kernel/sysctl.c

   1 /*
   2  * sysctl.c: General linux system control interface
   3  *
   4  * Begun 24 March 1995, Stephen Tweedie
   5  * Added /proc support, Dec 1995
   6  * Added bdflush entry and intvec min/max checking, 2/23/96, Tom Dyas.
   7  * Added hooks for /proc/sys/net (minor, minor patch), 96/4/1, Mike Shaver.
   8  * Added kernel/java-{interpreter,appletviewer}, 96/5/10, Mike Shaver.
   9  * Dynamic registration fixes, Stephen Tweedie.
  10  * Added kswapd-interval, ctrl-alt-del, printk stuff, 1/8/97, Chris Horn.
  11  * Made sysctl support optional via CONFIG_SYSCTL, 1/10/97, Chris
  12  *  Horn.
  13  * Added proc_doulongvec_ms_jiffies_minmax, 09/08/99, Carlos H. Bauer.
  14  * Added proc_doulongvec_minmax, 09/08/99, Carlos H. Bauer.
  15  * Changed linked lists to use list.h instead of lists.h, 02/24/00, Bill
  16  *  Wendling.
  17  * The list_for_each() macro wasn't appropriate for the sysctl loop.
  18  *  Removed it and replaced it with older style, 03/23/00, Bill Wendling
  19  */
  20
  21 #include <linux/module.h>
  22 #include <linux/aio.h>
  23 #include <linux/mm.h>
  24 #include <linux/swap.h>
  25 #include <linux/slab.h>
  26 #include <linux/sysctl.h>
  27 #include <linux/bitmap.h>
  28 #include <linux/signal.h>
  29 #include <linux/printk.h>
  30 #include <linux/proc_fs.h>
  31 #include <linux/security.h>
  32 #include <linux/ctype.h>
  33 #include <linux/kmemcheck.h>
  34 #include <linux/kmemleak.h>
  35 #include <linux/fs.h>
  36 #include <linux/init.h>
  37 #include <linux/kernel.h>
  38 #include <linux/kobject.h>
  39 #include <linux/net.h>
  40 #include <linux/sysrq.h>
  41 #include <linux/highuid.h>
  42 #include <linux/writeback.h>
  43 #include <linux/ratelimit.h>
  44 #include <linux/compaction.h>
  45 #include <linux/hugetlb.h>
  46 #include <linux/initrd.h>
  47 #include <linux/key.h>
  48 #include <linux/times.h>
  49 #include <linux/limits.h>
  50 #include <linux/dcache.h>
  51 #include <linux/dnotify.h>
  52 #include <linux/syscalls.h>
  53 #include <linux/vmstat.h>
  54 #include <linux/nfs_fs.h>
  55 #include <linux/acpi.h>
  56 #include <linux/reboot.h>
  57 #include <linux/ftrace.h>
  58 #include <linux/perf_event.h>
  59 #include <linux/kprobes.h>
  60 #include <linux/pipe_fs_i.h>
  61 #include <linux/oom.h>
  62 #include <linux/kmod.h>
  63 #include <linux/capability.h>
  64 #include <linux/binfmts.h>
  65 #include <linux/sched/sysctl.h>
  66 #include <linux/kexec.h>
  67 #include <linux/bpf.h>
  68 #include <linux/mount.h>
  69
  70 #include <asm/uaccess.h>
  71 #include <asm/processor.h>
  72
  73 #ifdef CONFIG_X86
  74 #include <asm/nmi.h>
  75 #include <asm/stacktrace.h>
  76 #include <asm/io.h>
  77 #endif
  78 #ifdef CONFIG_SPARC
  79 #include <asm/setup.h>
  80 #endif
  81 #ifdef CONFIG_BSD_PROCESS_ACCT
  82 #include <linux/acct.h>
  83 #endif
  84 #ifdef CONFIG_RT_MUTEXES
  85 #include <linux/rtmutex.h>
  86 #endif
  87 #if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_LOCK_STAT)
  88 #include <linux/lockdep.h>
  89 #endif
  90 #ifdef CONFIG_CHR_DEV_SG
  91 #include <scsi/sg.h>
  92 #endif
  93
  94 #ifdef CONFIG_LOCKUP_DETECTOR
  95 #include <linux/nmi.h>
  96 #endif
  97
  98 #if defined(CONFIG_SYSCTL)
  99
 100 /* External variables not in a header file. */
 101 extern int suid_dumpable;
 102 #ifdef CONFIG_COREDUMP
 103 extern int core_uses_pid;
 104 extern char core_pattern[];
 105 extern unsigned int core_pipe_limit;
 106 #endif
 107 extern int pid_max;
 108 extern int extra_free_kbytes;
 109 extern int pid_max_min, pid_max_max;
 110 extern int percpu_pagelist_fraction;
 111 extern int compat_log;
 112 extern int latencytop_enabled;
 113 extern int sysctl_nr_open_min, sysctl_nr_open_max;
 114 #ifndef CONFIG_MMU
 115 extern int sysctl_nr_trim_pages;
 116 #endif
 117
 118 /* Constants used for minimum and  maximum */
 119 #ifdef CONFIG_LOCKUP_DETECTOR
 120 static int sixty = 60;
 121 #endif
 122
 123 static int __maybe_unused neg_one = -1;
 124
 125 static int zero;
 126 static int __maybe_unused one = 1;
 127 static int __maybe_unused two = 2;
 128 static int __maybe_unused four = 4;
 129 static unsigned long one_ul = 1;
 130 static int one_hundred = 100;
 131 #ifdef CONFIG_PRINTK
 132 static int ten_thousand = 10000;
 133 #endif
 134
 135 /* this is needed for the proc_doulongvec_minmax of vm_dirty_bytes */
 136 static unsigned long dirty_bytes_min = 2 * PAGE_SIZE;
 137
 138 /* this is needed for the proc_dointvec_minmax for [fs_]overflow UID and GID */
 139 static int maxolduid = 65535;
 140 static int minolduid;
 141
 142 static int ngroups_max = NGROUPS_MAX;
 143 static const int cap_last_cap = CAP_LAST_CAP;
 144
 145 /*this is needed for proc_doulongvec_minmax of sysctl_hung_task_timeout_secs */
 146 #ifdef CONFIG_DETECT_HUNG_TASK
 147 static unsigned long hung_task_timeout_max = (LONG_MAX/HZ);
 148 #endif
 149
 150 #ifdef CONFIG_INOTIFY_USER
 151 #include <linux/inotify.h>
 152 #endif
 153 #ifdef CONFIG_SPARC
 154 #endif
 155
 156 #ifdef __hppa__
 157 extern int pwrsw_enabled;
 158 #endif
 159
 160 #ifdef CONFIG_SYSCTL_ARCH_UNALIGN_ALLOW
 161 extern int unaligned_enabled;
 162 #endif
 163
 164 #ifdef CONFIG_IA64
 165 extern int unaligned_dump_stack;
 166 #endif
 167
 168 #ifdef CONFIG_SYSCTL_ARCH_UNALIGN_NO_WARN
 169 extern int no_unaligned_warning;
 170 #endif
 171
 172 #ifdef CONFIG_PROC_SYSCTL
 173
 174 #define SYSCTL_WRITES_LEGACY    -1
 175 #define SYSCTL_WRITES_WARN       0
 176 #define SYSCTL_WRITES_STRICT     1
 177
 178 static int sysctl_writes_strict = SYSCTL_WRITES_STRICT;
 179
 180 static int proc_do_cad_pid(struct ctl_table *table, int write,
 181                   void __user *buffer, size_t *lenp, loff_t *ppos);
 182 static int proc_taint(struct ctl_table *table, int write,
 183                                void __user *buffer, size_t *lenp, loff_t *ppos);
 184 #endif
 185
 186 #ifdef CONFIG_PRINTK
 187 static int proc_dointvec_minmax_sysadmin(struct ctl_table *table, int write,
 188                                 void __user *buffer, size_t *lenp, loff_t *ppos);
 189 #endif
 190
 191 static int proc_dointvec_minmax_coredump(struct ctl_table *table, int write,
 192                 void __user *buffer, size_t *lenp, loff_t *ppos);
 193 #ifdef CONFIG_COREDUMP
 194 static int proc_dostring_coredump(struct ctl_table *table, int write,
 195                 void __user *buffer, size_t *lenp, loff_t *ppos);
 196 #endif
 197
 198 #ifdef CONFIG_MAGIC_SYSRQ
 199 /* Note: sysrq code uses it's own private copy */
 200 static int __sysrq_enabled = CONFIG_MAGIC_SYSRQ_DEFAULT_ENABLE;
 201
 202 static int sysrq_sysctl_handler(struct ctl_table *table, int write,
 203                                 void __user *buffer, size_t *lenp,
 204                                 loff_t *ppos)
 205 {
 206         int error;
 207
 208         error = proc_dointvec(table, write, buffer, lenp, ppos);
 209         if (error)
 210                 return error;
 211
 212         if (write)
 213                 sysrq_toggle_support(__sysrq_enabled);
 214
 215         return 0;
 216 }
 217
 218 #endif
 219
 220 static struct ctl_table kern_table[];
 221 static struct ctl_table vm_table[];
 222 static struct ctl_table fs_table[];
 223 static struct ctl_table debug_table[];
 224 static struct ctl_table dev_table[];
 225 extern struct ctl_table random_table[];
 226 #ifdef CONFIG_EPOLL
 227 extern struct ctl_table epoll_table[];
 228 #endif
 229
 230 #ifdef HAVE_ARCH_PICK_MMAP_LAYOUT
 231 int sysctl_legacy_va_layout;
 232 #endif
 233
 234 /* The default sysctl tables: */
 235
 236 static struct ctl_table sysctl_base_table[] = {
 237         {
 238                 .procname       = "kernel",
 239                 .mode           = 0555,
 240                 .child          = kern_table,
 241         },
 242         {
 243                 .procname       = "vm",
 244                 .mode           = 0555,
 245                 .child          = vm_table,
 246         },
 247         {
 248                 .procname       = "fs",
 249                 .mode           = 0555,
 250                 .child          = fs_table,
 251         },
 252         {
 253                 .procname       = "debug",
 254                 .mode           = 0555,
 255                 .child          = debug_table,
 256         },
 257         {
 258                 .procname       = "dev",
 259                 .mode           = 0555,
 260                 .child          = dev_table,
 261         },
 262         { }
 263 };
 264
 265 #ifdef CONFIG_SCHED_DEBUG
 266 static int min_sched_granularity_ns = 100000;           /* 100 usecs */
 267 static int max_sched_granularity_ns = NSEC_PER_SEC;     /* 1 second */
 268 static int min_wakeup_granularity_ns;                   /* 0 usecs */
 269 static int max_wakeup_granularity_ns = NSEC_PER_SEC;    /* 1 second */
 270 #ifdef CONFIG_SMP
 271 static int min_sched_tunable_scaling = SCHED_TUNABLESCALING_NONE;
 272 static int max_sched_tunable_scaling = SCHED_TUNABLESCALING_END-1;
 273 #endif /* CONFIG_SMP */
 274 #endif /* CONFIG_SCHED_DEBUG */
 275
 276 #ifdef CONFIG_COMPACTION
 277 static int min_extfrag_threshold;
 278 static int max_extfrag_threshold = 1000;
 279 #endif
 280
 281 static struct ctl_table kern_table[] = {
 282         {
 283                 .procname       = "sched_child_runs_first",
 284                 .data           = &sysctl_sched_child_runs_first,
 285                 .maxlen         = sizeof(unsigned int),
 286                 .mode           = 0644,
 287                 .proc_handler   = proc_dointvec,
 288         },
 289 #ifdef CONFIG_SCHED_DEBUG
 290         {
 291                 .procname       = "sched_min_granularity_ns",
 292                 .data           = &sysctl_sched_min_granularity,
 293                 .maxlen         = sizeof(unsigned int),
 294                 .mode           = 0644,
 295                 .proc_handler   = sched_proc_update_handler,
 296                 .extra1         = &min_sched_granularity_ns,
 297                 .extra2         = &max_sched_granularity_ns,
 298         },
 299         {
 300                 .procname       = "sched_latency_ns",
 301                 .data           = &sysctl_sched_latency,
 302                 .maxlen         = sizeof(unsigned int),
 303                 .mode           = 0644,
 304                 .proc_handler   = sched_proc_update_handler,
 305                 .extra1         = &min_sched_granularity_ns,
 306                 .extra2         = &max_sched_granularity_ns,
 307         },
 308         {
 309                 .procname       = "sched_sync_hint_enable",
 310                 .data           = &sysctl_sched_sync_hint_enable,
 311                 .maxlen         = sizeof(unsigned int),
 312                 .mode           = 0644,
 313                 .proc_handler   = proc_dointvec,
 314         },
 315 #ifdef CONFIG_SCHED_WALT
 316         {
 317                 .procname       = "sched_use_walt_cpu_util",
 318                 .data           = &sysctl_sched_use_walt_cpu_util,
 319                 .maxlen         = sizeof(unsigned int),
 320                 .mode           = 0644,
 321                 .proc_handler   = proc_dointvec,
 322         },
 323         {
 324                 .procname       = "sched_use_walt_task_util",
 325                 .data           = &sysctl_sched_use_walt_task_util,
 326                 .maxlen         = sizeof(unsigned int),
 327                 .mode           = 0644,
 328                 .proc_handler   = proc_dointvec,
 329         },
 330         {
 331                 .procname       = "sched_walt_init_task_load_pct",
 332                 .data           = &sysctl_sched_walt_init_task_load_pct,
 333                 .maxlen         = sizeof(unsigned int),
 334                 .mode           = 0644,
 335                 .proc_handler   = proc_dointvec,
 336         },
 337         {
 338                 .procname       = "sched_walt_cpu_high_irqload",
 339                 .data           = &sysctl_sched_walt_cpu_high_irqload,
 340                 .maxlen         = sizeof(unsigned int),
 341                 .mode           = 0644,
 342                 .proc_handler   = proc_dointvec,
 343         },
 344 #endif
 345         {
 346                 .procname       = "sched_cstate_aware",
 347                 .data           = &sysctl_sched_cstate_aware,
 348                 .maxlen         = sizeof(unsigned int),
 349                 .mode           = 0644,
 350                 .proc_handler   = proc_dointvec,
 351         },
 352         {
 353                 .procname       = "sched_wakeup_granularity_ns",
 354                 .data           = &sysctl_sched_wakeup_granularity,
 355                 .maxlen         = sizeof(unsigned int),
 356                 .mode           = 0644,
 357                 .proc_handler   = sched_proc_update_handler,
 358                 .extra1         = &min_wakeup_granularity_ns,
 359                 .extra2         = &max_wakeup_granularity_ns,
 360         },
 361 #ifdef CONFIG_SMP
 362         {
 363                 .procname       = "sched_tunable_scaling",
 364                 .data           = &sysctl_sched_tunable_scaling,
 365                 .maxlen         = sizeof(enum sched_tunable_scaling),
 366                 .mode           = 0644,
 367                 .proc_handler   = sched_proc_update_handler,
 368                 .extra1         = &min_sched_tunable_scaling,
 369                 .extra2         = &max_sched_tunable_scaling,
 370         },
 371         {
 372                 .procname       = "sched_migration_cost_ns",
 373                 .data           = &sysctl_sched_migration_cost,
 374                 .maxlen         = sizeof(unsigned int),
 375                 .mode           = 0644,
 376                 .proc_handler   = proc_dointvec,
 377         },
 378         {
 379                 .procname       = "sched_nr_migrate",
 380                 .data           = &sysctl_sched_nr_migrate,
 381                 .maxlen         = sizeof(unsigned int),
 382                 .mode           = 0644,
 383                 .proc_handler   = proc_dointvec,
 384         },
 385         {
 386                 .procname       = "sched_time_avg_ms",
 387                 .data           = &sysctl_sched_time_avg,
 388                 .maxlen         = sizeof(unsigned int),
 389                 .mode           = 0644,
 390                 .proc_handler   = proc_dointvec,
 391         },
 392         {
 393                 .procname       = "sched_shares_window_ns",
 394                 .data           = &sysctl_sched_shares_window,
 395                 .maxlen         = sizeof(unsigned int),
 396                 .mode           = 0644,
 397                 .proc_handler   = proc_dointvec,
 398         },
 399 #endif /* CONFIG_SMP */
 400 #ifdef CONFIG_NUMA_BALANCING
 401         {
 402                 .procname       = "numa_balancing_scan_delay_ms",
 403                 .data           = &sysctl_numa_balancing_scan_delay,
 404                 .maxlen         = sizeof(unsigned int),
 405                 .mode           = 0644,
 406                 .proc_handler   = proc_dointvec,
 407         },
 408         {
 409                 .procname       = "numa_balancing_scan_period_min_ms",
 410                 .data           = &sysctl_numa_balancing_scan_period_min,
 411                 .maxlen         = sizeof(unsigned int),
 412                 .mode           = 0644,
 413                 .proc_handler   = proc_dointvec,
 414         },
 415         {
 416                 .procname       = "numa_balancing_scan_period_max_ms",
 417                 .data           = &sysctl_numa_balancing_scan_period_max,
 418                 .maxlen         = sizeof(unsigned int),
 419                 .mode           = 0644,
 420                 .proc_handler   = proc_dointvec,
 421         },
 422         {
 423                 .procname       = "numa_balancing_scan_size_mb",
 424                 .data           = &sysctl_numa_balancing_scan_size,
 425                 .maxlen         = sizeof(unsigned int),
 426                 .mode           = 0644,
 427                 .proc_handler   = proc_dointvec_minmax,
 428                 .extra1         = &one,
 429         },
 430         {
 431                 .procname       = "numa_balancing",
 432                 .data           = NULL, /* filled in by handler */
 433                 .maxlen         = sizeof(unsigned int),
 434                 .mode           = 0644,
 435                 .proc_handler   = sysctl_numa_balancing,
 436                 .extra1         = &zero,
 437                 .extra2         = &one,
 438         },
 439 #endif /* CONFIG_NUMA_BALANCING */
 440 #endif /* CONFIG_SCHED_DEBUG */
 441         {
 442                 .procname       = "sched_rt_period_us",
 443                 .data           = &sysctl_sched_rt_period,
 444                 .maxlen         = sizeof(unsigned int),
 445                 .mode           = 0644,
 446                 .proc_handler   = sched_rt_handler,
 447         },
 448         {
 449                 .procname       = "sched_rt_runtime_us",
 450                 .data           = &sysctl_sched_rt_runtime,
 451                 .maxlen         = sizeof(int),
 452                 .mode           = 0644,
 453                 .proc_handler   = sched_rt_handler,
 454         },
 455         {
 456                 .procname       = "sched_rr_timeslice_ms",
 457                 .data           = &sched_rr_timeslice,
 458                 .maxlen         = sizeof(int),
 459                 .mode           = 0644,
 460                 .proc_handler   = sched_rr_handler,
 461         },
 462 #ifdef CONFIG_SCHED_AUTOGROUP
 463         {
 464                 .procname       = "sched_autogroup_enabled",
 465                 .data           = &sysctl_sched_autogroup_enabled,
 466                 .maxlen         = sizeof(unsigned int),
 467                 .mode           = 0644,
 468                 .proc_handler   = proc_dointvec_minmax,
 469                 .extra1         = &zero,
 470                 .extra2         = &one,
 471         },
 472 #endif
 473 #ifdef CONFIG_CFS_BANDWIDTH
 474         {
 475                 .procname       = "sched_cfs_bandwidth_slice_us",
 476                 .data           = &sysctl_sched_cfs_bandwidth_slice,
 477                 .maxlen         = sizeof(unsigned int),
 478                 .mode           = 0644,
 479                 .proc_handler   = proc_dointvec_minmax,
 480                 .extra1         = &one,
 481         },
 482 #endif
 483 #ifdef CONFIG_SCHED_TUNE
 484         {
 485                 .procname       = "sched_cfs_boost",
 486                 .data           = &sysctl_sched_cfs_boost,
 487                 .maxlen         = sizeof(sysctl_sched_cfs_boost),
 488 #ifdef CONFIG_CGROUP_SCHEDTUNE
 489                 .mode           = 0444,
 490 #else
 491                 .mode           = 0644,
 492 #endif
 493                 .proc_handler   = &sysctl_sched_cfs_boost_handler,
 494                 .extra1         = &zero,
 495                 .extra2         = &one_hundred,
 496         },
 497 #endif
 498 #ifdef CONFIG_PROVE_LOCKING
 499         {
 500                 .procname       = "prove_locking",
 501                 .data           = &prove_locking,
 502                 .maxlen         = sizeof(int),
 503                 .mode           = 0644,
 504                 .proc_handler   = proc_dointvec,
 505         },
 506 #endif
 507 #ifdef CONFIG_LOCK_STAT
 508         {
 509                 .procname       = "lock_stat",
 510                 .data           = &lock_stat,
 511                 .maxlen         = sizeof(int),
 512                 .mode           = 0644,
 513                 .proc_handler   = proc_dointvec,
 514         },
 515 #endif
 516         {
 517                 .procname       = "panic",
 518                 .data           = &panic_timeout,
 519                 .maxlen         = sizeof(int),
 520                 .mode           = 0644,
 521                 .proc_handler   = proc_dointvec,
 522         },
 523 #ifdef CONFIG_COREDUMP
 524         {
 525                 .procname       = "core_uses_pid",
 526                 .data           = &core_uses_pid,
 527                 .maxlen         = sizeof(int),
 528                 .mode           = 0644,
 529                 .proc_handler   = proc_dointvec,
 530         },
 531         {
 532                 .procname       = "core_pattern",
 533                 .data           = core_pattern,
 534                 .maxlen         = CORENAME_MAX_SIZE,
 535                 .mode           = 0644,
 536                 .proc_handler   = proc_dostring_coredump,
 537         },
 538         {
 539                 .procname       = "core_pipe_limit",
 540                 .data           = &core_pipe_limit,
 541                 .maxlen         = sizeof(unsigned int),
 542                 .mode           = 0644,
 543                 .proc_handler   = proc_dointvec,
 544         },
 545 #endif
 546 #ifdef CONFIG_PROC_SYSCTL
 547         {
 548                 .procname       = "tainted",
 549                 .maxlen         = sizeof(long),
 550                 .mode           = 0644,
 551                 .proc_handler   = proc_taint,
 552         },
 553         {
 554                 .procname       = "sysctl_writes_strict",
 555                 .data           = &sysctl_writes_strict,
 556                 .maxlen         = sizeof(int),
 557                 .mode           = 0644,
 558                 .proc_handler   = proc_dointvec_minmax,
 559                 .extra1         = &neg_one,
 560                 .extra2         = &one,
 561         },
 562 #endif
 563 #ifdef CONFIG_LATENCYTOP
 564         {
 565                 .procname       = "latencytop",
 566                 .data           = &latencytop_enabled,
 567                 .maxlen         = sizeof(int),
 568                 .mode           = 0644,
 569                 .proc_handler   = proc_dointvec,
 570         },
 571 #endif
 572 #ifdef CONFIG_BLK_DEV_INITRD
 573         {
 574                 .procname       = "real-root-dev",
 575                 .data           = &real_root_dev,
 576                 .maxlen         = sizeof(int),
 577                 .mode           = 0644,
 578                 .proc_handler   = proc_dointvec,
 579         },
 580 #endif
 581         {
 582                 .procname       = "print-fatal-signals",
 583                 .data           = &print_fatal_signals,
 584                 .maxlen         = sizeof(int),
 585                 .mode           = 0644,
 586                 .proc_handler   = proc_dointvec,
 587         },
 588 #ifdef CONFIG_SPARC
 589         {
 590                 .procname       = "reboot-cmd",
 591                 .data           = reboot_command,
 592                 .maxlen         = 256,
 593                 .mode           = 0644,
 594                 .proc_handler   = proc_dostring,
 595         },
 596         {
 597                 .procname       = "stop-a",
 598                 .data           = &stop_a_enabled,
 599                 .maxlen         = sizeof (int),
 600                 .mode           = 0644,
 601                 .proc_handler   = proc_dointvec,
 602         },
 603         {
 604                 .procname       = "scons-poweroff",
 605                 .data           = &scons_pwroff,
 606                 .maxlen         = sizeof (int),
 607                 .mode           = 0644,
 608                 .proc_handler   = proc_dointvec,
 609         },
 610 #endif
 611 #ifdef CONFIG_SPARC64
 612         {
 613                 .procname       = "tsb-ratio",
 614                 .data           = &sysctl_tsb_ratio,
 615                 .maxlen         = sizeof (int),
 616                 .mode           = 0644,
 617                 .proc_handler   = proc_dointvec,
 618         },
 619 #endif
 620 #ifdef __hppa__
 621         {
 622                 .procname       = "soft-power",
 623                 .data           = &pwrsw_enabled,
 624                 .maxlen         = sizeof (int),
 625                 .mode           = 0644,
 626                 .proc_handler   = proc_dointvec,
 627         },
 628 #endif
 629 #ifdef CONFIG_SYSCTL_ARCH_UNALIGN_ALLOW
 630         {
 631                 .procname       = "unaligned-trap",
 632                 .data           = &unaligned_enabled,
 633                 .maxlen         = sizeof (int),
 634                 .mode           = 0644,
 635                 .proc_handler   = proc_dointvec,
 636         },
 637 #endif
 638         {
 639                 .procname       = "ctrl-alt-del",
 640                 .data           = &C_A_D,
 641                 .maxlen         = sizeof(int),
 642                 .mode           = 0644,
 643                 .proc_handler   = proc_dointvec,
 644         },
 645 #ifdef CONFIG_FUNCTION_TRACER
 646         {
 647                 .procname       = "ftrace_enabled",
 648                 .data           = &ftrace_enabled,
 649                 .maxlen         = sizeof(int),
 650                 .mode           = 0644,
 651                 .proc_handler   = ftrace_enable_sysctl,
 652         },
 653 #endif
 654 #ifdef CONFIG_STACK_TRACER
 655         {
 656                 .procname       = "stack_tracer_enabled",
 657                 .data           = &stack_tracer_enabled,
 658                 .maxlen         = sizeof(int),
 659                 .mode           = 0644,
 660                 .proc_handler   = stack_trace_sysctl,
 661         },
 662 #endif
 663 #ifdef CONFIG_TRACING
 664         {
 665                 .procname       = "ftrace_dump_on_oops",
 666                 .data           = &ftrace_dump_on_oops,
 667                 .maxlen         = sizeof(int),
 668                 .mode           = 0644,
 669                 .proc_handler   = proc_dointvec,
 670         },
 671         {
 672                 .procname       = "traceoff_on_warning",
 673                 .data           = &__disable_trace_on_warning,
 674                 .maxlen         = sizeof(__disable_trace_on_warning),
 675                 .mode           = 0644,
 676                 .proc_handler   = proc_dointvec,
 677         },
 678         {
 679                 .procname       = "tracepoint_printk",
 680                 .data           = &tracepoint_printk,
 681                 .maxlen         = sizeof(tracepoint_printk),
 682                 .mode           = 0644,
 683                 .proc_handler   = proc_dointvec,
 684         },
 685 #endif
 686 #ifdef CONFIG_KEXEC_CORE
 687         {
 688                 .procname       = "kexec_load_disabled",
 689                 .data           = &kexec_load_disabled,
 690                 .maxlen         = sizeof(int),
 691                 .mode           = 0644,
 692                 /* only handle a transition from default "0" to "1" */
 693                 .proc_handler   = proc_dointvec_minmax,
 694                 .extra1         = &one,
 695                 .extra2         = &one,
 696         },
 697 #endif
 698 #ifdef CONFIG_MODULES
 699         {
 700                 .procname       = "modprobe",
 701                 .data           = &modprobe_path,
 702                 .maxlen         = KMOD_PATH_LEN,
 703                 .mode           = 0644,
 704                 .proc_handler   = proc_dostring,
 705         },
 706         {
 707                 .procname       = "modules_disabled",
 708                 .data           = &modules_disabled,
 709                 .maxlen         = sizeof(int),
 710                 .mode           = 0644,
 711                 /* only handle a transition from default "0" to "1" */
 712                 .proc_handler   = proc_dointvec_minmax,
 713                 .extra1         = &one,
 714                 .extra2         = &one,
 715         },
 716 #endif
 717 #ifdef CONFIG_UEVENT_HELPER
 718         {
 719                 .procname       = "hotplug",
 720                 .data           = &uevent_helper,
 721                 .maxlen         = UEVENT_HELPER_PATH_LEN,
 722                 .mode           = 0644,
 723                 .proc_handler   = proc_dostring,
 724         },
 725 #endif
 726 #ifdef CONFIG_CHR_DEV_SG
 727         {
 728                 .procname       = "sg-big-buff",
 729                 .data           = &sg_big_buff,
 730                 .maxlen         = sizeof (int),
 731                 .mode           = 0444,
 732                 .proc_handler   = proc_dointvec,
 733         },
 734 #endif
 735 #ifdef CONFIG_BSD_PROCESS_ACCT
 736         {
 737                 .procname       = "acct",
 738                 .data           = &acct_parm,
 739                 .maxlen         = 3*sizeof(int),
 740                 .mode           = 0644,
 741                 .proc_handler   = proc_dointvec,
 742         },
 743 #endif
 744 #ifdef CONFIG_MAGIC_SYSRQ
 745         {
 746                 .procname       = "sysrq",
 747                 .data           = &__sysrq_enabled,
 748                 .maxlen         = sizeof (int),
 749                 .mode           = 0644,
 750                 .proc_handler   = sysrq_sysctl_handler,
 751         },
 752 #endif
 753 #ifdef CONFIG_PROC_SYSCTL
 754         {
 755                 .procname       = "cad_pid",
 756                 .data           = NULL,
 757                 .maxlen         = sizeof (int),
 758                 .mode           = 0600,
 759                 .proc_handler   = proc_do_cad_pid,
 760         },
 761 #endif
 762         {
 763                 .procname       = "threads-max",
 764                 .data           = NULL,
 765                 .maxlen         = sizeof(int),
 766                 .mode           = 0644,
 767                 .proc_handler   = sysctl_max_threads,
 768         },
 769         {
 770                 .procname       = "random",
 771                 .mode           = 0555,
 772                 .child          = random_table,
 773         },
 774         {
 775                 .procname       = "usermodehelper",
 776                 .mode           = 0555,
 777                 .child          = usermodehelper_table,
 778         },
 779         {
 780                 .procname       = "overflowuid",
 781                 .data           = &overflowuid,
 782                 .maxlen         = sizeof(int),
 783                 .mode           = 0644,
 784                 .proc_handler   = proc_dointvec_minmax,
 785                 .extra1         = &minolduid,
 786                 .extra2         = &maxolduid,
 787         },
 788         {
 789                 .procname       = "overflowgid",
 790                 .data           = &overflowgid,
 791                 .maxlen         = sizeof(int),
 792                 .mode           = 0644,
 793                 .proc_handler   = proc_dointvec_minmax,
 794                 .extra1         = &minolduid,
 795                 .extra2         = &maxolduid,
 796         },
 797 #ifdef CONFIG_S390
 798 #ifdef CONFIG_MATHEMU
 799         {
 800                 .procname       = "ieee_emulation_warnings",
 801                 .data           = &sysctl_ieee_emulation_warnings,
 802                 .maxlen         = sizeof(int),
 803                 .mode           = 0644,
 804                 .proc_handler   = proc_dointvec,
 805         },
 806 #endif
 807         {
 808                 .procname       = "userprocess_debug",
 809                 .data           = &show_unhandled_signals,
 810                 .maxlen         = sizeof(int),
 811                 .mode           = 0644,
 812                 .proc_handler   = proc_dointvec,
 813         },
 814 #endif
 815         {
 816                 .procname       = "pid_max",
 817                 .data           = &pid_max,
 818                 .maxlen         = sizeof (int),
 819                 .mode           = 0644,
 820                 .proc_handler   = proc_dointvec_minmax,
 821                 .extra1         = &pid_max_min,
 822                 .extra2         = &pid_max_max,
 823         },
 824         {
 825                 .procname       = "panic_on_oops",
 826                 .data           = &panic_on_oops,
 827                 .maxlen         = sizeof(int),
 828                 .mode           = 0644,
 829                 .proc_handler   = proc_dointvec,
 830         },
 831 #if defined CONFIG_PRINTK
 832         {
 833                 .procname       = "printk",
 834                 .data           = &console_loglevel,
 835                 .maxlen         = 4*sizeof(int),
 836                 .mode           = 0644,
 837                 .proc_handler   = proc_dointvec,
 838         },
 839         {
 840                 .procname       = "printk_ratelimit",
 841                 .data           = &printk_ratelimit_state.interval,
 842                 .maxlen         = sizeof(int),
 843                 .mode           = 0644,
 844                 .proc_handler   = proc_dointvec_jiffies,
 845         },
 846         {
 847                 .procname       = "printk_ratelimit_burst",
 848                 .data           = &printk_ratelimit_state.burst,
 849                 .maxlen         = sizeof(int),
 850                 .mode           = 0644,
 851                 .proc_handler   = proc_dointvec,
 852         },
 853         {
 854                 .procname       = "printk_delay",
 855                 .data           = &printk_delay_msec,
 856                 .maxlen         = sizeof(int),
 857                 .mode           = 0644,
 858                 .proc_handler   = proc_dointvec_minmax,
 859                 .extra1         = &zero,
 860                 .extra2         = &ten_thousand,
 861         },
 862         {
 863                 .procname       = "dmesg_restrict",
 864                 .data           = &dmesg_restrict,
 865                 .maxlen         = sizeof(int),
 866                 .mode           = 0644,
 867                 .proc_handler   = proc_dointvec_minmax_sysadmin,
 868                 .extra1         = &zero,
 869                 .extra2         = &one,
 870         },
 871         {
 872                 .procname       = "kptr_restrict",
 873                 .data           = &kptr_restrict,
 874                 .maxlen         = sizeof(int),
 875                 .mode           = 0644,
 876                 .proc_handler   = proc_dointvec_minmax_sysadmin,
 877                 .extra1         = &zero,
 878                 .extra2         = &two,
 879         },
 880 #endif
 881         {
 882                 .procname       = "ngroups_max",
 883                 .data           = &ngroups_max,
 884                 .maxlen         = sizeof (int),
 885                 .mode           = 0444,
 886                 .proc_handler   = proc_dointvec,
 887         },
 888         {
 889                 .procname       = "cap_last_cap",
 890                 .data           = (void *)&cap_last_cap,
 891                 .maxlen         = sizeof(int),
 892                 .mode           = 0444,
 893                 .proc_handler   = proc_dointvec,
 894         },
 895 #if defined(CONFIG_LOCKUP_DETECTOR)
 896         {
 897                 .procname       = "watchdog",
 898                 .data           = &watchdog_user_enabled,
 899                 .maxlen         = sizeof (int),
 900                 .mode           = 0644,
 901                 .proc_handler   = proc_watchdog,
 902                 .extra1         = &zero,
 903                 .extra2         = &one,
 904         },
 905         {
 906                 .procname       = "watchdog_thresh",
 907                 .data           = &watchdog_thresh,
 908                 .maxlen         = sizeof(int),
 909                 .mode           = 0644,
 910                 .proc_handler   = proc_watchdog_thresh,
 911                 .extra1         = &zero,
 912                 .extra2         = &sixty,
 913         },
 914         {
 915                 .procname       = "nmi_watchdog",
 916                 .data           = &nmi_watchdog_enabled,
 917                 .maxlen         = sizeof (int),
 918                 .mode           = 0644,
 919                 .proc_handler   = proc_nmi_watchdog,
 920                 .extra1         = &zero,
 921 #if defined(CONFIG_HAVE_NMI_WATCHDOG) || defined(CONFIG_HARDLOCKUP_DETECTOR)
 922                 .extra2         = &one,
 923 #else
 924                 .extra2         = &zero,
 925 #endif
 926         },
 927         {
 928                 .procname       = "soft_watchdog",
 929                 .data           = &soft_watchdog_enabled,
 930                 .maxlen         = sizeof (int),
 931                 .mode           = 0644,
 932                 .proc_handler   = proc_soft_watchdog,
 933                 .extra1         = &zero,
 934                 .extra2         = &one,
 935         },
 936         {
 937                 .procname       = "watchdog_cpumask",
 938                 .data           = &watchdog_cpumask_bits,
 939                 .maxlen         = NR_CPUS,
 940                 .mode           = 0644,
 941                 .proc_handler   = proc_watchdog_cpumask,
 942         },
 943         {
 944                 .procname       = "softlockup_panic",
 945                 .data           = &softlockup_panic,
 946                 .maxlen         = sizeof(int),
 947                 .mode           = 0644,
 948                 .proc_handler   = proc_dointvec_minmax,
 949                 .extra1         = &zero,
 950                 .extra2         = &one,
 951         },
 952 #ifdef CONFIG_HARDLOCKUP_DETECTOR
 953         {
 954                 .procname       = "hardlockup_panic",
 955                 .data           = &hardlockup_panic,
 956                 .maxlen         = sizeof(int),
 957                 .mode           = 0644,
 958                 .proc_handler   = proc_dointvec_minmax,
 959                 .extra1         = &zero,
 960                 .extra2         = &one,
 961         },
 962 #endif
 963 #ifdef CONFIG_SMP
 964         {
 965                 .procname       = "softlockup_all_cpu_backtrace",
 966                 .data           = &sysctl_softlockup_all_cpu_backtrace,
 967                 .maxlen         = sizeof(int),
 968                 .mode           = 0644,
 969                 .proc_handler   = proc_dointvec_minmax,
 970                 .extra1         = &zero,
 971                 .extra2         = &one,
 972         },
 973         {
 974                 .procname       = "hardlockup_all_cpu_backtrace",
 975                 .data           = &sysctl_hardlockup_all_cpu_backtrace,
 976                 .maxlen         = sizeof(int),
 977                 .mode           = 0644,
 978                 .proc_handler   = proc_dointvec_minmax,
 979                 .extra1         = &zero,
 980                 .extra2         = &one,
 981         },
 982 #endif /* CONFIG_SMP */
 983 #endif
 984 #if defined(CONFIG_X86_LOCAL_APIC) && defined(CONFIG_X86)
 985         {
 986                 .procname       = "unknown_nmi_panic",
 987                 .data           = &unknown_nmi_panic,
 988                 .maxlen         = sizeof (int),
 989                 .mode           = 0644,
 990                 .proc_handler   = proc_dointvec,
 991         },
 992 #endif
 993 #if defined(CONFIG_X86)
 994         {
 995                 .procname       = "panic_on_unrecovered_nmi",
 996                 .data           = &panic_on_unrecovered_nmi,
 997                 .maxlen         = sizeof(int),
 998                 .mode           = 0644,
 999                 .proc_handler   = proc_dointvec,
1000         },
1001         {
1002                 .procname       = "panic_on_io_nmi",
1003                 .data           = &panic_on_io_nmi,
1004                 .maxlen         = sizeof(int),
1005                 .mode           = 0644,
1006                 .proc_handler   = proc_dointvec,
1007         },
1008 #ifdef CONFIG_DEBUG_STACKOVERFLOW
1009         {
1010                 .procname       = "panic_on_stackoverflow",
1011                 .data           = &sysctl_panic_on_stackoverflow,
1012                 .maxlen         = sizeof(int),
1013                 .mode           = 0644,
1014                 .proc_handler   = proc_dointvec,
1015         },
1016 #endif
1017         {
1018                 .procname       = "bootloader_type",
1019                 .data           = &bootloader_type,
1020                 .maxlen         = sizeof (int),
1021                 .mode           = 0444,
1022                 .proc_handler   = proc_dointvec,
1023         },
1024         {
1025                 .procname       = "bootloader_version",
1026                 .data           = &bootloader_version,
1027                 .maxlen         = sizeof (int),
1028                 .mode           = 0444,
1029                 .proc_handler   = proc_dointvec,
1030         },
1031         {
1032                 .procname       = "kstack_depth_to_print",
1033                 .data           = &kstack_depth_to_print,
1034                 .maxlen         = sizeof(int),
1035                 .mode           = 0644,
1036                 .proc_handler   = proc_dointvec,
1037         },
1038         {
1039                 .procname       = "io_delay_type",
1040                 .data           = &io_delay_type,
1041                 .maxlen         = sizeof(int),
1042                 .mode           = 0644,
1043                 .proc_handler   = proc_dointvec,
1044         },
1045 #endif
1046 #if defined(CONFIG_MMU)
1047         {
1048                 .procname       = "randomize_va_space",
1049                 .data           = &randomize_va_space,
1050                 .maxlen         = sizeof(int),
1051                 .mode           = 0644,
1052                 .proc_handler   = proc_dointvec,
1053         },
1054 #endif
1055 #if defined(CONFIG_S390) && defined(CONFIG_SMP)
1056         {
1057                 .procname       = "spin_retry",
1058                 .data           = &spin_retry,
1059                 .maxlen         = sizeof (int),
1060                 .mode           = 0644,
1061                 .proc_handler   = proc_dointvec,
1062         },
1063 #endif
1064 #if     defined(CONFIG_ACPI_SLEEP) && defined(CONFIG_X86)
1065         {
1066                 .procname       = "acpi_video_flags",
1067                 .data           = &acpi_realmode_flags,
1068                 .maxlen         = sizeof (unsigned long),
1069                 .mode           = 0644,
1070                 .proc_handler   = proc_doulongvec_minmax,
1071         },
1072 #endif
1073 #ifdef CONFIG_SYSCTL_ARCH_UNALIGN_NO_WARN
1074         {
1075                 .procname       = "ignore-unaligned-usertrap",
1076                 .data           = &no_unaligned_warning,
1077                 .maxlen         = sizeof (int),
1078                 .mode           = 0644,
1079                 .proc_handler   = proc_dointvec,
1080         },
1081 #endif
1082 #ifdef CONFIG_IA64
1083         {
1084                 .procname       = "unaligned-dump-stack",
1085                 .data           = &unaligned_dump_stack,
1086                 .maxlen         = sizeof (int),
1087                 .mode           = 0644,
1088                 .proc_handler   = proc_dointvec,
1089         },
1090 #endif
1091 #ifdef CONFIG_DETECT_HUNG_TASK
1092         {
1093                 .procname       = "hung_task_panic",
1094                 .data           = &sysctl_hung_task_panic,
1095                 .maxlen         = sizeof(int),
1096                 .mode           = 0644,
1097                 .proc_handler   = proc_dointvec_minmax,
1098                 .extra1         = &zero,
1099                 .extra2         = &one,
1100         },
1101         {
1102                 .procname       = "hung_task_check_count",
1103                 .data           = &sysctl_hung_task_check_count,
1104                 .maxlen         = sizeof(int),
1105                 .mode           = 0644,
1106                 .proc_handler   = proc_dointvec_minmax,
1107                 .extra1         = &zero,
1108         },
1109         {
1110                 .procname       = "hung_task_timeout_secs",
1111                 .data           = &sysctl_hung_task_timeout_secs,
1112                 .maxlen         = sizeof(unsigned long),
1113                 .mode           = 0644,
1114                 .proc_handler   = proc_dohung_task_timeout_secs,
1115                 .extra2         = &hung_task_timeout_max,
1116         },
1117         {
1118                 .procname       = "hung_task_warnings",
1119                 .data           = &sysctl_hung_task_warnings,
1120                 .maxlen         = sizeof(int),
1121                 .mode           = 0644,
1122                 .proc_handler   = proc_dointvec_minmax,
1123                 .extra1         = &neg_one,
1124         },
1125 #endif
1126 #ifdef CONFIG_COMPAT
1127         {
1128                 .procname       = "compat-log",
1129                 .data           = &compat_log,
1130                 .maxlen         = sizeof (int),
1131                 .mode           = 0644,
1132                 .proc_handler   = proc_dointvec,
1133         },
1134 #endif
1135 #ifdef CONFIG_RT_MUTEXES
1136         {
1137                 .procname       = "max_lock_depth",
1138                 .data           = &max_lock_depth,
1139                 .maxlen         = sizeof(int),
1140                 .mode           = 0644,
1141                 .proc_handler   = proc_dointvec,
1142         },
1143 #endif
1144         {
1145                 .procname       = "poweroff_cmd",
1146                 .data           = &poweroff_cmd,
1147                 .maxlen         = POWEROFF_CMD_PATH_LEN,
1148                 .mode           = 0644,
1149                 .proc_handler   = proc_dostring,
1150         },
1151 #ifdef CONFIG_KEYS
1152         {
1153                 .procname       = "keys",
1154                 .mode           = 0555,
1155                 .child          = key_sysctls,
1156         },
1157 #endif
1158 #ifdef CONFIG_PERF_EVENTS
1159         /*
1160          * User-space scripts rely on the existence of this file
1161          * as a feature check for perf_events being enabled.
1162          *
1163          * So it's an ABI, do not remove!
1164          */
1165         {
1166                 .procname       = "perf_event_paranoid",
1167                 .data           = &sysctl_perf_event_paranoid,
1168                 .maxlen         = sizeof(sysctl_perf_event_paranoid),
1169                 .mode           = 0644,
1170                 .proc_handler   = proc_dointvec,
1171         },
1172         {
1173                 .procname       = "perf_event_mlock_kb",
1174                 .data           = &sysctl_perf_event_mlock,
1175                 .maxlen         = sizeof(sysctl_perf_event_mlock),
1176                 .mode           = 0644,
1177                 .proc_handler   = proc_dointvec,
1178         },
1179         {
1180                 .procname       = "perf_event_max_sample_rate",
1181                 .data           = &sysctl_perf_event_sample_rate,
1182                 .maxlen         = sizeof(sysctl_perf_event_sample_rate),
1183                 .mode           = 0644,
1184                 .proc_handler   = perf_proc_update_handler,
1185                 .extra1         = &one,
1186         },
1187         {
1188                 .procname       = "perf_cpu_time_max_percent",
1189                 .data           = &sysctl_perf_cpu_time_max_percent,
1190                 .maxlen         = sizeof(sysctl_perf_cpu_time_max_percent),
1191                 .mode           = 0644,
1192                 .proc_handler   = perf_cpu_time_max_percent_handler,
1193                 .extra1         = &zero,
1194                 .extra2         = &one_hundred,
1195         },
1196 #endif
1197 #ifdef CONFIG_KMEMCHECK
1198         {
1199                 .procname       = "kmemcheck",
1200                 .data           = &kmemcheck_enabled,
1201                 .maxlen         = sizeof(int),
1202                 .mode           = 0644,
1203                 .proc_handler   = proc_dointvec,
1204         },
1205 #endif
1206         {
1207                 .procname       = "panic_on_warn",
1208                 .data           = &panic_on_warn,
1209                 .maxlen         = sizeof(int),
1210                 .mode           = 0644,
1211                 .proc_handler   = proc_dointvec_minmax,
1212                 .extra1         = &zero,
1213                 .extra2         = &one,
1214         },
1215 #if defined(CONFIG_SMP) && defined(CONFIG_NO_HZ_COMMON)
1216         {
1217                 .procname       = "timer_migration",
1218                 .data           = &sysctl_timer_migration,
1219                 .maxlen         = sizeof(unsigned int),
1220                 .mode           = 0644,
1221                 .proc_handler   = timer_migration_handler,
1222                 .extra1         = &zero,
1223                 .extra2         = &one,
1224         },
1225 #endif
1226 #ifdef CONFIG_BPF_SYSCALL
1227         {
1228                 .procname       = "unprivileged_bpf_disabled",
1229                 .data           = &sysctl_unprivileged_bpf_disabled,
1230                 .maxlen         = sizeof(sysctl_unprivileged_bpf_disabled),
1231                 .mode           = 0644,
1232                 /* only handle a transition from default "0" to "1" */
1233                 .proc_handler   = proc_dointvec_minmax,
1234                 .extra1         = &one,
1235                 .extra2         = &one,
1236         },
1237 #endif
1238         { }
1239 };
1240
1241 static struct ctl_table vm_table[] = {
1242         {
1243                 .procname       = "overcommit_memory",
1244                 .data           = &sysctl_overcommit_memory,
1245                 .maxlen         = sizeof(sysctl_overcommit_memory),
1246                 .mode           = 0644,
1247                 .proc_handler   = proc_dointvec_minmax,
1248                 .extra1         = &zero,
1249                 .extra2         = &two,
1250         },
1251         {
1252                 .procname       = "panic_on_oom",
1253                 .data           = &sysctl_panic_on_oom,
1254                 .maxlen         = sizeof(sysctl_panic_on_oom),
1255                 .mode           = 0644,
1256                 .proc_handler   = proc_dointvec_minmax,
1257                 .extra1         = &zero,
1258                 .extra2         = &two,
1259         },
1260         {
1261                 .procname       = "oom_kill_allocating_task",
1262                 .data           = &sysctl_oom_kill_allocating_task,
1263                 .maxlen         = sizeof(sysctl_oom_kill_allocating_task),
1264                 .mode           = 0644,
1265                 .proc_handler   = proc_dointvec,
1266         },
1267         {
1268                 .procname       = "oom_dump_tasks",
1269                 .data           = &sysctl_oom_dump_tasks,
1270                 .maxlen         = sizeof(sysctl_oom_dump_tasks),
1271                 .mode           = 0644,
1272                 .proc_handler   = proc_dointvec,
1273         },
1274         {
1275                 .procname       = "overcommit_ratio",
1276                 .data           = &sysctl_overcommit_ratio,
1277                 .maxlen         = sizeof(sysctl_overcommit_ratio),
1278                 .mode           = 0644,
1279                 .proc_handler   = overcommit_ratio_handler,
1280         },
1281         {
1282                 .procname       = "overcommit_kbytes",
1283                 .data           = &sysctl_overcommit_kbytes,
1284                 .maxlen         = sizeof(sysctl_overcommit_kbytes),
1285                 .mode           = 0644,
1286                 .proc_handler   = overcommit_kbytes_handler,
1287         },
1288         {
1289                 .procname       = "page-cluster",
1290                 .data           = &page_cluster,
1291                 .maxlen         = sizeof(int),
1292                 .mode           = 0644,
1293                 .proc_handler   = proc_dointvec_minmax,
1294                 .extra1         = &zero,
1295         },
1296         {
1297                 .procname       = "dirty_background_ratio",
1298                 .data           = &dirty_background_ratio,
1299                 .maxlen         = sizeof(dirty_background_ratio),
1300                 .mode           = 0644,
1301                 .proc_handler   = dirty_background_ratio_handler,
1302                 .extra1         = &zero,
1303                 .extra2         = &one_hundred,
1304         },
1305         {
1306                 .procname       = "dirty_background_bytes",
1307                 .data           = &dirty_background_bytes,
1308                 .maxlen         = sizeof(dirty_background_bytes),
1309                 .mode           = 0644,
1310                 .proc_handler   = dirty_background_bytes_handler,
1311                 .extra1         = &one_ul,
1312         },
1313         {
1314                 .procname       = "dirty_ratio",
1315                 .data           = &vm_dirty_ratio,
1316                 .maxlen         = sizeof(vm_dirty_ratio),
1317                 .mode           = 0644,
1318                 .proc_handler   = dirty_ratio_handler,
1319                 .extra1         = &zero,
1320                 .extra2         = &one_hundred,
1321         },
1322         {
1323                 .procname       = "dirty_bytes",
1324                 .data           = &vm_dirty_bytes,
1325                 .maxlen         = sizeof(vm_dirty_bytes),
1326                 .mode           = 0644,
1327                 .proc_handler   = dirty_bytes_handler,
1328                 .extra1         = &dirty_bytes_min,
1329         },
1330         {
1331                 .procname       = "dirty_writeback_centisecs",
1332                 .data           = &dirty_writeback_interval,
1333                 .maxlen         = sizeof(dirty_writeback_interval),
1334                 .mode           = 0644,
1335                 .proc_handler   = dirty_writeback_centisecs_handler,
1336         },
1337         {
1338                 .procname       = "dirty_expire_centisecs",
1339                 .data           = &dirty_expire_interval,
1340                 .maxlen         = sizeof(dirty_expire_interval),
1341                 .mode           = 0644,
1342                 .proc_handler   = proc_dointvec_minmax,
1343                 .extra1         = &zero,
1344         },
1345         {
1346                 .procname       = "dirtytime_expire_seconds",
1347                 .data           = &dirtytime_expire_interval,
1348                 .maxlen         = sizeof(dirty_expire_interval),
1349                 .mode           = 0644,
1350                 .proc_handler   = dirtytime_interval_handler,
1351                 .extra1         = &zero,
1352         },
1353         {
1354                 .procname       = "nr_pdflush_threads",
1355                 .mode           = 0444 /* read-only */,
1356                 .proc_handler   = pdflush_proc_obsolete,
1357         },
1358         {
1359                 .procname       = "swappiness",
1360                 .data           = &vm_swappiness,
1361                 .maxlen         = sizeof(vm_swappiness),
1362                 .mode           = 0644,
1363                 .proc_handler   = proc_dointvec_minmax,
1364                 .extra1         = &zero,
1365                 .extra2         = &one_hundred,
1366         },
1367 #ifdef CONFIG_HUGETLB_PAGE
1368         {
1369                 .procname       = "nr_hugepages",
1370                 .data           = NULL,
1371                 .maxlen         = sizeof(unsigned long),
1372                 .mode           = 0644,
1373                 .proc_handler   = hugetlb_sysctl_handler,
1374         },
1375 #ifdef CONFIG_NUMA
1376         {
1377                 .procname       = "nr_hugepages_mempolicy",
1378                 .data           = NULL,
1379                 .maxlen         = sizeof(unsigned long),
1380                 .mode           = 0644,
1381                 .proc_handler   = &hugetlb_mempolicy_sysctl_handler,
1382         },
1383 #endif
1384          {
1385                 .procname       = "hugetlb_shm_group",
1386                 .data           = &sysctl_hugetlb_shm_group,
1387                 .maxlen         = sizeof(gid_t),
1388                 .mode           = 0644,
1389                 .proc_handler   = proc_dointvec,
1390          },
1391          {
1392                 .procname       = "hugepages_treat_as_movable",
1393                 .data           = &hugepages_treat_as_movable,
1394                 .maxlen         = sizeof(int),
1395                 .mode           = 0644,
1396                 .proc_handler   = proc_dointvec,
1397         },
1398         {
1399                 .procname       = "nr_overcommit_hugepages",
1400                 .data           = NULL,
1401                 .maxlen         = sizeof(unsigned long),
1402                 .mode           = 0644,
1403                 .proc_handler   = hugetlb_overcommit_handler,
1404         },
1405 #endif
1406         {
1407                 .procname       = "lowmem_reserve_ratio",
1408                 .data           = &sysctl_lowmem_reserve_ratio,
1409                 .maxlen         = sizeof(sysctl_lowmem_reserve_ratio),
1410                 .mode           = 0644,
1411                 .proc_handler   = lowmem_reserve_ratio_sysctl_handler,
1412         },
1413         {
1414                 .procname       = "drop_caches",
1415                 .data           = &sysctl_drop_caches,
1416                 .maxlen         = sizeof(int),
1417                 .mode           = 0644,
1418                 .proc_handler   = drop_caches_sysctl_handler,
1419                 .extra1         = &one,
1420                 .extra2         = &four,
1421         },
1422 #ifdef CONFIG_COMPACTION
1423         {
1424                 .procname       = "compact_memory",
1425                 .data           = &sysctl_compact_memory,
1426                 .maxlen         = sizeof(int),
1427                 .mode           = 0200,
1428                 .proc_handler   = sysctl_compaction_handler,
1429         },
1430         {
1431                 .procname       = "extfrag_threshold",
1432                 .data           = &sysctl_extfrag_threshold,
1433                 .maxlen         = sizeof(int),
1434                 .mode           = 0644,
1435                 .proc_handler   = sysctl_extfrag_handler,
1436                 .extra1         = &min_extfrag_threshold,
1437                 .extra2         = &max_extfrag_threshold,
1438         },
1439         {
1440                 .procname       = "compact_unevictable_allowed",
1441                 .data           = &sysctl_compact_unevictable_allowed,
1442                 .maxlen         = sizeof(int),
1443                 .mode           = 0644,
1444                 .proc_handler   = proc_dointvec,
1445                 .extra1         = &zero,
1446                 .extra2         = &one,
1447         },
1448
1449 #endif /* CONFIG_COMPACTION */
1450         {
1451                 .procname       = "min_free_kbytes",
1452                 .data           = &min_free_kbytes,
1453                 .maxlen         = sizeof(min_free_kbytes),
1454                 .mode           = 0644,
1455                 .proc_handler   = min_free_kbytes_sysctl_handler,
1456                 .extra1         = &zero,
1457         },
1458         {
1459                 .procname       = "extra_free_kbytes",
1460                 .data           = &extra_free_kbytes,
1461                 .maxlen         = sizeof(extra_free_kbytes),
1462                 .mode           = 0644,
1463                 .proc_handler   = min_free_kbytes_sysctl_handler,
1464                 .extra1         = &zero,
1465         },
1466         {
1467                 .procname       = "percpu_pagelist_fraction",
1468                 .data           = &percpu_pagelist_fraction,
1469                 .maxlen         = sizeof(percpu_pagelist_fraction),
1470                 .mode           = 0644,
1471                 .proc_handler   = percpu_pagelist_fraction_sysctl_handler,
1472                 .extra1         = &zero,
1473         },
1474 #ifdef CONFIG_MMU
1475         {
1476                 .procname       = "max_map_count",
1477                 .data           = &sysctl_max_map_count,
1478                 .maxlen         = sizeof(sysctl_max_map_count),
1479                 .mode           = 0644,
1480                 .proc_handler   = proc_dointvec_minmax,
1481                 .extra1         = &zero,
1482         },
1483 #else
1484         {
1485                 .procname       = "nr_trim_pages",
1486                 .data           = &sysctl_nr_trim_pages,
1487                 .maxlen         = sizeof(sysctl_nr_trim_pages),
1488                 .mode           = 0644,
1489                 .proc_handler   = proc_dointvec_minmax,
1490                 .extra1         = &zero,
1491         },
1492 #endif
1493         {
1494                 .procname       = "laptop_mode",
1495                 .data           = &laptop_mode,
1496                 .maxlen         = sizeof(laptop_mode),
1497                 .mode           = 0644,
1498                 .proc_handler   = proc_dointvec_jiffies,
1499         },
1500         {
1501                 .procname       = "block_dump",
1502                 .data           = &block_dump,
1503                 .maxlen         = sizeof(block_dump),
1504                 .mode           = 0644,
1505                 .proc_handler   = proc_dointvec,
1506                 .extra1         = &zero,
1507         },
1508         {
1509                 .procname       = "vfs_cache_pressure",
1510                 .data           = &sysctl_vfs_cache_pressure,
1511                 .maxlen         = sizeof(sysctl_vfs_cache_pressure),
1512                 .mode           = 0644,
1513                 .proc_handler   = proc_dointvec,
1514                 .extra1         = &zero,
1515         },
1516 #ifdef HAVE_ARCH_PICK_MMAP_LAYOUT
1517         {
1518                 .procname       = "legacy_va_layout",
1519                 .data           = &sysctl_legacy_va_layout,
1520                 .maxlen         = sizeof(sysctl_legacy_va_layout),
1521                 .mode           = 0644,
1522                 .proc_handler   = proc_dointvec,
1523                 .extra1         = &zero,
1524         },
1525 #endif
1526 #ifdef CONFIG_NUMA
1527         {
1528                 .procname       = "zone_reclaim_mode",
1529                 .data           = &zone_reclaim_mode,
1530                 .maxlen         = sizeof(zone_reclaim_mode),
1531                 .mode           = 0644,
1532                 .proc_handler   = proc_dointvec,
1533                 .extra1         = &zero,
1534         },
1535         {
1536                 .procname       = "min_unmapped_ratio",
1537                 .data           = &sysctl_min_unmapped_ratio,
1538                 .maxlen         = sizeof(sysctl_min_unmapped_ratio),
1539                 .mode           = 0644,
1540                 .proc_handler   = sysctl_min_unmapped_ratio_sysctl_handler,
1541                 .extra1         = &zero,
1542                 .extra2         = &one_hundred,
1543         },
1544         {
1545                 .procname       = "min_slab_ratio",
1546                 .data           = &sysctl_min_slab_ratio,
1547                 .maxlen         = sizeof(sysctl_min_slab_ratio),
1548                 .mode           = 0644,
1549                 .proc_handler   = sysctl_min_slab_ratio_sysctl_handler,
1550                 .extra1         = &zero,
1551                 .extra2         = &one_hundred,
1552         },
1553 #endif
1554 #ifdef CONFIG_SMP
1555         {
1556                 .procname       = "stat_interval",
1557                 .data           = &sysctl_stat_interval,
1558                 .maxlen         = sizeof(sysctl_stat_interval),
1559                 .mode           = 0644,
1560                 .proc_handler   = proc_dointvec_jiffies,
1561         },
1562 #endif
1563 #ifdef CONFIG_MMU
1564         {
1565                 .procname       = "mmap_min_addr",
1566                 .data           = &dac_mmap_min_addr,
1567                 .maxlen         = sizeof(unsigned long),
1568                 .mode           = 0644,
1569                 .proc_handler   = mmap_min_addr_handler,
1570         },
1571 #endif
1572 #ifdef CONFIG_NUMA
1573         {
1574                 .procname       = "numa_zonelist_order",
1575                 .data           = &numa_zonelist_order,
1576                 .maxlen         = NUMA_ZONELIST_ORDER_LEN,
1577                 .mode           = 0644,
1578                 .proc_handler   = numa_zonelist_order_handler,
1579         },
1580 #endif
1581 #if (defined(CONFIG_X86_32) && !defined(CONFIG_UML))|| \
1582    (defined(CONFIG_SUPERH) && defined(CONFIG_VSYSCALL))
1583         {
1584                 .procname       = "vdso_enabled",
1585 #ifdef CONFIG_X86_32
1586                 .data           = &vdso32_enabled,
1587                 .maxlen         = sizeof(vdso32_enabled),
1588 #else
1589                 .data           = &vdso_enabled,
1590                 .maxlen         = sizeof(vdso_enabled),
1591 #endif
1592                 .mode           = 0644,
1593                 .proc_handler   = proc_dointvec,
1594                 .extra1         = &zero,
1595         },
1596 #endif
1597 #ifdef CONFIG_HIGHMEM
1598         {
1599                 .procname       = "highmem_is_dirtyable",
1600                 .data           = &vm_highmem_is_dirtyable,
1601                 .maxlen         = sizeof(vm_highmem_is_dirtyable),
1602                 .mode           = 0644,
1603                 .proc_handler   = proc_dointvec_minmax,
1604                 .extra1         = &zero,
1605                 .extra2         = &one,
1606         },
1607 #endif
1608 #ifdef CONFIG_MEMORY_FAILURE
1609         {
1610                 .procname       = "memory_failure_early_kill",
1611                 .data           = &sysctl_memory_failure_early_kill,
1612                 .maxlen         = sizeof(sysctl_memory_failure_early_kill),
1613                 .mode           = 0644,
1614                 .proc_handler   = proc_dointvec_minmax,
1615                 .extra1         = &zero,
1616                 .extra2         = &one,
1617         },
1618         {
1619                 .procname       = "memory_failure_recovery",
1620                 .data           = &sysctl_memory_failure_recovery,
1621                 .maxlen         = sizeof(sysctl_memory_failure_recovery),
1622                 .mode           = 0644,
1623                 .proc_handler   = proc_dointvec_minmax,
1624                 .extra1         = &zero,
1625                 .extra2         = &one,
1626         },
1627 #endif
1628         {
1629                 .procname       = "user_reserve_kbytes",
1630                 .data           = &sysctl_user_reserve_kbytes,
1631                 .maxlen         = sizeof(sysctl_user_reserve_kbytes),
1632                 .mode           = 0644,
1633                 .proc_handler   = proc_doulongvec_minmax,
1634         },
1635         {
1636                 .procname       = "admin_reserve_kbytes",
1637                 .data           = &sysctl_admin_reserve_kbytes,
1638                 .maxlen         = sizeof(sysctl_admin_reserve_kbytes),
1639                 .mode           = 0644,
1640                 .proc_handler   = proc_doulongvec_minmax,
1641         },
1642 #ifdef CONFIG_HAVE_ARCH_MMAP_RND_BITS
1643         {
1644                 .procname       = "mmap_rnd_bits",
1645                 .data           = &mmap_rnd_bits,
1646                 .maxlen         = sizeof(mmap_rnd_bits),
1647                 .mode           = 0600,
1648                 .proc_handler   = proc_dointvec_minmax,
1649                 .extra1         = (void *)&mmap_rnd_bits_min,
1650                 .extra2         = (void *)&mmap_rnd_bits_max,
1651         },
1652 #endif
1653 #ifdef CONFIG_HAVE_ARCH_MMAP_RND_COMPAT_BITS
1654         {
1655                 .procname       = "mmap_rnd_compat_bits",
1656                 .data           = &mmap_rnd_compat_bits,
1657                 .maxlen         = sizeof(mmap_rnd_compat_bits),
1658                 .mode           = 0600,
1659                 .proc_handler   = proc_dointvec_minmax,
1660                 .extra1         = (void *)&mmap_rnd_compat_bits_min,
1661                 .extra2         = (void *)&mmap_rnd_compat_bits_max,
1662         },
1663 #endif
1664         { }
1665 };
1666
1667 static struct ctl_table fs_table[] = {
1668         {
1669                 .procname       = "inode-nr",
1670                 .data           = &inodes_stat,
1671                 .maxlen         = 2*sizeof(long),
1672                 .mode           = 0444,
1673                 .proc_handler   = proc_nr_inodes,
1674         },
1675         {
1676                 .procname       = "inode-state",
1677                 .data           = &inodes_stat,
1678                 .maxlen         = 7*sizeof(long),
1679                 .mode           = 0444,
1680                 .proc_handler   = proc_nr_inodes,
1681         },
1682         {
1683                 .procname       = "file-nr",
1684                 .data           = &files_stat,
1685                 .maxlen         = sizeof(files_stat),
1686                 .mode           = 0444,
1687                 .proc_handler   = proc_nr_files,
1688         },
1689         {
1690                 .procname       = "file-max",
1691                 .data           = &files_stat.max_files,
1692                 .maxlen         = sizeof(files_stat.max_files),
1693                 .mode           = 0644,
1694                 .proc_handler   = proc_doulongvec_minmax,
1695         },
1696         {
1697                 .procname       = "nr_open",
1698                 .data           = &sysctl_nr_open,
1699                 .maxlen         = sizeof(int),
1700                 .mode           = 0644,
1701                 .proc_handler   = proc_dointvec_minmax,
1702                 .extra1         = &sysctl_nr_open_min,
1703                 .extra2         = &sysctl_nr_open_max,
1704         },
1705         {
1706                 .procname       = "dentry-state",
1707                 .data           = &dentry_stat,
1708                 .maxlen         = 6*sizeof(long),
1709                 .mode           = 0444,
1710                 .proc_handler   = proc_nr_dentry,
1711         },
1712         {
1713                 .procname       = "overflowuid",
1714                 .data           = &fs_overflowuid,
1715                 .maxlen         = sizeof(int),
1716                 .mode           = 0644,
1717                 .proc_handler   = proc_dointvec_minmax,
1718                 .extra1         = &minolduid,
1719                 .extra2         = &maxolduid,
1720         },
1721         {
1722                 .procname       = "overflowgid",
1723                 .data           = &fs_overflowgid,
1724                 .maxlen         = sizeof(int),
1725                 .mode           = 0644,
1726                 .proc_handler   = proc_dointvec_minmax,
1727                 .extra1         = &minolduid,
1728                 .extra2         = &maxolduid,
1729         },
1730 #ifdef CONFIG_FILE_LOCKING
1731         {
1732                 .procname       = "leases-enable",
1733                 .data           = &leases_enable,
1734                 .maxlen         = sizeof(int),
1735                 .mode           = 0644,
1736                 .proc_handler   = proc_dointvec,
1737         },
1738 #endif
1739 #ifdef CONFIG_DNOTIFY
1740         {
1741                 .procname       = "dir-notify-enable",
1742                 .data           = &dir_notify_enable,
1743                 .maxlen         = sizeof(int),
1744                 .mode           = 0644,
1745                 .proc_handler   = proc_dointvec,
1746         },
1747 #endif
1748 #ifdef CONFIG_MMU
1749 #ifdef CONFIG_FILE_LOCKING
1750         {
1751                 .procname       = "lease-break-time",
1752                 .data           = &lease_break_time,
1753                 .maxlen         = sizeof(int),
1754                 .mode           = 0644,
1755                 .proc_handler   = proc_dointvec,
1756         },
1757 #endif
1758 #ifdef CONFIG_AIO
1759         {
1760                 .procname       = "aio-nr",
1761                 .data           = &aio_nr,
1762                 .maxlen         = sizeof(aio_nr),
1763                 .mode           = 0444,
1764                 .proc_handler   = proc_doulongvec_minmax,
1765         },
1766         {
1767                 .procname       = "aio-max-nr",
1768                 .data           = &aio_max_nr,
1769                 .maxlen         = sizeof(aio_max_nr),
1770                 .mode           = 0644,
1771                 .proc_handler   = proc_doulongvec_minmax,
1772         },
1773 #endif /* CONFIG_AIO */
1774 #ifdef CONFIG_INOTIFY_USER
1775         {
1776                 .procname       = "inotify",
1777                 .mode           = 0555,
1778                 .child          = inotify_table,
1779         },
1780 #endif
1781 #ifdef CONFIG_EPOLL
1782         {
1783                 .procname       = "epoll",
1784                 .mode           = 0555,
1785                 .child          = epoll_table,
1786         },
1787 #endif
1788 #endif
1789         {
1790                 .procname       = "protected_symlinks",
1791                 .data           = &sysctl_protected_symlinks,
1792                 .maxlen         = sizeof(int),
1793                 .mode           = 0600,
1794                 .proc_handler   = proc_dointvec_minmax,
1795                 .extra1         = &zero,
1796                 .extra2         = &one,
1797         },
1798         {
1799                 .procname       = "protected_hardlinks",
1800                 .data           = &sysctl_protected_hardlinks,
1801                 .maxlen         = sizeof(int),
1802                 .mode           = 0600,
1803                 .proc_handler   = proc_dointvec_minmax,
1804                 .extra1         = &zero,
1805                 .extra2         = &one,
1806         },
1807         {
1808                 .procname       = "suid_dumpable",
1809                 .data           = &suid_dumpable,
1810                 .maxlen         = sizeof(int),
1811                 .mode           = 0644,
1812                 .proc_handler   = proc_dointvec_minmax_coredump,
1813                 .extra1         = &zero,
1814                 .extra2         = &two,
1815         },
1816 #if defined(CONFIG_BINFMT_MISC) || defined(CONFIG_BINFMT_MISC_MODULE)
1817         {
1818                 .procname       = "binfmt_misc",
1819                 .mode           = 0555,
1820                 .child          = sysctl_mount_point,
1821         },
1822 #endif
1823         {
1824                 .procname       = "pipe-max-size",
1825                 .data           = &pipe_max_size,
1826                 .maxlen         = sizeof(int),
1827                 .mode           = 0644,
1828                 .proc_handler   = &pipe_proc_fn,
1829                 .extra1         = &pipe_min_size,
1830         },
1831         {
1832                 .procname       = "pipe-user-pages-hard",
1833                 .data           = &pipe_user_pages_hard,
1834                 .maxlen         = sizeof(pipe_user_pages_hard),
1835                 .mode           = 0644,
1836                 .proc_handler   = proc_doulongvec_minmax,
1837         },
1838         {
1839                 .procname       = "pipe-user-pages-soft",
1840                 .data           = &pipe_user_pages_soft,
1841                 .maxlen         = sizeof(pipe_user_pages_soft),
1842                 .mode           = 0644,
1843                 .proc_handler   = proc_doulongvec_minmax,
1844         },
1845         {
1846                 .procname       = "mount-max",
1847                 .data           = &sysctl_mount_max,
1848                 .maxlen         = sizeof(unsigned int),
1849                 .mode           = 0644,
1850                 .proc_handler   = proc_dointvec_minmax,
1851                 .extra1         = &one,
1852         },
1853         { }
1854 };
1855
1856 static struct ctl_table debug_table[] = {
1857 #ifdef CONFIG_SYSCTL_EXCEPTION_TRACE
1858         {
1859                 .procname       = "exception-trace",
1860                 .data           = &show_unhandled_signals,
1861                 .maxlen         = sizeof(int),
1862                 .mode           = 0644,
1863                 .proc_handler   = proc_dointvec
1864         },
1865 #endif
1866 #if defined(CONFIG_OPTPROBES)
1867         {
1868                 .procname       = "kprobes-optimization",
1869                 .data           = &sysctl_kprobes_optimization,
1870                 .maxlen         = sizeof(int),
1871                 .mode           = 0644,
1872                 .proc_handler   = proc_kprobes_optimization_handler,
1873                 .extra1         = &zero,
1874                 .extra2         = &one,
1875         },
1876 #endif
1877         { }
1878 };
1879
1880 static struct ctl_table dev_table[] = {
1881         { }
1882 };
1883
1884 int __init sysctl_init(void)
1885 {
1886         struct ctl_table_header *hdr;
1887
1888         hdr = register_sysctl_table(sysctl_base_table);
1889         kmemleak_not_leak(hdr);
1890         return 0;
1891 }
1892
1893 #endif /* CONFIG_SYSCTL */
1894
1895 /*
1896  * /proc/sys support
1897  */
1898
1899 #ifdef CONFIG_PROC_SYSCTL
1900
1901 static int _proc_do_string(char *data, int maxlen, int write,
1902                            char __user *buffer,
1903                            size_t *lenp, loff_t *ppos)
1904 {
1905         size_t len;
1906         char __user *p;
1907         char c;
1908
1909         if (!data || !maxlen || !*lenp) {
1910                 *lenp = 0;
1911                 return 0;
1912         }
1913
1914         if (write) {
1915                 if (sysctl_writes_strict == SYSCTL_WRITES_STRICT) {
1916                         /* Only continue writes not past the end of buffer. */
1917                         len = strlen(data);
1918                         if (len > maxlen - 1)
1919                                 len = maxlen - 1;
1920
1921                         if (*ppos > len)
1922                                 return 0;
1923                         len = *ppos;
1924                 } else {
1925                         /* Start writing from beginning of buffer. */
1926                         len = 0;
1927                 }
1928
1929                 *ppos += *lenp;
1930                 p = buffer;
1931                 while ((p - buffer) < *lenp && len < maxlen - 1) {
1932                         if (get_user(c, p++))
1933                                 return -EFAULT;
1934                         if (c == 0 || c == '\n')
1935                                 break;
1936                         data[len++] = c;
1937                 }
1938                 data[len] = 0;
1939         } else {
1940                 len = strlen(data);
1941                 if (len > maxlen)
1942                         len = maxlen;
1943
1944                 if (*ppos > len) {
1945                         *lenp = 0;
1946                         return 0;
1947                 }
1948
1949                 data += *ppos;
1950                 len  -= *ppos;
1951
1952                 if (len > *lenp)
1953                         len = *lenp;
1954                 if (len)
1955                         if (copy_to_user(buffer, data, len))
1956                                 return -EFAULT;
1957                 if (len < *lenp) {
1958                         if (put_user('\n', buffer + len))
1959                                 return -EFAULT;
1960                         len++;
1961                 }
1962                 *lenp = len;
1963                 *ppos += len;
1964         }
1965         return 0;
1966 }
1967
1968 static void warn_sysctl_write(struct ctl_table *table)
1969 {
1970         pr_warn_once("%s wrote to %s when file position was not 0!\n"
1971                 "This will not be supported in the future. To silence this\n"
1972                 "warning, set kernel.sysctl_writes_strict = -1\n",
1973                 current->comm, table->procname);
1974 }
1975
1976 /**
1977  * proc_dostring - read a string sysctl
1978  * @table: the sysctl table
1979  * @write: %TRUE if this is a write to the sysctl file
1980  * @buffer: the user buffer
1981  * @lenp: the size of the user buffer
1982  * @ppos: file position
1983  *
1984  * Reads/writes a string from/to the user buffer. If the kernel
1985  * buffer provided is not large enough to hold the string, the
1986  * string is truncated. The copied string is %NULL-terminated.
1987  * If the string is being read by the user process, it is copied
1988  * and a newline '\n' is added. It is truncated if the buffer is
1989  * not large enough.
1990  *
1991  * Returns 0 on success.
1992  */
1993 int proc_dostring(struct ctl_table *table, int write,
1994                   void __user *buffer, size_t *lenp, loff_t *ppos)
1995 {
1996         if (write && *ppos && sysctl_writes_strict == SYSCTL_WRITES_WARN)
1997                 warn_sysctl_write(table);
1998
1999         return _proc_do_string((char *)(table->data), table->maxlen, write,
2000                                (char __user *)buffer, lenp, ppos);
2001 }
2002
2003 static size_t proc_skip_spaces(char **buf)
2004 {
2005         size_t ret;
2006         char *tmp = skip_spaces(*buf);
2007         ret = tmp - *buf;
2008         *buf = tmp;
2009         return ret;
2010 }
2011
2012 static void proc_skip_char(char **buf, size_t *size, const char v)
2013 {
2014         while (*size) {
2015                 if (**buf != v)
2016                         break;
2017                 (*size)--;
2018                 (*buf)++;
2019         }
2020 }
2021
2022 #define TMPBUFLEN 22
2023 /**
2024  * proc_get_long - reads an ASCII formatted integer from a user buffer
2025  *
2026  * @buf: a kernel buffer
2027  * @size: size of the kernel buffer
2028  * @val: this is where the number will be stored
2029  * @neg: set to %TRUE if number is negative
2030  * @perm_tr: a vector which contains the allowed trailers
2031  * @perm_tr_len: size of the perm_tr vector
2032  * @tr: pointer to store the trailer character
2033  *
2034  * In case of success %0 is returned and @buf and @size are updated with
2035  * the amount of bytes read. If @tr is non-NULL and a trailing
2036  * character exists (size is non-zero after returning from this
2037  * function), @tr is updated with the trailing character.
2038  */
2039 static int proc_get_long(char **buf, size_t *size,
2040                           unsigned long *val, bool *neg,
2041                           const char *perm_tr, unsigned perm_tr_len, char *tr)
2042 {
2043         int len;
2044         char *p, tmp[TMPBUFLEN];
2045
2046         if (!*size)
2047                 return -EINVAL;
2048
2049         len = *size;
2050         if (len > TMPBUFLEN - 1)
2051                 len = TMPBUFLEN - 1;
2052
2053         memcpy(tmp, *buf, len);
2054
2055         tmp[len] = 0;
2056         p = tmp;
2057         if (*p == '-' && *size > 1) {
2058                 *neg = true;
2059                 p++;
2060         } else
2061                 *neg = false;
2062         if (!isdigit(*p))
2063                 return -EINVAL;
2064
2065         *val = simple_strtoul(p, &p, 0);
2066
2067         len = p - tmp;
2068
2069         /* We don't know if the next char is whitespace thus we may accept
2070          * invalid integers (e.g. 1234...a) or two integers instead of one
2071          * (e.g. 123...1). So lets not allow such large numbers. */
2072         if (len == TMPBUFLEN - 1)
2073                 return -EINVAL;
2074
2075         if (len < *size && perm_tr_len && !memchr(perm_tr, *p, perm_tr_len))
2076                 return -EINVAL;
2077
2078         if (tr && (len < *size))
2079                 *tr = *p;
2080
2081         *buf += len;
2082         *size -= len;
2083
2084         return 0;
2085 }
2086
2087 /**
2088  * proc_put_long - converts an integer to a decimal ASCII formatted string
2089  *
2090  * @buf: the user buffer
2091  * @size: the size of the user buffer
2092  * @val: the integer to be converted
2093  * @neg: sign of the number, %TRUE for negative
2094  *
2095  * In case of success %0 is returned and @buf and @size are updated with
2096  * the amount of bytes written.
2097  */
2098 static int proc_put_long(void __user **buf, size_t *size, unsigned long val,
2099                           bool neg)
2100 {
2101         int len;
2102         char tmp[TMPBUFLEN], *p = tmp;
2103
2104         sprintf(p, "%s%lu", neg ? "-" : "", val);
2105         len = strlen(tmp);
2106         if (len > *size)
2107                 len = *size;
2108         if (copy_to_user(*buf, tmp, len))
2109                 return -EFAULT;
2110         *size -= len;
2111         *buf += len;
2112         return 0;
2113 }
2114 #undef TMPBUFLEN
2115
2116 static int proc_put_char(void __user **buf, size_t *size, char c)
2117 {
2118         if (*size) {
2119                 char __user **buffer = (char __user **)buf;
2120                 if (put_user(c, *buffer))
2121                         return -EFAULT;
2122                 (*size)--, (*buffer)++;
2123                 *buf = *buffer;
2124         }
2125         return 0;
2126 }
2127
2128 static int do_proc_dointvec_conv(bool *negp, unsigned long *lvalp,
2129                                  int *valp,
2130                                  int write, void *data)
2131 {
2132         if (write) {
2133                 if (*negp) {
2134                         if (*lvalp > (unsigned long) INT_MAX + 1)
2135                                 return -EINVAL;
2136                         *valp = -*lvalp;
2137                 } else {
2138                         if (*lvalp > (unsigned long) INT_MAX)
2139                                 return -EINVAL;
2140                         *valp = *lvalp;
2141                 }
2142         } else {
2143                 int val = *valp;
2144                 if (val < 0) {
2145                         *negp = true;
2146                         *lvalp = -(unsigned long)val;
2147                 } else {
2148                         *negp = false;
2149                         *lvalp = (unsigned long)val;
2150                 }
2151         }
2152         return 0;
2153 }
2154
2155 static int do_proc_douintvec_conv(bool *negp, unsigned long *lvalp,
2156                                  int *valp,
2157                                  int write, void *data)
2158 {
2159         if (write) {
2160                 if (*negp)
2161                         return -EINVAL;
2162                 if (*lvalp > UINT_MAX)
2163                         return -EINVAL;
2164                 *valp = *lvalp;
2165         } else {
2166                 unsigned int val = *valp;
2167                 *negp = false;
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);