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
  69 #include <asm/uaccess.h>
  70 #include <asm/processor.h>
  71
  72 #ifdef CONFIG_X86
  73 #include <asm/nmi.h>
  74 #include <asm/stacktrace.h>
  75 #include <asm/io.h>
  76 #endif
  77 #ifdef CONFIG_SPARC
  78 #include <asm/setup.h>
  79 #endif
  80 #ifdef CONFIG_BSD_PROCESS_ACCT
  81 #include <linux/acct.h>
  82 #endif
  83 #ifdef CONFIG_RT_MUTEXES
  84 #include <linux/rtmutex.h>
  85 #endif
  86 #if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_LOCK_STAT)
  87 #include <linux/lockdep.h>
  88 #endif
  89 #ifdef CONFIG_CHR_DEV_SG
  90 #include <scsi/sg.h>
  91 #endif
  92
  93 #ifdef CONFIG_LOCKUP_DETECTOR
  94 #include <linux/nmi.h>
  95 #endif
  96
  97 #if defined(CONFIG_SYSCTL)
  98
  99 /* External variables not in a header file. */
 100 extern int suid_dumpable;
 101 #ifdef CONFIG_COREDUMP
 102 extern int core_uses_pid;
 103 extern char core_pattern[];
 104 extern unsigned int core_pipe_limit;
 105 #endif
 106 extern int pid_max;
 107 extern int extra_free_kbytes;
 108 extern int pid_max_min, pid_max_max;
 109 extern int percpu_pagelist_fraction;
 110 extern int compat_log;
 111 extern int latencytop_enabled;
 112 extern int sysctl_nr_open_min, sysctl_nr_open_max;
 113 #ifndef CONFIG_MMU
 114 extern int sysctl_nr_trim_pages;
 115 #endif
 116
 117 /* Constants used for minimum and  maximum */
 118 #ifdef CONFIG_LOCKUP_DETECTOR
 119 static int sixty = 60;
 120 #endif
 121
 122 static int __maybe_unused neg_one = -1;
 123
 124 static int zero;
 125 static int __maybe_unused one = 1;
 126 static int __maybe_unused two = 2;
 127 static int __maybe_unused four = 4;
 128 static unsigned long one_ul = 1;
 129 static int one_hundred = 100;
 130 #ifdef CONFIG_PRINTK
 131 static int ten_thousand = 10000;
 132 #endif
 133
 134 /* this is needed for the proc_doulongvec_minmax of vm_dirty_bytes */
 135 static unsigned long dirty_bytes_min = 2 * PAGE_SIZE;
 136
 137 /* this is needed for the proc_dointvec_minmax for [fs_]overflow UID and GID */
 138 static int maxolduid = 65535;
 139 static int minolduid;
 140
 141 static int ngroups_max = NGROUPS_MAX;
 142 static const int cap_last_cap = CAP_LAST_CAP;
 143
 144 /*this is needed for proc_doulongvec_minmax of sysctl_hung_task_timeout_secs */
 145 #ifdef CONFIG_DETECT_HUNG_TASK
 146 static unsigned long hung_task_timeout_max = (LONG_MAX/HZ);
 147 #endif
 148
 149 #ifdef CONFIG_INOTIFY_USER
 150 #include <linux/inotify.h>
 151 #endif
 152 #ifdef CONFIG_SPARC
 153 #endif
 154
 155 #ifdef __hppa__
 156 extern int pwrsw_enabled;
 157 #endif
 158
 159 #ifdef CONFIG_SYSCTL_ARCH_UNALIGN_ALLOW
 160 extern int unaligned_enabled;
 161 #endif
 162
 163 #ifdef CONFIG_IA64
 164 extern int unaligned_dump_stack;
 165 #endif
 166
 167 #ifdef CONFIG_SYSCTL_ARCH_UNALIGN_NO_WARN
 168 extern int no_unaligned_warning;
 169 #endif
 170
 171 #ifdef CONFIG_PROC_SYSCTL
 172
 173 #define SYSCTL_WRITES_LEGACY    -1
 174 #define SYSCTL_WRITES_WARN       0
 175 #define SYSCTL_WRITES_STRICT     1
 176
 177 static int sysctl_writes_strict = SYSCTL_WRITES_WARN;
 178
 179 static int proc_do_cad_pid(struct ctl_table *table, int write,
 180                   void __user *buffer, size_t *lenp, loff_t *ppos);
 181 static int proc_taint(struct ctl_table *table, int write,
 182                                void __user *buffer, size_t *lenp, loff_t *ppos);
 183 #endif
 184
 185 #ifdef CONFIG_PRINTK
 186 static int proc_dointvec_minmax_sysadmin(struct ctl_table *table, int write,
 187                                 void __user *buffer, size_t *lenp, loff_t *ppos);
 188 #endif
 189
 190 static int proc_dointvec_minmax_coredump(struct ctl_table *table, int write,
 191                 void __user *buffer, size_t *lenp, loff_t *ppos);
 192 #ifdef CONFIG_COREDUMP
 193 static int proc_dostring_coredump(struct ctl_table *table, int write,
 194                 void __user *buffer, size_t *lenp, loff_t *ppos);
 195 #endif
 196
 197 #ifdef CONFIG_MAGIC_SYSRQ
 198 /* Note: sysrq code uses it's own private copy */
 199 static int __sysrq_enabled = CONFIG_MAGIC_SYSRQ_DEFAULT_ENABLE;
 200
 201 static int sysrq_sysctl_handler(struct ctl_table *table, int write,
 202                                 void __user *buffer, size_t *lenp,
 203                                 loff_t *ppos)
 204 {
 205         int error;
 206
 207         error = proc_dointvec(table, write, buffer, lenp, ppos);
 208         if (error)
 209                 return error;
 210
 211         if (write)
 212                 sysrq_toggle_support(__sysrq_enabled);
 213
 214         return 0;
 215 }
 216
 217 #endif
 218
 219 static struct ctl_table kern_table[];
 220 static struct ctl_table vm_table[];
 221 static struct ctl_table fs_table[];
 222 static struct ctl_table debug_table[];
 223 static struct ctl_table dev_table[];
 224 extern struct ctl_table random_table[];
 225 #ifdef CONFIG_EPOLL
 226 extern struct ctl_table epoll_table[];
 227 #endif
 228
 229 #ifdef HAVE_ARCH_PICK_MMAP_LAYOUT
 230 int sysctl_legacy_va_layout;
 231 #endif
 232
 233 /* The default sysctl tables: */
 234
 235 static struct ctl_table sysctl_base_table[] = {
 236         {
 237                 .procname       = "kernel",
 238                 .mode           = 0555,
 239                 .child          = kern_table,
 240         },
 241         {
 242                 .procname       = "vm",
 243                 .mode           = 0555,
 244                 .child          = vm_table,
 245         },
 246         {
 247                 .procname       = "fs",
 248                 .mode           = 0555,
 249                 .child          = fs_table,
 250         },
 251         {
 252                 .procname       = "debug",
 253                 .mode           = 0555,
 254                 .child          = debug_table,
 255         },
 256         {
 257                 .procname       = "dev",
 258                 .mode           = 0555,
 259                 .child          = dev_table,
 260         },
 261         { }
 262 };
 263
 264 #ifdef CONFIG_SCHED_DEBUG
 265 static int min_sched_granularity_ns = 100000;           /* 100 usecs */
 266 static int max_sched_granularity_ns = NSEC_PER_SEC;     /* 1 second */
 267 static int min_wakeup_granularity_ns;                   /* 0 usecs */
 268 static int max_wakeup_granularity_ns = NSEC_PER_SEC;    /* 1 second */
 269 #ifdef CONFIG_SMP
 270 static int min_sched_tunable_scaling = SCHED_TUNABLESCALING_NONE;
 271 static int max_sched_tunable_scaling = SCHED_TUNABLESCALING_END-1;
 272 #endif /* CONFIG_SMP */
 273 #endif /* CONFIG_SCHED_DEBUG */
 274
 275 #ifdef CONFIG_COMPACTION
 276 static int min_extfrag_threshold;
 277 static int max_extfrag_threshold = 1000;
 278 #endif
 279
 280 static struct ctl_table kern_table[] = {
 281         {
 282                 .procname       = "sched_child_runs_first",
 283                 .data           = &sysctl_sched_child_runs_first,
 284                 .maxlen         = sizeof(unsigned int),
 285                 .mode           = 0644,
 286                 .proc_handler   = proc_dointvec,
 287         },
 288 #ifdef CONFIG_SCHED_DEBUG
 289         {
 290                 .procname       = "sched_min_granularity_ns",
 291                 .data           = &sysctl_sched_min_granularity,
 292                 .maxlen         = sizeof(unsigned int),
 293                 .mode           = 0644,
 294                 .proc_handler   = sched_proc_update_handler,
 295                 .extra1         = &min_sched_granularity_ns,
 296                 .extra2         = &max_sched_granularity_ns,
 297         },
 298         {
 299                 .procname       = "sched_latency_ns",
 300                 .data           = &sysctl_sched_latency,
 301                 .maxlen         = sizeof(unsigned int),
 302                 .mode           = 0644,
 303                 .proc_handler   = sched_proc_update_handler,
 304                 .extra1         = &min_sched_granularity_ns,
 305                 .extra2         = &max_sched_granularity_ns,
 306         },
 307         {
 308                 .procname       = "sched_is_big_little",
 309                 .data           = &sysctl_sched_is_big_little,
 310                 .maxlen         = sizeof(unsigned int),
 311                 .mode           = 0644,
 312                 .proc_handler   = proc_dointvec,
 313         },
 314 #ifdef CONFIG_SCHED_WALT
 315         {
 316                 .procname       = "sched_use_walt_cpu_util",
 317                 .data           = &sysctl_sched_use_walt_cpu_util,
 318                 .maxlen         = sizeof(unsigned int),
 319                 .mode           = 0644,
 320                 .proc_handler   = proc_dointvec,
 321         },
 322         {
 323                 .procname       = "sched_use_walt_task_util",
 324                 .data           = &sysctl_sched_use_walt_task_util,
 325                 .maxlen         = sizeof(unsigned int),
 326                 .mode           = 0644,
 327                 .proc_handler   = proc_dointvec,
 328         },
 329         {
 330                 .procname       = "sched_walt_init_task_load_pct",
 331                 .data           = &sysctl_sched_walt_init_task_load_pct,
 332                 .maxlen         = sizeof(unsigned int),
 333                 .mode           = 0644,
 334                 .proc_handler   = proc_dointvec,
 335         },
 336         {
 337                 .procname       = "sched_walt_cpu_high_irqload",
 338                 .data           = &sysctl_sched_walt_cpu_high_irqload,
 339                 .maxlen         = sizeof(unsigned int),
 340                 .mode           = 0644,
 341                 .proc_handler   = proc_dointvec,
 342         },
 343 #endif
 344         {
 345                 .procname       = "sched_sync_hint_enable",
 346                 .data           = &sysctl_sched_sync_hint_enable,
 347                 .maxlen         = sizeof(unsigned int),
 348                 .mode           = 0644,
 349                 .proc_handler   = proc_dointvec,
 350         },
 351         {
 352                 .procname       = "sched_initial_task_util",
 353                 .data           = &sysctl_sched_initial_task_util,
 354                 .maxlen         = sizeof(unsigned int),
 355                 .mode           = 0644,
 356                 .proc_handler   = proc_dointvec,
 357         },
 358         {
 359                 .procname       = "sched_cstate_aware",
 360                 .data           = &sysctl_sched_cstate_aware,
 361                 .maxlen         = sizeof(unsigned int),
 362                 .mode           = 0644,
 363                 .proc_handler   = proc_dointvec,
 364         },
 365         {
 366                 .procname       = "sched_wakeup_granularity_ns",
 367                 .data           = &sysctl_sched_wakeup_granularity,
 368                 .maxlen         = sizeof(unsigned int),
 369                 .mode           = 0644,
 370                 .proc_handler   = sched_proc_update_handler,
 371                 .extra1         = &min_wakeup_granularity_ns,
 372                 .extra2         = &max_wakeup_granularity_ns,
 373         },
 374 #ifdef CONFIG_SMP
 375         {
 376                 .procname       = "sched_tunable_scaling",
 377                 .data           = &sysctl_sched_tunable_scaling,
 378                 .maxlen         = sizeof(enum sched_tunable_scaling),
 379                 .mode           = 0644,
 380                 .proc_handler   = sched_proc_update_handler,
 381                 .extra1         = &min_sched_tunable_scaling,
 382                 .extra2         = &max_sched_tunable_scaling,
 383         },
 384         {
 385                 .procname       = "sched_migration_cost_ns",
 386                 .data           = &sysctl_sched_migration_cost,
 387                 .maxlen         = sizeof(unsigned int),
 388                 .mode           = 0644,
 389                 .proc_handler   = proc_dointvec,
 390         },
 391         {
 392                 .procname       = "sched_nr_migrate",
 393                 .data           = &sysctl_sched_nr_migrate,
 394                 .maxlen         = sizeof(unsigned int),
 395                 .mode           = 0644,
 396                 .proc_handler   = proc_dointvec,
 397         },
 398         {
 399                 .procname       = "sched_time_avg_ms",
 400                 .data           = &sysctl_sched_time_avg,
 401                 .maxlen         = sizeof(unsigned int),
 402                 .mode           = 0644,
 403                 .proc_handler   = proc_dointvec,
 404         },
 405         {
 406                 .procname       = "sched_shares_window_ns",
 407                 .data           = &sysctl_sched_shares_window,
 408                 .maxlen         = sizeof(unsigned int),
 409                 .mode           = 0644,
 410                 .proc_handler   = proc_dointvec,
 411         },
 412 #endif /* CONFIG_SMP */
 413 #ifdef CONFIG_NUMA_BALANCING
 414         {
 415                 .procname       = "numa_balancing_scan_delay_ms",
 416                 .data           = &sysctl_numa_balancing_scan_delay,
 417                 .maxlen         = sizeof(unsigned int),
 418                 .mode           = 0644,
 419                 .proc_handler   = proc_dointvec,
 420         },
 421         {
 422                 .procname       = "numa_balancing_scan_period_min_ms",
 423                 .data           = &sysctl_numa_balancing_scan_period_min,
 424                 .maxlen         = sizeof(unsigned int),
 425                 .mode           = 0644,
 426                 .proc_handler   = proc_dointvec,
 427         },
 428         {
 429                 .procname       = "numa_balancing_scan_period_max_ms",
 430                 .data           = &sysctl_numa_balancing_scan_period_max,
 431                 .maxlen         = sizeof(unsigned int),
 432                 .mode           = 0644,
 433                 .proc_handler   = proc_dointvec,
 434         },
 435         {
 436                 .procname       = "numa_balancing_scan_size_mb",
 437                 .data           = &sysctl_numa_balancing_scan_size,
 438                 .maxlen         = sizeof(unsigned int),
 439                 .mode           = 0644,
 440                 .proc_handler   = proc_dointvec_minmax,
 441                 .extra1         = &one,
 442         },
 443         {
 444                 .procname       = "numa_balancing",
 445                 .data           = NULL, /* filled in by handler */
 446                 .maxlen         = sizeof(unsigned int),
 447                 .mode           = 0644,
 448                 .proc_handler   = sysctl_numa_balancing,
 449                 .extra1         = &zero,
 450                 .extra2         = &one,
 451         },
 452 #endif /* CONFIG_NUMA_BALANCING */
 453 #endif /* CONFIG_SCHED_DEBUG */
 454         {
 455                 .procname       = "sched_rt_period_us",
 456                 .data           = &sysctl_sched_rt_period,
 457                 .maxlen         = sizeof(unsigned int),
 458                 .mode           = 0644,
 459                 .proc_handler   = sched_rt_handler,
 460         },
 461         {
 462                 .procname       = "sched_rt_runtime_us",
 463                 .data           = &sysctl_sched_rt_runtime,
 464                 .maxlen         = sizeof(int),
 465                 .mode           = 0644,
 466                 .proc_handler   = sched_rt_handler,
 467         },
 468         {
 469                 .procname       = "sched_rr_timeslice_ms",
 470                 .data           = &sched_rr_timeslice,
 471                 .maxlen         = sizeof(int),
 472                 .mode           = 0644,
 473                 .proc_handler   = sched_rr_handler,
 474         },
 475 #ifdef CONFIG_SCHED_AUTOGROUP
 476         {
 477                 .procname       = "sched_autogroup_enabled",
 478                 .data           = &sysctl_sched_autogroup_enabled,
 479                 .maxlen         = sizeof(unsigned int),
 480                 .mode           = 0644,
 481                 .proc_handler   = proc_dointvec_minmax,
 482                 .extra1         = &zero,
 483                 .extra2         = &one,
 484         },
 485 #endif
 486 #ifdef CONFIG_CFS_BANDWIDTH
 487         {
 488                 .procname       = "sched_cfs_bandwidth_slice_us",
 489                 .data           = &sysctl_sched_cfs_bandwidth_slice,
 490                 .maxlen         = sizeof(unsigned int),
 491                 .mode           = 0644,
 492                 .proc_handler   = proc_dointvec_minmax,
 493                 .extra1         = &one,
 494         },
 495 #endif
 496 #ifdef CONFIG_SCHED_TUNE
 497         {
 498                 .procname       = "sched_cfs_boost",
 499                 .data           = &sysctl_sched_cfs_boost,
 500                 .maxlen         = sizeof(sysctl_sched_cfs_boost),
 501 #ifdef CONFIG_CGROUP_SCHEDTUNE
 502                 .mode           = 0444,
 503 #else
 504                 .mode           = 0644,
 505 #endif
 506                 .proc_handler   = &sysctl_sched_cfs_boost_handler,
 507                 .extra1         = &zero,
 508                 .extra2         = &one_hundred,
 509         },
 510 #endif
 511 #ifdef CONFIG_PROVE_LOCKING
 512         {
 513                 .procname       = "prove_locking",
 514                 .data           = &prove_locking,
 515                 .maxlen         = sizeof(int),
 516                 .mode           = 0644,
 517                 .proc_handler   = proc_dointvec,
 518         },
 519 #endif
 520 #ifdef CONFIG_LOCK_STAT
 521         {
 522                 .procname       = "lock_stat",
 523                 .data           = &lock_stat,
 524                 .maxlen         = sizeof(int),
 525                 .mode           = 0644,
 526                 .proc_handler   = proc_dointvec,
 527         },
 528 #endif
 529         {
 530                 .procname       = "panic",
 531                 .data           = &panic_timeout,
 532                 .maxlen         = sizeof(int),
 533                 .mode           = 0644,
 534                 .proc_handler   = proc_dointvec,
 535         },
 536 #ifdef CONFIG_COREDUMP
 537         {
 538                 .procname       = "core_uses_pid",
 539                 .data           = &core_uses_pid,
 540                 .maxlen         = sizeof(int),
 541                 .mode           = 0644,
 542                 .proc_handler   = proc_dointvec,
 543         },
 544         {
 545                 .procname       = "core_pattern",
 546                 .data           = core_pattern,
 547                 .maxlen         = CORENAME_MAX_SIZE,
 548                 .mode           = 0644,
 549                 .proc_handler   = proc_dostring_coredump,
 550         },
 551         {
 552                 .procname       = "core_pipe_limit",
 553                 .data           = &core_pipe_limit,
 554                 .maxlen         = sizeof(unsigned int),
 555                 .mode           = 0644,
 556                 .proc_handler   = proc_dointvec,
 557         },
 558 #endif
 559 #ifdef CONFIG_PROC_SYSCTL
 560         {
 561                 .procname       = "tainted",
 562                 .maxlen         = sizeof(long),
 563                 .mode           = 0644,
 564                 .proc_handler   = proc_taint,
 565         },
 566         {
 567                 .procname       = "sysctl_writes_strict",
 568                 .data           = &sysctl_writes_strict,
 569                 .maxlen         = sizeof(int),
 570                 .mode           = 0644,
 571                 .proc_handler   = proc_dointvec_minmax,
 572                 .extra1         = &neg_one,
 573                 .extra2         = &one,
 574         },
 575 #endif
 576 #ifdef CONFIG_LATENCYTOP
 577         {
 578                 .procname       = "latencytop",
 579                 .data           = &latencytop_enabled,
 580                 .maxlen         = sizeof(int),
 581                 .mode           = 0644,
 582                 .proc_handler   = proc_dointvec,
 583         },
 584 #endif
 585 #ifdef CONFIG_BLK_DEV_INITRD
 586         {
 587                 .procname       = "real-root-dev",
 588                 .data           = &real_root_dev,
 589                 .maxlen         = sizeof(int),
 590                 .mode           = 0644,
 591                 .proc_handler   = proc_dointvec,
 592         },
 593 #endif
 594         {
 595                 .procname       = "print-fatal-signals",
 596                 .data           = &print_fatal_signals,
 597                 .maxlen         = sizeof(int),
 598                 .mode           = 0644,
 599                 .proc_handler   = proc_dointvec,
 600         },
 601 #ifdef CONFIG_SPARC
 602         {
 603                 .procname       = "reboot-cmd",
 604                 .data           = reboot_command,
 605                 .maxlen         = 256,
 606                 .mode           = 0644,
 607                 .proc_handler   = proc_dostring,
 608         },
 609         {
 610                 .procname       = "stop-a",
 611                 .data           = &stop_a_enabled,
 612                 .maxlen         = sizeof (int),
 613                 .mode           = 0644,
 614                 .proc_handler   = proc_dointvec,
 615         },
 616         {
 617                 .procname       = "scons-poweroff",
 618                 .data           = &scons_pwroff,
 619                 .maxlen         = sizeof (int),
 620                 .mode           = 0644,
 621                 .proc_handler   = proc_dointvec,
 622         },
 623 #endif
 624 #ifdef CONFIG_SPARC64
 625         {
 626                 .procname       = "tsb-ratio",
 627                 .data           = &sysctl_tsb_ratio,
 628                 .maxlen         = sizeof (int),
 629                 .mode           = 0644,
 630                 .proc_handler   = proc_dointvec,
 631         },
 632 #endif
 633 #ifdef __hppa__
 634         {
 635                 .procname       = "soft-power",
 636                 .data           = &pwrsw_enabled,
 637                 .maxlen         = sizeof (int),
 638                 .mode           = 0644,
 639                 .proc_handler   = proc_dointvec,
 640         },
 641 #endif
 642 #ifdef CONFIG_SYSCTL_ARCH_UNALIGN_ALLOW
 643         {
 644                 .procname       = "unaligned-trap",
 645                 .data           = &unaligned_enabled,
 646                 .maxlen         = sizeof (int),
 647                 .mode           = 0644,
 648                 .proc_handler   = proc_dointvec,
 649         },
 650 #endif
 651         {
 652                 .procname       = "ctrl-alt-del",
 653                 .data           = &C_A_D,
 654                 .maxlen         = sizeof(int),
 655                 .mode           = 0644,
 656                 .proc_handler   = proc_dointvec,
 657         },
 658 #ifdef CONFIG_FUNCTION_TRACER
 659         {
 660                 .procname       = "ftrace_enabled",
 661                 .data           = &ftrace_enabled,
 662                 .maxlen         = sizeof(int),
 663                 .mode           = 0644,
 664                 .proc_handler   = ftrace_enable_sysctl,
 665         },
 666 #endif
 667 #ifdef CONFIG_STACK_TRACER
 668         {
 669                 .procname       = "stack_tracer_enabled",
 670                 .data           = &stack_tracer_enabled,
 671                 .maxlen         = sizeof(int),
 672                 .mode           = 0644,
 673                 .proc_handler   = stack_trace_sysctl,
 674         },
 675 #endif
 676 #ifdef CONFIG_TRACING
 677         {
 678                 .procname       = "ftrace_dump_on_oops",
 679                 .data           = &ftrace_dump_on_oops,
 680                 .maxlen         = sizeof(int),
 681                 .mode           = 0644,
 682                 .proc_handler   = proc_dointvec,
 683         },
 684         {
 685                 .procname       = "traceoff_on_warning",
 686                 .data           = &__disable_trace_on_warning,
 687                 .maxlen         = sizeof(__disable_trace_on_warning),
 688                 .mode           = 0644,
 689                 .proc_handler   = proc_dointvec,
 690         },
 691         {
 692                 .procname       = "tracepoint_printk",
 693                 .data           = &tracepoint_printk,
 694                 .maxlen         = sizeof(tracepoint_printk),
 695                 .mode           = 0644,
 696                 .proc_handler   = proc_dointvec,
 697         },
 698 #endif
 699 #ifdef CONFIG_KEXEC_CORE
 700         {
 701                 .procname       = "kexec_load_disabled",
 702                 .data           = &kexec_load_disabled,
 703                 .maxlen         = sizeof(int),
 704                 .mode           = 0644,
 705                 /* only handle a transition from default "0" to "1" */
 706                 .proc_handler   = proc_dointvec_minmax,
 707                 .extra1         = &one,
 708                 .extra2         = &one,
 709         },
 710 #endif
 711 #ifdef CONFIG_MODULES
 712         {
 713                 .procname       = "modprobe",
 714                 .data           = &modprobe_path,
 715                 .maxlen         = KMOD_PATH_LEN,
 716                 .mode           = 0644,
 717                 .proc_handler   = proc_dostring,
 718         },
 719         {
 720                 .procname       = "modules_disabled",
 721                 .data           = &modules_disabled,
 722                 .maxlen         = sizeof(int),
 723                 .mode           = 0644,
 724                 /* only handle a transition from default "0" to "1" */
 725                 .proc_handler   = proc_dointvec_minmax,
 726                 .extra1         = &one,
 727                 .extra2         = &one,
 728         },
 729 #endif
 730 #ifdef CONFIG_UEVENT_HELPER
 731         {
 732                 .procname       = "hotplug",
 733                 .data           = &uevent_helper,
 734                 .maxlen         = UEVENT_HELPER_PATH_LEN,
 735                 .mode           = 0644,
 736                 .proc_handler   = proc_dostring,
 737         },
 738 #endif
 739 #ifdef CONFIG_CHR_DEV_SG
 740         {
 741                 .procname       = "sg-big-buff",
 742                 .data           = &sg_big_buff,
 743                 .maxlen         = sizeof (int),
 744                 .mode           = 0444,
 745                 .proc_handler   = proc_dointvec,
 746         },
 747 #endif
 748 #ifdef CONFIG_BSD_PROCESS_ACCT
 749         {
 750                 .procname       = "acct",
 751                 .data           = &acct_parm,
 752                 .maxlen         = 3*sizeof(int),
 753                 .mode           = 0644,
 754                 .proc_handler   = proc_dointvec,
 755         },
 756 #endif
 757 #ifdef CONFIG_MAGIC_SYSRQ
 758         {
 759                 .procname       = "sysrq",
 760                 .data           = &__sysrq_enabled,
 761                 .maxlen         = sizeof (int),
 762                 .mode           = 0644,
 763                 .proc_handler   = sysrq_sysctl_handler,
 764         },
 765 #endif
 766 #ifdef CONFIG_PROC_SYSCTL
 767         {
 768                 .procname       = "cad_pid",
 769                 .data           = NULL,
 770                 .maxlen         = sizeof (int),
 771                 .mode           = 0600,
 772                 .proc_handler   = proc_do_cad_pid,
 773         },
 774 #endif
 775         {
 776                 .procname       = "threads-max",
 777                 .data           = NULL,
 778                 .maxlen         = sizeof(int),
 779                 .mode           = 0644,
 780                 .proc_handler   = sysctl_max_threads,
 781         },
 782         {
 783                 .procname       = "random",
 784                 .mode           = 0555,
 785                 .child          = random_table,
 786         },
 787         {
 788                 .procname       = "usermodehelper",
 789                 .mode           = 0555,
 790                 .child          = usermodehelper_table,
 791         },
 792         {
 793                 .procname       = "overflowuid",
 794                 .data           = &overflowuid,
 795                 .maxlen         = sizeof(int),
 796                 .mode           = 0644,
 797                 .proc_handler   = proc_dointvec_minmax,
 798                 .extra1         = &minolduid,
 799                 .extra2         = &maxolduid,
 800         },
 801         {
 802                 .procname       = "overflowgid",
 803                 .data           = &overflowgid,
 804                 .maxlen         = sizeof(int),
 805                 .mode           = 0644,
 806                 .proc_handler   = proc_dointvec_minmax,
 807                 .extra1         = &minolduid,
 808                 .extra2         = &maxolduid,
 809         },
 810 #ifdef CONFIG_S390
 811 #ifdef CONFIG_MATHEMU
 812         {
 813                 .procname       = "ieee_emulation_warnings",
 814                 .data           = &sysctl_ieee_emulation_warnings,
 815                 .maxlen         = sizeof(int),
 816                 .mode           = 0644,
 817                 .proc_handler   = proc_dointvec,
 818         },
 819 #endif
 820         {
 821                 .procname       = "userprocess_debug",
 822                 .data           = &show_unhandled_signals,
 823                 .maxlen         = sizeof(int),
 824                 .mode           = 0644,
 825                 .proc_handler   = proc_dointvec,
 826         },
 827 #endif
 828         {
 829                 .procname       = "pid_max",
 830                 .data           = &pid_max,
 831                 .maxlen         = sizeof (int),
 832                 .mode           = 0644,
 833                 .proc_handler   = proc_dointvec_minmax,
 834                 .extra1         = &pid_max_min,
 835                 .extra2         = &pid_max_max,
 836         },
 837         {
 838                 .procname       = "panic_on_oops",
 839                 .data           = &panic_on_oops,
 840                 .maxlen         = sizeof(int),
 841                 .mode           = 0644,
 842                 .proc_handler   = proc_dointvec,
 843         },
 844 #if defined CONFIG_PRINTK
 845         {
 846                 .procname       = "printk",
 847                 .data           = &console_loglevel,
 848                 .maxlen         = 4*sizeof(int),
 849                 .mode           = 0644,
 850                 .proc_handler   = proc_dointvec,
 851         },
 852         {
 853                 .procname       = "printk_ratelimit",
 854                 .data           = &printk_ratelimit_state.interval,
 855                 .maxlen         = sizeof(int),
 856                 .mode           = 0644,
 857                 .proc_handler   = proc_dointvec_jiffies,
 858         },
 859         {
 860                 .procname       = "printk_ratelimit_burst",
 861                 .data           = &printk_ratelimit_state.burst,
 862                 .maxlen         = sizeof(int),
 863                 .mode           = 0644,
 864                 .proc_handler   = proc_dointvec,
 865         },
 866         {
 867                 .procname       = "printk_delay",
 868                 .data           = &printk_delay_msec,
 869                 .maxlen         = sizeof(int),
 870                 .mode           = 0644,
 871                 .proc_handler   = proc_dointvec_minmax,
 872                 .extra1         = &zero,
 873                 .extra2         = &ten_thousand,
 874         },
 875         {
 876                 .procname       = "dmesg_restrict",
 877                 .data           = &dmesg_restrict,
 878                 .maxlen         = sizeof(int),
 879                 .mode           = 0644,
 880                 .proc_handler   = proc_dointvec_minmax_sysadmin,
 881                 .extra1         = &zero,
 882                 .extra2         = &one,
 883         },
 884         {
 885                 .procname       = "kptr_restrict",
 886                 .data           = &kptr_restrict,
 887                 .maxlen         = sizeof(int),
 888                 .mode           = 0644,
 889                 .proc_handler   = proc_dointvec_minmax_sysadmin,
 890                 .extra1         = &zero,
 891                 .extra2         = &two,
 892         },
 893 #endif
 894         {
 895                 .procname       = "ngroups_max",
 896                 .data           = &ngroups_max,
 897                 .maxlen         = sizeof (int),
 898                 .mode           = 0444,
 899                 .proc_handler   = proc_dointvec,
 900         },
 901         {
 902                 .procname       = "cap_last_cap",
 903                 .data           = (void *)&cap_last_cap,
 904                 .maxlen         = sizeof(int),
 905                 .mode           = 0444,
 906                 .proc_handler   = proc_dointvec,
 907         },
 908 #if defined(CONFIG_LOCKUP_DETECTOR)
 909         {
 910                 .procname       = "watchdog",
 911                 .data           = &watchdog_user_enabled,
 912                 .maxlen         = sizeof (int),
 913                 .mode           = 0644,
 914                 .proc_handler   = proc_watchdog,
 915                 .extra1         = &zero,
 916                 .extra2         = &one,
 917         },
 918         {
 919                 .procname       = "watchdog_thresh",
 920                 .data           = &watchdog_thresh,
 921                 .maxlen         = sizeof(int),
 922                 .mode           = 0644,
 923                 .proc_handler   = proc_watchdog_thresh,
 924                 .extra1         = &zero,
 925                 .extra2         = &sixty,
 926         },
 927         {
 928                 .procname       = "nmi_watchdog",
 929                 .data           = &nmi_watchdog_enabled,
 930                 .maxlen         = sizeof (int),
 931                 .mode           = 0644,
 932                 .proc_handler   = proc_nmi_watchdog,
 933                 .extra1         = &zero,
 934 #if defined(CONFIG_HAVE_NMI_WATCHDOG) || defined(CONFIG_HARDLOCKUP_DETECTOR)
 935                 .extra2         = &one,
 936 #else
 937                 .extra2         = &zero,
 938 #endif
 939         },
 940         {
 941                 .procname       = "soft_watchdog",
 942                 .data           = &soft_watchdog_enabled,
 943                 .maxlen         = sizeof (int),
 944                 .mode           = 0644,
 945                 .proc_handler   = proc_soft_watchdog,
 946                 .extra1         = &zero,
 947                 .extra2         = &one,
 948         },
 949         {
 950                 .procname       = "watchdog_cpumask",
 951                 .data           = &watchdog_cpumask_bits,
 952                 .maxlen         = NR_CPUS,
 953                 .mode           = 0644,
 954                 .proc_handler   = proc_watchdog_cpumask,
 955         },
 956         {
 957                 .procname       = "softlockup_panic",
 958                 .data           = &softlockup_panic,
 959                 .maxlen         = sizeof(int),
 960                 .mode           = 0644,
 961                 .proc_handler   = proc_dointvec_minmax,
 962                 .extra1         = &zero,
 963                 .extra2         = &one,
 964         },
 965 #ifdef CONFIG_HARDLOCKUP_DETECTOR
 966         {
 967                 .procname       = "hardlockup_panic",
 968                 .data           = &hardlockup_panic,
 969                 .maxlen         = sizeof(int),
 970                 .mode           = 0644,
 971                 .proc_handler   = proc_dointvec_minmax,
 972                 .extra1         = &zero,
 973                 .extra2         = &one,
 974         },
 975 #endif
 976 #ifdef CONFIG_SMP
 977         {
 978                 .procname       = "softlockup_all_cpu_backtrace",
 979                 .data           = &sysctl_softlockup_all_cpu_backtrace,
 980                 .maxlen         = sizeof(int),
 981                 .mode           = 0644,
 982                 .proc_handler   = proc_dointvec_minmax,
 983                 .extra1         = &zero,
 984                 .extra2         = &one,
 985         },
 986         {
 987                 .procname       = "hardlockup_all_cpu_backtrace",
 988                 .data           = &sysctl_hardlockup_all_cpu_backtrace,
 989                 .maxlen         = sizeof(int),
 990                 .mode           = 0644,
 991                 .proc_handler   = proc_dointvec_minmax,
 992                 .extra1         = &zero,
 993                 .extra2         = &one,
 994         },
 995 #endif /* CONFIG_SMP */
 996 #endif
 997 #if defined(CONFIG_X86_LOCAL_APIC) && defined(CONFIG_X86)
 998         {
 999                 .procname       = "unknown_nmi_panic",
1000                 .data           = &unknown_nmi_panic,
1001                 .maxlen         = sizeof (int),
1002                 .mode           = 0644,
1003                 .proc_handler   = proc_dointvec,
1004         },
1005 #endif
1006 #if defined(CONFIG_X86)
1007         {
1008                 .procname       = "panic_on_unrecovered_nmi",
1009                 .data           = &panic_on_unrecovered_nmi,
1010                 .maxlen         = sizeof(int),
1011                 .mode           = 0644,
1012                 .proc_handler   = proc_dointvec,
1013         },
1014         {
1015                 .procname       = "panic_on_io_nmi",
1016                 .data           = &panic_on_io_nmi,
1017                 .maxlen         = sizeof(int),
1018                 .mode           = 0644,
1019                 .proc_handler   = proc_dointvec,
1020         },
1021 #ifdef CONFIG_DEBUG_STACKOVERFLOW
1022         {
1023                 .procname       = "panic_on_stackoverflow",
1024                 .data           = &sysctl_panic_on_stackoverflow,
1025                 .maxlen         = sizeof(int),
1026                 .mode           = 0644,
1027                 .proc_handler   = proc_dointvec,
1028         },
1029 #endif
1030         {
1031                 .procname       = "bootloader_type",
1032                 .data           = &bootloader_type,
1033                 .maxlen         = sizeof (int),
1034                 .mode           = 0444,
1035                 .proc_handler   = proc_dointvec,
1036         },
1037         {
1038                 .procname       = "bootloader_version",
1039                 .data           = &bootloader_version,
1040                 .maxlen         = sizeof (int),
1041                 .mode           = 0444,
1042                 .proc_handler   = proc_dointvec,
1043         },
1044         {
1045                 .procname       = "kstack_depth_to_print",
1046                 .data           = &kstack_depth_to_print,
1047                 .maxlen         = sizeof(int),
1048                 .mode           = 0644,
1049                 .proc_handler   = proc_dointvec,
1050         },
1051         {
1052                 .procname       = "io_delay_type",
1053                 .data           = &io_delay_type,
1054                 .maxlen         = sizeof(int),
1055                 .mode           = 0644,
1056                 .proc_handler   = proc_dointvec,
1057         },
1058 #endif
1059 #if defined(CONFIG_MMU)
1060         {
1061                 .procname       = "randomize_va_space",
1062                 .data           = &randomize_va_space,
1063                 .maxlen         = sizeof(int),
1064                 .mode           = 0644,
1065                 .proc_handler   = proc_dointvec,
1066         },
1067 #endif
1068 #if defined(CONFIG_S390) && defined(CONFIG_SMP)
1069         {
1070                 .procname       = "spin_retry",
1071                 .data           = &spin_retry,
1072                 .maxlen         = sizeof (int),
1073                 .mode           = 0644,
1074                 .proc_handler   = proc_dointvec,
1075         },
1076 #endif
1077 #if     defined(CONFIG_ACPI_SLEEP) && defined(CONFIG_X86)
1078         {
1079                 .procname       = "acpi_video_flags",
1080                 .data           = &acpi_realmode_flags,
1081                 .maxlen         = sizeof (unsigned long),
1082                 .mode           = 0644,
1083                 .proc_handler   = proc_doulongvec_minmax,
1084         },
1085 #endif
1086 #ifdef CONFIG_SYSCTL_ARCH_UNALIGN_NO_WARN
1087         {
1088                 .procname       = "ignore-unaligned-usertrap",
1089                 .data           = &no_unaligned_warning,
1090                 .maxlen         = sizeof (int),
1091                 .mode           = 0644,
1092                 .proc_handler   = proc_dointvec,
1093         },
1094 #endif
1095 #ifdef CONFIG_IA64
1096         {
1097                 .procname       = "unaligned-dump-stack",
1098                 .data           = &unaligned_dump_stack,
1099                 .maxlen         = sizeof (int),
1100                 .mode           = 0644,
1101                 .proc_handler   = proc_dointvec,
1102         },
1103 #endif
1104 #ifdef CONFIG_DETECT_HUNG_TASK
1105         {
1106                 .procname       = "hung_task_panic",
1107                 .data           = &sysctl_hung_task_panic,
1108                 .maxlen         = sizeof(int),
1109                 .mode           = 0644,
1110                 .proc_handler   = proc_dointvec_minmax,
1111                 .extra1         = &zero,
1112                 .extra2         = &one,
1113         },
1114         {
1115                 .procname       = "hung_task_check_count",
1116                 .data           = &sysctl_hung_task_check_count,
1117                 .maxlen         = sizeof(int),
1118                 .mode           = 0644,
1119                 .proc_handler   = proc_dointvec_minmax,
1120                 .extra1         = &zero,
1121         },
1122         {
1123                 .procname       = "hung_task_timeout_secs",
1124                 .data           = &sysctl_hung_task_timeout_secs,
1125                 .maxlen         = sizeof(unsigned long),
1126                 .mode           = 0644,
1127                 .proc_handler   = proc_dohung_task_timeout_secs,
1128                 .extra2         = &hung_task_timeout_max,
1129         },
1130         {
1131                 .procname       = "hung_task_warnings",
1132                 .data           = &sysctl_hung_task_warnings,
1133                 .maxlen         = sizeof(int),
1134                 .mode           = 0644,
1135                 .proc_handler   = proc_dointvec_minmax,
1136                 .extra1         = &neg_one,
1137         },
1138 #endif
1139 #ifdef CONFIG_COMPAT
1140         {
1141                 .procname       = "compat-log",
1142                 .data           = &compat_log,
1143                 .maxlen         = sizeof (int),
1144                 .mode           = 0644,
1145                 .proc_handler   = proc_dointvec,
1146         },
1147 #endif
1148 #ifdef CONFIG_RT_MUTEXES
1149         {
1150                 .procname       = "max_lock_depth",
1151                 .data           = &max_lock_depth,
1152                 .maxlen         = sizeof(int),
1153                 .mode           = 0644,
1154                 .proc_handler   = proc_dointvec,
1155         },
1156 #endif
1157         {
1158                 .procname       = "poweroff_cmd",
1159                 .data           = &poweroff_cmd,
1160                 .maxlen         = POWEROFF_CMD_PATH_LEN,
1161                 .mode           = 0644,
1162                 .proc_handler   = proc_dostring,
1163         },
1164 #ifdef CONFIG_KEYS
1165         {
1166                 .procname       = "keys",
1167                 .mode           = 0555,
1168                 .child          = key_sysctls,
1169         },
1170 #endif
1171 #ifdef CONFIG_PERF_EVENTS
1172         /*
1173          * User-space scripts rely on the existence of this file
1174          * as a feature check for perf_events being enabled.
1175          *
1176          * So it's an ABI, do not remove!
1177          */
1178         {
1179                 .procname       = "perf_event_paranoid",
1180                 .data           = &sysctl_perf_event_paranoid,
1181                 .maxlen         = sizeof(sysctl_perf_event_paranoid),
1182                 .mode           = 0644,
1183                 .proc_handler   = proc_dointvec,
1184         },
1185         {
1186                 .procname       = "perf_event_mlock_kb",
1187                 .data           = &sysctl_perf_event_mlock,
1188                 .maxlen         = sizeof(sysctl_perf_event_mlock),
1189                 .mode           = 0644,
1190                 .proc_handler   = proc_dointvec,
1191         },
1192         {
1193                 .procname       = "perf_event_max_sample_rate",
1194                 .data           = &sysctl_perf_event_sample_rate,
1195                 .maxlen         = sizeof(sysctl_perf_event_sample_rate),
1196                 .mode           = 0644,
1197                 .proc_handler   = perf_proc_update_handler,
1198                 .extra1         = &one,
1199         },
1200         {
1201                 .procname       = "perf_cpu_time_max_percent",
1202                 .data           = &sysctl_perf_cpu_time_max_percent,
1203                 .maxlen         = sizeof(sysctl_perf_cpu_time_max_percent),
1204                 .mode           = 0644,
1205                 .proc_handler   = perf_cpu_time_max_percent_handler,
1206                 .extra1         = &zero,
1207                 .extra2         = &one_hundred,
1208         },
1209 #endif
1210 #ifdef CONFIG_KMEMCHECK
1211         {
1212                 .procname       = "kmemcheck",
1213                 .data           = &kmemcheck_enabled,
1214                 .maxlen         = sizeof(int),
1215                 .mode           = 0644,
1216                 .proc_handler   = proc_dointvec,
1217         },
1218 #endif
1219         {
1220                 .procname       = "panic_on_warn",
1221                 .data           = &panic_on_warn,
1222                 .maxlen         = sizeof(int),
1223                 .mode           = 0644,
1224                 .proc_handler   = proc_dointvec_minmax,
1225                 .extra1         = &zero,
1226                 .extra2         = &one,
1227         },
1228 #if defined(CONFIG_SMP) && defined(CONFIG_NO_HZ_COMMON)
1229         {
1230                 .procname       = "timer_migration",
1231                 .data           = &sysctl_timer_migration,
1232                 .maxlen         = sizeof(unsigned int),
1233                 .mode           = 0644,
1234                 .proc_handler   = timer_migration_handler,
1235         },
1236 #endif
1237 #ifdef CONFIG_BPF_SYSCALL
1238         {
1239                 .procname       = "unprivileged_bpf_disabled",
1240                 .data           = &sysctl_unprivileged_bpf_disabled,
1241                 .maxlen         = sizeof(sysctl_unprivileged_bpf_disabled),
1242                 .mode           = 0644,
1243                 /* only handle a transition from default "0" to "1" */
1244                 .proc_handler   = proc_dointvec_minmax,
1245                 .extra1         = &one,
1246                 .extra2         = &one,
1247         },
1248 #endif
1249         { }
1250 };
1251
1252 static struct ctl_table vm_table[] = {
1253         {
1254                 .procname       = "overcommit_memory",
1255                 .data           = &sysctl_overcommit_memory,
1256                 .maxlen         = sizeof(sysctl_overcommit_memory),
1257                 .mode           = 0644,
1258                 .proc_handler   = proc_dointvec_minmax,
1259                 .extra1         = &zero,
1260                 .extra2         = &two,
1261         },
1262         {
1263                 .procname       = "panic_on_oom",
1264                 .data           = &sysctl_panic_on_oom,
1265                 .maxlen         = sizeof(sysctl_panic_on_oom),
1266                 .mode           = 0644,
1267                 .proc_handler   = proc_dointvec_minmax,
1268                 .extra1         = &zero,
1269                 .extra2         = &two,
1270         },
1271         {
1272                 .procname       = "oom_kill_allocating_task",
1273                 .data           = &sysctl_oom_kill_allocating_task,
1274                 .maxlen         = sizeof(sysctl_oom_kill_allocating_task),
1275                 .mode           = 0644,
1276                 .proc_handler   = proc_dointvec,
1277         },
1278         {
1279                 .procname       = "oom_dump_tasks",
1280                 .data           = &sysctl_oom_dump_tasks,
1281                 .maxlen         = sizeof(sysctl_oom_dump_tasks),
1282                 .mode           = 0644,
1283                 .proc_handler   = proc_dointvec,
1284         },
1285         {
1286                 .procname       = "overcommit_ratio",
1287                 .data           = &sysctl_overcommit_ratio,
1288                 .maxlen         = sizeof(sysctl_overcommit_ratio),
1289                 .mode           = 0644,
1290                 .proc_handler   = overcommit_ratio_handler,
1291         },
1292         {
1293                 .procname       = "overcommit_kbytes",
1294                 .data           = &sysctl_overcommit_kbytes,
1295                 .maxlen         = sizeof(sysctl_overcommit_kbytes),
1296                 .mode           = 0644,
1297                 .proc_handler   = overcommit_kbytes_handler,
1298         },
1299         {
1300                 .procname       = "page-cluster",
1301                 .data           = &page_cluster,
1302                 .maxlen         = sizeof(int),
1303                 .mode           = 0644,
1304                 .proc_handler   = proc_dointvec_minmax,
1305                 .extra1         = &zero,
1306         },
1307         {
1308                 .procname       = "dirty_background_ratio",
1309                 .data           = &dirty_background_ratio,
1310                 .maxlen         = sizeof(dirty_background_ratio),
1311                 .mode           = 0644,
1312                 .proc_handler   = dirty_background_ratio_handler,
1313                 .extra1         = &zero,
1314                 .extra2         = &one_hundred,
1315         },
1316         {
1317                 .procname       = "dirty_background_bytes",
1318                 .data           = &dirty_background_bytes,
1319                 .maxlen         = sizeof(dirty_background_bytes),
1320                 .mode           = 0644,
1321                 .proc_handler   = dirty_background_bytes_handler,
1322                 .extra1         = &one_ul,
1323         },
1324         {
1325                 .procname       = "dirty_ratio",
1326                 .data           = &vm_dirty_ratio,
1327                 .maxlen         = sizeof(vm_dirty_ratio),
1328                 .mode           = 0644,
1329                 .proc_handler   = dirty_ratio_handler,
1330                 .extra1         = &zero,
1331                 .extra2         = &one_hundred,
1332         },
1333         {
1334                 .procname       = "dirty_bytes",
1335                 .data           = &vm_dirty_bytes,
1336                 .maxlen         = sizeof(vm_dirty_bytes),
1337                 .mode           = 0644,
1338                 .proc_handler   = dirty_bytes_handler,
1339                 .extra1         = &dirty_bytes_min,
1340         },
1341         {
1342                 .procname       = "dirty_writeback_centisecs",
1343                 .data           = &dirty_writeback_interval,
1344                 .maxlen         = sizeof(dirty_writeback_interval),
1345                 .mode           = 0644,
1346                 .proc_handler   = dirty_writeback_centisecs_handler,
1347         },
1348         {
1349                 .procname       = "dirty_expire_centisecs",
1350                 .data           = &dirty_expire_interval,
1351                 .maxlen         = sizeof(dirty_expire_interval),
1352                 .mode           = 0644,
1353                 .proc_handler   = proc_dointvec_minmax,
1354                 .extra1         = &zero,
1355         },
1356         {
1357                 .procname       = "dirtytime_expire_seconds",
1358                 .data           = &dirtytime_expire_interval,
1359                 .maxlen         = sizeof(dirty_expire_interval),
1360                 .mode           = 0644,
1361                 .proc_handler   = dirtytime_interval_handler,
1362                 .extra1         = &zero,
1363         },
1364         {
1365                 .procname       = "nr_pdflush_threads",
1366                 .mode           = 0444 /* read-only */,
1367                 .proc_handler   = pdflush_proc_obsolete,
1368         },
1369         {
1370                 .procname       = "swappiness",
1371                 .data           = &vm_swappiness,
1372                 .maxlen         = sizeof(vm_swappiness),
1373                 .mode           = 0644,
1374                 .proc_handler   = proc_dointvec_minmax,
1375                 .extra1         = &zero,
1376                 .extra2         = &one_hundred,
1377         },
1378 #ifdef CONFIG_HUGETLB_PAGE
1379         {
1380                 .procname       = "nr_hugepages",
1381                 .data           = NULL,
1382                 .maxlen         = sizeof(unsigned long),
1383                 .mode           = 0644,
1384                 .proc_handler   = hugetlb_sysctl_handler,
1385         },
1386 #ifdef CONFIG_NUMA
1387         {
1388                 .procname       = "nr_hugepages_mempolicy",
1389                 .data           = NULL,
1390                 .maxlen         = sizeof(unsigned long),
1391                 .mode           = 0644,
1392                 .proc_handler   = &hugetlb_mempolicy_sysctl_handler,
1393         },
1394 #endif
1395          {
1396                 .procname       = "hugetlb_shm_group",
1397                 .data           = &sysctl_hugetlb_shm_group,
1398                 .maxlen         = sizeof(gid_t),
1399                 .mode           = 0644,
1400                 .proc_handler   = proc_dointvec,
1401          },
1402          {
1403                 .procname       = "hugepages_treat_as_movable",
1404                 .data           = &hugepages_treat_as_movable,
1405                 .maxlen         = sizeof(int),
1406                 .mode           = 0644,
1407                 .proc_handler   = proc_dointvec,
1408         },
1409         {
1410                 .procname       = "nr_overcommit_hugepages",
1411                 .data           = NULL,
1412                 .maxlen         = sizeof(unsigned long),
1413                 .mode           = 0644,
1414                 .proc_handler   = hugetlb_overcommit_handler,
1415         },
1416 #endif
1417         {
1418                 .procname       = "lowmem_reserve_ratio",
1419                 .data           = &sysctl_lowmem_reserve_ratio,
1420                 .maxlen         = sizeof(sysctl_lowmem_reserve_ratio),
1421                 .mode           = 0644,
1422                 .proc_handler   = lowmem_reserve_ratio_sysctl_handler,
1423         },
1424         {
1425                 .procname       = "drop_caches",
1426                 .data           = &sysctl_drop_caches,
1427                 .maxlen         = sizeof(int),
1428                 .mode           = 0644,
1429                 .proc_handler   = drop_caches_sysctl_handler,
1430                 .extra1         = &one,
1431                 .extra2         = &four,
1432         },
1433 #ifdef CONFIG_COMPACTION
1434         {
1435                 .procname       = "compact_memory",
1436                 .data           = &sysctl_compact_memory,
1437                 .maxlen         = sizeof(int),
1438                 .mode           = 0200,
1439                 .proc_handler   = sysctl_compaction_handler,
1440         },
1441         {
1442                 .procname       = "extfrag_threshold",
1443                 .data           = &sysctl_extfrag_threshold,
1444                 .maxlen         = sizeof(int),
1445                 .mode           = 0644,
1446                 .proc_handler   = sysctl_extfrag_handler,
1447                 .extra1         = &min_extfrag_threshold,
1448                 .extra2         = &max_extfrag_threshold,
1449         },
1450         {
1451                 .procname       = "compact_unevictable_allowed",
1452                 .data           = &sysctl_compact_unevictable_allowed,
1453                 .maxlen         = sizeof(int),
1454                 .mode           = 0644,
1455                 .proc_handler   = proc_dointvec,
1456                 .extra1         = &zero,
1457                 .extra2         = &one,
1458         },
1459
1460 #endif /* CONFIG_COMPACTION */
1461         {
1462                 .procname       = "min_free_kbytes",
1463                 .data           = &min_free_kbytes,
1464                 .maxlen         = sizeof(min_free_kbytes),
1465                 .mode           = 0644,
1466                 .proc_handler   = min_free_kbytes_sysctl_handler,
1467                 .extra1         = &zero,
1468         },
1469         {
1470                 .procname       = "extra_free_kbytes",
1471                 .data           = &extra_free_kbytes,
1472                 .maxlen         = sizeof(extra_free_kbytes),
1473                 .mode           = 0644,
1474                 .proc_handler   = min_free_kbytes_sysctl_handler,
1475                 .extra1         = &zero,
1476         },
1477         {
1478                 .procname       = "percpu_pagelist_fraction",
1479                 .data           = &percpu_pagelist_fraction,
1480                 .maxlen         = sizeof(percpu_pagelist_fraction),
1481                 .mode           = 0644,
1482                 .proc_handler   = percpu_pagelist_fraction_sysctl_handler,
1483                 .extra1         = &zero,
1484         },
1485 #ifdef CONFIG_MMU
1486         {
1487                 .procname       = "max_map_count",
1488                 .data           = &sysctl_max_map_count,
1489                 .maxlen         = sizeof(sysctl_max_map_count),
1490                 .mode           = 0644,
1491                 .proc_handler   = proc_dointvec_minmax,
1492                 .extra1         = &zero,
1493         },
1494 #else
1495         {
1496                 .procname       = "nr_trim_pages",
1497                 .data           = &sysctl_nr_trim_pages,
1498                 .maxlen         = sizeof(sysctl_nr_trim_pages),
1499                 .mode           = 0644,
1500                 .proc_handler   = proc_dointvec_minmax,
1501                 .extra1         = &zero,
1502         },
1503 #endif
1504         {
1505                 .procname       = "laptop_mode",
1506                 .data           = &laptop_mode,
1507                 .maxlen         = sizeof(laptop_mode),
1508                 .mode           = 0644,
1509                 .proc_handler   = proc_dointvec_jiffies,
1510         },
1511         {
1512                 .procname       = "block_dump",
1513                 .data           = &block_dump,
1514                 .maxlen         = sizeof(block_dump),
1515                 .mode           = 0644,
1516                 .proc_handler   = proc_dointvec,
1517                 .extra1         = &zero,
1518         },
1519         {
1520                 .procname       = "vfs_cache_pressure",
1521                 .data           = &sysctl_vfs_cache_pressure,
1522                 .maxlen         = sizeof(sysctl_vfs_cache_pressure),
1523                 .mode           = 0644,
1524                 .proc_handler   = proc_dointvec,
1525                 .extra1         = &zero,
1526         },
1527 #ifdef HAVE_ARCH_PICK_MMAP_LAYOUT
1528         {
1529                 .procname       = "legacy_va_layout",
1530                 .data           = &sysctl_legacy_va_layout,
1531                 .maxlen         = sizeof(sysctl_legacy_va_layout),
1532                 .mode           = 0644,
1533                 .proc_handler   = proc_dointvec,
1534                 .extra1         = &zero,
1535         },
1536 #endif
1537 #ifdef CONFIG_NUMA
1538         {
1539                 .procname       = "zone_reclaim_mode",
1540                 .data           = &zone_reclaim_mode,
1541                 .maxlen         = sizeof(zone_reclaim_mode),
1542                 .mode           = 0644,
1543                 .proc_handler   = proc_dointvec,
1544                 .extra1         = &zero,
1545         },
1546         {
1547                 .procname       = "min_unmapped_ratio",
1548                 .data           = &sysctl_min_unmapped_ratio,
1549                 .maxlen         = sizeof(sysctl_min_unmapped_ratio),
1550                 .mode           = 0644,
1551                 .proc_handler   = sysctl_min_unmapped_ratio_sysctl_handler,
1552                 .extra1         = &zero,
1553                 .extra2         = &one_hundred,
1554         },
1555         {
1556                 .procname       = "min_slab_ratio",
1557                 .data           = &sysctl_min_slab_ratio,
1558                 .maxlen         = sizeof(sysctl_min_slab_ratio),
1559                 .mode           = 0644,
1560                 .proc_handler   = sysctl_min_slab_ratio_sysctl_handler,
1561                 .extra1         = &zero,
1562                 .extra2         = &one_hundred,
1563         },
1564 #endif
1565 #ifdef CONFIG_SMP
1566         {
1567                 .procname       = "stat_interval",
1568                 .data           = &sysctl_stat_interval,
1569                 .maxlen         = sizeof(sysctl_stat_interval),
1570                 .mode           = 0644,
1571                 .proc_handler   = proc_dointvec_jiffies,
1572         },
1573 #endif
1574 #ifdef CONFIG_MMU
1575         {
1576                 .procname       = "mmap_min_addr",
1577                 .data           = &dac_mmap_min_addr,
1578                 .maxlen         = sizeof(unsigned long),
1579                 .mode           = 0644,
1580                 .proc_handler   = mmap_min_addr_handler,
1581         },
1582 #endif
1583 #ifdef CONFIG_NUMA
1584         {
1585                 .procname       = "numa_zonelist_order",
1586                 .data           = &numa_zonelist_order,
1587                 .maxlen         = NUMA_ZONELIST_ORDER_LEN,
1588                 .mode           = 0644,
1589                 .proc_handler   = numa_zonelist_order_handler,
1590         },
1591 #endif
1592 #if (defined(CONFIG_X86_32) && !defined(CONFIG_UML))|| \
1593    (defined(CONFIG_SUPERH) && defined(CONFIG_VSYSCALL))
1594         {
1595                 .procname       = "vdso_enabled",
1596 #ifdef CONFIG_X86_32
1597                 .data           = &vdso32_enabled,
1598                 .maxlen         = sizeof(vdso32_enabled),
1599 #else
1600                 .data           = &vdso_enabled,
1601                 .maxlen         = sizeof(vdso_enabled),
1602 #endif
1603                 .mode           = 0644,
1604                 .proc_handler   = proc_dointvec,
1605                 .extra1         = &zero,
1606         },
1607 #endif
1608 #ifdef CONFIG_HIGHMEM
1609         {
1610                 .procname       = "highmem_is_dirtyable",
1611                 .data           = &vm_highmem_is_dirtyable,
1612                 .maxlen         = sizeof(vm_highmem_is_dirtyable),
1613                 .mode           = 0644,
1614                 .proc_handler   = proc_dointvec_minmax,
1615                 .extra1         = &zero,
1616                 .extra2         = &one,
1617         },
1618 #endif
1619 #ifdef CONFIG_MEMORY_FAILURE
1620         {
1621                 .procname       = "memory_failure_early_kill",
1622                 .data           = &sysctl_memory_failure_early_kill,
1623                 .maxlen         = sizeof(sysctl_memory_failure_early_kill),
1624                 .mode           = 0644,
1625                 .proc_handler   = proc_dointvec_minmax,
1626                 .extra1         = &zero,
1627                 .extra2         = &one,
1628         },
1629         {
1630                 .procname       = "memory_failure_recovery",
1631                 .data           = &sysctl_memory_failure_recovery,
1632                 .maxlen         = sizeof(sysctl_memory_failure_recovery),
1633                 .mode           = 0644,
1634                 .proc_handler   = proc_dointvec_minmax,
1635                 .extra1         = &zero,
1636                 .extra2         = &one,
1637         },
1638 #endif
1639         {
1640                 .procname       = "user_reserve_kbytes",
1641                 .data           = &sysctl_user_reserve_kbytes,
1642                 .maxlen         = sizeof(sysctl_user_reserve_kbytes),
1643                 .mode           = 0644,
1644                 .proc_handler   = proc_doulongvec_minmax,
1645         },
1646         {
1647                 .procname       = "admin_reserve_kbytes",
1648                 .data           = &sysctl_admin_reserve_kbytes,
1649                 .maxlen         = sizeof(sysctl_admin_reserve_kbytes),
1650                 .mode           = 0644,
1651                 .proc_handler   = proc_doulongvec_minmax,
1652         },
1653 #ifdef CONFIG_HAVE_ARCH_MMAP_RND_BITS
1654         {
1655                 .procname       = "mmap_rnd_bits",
1656                 .data           = &mmap_rnd_bits,
1657                 .maxlen         = sizeof(mmap_rnd_bits),
1658                 .mode           = 0600,
1659                 .proc_handler   = proc_dointvec_minmax,
1660                 .extra1         = (void *)&mmap_rnd_bits_min,
1661                 .extra2         = (void *)&mmap_rnd_bits_max,
1662         },
1663 #endif
1664 #ifdef CONFIG_HAVE_ARCH_MMAP_RND_COMPAT_BITS
1665         {
1666                 .procname       = "mmap_rnd_compat_bits",
1667                 .data           = &mmap_rnd_compat_bits,
1668                 .maxlen         = sizeof(mmap_rnd_compat_bits),
1669                 .mode           = 0600,
1670                 .proc_handler   = proc_dointvec_minmax,
1671                 .extra1         = (void *)&mmap_rnd_compat_bits_min,
1672                 .extra2         = (void *)&mmap_rnd_compat_bits_max,
1673         },
1674 #endif
1675         { }
1676 };
1677
1678 static struct ctl_table fs_table[] = {
1679         {
1680                 .procname       = "inode-nr",
1681                 .data           = &inodes_stat,
1682                 .maxlen         = 2*sizeof(long),
1683                 .mode           = 0444,
1684                 .proc_handler   = proc_nr_inodes,
1685         },
1686         {
1687                 .procname       = "inode-state",
1688                 .data           = &inodes_stat,
1689                 .maxlen         = 7*sizeof(long),
1690                 .mode           = 0444,
1691                 .proc_handler   = proc_nr_inodes,
1692         },
1693         {
1694                 .procname       = "file-nr",
1695                 .data           = &files_stat,
1696                 .maxlen         = sizeof(files_stat),
1697                 .mode           = 0444,
1698                 .proc_handler   = proc_nr_files,
1699         },
1700         {
1701                 .procname       = "file-max",
1702                 .data           = &files_stat.max_files,
1703                 .maxlen         = sizeof(files_stat.max_files),
1704                 .mode           = 0644,
1705                 .proc_handler   = proc_doulongvec_minmax,
1706         },
1707         {
1708                 .procname       = "nr_open",
1709                 .data           = &sysctl_nr_open,
1710                 .maxlen         = sizeof(int),
1711                 .mode           = 0644,
1712                 .proc_handler   = proc_dointvec_minmax,
1713                 .extra1         = &sysctl_nr_open_min,
1714                 .extra2         = &sysctl_nr_open_max,
1715         },
1716         {
1717                 .procname       = "dentry-state",
1718                 .data           = &dentry_stat,
1719                 .maxlen         = 6*sizeof(long),
1720                 .mode           = 0444,
1721                 .proc_handler   = proc_nr_dentry,
1722         },
1723         {
1724                 .procname       = "overflowuid",
1725                 .data           = &fs_overflowuid,
1726                 .maxlen         = sizeof(int),
1727                 .mode           = 0644,
1728                 .proc_handler   = proc_dointvec_minmax,
1729                 .extra1         = &minolduid,
1730                 .extra2         = &maxolduid,
1731         },
1732         {
1733                 .procname       = "overflowgid",
1734                 .data           = &fs_overflowgid,
1735                 .maxlen         = sizeof(int),
1736                 .mode           = 0644,
1737                 .proc_handler   = proc_dointvec_minmax,
1738                 .extra1         = &minolduid,
1739                 .extra2         = &maxolduid,
1740         },
1741 #ifdef CONFIG_FILE_LOCKING
1742         {
1743                 .procname       = "leases-enable",
1744                 .data           = &leases_enable,
1745                 .maxlen         = sizeof(int),
1746                 .mode           = 0644,
1747                 .proc_handler   = proc_dointvec,
1748         },
1749 #endif
1750 #ifdef CONFIG_DNOTIFY
1751         {
1752                 .procname       = "dir-notify-enable",
1753                 .data           = &dir_notify_enable,
1754                 .maxlen         = sizeof(int),
1755                 .mode           = 0644,
1756                 .proc_handler   = proc_dointvec,
1757         },
1758 #endif
1759 #ifdef CONFIG_MMU
1760 #ifdef CONFIG_FILE_LOCKING
1761         {
1762                 .procname       = "lease-break-time",
1763                 .data           = &lease_break_time,
1764                 .maxlen         = sizeof(int),
1765                 .mode           = 0644,
1766                 .proc_handler   = proc_dointvec,
1767         },
1768 #endif
1769 #ifdef CONFIG_AIO
1770         {
1771                 .procname       = "aio-nr",
1772                 .data           = &aio_nr,
1773                 .maxlen         = sizeof(aio_nr),
1774                 .mode           = 0444,
1775                 .proc_handler   = proc_doulongvec_minmax,
1776         },
1777         {
1778                 .procname       = "aio-max-nr",
1779                 .data           = &aio_max_nr,
1780                 .maxlen         = sizeof(aio_max_nr),
1781                 .mode           = 0644,
1782                 .proc_handler   = proc_doulongvec_minmax,
1783         },
1784 #endif /* CONFIG_AIO */
1785 #ifdef CONFIG_INOTIFY_USER
1786         {
1787                 .procname       = "inotify",
1788                 .mode           = 0555,
1789                 .child          = inotify_table,
1790         },
1791 #endif
1792 #ifdef CONFIG_EPOLL
1793         {
1794                 .procname       = "epoll",
1795                 .mode           = 0555,
1796                 .child          = epoll_table,
1797         },
1798 #endif
1799 #endif
1800         {
1801                 .procname       = "protected_symlinks",
1802                 .data           = &sysctl_protected_symlinks,
1803                 .maxlen         = sizeof(int),
1804                 .mode           = 0600,
1805                 .proc_handler   = proc_dointvec_minmax,
1806                 .extra1         = &zero,
1807                 .extra2         = &one,
1808         },
1809         {
1810                 .procname       = "protected_hardlinks",
1811                 .data           = &sysctl_protected_hardlinks,
1812                 .maxlen         = sizeof(int),
1813                 .mode           = 0600,
1814                 .proc_handler   = proc_dointvec_minmax,
1815                 .extra1         = &zero,
1816                 .extra2         = &one,
1817         },
1818         {
1819                 .procname       = "suid_dumpable",
1820                 .data           = &suid_dumpable,
1821                 .maxlen         = sizeof(int),
1822                 .mode           = 0644,
1823                 .proc_handler   = proc_dointvec_minmax_coredump,
1824                 .extra1         = &zero,
1825                 .extra2         = &two,
1826         },
1827 #if defined(CONFIG_BINFMT_MISC) || defined(CONFIG_BINFMT_MISC_MODULE)
1828         {
1829                 .procname       = "binfmt_misc",
1830                 .mode           = 0555,
1831                 .child          = sysctl_mount_point,
1832         },
1833 #endif
1834         {
1835                 .procname       = "pipe-max-size",
1836                 .data           = &pipe_max_size,
1837                 .maxlen         = sizeof(int),
1838                 .mode           = 0644,
1839                 .proc_handler   = &pipe_proc_fn,
1840                 .extra1         = &pipe_min_size,
1841         },
1842         {
1843                 .procname       = "pipe-user-pages-hard",
1844                 .data           = &pipe_user_pages_hard,
1845                 .maxlen         = sizeof(pipe_user_pages_hard),
1846                 .mode           = 0644,
1847                 .proc_handler   = proc_doulongvec_minmax,
1848         },
1849         {
1850                 .procname       = "pipe-user-pages-soft",
1851                 .data           = &pipe_user_pages_soft,
1852                 .maxlen         = sizeof(pipe_user_pages_soft),
1853                 .mode           = 0644,
1854                 .proc_handler   = proc_doulongvec_minmax,
1855         },
1856         { }
1857 };
1858
1859 static struct ctl_table debug_table[] = {
1860 #ifdef CONFIG_SYSCTL_EXCEPTION_TRACE
1861         {
1862                 .procname       = "exception-trace",
1863                 .data           = &show_unhandled_signals,
1864                 .maxlen         = sizeof(int),
1865                 .mode           = 0644,
1866                 .proc_handler   = proc_dointvec
1867         },
1868 #endif
1869 #if defined(CONFIG_OPTPROBES)
1870         {
1871                 .procname       = "kprobes-optimization",
1872                 .data           = &sysctl_kprobes_optimization,
1873                 .maxlen         = sizeof(int),
1874                 .mode           = 0644,
1875                 .proc_handler   = proc_kprobes_optimization_handler,
1876                 .extra1         = &zero,
1877                 .extra2         = &one,
1878         },
1879 #endif
1880         { }
1881 };
1882
1883 static struct ctl_table dev_table[] = {
1884         { }
1885 };
1886
1887 int __init sysctl_init(void)
1888 {
1889         struct ctl_table_header *hdr;
1890
1891         hdr = register_sysctl_table(sysctl_base_table);
1892         kmemleak_not_leak(hdr);
1893         return 0;
1894 }
1895
1896 #endif /* CONFIG_SYSCTL */
1897
1898 /*
1899  * /proc/sys support
1900  */
1901
1902 #ifdef CONFIG_PROC_SYSCTL
1903
1904 static int _proc_do_string(char *data, int maxlen, int write,
1905                            char __user *buffer,
1906                            size_t *lenp, loff_t *ppos)
1907 {
1908         size_t len;
1909         char __user *p;
1910         char c;
1911
1912         if (!data || !maxlen || !*lenp) {
1913                 *lenp = 0;
1914                 return 0;
1915         }
1916
1917         if (write) {
1918                 if (sysctl_writes_strict == SYSCTL_WRITES_STRICT) {
1919                         /* Only continue writes not past the end of buffer. */
1920                         len = strlen(data);
1921                         if (len > maxlen - 1)
1922                                 len = maxlen - 1;
1923
1924                         if (*ppos > len)
1925                                 return 0;
1926                         len = *ppos;
1927                 } else {
1928                         /* Start writing from beginning of buffer. */
1929                         len = 0;
1930                 }
1931
1932                 *ppos += *lenp;
1933                 p = buffer;
1934                 while ((p - buffer) < *lenp && len < maxlen - 1) {
1935                         if (get_user(c, p++))
1936                                 return -EFAULT;
1937                         if (c == 0 || c == '\n')
1938                                 break;
1939                         data[len++] = c;
1940                 }
1941                 data[len] = 0;
1942         } else {
1943                 len = strlen(data);
1944                 if (len > maxlen)
1945                         len = maxlen;
1946
1947                 if (*ppos > len) {
1948                         *lenp = 0;
1949                         return 0;
1950                 }
1951
1952                 data += *ppos;
1953                 len  -= *ppos;
1954
1955                 if (len > *lenp)
1956                         len = *lenp;
1957                 if (len)
1958                         if (copy_to_user(buffer, data, len))
1959                                 return -EFAULT;
1960                 if (len < *lenp) {
1961                         if (put_user('\n', buffer + len))
1962                                 return -EFAULT;
1963                         len++;
1964                 }
1965                 *lenp = len;
1966                 *ppos += len;
1967         }
1968         return 0;
1969 }
1970
1971 static void warn_sysctl_write(struct ctl_table *table)
1972 {
1973         pr_warn_once("%s wrote to %s when file position was not 0!\n"
1974                 "This will not be supported in the future. To silence this\n"
1975                 "warning, set kernel.sysctl_writes_strict = -1\n",
1976                 current->comm, table->procname);
1977 }
1978
1979 /**
1980  * proc_dostring - read a string sysctl
1981  * @table: the sysctl table
1982  * @write: %TRUE if this is a write to the sysctl file
1983  * @buffer: the user buffer
1984  * @lenp: the size of the user buffer
1985  * @ppos: file position
1986  *
1987  * Reads/writes a string from/to the user buffer. If the kernel
1988  * buffer provided is not large enough to hold the string, the
1989  * string is truncated. The copied string is %NULL-terminated.
1990  * If the string is being read by the user process, it is copied
1991  * and a newline '\n' is added. It is truncated if the buffer is
1992  * not large enough.
1993  *
1994  * Returns 0 on success.
1995  */
1996 int proc_dostring(struct ctl_table *table, int write,
1997                   void __user *buffer, size_t *lenp, loff_t *ppos)
1998 {
1999         if (write && *ppos && sysctl_writes_strict == SYSCTL_WRITES_WARN)
2000                 warn_sysctl_write(table);
2001
2002         return _proc_do_string((char *)(table->data), table->maxlen, write,
2003                                (char __user *)buffer, lenp, ppos);
2004 }
2005
2006 static size_t proc_skip_spaces(char **buf)
2007 {
2008         size_t ret;
2009         char *tmp = skip_spaces(*buf);
2010         ret = tmp - *buf;
2011         *buf = tmp;
2012         return ret;
2013 }
2014
2015 static void proc_skip_char(char **buf, size_t *size, const char v)
2016 {
2017         while (*size) {
2018                 if (**buf != v)
2019                         break;
2020                 (*size)--;
2021                 (*buf)++;
2022         }
2023 }
2024
2025 #define TMPBUFLEN 22
2026 /**
2027  * proc_get_long - reads an ASCII formatted integer from a user buffer
2028  *
2029  * @buf: a kernel buffer
2030  * @size: size of the kernel buffer
2031  * @val: this is where the number will be stored
2032  * @neg: set to %TRUE if number is negative
2033  * @perm_tr: a vector which contains the allowed trailers
2034  * @perm_tr_len: size of the perm_tr vector
2035  * @tr: pointer to store the trailer character
2036  *
2037  * In case of success %0 is returned and @buf and @size are updated with
2038  * the amount of bytes read. If @tr is non-NULL and a trailing
2039  * character exists (size is non-zero after returning from this
2040  * function), @tr is updated with the trailing character.
2041  */
2042 static int proc_get_long(char **buf, size_t *size,
2043                           unsigned long *val, bool *neg,
2044                           const char *perm_tr, unsigned perm_tr_len, char *tr)
2045 {
2046         int len;
2047         char *p, tmp[TMPBUFLEN];
2048
2049         if (!*size)
2050                 return -EINVAL;
2051
2052         len = *size;
2053         if (len > TMPBUFLEN - 1)
2054                 len = TMPBUFLEN - 1;
2055
2056         memcpy(tmp, *buf, len);
2057
2058         tmp[len] = 0;
2059         p = tmp;
2060         if (*p == '-' && *size > 1) {
2061                 *neg = true;
2062                 p++;
2063         } else
2064                 *neg = false;
2065         if (!isdigit(*p))
2066                 return -EINVAL;
2067
2068         *val = simple_strtoul(p, &p, 0);
2069
2070         len = p - tmp;
2071
2072         /* We don't know if the next char is whitespace thus we may accept
2073          * invalid integers (e.g. 1234...a) or two integers instead of one
2074          * (e.g. 123...1). So lets not allow such large numbers. */
2075         if (len == TMPBUFLEN - 1)
2076                 return -EINVAL;
2077
2078         if (len < *size && perm_tr_len && !memchr(perm_tr, *p, perm_tr_len))
2079                 return -EINVAL;
2080
2081         if (tr && (len < *size))
2082                 *tr = *p;
2083
2084         *buf += len;
2085         *size -= len;
2086
2087         return 0;
2088 }
2089
2090 /**
2091  * proc_put_long - converts an integer to a decimal ASCII formatted string
2092  *
2093  * @buf: the user buffer
2094  * @size: the size of the user buffer
2095  * @val: the integer to be converted
2096  * @neg: sign of the number, %TRUE for negative
2097  *
2098  * In case of success %0 is returned and @buf and @size are updated with
2099  * the amount of bytes written.
2100  */
2101 static int proc_put_long(void __user **buf, size_t *size, unsigned long val,
2102                           bool neg)
2103 {
2104         int len;
2105         char tmp[TMPBUFLEN], *p = tmp;
2106
2107         sprintf(p, "%s%lu", neg ? "-" : "", val);
2108         len = strlen(tmp);
2109         if (len > *size)
2110                 len = *size;
2111         if (copy_to_user(*buf, tmp, len))
2112                 return -EFAULT;
2113         *size -= len;
2114         *buf += len;
2115         return 0;
2116 }
2117 #undef TMPBUFLEN
2118
2119 static int proc_put_char(void __user **buf, size_t *size, char c)
2120 {
2121         if (*size) {
2122                 char __user **buffer = (char __user **)buf;
2123                 if (put_user(c, *buffer))
2124                         return -EFAULT;
2125                 (*size)--, (*buffer)++;
2126                 *buf = *buffer;
2127         }
2128         return 0;
2129 }
2130
2131 static int do_proc_dointvec_conv(bool *negp, unsigned long *lvalp,
2132                                  int *valp,
2133                                  int write, void *data)
2134 {
2135         if (write) {
2136                 if (*negp) {
2137                         if (*lvalp > (unsigned long) INT_MAX + 1)
2138                                 return -EINVAL;
2139                         *valp = -*lvalp;
2140                 } else {
2141                         if (*lvalp > (unsigned long) INT_MAX)
2142                                 return -EINVAL;
2143                         *valp = *lvalp;
2144                 }
2145         } else {
2146                 int val = *valp;
2147                 if (val < 0) {
2148                         *negp = true;
2149                         *lvalp = -(unsigned long)val;
2150                 } else {
2151                         *negp = false;
2152                         *lvalp = (unsigned long)val;
2153                 }
2154         }
2155         return 0;
2156 }
2157
2158 static int do_proc_douintvec_conv(bool *negp, unsigned long *lvalp,
2159                                  int *valp,
2160                                  int write, void *data)
2161 {
2162         if (write) {
2163                 if (*negp)
2164                         return -EINVAL;
2165                 *valp = *lvalp;
2166         } else {
2167                 unsigned int val = *valp;
2168                 *lvalp = (unsigned long)val;
2169         }
2170         return 0;
2171 }
2172
2173 static const char proc_wspace_sep[] = { ' ', '\t', '\n' };
2174
2175 static int __do_proc_dointvec(void *tbl_data, struct ctl_table *table,
2176                   int write, void __user *buffer,
2177                   size_t *lenp, loff_t *ppos,
2178                   int (*conv)(bool *negp, unsigned long *lvalp, int *valp,
2179                               int write, void *data),
2180                   void *data)
2181 {
2182         int *i, vleft, first = 1, err = 0;
2183         unsigned long page = 0;
2184         size_t left;
2185         char *kbuf;
2186
2187         if (!tbl_data || !table->maxlen || !*lenp || (*ppos && !write)) {
2188                 *lenp = 0;
2189                 return 0;
2190         }
2191
2192         i = (int *) tbl_data;
2193         vleft = table->maxlen / sizeof(*i);
2194         left = *lenp;
2195
2196         if (!conv)
2197                 conv = do_proc_dointvec_conv;
2198
2199         if (write) {
2200                 if (*ppos) {
2201                         switch (sysctl_writes_strict) {
2202                         case SYSCTL_WRITES_STRICT:
2203                                 goto out;
2204                         case SYSCTL_WRITES_WARN:
2205                                 warn_sysctl_write(table);
2206                                 break;
2207                         default:
2208                                 break;
2209                         }
2210                 }
2211
2212                 if (left > PAGE_SIZE - 1)
2213                         left = PAGE_SIZE - 1;
2214                 page = __get_free_page(GFP_TEMPORARY);
2215                 kbuf = (char *) page;
2216                 if (!kbuf)
2217                         return -ENOMEM;
2218                 if (copy_from_user(kbuf, buffer, left)) {
2219                         err = -EFAULT;
2220                         goto free;
2221                 }
2222                 kbuf[left] = 0;
2223         }
2224
2225         for (; left && vleft--; i++, first=0) {
2226                 unsigned long lval;
2227                 bool neg;
2228
2229                 if (write) {
2230                         left -= proc_skip_spaces(&kbuf);
2231
2232                         if (!left)
2233                                 break;
2234                         err = proc_get_long(&kbuf, &left, &lval, &neg,
2235                                              proc_wspace_sep,
2236                                              sizeof(proc_wspace_sep), NULL);
2237                         if (err)
2238                                 break;
2239                         if (conv(&neg, &lval, i, 1, data)) {
2240                                 err = -EINVAL;
2241                                 break;
2242                         }
2243                 } else {
2244                         if (conv(&neg, &lval, i, 0, data)) {
2245                                 err = -EINVAL;
2246                                 break;
2247                         }
2248                         if (!first)
2249                                 err = proc_put_char(&buffer, &left, '\t');
2250                         if (err)
2251                                 break;
2252                         err = proc_put_long(&buffer, &left, lval, neg);
2253                         if (err)
2254                                 break;
2255                 }
2256         }
2257
2258         if (!write && !first && left && !err)
2259                 err = proc_put_char(&buffer, &left, '\n');
2260         if (write && !err && left)
2261                 left -= proc_skip_spaces(&kbuf);
2262 free:
2263         if (write) {
2264                 free_page(page);
2265                 if (first)
2266                         return err ? : -EINVAL;
2267         }
2268         *lenp -= left;
2269 out:
2270         *ppos += *lenp;
2271         return err;
2272 }
2273
2274 static int do_proc_dointvec(struct ctl_table *table, int write,
2275                   void __user *buffer, size_t *lenp, loff_t *ppos,
2276                   int (*conv)(bool *negp, unsigned long *lvalp, int *valp,
2277                               int write, void *data),
2278                   void *data)
2279 {
2280         return __do_proc_dointvec(table->data, table, write,
2281                         buffer, lenp, ppos, conv, data);
2282 }
2283
2284 /**
2285  * proc_dointvec - read a vector of integers
2286  * @table: the sysctl table
2287  * @write: %TRUE if this is a write to the sysctl file
2288  * @buffer: the user buffer
2289  * @lenp: the size of the user buffer
2290  * @ppos: file position
2291  *
2292  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2293  * values from/to the user buffer, treated as an ASCII string.
2294  *
2295  * Returns 0 on success.
2296  */
2297 int proc_dointvec(struct ctl_table *table, int write,
2298                      void __user *buffer, size_t *lenp, loff_t *ppos)
2299 {
2300         return do_proc_dointvec(table, write, buffer, lenp, ppos, NULL, NULL);
2301 }
2302
2303 /**
2304  * proc_douintvec - read a vector of unsigned integers
2305  * @table: the sysctl table
2306  * @write: %TRUE if this is a write to the sysctl file
2307  * @buffer: the user buffer
2308  * @lenp: the size of the user buffer
2309  * @ppos: file position
2310  *
2311  * Reads/writes up to table->maxlen/sizeof(unsigned int) unsigned integer
2312  * values from/to the user buffer, treated as an ASCII string.
2313  *
2314  * Returns 0 on success.
2315  */
2316 int proc_douintvec(struct ctl_table *table, int write,
2317                      void __user *buffer, size_t *lenp, loff_t *ppos)
2318 {
2319         return do_proc_dointvec(table, write, buffer, lenp, ppos,
2320                                 do_proc_douintvec_conv, NULL);
2321 }
2322
2323 /*
2324  * Taint values can only be increased
2325  * This means we can safely use a temporary.
2326  */
2327 static int proc_taint(struct ctl_table *table, int write,
2328                                void __user *buffer, size_t *lenp, loff_t *ppos)
2329 {
2330         struct ctl_table t;
2331         unsigned long tmptaint = get_taint();
2332         int err;
2333
2334         if (write && !capable(CAP_SYS_ADMIN))
2335                 return -EPERM;
2336
2337         t = *table;
2338         t.data = &tmptaint;
2339         err = proc_doulongvec_minmax(&t, write, buffer, lenp, ppos);
2340         if (err < 0)
2341                 return err;
2342
2343         if (write) {
2344                 /*
2345                  * Poor man's atomic or. Not worth adding a primitive
2346                  * to everyone's atomic.h for this
2347                  */
2348                 int i;
2349                 for (i = 0; i < BITS_PER_LONG && tmptaint >> i; i++) {
2350                         if ((tmptaint >> i) & 1)
2351                                 add_taint(i, LOCKDEP_STILL_OK);
2352                 }
2353         }
2354
2355         return err;
2356 }
2357
2358 #ifdef CONFIG_PRINTK
2359 static int proc_dointvec_minmax_sysadmin(struct ctl_table *table, int write,
2360                                 void __user *buffer, size_t *lenp, loff_t *ppos)
2361 {
2362         if (write && !capable(CAP_SYS_ADMIN))
2363                 return -EPERM;
2364
2365         return proc_dointvec_minmax(table, write, buffer, lenp, ppos);
2366 }
2367 #endif
2368
2369 struct do_proc_dointvec_minmax_conv_param {
2370         int *min;
2371         int *max;
2372 };
2373
2374 static int do_proc_dointvec_minmax_conv(bool *negp, unsigned long *lvalp,
2375                                         int *valp,
2376                                         int write, void *data)
2377 {
2378         struct do_proc_dointvec_minmax_conv_param *param = data;
2379         if (write) {
2380                 int val = *negp ? -*lvalp : *lvalp;
2381                 if ((param->min && *param->min > val) ||
2382                     (param->max && *param->max < val))
2383                         return -EINVAL;
2384                 *valp = val;
2385         } else {
2386                 int val = *valp;
2387                 if (val < 0) {
2388                         *negp = true;
2389                         *lvalp = -(unsigned long)val;
2390                 } else {
2391                         *negp = false;
2392                         *lvalp = (unsigned long)val;
2393                 }
2394         }
2395         return 0;
2396 }
2397
2398 /**
2399  * proc_dointvec_minmax - read a vector of integers with min/max values
2400  * @table: the sysctl table
2401  * @write: %TRUE if this is a write to the sysctl file
2402  * @buffer: the user buffer
2403  * @lenp: the size of the user buffer
2404  * @ppos: file position
2405  *
2406  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2407  * values from/to the user buffer, treated as an ASCII string.
2408  *
2409  * This routine will ensure the values are within the range specified by
2410  * table->extra1 (min) and table->extra2 (max).
2411  *
2412  * Returns 0 on success.
2413  */
2414 int proc_dointvec_minmax(struct ctl_table *table, int write,
2415                   void __user *buffer, size_t *lenp, loff_t *ppos)
2416 {
2417         struct do_proc_dointvec_minmax_conv_param param = {
2418                 .min = (int *) table->extra1,
2419                 .max = (int *) table->extra2,
2420         };
2421         return do_proc_dointvec(table, write, buffer, lenp, ppos,
2422                                 do_proc_dointvec_minmax_conv, &param);
2423 }
2424
2425 static void validate_coredump_safety(void)
2426 {
2427 #ifdef CONFIG_COREDUMP
2428         if (suid_dumpable == SUID_DUMP_ROOT &&
2429             core_pattern[0] != '/' && core_pattern[0] != '|') {
2430                 printk(KERN_WARNING "Unsafe core_pattern used with "\
2431                         "suid_dumpable=2. Pipe handler or fully qualified "\
2432                         "core dump path required.\n");
2433         }
2434 #endif
2435 }
2436
2437 static int proc_dointvec_minmax_coredump(struct ctl_table *table, int write,
2438                 void __user *buffer, size_t *lenp, loff_t *ppos)
2439 {
2440         int error = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
2441         if (!error)
2442                 validate_coredump_safety();
2443         return error;
2444 }
2445
2446 #ifdef CONFIG_COREDUMP
2447 static int proc_dostring_coredump(struct ctl_table *table, int write,
2448                   void __user *buffer, size_t *lenp, loff_t *ppos)
2449 {
2450         int error = proc_dostring(table, write, buffer, lenp, ppos);
2451         if (!error)
2452                 validate_coredump_safety();
2453         return error;
2454 }
2455 #endif
2456
2457 static int __do_proc_doulongvec_minmax(void *data, struct ctl_table *table, int write,
2458                                      void __user *buffer,
2459                                      size_t *lenp, loff_t *ppos,
2460                                      unsigned long convmul,
2461                                      unsigned long convdiv)
2462 {
2463         unsigned long *i, *min, *max;
2464         int vleft, first = 1, err = 0;
2465         unsigned long page = 0;
2466         size_t left;
2467         char *kbuf;
2468
2469         if (!data || !table->maxlen || !*lenp || (*ppos && !write)) {
2470                 *lenp = 0;
2471                 return 0;
2472         }
2473
2474         i = (unsigned long *) data;
2475         min = (unsigned long *) table->extra1;
2476         max = (unsigned long *) table->extra2;
2477         vleft = table->maxlen / sizeof(unsigned long);
2478         left = *lenp;
2479
2480         if (write) {
2481                 if (*ppos) {
2482                         switch (sysctl_writes_strict) {
2483                         case SYSCTL_WRITES_STRICT:
2484                                 goto out;
2485                         case SYSCTL_WRITES_WARN:
2486                                 warn_sysctl_write(table);
2487                                 break;
2488                         default:
2489                                 break;
2490                         }
2491                 }
2492
2493                 if (left > PAGE_SIZE - 1)
2494                         left = PAGE_SIZE - 1;
2495                 page = __get_free_page(GFP_TEMPORARY);
2496                 kbuf = (char *) page;
2497                 if (!kbuf)
2498                         return -ENOMEM;
2499                 if (copy_from_user(kbuf, buffer, left)) {
2500                         err = -EFAULT;
2501                         goto free;
2502                 }
2503                 kbuf[left] = 0;
2504         }
2505
2506         for (; left && vleft--; i++, first = 0) {
2507                 unsigned long val;
2508
2509                 if (write) {
2510                         bool neg;
2511
2512                         left -= proc_skip_spaces(&kbuf);
2513
2514                         err = proc_get_long(&kbuf, &left, &val, &neg,
2515                                              proc_wspace_sep,
2516                                              sizeof(proc_wspace_sep), NULL);
2517                         if (err)
2518                                 break;
2519                         if (neg)
2520                                 continue;
2521                         val = convmul * val / convdiv;
2522                         if ((min && val < *min) || (max && val > *max))
2523                                 continue;
2524                         *i = val;
2525                 } else {
2526                         val = convdiv * (*i) / convmul;
2527                         if (!first) {
2528                                 err = proc_put_char(&buffer, &left, '\t');
2529                                 if (err)
2530                                         break;
2531                         }
2532                         err = proc_put_long(&buffer, &left, val, false);
2533                         if (err)
2534                                 break;
2535                 }
2536         }
2537
2538         if (!write && !first && left && !err)
2539                 err = proc_put_char(&buffer, &left, '\n');
2540         if (write && !err)
2541                 left -= proc_skip_spaces(&kbuf);
2542 free:
2543         if (write) {
2544                 free_page(page);
2545                 if (first)
2546                         return err ? : -EINVAL;
2547         }
2548         *lenp -= left;
2549 out:
2550         *ppos += *lenp;
2551         return err;
2552 }
2553
2554 static int do_proc_doulongvec_minmax(struct ctl_table *table, int write,
2555                                      void __user *buffer,
2556                                      size_t *lenp, loff_t *ppos,
2557                                      unsigned long convmul,
2558                                      unsigned long convdiv)
2559 {
2560         return __do_proc_doulongvec_minmax(table->data, table, write,
2561                         buffer, lenp, ppos, convmul, convdiv);
2562 }
2563
2564 /**
2565  * proc_doulongvec_minmax - read a vector of long integers with min/max values
2566  * @table: the sysctl table
2567  * @write: %TRUE if this is a write to the sysctl file
2568  * @buffer: the user buffer
2569  * @lenp: the size of the user buffer
2570  * @ppos: file position
2571  *
2572  * Reads/writes up to table->maxlen/sizeof(unsigned long) unsigned long
2573  * values from/to the user buffer, treated as an ASCII string.
2574  *
2575  * This routine will ensure the values are within the range specified by
2576  * table->extra1 (min) and table->extra2 (max).
2577  *
2578  * Returns 0 on success.
2579  */
2580 int proc_doulongvec_minmax(struct ctl_table *table, int write,
2581                            void __user *buffer, size_t *lenp, loff_t *ppos)
2582 {
2583     return do_proc_doulongvec_minmax(table, write, buffer, lenp, ppos, 1l, 1l);
2584 }
2585
2586 /**
2587  * proc_doulongvec_ms_jiffies_minmax - read a vector of millisecond values with min/max values
2588  * @table: the sysctl table
2589  * @write: %TRUE if this is a write to the sysctl file
2590  * @buffer: the user buffer
2591  * @lenp: the size of the user buffer
2592  * @ppos: file position
2593  *
2594  * Reads/writes up to table->maxlen/sizeof(unsigned long) unsigned long
2595  * values from/to the user buffer, treated as an ASCII string. The values
2596  * are treated as milliseconds, and converted to jiffies when they are stored.
2597  *
2598  * This routine will ensure the values are within the range specified by
2599  * table->extra1 (min) and table->extra2 (max).
2600  *
2601  * Returns 0 on success.
2602  */
2603 int proc_doulongvec_ms_jiffies_minmax(struct ctl_table *table, int write,
2604                                       void __user *buffer,
2605                                       size_t *lenp, loff_t *ppos)
2606 {
2607     return do_proc_doulongvec_minmax(table, write, buffer,
2608                                      lenp, ppos, HZ, 1000l);
2609 }
2610
2611
2612 static int do_proc_dointvec_jiffies_conv(bool *negp, unsigned long *lvalp,
2613                                          int *valp,
2614                                          int write, void *data)
2615 {
2616         if (write) {
2617                 if (*lvalp > LONG_MAX / HZ)
2618                         return 1;
2619                 *valp = *negp ? -(*lvalp*HZ) : (*lvalp*HZ);
2620         } else {
2621                 int val = *valp;
2622                 unsigned long lval;
2623                 if (val < 0) {
2624                         *negp = true;
2625                         lval = -(unsigned long)val;
2626                 } else {
2627                         *negp = false;
2628                         lval = (unsigned long)val;
2629                 }
2630                 *lvalp = lval / HZ;
2631         }
2632         return 0;
2633 }
2634
2635 static int do_proc_dointvec_userhz_jiffies_conv(bool *negp, unsigned long *lvalp,
2636                                                 int *valp,
2637                                                 int write, void *data)
2638 {
2639         if (write) {
2640                 if (USER_HZ < HZ && *lvalp > (LONG_MAX / HZ) * USER_HZ)
2641                         return 1;
2642                 *valp = clock_t_to_jiffies(*negp ? -*lvalp : *lvalp);
2643         } else {
2644                 int val = *valp;
2645                 unsigned long lval;
2646                 if (val < 0) {
2647                         *negp = true;
2648                         lval = -(unsigned long)val;
2649                 } else {
2650                         *negp = false;
2651                         lval = (unsigned long)val;
2652                 }
2653                 *lvalp = jiffies_to_clock_t(lval);
2654         }
2655         return 0;
2656 }
2657
2658 static int do_proc_dointvec_ms_jiffies_conv(bool *negp, unsigned long *lvalp,
2659                                             int *valp,
2660                                             int write, void *data)
2661 {
2662         if (write) {
2663                 unsigned long jif = msecs_to_jiffies(*negp ? -*lvalp : *lvalp);
2664
2665                 if (jif > INT_MAX)
2666                         return 1;
2667                 *valp = (int)jif;
2668         } else {
2669                 int val = *valp;
2670                 unsigned long lval;
2671                 if (val < 0) {
2672                         *negp = true;
2673                         lval = -(unsigned long)val;
2674                 } else {
2675                         *negp = false;
2676                         lval = (unsigned long)val;
2677                 }
2678                 *lvalp = jiffies_to_msecs(lval);
2679         }
2680         return 0;
2681 }
2682
2683 /**
2684  * proc_dointvec_jiffies - read a vector of integers as seconds
2685  * @table: the sysctl table
2686  * @write: %TRUE if this is a write to the sysctl file
2687  * @buffer: the user buffer
2688  * @lenp: the size of the user buffer
2689  * @ppos: file position
2690  *
2691  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2692  * values from/to the user buffer, treated as an ASCII string.
2693  * The values read are assumed to be in seconds, and are converted into
2694  * jiffies.
2695  *
2696  * Returns 0 on success.
2697  */
2698 int proc_dointvec_jiffies(struct ctl_table *table, int write,
2699                           void __user *buffer, size_t *lenp, loff_t *ppos)
2700 {
2701     return do_proc_dointvec(table,write,buffer,lenp,ppos,
2702                             do_proc_dointvec_jiffies_conv,NULL);
2703 }
2704
2705 /**
2706  * proc_dointvec_userhz_jiffies - read a vector of integers as 1/USER_HZ seconds
2707  * @table: the sysctl table
2708  * @write: %TRUE if this is a write to the sysctl file
2709  * @buffer: the user buffer
2710  * @lenp: the size of the user buffer
2711  * @ppos: pointer to the file position
2712  *
2713  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2714  * values from/to the user buffer, treated as an ASCII string.
2715  * The values read are assumed to be in 1/USER_HZ seconds, and
2716  * are converted into jiffies.
2717  *
2718  * Returns 0 on success.
2719  */
2720 int proc_dointvec_userhz_jiffies(struct ctl_table *table, int write,
2721                                  void __user *buffer, size_t *lenp, loff_t *ppos)
2722 {
2723     return do_proc_dointvec(table,write,buffer,lenp,ppos,
2724                             do_proc_dointvec_userhz_jiffies_conv,NULL);
2725 }
2726
2727 /**
2728  * proc_dointvec_ms_jiffies - read a vector of integers as 1 milliseconds
2729  * @table: the sysctl table
2730  * @write: %TRUE if this is a write to the sysctl file
2731  * @buffer: the user buffer
2732  * @lenp: the size of the user buffer
2733  * @ppos: file position
2734  * @ppos: the current position in the file
2735  *
2736  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2737  * values from/to the user buffer, treated as an ASCII string.
2738  * The values read are assumed to be in 1/1000 seconds, and
2739  * are converted into jiffies.
2740  *
2741  * Returns 0 on success.
2742  */
2743 int proc_dointvec_ms_jiffies(struct ctl_table *table, int write,
2744                              void __user *buffer, size_t *lenp, loff_t *ppos)
2745 {
2746         return do_proc_dointvec(table, write, buffer, lenp, ppos,
2747                                 do_proc_dointvec_ms_jiffies_conv, NULL);
2748 }
2749
2750 static int proc_do_cad_pid(struct ctl_table *table, int write,
2751                            void __user *buffer, size_t *lenp, loff_t *ppos)
2752 {
2753         struct pid *new_pid;
2754         pid_t tmp;
2755         int r;
2756
2757         tmp = pid_vnr(cad_pid);
2758
2759         r = __do_proc_dointvec(&tmp, table, write, buffer,
2760                                lenp, ppos, NULL, NULL);
2761         if (r || !write)
2762                 return r;
2763
2764         new_pid = find_get_pid(tmp);
2765         if (!new_pid)
2766                 return -ESRCH;
2767
2768         put_pid(xchg(&cad_pid, new_pid));
2769         return 0;
2770 }
2771
2772 /**
2773  * proc_do_large_bitmap - read/write from/to a large bitmap
2774  * @table: the sysctl table
2775  * @write: %TRUE if this is a write to the sysctl file
2776  * @buffer: the user buffer
2777  * @lenp: the size of the user buffer
2778  * @ppos: file position
2779  *
2780  * The bitmap is stored at table->data and the bitmap length (in bits)
2781  * in table->maxlen.
2782  *
2783  * We use a range comma separated format (e.g. 1,3-4,10-10) so that
2784  * large bitmaps may be represented in a compact manner. Writing into
2785  * the file will clear the bitmap then update it with the given input.
2786  *
2787  * Returns 0 on success.
2788  */
2789 int proc_do_large_bitmap(struct ctl_table *table, int write,
2790                          void __user *buffer, size_t *lenp, loff_t *ppos)
2791 {
2792         int err = 0;
2793         bool first = 1;
2794         size_t left = *lenp;
2795         unsigned long bitmap_len = table->maxlen;
2796         unsigned long *bitmap = *(unsigned long **) table->data;
2797         unsigned long *tmp_bitmap = NULL;
2798         char tr_a[] = { '-', ',', '\n' }, tr_b[] = { ',', '\n', 0 }, c;
2799
2800         if (!bitmap || !bitmap_len || !left || (*ppos && !write)) {
2801                 *lenp = 0;
2802                 return 0;
2803         }
2804
2805         if (write) {
2806                 unsigned long page = 0;
2807                 char *kbuf;
2808
2809                 if (left > PAGE_SIZE - 1)
2810                         left = PAGE_SIZE - 1;
2811
2812                 page = __get_free_page(GFP_TEMPORARY);
2813                 kbuf = (char *) page;
2814                 if (!kbuf)
2815                         return -ENOMEM;
2816                 if (copy_from_user(kbuf, buffer, left)) {
2817                         free_page(page);
2818                         return -EFAULT;
2819                 }
2820                 kbuf[left] = 0;
2821
2822                 tmp_bitmap = kzalloc(BITS_TO_LONGS(bitmap_len) * sizeof(unsigned long),
2823                                      GFP_KERNEL);
2824                 if (!tmp_bitmap) {
2825                         free_page(page);
2826                         return -ENOMEM;
2827                 }
2828                 proc_skip_char(&kbuf, &left, '\n');
2829                 while (!err && left) {
2830                         unsigned long val_a, val_b;
2831                         bool neg;
2832
2833                         err = proc_get_long(&kbuf, &left, &val_a, &neg, tr_a,
2834                                              sizeof(tr_a), &c);
2835                         if (err)
2836                                 break;
2837                         if (val_a >= bitmap_len || neg) {
2838                                 err = -EINVAL;
2839                                 break;
2840                         }
2841
2842                         val_b = val_a;
2843                         if (left) {
2844                                 kbuf++;
2845                                 left--;
2846                         }
2847
2848                         if (c == '-') {
2849                                 err = proc_get_long(&kbuf, &left, &val_b,
2850                                                      &neg, tr_b, sizeof(tr_b),
2851                                                      &c);
2852                                 if (err)
2853                                         break;
2854                                 if (val_b >= bitmap_len || neg ||
2855                                     val_a > val_b) {
2856                                         err = -EINVAL;
2857                                         break;
2858                                 }
2859                                 if (left) {
2860                                         kbuf++;
2861                                         left--;
2862                                 }
2863                         }
2864
2865                         bitmap_set(tmp_bitmap, val_a, val_b - val_a + 1);
2866                         first = 0;
2867                         proc_skip_char(&kbuf, &left, '\n');
2868                 }
2869                 free_page(page);
2870         } else {
2871                 unsigned long bit_a, bit_b = 0;
2872
2873                 while (left) {
2874                         bit_a = find_next_bit(bitmap, bitmap_len, bit_b);
2875                         if (bit_a >= bitmap_len)
2876                                 break;
2877                         bit_b = find_next_zero_bit(bitmap, bitmap_len,
2878                                                    bit_a + 1) - 1;
2879
2880                         if (!first) {
2881                                 err = proc_put_char(&buffer, &left, ',');
2882                                 if (err)
2883                                         break;
2884                         }
2885                         err = proc_put_long(&buffer, &left, bit_a, false);
2886                         if (err)
2887                                 break;
2888                         if (bit_a != bit_b) {
2889                                 err = proc_put_char(&buffer, &left, '-');
2890                                 if (err)
2891                                         break;
2892                                 err = proc_put_long(&buffer, &left, bit_b, false);
2893                                 if (err)
2894                                         break;
2895                         }
2896
2897                         first = 0; bit_b++;
2898                 }
2899                 if (!err)
2900                         err = proc_put_char(&buffer, &left, '\n');
2901         }
2902
2903         if (!err) {
2904                 if (write) {
2905                         if (*ppos)
2906                                 bitmap_or(bitmap, bitmap, tmp_bitmap, bitmap_len);
2907                         else
2908                                 bitmap_copy(bitmap, tmp_bitmap, bitmap_len);
2909                 }
2910                 kfree(tmp_bitmap);
2911                 *lenp -= left;
2912                 *ppos += *lenp;
2913                 return 0;
2914         } else {
2915                 kfree(tmp_bitmap);
2916                 return err;
2917         }
2918 }
2919
2920 #else /* CONFIG_PROC_SYSCTL */
2921
2922 int proc_dostring(struct ctl_table *table, int write,
2923                   void __user *buffer, size_t *lenp, loff_t *ppos)
2924 {
2925         return -ENOSYS;
2926 }
2927
2928 int proc_dointvec(struct ctl_table *table, int write,
2929                   void __user *buffer, size_t *lenp, loff_t *ppos)
2930 {
2931         return -ENOSYS;
2932 }
2933
2934 int proc_douintvec(struct ctl_table *table, int write,
2935                   void __user *buffer, size_t *lenp, loff_t *ppos)
2936 {
2937         return -ENOSYS;
2938 }
2939
2940 int proc_dointvec_minmax(struct ctl_table *table, int write,
2941                     void __user *buffer, size_t *lenp, loff_t *ppos)
2942 {
2943         return -ENOSYS;
2944 }
2945
2946 int proc_dointvec_jiffies(struct ctl_table *table, int write,
2947                     void __user *buffer, size_t *lenp, loff_t *ppos)
2948 {
2949         return -ENOSYS;
2950 }
2951
2952 int proc_dointvec_userhz_jiffies(struct ctl_table *table, int write,
2953                     void __user *buffer, size_t *lenp, loff_t *ppos)
2954 {
2955         return -ENOSYS;
2956 }
2957
2958 int proc_dointvec_ms_jiffies(struct ctl_table *table, int write,
2959                              void __user *buffer, size_t *lenp, loff_t *ppos)
2960 {
2961         return -ENOSYS;
2962 }
2963
2964 int proc_doulongvec_minmax(struct ctl_table *table, int write,
2965                     void __user *buffer, size_t *lenp, loff_t *ppos)
2966 {
2967         return -ENOSYS;
2968 }
2969
2970 int proc_doulongvec_ms_jiffies_minmax(struct ctl_table *table, int write,
2971                                       void __user *buffer,
2972                                       size_t *lenp, loff_t *ppos)
2973 {
2974     return -ENOSYS;
2975 }
2976
2977
2978 #endif /* CONFIG_PROC_SYSCTL */
2979
2980 /*
2981  * No sense putting this after each symbol definition, twice,
2982  * exception granted :-)
2983  */
2984 EXPORT_SYMBOL(proc_dointvec);
2985 EXPORT_SYMBOL(proc_douintvec);
2986 EXPORT_SYMBOL(proc_dointvec_jiffies);
2987 EXPORT_SYMBOL(proc_dointvec_minmax);
2988 EXPORT_SYMBOL(proc_dointvec_userhz_jiffies);
2989 EXPORT_SYMBOL(proc_dointvec_ms_jiffies);
2990 EXPORT_SYMBOL(proc_dostring);
2991 EXPORT_SYMBOL(proc_doulongvec_minmax);
2992 EXPORT_SYMBOL(proc_doulongvec_ms_jiffies_minmax);