this is needed for the proc_dointvec_minmax for [fs_]overflow UID and GID */
 139 static int maxolduid = 65535;
 140 static int minolduid;
 141
 142 static int ngroups_max = NGROUPS_MAX;
 143 static const int cap_last_cap = CAP_LAST_CAP;
 144
 145 /*this is needed for proc_doulongvec_minmax of sysctl_hung_task_timeout_secs */
 146 #ifdef CONFIG_DETECT_HUNG_TASK
 147 static unsigned long hung_task_timeout_max = (LONG_MAX/HZ);
 148 #endif
 149
 150 #ifdef CONFIG_INOTIFY_USER
 151 #include <linux/inotify.h>
 152 #endif
 153 #ifdef CONFIG_SPARC
 154 #endif
 155
 156 #ifdef __hppa__
 157 extern int pwrsw_enabled;
 158 #endif
 159
 160 #ifdef CONFIG_SYSCTL_ARCH_UNALIGN_ALLOW
 161 extern int unaligned_enabled;
 162 #endif
 163
 164 #ifdef CONFIG_IA64
 165 extern int unaligned_dump_stack;
 166 #endif
 167
 168 #ifdef CONFIG_SYSCTL_ARCH_UNALIGN_NO_WARN
 169 extern int no_unaligned_warning;
 170 #endif
 171
 172 #ifdef CONFIG_PROC_SYSCTL
 173
 174 #define SYSCTL_WRITES_LEGACY    -1
 175 #define SYSCTL_WRITES_WARN       0
 176 #define SYSCTL_WRITES_STRICT     1
 177
 178 static int sysctl_writes_strict = SYSCTL_WRITES_WARN;
 179
 180 static int proc_do_cad_pid(struct ctl_table *table, int write,
 181                   void __user *buffer, size_t *lenp, loff_t *ppos);
 182 static int proc_taint(struct ctl_table *table, int write,
 183                                void __user *buffer, size_t *lenp, loff_t *ppos);
 184 #endif
 185
 186 #ifdef CONFIG_PRINTK
 187 static int proc_dointvec_minmax_sysadmin(struct ctl_table *table, int write,
 188                                 void __user *buffer, size_t *lenp, loff_t *ppos);
 189 #endif
 190
 191 static int proc_dointvec_minmax_coredump(struct ctl_table *table, int write,
 192                 void __user *buffer, size_t *lenp, loff_t *ppos);
 193 #ifdef CONFIG_COREDUMP
 194 static int proc_dostring_coredump(struct ctl_table *table, int write,
 195                 void __user *buffer, size_t *lenp, loff_t *ppos);
 196 #endif
 197
 198 #ifdef CONFIG_MAGIC_SYSRQ
 199 /* Note: sysrq code uses it's own private copy */
 200 static int __sysrq_enabled = CONFIG_MAGIC_SYSRQ_DEFAULT_ENABLE;
 201
 202 static int sysrq_sysctl_handler(struct ctl_table *table, int write,
 203                                 void __user *buffer, size_t *lenp,
 204                                 loff_t *ppos)
 205 {
 206         int error;
 207
 208         error = proc_dointvec(table, write, buffer, lenp, ppos);
 209         if (error)
 210                 return error;
 211
 212         if (write)
 213                 sysrq_toggle_support(__sysrq_enabled);
 214
 215         return 0;
 216 }
 217
 218 #endif
 219
 220 static struct ctl_table kern_table[];
 221 static struct ctl_table vm_table[];
 222 static struct ctl_table fs_table[];
 223 static struct ctl_table debug_table[];
 224 static struct ctl_table dev_table[];
 225 extern struct ctl_table random_table[];
 226 #ifdef CONFIG_EPOLL
 227 extern struct ctl_table epoll_table[];
 228 #endif
 229
 230 #ifdef HAVE_ARCH_PICK_MMAP_LAYOUT
 231 int sysctl_legacy_va_layout;
 232 #endif
 233
 234 /* The default sysctl tables: */
 235
 236 static struct ctl_table sysctl_base_table[] = {
 237         {
 238                 .procname       = "kernel",
 239                 .mode           = 0555,
 240                 .child          = kern_table,
 241         },
 242         {
 243                 .procname       = "vm",
 244                 .mode           = 0555,
 245                 .child          = vm_table,
 246         },
 247         {
 248                 .procname       = "fs",
 249                 .mode           = 0555,
 250                 .child          = fs_table,
 251         },
 252         {
 253                 .procname       = "debug",
 254                 .mode           = 0555,
 255                 .child          = debug_table,
 256         },
 257         {
 258                 .procname       = "dev",
 259                 .mode           = 0555,
 260                 .child          = dev_table,
 261         },
 262         { }
 263 };
 264
 265 #ifdef CONFIG_SCHED_DEBUG
 266 static int min_sched_granularity_ns = 100000;           /* 100 usecs */
 267 static int max_sched_granularity_ns = NSEC_PER_SEC;     /* 1 second */
 268 static int min_wakeup_granularity_ns;                   /* 0 usecs */
 269 static int max_wakeup_granularity_ns = NSEC_PER_SEC;    /* 1 second */
 270 #ifdef CONFIG_SMP
 271 static int min_sched_tunable_scaling = SCHED_TUNABLESCALING_NONE;
 272 static int max_sched_tunable_scaling = SCHED_TUNABLESCALING_END-1;
 273 #endif /* CONFIG_SMP */
 274 #endif /* CONFIG_SCHED_DEBUG */
 275
 276 #ifdef CONFIG_COMPACTION
 277 static int min_extfrag_threshold;
 278 static int max_extfrag_threshold = 1000;
 279 #endif
 280
 281 static struct ctl_table kern_table[] = {
 282         {
 283                 .procname       = "sched_child_runs_first",
 284                 .data           = &sysctl_sched_child_runs_first,
 285                 .maxlen         = sizeof(unsigned int),
 286                 .mode           = 0644,
 287                 .proc_handler   = proc_dointvec,
 288         },
 289 #ifdef CONFIG_SCHED_DEBUG
 290         {
 291                 .procname       = "sched_min_granularity_ns",
 292                 .data           = &sysctl_sched_min_granularity,
 293                 .maxlen         = sizeof(unsigned int),
 294                 .mode           = 0644,
 295                 .proc_handler   = sched_proc_update_handler,
 296                 .extra1         = &min_sched_granularity_ns,
 297                 .extra2         = &max_sched_granularity_ns,
 298         },
 299         {
 300                 .procname       = "sched_latency_ns",
 301                 .data           = &sysctl_sched_latency,
 302                 .maxlen         = sizeof(unsigned int),
 303                 .mode           = 0644,
 304                 .proc_handler   = sched_proc_update_handler,
 305                 .extra1         = &min_sched_granularity_ns,
 306                 .extra2         = &max_sched_granularity_ns,
 307         },
 308         {
 309                 .procname       = "sched_is_big_little",
 310                 .data           = &sysctl_sched_is_big_little,
 311                 .maxlen         = sizeof(unsigned int),
 312                 .mode           = 0644,
 313                 .proc_handler   = proc_dointvec,
 314         },
 315 #ifdef CONFIG_SCHED_WALT
 316         {
 317                 .procname       = "sched_use_walt_cpu_util",
 318                 .data           = &sysctl_sched_use_walt_cpu_util,
 319                 .maxlen         = sizeof(unsigned int),
 320                 .mode           = 0644,
 321                 .proc_handler   = proc_dointvec,
 322         },
 323         {
 324                 .procname       = "sched_use_walt_task_util",
 325                 .data           = &sysctl_sched_use_walt_task_util,
 326                 .maxlen         = sizeof(unsigned int),
 327                 .mode           = 0644,
 328                 .proc_handler   = proc_dointvec,
 329         },
 330         {
 331                 .procname       = "sched_walt_init_task_load_pct",
 332                 .data           = &sysctl_sched_walt_init_task_load_pct,
 333                 .maxlen         = sizeof(unsigned int),
 334                 .mode           = 0644,
 335                 .proc_handler   = proc_dointvec,
 336         },
 337         {
 338                 .procname       = "sched_walt_cpu_high_irqload",
 339                 .data           = &sysctl_sched_walt_cpu_high_irqload,
 340                 .maxlen         = sizeof(unsigned int),
 341                 .mode           = 0644,
 342                 .proc_handler   = proc_dointvec,
 343         },
 344 #endif
 345         {
 346                 .procname       = "sched_sync_hint_enable",
 347                 .data           = &sysctl_sched_sync_hint_enable,
 348                 .maxlen         = sizeof(unsigned int),
 349                 .mode           = 0644,
 350                 .proc_handler   = proc_dointvec,
 351         },
 352         {
 353                 .procname       = "sched_initial_task_util",
 354                 .data           = &sysctl_sched_initial_task_util,
 355                 .maxlen         = sizeof(unsigned int),
 356                 .mode           = 0644,
 357                 .proc_handler   = proc_dointvec,
 358         },
 359         {
 360                 .procname       = "sched_cstate_aware",
 361                 .data           = &sysctl_sched_cstate_aware,
 362                 .maxlen         = sizeof(unsigned int),
 363                 .mode           = 0644,
 364                 .proc_handler   = proc_dointvec,
 365         },
 366         {
 367                 .procname       = "sched_wakeup_granularity_ns",
 368                 .data           = &sysctl_sched_wakeup_granularity,
 369                 .maxlen         = sizeof(unsigned int),
 370                 .mode           = 0644,
 371                 .proc_handler   = sched_proc_update_handler,
 372                 .extra1         = &min_wakeup_granularity_ns,
 373                 .extra2         = &max_wakeup_granularity_ns,
 374         },
 375 #ifdef CONFIG_SMP
 376         {
 377                 .procname       = "sched_tunable_scaling",
 378                 .data           = &sysctl_sched_tunable_scaling,
 379                 .maxlen         = sizeof(enum sched_tunable_scaling),
 380                 .mode           = 0644,
 381                 .proc_handler   = sched_proc_update_handler,
 382                 .extra1         = &min_sched_tunable_scaling,
 383                 .extra2         = &max_sched_tunable_scaling,
 384         },
 385         {
 386                 .procname       = "sched_migration_cost_ns",
 387                 .data           = &sysctl_sched_migration_cost,
 388                 .maxlen         = sizeof(unsigned int),
 389                 .mode           = 0644,
 390                 .proc_handler   = proc_dointvec,
 391         },
 392         {
 393                 .procname       = "sched_nr_migrate",
 394                 .data           = &sysctl_sched_nr_migrate,
 395                 .maxlen         = sizeof(unsigned int),
 396                 .mode           = 0644,
 397                 .proc_handler   = proc_dointvec,
 398         },
 399         {
 400                 .procname       = "sched_time_avg_ms",
 401                 .data           = &sysctl_sched_time_avg,
 402                 .maxlen         = sizeof(unsigned int),
 403                 .mode           = 0644,
 404                 .proc_handler   = proc_dointvec,
 405         },
 406         {
 407                 .procname       = "sched_shares_window_ns",
 408                 .data           = &sysctl_sched_shares_window,
 409                 .maxlen         = sizeof(unsigned int),
 410                 .mode           = 0644,
 411                 .proc_handler   = proc_dointvec,
 412         },
 413 #endif /* CONFIG_SMP */
 414 #ifdef CONFIG_NUMA_BALANCING
 415         {
 416                 .procname       = "numa_balancing_scan_delay_ms",
 417                 .data           = &sysctl_numa_balancing_scan_delay,
 418                 .maxlen         = sizeof(unsigned int),
 419                 .mode           = 0644,
 420                 .proc_handler   = proc_dointvec,
 421         },
 422         {
 423                 .procname       = "numa_balancing_scan_period_min_ms",
 424                 .data           = &sysctl_numa_balancing_scan_period_min,
 425                 .maxlen         = sizeof(unsigned int),
 426                 .mode           = 0644,
 427                 .proc_handler   = proc_dointvec,
 428         },
 429         {
 430                 .procname       = "numa_balancing_scan_period_max_ms",
 431                 .data           = &sysctl_numa_balancing_scan_period_max,
 432                 .maxlen         = sizeof(unsigned int),
 433                 .mode           = 0644,
 434                 .proc_handler   = proc_dointvec,
 435         },
 436         {
 437                 .procname       = "numa_balancing_scan_size_mb",
 438                 .data           = &sysctl_numa_balancing_scan_size,
 439                 .maxlen         = sizeof(unsigned int),
 440                 .mode           = 0644,
 441                 .proc_handler   = proc_dointvec_minmax,
 442                 .extra1         = &one,
 443         },
 444         {
 445                 .procname       = "numa_balancing",
 446                 .data           = NULL, /* filled in by handler */
 447                 .maxlen         = sizeof(unsigned int),
 448                 .mode           = 0644,
 449                 .proc_handler   = sysctl_numa_balancing,
 450                 .extra1         = &zero,
 451                 .extra2         = &one,
 452         },
 453 #endif /* CONFIG_NUMA_BALANCING */
 454 #endif /* CONFIG_SCHED_DEBUG */
 455         {
 456                 .procname       = "sched_rt_period_us",
 457                 .data           = &sysctl_sched_rt_period,
 458                 .maxlen         = sizeof(unsigned int),
 459                 .mode           = 0644,
 460                 .proc_handler   = sched_rt_handler,
 461         },
 462         {
 463                 .procname       = "sched_rt_runtime_us",
 464                 .data           = &sysctl_sched_rt_runtime,
 465                 .maxlen         = sizeof(int),
 466                 .mode           = 0644,
 467                 .proc_handler   = sched_rt_handler,
 468         },
 469         {
 470                 .procname       = "sched_rr_timeslice_ms",
 471                 .data           = &sched_rr_timeslice,
 472                 .maxlen         = sizeof(int),
 473                 .mode           = 0644,
 474                 .proc_handler   = sched_rr_handler,
 475         },
 476 #ifdef CONFIG_SCHED_AUTOGROUP
 477         {
 478                 .procname       = "sched_autogroup_enabled",
 479                 .data           = &sysctl_sched_autogroup_enabled,
 480                 .maxlen         = sizeof(unsigned int),
 481                 .mode           = 0644,
 482                 .proc_handler   = proc_dointvec_minmax,
 483                 .extra1         = &zero,
 484                 .extra2         = &one,
 485         },
 486 #endif
 487 #ifdef CONFIG_CFS_BANDWIDTH
 488         {
 489                 .procname       = "sched_cfs_bandwidth_slice_us",
 490                 .data           = &sysctl_sched_cfs_bandwidth_slice,
 491                 .maxlen         = sizeof(unsigned int),
 492                 .mode           = 0644,
 493                 .proc_handler   = proc_dointvec_minmax,
 494                 .extra1         = &one,
 495         },
 496 #endif
 497 #ifdef CONFIG_SCHED_TUNE
 498         {
 499                 .procname       = "sched_cfs_boost",
 500                 .data           = &sysctl_sched_cfs_boost,
 501                 .maxlen         = sizeof(sysctl_sched_cfs_boost),
 502 #ifdef CONFIG_CGROUP_SCHEDTUNE
 503                 .mode           = 0444,
 504 #else
 505                 .mode           = 0644,
 506 #endif
 507                 .proc_handler   = &sysctl_sched_cfs_boost_handler,
 508                 .extra1         = &zero,
 509                 .extra2         = &one_hundred,
 510         },
 511 #endif
 512 #ifdef CONFIG_PROVE_LOCKING
 513         {
 514                 .procname       = "prove_locking",
 515                 .data           = &prove_locking,
 516                 .maxlen         = sizeof(int),
 517                 .mode           = 0644,
 518                 .proc_handler   = proc_dointvec,
 519         },
 520 #endif
 521 #ifdef CONFIG_LOCK_STAT
 522         {
 523                 .procname       = "lock_stat",
 524                 .data           = &lock_stat,
 525                 .maxlen         = sizeof(int),
 526                 .mode           = 0644,
 527                 .proc_handler   = proc_dointvec,
 528         },
 529 #endif
 530         {
 531                 .procname       = "panic",
 532                 .data           = &panic_timeout,
 533                 .maxlen         = sizeof(int),
 534                 .mode           = 0644,
 535                 .proc_handler   = proc_dointvec,
 536         },
 537 #ifdef CONFIG_COREDUMP
 538         {
 539                 .procname       = "core_uses_pid",
 540                 .data           = &core_uses_pid,
 541                 .maxlen         = sizeof(int),
 542                 .mode           = 0644,
 543                 .proc_handler   = proc_dointvec,
 544         },
 545         {
 546                 .procname       = "core_pattern",
 547                 .data           = core_pattern,
 548                 .maxlen         = CORENAME_MAX_SIZE,
 549                 .mode           = 0644,
 550                 .proc_handler   = proc_dostring_coredump,
 551         },
 552         {
 553                 .procname       = "core_pipe_limit",
 554                 .data           = &core_pipe_limit,
 555                 .maxlen         = sizeof(unsigned int),
 556                 .mode           = 0644,
 557                 .proc_handler   = proc_dointvec,
 558         },
 559 #endif
 560 #ifdef CONFIG_PROC_SYSCTL
 561         {
 562                 .procname       = "tainted",
 563                 .maxlen         = sizeof(long),
 564                 .mode           = 0644,
 565                 .proc_handler   = proc_taint,
 566         },
 567         {
 568                 .procname       = "sysctl_writes_strict",
 569                 .data           = &sysctl_writes_strict,
 570                 .maxlen         = sizeof(int),
 571                 .mode           = 0644,
 572                 .proc_handler   = proc_dointvec_minmax,
 573                 .extra1         = &neg_one,
 574                 .extra2         = &one,
 575         },
 576 #endif
 577 #ifdef CONFIG_LATENCYTOP
 578         {
 579                 .procname       = "latencytop",
 580                 .data           = &latencytop_enabled,
 581                 .maxlen         = sizeof(int),
 582                 .mode           = 0644,
 583                 .proc_handler   = proc_dointvec,
 584         },
 585 #endif
 586 #ifdef CONFIG_BLK_DEV_INITRD
 587         {
 588                 .procname       = "real-root-dev",
 589                 .data           = &real_root_dev,
 590                 .maxlen         = sizeof(int),
 591                 .mode           = 0644,
 592                 .proc_handler   = proc_dointvec,
 593         },
 594 #endif
 595         {
 596                 .procname       = "print-fatal-signals",
 597                 .data           = &print_fatal_signals,
 598                 .maxlen         = sizeof(int),
 599                 .mode           = 0644,
 600                 .proc_handler   = proc_dointvec,
 601         },
 602 #ifdef CONFIG_SPARC
 603         {
 604                 .procname       = "reboot-cmd",
 605                 .data           = reboot_command,
 606                 .maxlen         = 256,
 607                 .mode           = 0644,
 608                 .proc_handler   = proc_dostring,
 609         },
 610         {
 611                 .procname       = "stop-a",
 612                 .data           = &stop_a_enabled,
 613                 .maxlen         = sizeof (int),
 614                 .mode           = 0644,
 615                 .proc_handler   = proc_dointvec,
 616         },
 617         {
 618                 .procname       = "scons-poweroff",
 619                 .data           = &scons_pwroff,
 620                 .maxlen         = sizeof (int),
 621                 .mode           = 0644,
 622                 .proc_handler   = proc_dointvec,
 623         },
 624 #endif
 625 #ifdef CONFIG_SPARC64
 626         {
 627                 .procname       = "tsb-ratio",
 628                 .data           = &sysctl_tsb_ratio,
 629                 .maxlen         = sizeof (int),
 630                 .mode           = 0644,
 631                 .proc_handler   = proc_dointvec,
 632         },
 633 #endif
 634 #ifdef __hppa__
 635         {
 636                 .procname       = "soft-power",
 637                 .data           = &pwrsw_enabled,
 638                 .maxlen         = sizeof (int),
 639                 .mode           = 0644,
 640                 .proc_handler   = proc_dointvec,
 641         },
 642 #endif
 643 #ifdef CONFIG_SYSCTL_ARCH_UNALIGN_ALLOW
 644         {
 645                 .procname       = "unaligned-trap",
 646                 .data           = &unaligned_enabled,
 647                 .maxlen         = sizeof (int),
 648                 .mode           = 0644,
 649                 .proc_handler   = proc_dointvec,
 650         },
 651 #endif
 652         {
 653                 .procname       = "ctrl-alt-del",
 654                 .data           = &C_A_D,
 655                 .maxlen         = sizeof(int),
 656                 .mode           = 0644,
 657                 .proc_handler   = proc_dointvec,
 658         },
 659 #ifdef CONFIG_FUNCTION_TRACER
 660         {
 661                 .procname       = "ftrace_enabled",
 662                 .data           = &ftrace_enabled,
 663                 .maxlen         = sizeof(int),
 664                 .mode           = 0644,
 665                 .proc_handler   = ftrace_enable_sysctl,
 666         },
 667 #endif
 668 #ifdef CONFIG_STACK_TRACER
 669         {
 670                 .procname       = "stack_tracer_enabled",
 671                 .data           = &stack_tracer_enabled,
 672                 .maxlen         = sizeof(int),
 673                 .mode           = 0644,
 674                 .proc_handler   = stack_trace_sysctl,
 675         },
 676 #endif
 677 #ifdef CONFIG_TRACING
 678         {
 679                 .procname       = "ftrace_dump_on_oops",
 680                 .data           = &ftrace_dump_on_oops,
 681                 .maxlen         = sizeof(int),
 682                 .mode           = 0644,
 683                 .proc_handler   = proc_dointvec,
 684         },
 685         {
 686                 .procname       = "traceoff_on_warning",
 687                 .data           = &__disable_trace_on_warning,
 688                 .maxlen         = sizeof(__disable_trace_on_warning),
 689                 .mode           = 0644,
 690                 .proc_handler   = proc_dointvec,
 691         },
 692         {
 693                 .procname       = "tracepoint_printk",
 694                 .data           = &tracepoint_printk,
 695                 .maxlen         = sizeof(tracepoint_printk),
 696                 .mode           = 0644,
 697                 .proc_handler   = proc_dointvec,
 698         },
 699 #endif
 700 #ifdef CONFIG_KEXEC_CORE
 701         {
 702                 .procname       = "kexec_load_disabled",
 703                 .data           = &kexec_load_disabled,
 704                 .maxlen         = sizeof(int),
 705                 .mode           = 0644,
 706                 /* only handle a transition from default "0" to "1" */
 707                 .proc_handler   = proc_dointvec_minmax,
 708                 .extra1         = &one,
 709                 .extra2         = &one,
 710         },
 711 #endif
 712 #ifdef CONFIG_MODULES
 713         {
 714                 .procname       = "modprobe",
 715                 .data           = &modprobe_path,
 716                 .maxlen         = KMOD_PATH_LEN,
 717                 .mode           = 0644,
 718                 .proc_handler   = proc_dostring,
 719         },
 720         {
 721                 .procname       = "modules_disabled",
 722                 .data           = &modules_disabled,
 723                 .maxlen         = sizeof(int),
 724                 .mode           = 0644,
 725                 /* only handle a transition from default "0" to "1" */
 726                 .proc_handler   = proc_dointvec_minmax,
 727                 .extra1         = &one,
 728                 .extra2         = &one,
 729         },
 730 #endif
 731 #ifdef CONFIG_UEVENT_HELPER
 732         {
 733                 .procname       = "hotplug",
 734                 .data           = &uevent_helper,
 735                 .maxlen         = UEVENT_HELPER_PATH_LEN,
 736                 .mode           = 0644,
 737                 .proc_handler   = proc_dostring,
 738         },
 739 #endif
 740 #ifdef CONFIG_CHR_DEV_SG
 741         {
 742                 .procname       = "sg-big-buff",
 743                 .data           = &sg_big_buff,
 744                 .maxlen         = sizeof (int),
 745                 .mode           = 0444,
 746                 .proc_handler   = proc_dointvec,
 747         },
 748 #endif
 749 #ifdef CONFIG_BSD_PROCESS_ACCT
 750         {
 751                 .procname       = "acct",
 752                 .data           = &acct_parm,
 753                 .maxlen         = 3*sizeof(int),
 754                 .mode           = 0644,
 755                 .proc_handler   = proc_dointvec,
 756         },
 757 #endif
 758 #ifdef CONFIG_MAGIC_SYSRQ
 759         {
 760                 .procname       = "sysrq",
 761                 .data           = &__sysrq_enabled,
 762                 .maxlen         = sizeof (int),
 763                 .mode           = 0644,
 764                 .proc_handler   = sysrq_sysctl_handler,
 765         },
 766 #endif
 767 #ifdef CONFIG_PROC_SYSCTL
 768         {
 769                 .procname       = "cad_pid",
 770                 .data           = NULL,
 771                 .maxlen         = sizeof (int),
 772                 .mode           = 0600,
 773                 .proc_handler   = proc_do_cad_pid,
 774         },
 775 #endif
 776         {
 777                 .procname       = "threads-max",
 778                 .data           = NULL,
 779                 .maxlen         = sizeof(int),
 780                 .mode           = 0644,
 781                 .proc_handler   = sysctl_max_threads,
 782         },
 783         {
 784                 .procname       = "random",
 785                 .mode           = 0555,
 786                 .child          = random_table,
 787         },
 788         {
 789                 .procname       = "usermodehelper",
 790                 .mode           = 0555,
 791                 .child          = usermodehelper_table,
 792         },
 793         {
 794                 .procname       = "overflowuid",
 795                 .data           = &overflowuid,
 796                 .maxlen         = sizeof(int),
 797                 .mode           = 0644,
 798                 .proc_handler   = proc_dointvec_minmax,
 799                 .extra1         = &minolduid,
 800                 .extra2         = &maxolduid,
 801         },
 802         {
 803                 .procname       = "overflowgid",
 804                 .data           = &overflowgid,
 805                 .maxlen         = sizeof(int),
 806                 .mode           = 0644,
 807                 .proc_handler   = proc_dointvec_minmax,
 808                 .extra1         = &minolduid,
 809                 .extra2         = &maxolduid,
 810         },
 811 #ifdef CONFIG_S390
 812 #ifdef CONFIG_MATHEMU
 813         {
 814                 .procname       = "ieee_emulation_warnings",
 815                 .data           = &sysctl_ieee_emulation_warnings,
 816                 .maxlen         = sizeof(int),
 817                 .mode           = 0644,
 818                 .proc_handler   = proc_dointvec,
 819         },
 820 #endif
 821         {
 822                 .procname       = "userprocess_debug",
 823                 .data           = &show_unhandled_signals,
 824                 .maxlen         = sizeof(int),
 825                 .mode           = 0644,
 826                 .proc_handler   = proc_dointvec,
 827         },
 828 #endif
 829         {
 830                 .procname       = "pid_max",
 831                 .data           = &pid_max,
 832                 .maxlen         = sizeof (int),
 833                 .mode           = 0644,
 834                 .proc_handler   = proc_dointvec_minmax,
 835                 .extra1         = &pid_max_min,
 836                 .extra2         = &pid_max_max,
 837         },
 838         {
 839                 .procname       = "panic_on_oops",
 840                 .data           = &panic_on_oops,
 841                 .maxlen         = sizeof(int),
 842                 .mode           = 0644,
 843                 .proc_handler   = proc_dointvec,
 844         },
 845 #if defined CONFIG_PRINTK
 846         {
 847                 .procname       = "printk",
 848                 .data           = &console_loglevel,
 849                 .maxlen         = 4*sizeof(int),
 850                 .mode           = 0644,
 851                 .proc_handler   = proc_dointvec,
 852         },
 853         {
 854                 .procname       = "printk_ratelimit",
 855                 .data           = &printk_ratelimit_state.interval,
 856                 .maxlen         = sizeof(int),
 857                 .mode           = 0644,
 858                 .proc_handler   = proc_dointvec_jiffies,
 859         },
 860         {
 861                 .procname       = "printk_ratelimit_burst",
 862                 .data           = &printk_ratelimit_state.burst,
 863                 .maxlen         = sizeof(int),
 864                 .mode           = 0644,
 865                 .proc_handler   = proc_dointvec,
 866         },
 867         {
 868                 .procname       = "printk_delay",
 869                 .data           = &printk_delay_msec,
 870                 .maxlen         = sizeof(int),
 871                 .mode           = 0644,
 872                 .proc_handler   = proc_dointvec_minmax,
 873                 .extra1         = &zero,
 874                 .extra2         = &ten_thousand,
 875         },
 876         {
 877                 .procname       = "dmesg_restrict",
 878                 .data           = &dmesg_restrict,
 879                 .maxlen         = sizeof(int),
 880                 .mode           = 0644,
 881                 .proc_handler   = proc_dointvec_minmax_sysadmin,
 882                 .extra1         = &zero,
 883                 .extra2         = &one,
 884         },
 885         {
 886                 .procname       = "kptr_restrict",
 887                 .data           = &kptr_restrict,
 888                 .maxlen         = sizeof(int),
 889                 .mode           = 0644,
 890                 .proc_handler   = proc_dointvec_minmax_sysadmin,
 891                 .extra1         = &zero,
 892                 .extra2         = &two,
 893         },
 894 #endif
 895         {
 896                 .procname       = "ngroups_max",
 897                 .data           = &ngroups_max,
 898                 .maxlen         = sizeof (int),
 899                 .mode           = 0444,
 900                 .proc_handler   = proc_dointvec,
 901         },
 902         {
 903                 .procname       = "cap_last_cap",
 904                 .data           = (void *)&cap_last_cap,
 905                 .maxlen         = sizeof(int),
 906                 .mode           = 0444,
 907                 .proc_handler   = proc_dointvec,
 908         },
 909 #if defined(CONFIG_LOCKUP_DETECTOR)
 910         {
 911                 .procname       = "watchdog",
 912                 .data           = &watchdog_user_enabled,
 913                 .maxlen         = sizeof (int),
 914                 .mode           = 0644,
 915                 .proc_handler   = proc_watchdog,
 916                 .extra1         = &zero,
 917                 .extra2         = &one,
 918         },
 919         {
 920                 .procname       = "watchdog_thresh",
 921                 .data           = &watchdog_thresh,
 922                 .maxlen         = sizeof(int),
 923                 .mode           = 0644,
 924                 .proc_handler   = proc_watchdog_thresh,
 925                 .extra1         = &zero,
 926                 .extra2         = &sixty,
 927         },
 928         {
 929                 .procname       = "nmi_watchdog",
 930                 .data           = &nmi_watchdog_enabled,
 931                 .maxlen         = sizeof (int),
 932                 .mode           = 0644,
 933                 .proc_handler   = proc_nmi_watchdog,
 934                 .extra1         = &zero,
 935 #if defined(CONFIG_HAVE_NMI_WATCHDOG) || defined(CONFIG_HARDLOCKUP_DETECTOR)
 936                 .extra2         = &one,
 937 #else
 938                 .extra2         = &zero,
 939 #endif
 940         },
 941         {
 942                 .procname       = "soft_watchdog",
 943                 .data           = &soft_watchdog_enabled,
 944                 .maxlen         = sizeof (int),
 945                 .mode           = 0644,
 946                 .proc_handler   = proc_soft_watchdog,
 947                 .extra1         = &zero,
 948                 .extra2         = &one,
 949         },
 950         {
 951                 .procname       = "watchdog_cpumask",
 952                 .data           = &watchdog_cpumask_bits,
 953                 .maxlen         = NR_CPUS,
 954                 .mode           = 0644,
 955                 .proc_handler   = proc_watchdog_cpumask,
 956         },
 957         {
 958                 .procname       = "softlockup_panic",
 959                 .data           = &softlockup_panic,
 960                 .maxlen         = sizeof(int),
 961                 .mode           = 0644,
 962                 .proc_handler   = proc_dointvec_minmax,
 963                 .extra1         = &zero,
 964                 .extra2         = &one,
 965         },
 966 #ifdef CONFIG_HARDLOCKUP_DETECTOR
 967         {
 968                 .procname       = "hardlockup_panic",
 969                 .data           = &hardlockup_panic,
 970                 .maxlen         = sizeof(int),
 971                 .mode           = 0644,
 972                 .proc_handler   = proc_dointvec_minmax,
 973                 .extra1         = &zero,
 974                 .extra2         = &one,
 975         },
 976 #endif
 977 #ifdef CONFIG_SMP
 978         {
 979                 .procname       = "softlockup_all_cpu_backtrace",
 980                 .data           = &sysctl_softlockup_all_cpu_backtrace,
 981                 .maxlen         = sizeof(int),
 982                 .mode           = 0644,
 983                 .proc_handler   = proc_dointvec_minmax,
 984                 .extra1         = &zero,
 985                 .extra2         = &one,
 986         },
 987         {
 988                 .procname       = "hardlockup_all_cpu_backtrace",
 989                 .data           = &sysctl_hardlockup_all_cpu_backtrace,
 990                 .maxlen         = sizeof(int),
 991                 .mode           = 0644,
 992                 .proc_handler   = proc_dointvec_minmax,
 993                 .extra1         = &zero,
 994                 .extra2         = &one,
 995         },
 996 #endif /* CONFIG_SMP */
 997 #endif
 998 #if defined(CONFIG_X86_LOCAL_APIC) && defined(CONFIG_X86)
 999         {
1000                 .procname       = "unknown_nmi_panic",
1001                 .data           = &unknown_nmi_panic,
1002                 .maxlen         = sizeof (int),
1003                 .mode           = 0644,
1004                 .proc_handler   = proc_dointvec,
1005         },
1006 #endif
1007 #if defined(CONFIG_X86)
1008         {
1009                 .procname       = "panic_on_unrecovered_nmi",
1010                 .data           = &panic_on_unrecovered_nmi,
1011                 .maxlen         = sizeof(int),
1012                 .mode           = 0644,
1013                 .proc_handler   = proc_dointvec,
1014         },
1015         {
1016                 .procname       = "panic_on_io_nmi",
1017                 .data           = &panic_on_io_nmi,
1018                 .maxlen         = sizeof(int),
1019                 .mode           = 0644,
1020                 .proc_handler   = proc_dointvec,
1021         },
1022 #ifdef CONFIG_DEBUG_STACKOVERFLOW
1023         {
1024                 .procname       = "panic_on_stackoverflow",
1025                 .data           = &sysctl_panic_on_stackoverflow,
1026                 .maxlen         = sizeof(int),
1027                 .mode           = 0644,
1028                 .proc_handler   = proc_dointvec,
1029         },
1030 #endif
1031         {
1032                 .procname       = "bootloader_type",
1033                 .data           = &bootloader_type,
1034                 .maxlen         = sizeof (int),
1035                 .mode           = 0444,
1036                 .proc_handler   = proc_dointvec,
1037         },
1038         {
1039                 .procname       = "bootloader_version",
1040                 .data           = &bootloader_version,
1041                 .maxlen         = sizeof (int),
1042                 .mode           = 0444,
1043                 .proc_handler   = proc_dointvec,
1044         },
1045         {
1046                 .procname       = "kstack_depth_to_print",
1047                 .data           = &kstack_depth_to_print,
1048                 .maxlen         = sizeof(int),
1049                 .mode           = 0644,
1050                 .proc_handler   = proc_dointvec,
1051         },
1052         {
1053                 .procname       = "io_delay_type",
1054                 .data           = &io_delay_type,
1055                 .maxlen         = sizeof(int),
1056                 .mode           = 0644,
1057                 .proc_handler   = proc_dointvec,
1058         },
1059 #endif
1060 #if defined(CONFIG_MMU)
1061         {
1062                 .procname       = "randomize_va_space",
1063                 .data           = &randomize_va_space,
1064                 .maxlen         = sizeof(int),
1065                 .mode           = 0644,
1066                 .proc_handler   = proc_dointvec,
1067         },
1068 #endif
1069 #if defined(CONFIG_S390) && defined(CONFIG_SMP)
1070         {
1071                 .procname       = "spin_retry",
1072                 .data           = &spin_retry,
1073                 .maxlen         = sizeof (int),
1074                 .mode           = 0644,
1075                 .proc_handler   = proc_dointvec,
1076         },
1077 #endif
1078 #if     defined(CONFIG_ACPI_SLEEP) && defined(CONFIG_X86)
1079         {
1080                 .procname       = "acpi_video_flags",
1081                 .data           = &acpi_realmode_flags,
1082                 .maxlen         = sizeof (unsigned long),
1083                 .mode           = 0644,
1084                 .proc_handler   = proc_doulongvec_minmax,
1085         },
1086 #endif
1087 #ifdef CONFIG_SYSCTL_ARCH_UNALIGN_NO_WARN
1088         {
1089                 .procname       = "ignore-unaligned-usertrap",
1090                 .data           = &no_unaligned_warning,
1091                 .maxlen         = sizeof (int),
1092                 .mode           = 0644,
1093                 .proc_handler   = proc_dointvec,
1094         },
1095 #endif
1096 #ifdef CONFIG_IA64
1097         {
1098                 .procname       = "unaligned-dump-stack",
1099                 .data           = &unaligned_dump_stack,
1100                 .maxlen         = sizeof (int),
1101                 .mode           = 0644,
1102                 .proc_handler   = proc_dointvec,
1103         },
1104 #endif
1105 #ifdef CONFIG_DETECT_HUNG_TASK
1106         {
1107                 .procname       = "hung_task_panic",
1108                 .data           = &sysctl_hung_task_panic,
1109                 .maxlen         = sizeof(int),
1110                 .mode           = 0644,
1111                 .proc_handler   = proc_dointvec_minmax,
1112                 .extra1         = &zero,
1113                 .extra2         = &one,
1114         },
1115         {
1116                 .procname       = "hung_task_check_count",
1117                 .data           = &sysctl_hung_task_check_count,
1118                 .maxlen         = sizeof(int),
1119                 .mode           = 0644,
1120                 .proc_handler   = proc_dointvec_minmax,
1121                 .extra1         = &zero,
1122         },
1123         {
1124                 .procname       = "hung_task_timeout_secs",
1125                 .data           = &sysctl_hung_task_timeout_secs,
1126                 .maxlen         = sizeof(unsigned long),
1127                 .mode           = 0644,
1128                 .proc_handler   = proc_dohung_task_timeout_secs,
1129                 .extra2         = &hung_task_timeout_max,
1130         },
1131         {
1132                 .procname       = "hung_task_warnings",
1133                 .data           = &sysctl_hung_task_warnings,
1134                 .maxlen         = sizeof(int),
1135                 .mode           = 0644,
1136                 .proc_handler   = proc_dointvec_minmax,
1137                 .extra1         = &neg_one,
1138         },
1139 #endif
1140 #ifdef CONFIG_COMPAT
1141         {
1142                 .procname       = "compat-log",
1143                 .data           = &compat_log,
1144                 .maxlen         = sizeof (int),
1145                 .mode           = 0644,
1146                 .proc_handler   = proc_dointvec,
1147         },
1148 #endif
1149 #ifdef CONFIG_RT_MUTEXES
1150         {
1151                 .procname       = "max_lock_depth",
1152                 .data           = &max_lock_depth,
1153                 .maxlen         = sizeof(int),
1154                 .mode           = 0644,
1155                 .proc_handler   = proc_dointvec,
1156         },
1157 #endif
1158         {
1159                 .procname       = "poweroff_cmd",
1160                 .data           = &poweroff_cmd,
1161                 .maxlen         = POWEROFF_CMD_PATH_LEN,
1162                 .mode           = 0644,
1163                 .proc_handler   = proc_dostring,
1164         },
1165 #ifdef CONFIG_KEYS
1166         {
1167                 .procname       = "keys",
1168                 .mode           = 0555,
1169                 .child          = key_sysctls,
1170         },
1171 #endif
1172 #ifdef CONFIG_PERF_EVENTS
1173         /*
1174          * User-space scripts rely on the existence of this file
1175          * as a feature check for perf_events being enabled.
1176          *
1177          * So it's an ABI, do not remove!
1178          */
1179         {
1180                 .procname       = "perf_event_paranoid",
1181                 .data           = &sysctl_perf_event_paranoid,
1182                 .maxlen         = sizeof(sysctl_perf_event_paranoid),
1183                 .mode           = 0644,
1184                 .proc_handler   = proc_dointvec,
1185         },
1186         {
1187                 .procname       = "perf_event_mlock_kb",
1188                 .data           = &sysctl_perf_event_mlock,
1189                 .maxlen         = sizeof(sysctl_perf_event_mlock),
1190                 .mode           = 0644,
1191                 .proc_handler   = proc_dointvec,
1192         },
1193         {
1194                 .procname       = "perf_event_max_sample_rate",
1195                 .data           = &sysctl_perf_event_sample_rate,
1196                 .maxlen         = sizeof(sysctl_perf_event_sample_rate),
1197                 .mode           = 0644,
1198                 .proc_handler   = perf_proc_update_handler,
1199                 .extra1         = &one,
1200         },
1201         {
1202                 .procname       = "perf_cpu_time_max_percent",
1203                 .data           = &sysctl_perf_cpu_time_max_percent,
1204                 .maxlen         = sizeof(sysctl_perf_cpu_time_max_percent),
1205                 .mode           = 0644,
1206                 .proc_handler   = perf_cpu_time_max_percent_handler,
1207                 .extra1         = &zero,
1208                 .extra2         = &one_hundred,
1209         },
1210 #endif
1211 #ifdef CONFIG_KMEMCHECK
1212         {
1213                 .procname       = "kmemcheck",
1214                 .data           = &kmemcheck_enabled,
1215                 .maxlen         = sizeof(int),
1216                 .mode           = 0644,
1217                 .proc_handler   = proc_dointvec,
1218         },
1219 #endif
1220         {
1221                 .procname       = "panic_on_warn",
1222                 .data           = &panic_on_warn,
1223                 .maxlen         = sizeof(int),
1224                 .mode           = 0644,
1225                 .proc_handler   = proc_dointvec_minmax,
1226                 .extra1         = &zero,
1227                 .extra2         = &one,
1228         },
1229 #if defined(CONFIG_SMP) && defined(CONFIG_NO_HZ_COMMON)
1230         {
1231                 .procname       = "timer_migration",
1232                 .data           = &sysctl_timer_migration,
1233                 .maxlen         = sizeof(unsigned int),
1234                 .mode           = 0644,
1235                 .proc_handler   = timer_migration_handler,
1236         },
1237 #endif
1238 #ifdef CONFIG_BPF_SYSCALL
1239         {
1240                 .procname       = "unprivileged_bpf_disabled",
1241                 .data           = &sysctl_unprivileged_bpf_disabled,
1242                 .maxlen         = sizeof(sysctl_unprivileged_bpf_disabled),
1243                 .mode           = 0644,
1244                 /* only handle a transition from default "0" to "1" */
1245                 .proc_handler   = proc_dointvec_minmax,
1246                 .extra1         = &one,
1247                 .extra2         = &one,
1248         },
1249 #endif
1250         { }
1251 };
1252
1253 static struct ctl_table vm_table[] = {
1254         {
1255                 .procname       = "overcommit_memory",
1256                 .data           = &sysctl_overcommit_memory,
1257                 .maxlen         = sizeof(sysctl_overcommit_memory),
1258                 .mode           = 0644,
1259                 .proc_handler   = proc_dointvec_minmax,
1260                 .extra1         = &zero,
1261                 .extra2         = &two,
1262         },
1263         {
1264                 .procname       = "panic_on_oom",
1265                 .data           = &sysctl_panic_on_oom,
1266                 .maxlen         = sizeof(sysctl_panic_on_oom),
1267                 .mode           = 0644,
1268                 .proc_handler   = proc_dointvec_minmax,
1269                 .extra1         = &zero,
1270                 .extra2         = &two,
1271         },
1272         {
1273                 .procname       = "oom_kill_allocating_task",
1274                 .data           = &sysctl_oom_kill_allocating_task,
1275                 .maxlen         = sizeof(sysctl_oom_kill_allocating_task),
1276                 .mode           = 0644,
1277                 .proc_handler   = proc_dointvec,
1278         },
1279         {
1280                 .procname       = "oom_dump_tasks",
1281                 .data           = &sysctl_oom_dump_tasks,
1282                 .maxlen         = sizeof(sysctl_oom_dump_tasks),
1283                 .mode           = 0644,
1284                 .proc_handler   = proc_dointvec,
1285         },
1286         {
1287                 .procname       = "overcommit_ratio",
1288                 .data           = &sysctl_overcommit_ratio,
1289                 .maxlen         = sizeof(sysctl_overcommit_ratio),
1290                 .mode           = 0644,
1291                 .proc_handler   = overcommit_ratio_handler,
1292         },
1293         {
1294                 .procname       = "overcommit_kbytes",
1295                 .data           = &sysctl_overcommit_kbytes,
1296                 .maxlen         = sizeof(sysctl_overcommit_kbytes),
1297                 .mode           = 0644,
1298                 .proc_handler   = overcommit_kbytes_handler,
1299         },
1300         {
1301                 .procname       = "page-cluster",
1302                 .data           = &page_cluster,
1303                 .maxlen         = sizeof(int),
1304                 .mode           = 0644,
1305                 .proc_handler   = proc_dointvec_minmax,
1306                 .extra1         = &zero,
1307         },
1308         {
1309                 .procname       = "dirty_background_ratio",
1310                 .data           = &dirty_background_ratio,
1311                 .maxlen         = sizeof(dirty_background_ratio),
1312                 .mode           = 0644,
1313                 .proc_handler   = dirty_background_ratio_handler,
1314                 .extra1         = &zero,
1315                 .extra2         = &one_hundred,
1316         },
1317         {
1318                 .procname       = "dirty_background_bytes",
1319                 .data           = &dirty_background_bytes,
1320                 .maxlen         = sizeof(dirty_background_bytes),
1321                 .mode           = 0644,
1322                 .proc_handler   = dirty_background_bytes_handler,
1323                 .extra1         = &one_ul,
1324         },
1325         {
1326                 .procname       = "dirty_ratio",
1327                 .data           = &vm_dirty_ratio,
1328                 .maxlen         = sizeof(vm_dirty_ratio),
1329                 .mode           = 0644,
1330                 .proc_handler   = dirty_ratio_handler,
1331                 .extra1         = &zero,
1332                 .extra2         = &one_hundred,
1333         },
1334         {
1335                 .procname       = "dirty_bytes",
1336                 .data           = &vm_dirty_bytes,
1337                 .maxlen         = sizeof(vm_dirty_bytes),
1338                 .mode           = 0644,
1339                 .proc_handler   = dirty_bytes_handler,
1340                 .extra1         = &dirty_bytes_min,
1341         },
1342         {
1343                 .procname       = "dirty_writeback_centisecs",
1344                 .data           = &dirty_writeback_interval,
1345                 .maxlen         = sizeof(dirty_writeback_interval),
1346                 .mode           = 0644,
1347                 .proc_handler   = dirty_writeback_centisecs_handler,
1348         },
1349         {
1350                 .procname       = "dirty_expire_centisecs",
1351                 .data           = &dirty_expire_interval,
1352                 .maxlen         = sizeof(dirty_expire_interval),
1353                 .mode           = 0644,
1354                 .proc_handler   = proc_dointvec_minmax,
1355                 .extra1         = &zero,
1356         },
1357         {
1358                 .procname       = "dirtytime_expire_seconds",
1359                 .data           = &dirtytime_expire_interval,
1360                 .maxlen         = sizeof(dirty_expire_interval),
1361                 .mode           = 0644,
1362                 .proc_handler   = dirtytime_interval_handler,
1363                 .extra1         = &zero,
1364         },
1365         {
1366                 .procname       = "nr_pdflush_threads",
1367                 .mode           = 0444 /* read-only */,
1368                 .proc_handler   = pdflush_proc_obsolete,
1369         },
1370         {
1371                 .procname       = "swappiness",
1372                 .data           = &vm_swappiness,
1373                 .maxlen         = sizeof(vm_swappiness),
1374                 .mode           = 0644,
1375                 .proc_handler   = proc_dointvec_minmax,
1376                 .extra1         = &zero,
1377                 .extra2         = &one_hundred,
1378         },
1379 #ifdef CONFIG_HUGETLB_PAGE
1380         {
1381                 .procname       = "nr_hugepages",
1382                 .data           = NULL,
1383                 .maxlen         = sizeof(unsigned long),
1384                 .mode           = 0644,
1385                 .proc_handler   = hugetlb_sysctl_handler,
1386         },
1387 #ifdef CONFIG_NUMA
1388         {
1389                 .procname       = "nr_hugepages_mempolicy",
1390                 .data           = NULL,
1391                 .maxlen         = sizeof(unsigned long),
1392                 .mode           = 0644,
1393                 .proc_handler   = &hugetlb_mempolicy_sysctl_handler,
1394         },
1395 #endif
1396          {
1397                 .procname       = "hugetlb_shm_group",
1398                 .data           = &sysctl_hugetlb_shm_group,
1399                 .maxlen         = sizeof(gid_t),
1400                 .mode           = 0644,
1401                 .proc_handler   = proc_dointvec,
1402          },
1403          {
1404                 .procname       = "hugepages_treat_as_movable",
1405                 .data           = &hugepages_treat_as_movable,
1406                 .maxlen         = sizeof(int),
1407                 .mode           = 0644,
1408                 .proc_handler   = proc_dointvec,
1409         },
1410         {
1411                 .procname       = "nr_overcommit_hugepages",
1412                 .data           = NULL,
1413                 .maxlen         = sizeof(unsigned long),
1414                 .mode           = 0644,
1415                 .proc_handler   = hugetlb_overcommit_handler,
1416         },
1417 #endif
1418         {
1419                 .procname       = "lowmem_reserve_ratio",
1420                 .data           = &sysctl_lowmem_reserve_ratio,
1421                 .maxlen         = sizeof(sysctl_lowmem_reserve_ratio),
1422                 .mode           = 0644,
1423                 .proc_handler   = lowmem_reserve_ratio_sysctl_handler,
1424         },
1425         {
1426                 .procname       = "drop_caches",
1427                 .data           = &sysctl_drop_caches,
1428                 .maxlen         = sizeof(int),
1429                 .mode           = 0644,
1430                 .proc_handler   = drop_caches_sysctl_handler,
1431                 .extra1         = &one,
1432                 .extra2         = &four,
1433         },
1434 #ifdef CONFIG_COMPACTION
1435         {
1436                 .procname       = "compact_memory",
1437                 .data           = &sysctl_compact_memory,
1438                 .maxlen         = sizeof(int),
1439                 .mode           = 0200,
1440                 .proc_handler   = sysctl_compaction_handler,
1441         },
1442         {
1443                 .procname       = "extfrag_threshold",
1444                 .data           = &sysctl_extfrag_threshold,
1445                 .maxlen         = sizeof(int),
1446                 .mode           = 0644,
1447                 .proc_handler   = sysctl_extfrag_handler,
1448                 .extra1         = &min_extfrag_threshold,
1449                 .extra2         = &max_extfrag_threshold,
1450         },
1451         {
1452                 .procname       = "compact_unevictable_allowed",
1453                 .data           = &sysctl_compact_unevictable_allowed,
1454                 .maxlen         = sizeof(int),
1455                 .mode           = 0644,
1456                 .proc_handler   = proc_dointvec,
1457                 .extra1         = &zero,
1458                 .extra2         = &one,
1459         },
1460
1461 #endif /* CONFIG_COMPACTION */
1462         {
1463                 .procname       = "min_free_kbytes",
1464                 .data           = &min_free_kbytes,
1465                 .maxlen         = sizeof(min_free_kbytes),
1466                 .mode           = 0644,
1467                 .proc_handler   = min_free_kbytes_sysctl_handler,
1468                 .extra1         = &zero,
1469         },
1470         {
1471                 .procname       = "extra_free_kbytes",
1472                 .data           = &extra_free_kbytes,
1473                 .maxlen         = sizeof(extra_free_kbytes),
1474                 .mode           = 0644,
1475                 .proc_handler   = min_free_kbytes_sysctl_handler,
1476                 .extra1         = &zero,
1477         },
1478         {
1479                 .procname       = "percpu_pagelist_fraction",
1480                 .data           = &percpu_pagelist_fraction,
1481                 .maxlen         = sizeof(percpu_pagelist_fraction),
1482                 .mode           = 0644,
1483                 .proc_handler   = percpu_pagelist_fraction_sysctl_handler,
1484                 .extra1         = &zero,
1485         },
1486 #ifdef CONFIG_MMU
1487         {
1488                 .procname       = "max_map_count",
1489                 .data           = &sysctl_max_map_count,
1490                 .maxlen         = sizeof(sysctl_max_map_count),
1491                 .mode           = 0644,
1492                 .proc_handler   = proc_dointvec_minmax,
1493                 .extra1         = &zero,
1494         },
1495 #else
1496         {
1497                 .procname       = "nr_trim_pages",
1498                 .data           = &sysctl_nr_trim_pages,
1499                 .maxlen         = sizeof(sysctl_nr_trim_pages),
1500                 .mode           = 0644,
1501                 .proc_handler   = proc_dointvec_minmax,
1502                 .extra1         = &zero,
1503         },
1504 #endif
1505         {
1506                 .procname       = "laptop_mode",
1507                 .data           = &laptop_mode,
1508                 .maxlen         = sizeof(laptop_mode),
1509                 .mode           = 0644,
1510                 .proc_handler   = proc_dointvec_jiffies,
1511         },
1512         {
1513                 .procname       = "block_dump",
1514                 .data           = &block_dump,
1515                 .maxlen         = sizeof(block_dump),
1516                 .mode           = 0644,
1517                 .proc_handler   = proc_dointvec,
1518                 .extra1         = &zero,
1519         },
1520         {
1521                 .procname       = "vfs_cache_pressure",
1522                 .data           = &sysctl_vfs_cache_pressure,
1523                 .maxlen         = sizeof(sysctl_vfs_cache_pressure),
1524                 .mode           = 0644,
1525                 .proc_handler   = proc_dointvec,
1526                 .extra1         = &zero,
1527         },
1528 #ifdef HAVE_ARCH_PICK_MMAP_LAYOUT
1529         {
1530                 .procname       = "legacy_va_layout",
1531                 .data           = &sysctl_legacy_va_layout,
1532                 .maxlen         = sizeof(sysctl_legacy_va_layout),
1533                 .mode           = 0644,
1534                 .proc_handler   = proc_dointvec,
1535                 .extra1         = &zero,
1536         },
1537 #endif
1538 #ifdef CONFIG_NUMA
1539         {
1540                 .procname       = "zone_reclaim_mode",
1541                 .data           = &zone_reclaim_mode,
1542                 .maxlen         = sizeof(zone_reclaim_mode),
1543                 .mode           = 0644,
1544                 .proc_handler   = proc_dointvec,
1545                 .extra1         = &zero,
1546         },
1547         {
1548                 .procname       = "min_unmapped_ratio",
1549                 .data           = &sysctl_min_unmapped_ratio,
1550                 .maxlen         = sizeof(sysctl_min_unmapped_ratio),
1551                 .mode           = 0644,
1552                 .proc_handler   = sysctl_min_unmapped_ratio_sysctl_handler,
1553                 .extra1         = &zero,
1554                 .extra2         = &one_hundred,
1555         },
1556         {
1557                 .procname       = "min_slab_ratio",
1558                 .data           = &sysctl_min_slab_ratio,
1559                 .maxlen         = sizeof(sysctl_min_slab_ratio),
1560                 .mode           = 0644,
1561                 .proc_handler   = sysctl_min_slab_ratio_sysctl_handler,
1562                 .extra1         = &zero,
1563                 .extra2         = &one_hundred,
1564         },
1565 #endif
1566 #ifdef CONFIG_SMP
1567         {
1568                 .procname       = "stat_interval",
1569                 .data           = &sysctl_stat_interval,
1570                 .maxlen         = sizeof(sysctl_stat_interval),
1571                 .mode           = 0644,
1572                 .proc_handler   = proc_dointvec_jiffies,
1573         },
1574 #endif
1575 #ifdef CONFIG_MMU
1576         {
1577                 .procname       = "mmap_min_addr",
1578                 .data           = &dac_mmap_min_addr,
1579                 .maxlen         = sizeof(unsigned long),
1580                 .mode           = 0644,
1581                 .proc_handler   = mmap_min_addr_handler,
1582         },
1583 #endif
1584 #ifdef CONFIG_NUMA
1585         {
1586                 .procname       = "numa_zonelist_order",
1587                 .data           = &numa_zonelist_order,
1588                 .maxlen         = NUMA_ZONELIST_ORDER_LEN,
1589                 .mode           = 0644,
1590                 .proc_handler   = numa_zonelist_order_handler,
1591         },
1592 #endif
1593 #if (defined(CONFIG_X86_32) && !defined(CONFIG_UML))|| \
1594    (defined(CONFIG_SUPERH) && defined(CONFIG_VSYSCALL))
1595         {
1596                 .procname       = "vdso_enabled",
1597 #ifdef CONFIG_X86_32
1598                 .data           = &vdso32_enabled,
1599                 .maxlen         = sizeof(vdso32_enabled),
1600 #else
1601                 .data           = &vdso_enabled,
1602                 .maxlen         = sizeof(vdso_enabled),
1603 #endif
1604                 .mode           = 0644,
1605                 .proc_handler   = proc_dointvec,
1606                 .extra1         = &zero,
1607         },
1608 #endif
1609 #ifdef CONFIG_HIGHMEM
1610         {
1611                 .procname       = "highmem_is_dirtyable",
1612                 .data           = &vm_highmem_is_dirtyable,
1613                 .maxlen         = sizeof(vm_highmem_is_dirtyable),
1614                 .mode           = 0644,
1615                 .proc_handler   = proc_dointvec_minmax,
1616                 .extra1         = &zero,
1617                 .extra2         = &one,
1618         },
1619 #endif
1620 #ifdef CONFIG_MEMORY_FAILURE
1621         {
1622                 .procname       = "memory_failure_early_kill",
1623                 .data           = &sysctl_memory_failure_early_kill,
1624                 .maxlen         = sizeof(sysctl_memory_failure_early_kill),
1625                 .mode           = 0644,
1626                 .proc_handler   = proc_dointvec_minmax,
1627                 .extra1         = &zero,
1628                 .extra2         = &one,
1629         },
1630         {
1631                 .procname       = "memory_failure_recovery",
1632                 .data           = &sysctl_memory_failure_recovery,
1633                 .maxlen         = sizeof(sysctl_memory_failure_recovery),
1634                 .mode           = 0644,
1635                 .proc_handler   = proc_dointvec_minmax,
1636                 .extra1         = &zero,
1637                 .extra2         = &one,
1638         },
1639 #endif
1640         {
1641                 .procname       = "user_reserve_kbytes",
1642                 .data           = &sysctl_user_reserve_kbytes,
1643                 .maxlen         = sizeof(sysctl_user_reserve_kbytes),
1644                 .mode           = 0644,
1645                 .proc_handler   = proc_doulongvec_minmax,
1646         },
1647         {
1648                 .procname       = "admin_reserve_kbytes",
1649                 .data           = &sysctl_admin_reserve_kbytes,
1650                 .maxlen         = sizeof(sysctl_admin_reserve_kbytes),
1651                 .mode           = 0644,
1652                 .proc_handler   = proc_doulongvec_minmax,
1653         },
1654 #ifdef CONFIG_HAVE_ARCH_MMAP_RND_BITS
1655         {
1656                 .procname       = "mmap_rnd_bits",
1657                 .data           = &mmap_rnd_bits,
1658                 .maxlen         = sizeof(mmap_rnd_bits),
1659                 .mode           = 0600,
1660                 .proc_handler   = proc_dointvec_minmax,
1661                 .extra1         = (void *)&mmap_rnd_bits_min,
1662                 .extra2         = (void *)&mmap_rnd_bits_max,
1663         },
1664 #endif
1665 #ifdef CONFIG_HAVE_ARCH_MMAP_RND_COMPAT_BITS
1666         {
1667                 .procname       = "mmap_rnd_compat_bits",
1668                 .data           = &mmap_rnd_compat_bits,
1669                 .maxlen         = sizeof(mmap_rnd_compat_bits),
1670                 .mode           = 0600,
1671                 .proc_handler   = proc_dointvec_minmax,
1672                 .extra1         = (void *)&mmap_rnd_compat_bits_min,
1673                 .extra2         = (void *)&mmap_rnd_compat_bits_max,
1674         },
1675 #endif
1676         { }
1677 };
1678
1679 static struct ctl_table fs_table[] = {
1680         {
1681                 .procname       = "inode-nr",
1682                 .data           = &inodes_stat,
1683                 .maxlen         = 2*sizeof(long),
1684                 .mode           = 0444,
1685                 .proc_handler   = proc_nr_inodes,
1686         },
1687         {
1688                 .procname       = "inode-state",
1689                 .data           = &inodes_stat,
1690                 .maxlen         = 7*sizeof(long),
1691                 .mode           = 0444,
1692                 .proc_handler   = proc_nr_inodes,
1693         },
1694         {
1695                 .procname       = "file-nr",
1696                 .data           = &files_stat,
1697                 .maxlen         = sizeof(files_stat),
1698                 .mode           = 0444,
1699                 .proc_handler   = proc_nr_files,
1700         },
1701         {
1702                 .procname       = "file-max",
1703                 .data           = &files_stat.max_files,
1704                 .maxlen         = sizeof(files_stat.max_files),
1705                 .mode           = 0644,
1706                 .proc_handler   = proc_doulongvec_minmax,
1707         },
1708         {
1709                 .procname       = "nr_open",
1710                 .data           = &sysctl_nr_open,
1711                 .maxlen         = sizeof(int),
1712                 .mode           = 0644,
1713                 .proc_handler   = proc_dointvec_minmax,
1714                 .extra1         = &sysctl_nr_open_min,
1715                 .extra2         = &sysctl_nr_open_max,
1716         },
1717         {
1718                 .procname       = "dentry-state",
1719                 .data           = &dentry_stat,
1720                 .maxlen         = 6*sizeof(long),
1721                 .mode           = 0444,
1722                 .proc_handler   = proc_nr_dentry,
1723         },
1724         {
1725                 .procname       = "overflowuid",
1726                 .data           = &fs_overflowuid,
1727                 .maxlen         = sizeof(int),
1728                 .mode           = 0644,
1729                 .proc_handler   = proc_dointvec_minmax,
1730                 .extra1         = &minolduid,
1731                 .extra2         = &maxolduid,
1732         },
1733         {
1734                 .procname       = "overflowgid",
1735                 .data           = &fs_overflowgid,
1736                 .maxlen         = sizeof(int),
1737                 .mode           = 0644,
1738                 .proc_handler   = proc_dointvec_minmax,
1739                 .extra1         = &minolduid,
1740                 .extra2         = &maxolduid,
1741         },
1742 #ifdef CONFIG_FILE_LOCKING
1743         {
1744                 .procname       = "leases-enable",
1745                 .data           = &leases_enable,
1746                 .maxlen         = sizeof(int),
1747                 .mode           = 0644,
1748                 .proc_handler   = proc_dointvec,
1749         },
1750 #endif
1751 #ifdef CONFIG_DNOTIFY
1752         {
1753                 .procname       = "dir-notify-enable",
1754                 .data           = &dir_notify_enable,
1755                 .maxlen         = sizeof(int),
1756                 .mode           = 0644,
1757                 .proc_handler   = proc_dointvec,
1758         },
1759 #endif
1760 #ifdef CONFIG_MMU
1761 #ifdef CONFIG_FILE_LOCKING
1762         {
1763                 .procname       = "lease-break-time",
1764                 .data           = &lease_break_time,
1765                 .maxlen         = sizeof(int),
1766                 .mode           = 0644,
1767                 .proc_handler   = proc_dointvec,
1768         },
1769 #endif
1770 #ifdef CONFIG_AIO
1771         {
1772                 .procname       = "aio-nr",
1773                 .data           = &aio_nr,
1774                 .maxlen         = sizeof(aio_nr),
1775                 .mode           = 0444,
1776                 .proc_handler   = proc_doulongvec_minmax,
1777         },
1778         {
1779                 .procname       = "aio-max-nr",
1780                 .data           = &aio_max_nr,
1781                 .maxlen         = sizeof(aio_max_nr),
1782                 .mode           = 0644,
1783                 .proc_handler   = proc_doulongvec_minmax,
1784         },
1785 #endif /* CONFIG_AIO */
1786 #ifdef CONFIG_INOTIFY_USER
1787         {
1788                 .procname       = "inotify",
1789                 .mode           = 0555,
1790                 .child          = inotify_table,
1791         },
1792 #endif
1793 #ifdef CONFIG_EPOLL
1794         {
1795                 .procname       = "epoll",
1796                 .mode           = 0555,
1797                 .child          = epoll_table,
1798         },
1799 #endif
1800 #endif
1801         {
1802                 .procname       = "protected_symlinks",
1803                 .data           = &sysctl_protected_symlinks,
1804                 .maxlen         = sizeof(int),
1805                 .mode           = 0600,
1806                 .proc_handler   = proc_dointvec_minmax,
1807                 .extra1         = &zero,
1808                 .extra2         = &one,
1809         },
1810         {
1811                 .procname       = "protected_hardlinks",
1812                 .data           = &sysctl_protected_hardlinks,
1813                 .maxlen         = sizeof(int),
1814                 .mode           = 0600,
1815                 .proc_handler   = proc_dointvec_minmax,
1816                 .extra1         = &zero,
1817                 .extra2         = &one,
1818         },
1819         {
1820                 .procname       = "suid_dumpable",
1821                 .data           = &suid_dumpable,
1822                 .maxlen         = sizeof(int),
1823                 .mode           = 0644,
1824                 .proc_handler   = proc_dointvec_minmax_coredump,
1825                 .extra1         = &zero,
1826                 .extra2         = &two,
1827         },
1828 #if defined(CONFIG_BINFMT_MISC) || defined(CONFIG_BINFMT_MISC_MODULE)
1829         {
1830                 .procname       = "binfmt_misc",
1831                 .mode           = 0555,
1832                 .child          = sysctl_mount_point,
1833         },
1834 #endif
1835         {
1836                 .procname       = "pipe-max-size",
1837                 .data           = &pipe_max_size,
1838                 .maxlen         = sizeof(int),
1839                 .mode           = 0644,
1840                 .proc_handler   = &pipe_proc_fn,
1841                 .extra1         = &pipe_min_size,
1842         },
1843         {
1844                 .procname       = "pipe-user-pages-hard",
1845                 .data           = &pipe_user_pages_hard,
1846                 .maxlen         = sizeof(pipe_user_pages_hard),
1847                 .mode           = 0644,
1848                 .proc_handler   = proc_doulongvec_minmax,
1849         },
1850         {
1851                 .procname       = "pipe-user-pages-soft",
1852                 .data           = &pipe_user_pages_soft,
1853                 .maxlen         = sizeof(pipe_user_pages_soft),
1854                 .mode           = 0644,
1855                 .proc_handler   = proc_doulongvec_minmax,
1856         },
1857         {
1858                 .procname       = "mount-max",
1859                 .data           = &sysctl_mount_max,
1860                 .maxlen         = sizeof(unsigned int),
1861                 .mode           = 0644,
1862                 .proc_handler   = proc_dointvec_minmax,
1863                 .extra1         = &one,
1864         },
1865         { }
1866 };
1867
1868 static struct ctl_table debug_table[] = {
1869 #ifdef CONFIG_SYSCTL_EXCEPTION_TRACE
1870         {
1871                 .procname       = "exception-trace",
1872                 .data           = &show_unhandled_signals,
1873                 .maxlen         = sizeof(int),
1874                 .mode           = 0644,
1875                 .proc_handler   = proc_dointvec
1876         },
1877 #endif
1878 #if defined(CONFIG_OPTPROBES)
1879         {
1880                 .procname       = "kprobes-optimization",
1881                 .data           = &sysctl_kprobes_optimization,
1882                 .maxlen         = sizeof(int),
1883                 .mode           = 0644,
1884                 .proc_handler   = proc_kprobes_optimization_handler,
1885                 .extra1         = &zero,
1886                 .extra2         = &one,
1887         },
1888 #endif
1889         { }
1890 };
1891
1892 static struct ctl_table dev_table[] = {
1893         { }
1894 };
1895
1896 int __init sysctl_init(void)
1897 {
1898         struct ctl_table_header *hdr;
1899
1900         hdr = register_sysctl_table(sysctl_base_table);
1901         kmemleak_not_leak(hdr);
1902         return 0;
1903 }
1904
1905 #endif /* CONFIG_SYSCTL */
1906
1907 /*
1908  * /proc/sys support
1909  */
1910
1911 #ifdef CONFIG_PROC_SYSCTL
1912
1913 static int _proc_do_string(char *data, int maxlen, int write,
1914                            char __user *buffer,
1915                            size_t *lenp, loff_t *ppos)
1916 {
1917         size_t len;
1918         char __user *p;
1919         char c;
1920
1921         if (!data || !maxlen || !*lenp) {
1922                 *lenp = 0;
1923                 return 0;
1924         }
1925
1926         if (write) {
1927                 if (sysctl_writes_strict == SYSCTL_WRITES_STRICT) {
1928                         /* Only continue writes not past the end of buffer. */
1929                         len = strlen(data);
1930                         if (len > maxlen - 1)
1931                                 len = maxlen - 1;
1932
1933                         if (*ppos > len)
1934                                 return 0;
1935                         len = *ppos;
1936                 } else {
1937                         /* Start writing from beginning of buffer. */
1938                         len = 0;
1939                 }
1940
1941                 *ppos += *lenp;
1942                 p = buffer;
1943                 while ((p - buffer) < *lenp && len < maxlen - 1) {
1944                         if (get_user(c, p++))
1945                                 return -EFAULT;
1946                         if (c == 0 || c == '\n')
1947                                 break;
1948                         data[len++] = c;
1949                 }
1950                 data[len] = 0;
1951         } else {
1952                 len = strlen(data);
1953                 if (len > maxlen)
1954                         len = maxlen;
1955
1956                 if (*ppos > len) {
1957                         *lenp = 0;
1958                         return 0;
1959                 }
1960
1961                 data += *ppos;
1962                 len  -= *ppos;
1963
1964                 if (len > *lenp)
1965                         len = *lenp;
1966                 if (len)
1967                         if (copy_to_user(buffer, data, len))
1968                                 return -EFAULT;
1969                 if (len < *lenp) {
1970                         if (put_user('\n', buffer + len))
1971                                 return -EFAULT;
1972                         len++;
1973                 }
1974                 *lenp = len;
1975                 *ppos += len;
1976         }
1977         return 0;
1978 }
1979
1980 static void warn_sysctl_write(struct ctl_table *table)
1981 {
1982         pr_warn_once("%s wrote to %s when file position was not 0!\n"
1983                 "This will not be supported in the future. To silence this\n"
1984                 "warning, set kernel.sysctl_writes_strict = -1\n",
1985                 current->comm, table->procname);
1986 }
1987
1988 /**
1989  * proc_dostring - read a string sysctl
1990  * @table: the sysctl table
1991  * @write: %TRUE if this is a write to the sysctl file
1992  * @buffer: the user buffer
1993  * @lenp: the size of the user buffer
1994  * @ppos: file position
1995  *
1996  * Reads/writes a string from/to the user buffer. If the kernel
1997  * buffer provided is not large enough to hold the string, the
1998  * string is truncated. The copied string is %NULL-terminated.
1999  * If the string is being read by the user process, it is copied
2000  * and a newline '\n' is added. It is truncated if the buffer is
2001  * not large enough.
2002  *
2003  * Returns 0 on success.
2004  */
2005 int proc_dostring(struct ctl_table *table, int write,
2006                   void __user *buffer, size_t *lenp, loff_t *ppos)
2007 {
2008         if (write && *ppos && sysctl_writes_strict == SYSCTL_WRITES_WARN)
2009                 warn_sysctl_write(table);
2010
2011         return _proc_do_string((char *)(table->data), table->maxlen, write,
2012                                (char __user *)buffer, lenp, ppos);
2013 }
2014
2015 static size_t proc_skip_spaces(char **buf)
2016 {
2017         size_t ret;
2018         char *tmp = skip_spaces(*buf);
2019         ret = tmp - *buf;
2020         *buf = tmp;
2021         return ret;
2022 }
2023
2024 static void proc_skip_char(char **buf, size_t *size, const char v)
2025 {
2026         while (*size) {
2027                 if (**buf != v)
2028                         break;
2029                 (*size)--;
2030                 (*buf)++;
2031         }
2032 }
2033
2034 #define TMPBUFLEN 22
2035 /**
2036  * proc_get_long - reads an ASCII formatted integer from a user buffer
2037  *
2038  * @buf: a kernel buffer
2039  * @size: size of the kernel buffer
2040  * @val: this is where the number will be stored
2041  * @neg: set to %TRUE if number is negative
2042  * @perm_tr: a vector which contains the allowed trailers
2043  * @perm_tr_len: size of the perm_tr vector
2044  * @tr: pointer to store the trailer character
2045  *
2046  * In case of success %0 is returned and @buf and @size are updated with
2047  * the amount of bytes read. If @tr is non-NULL and a trailing
2048  * character exists (size is non-zero after returning from this
2049  * function), @tr is updated with the trailing character.
2050  */
2051 static int proc_get_long(char **buf, size_t *size,
2052                           unsigned long *val, bool *neg,
2053                           const char *perm_tr, unsigned perm_tr_len, char *tr)
2054 {
2055         int len;
2056         char *p, tmp[TMPBUFLEN];
2057
2058         if (!*size)
2059                 return -EINVAL;
2060
2061         len = *size;
2062         if (len > TMPBUFLEN - 1)
2063                 len = TMPBUFLEN - 1;
2064
2065         memcpy(tmp, *buf, len);
2066
2067         tmp[len] = 0;
2068         p = tmp;
2069         if (*p == '-' && *size > 1) {
2070                 *neg = true;
2071                 p++;
2072         } else
2073                 *neg = false;
2074         if (!isdigit(*p))
2075                 return -EINVAL;
2076
2077         *val = simple_strtoul(p, &p, 0);
2078
2079         len = p - tmp;
2080
2081         /* We don't know if the next char is whitespace thus we may accept
2082          * invalid integers (e.g. 1234...a) or two integers instead of one
2083          * (e.g. 123...1). So lets not allow such large numbers. */
2084         if (len == TMPBUFLEN - 1)
2085                 return -EINVAL;
2086
2087         if (len < *size && perm_tr_len && !memchr(perm_tr, *p, perm_tr_len))
2088                 return -EINVAL;
2089
2090         if (tr && (len < *size))
2091                 *tr = *p;
2092
2093         *buf += len;
2094         *size -= len;
2095
2096         return 0;
2097 }
2098
2099 /**
2100  * proc_put_long - converts an integer to a decimal ASCII formatted string
2101  *
2102  * @buf: the user buffer
2103  * @size: the size of the user buffer
2104  * @val: the integer to be converted
2105  * @neg: sign of the number, %TRUE for negative
2106  *
2107  * In case of success %0 is returned and @buf and @size are updated with
2108  * the amount of bytes written.
2109  */
2110 static int proc_put_long(void __user **buf, size_t *size, unsigned long val,
2111                           bool neg)
2112 {
2113         int len;
2114         char tmp[TMPBUFLEN], *p = tmp;
2115
2116         sprintf(p, "%s%lu", neg ? "-" : "", val);
2117         len = strlen(tmp);
2118         if (len > *size)
2119                 len = *size;
2120         if (copy_to_user(*buf, tmp, len))
2121                 return -EFAULT;
2122         *size -= len;
2123         *buf += len;
2124         return 0;
2125 }
2126 #undef TMPBUFLEN
2127
2128 static int proc_put_char(void __user **buf, size_t *size, char c)
2129 {
2130         if (*size) {
2131                 char __user **buffer = (char __user **)buf;
2132                 if (put_user(c, *buffer))
2133                         return -EFAULT;
2134                 (*size)--, (*buffer)++;
2135                 *buf = *buffer;
2136         }
2137         return 0;
2138 }
2139
2140 static int do_proc_dointvec_conv(bool *negp, unsigned long *lvalp,
2141                                  int *valp,
2142                                  int write, void *data)
2143 {
2144         if (write) {
2145                 if (*negp) {
2146                         if (*lvalp > (unsigned long) INT_MAX + 1)
2147                                 return -EINVAL;
2148                         *valp = -*lvalp;
2149                 } else {
2150                         if (*lvalp > (unsigned long) INT_MAX)
2151                                 return -EINVAL;
2152                         *valp = *lvalp;
2153                 }
2154         } else {
2155                 int val = *valp;
2156                 if (val < 0) {
2157                         *negp = true;
2158                         *lvalp = -(unsigned long)val;
2159                 } else {
2160                         *negp = false;
2161                         *lvalp = (unsigned long)val;
2162                 }
2163         }
2164         return 0;
2165 }
2166
2167 static int do_proc_douintvec_conv(bool *negp, unsigned long *lvalp,
2168                                  int *valp,
2169                                  int write, void *data)
2170 {
2171         if (write) {
2172                 if (*negp)
2173                         return -EINVAL;
2174                 *valp = *lvalp;
2175         } else {
2176                 unsigned int val = *valp;
2177                 *lvalp = (unsigned long)val;
2178         }
2179         return 0;
2180 }
2181
2182 static const char proc_wspace_sep[] = { ' ', '\t', '\n' };
2183
2184 static int __do_proc_dointvec(void *tbl_data, struct ctl_table *table,
2185                   int write, void __user *buffer,
2186                   size_t *lenp, loff_t *ppos,
2187                   int (*conv)(bool *negp, unsigned long *lvalp, int *valp,
2188                               int write, void *data),
2189                   void *data)
2190 {
2191         int *i, vleft, first = 1, err = 0;
2192         unsigned long page = 0;
2193         size_t left;
2194         char *kbuf;
2195
2196         if (!tbl_data || !table->maxlen || !*lenp || (*ppos && !write)) {
2197                 *lenp = 0;
2198                 return 0;
2199         }
2200
2201         i = (int *) tbl_data;
2202         vleft = table->maxlen / sizeof(*i);
2203         left = *lenp;
2204
2205         if (!conv)
2206                 conv = do_proc_dointvec_conv;
2207
2208         if (write) {
2209                 if (*ppos) {
2210                         switch (sysctl_writes_strict) {
2211                         case SYSCTL_WRITES_STRICT:
2212                                 goto out;
2213                         case SYSCTL_WRITES_WARN:
2214                                 warn_sysctl_write(table);
2215                                 break;
2216                         default:
2217                                 break;
2218                         }
2219                 }
2220
2221                 if (left > PAGE_SIZE - 1)
2222                         left = PAGE_SIZE - 1;
2223                 page = __get_free_page(GFP_TEMPORARY);
2224                 kbuf = (char *) page;
2225                 if (!kbuf)
2226                         return -ENOMEM;
2227                 if (copy_from_user(kbuf, buffer, left)) {
2228                         err = -EFAULT;
2229                         goto free;
2230                 }
2231                 kbuf[left] = 0;
2232         }
2233
2234         for (; left && vleft--; i++, first=0) {
2235                 unsigned long lval;
2236                 bool neg;
2237
2238                 if (write) {
2239                         left -= proc_skip_spaces(&kbuf);
2240
2241                         if (!left)
2242                                 break;
2243                         err = proc_get_long(&kbuf, &left, &lval, &neg,
2244                                              proc_wspace_sep,
2245                                              sizeof(proc_wspace_sep), NULL);
2246                         if (err)
2247                                 break;
2248                         if (conv(&neg, &lval, i, 1, data)) {
2249                                 err = -EINVAL;
2250                                 break;
2251                         }
2252                 } else {
2253                         if (conv(&neg, &lval, i, 0, data)) {
2254                                 err = -EINVAL;
2255                                 break;
2256                         }
2257                         if (!first)
2258                                 err = proc_put_char(&buffer, &left, '\t');
2259                         if (err)
2260                                 break;
2261                         err = proc_put_long(&buffer, &left, lval, neg);
2262                         if (err)
2263                                 break;
2264                 }
2265         }
2266
2267         if (!write && !first && left && !err)
2268                 err = proc_put_char(&buffer, &left, '\n');
2269         if (write && !err && left)
2270                 left -= proc_skip_spaces(&kbuf);
2271 free:
2272         if (write) {
2273                 free_page(page);
2274                 if (first)
2275                         return err ? : -EINVAL;
2276         }
2277         *lenp -= left;
2278 out:
2279         *ppos += *lenp;
2280         return err;
2281 }
2282
2283 static int do_proc_dointvec(struct ctl_table *table, int write,
2284                   void __user *buffer, size_t *lenp, loff_t *ppos,
2285                   int (*conv)(bool *negp, unsigned long *lvalp, int *valp,
2286                               int write, void *data),
2287                   void *data)
2288 {
2289         return __do_proc_dointvec(table->data, table, write,
2290                         buffer, lenp, ppos, conv, data);
2291 }
2292
2293 /**
2294  * proc_dointvec - read a vector of integers
2295  * @table: the sysctl table
2296  * @write: %TRUE if this is a write to the sysctl file
2297  * @buffer: the user buffer
2298  * @lenp: the size of the user buffer
2299  * @ppos: file position
2300  *
2301  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2302  * values from/to the user buffer, treated as an ASCII string.
2303  *
2304  * Returns 0 on success.
2305  */
2306 int proc_dointvec(struct ctl_table *table, int write,
2307                      void __user *buffer, size_t *lenp, loff_t *ppos)
2308 {
2309         return do_proc_dointvec(table, write, buffer, lenp, ppos, NULL, NULL);
2310 }
2311
2312 /**
2313  * proc_douintvec - read a vector of unsigned integers
2314  * @table: the sysctl table
2315  * @write: %TRUE if this is a write to the sysctl file
2316  * @buffer: the user buffer
2317  * @lenp: the size of the user buffer
2318  * @ppos: file position
2319  *
2320  * Reads/writes up to table->maxlen/sizeof(unsigned int) unsigned integer
2321  * values from/to the user buffer, treated as an ASCII string.
2322  *
2323  * Returns 0 on success.
2324  */
2325 int proc_douintvec(struct ctl_table *table, int write,
2326                      void __user *buffer, size_t *lenp, loff_t *ppos)
2327 {
2328         return do_proc_dointvec(table, write, buffer, lenp, ppos,
2329                                 do_proc_douintvec_conv, NULL);
2330 }
2331
2332 /*
2333  * Taint values can only be increased
2334  * This means we can safely use a temporary.
2335  */
2336 static int proc_taint(struct ctl_table *table, int write,
2337                                void __user *buffer, size_t *lenp, loff_t *ppos)
2338 {
2339         struct ctl_table t;
2340         unsigned long tmptaint = get_taint();
2341         int err;
2342
2343         if (write && !capable(CAP_SYS_ADMIN))
2344                 return -EPERM;
2345
2346         t = *table;
2347         t.data = &tmptaint;
2348         err = proc_doulongvec_minmax(&t, write, buffer, lenp, ppos);
2349         if (err < 0)
2350                 return err;
2351
2352         if (write) {
2353                 /*
2354                  * Poor man's atomic or. Not worth adding a primitive
2355                  * to everyone's atomic.h for this
2356                  */
2357                 int i;
2358                 for (i = 0; i < BITS_PER_LONG && tmptaint >> i; i++) {
2359                         if ((tmptaint >> i) & 1)
2360                                 add_taint(i, LOCKDEP_STILL_OK);
2361                 }
2362         }
2363
2364         return err;
2365 }
2366
2367 #ifdef CONFIG_PRINTK
2368 static int proc_dointvec_minmax_sysadmin(struct ctl_table *table, int write,
2369                                 void __user *buffer, size_t *lenp, loff_t *ppos)
2370 {
2371         if (write && !capable(CAP_SYS_ADMIN))
2372                 return -EPERM;
2373
2374         return proc_dointvec_minmax(table, write, buffer, lenp, ppos);
2375 }
2376 #endif
2377
2378 struct do_proc_dointvec_minmax_conv_param {
2379         int *min;
2380         int *max;
2381 };
2382
2383 static int do_proc_dointvec_minmax_conv(bool *negp, unsigned long *lvalp,
2384                                         int *valp,
2385                                         int write, void *data)
2386 {
2387         struct do_proc_dointvec_minmax_conv_param *param = data;
2388         if (write) {
2389                 int val = *negp ? -*lvalp : *lvalp;
2390                 if ((param->min && *param->min > val) ||
2391                     (param->max && *param->max < val))
2392                         return -EINVAL;
2393                 *valp = val;
2394         } else {
2395                 int val = *valp;
2396                 if (val < 0) {
2397                         *negp = true;
2398                         *lvalp = -(unsigned long)val;
2399                 } else {
2400                         *negp = false;
2401                         *lvalp = (unsigned long)val;
2402                 }
2403         }
2404         return 0;
2405 }
2406
2407 /**
2408  * proc_dointvec_minmax - read a vector of integers with min/max values
2409  * @table: the sysctl table
2410  * @write: %TRUE if this is a write to the sysctl file
2411  * @buffer: the user buffer
2412  * @lenp: the size of the user buffer
2413  * @ppos: file position
2414  *
2415  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2416  * values from/to the user buffer, treated as an ASCII string.
2417  *
2418  * This routine will ensure the values are within the range specified by
2419  * table->extra1 (min) and table->extra2 (max).
2420  *
2421  * Returns 0 on success.
2422  */
2423 int proc_dointvec_minmax(struct ctl_table *table, int write,
2424                   void __user *buffer, size_t *lenp, loff_t *ppos)
2425 {
2426         struct do_proc_dointvec_minmax_conv_param param = {
2427                 .min = (int *) table->extra1,
2428                 .max = (int *) table->extra2,
2429         };
2430         return do_proc_dointvec(table, write, buffer, lenp, ppos,
2431                                 do_proc_dointvec_minmax_conv, &param);
2432 }
2433
2434 static void validate_coredump_safety(void)
2435 {
2436 #ifdef CONFIG_COREDUMP
2437         if (suid_dumpable == SUID_DUMP_ROOT &&
2438             core_pattern[0] != '/' && core_pattern[0] != '|') {
2439                 printk(KERN_WARNING "Unsafe core_pattern used with "\
2440                         "suid_dumpable=2. Pipe handler or fully qualified "\
2441                         "core dump path required.\n");
2442         }
2443 #endif
2444 }
2445
2446 static int proc_dointvec_minmax_coredump(struct ctl_table *table, int write,
2447                 void __user *buffer, size_t *lenp, loff_t *ppos)
2448 {
2449         int error = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
2450         if (!error)
2451                 validate_coredump_safety();
2452         return error;
2453 }
2454
2455 #ifdef CONFIG_COREDUMP
2456 static int proc_dostring_coredump(struct ctl_table *table, int write,
2457                   void __user *buffer, size_t *lenp, loff_t *ppos)
2458 {
2459         int error = proc_dostring(table, write, buffer, lenp, ppos);
2460         if (!error)
2461                 validate_coredump_safety();
2462         return error;
2463 }
2464 #endif
2465
2466 static int __do_proc_doulongvec_minmax(void *data, struct ctl_table *table, int write,
2467                                      void __user *buffer,
2468                                      size_t *lenp, loff_t *ppos,
2469                                      unsigned long convmul,
2470                                      unsigned long convdiv)
2471 {
2472         unsigned long *i, *min, *max;
2473         int vleft, first = 1, err = 0;
2474         unsigned long page = 0;
2475         size_t left;
2476         char *kbuf;
2477
2478         if (!data || !table->maxlen || !*lenp || (*ppos && !write)) {
2479                 *lenp = 0;
2480                 return 0;
2481         }
2482
2483         i = (unsigned long *) data;
2484         min = (unsigned long *) table->extra1;
2485         max = (unsigned long *) table->extra2;
2486         vleft = table->maxlen / sizeof(unsigned long);
2487         left = *lenp;
2488
2489         if (write) {
2490                 if (*ppos) {
2491                         switch (sysctl_writes_strict) {
2492                         case SYSCTL_WRITES_STRICT:
2493                                 goto out;
2494                         case SYSCTL_WRITES_WARN:
2495                                 warn_sysctl_write(table);
2496                                 break;
2497                         default:
2498                                 break;
2499                         }
2500                 }
2501
2502                 if (left > PAGE_SIZE - 1)
2503                         left = PAGE_SIZE - 1;
2504                 page = __get_free_page(GFP_TEMPORARY);
2505                 kbuf = (char *) page;
2506                 if (!kbuf)
2507                         return -ENOMEM;
2508                 if (copy_from_user(kbuf, buffer, left)) {
2509                         err = -EFAULT;
2510                         goto free;
2511                 }
2512                 kbuf[left] = 0;
2513         }
2514
2515         for (; left && vleft--; i++, first = 0) {
2516                 unsigned long val;
2517
2518                 if (write) {
2519                         bool neg;
2520
2521                         left -= proc_skip_spaces(&kbuf);
2522
2523                         err = proc_get_long(&kbuf, &left, &val, &neg,
2524                                              proc_wspace_sep,
2525                                              sizeof(proc_wspace_sep), NULL);
2526                         if (err)
2527                                 break;
2528                         if (neg)
2529                                 continue;
2530                         val = convmul * val / convdiv;
2531                         if ((min && val < *min) || (max && val > *max))
2532                                 continue;
2533                         *i = val;
2534                 } else {
2535                         val = convdiv * (*i) / convmul;
2536                         if (!first) {
2537                                 err = proc_put_char(&buffer, &left, '\t');
2538                                 if (err)
2539                                         break;
2540                         }
2541                         err = proc_put_long(&buffer, &left, val, false);
2542                         if (err)
2543                                 break;
2544                 }
2545         }
2546
2547         if (!write && !first && left && !err)
2548                 err = proc_put_char(&buffer, &left, '\n');
2549         if (write && !err)
2550                 left -= proc_skip_spaces(&kbuf);
2551 free:
2552         if (write) {
2553                 free_page(page);
2554                 if (first)
2555                         return err ? : -EINVAL;
2556         }
2557         *lenp -= left;
2558 out:
2559         *ppos += *lenp;
2560         return err;
2561 }
2562
2563 static int do_proc_doulongvec_minmax(struct ctl_table *table, int write,
2564                                      void __user *buffer,
2565                                      size_t *lenp, loff_t *ppos,
2566                                      unsigned long convmul,
2567                                      unsigned long convdiv)
2568 {
2569         return __do_proc_doulongvec_minmax(table->data, table, write,
2570                         buffer, lenp, ppos, convmul, convdiv);
2571 }
2572
2573 /**
2574  * proc_doulongvec_minmax - read a vector of long integers with min/max values
2575  * @table: the sysctl table
2576  * @write: %TRUE if this is a write to the sysctl file
2577  * @buffer: the user buffer
2578  * @lenp: the size of the user buffer
2579  * @ppos: file position
2580  *
2581  * Reads/writes up to table->maxlen/sizeof(unsigned long) unsigned long
2582  * values from/to the user buffer, treated as an ASCII string.
2583  *
2584  * This routine will ensure the values are within the range specified by
2585  * table->extra1 (min) and table->extra2 (max).
2586  *
2587  * Returns 0 on success.
2588  */
2589 int proc_doulongvec_minmax(struct ctl_table *table, int write,
2590                            void __user *buffer, size_t *lenp, loff_t *ppos)
2591 {
2592     return do_proc_doulongvec_minmax(table, write, buffer, lenp, ppos, 1l, 1l);
2593 }
2594
2595 /**
2596  * proc_doulongvec_ms_jiffies_minmax - read a vector of millisecond values with min/max values
2597  * @table: the sysctl table
2598  * @write: %TRUE if this is a write to the sysctl file
2599  * @buffer: the user buffer
2600  * @lenp: the size of the user buffer
2601  * @ppos: file position
2602  *
2603  * Reads/writes up to table->maxlen/sizeof(unsigned long) unsigned long
2604  * values from/to the user buffer, treated as an ASCII string. The values
2605  * are treated as milliseconds, and converted to jiffies when they are stored.
2606  *
2607  * This routine will ensure the values are within the range specified by
2608  * table->extra1 (min) and table->extra2 (max).
2609  *
2610  * Returns 0 on success.
2611  */
2612 int proc_doulongvec_ms_jiffies_minmax(struct ctl_table *table, int write,
2613                                       void __user *buffer,
2614                                       size_t *lenp, loff_t *ppos)
2615 {
2616     return do_proc_doulongvec_minmax(table, write, buffer,
2617                                      lenp, ppos, HZ, 1000l);
2618 }
2619
2620
2621 static int do_proc_dointvec_jiffies_conv(bool *negp, unsigned long *lvalp,
2622                                          int *valp,
2623                                          int write, void *data)
2624 {
2625         if (write) {
2626                 if (*lvalp > LONG_MAX / HZ)
2627                         return 1;
2628                 *valp = *negp ? -(*lvalp*HZ) : (*lvalp*HZ);
2629         } else {
2630                 int val = *valp;
2631                 unsigned long lval;
2632                 if (val < 0) {
2633                         *negp = true;
2634                         lval = -(unsigned long)val;
2635                 } else {
2636                         *negp = false;
2637                         lval = (unsigned long)val;
2638                 }
2639                 *lvalp = lval / HZ;
2640         }
2641         return 0;
2642 }
2643
2644 static int do_proc_dointvec_userhz_jiffies_conv(bool *negp, unsigned long *lvalp,
2645                                                 int *valp,
2646                                                 int write, void *data)
2647 {
2648         if (write) {
2649                 if (USER_HZ < HZ && *lvalp > (LONG_MAX / HZ) * USER_HZ)
2650                         return 1;
2651                 *valp = clock_t_to_jiffies(*negp ? -*lvalp : *lvalp);
2652         } else {
2653                 int val = *valp;
2654                 unsigned long lval;
2655                 if (val < 0) {
2656                         *negp = true;
2657                         lval = -(unsigned long)val;
2658                 } else {
2659                         *negp = false;
2660                         lval = (unsigned long)val;
2661                 }
2662                 *lvalp = jiffies_to_clock_t(lval);
2663         }
2664         return 0;
2665 }
2666
2667 static int do_proc_dointvec_ms_jiffies_conv(bool *negp, unsigned long *lvalp,
2668                                             int *valp,
2669                                             int write, void *data)
2670 {
2671         if (write) {
2672                 unsigned long jif = msecs_to_jiffies(*negp ? -*lvalp : *lvalp);
2673
2674                 if (jif > INT_MAX)
2675                         return 1;
2676                 *valp = (int)jif;
2677         } else {
2678                 int val = *valp;
2679                 unsigned long lval;
2680                 if (val < 0) {
2681                         *negp = true;
2682                         lval = -(unsigned long)val;
2683                 } else {
2684                         *negp = false;
2685                         lval = (unsigned long)val;
2686                 }
2687                 *lvalp = jiffies_to_msecs(lval);
2688         }
2689         return 0;
2690 }
2691
2692 /**
2693  * proc_dointvec_jiffies - read a vector of integers as seconds
2694  * @table: the sysctl table
2695  * @write: %TRUE if this is a write to the sysctl file
2696  * @buffer: the user buffer
2697  * @lenp: the size of the user buffer
2698  * @ppos: file position
2699  *
2700  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2701  * values from/to the user buffer, treated as an ASCII string.
2702  * The values read are assumed to be in seconds, and are converted into
2703  * jiffies.
2704  *
2705  * Returns 0 on success.
2706  */
2707 int proc_dointvec_jiffies(struct ctl_table *table, int write,
2708                           void __user *buffer, size_t *lenp, loff_t *ppos)
2709 {
2710     return do_proc_dointvec(table,write,buffer,lenp,ppos,
2711                             do_proc_dointvec_jiffies_conv,NULL);
2712 }
2713
2714 /**
2715  * proc_dointvec_userhz_jiffies - read a vector of integers as 1/USER_HZ seconds
2716  * @table: the sysctl table
2717  * @write: %TRUE if this is a write to the sysctl file
2718  * @buffer: the user buffer
2719  * @lenp: the size of the user buffer
2720  * @ppos: pointer to the file position
2721  *
2722  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2723  * values from/to the user buffer, treated as an ASCII string.
2724  * The values read are assumed to be in 1/USER_HZ seconds, and
2725  * are converted into jiffies.
2726  *
2727  * Returns 0 on success.
2728  */
2729 int proc_dointvec_userhz_jiffies(struct ctl_table *table, int write,
2730                                  void __user *buffer, size_t *lenp, loff_t *ppos)
2731 {
2732     return do_proc_dointvec(table,write,buffer,lenp,ppos,
2733                             do_proc_dointvec_userhz_jiffies_conv,NULL);
2734 }
2735
2736 /**
2737  * proc_dointvec_ms_jiffies - read a vector of integers as 1 milliseconds
2738  * @table: the sysctl table
2739  * @write: %TRUE if this is a write to the sysctl file
2740  * @buffer: the user buffer
2741  * @lenp: the size of the user buffer
2742  * @ppos: file position
2743  * @ppos: the current position in the file
2744  *
2745  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2746  * values from/to the user buffer, treated as an ASCII string.
2747  * The values read are assumed to be in 1/1000 seconds, and
2748  * are converted into jiffies.
2749  *
2750  * Returns 0 on success.
2751  */
2752 int proc_dointvec_ms_jiffies(struct ctl_table *table, int write,
2753                              void __user *buffer, size_t *lenp, loff_t *ppos)
2754 {
2755         return do_proc_dointvec(table, write, buffer, lenp, ppos,
2756                                 do_proc_dointvec_ms_jiffies_conv, NULL);
2757 }
2758
2759 static int proc_do_cad_pid(struct ctl_table *table, int write,
2760                            void __user *buffer, size_t *lenp, loff_t *ppos)
2761 {
2762         struct pid *new_pid;
2763         pid_t tmp;
2764         int r;
2765
2766         tmp = pid_vnr(cad_pid);
2767
2768         r = __do_proc_dointvec(&tmp, table, write, buffer,
2769                                lenp, ppos, NULL, NULL);
2770         if (r || !write)
2771                 return r;
2772
2773         new_pid = find_get_pid(tmp);
2774         if (!new_pid)
2775                 return -ESRCH;
2776
2777         put_pid(xchg(&cad_pid, new_pid));
2778         return 0;
2779 }
2780
2781 /**
2782  * proc_do_large_bitmap - read/write from/to a large bitmap
2783  * @table: the sysctl table
2784  * @write: %TRUE if this is a write to the sysctl file
2785  * @buffer: the user buffer
2786  * @lenp: the size of the user buffer
2787  * @ppos: file position
2788  *
2789  * The bitmap is stored at table->data and the bitmap length (in bits)
2790  * in table->maxlen.
2791  *
2792  * We use a range comma separated format (e.g. 1,3-4,10-10) so that
2793  * large bitmaps may be represented in a compact manner. Writing into
2794  * the file will clear the bitmap then update it with the given input.
2795  *
2796  * Returns 0 on success.
2797  */
2798 int proc_do_large_bitmap(struct ctl_table *table, int write,
2799                          void __user *buffer, size_t *lenp, loff_t *ppos)
2800 {
2801         int err = 0;
2802         bool first = 1;
2803         size_t left = *lenp;
2804         unsigned long bitmap_len = table->maxlen;
2805         unsigned long *bitmap = *(unsigned long **) table->data;
2806         unsigned long *tmp_bitmap = NULL;
2807         char tr_a[] = { '-', ',', '\n' }, tr_b[] = { ',', '\n', 0 }, c;
2808
2809         if (!bitmap || !bitmap_len || !left || (*ppos && !write)) {
2810                 *lenp = 0;
2811                 return 0;
2812         }
2813
2814         if (write) {
2815                 unsigned long page = 0;
2816                 char *kbuf;
2817
2818                 if (left > PAGE_SIZE - 1)
2819                         left = PAGE_SIZE - 1;
2820
2821                 page = __get_free_page(GFP_TEMPORARY);
2822                 kbuf = (char *) page;
2823                 if (!kbuf)
2824                         return -ENOMEM;
2825                 if (copy_from_user(kbuf, buffer, left)) {
2826                         free_page(page);
2827                         return -EFAULT;
2828                 }
2829                 kbuf[left] = 0;
2830
2831                 tmp_bitmap = kzalloc(BITS_TO_LONGS(bitmap_len) * sizeof(unsigned long),
2832                                      GFP_KERNEL);
2833                 if (!tmp_bitmap) {
2834                         free_page(page);
2835                         return -ENOMEM;
2836                 }
2837                 proc_skip_char(&kbuf, &left, '\n');
2838                 while (!err && left) {
2839                         unsigned long val_a, val_b;
2840                         bool neg;
2841
2842                         err = proc_get_long(&kbuf, &left, &val_a, &neg, tr_a,
2843                                              sizeof(tr_a), &c);
2844                         if (err)
2845                                 break;
2846                         if (val_a >= bitmap_len || neg) {
2847                                 err = -EINVAL;
2848                                 break;
2849                         }
2850
2851                         val_b = val_a;
2852                         if (left) {
2853                                 kbuf++;
2854                                 left--;
2855                         }
2856
2857                         if (c == '-') {
2858                                 err = proc_get_long(&kbuf, &left, &val_b,
2859                                                      &neg, tr_b, sizeof(tr_b),
2860                                                      &c);
2861                                 if (err)
2862                                         break;
2863                                 if (val_b >= bitmap_len || neg ||
2864                                     val_a > val_b) {
2865                                         err = -EINVAL;
2866                                         break;
2867                                 }
2868                                 if (left) {
2869                                         kbuf++;
2870                                         left--;
2871                                 }
2872                         }
2873
2874                         bitmap_set(tmp_bitmap, val_a, val_b - val_a + 1);
2875                         first = 0;
2876                         proc_skip_char(&kbuf, &left, '\n');
2877                 }
2878                 free_page(page);
2879         } else {
2880                 unsigned long bit_a, bit_b = 0;
2881
2882                 while (left) {
2883                         bit_a = find_next_bit(bitmap, bitmap_len, bit_b);
2884                         if (bit_a >= bitmap_len)
2885                                 break;
2886                         bit_b = find_next_zero_bit(bitmap, bitmap_len,
2887                                                    bit_a + 1) - 1;
2888
2889                         if (!first) {
2890                                 err = proc_put_char(&buffer, &left, ',');
2891                                 if (err)
2892                                         break;
2893                         }
2894                         err = proc_put_long(&buffer, &left, bit_a, false);
2895                         if (err)
2896                                 break;
2897                         if (bit_a != bit_b) {
2898                                 err = proc_put_char(&buffer, &left, '-');
2899                                 if (err)
2900                                         break;
2901                                 err = proc_put_long(&buffer, &left, bit_b, false);
2902                                 if (err)
2903                                         break;
2904                         }
2905
2906                         first = 0; bit_b++;
2907                 }
2908                 if (!err)
2909                         err = proc_put_char(&buffer, &left, '\n');
2910         }
2911
2912         if (!err) {
2913                 if (write) {
2914                         if (*ppos)
2915                                 bitmap_or(bitmap, bitmap, tmp_bitmap, bitmap_len);
2916                         else
2917                                 bitmap_copy(bitmap, tmp_bitmap, bitmap_len);
2918                 }
2919                 kfree(tmp_bitmap);
2920                 *lenp -= left;
2921                 *ppos += *lenp;
2922                 return 0;
2923         } else {
2924                 kfree(tmp_bitmap);
2925                 return err;
2926         }
2927 }
2928
2929 #else /* CONFIG_PROC_SYSCTL */
2930
2931 int proc_dostring(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(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_douintvec(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_dointvec_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_dointvec_jiffies(struct ctl_table *table, int write,
2956                     void __user *buffer, size_t *lenp, loff_t *ppos)
2957 {
2958         return -ENOSYS;
2959 }
2960
2961 int proc_dointvec_userhz_jiffies(struct ctl_table *table, int write,
2962                     void __user *buffer, size_t *lenp, loff_t *ppos)
2963 {
2964         return -ENOSYS;
2965 }
2966
2967 int proc_dointvec_ms_jiffies(struct ctl_table *table, int write,
2968                              void __user *buffer, size_t *lenp, loff_t *ppos)
2969 {
2970         return -ENOSYS;
2971 }
2972
2973 int proc_doulongvec_minmax(struct ctl_table *table, int write,
2974                     void __user *buffer, size_t *lenp, loff_t *ppos)
2975 {
2976         return -ENOSYS;
2977 }
2978
2979 int proc_doulongvec_ms_jiffies_minmax(struct ctl_table *table, int write,
2980                                       void __user *buffer,
2981                                       size_t *lenp, loff_t *ppos)
2982 {
2983     return -ENOSYS;
2984 }
2985
2986
2987 #endif /* CONFIG_PROC_SYSCTL */
2988
2989 /*
2990  * No sense putting this after each symbol definition, twice,
2991  * exception granted :-)
2992  */
2993 EXPORT_SYMBOL(proc_dointvec);
2994 EXPORT_SYMBOL(proc_douintvec);
2995 EXPORT_SYMBOL(proc_dointvec_jiffies);
2996 EXPORT_SYMBOL(proc_dointvec_minmax);
2997 EXPORT_SYMBOL(proc_dointvec_userhz_jiffies);
2998 EXPORT_SYMBOL(proc_dointvec_ms_jiffies);
2999 EXPORT_SYMBOL(proc_dostring);
3000 EXPORT_SYMBOL(proc_doulongvec_minmax);
3001 EXPORT_SYMBOL(proc_doulongvec_ms_jiffies_minmax);