kernel/sysctl.c

   1 /*
   2  * sysctl.c: General linux system control interface
   3  *
   4  * Begun 24 March 1995, Stephen Tweedie
   5  * Added /proc support, Dec 1995
   6  * Added bdflush entry and intvec min/max checking, 2/23/96, Tom Dyas.
   7  * Added hooks for /proc/sys/net (minor, minor patch), 96/4/1, Mike Shaver.
   8  * Added kernel/java-{interpreter,appletviewer}, 96/5/10, Mike Shaver.
   9  * Dynamic registration fixes, Stephen Tweedie.
  10  * Added kswapd-interval, ctrl-alt-del, printk stuff, 1/8/97, Chris Horn.
  11  * Made sysctl support optional via CONFIG_SYSCTL, 1/10/97, Chris
  12  *  Horn.
  13  * Added proc_doulongvec_ms_jiffies_minmax, 09/08/99, Carlos H. Bauer.
  14  * Added proc_doulongvec_minmax, 09/08/99, Carlos H. Bauer.
  15  * Changed linked lists to use list.h instead of lists.h, 02/24/00, Bill
  16  *  Wendling.
  17  * The list_for_each() macro wasn't appropriate for the sysctl loop.
  18  *  Removed it and replaced it with older style, 03/23/00, Bill Wendling
  19  */
  20
  21 #include <linux/module.h>
  22 #include <linux/mm.h>
  23 #include <linux/swap.h>
  24 #include <linux/slab.h>
  25 #include <linux/sysctl.h>
  26 #include <linux/proc_fs.h>
  27 #include <linux/security.h>
  28 #include <linux/ctype.h>
  29 #include <linux/utsname.h>
  30 #include <linux/smp_lock.h>
  31 #include <linux/fs.h>
  32 #include <linux/init.h>
  33 #include <linux/kernel.h>
  34 #include <linux/kobject.h>
  35 #include <linux/net.h>
  36 #include <linux/sysrq.h>
  37 #include <linux/highuid.h>
  38 #include <linux/writeback.h>
  39 #include <linux/hugetlb.h>
  40 #include <linux/initrd.h>
  41 #include <linux/key.h>
  42 #include <linux/times.h>
  43 #include <linux/limits.h>
  44 #include <linux/dcache.h>
  45 #include <linux/syscalls.h>
  46 #include <linux/vmstat.h>
  47 #include <linux/nfs_fs.h>
  48 #include <linux/acpi.h>
  49 #include <linux/reboot.h>
  50 #include <linux/ftrace.h>
  51 #include <linux/slow-work.h>
  52
  53 #include <asm/uaccess.h>
  54 #include <asm/processor.h>
  55
  56 #ifdef CONFIG_X86
  57 #include <asm/nmi.h>
  58 #include <asm/stacktrace.h>
  59 #include <asm/io.h>
  60 #endif
  61
  62 static int deprecated_sysctl_warning(struct __sysctl_args *args);
  63
  64 #if defined(CONFIG_SYSCTL)
  65
  66 /* External variables not in a header file. */
  67 extern int C_A_D;
  68 extern int print_fatal_signals;
  69 extern int sysctl_overcommit_memory;
  70 extern int sysctl_overcommit_ratio;
  71 extern int sysctl_panic_on_oom;
  72 extern int sysctl_oom_kill_allocating_task;
  73 extern int sysctl_oom_dump_tasks;
  74 extern int max_threads;
  75 extern int core_uses_pid;
  76 extern int suid_dumpable;
  77 extern char core_pattern[];
  78 extern int pid_max;
  79 extern int min_free_kbytes;
  80 extern int pid_max_min, pid_max_max;
  81 extern int sysctl_drop_caches;
  82 extern int percpu_pagelist_fraction;
  83 extern int compat_log;
  84 extern int latencytop_enabled;
  85 extern int sysctl_nr_open_min, sysctl_nr_open_max;
  86 #ifndef CONFIG_MMU
  87 extern int sysctl_nr_trim_pages;
  88 #endif
  89 #ifdef CONFIG_RCU_TORTURE_TEST
  90 extern int rcutorture_runnable;
  91 #endif /* #ifdef CONFIG_RCU_TORTURE_TEST */
  92
  93 /* Constants used for minimum and  maximum */
  94 #ifdef CONFIG_DETECT_SOFTLOCKUP
  95 static int sixty = 60;
  96 static int neg_one = -1;
  97 #endif
  98
  99 static int zero;
 100 static int __maybe_unused one = 1;
 101 static int __maybe_unused two = 2;
 102 static unsigned long one_ul = 1;
 103 static int one_hundred = 100;
 104
 105 /* this is needed for the proc_dointvec_minmax for [fs_]overflow UID and GID */
 106 static int maxolduid = 65535;
 107 static int minolduid;
 108 static int min_percpu_pagelist_fract = 8;
 109
 110 static int ngroups_max = NGROUPS_MAX;
 111
 112 #ifdef CONFIG_MODULES
 113 extern char modprobe_path[];
 114 #endif
 115 #ifdef CONFIG_CHR_DEV_SG
 116 extern int sg_big_buff;
 117 #endif
 118
 119 #ifdef CONFIG_SPARC
 120 #include <asm/system.h>
 121 #endif
 122
 123 #ifdef CONFIG_SPARC64
 124 extern int sysctl_tsb_ratio;
 125 #endif
 126
 127 #ifdef __hppa__
 128 extern int pwrsw_enabled;
 129 extern int unaligned_enabled;
 130 #endif
 131
 132 #ifdef CONFIG_S390
 133 #ifdef CONFIG_MATHEMU
 134 extern int sysctl_ieee_emulation_warnings;
 135 #endif
 136 extern int sysctl_userprocess_debug;
 137 extern int spin_retry;
 138 #endif
 139
 140 #ifdef CONFIG_BSD_PROCESS_ACCT
 141 extern int acct_parm[];
 142 #endif
 143
 144 #ifdef CONFIG_IA64
 145 extern int no_unaligned_warning;
 146 extern int unaligned_dump_stack;
 147 #endif
 148
 149 #ifdef CONFIG_RT_MUTEXES
 150 extern int max_lock_depth;
 151 #endif
 152
 153 #ifdef CONFIG_PROC_SYSCTL
 154 static int proc_do_cad_pid(struct ctl_table *table, int write, struct file *filp,
 155                   void __user *buffer, size_t *lenp, loff_t *ppos);
 156 static int proc_taint(struct ctl_table *table, int write, struct file *filp,
 157                                void __user *buffer, size_t *lenp, loff_t *ppos);
 158 #endif
 159
 160 static struct ctl_table root_table[];
 161 static struct ctl_table_root sysctl_table_root;
 162 static struct ctl_table_header root_table_header = {
 163         .count = 1,
 164         .ctl_table = root_table,
 165         .ctl_entry = LIST_HEAD_INIT(sysctl_table_root.default_set.list),
 166         .root = &sysctl_table_root,
 167         .set = &sysctl_table_root.default_set,
 168 };
 169 static struct ctl_table_root sysctl_table_root = {
 170         .root_list = LIST_HEAD_INIT(sysctl_table_root.root_list),
 171         .default_set.list = LIST_HEAD_INIT(root_table_header.ctl_entry),
 172 };
 173
 174 static struct ctl_table kern_table[];
 175 static struct ctl_table vm_table[];
 176 static struct ctl_table fs_table[];
 177 static struct ctl_table debug_table[];
 178 static struct ctl_table dev_table[];
 179 extern struct ctl_table random_table[];
 180 #ifdef CONFIG_INOTIFY_USER
 181 extern struct ctl_table inotify_table[];
 182 #endif
 183 #ifdef CONFIG_EPOLL
 184 extern struct ctl_table epoll_table[];
 185 #endif
 186
 187 #ifdef HAVE_ARCH_PICK_MMAP_LAYOUT
 188 int sysctl_legacy_va_layout;
 189 #endif
 190
 191 extern int prove_locking;
 192 extern int lock_stat;
 193
 194 /* The default sysctl tables: */
 195
 196 static struct ctl_table root_table[] = {
 197         {
 198                 .ctl_name       = CTL_KERN,
 199                 .procname       = "kernel",
 200                 .mode           = 0555,
 201                 .child          = kern_table,
 202         },
 203         {
 204                 .ctl_name       = CTL_VM,
 205                 .procname       = "vm",
 206                 .mode           = 0555,
 207                 .child          = vm_table,
 208         },
 209         {
 210                 .ctl_name       = CTL_FS,
 211                 .procname       = "fs",
 212                 .mode           = 0555,
 213                 .child          = fs_table,
 214         },
 215         {
 216                 .ctl_name       = CTL_DEBUG,
 217                 .procname       = "debug",
 218                 .mode           = 0555,
 219                 .child          = debug_table,
 220         },
 221         {
 222                 .ctl_name       = CTL_DEV,
 223                 .procname       = "dev",
 224                 .mode           = 0555,
 225                 .child          = dev_table,
 226         },
 227 /*
 228  * NOTE: do not add new entries to this table unless you have read
 229  * Documentation/sysctl/ctl_unnumbered.txt
 230  */
 231         { .ctl_name = 0 }
 232 };
 233
 234 #ifdef CONFIG_SCHED_DEBUG
 235 static int min_sched_granularity_ns = 100000;           /* 100 usecs */
 236 static int max_sched_granularity_ns = NSEC_PER_SEC;     /* 1 second */
 237 static int min_wakeup_granularity_ns;                   /* 0 usecs */
 238 static int max_wakeup_granularity_ns = NSEC_PER_SEC;    /* 1 second */
 239 #endif
 240
 241 static struct ctl_table kern_table[] = {
 242 #ifdef CONFIG_SCHED_DEBUG
 243         {
 244                 .ctl_name       = CTL_UNNUMBERED,
 245                 .procname       = "sched_min_granularity_ns",
 246                 .data           = &sysctl_sched_min_granularity,
 247                 .maxlen         = sizeof(unsigned int),
 248                 .mode           = 0644,
 249                 .proc_handler   = &sched_nr_latency_handler,
 250                 .strategy       = &sysctl_intvec,
 251                 .extra1         = &min_sched_granularity_ns,
 252                 .extra2         = &max_sched_granularity_ns,
 253         },
 254         {
 255                 .ctl_name       = CTL_UNNUMBERED,
 256                 .procname       = "sched_latency_ns",
 257                 .data           = &sysctl_sched_latency,
 258                 .maxlen         = sizeof(unsigned int),
 259                 .mode           = 0644,
 260                 .proc_handler   = &sched_nr_latency_handler,
 261                 .strategy       = &sysctl_intvec,
 262                 .extra1         = &min_sched_granularity_ns,
 263                 .extra2         = &max_sched_granularity_ns,
 264         },
 265         {
 266                 .ctl_name       = CTL_UNNUMBERED,
 267                 .procname       = "sched_wakeup_granularity_ns",
 268                 .data           = &sysctl_sched_wakeup_granularity,
 269                 .maxlen         = sizeof(unsigned int),
 270                 .mode           = 0644,
 271                 .proc_handler   = &proc_dointvec_minmax,
 272                 .strategy       = &sysctl_intvec,
 273                 .extra1         = &min_wakeup_granularity_ns,
 274                 .extra2         = &max_wakeup_granularity_ns,
 275         },
 276         {
 277                 .ctl_name       = CTL_UNNUMBERED,
 278                 .procname       = "sched_shares_ratelimit",
 279                 .data           = &sysctl_sched_shares_ratelimit,
 280                 .maxlen         = sizeof(unsigned int),
 281                 .mode           = 0644,
 282                 .proc_handler   = &proc_dointvec,
 283         },
 284         {
 285                 .ctl_name       = CTL_UNNUMBERED,
 286                 .procname       = "sched_shares_thresh",
 287                 .data           = &sysctl_sched_shares_thresh,
 288                 .maxlen         = sizeof(unsigned int),
 289                 .mode           = 0644,
 290                 .proc_handler   = &proc_dointvec_minmax,
 291                 .strategy       = &sysctl_intvec,
 292                 .extra1         = &zero,
 293         },
 294         {
 295                 .ctl_name       = CTL_UNNUMBERED,
 296                 .procname       = "sched_child_runs_first",
 297                 .data           = &sysctl_sched_child_runs_first,
 298                 .maxlen         = sizeof(unsigned int),
 299                 .mode           = 0644,
 300                 .proc_handler   = &proc_dointvec,
 301         },
 302         {
 303                 .ctl_name       = CTL_UNNUMBERED,
 304                 .procname       = "sched_features",
 305                 .data           = &sysctl_sched_features,
 306                 .maxlen         = sizeof(unsigned int),
 307                 .mode           = 0644,
 308                 .proc_handler   = &proc_dointvec,
 309         },
 310         {
 311                 .ctl_name       = CTL_UNNUMBERED,
 312                 .procname       = "sched_migration_cost",
 313                 .data           = &sysctl_sched_migration_cost,
 314                 .maxlen         = sizeof(unsigned int),
 315                 .mode           = 0644,
 316                 .proc_handler   = &proc_dointvec,
 317         },
 318         {
 319                 .ctl_name       = CTL_UNNUMBERED,
 320                 .procname       = "sched_nr_migrate",
 321                 .data           = &sysctl_sched_nr_migrate,
 322                 .maxlen         = sizeof(unsigned int),
 323                 .mode           = 0644,
 324                 .proc_handler   = &proc_dointvec,
 325         },
 326 #endif
 327         {
 328                 .ctl_name       = CTL_UNNUMBERED,
 329                 .procname       = "sched_rt_period_us",
 330                 .data           = &sysctl_sched_rt_period,
 331                 .maxlen         = sizeof(unsigned int),
 332                 .mode           = 0644,
 333                 .proc_handler   = &sched_rt_handler,
 334         },
 335         {
 336                 .ctl_name       = CTL_UNNUMBERED,
 337                 .procname       = "sched_rt_runtime_us",
 338                 .data           = &sysctl_sched_rt_runtime,
 339                 .maxlen         = sizeof(int),
 340                 .mode           = 0644,
 341                 .proc_handler   = &sched_rt_handler,
 342         },
 343         {
 344                 .ctl_name       = CTL_UNNUMBERED,
 345                 .procname       = "sched_compat_yield",
 346                 .data           = &sysctl_sched_compat_yield,
 347                 .maxlen         = sizeof(unsigned int),
 348                 .mode           = 0644,
 349                 .proc_handler   = &proc_dointvec,
 350         },
 351 #ifdef CONFIG_PROVE_LOCKING
 352         {
 353                 .ctl_name       = CTL_UNNUMBERED,
 354                 .procname       = "prove_locking",
 355                 .data           = &prove_locking,
 356                 .maxlen         = sizeof(int),
 357                 .mode           = 0644,
 358                 .proc_handler   = &proc_dointvec,
 359         },
 360 #endif
 361 #ifdef CONFIG_LOCK_STAT
 362         {
 363                 .ctl_name       = CTL_UNNUMBERED,
 364                 .procname       = "lock_stat",
 365                 .data           = &lock_stat,
 366                 .maxlen         = sizeof(int),
 367                 .mode           = 0644,
 368                 .proc_handler   = &proc_dointvec,
 369         },
 370 #endif
 371         {
 372                 .ctl_name       = KERN_PANIC,
 373                 .procname       = "panic",
 374                 .data           = &panic_timeout,
 375                 .maxlen         = sizeof(int),
 376                 .mode           = 0644,
 377                 .proc_handler   = &proc_dointvec,
 378         },
 379         {
 380                 .ctl_name       = KERN_CORE_USES_PID,
 381                 .procname       = "core_uses_pid",
 382                 .data           = &core_uses_pid,
 383                 .maxlen         = sizeof(int),
 384                 .mode           = 0644,
 385                 .proc_handler   = &proc_dointvec,
 386         },
 387         {
 388                 .ctl_name       = KERN_CORE_PATTERN,
 389                 .procname       = "core_pattern",
 390                 .data           = core_pattern,
 391                 .maxlen         = CORENAME_MAX_SIZE,
 392                 .mode           = 0644,
 393                 .proc_handler   = &proc_dostring,
 394                 .strategy       = &sysctl_string,
 395         },
 396 #ifdef CONFIG_PROC_SYSCTL
 397         {
 398                 .procname       = "tainted",
 399                 .maxlen         = sizeof(long),
 400                 .mode           = 0644,
 401                 .proc_handler   = &proc_taint,
 402         },
 403 #endif
 404 #ifdef CONFIG_LATENCYTOP
 405         {
 406                 .procname       = "latencytop",
 407                 .data           = &latencytop_enabled,
 408                 .maxlen         = sizeof(int),
 409                 .mode           = 0644,
 410                 .proc_handler   = &proc_dointvec,
 411         },
 412 #endif
 413 #ifdef CONFIG_BLK_DEV_INITRD
 414         {
 415                 .ctl_name       = KERN_REALROOTDEV,
 416                 .procname       = "real-root-dev",
 417                 .data           = &real_root_dev,
 418                 .maxlen         = sizeof(int),
 419                 .mode           = 0644,
 420                 .proc_handler   = &proc_dointvec,
 421         },
 422 #endif
 423         {
 424                 .ctl_name       = CTL_UNNUMBERED,
 425                 .procname       = "print-fatal-signals",
 426                 .data           = &print_fatal_signals,
 427                 .maxlen         = sizeof(int),
 428                 .mode           = 0644,
 429                 .proc_handler   = &proc_dointvec,
 430         },
 431 #ifdef CONFIG_SPARC
 432         {
 433                 .ctl_name       = KERN_SPARC_REBOOT,
 434                 .procname       = "reboot-cmd",
 435                 .data           = reboot_command,
 436                 .maxlen         = 256,
 437                 .mode           = 0644,
 438                 .proc_handler   = &proc_dostring,
 439                 .strategy       = &sysctl_string,
 440         },
 441         {
 442                 .ctl_name       = KERN_SPARC_STOP_A,
 443                 .procname       = "stop-a",
 444                 .data           = &stop_a_enabled,
 445                 .maxlen         = sizeof (int),
 446                 .mode           = 0644,
 447                 .proc_handler   = &proc_dointvec,
 448         },
 449         {
 450                 .ctl_name       = KERN_SPARC_SCONS_PWROFF,
 451                 .procname       = "scons-poweroff",
 452                 .data           = &scons_pwroff,
 453                 .maxlen         = sizeof (int),
 454                 .mode           = 0644,
 455                 .proc_handler   = &proc_dointvec,
 456         },
 457 #endif
 458 #ifdef CONFIG_SPARC64
 459         {
 460                 .ctl_name       = CTL_UNNUMBERED,
 461                 .procname       = "tsb-ratio",
 462                 .data           = &sysctl_tsb_ratio,
 463                 .maxlen         = sizeof (int),
 464                 .mode           = 0644,
 465                 .proc_handler   = &proc_dointvec,
 466         },
 467 #endif
 468 #ifdef __hppa__
 469         {
 470                 .ctl_name       = KERN_HPPA_PWRSW,
 471                 .procname       = "soft-power",
 472                 .data           = &pwrsw_enabled,
 473                 .maxlen         = sizeof (int),
 474                 .mode           = 0644,
 475                 .proc_handler   = &proc_dointvec,
 476         },
 477         {
 478                 .ctl_name       = KERN_HPPA_UNALIGNED,
 479                 .procname       = "unaligned-trap",
 480                 .data           = &unaligned_enabled,
 481                 .maxlen         = sizeof (int),
 482                 .mode           = 0644,
 483                 .proc_handler   = &proc_dointvec,
 484         },
 485 #endif
 486         {
 487                 .ctl_name       = KERN_CTLALTDEL,
 488                 .procname       = "ctrl-alt-del",
 489                 .data           = &C_A_D,
 490                 .maxlen         = sizeof(int),
 491                 .mode           = 0644,
 492                 .proc_handler   = &proc_dointvec,
 493         },
 494 #ifdef CONFIG_FUNCTION_TRACER
 495         {
 496                 .ctl_name       = CTL_UNNUMBERED,
 497                 .procname       = "ftrace_enabled",
 498                 .data           = &ftrace_enabled,
 499                 .maxlen         = sizeof(int),
 500                 .mode           = 0644,
 501                 .proc_handler   = &ftrace_enable_sysctl,
 502         },
 503 #endif
 504 #ifdef CONFIG_STACK_TRACER
 505         {
 506                 .ctl_name       = CTL_UNNUMBERED,
 507                 .procname       = "stack_tracer_enabled",
 508                 .data           = &stack_tracer_enabled,
 509                 .maxlen         = sizeof(int),
 510                 .mode           = 0644,
 511                 .proc_handler   = &stack_trace_sysctl,
 512         },
 513 #endif
 514 #ifdef CONFIG_TRACING
 515         {
 516                 .ctl_name       = CTL_UNNUMBERED,
 517                 .procname       = "ftrace_dump_on_oops",
 518                 .data           = &ftrace_dump_on_oops,
 519                 .maxlen         = sizeof(int),
 520                 .mode           = 0644,
 521                 .proc_handler   = &proc_dointvec,
 522         },
 523 #endif
 524 #ifdef CONFIG_MODULES
 525         {
 526                 .ctl_name       = KERN_MODPROBE,
 527                 .procname       = "modprobe",
 528                 .data           = &modprobe_path,
 529                 .maxlen         = KMOD_PATH_LEN,
 530                 .mode           = 0644,
 531                 .proc_handler   = &proc_dostring,
 532                 .strategy       = &sysctl_string,
 533         },
 534 #endif
 535 #if defined(CONFIG_HOTPLUG) && defined(CONFIG_NET)
 536         {
 537                 .ctl_name       = KERN_HOTPLUG,
 538                 .procname       = "hotplug",
 539                 .data           = &uevent_helper,
 540                 .maxlen         = UEVENT_HELPER_PATH_LEN,
 541                 .mode           = 0644,
 542                 .proc_handler   = &proc_dostring,
 543                 .strategy       = &sysctl_string,
 544         },
 545 #endif
 546 #ifdef CONFIG_CHR_DEV_SG
 547         {
 548                 .ctl_name       = KERN_SG_BIG_BUFF,
 549                 .procname       = "sg-big-buff",
 550                 .data           = &sg_big_buff,
 551                 .maxlen         = sizeof (int),
 552                 .mode           = 0444,
 553                 .proc_handler   = &proc_dointvec,
 554         },
 555 #endif
 556 #ifdef CONFIG_BSD_PROCESS_ACCT
 557         {
 558                 .ctl_name       = KERN_ACCT,
 559                 .procname       = "acct",
 560                 .data           = &acct_parm,
 561                 .maxlen         = 3*sizeof(int),
 562                 .mode           = 0644,
 563                 .proc_handler   = &proc_dointvec,
 564         },
 565 #endif
 566 #ifdef CONFIG_MAGIC_SYSRQ
 567         {
 568                 .ctl_name       = KERN_SYSRQ,
 569                 .procname       = "sysrq",
 570                 .data           = &__sysrq_enabled,
 571                 .maxlen         = sizeof (int),
 572                 .mode           = 0644,
 573                 .proc_handler   = &proc_dointvec,
 574         },
 575 #endif
 576 #ifdef CONFIG_PROC_SYSCTL
 577         {
 578                 .procname       = "cad_pid",
 579                 .data           = NULL,
 580                 .maxlen         = sizeof (int),
 581                 .mode           = 0600,
 582                 .proc_handler   = &proc_do_cad_pid,
 583         },
 584 #endif
 585         {
 586                 .ctl_name       = KERN_MAX_THREADS,
 587                 .procname       = "threads-max",
 588                 .data           = &max_threads,
 589                 .maxlen         = sizeof(int),
 590                 .mode           = 0644,
 591                 .proc_handler   = &proc_dointvec,
 592         },
 593         {
 594                 .ctl_name       = KERN_RANDOM,
 595                 .procname       = "random",
 596                 .mode           = 0555,
 597                 .child          = random_table,
 598         },
 599         {
 600                 .ctl_name       = KERN_OVERFLOWUID,
 601                 .procname       = "overflowuid",
 602                 .data           = &overflowuid,
 603                 .maxlen         = sizeof(int),
 604                 .mode           = 0644,
 605                 .proc_handler   = &proc_dointvec_minmax,
 606                 .strategy       = &sysctl_intvec,
 607                 .extra1         = &minolduid,
 608                 .extra2         = &maxolduid,
 609         },
 610         {
 611                 .ctl_name       = KERN_OVERFLOWGID,
 612                 .procname       = "overflowgid",
 613                 .data           = &overflowgid,
 614                 .maxlen         = sizeof(int),
 615                 .mode           = 0644,
 616                 .proc_handler   = &proc_dointvec_minmax,
 617                 .strategy       = &sysctl_intvec,
 618                 .extra1         = &minolduid,
 619                 .extra2         = &maxolduid,
 620         },
 621 #ifdef CONFIG_S390
 622 #ifdef CONFIG_MATHEMU
 623         {
 624                 .ctl_name       = KERN_IEEE_EMULATION_WARNINGS,
 625                 .procname       = "ieee_emulation_warnings",
 626                 .data           = &sysctl_ieee_emulation_warnings,
 627                 .maxlen         = sizeof(int),
 628                 .mode           = 0644,
 629                 .proc_handler   = &proc_dointvec,
 630         },
 631 #endif
 632         {
 633                 .ctl_name       = KERN_S390_USER_DEBUG_LOGGING,
 634                 .procname       = "userprocess_debug",
 635                 .data           = &sysctl_userprocess_debug,
 636                 .maxlen         = sizeof(int),
 637                 .mode           = 0644,
 638                 .proc_handler   = &proc_dointvec,
 639         },
 640 #endif
 641         {
 642                 .ctl_name       = KERN_PIDMAX,
 643                 .procname       = "pid_max",
 644                 .data           = &pid_max,
 645                 .maxlen         = sizeof (int),
 646                 .mode           = 0644,
 647                 .proc_handler   = &proc_dointvec_minmax,
 648                 .strategy       = sysctl_intvec,
 649                 .extra1         = &pid_max_min,
 650                 .extra2         = &pid_max_max,
 651         },
 652         {
 653                 .ctl_name       = KERN_PANIC_ON_OOPS,
 654                 .procname       = "panic_on_oops",
 655                 .data           = &panic_on_oops,
 656                 .maxlen         = sizeof(int),
 657                 .mode           = 0644,
 658                 .proc_handler   = &proc_dointvec,
 659         },
 660 #if defined CONFIG_PRINTK
 661         {
 662                 .ctl_name       = KERN_PRINTK,
 663                 .procname       = "printk",
 664                 .data           = &console_loglevel,
 665                 .maxlen         = 4*sizeof(int),
 666                 .mode           = 0644,
 667                 .proc_handler   = &proc_dointvec,
 668         },
 669         {
 670                 .ctl_name       = KERN_PRINTK_RATELIMIT,
 671                 .procname       = "printk_ratelimit",
 672                 .data           = &printk_ratelimit_state.interval,
 673                 .maxlen         = sizeof(int),
 674                 .mode           = 0644,
 675                 .proc_handler   = &proc_dointvec_jiffies,
 676                 .strategy       = &sysctl_jiffies,
 677         },
 678         {
 679                 .ctl_name       = KERN_PRINTK_RATELIMIT_BURST,
 680                 .procname       = "printk_ratelimit_burst",
 681                 .data           = &printk_ratelimit_state.burst,
 682                 .maxlen         = sizeof(int),
 683                 .mode           = 0644,
 684                 .proc_handler   = &proc_dointvec,
 685         },
 686 #endif
 687         {
 688                 .ctl_name       = KERN_NGROUPS_MAX,
 689                 .procname       = "ngroups_max",
 690                 .data           = &ngroups_max,
 691                 .maxlen         = sizeof (int),
 692                 .mode           = 0444,
 693                 .proc_handler   = &proc_dointvec,
 694         },
 695 #if defined(CONFIG_X86_LOCAL_APIC) && defined(CONFIG_X86)
 696         {
 697                 .ctl_name       = KERN_UNKNOWN_NMI_PANIC,
 698                 .procname       = "unknown_nmi_panic",
 699                 .data           = &unknown_nmi_panic,
 700                 .maxlen         = sizeof (int),
 701                 .mode           = 0644,
 702                 .proc_handler   = &proc_dointvec,
 703         },
 704         {
 705                 .procname       = "nmi_watchdog",
 706                 .data           = &nmi_watchdog_enabled,
 707                 .maxlen         = sizeof (int),
 708                 .mode           = 0644,
 709                 .proc_handler   = &proc_nmi_enabled,
 710         },
 711 #endif
 712 #if defined(CONFIG_X86)
 713         {
 714                 .ctl_name       = KERN_PANIC_ON_NMI,
 715                 .procname       = "panic_on_unrecovered_nmi",
 716                 .data           = &panic_on_unrecovered_nmi,
 717                 .maxlen         = sizeof(int),
 718                 .mode           = 0644,
 719                 .proc_handler   = &proc_dointvec,
 720         },
 721         {
 722                 .ctl_name       = KERN_BOOTLOADER_TYPE,
 723                 .procname       = "bootloader_type",
 724                 .data           = &bootloader_type,
 725                 .maxlen         = sizeof (int),
 726                 .mode           = 0444,
 727                 .proc_handler   = &proc_dointvec,
 728         },
 729         {
 730                 .ctl_name       = CTL_UNNUMBERED,
 731                 .procname       = "kstack_depth_to_print",
 732                 .data           = &kstack_depth_to_print,
 733                 .maxlen         = sizeof(int),
 734                 .mode           = 0644,
 735                 .proc_handler   = &proc_dointvec,
 736         },
 737         {
 738                 .ctl_name       = CTL_UNNUMBERED,
 739                 .procname       = "io_delay_type",
 740                 .data           = &io_delay_type,
 741                 .maxlen         = sizeof(int),
 742                 .mode           = 0644,
 743                 .proc_handler   = &proc_dointvec,
 744         },
 745 #endif
 746 #if defined(CONFIG_MMU)
 747         {
 748                 .ctl_name       = KERN_RANDOMIZE,
 749                 .procname       = "randomize_va_space",
 750                 .data           = &randomize_va_space,
 751                 .maxlen         = sizeof(int),
 752                 .mode           = 0644,
 753                 .proc_handler   = &proc_dointvec,
 754         },
 755 #endif
 756 #if defined(CONFIG_S390) && defined(CONFIG_SMP)
 757         {
 758                 .ctl_name       = KERN_SPIN_RETRY,
 759                 .procname       = "spin_retry",
 760                 .data           = &spin_retry,
 761                 .maxlen         = sizeof (int),
 762                 .mode           = 0644,
 763                 .proc_handler   = &proc_dointvec,
 764         },
 765 #endif
 766 #if     defined(CONFIG_ACPI_SLEEP) && defined(CONFIG_X86)
 767         {
 768                 .procname       = "acpi_video_flags",
 769                 .data           = &acpi_realmode_flags,
 770                 .maxlen         = sizeof (unsigned long),
 771                 .mode           = 0644,
 772                 .proc_handler   = &proc_doulongvec_minmax,
 773         },
 774 #endif
 775 #ifdef CONFIG_IA64
 776         {
 777                 .ctl_name       = KERN_IA64_UNALIGNED,
 778                 .procname       = "ignore-unaligned-usertrap",
 779                 .data           = &no_unaligned_warning,
 780                 .maxlen         = sizeof (int),
 781                 .mode           = 0644,
 782                 .proc_handler   = &proc_dointvec,
 783         },
 784         {
 785                 .ctl_name       = CTL_UNNUMBERED,
 786                 .procname       = "unaligned-dump-stack",
 787                 .data           = &unaligned_dump_stack,
 788                 .maxlen         = sizeof (int),
 789                 .mode           = 0644,
 790                 .proc_handler   = &proc_dointvec,
 791         },
 792 #endif
 793 #ifdef CONFIG_DETECT_SOFTLOCKUP
 794         {
 795                 .ctl_name       = CTL_UNNUMBERED,
 796                 .procname       = "softlockup_panic",
 797                 .data           = &softlockup_panic,
 798                 .maxlen         = sizeof(int),
 799                 .mode           = 0644,
 800                 .proc_handler   = &proc_dointvec_minmax,
 801                 .strategy       = &sysctl_intvec,
 802                 .extra1         = &zero,
 803                 .extra2         = &one,
 804         },
 805         {
 806                 .ctl_name       = CTL_UNNUMBERED,
 807                 .procname       = "softlockup_thresh",
 808                 .data           = &softlockup_thresh,
 809                 .maxlen         = sizeof(int),
 810                 .mode           = 0644,
 811                 .proc_handler   = &proc_dosoftlockup_thresh,
 812                 .strategy       = &sysctl_intvec,
 813                 .extra1         = &neg_one,
 814                 .extra2         = &sixty,
 815         },
 816 #endif
 817 #ifdef CONFIG_DETECT_HUNG_TASK
 818         {
 819                 .ctl_name       = CTL_UNNUMBERED,
 820                 .procname       = "hung_task_panic",
 821                 .data           = &sysctl_hung_task_panic,
 822                 .maxlen         = sizeof(int),
 823                 .mode           = 0644,
 824                 .proc_handler   = &proc_dointvec_minmax,
 825                 .strategy       = &sysctl_intvec,
 826                 .extra1         = &zero,
 827                 .extra2         = &one,
 828         },
 829         {
 830                 .ctl_name       = CTL_UNNUMBERED,
 831                 .procname       = "hung_task_check_count",
 832                 .data           = &sysctl_hung_task_check_count,
 833                 .maxlen         = sizeof(unsigned long),
 834                 .mode           = 0644,
 835                 .proc_handler   = &proc_doulongvec_minmax,
 836                 .strategy       = &sysctl_intvec,
 837         },
 838         {
 839                 .ctl_name       = CTL_UNNUMBERED,
 840                 .procname       = "hung_task_timeout_secs",
 841                 .data           = &sysctl_hung_task_timeout_secs,
 842                 .maxlen         = sizeof(unsigned long),
 843                 .mode           = 0644,
 844                 .proc_handler   = &proc_dohung_task_timeout_secs,
 845                 .strategy       = &sysctl_intvec,
 846         },
 847         {
 848                 .ctl_name       = CTL_UNNUMBERED,
 849                 .procname       = "hung_task_warnings",
 850                 .data           = &sysctl_hung_task_warnings,
 851                 .maxlen         = sizeof(unsigned long),
 852                 .mode           = 0644,
 853                 .proc_handler   = &proc_doulongvec_minmax,
 854                 .strategy       = &sysctl_intvec,
 855         },
 856 #endif
 857 #ifdef CONFIG_COMPAT
 858         {
 859                 .ctl_name       = KERN_COMPAT_LOG,
 860                 .procname       = "compat-log",
 861                 .data           = &compat_log,
 862                 .maxlen         = sizeof (int),
 863                 .mode           = 0644,
 864                 .proc_handler   = &proc_dointvec,
 865         },
 866 #endif
 867 #ifdef CONFIG_RT_MUTEXES
 868         {
 869                 .ctl_name       = KERN_MAX_LOCK_DEPTH,
 870                 .procname       = "max_lock_depth",
 871                 .data           = &max_lock_depth,
 872                 .maxlen         = sizeof(int),
 873                 .mode           = 0644,
 874                 .proc_handler   = &proc_dointvec,
 875         },
 876 #endif
 877         {
 878                 .ctl_name       = CTL_UNNUMBERED,
 879                 .procname       = "poweroff_cmd",
 880                 .data           = &poweroff_cmd,
 881                 .maxlen         = POWEROFF_CMD_PATH_LEN,
 882                 .mode           = 0644,
 883                 .proc_handler   = &proc_dostring,
 884                 .strategy       = &sysctl_string,
 885         },
 886 #ifdef CONFIG_KEYS
 887         {
 888                 .ctl_name       = CTL_UNNUMBERED,
 889                 .procname       = "keys",
 890                 .mode           = 0555,
 891                 .child          = key_sysctls,
 892         },
 893 #endif
 894 #ifdef CONFIG_RCU_TORTURE_TEST
 895         {
 896                 .ctl_name       = CTL_UNNUMBERED,
 897                 .procname       = "rcutorture_runnable",
 898                 .data           = &rcutorture_runnable,
 899                 .maxlen         = sizeof(int),
 900                 .mode           = 0644,
 901                 .proc_handler   = &proc_dointvec,
 902         },
 903 #endif
 904 #ifdef CONFIG_UNEVICTABLE_LRU
 905         {
 906                 .ctl_name       = CTL_UNNUMBERED,
 907                 .procname       = "scan_unevictable_pages",
 908                 .data           = &scan_unevictable_pages,
 909                 .maxlen         = sizeof(scan_unevictable_pages),
 910                 .mode           = 0644,
 911                 .proc_handler   = &scan_unevictable_handler,
 912         },
 913 #endif
 914 #ifdef CONFIG_SLOW_WORK
 915         {
 916                 .ctl_name       = CTL_UNNUMBERED,
 917                 .procname       = "slow-work",
 918                 .mode           = 0555,
 919                 .child          = slow_work_sysctls,
 920         },
 921 #endif
 922 /*
 923  * NOTE: do not add new entries to this table unless you have read
 924  * Documentation/sysctl/ctl_unnumbered.txt
 925  */
 926         { .ctl_name = 0 }
 927 };
 928
 929 static struct ctl_table vm_table[] = {
 930         {
 931                 .ctl_name       = VM_OVERCOMMIT_MEMORY,
 932                 .procname       = "overcommit_memory",
 933                 .data           = &sysctl_overcommit_memory,
 934                 .maxlen         = sizeof(sysctl_overcommit_memory),
 935                 .mode           = 0644,
 936                 .proc_handler   = &proc_dointvec,
 937         },
 938         {
 939                 .ctl_name       = VM_PANIC_ON_OOM,
 940                 .procname       = "panic_on_oom",
 941                 .data           = &sysctl_panic_on_oom,
 942                 .maxlen         = sizeof(sysctl_panic_on_oom),
 943                 .mode           = 0644,
 944                 .proc_handler   = &proc_dointvec,
 945         },
 946         {
 947                 .ctl_name       = CTL_UNNUMBERED,
 948                 .procname       = "oom_kill_allocating_task",
 949                 .data           = &sysctl_oom_kill_allocating_task,
 950                 .maxlen         = sizeof(sysctl_oom_kill_allocating_task),
 951                 .mode           = 0644,
 952                 .proc_handler   = &proc_dointvec,
 953         },
 954         {
 955                 .ctl_name       = CTL_UNNUMBERED,
 956                 .procname       = "oom_dump_tasks",
 957                 .data           = &sysctl_oom_dump_tasks,
 958                 .maxlen         = sizeof(sysctl_oom_dump_tasks),
 959                 .mode           = 0644,
 960                 .proc_handler   = &proc_dointvec,
 961         },
 962         {
 963                 .ctl_name       = VM_OVERCOMMIT_RATIO,
 964                 .procname       = "overcommit_ratio",
 965                 .data           = &sysctl_overcommit_ratio,
 966                 .maxlen         = sizeof(sysctl_overcommit_ratio),
 967                 .mode           = 0644,
 968                 .proc_handler   = &proc_dointvec,
 969         },
 970         {
 971                 .ctl_name       = VM_PAGE_CLUSTER,
 972                 .procname       = "page-cluster",
 973                 .data           = &page_cluster,
 974                 .maxlen         = sizeof(int),
 975                 .mode           = 0644,
 976                 .proc_handler   = &proc_dointvec,
 977         },
 978         {
 979                 .ctl_name       = VM_DIRTY_BACKGROUND,
 980                 .procname       = "dirty_background_ratio",
 981                 .data           = &dirty_background_ratio,
 982                 .maxlen         = sizeof(dirty_background_ratio),
 983                 .mode           = 0644,
 984                 .proc_handler   = &dirty_background_ratio_handler,
 985                 .strategy       = &sysctl_intvec,
 986                 .extra1         = &zero,
 987                 .extra2         = &one_hundred,
 988         },
 989         {
 990                 .ctl_name       = CTL_UNNUMBERED,
 991                 .procname       = "dirty_background_bytes",
 992                 .data           = &dirty_background_bytes,
 993                 .maxlen         = sizeof(dirty_background_bytes),
 994                 .mode           = 0644,
 995                 .proc_handler   = &dirty_background_bytes_handler,
 996                 .strategy       = &sysctl_intvec,
 997                 .extra1         = &one_ul,
 998         },
 999         {
1000                 .ctl_name       = VM_DIRTY_RATIO,
1001                 .procname       = "dirty_ratio",
1002                 .data           = &vm_dirty_ratio,
1003                 .maxlen         = sizeof(vm_dirty_ratio),
1004                 .mode           = 0644,
1005                 .proc_handler   = &dirty_ratio_handler,
1006                 .strategy       = &sysctl_intvec,
1007                 .extra1         = &zero,
1008                 .extra2         = &one_hundred,
1009         },
1010         {
1011                 .ctl_name       = CTL_UNNUMBERED,
1012                 .procname       = "dirty_bytes",
1013                 .data           = &vm_dirty_bytes,
1014                 .maxlen         = sizeof(vm_dirty_bytes),
1015                 .mode           = 0644,
1016                 .proc_handler   = &dirty_bytes_handler,
1017                 .strategy       = &sysctl_intvec,
1018                 .extra1         = &one_ul,
1019         },
1020         {
1021                 .procname       = "dirty_writeback_centisecs",
1022                 .data           = &dirty_writeback_interval,
1023                 .maxlen         = sizeof(dirty_writeback_interval),
1024                 .mode           = 0644,
1025                 .proc_handler   = &dirty_writeback_centisecs_handler,
1026         },
1027         {
1028                 .procname       = "dirty_expire_centisecs",
1029                 .data           = &dirty_expire_interval,
1030                 .maxlen         = sizeof(dirty_expire_interval),
1031                 .mode           = 0644,
1032                 .proc_handler   = &proc_dointvec,
1033         },
1034         {
1035                 .ctl_name       = VM_NR_PDFLUSH_THREADS,
1036                 .procname       = "nr_pdflush_threads",
1037                 .data           = &nr_pdflush_threads,
1038                 .maxlen         = sizeof nr_pdflush_threads,
1039                 .mode           = 0444 /* read-only*/,
1040                 .proc_handler   = &proc_dointvec,
1041         },
1042         {
1043                 .ctl_name       = VM_SWAPPINESS,
1044                 .procname       = "swappiness",
1045                 .data           = &vm_swappiness,
1046                 .maxlen         = sizeof(vm_swappiness),
1047                 .mode           = 0644,
1048                 .proc_handler   = &proc_dointvec_minmax,
1049                 .strategy       = &sysctl_intvec,
1050                 .extra1         = &zero,
1051                 .extra2         = &one_hundred,
1052         },
1053 #ifdef CONFIG_HUGETLB_PAGE
1054          {
1055                 .procname       = "nr_hugepages",
1056                 .data           = NULL,
1057                 .maxlen         = sizeof(unsigned long),
1058                 .mode           = 0644,
1059                 .proc_handler   = &hugetlb_sysctl_handler,
1060                 .extra1         = (void *)&hugetlb_zero,
1061                 .extra2         = (void *)&hugetlb_infinity,
1062          },
1063          {
1064                 .ctl_name       = VM_HUGETLB_GROUP,
1065                 .procname       = "hugetlb_shm_group",
1066                 .data           = &sysctl_hugetlb_shm_group,
1067                 .maxlen         = sizeof(gid_t),
1068                 .mode           = 0644,
1069                 .proc_handler   = &proc_dointvec,
1070          },
1071          {
1072                 .ctl_name       = CTL_UNNUMBERED,
1073                 .procname       = "hugepages_treat_as_movable",
1074                 .data           = &hugepages_treat_as_movable,
1075                 .maxlen         = sizeof(int),
1076                 .mode           = 0644,
1077                 .proc_handler   = &hugetlb_treat_movable_handler,
1078         },
1079         {
1080                 .ctl_name       = CTL_UNNUMBERED,
1081                 .procname       = "nr_overcommit_hugepages",
1082                 .data           = NULL,
1083                 .maxlen         = sizeof(unsigned long),
1084                 .mode           = 0644,
1085                 .proc_handler   = &hugetlb_overcommit_handler,
1086                 .extra1         = (void *)&hugetlb_zero,
1087                 .extra2         = (void *)&hugetlb_infinity,
1088         },
1089 #endif
1090         {
1091                 .ctl_name       = VM_LOWMEM_RESERVE_RATIO,
1092                 .procname       = "lowmem_reserve_ratio",
1093                 .data           = &sysctl_lowmem_reserve_ratio,
1094                 .maxlen         = sizeof(sysctl_lowmem_reserve_ratio),
1095                 .mode           = 0644,
1096                 .proc_handler   = &lowmem_reserve_ratio_sysctl_handler,
1097                 .strategy       = &sysctl_intvec,
1098         },
1099         {
1100                 .ctl_name       = VM_DROP_PAGECACHE,
1101                 .procname       = "drop_caches",
1102                 .data           = &sysctl_drop_caches,
1103                 .maxlen         = sizeof(int),
1104                 .mode           = 0644,
1105                 .proc_handler   = drop_caches_sysctl_handler,
1106                 .strategy       = &sysctl_intvec,
1107         },
1108         {
1109                 .ctl_name       = VM_MIN_FREE_KBYTES,
1110                 .procname       = "min_free_kbytes",
1111                 .data           = &min_free_kbytes,
1112                 .maxlen         = sizeof(min_free_kbytes),
1113                 .mode           = 0644,
1114                 .proc_handler   = &min_free_kbytes_sysctl_handler,
1115                 .strategy       = &sysctl_intvec,
1116                 .extra1         = &zero,
1117         },
1118         {
1119                 .ctl_name       = VM_PERCPU_PAGELIST_FRACTION,
1120                 .procname       = "percpu_pagelist_fraction",
1121                 .data           = &percpu_pagelist_fraction,
1122                 .maxlen         = sizeof(percpu_pagelist_fraction),
1123                 .mode           = 0644,
1124                 .proc_handler   = &percpu_pagelist_fraction_sysctl_handler,
1125                 .strategy       = &sysctl_intvec,
1126                 .extra1         = &min_percpu_pagelist_fract,
1127         },
1128 #ifdef CONFIG_MMU
1129         {
1130                 .ctl_name       = VM_MAX_MAP_COUNT,
1131                 .procname       = "max_map_count",
1132                 .data           = &sysctl_max_map_count,
1133                 .maxlen         = sizeof(sysctl_max_map_count),
1134                 .mode           = 0644,
1135                 .proc_handler   = &proc_dointvec
1136         },
1137 #else
1138         {
1139                 .ctl_name       = CTL_UNNUMBERED,
1140                 .procname       = "nr_trim_pages",
1141                 .data           = &sysctl_nr_trim_pages,
1142                 .maxlen         = sizeof(sysctl_nr_trim_pages),
1143                 .mode           = 0644,
1144                 .proc_handler   = &proc_dointvec_minmax,
1145                 .strategy       = &sysctl_intvec,
1146                 .extra1         = &zero,
1147         },
1148 #endif
1149         {
1150                 .ctl_name       = VM_LAPTOP_MODE,
1151                 .procname       = "laptop_mode",
1152                 .data           = &laptop_mode,
1153                 .maxlen         = sizeof(laptop_mode),
1154                 .mode           = 0644,
1155                 .proc_handler   = &proc_dointvec_jiffies,
1156                 .strategy       = &sysctl_jiffies,
1157         },
1158         {
1159                 .ctl_name       = VM_BLOCK_DUMP,
1160                 .procname       = "block_dump",
1161                 .data           = &block_dump,
1162                 .maxlen         = sizeof(block_dump),
1163                 .mode           = 0644,
1164                 .proc_handler   = &proc_dointvec,
1165                 .strategy       = &sysctl_intvec,
1166                 .extra1         = &zero,
1167         },
1168         {
1169                 .ctl_name       = VM_VFS_CACHE_PRESSURE,
1170                 .procname       = "vfs_cache_pressure",
1171                 .data           = &sysctl_vfs_cache_pressure,
1172                 .maxlen         = sizeof(sysctl_vfs_cache_pressure),
1173                 .mode           = 0644,
1174                 .proc_handler   = &proc_dointvec,
1175                 .strategy       = &sysctl_intvec,
1176                 .extra1         = &zero,
1177         },
1178 #ifdef HAVE_ARCH_PICK_MMAP_LAYOUT
1179         {
1180                 .ctl_name       = VM_LEGACY_VA_LAYOUT,
1181                 .procname       = "legacy_va_layout",
1182                 .data           = &sysctl_legacy_va_layout,
1183                 .maxlen         = sizeof(sysctl_legacy_va_layout),
1184                 .mode           = 0644,
1185                 .proc_handler   = &proc_dointvec,
1186                 .strategy       = &sysctl_intvec,
1187                 .extra1         = &zero,
1188         },
1189 #endif
1190 #ifdef CONFIG_NUMA
1191         {
1192                 .ctl_name       = VM_ZONE_RECLAIM_MODE,
1193                 .procname       = "zone_reclaim_mode",
1194                 .data           = &zone_reclaim_mode,
1195                 .maxlen         = sizeof(zone_reclaim_mode),
1196                 .mode           = 0644,
1197                 .proc_handler   = &proc_dointvec,
1198                 .strategy       = &sysctl_intvec,
1199                 .extra1         = &zero,
1200         },
1201         {
1202                 .ctl_name       = VM_MIN_UNMAPPED,
1203                 .procname       = "min_unmapped_ratio",
1204                 .data           = &sysctl_min_unmapped_ratio,
1205                 .maxlen         = sizeof(sysctl_min_unmapped_ratio),
1206                 .mode           = 0644,
1207                 .proc_handler   = &sysctl_min_unmapped_ratio_sysctl_handler,
1208                 .strategy       = &sysctl_intvec,
1209                 .extra1         = &zero,
1210                 .extra2         = &one_hundred,
1211         },
1212         {
1213                 .ctl_name       = VM_MIN_SLAB,
1214                 .procname       = "min_slab_ratio",
1215                 .data           = &sysctl_min_slab_ratio,
1216                 .maxlen         = sizeof(sysctl_min_slab_ratio),
1217                 .mode           = 0644,
1218                 .proc_handler   = &sysctl_min_slab_ratio_sysctl_handler,
1219                 .strategy       = &sysctl_intvec,
1220                 .extra1         = &zero,
1221                 .extra2         = &one_hundred,
1222         },
1223 #endif
1224 #ifdef CONFIG_SMP
1225         {
1226                 .ctl_name       = CTL_UNNUMBERED,
1227                 .procname       = "stat_interval",
1228                 .data           = &sysctl_stat_interval,
1229                 .maxlen         = sizeof(sysctl_stat_interval),
1230                 .mode           = 0644,
1231                 .proc_handler   = &proc_dointvec_jiffies,
1232                 .strategy       = &sysctl_jiffies,
1233         },
1234 #endif
1235 #ifdef CONFIG_SECURITY
1236         {
1237                 .ctl_name       = CTL_UNNUMBERED,
1238                 .procname       = "mmap_min_addr",
1239                 .data           = &mmap_min_addr,
1240                 .maxlen         = sizeof(unsigned long),
1241                 .mode           = 0644,
1242                 .proc_handler   = &proc_doulongvec_minmax,
1243         },
1244 #endif
1245 #ifdef CONFIG_NUMA
1246         {
1247                 .ctl_name       = CTL_UNNUMBERED,
1248                 .procname       = "numa_zonelist_order",
1249                 .data           = &numa_zonelist_order,
1250                 .maxlen         = NUMA_ZONELIST_ORDER_LEN,
1251                 .mode           = 0644,
1252                 .proc_handler   = &numa_zonelist_order_handler,
1253                 .strategy       = &sysctl_string,
1254         },
1255 #endif
1256 #if (defined(CONFIG_X86_32) && !defined(CONFIG_UML))|| \
1257    (defined(CONFIG_SUPERH) && defined(CONFIG_VSYSCALL))
1258         {
1259                 .ctl_name       = VM_VDSO_ENABLED,
1260                 .procname       = "vdso_enabled",
1261                 .data           = &vdso_enabled,
1262                 .maxlen         = sizeof(vdso_enabled),
1263                 .mode           = 0644,
1264                 .proc_handler   = &proc_dointvec,
1265                 .strategy       = &sysctl_intvec,
1266                 .extra1         = &zero,
1267         },
1268 #endif
1269 #ifdef CONFIG_HIGHMEM
1270         {
1271                 .ctl_name       = CTL_UNNUMBERED,
1272                 .procname       = "highmem_is_dirtyable",
1273                 .data           = &vm_highmem_is_dirtyable,
1274                 .maxlen         = sizeof(vm_highmem_is_dirtyable),
1275                 .mode           = 0644,
1276                 .proc_handler   = &proc_dointvec_minmax,
1277                 .strategy       = &sysctl_intvec,
1278                 .extra1         = &zero,
1279                 .extra2         = &one,
1280         },
1281 #endif
1282 /*
1283  * NOTE: do not add new entries to this table unless you have read
1284  * Documentation/sysctl/ctl_unnumbered.txt
1285  */
1286         { .ctl_name = 0 }
1287 };
1288
1289 #if defined(CONFIG_BINFMT_MISC) || defined(CONFIG_BINFMT_MISC_MODULE)
1290 static struct ctl_table binfmt_misc_table[] = {
1291         { .ctl_name = 0 }
1292 };
1293 #endif
1294
1295 static struct ctl_table fs_table[] = {
1296         {
1297                 .ctl_name       = FS_NRINODE,
1298                 .procname       = "inode-nr",
1299                 .data           = &inodes_stat,
1300                 .maxlen         = 2*sizeof(int),
1301                 .mode           = 0444,
1302                 .proc_handler   = &proc_dointvec,
1303         },
1304         {
1305                 .ctl_name       = FS_STATINODE,
1306                 .procname       = "inode-state",
1307                 .data           = &inodes_stat,
1308                 .maxlen         = 7*sizeof(int),
1309                 .mode           = 0444,
1310                 .proc_handler   = &proc_dointvec,
1311         },
1312         {
1313                 .procname       = "file-nr",
1314                 .data           = &files_stat,
1315                 .maxlen         = 3*sizeof(int),
1316                 .mode           = 0444,
1317                 .proc_handler   = &proc_nr_files,
1318         },
1319         {
1320                 .ctl_name       = FS_MAXFILE,
1321                 .procname       = "file-max",
1322                 .data           = &files_stat.max_files,
1323                 .maxlen         = sizeof(int),
1324                 .mode           = 0644,
1325                 .proc_handler   = &proc_dointvec,
1326         },
1327         {
1328                 .ctl_name       = CTL_UNNUMBERED,
1329                 .procname       = "nr_open",
1330                 .data           = &sysctl_nr_open,
1331                 .maxlen         = sizeof(int),
1332                 .mode           = 0644,
1333                 .proc_handler   = &proc_dointvec_minmax,
1334                 .extra1         = &sysctl_nr_open_min,
1335                 .extra2         = &sysctl_nr_open_max,
1336         },
1337         {
1338                 .ctl_name       = FS_DENTRY,
1339                 .procname       = "dentry-state",
1340                 .data           = &dentry_stat,
1341                 .maxlen         = 6*sizeof(int),
1342                 .mode           = 0444,
1343                 .proc_handler   = &proc_dointvec,
1344         },
1345         {
1346                 .ctl_name       = FS_OVERFLOWUID,
1347                 .procname       = "overflowuid",
1348                 .data           = &fs_overflowuid,
1349                 .maxlen         = sizeof(int),
1350                 .mode           = 0644,
1351                 .proc_handler   = &proc_dointvec_minmax,
1352                 .strategy       = &sysctl_intvec,
1353                 .extra1         = &minolduid,
1354                 .extra2         = &maxolduid,
1355         },
1356         {
1357                 .ctl_name       = FS_OVERFLOWGID,
1358                 .procname       = "overflowgid",
1359                 .data           = &fs_overflowgid,
1360                 .maxlen         = sizeof(int),
1361                 .mode           = 0644,
1362                 .proc_handler   = &proc_dointvec_minmax,
1363                 .strategy       = &sysctl_intvec,
1364                 .extra1         = &minolduid,
1365                 .extra2         = &maxolduid,
1366         },
1367 #ifdef CONFIG_FILE_LOCKING
1368         {
1369                 .ctl_name       = FS_LEASES,
1370                 .procname       = "leases-enable",
1371                 .data           = &leases_enable,
1372                 .maxlen         = sizeof(int),
1373                 .mode           = 0644,
1374                 .proc_handler   = &proc_dointvec,
1375         },
1376 #endif
1377 #ifdef CONFIG_DNOTIFY
1378         {
1379                 .ctl_name       = FS_DIR_NOTIFY,
1380                 .procname       = "dir-notify-enable",
1381                 .data           = &dir_notify_enable,
1382                 .maxlen         = sizeof(int),
1383                 .mode           = 0644,
1384                 .proc_handler   = &proc_dointvec,
1385         },
1386 #endif
1387 #ifdef CONFIG_MMU
1388 #ifdef CONFIG_FILE_LOCKING
1389         {
1390                 .ctl_name       = FS_LEASE_TIME,
1391                 .procname       = "lease-break-time",
1392                 .data           = &lease_break_time,
1393                 .maxlen         = sizeof(int),
1394                 .mode           = 0644,
1395                 .proc_handler   = &proc_dointvec,
1396         },
1397 #endif
1398 #ifdef CONFIG_AIO
1399         {
1400                 .procname       = "aio-nr",
1401                 .data           = &aio_nr,
1402                 .maxlen         = sizeof(aio_nr),
1403                 .mode           = 0444,
1404                 .proc_handler   = &proc_doulongvec_minmax,
1405         },
1406         {
1407                 .procname       = "aio-max-nr",
1408                 .data           = &aio_max_nr,
1409                 .maxlen         = sizeof(aio_max_nr),
1410                 .mode           = 0644,
1411                 .proc_handler   = &proc_doulongvec_minmax,
1412         },
1413 #endif /* CONFIG_AIO */
1414 #ifdef CONFIG_INOTIFY_USER
1415         {
1416                 .ctl_name       = FS_INOTIFY,
1417                 .procname       = "inotify",
1418                 .mode           = 0555,
1419                 .child          = inotify_table,
1420         },
1421 #endif
1422 #ifdef CONFIG_EPOLL
1423         {
1424                 .procname       = "epoll",
1425                 .mode           = 0555,
1426                 .child          = epoll_table,
1427         },
1428 #endif
1429 #endif
1430         {
1431                 .ctl_name       = KERN_SETUID_DUMPABLE,
1432                 .procname       = "suid_dumpable",
1433                 .data           = &suid_dumpable,
1434                 .maxlen         = sizeof(int),
1435                 .mode           = 0644,
1436                 .proc_handler   = &proc_dointvec_minmax,
1437                 .strategy       = &sysctl_intvec,
1438                 .extra1         = &zero,
1439                 .extra2         = &two,
1440         },
1441 #if defined(CONFIG_BINFMT_MISC) || defined(CONFIG_BINFMT_MISC_MODULE)
1442         {
1443                 .ctl_name       = CTL_UNNUMBERED,
1444                 .procname       = "binfmt_misc",
1445                 .mode           = 0555,
1446                 .child          = binfmt_misc_table,
1447         },
1448 #endif
1449 /*
1450  * NOTE: do not add new entries to this table unless you have read
1451  * Documentation/sysctl/ctl_unnumbered.txt
1452  */
1453         { .ctl_name = 0 }
1454 };
1455
1456 static struct ctl_table debug_table[] = {
1457 #if defined(CONFIG_X86) || defined(CONFIG_PPC)
1458         {
1459                 .ctl_name       = CTL_UNNUMBERED,
1460                 .procname       = "exception-trace",
1461                 .data           = &show_unhandled_signals,
1462                 .maxlen         = sizeof(int),
1463                 .mode           = 0644,
1464                 .proc_handler   = proc_dointvec
1465         },
1466 #endif
1467         { .ctl_name = 0 }
1468 };
1469
1470 static struct ctl_table dev_table[] = {
1471         { .ctl_name = 0 }
1472 };
1473
1474 static DEFINE_SPINLOCK(sysctl_lock);
1475
1476 /* called under sysctl_lock */
1477 static int use_table(struct ctl_table_header *p)
1478 {
1479         if (unlikely(p->unregistering))
1480                 return 0;
1481         p->used++;
1482         return 1;
1483 }
1484
1485 /* called under sysctl_lock */
1486 static void unuse_table(struct ctl_table_header *p)
1487 {
1488         if (!--p->used)
1489                 if (unlikely(p->unregistering))
1490                         complete(p->unregistering);
1491 }
1492
1493 /* called under sysctl_lock, will reacquire if has to wait */
1494 static void start_unregistering(struct ctl_table_header *p)
1495 {
1496         /*
1497          * if p->used is 0, nobody will ever touch that entry again;
1498          * we'll eliminate all paths to it before dropping sysctl_lock
1499          */
1500         if (unlikely(p->used)) {
1501                 struct completion wait;
1502                 init_completion(&wait);
1503                 p->unregistering = &wait;
1504                 spin_unlock(&sysctl_lock);
1505                 wait_for_completion(&wait);
1506                 spin_lock(&sysctl_lock);
1507         } else {
1508                 /* anything non-NULL; we'll never dereference it */
1509                 p->unregistering = ERR_PTR(-EINVAL);
1510         }
1511         /*
1512          * do not remove from the list until nobody holds it; walking the
1513          * list in do_sysctl() relies on that.
1514          */
1515         list_del_init(&p->ctl_entry);
1516 }
1517
1518 void sysctl_head_get(struct ctl_table_header *head)
1519 {
1520         spin_lock(&sysctl_lock);
1521         head->count++;
1522         spin_unlock(&sysctl_lock);
1523 }
1524
1525 void sysctl_head_put(struct ctl_table_header *head)
1526 {
1527         spin_lock(&sysctl_lock);
1528         if (!--head->count)
1529                 kfree(head);
1530         spin_unlock(&sysctl_lock);
1531 }
1532
1533 struct ctl_table_header *sysctl_head_grab(struct ctl_table_header *head)
1534 {
1535         if (!head)
1536                 BUG();
1537         spin_lock(&sysctl_lock);
1538         if (!use_table(head))
1539                 head = ERR_PTR(-ENOENT);
1540         spin_unlock(&sysctl_lock);
1541         return head;
1542 }
1543
1544 void sysctl_head_finish(struct ctl_table_header *head)
1545 {
1546         if (!head)
1547                 return;
1548         spin_lock(&sysctl_lock);
1549         unuse_table(head);
1550         spin_unlock(&sysctl_lock);
1551 }
1552
1553 static struct ctl_table_set *
1554 lookup_header_set(struct ctl_table_root *root, struct nsproxy *namespaces)
1555 {
1556         struct ctl_table_set *set = &root->default_set;
1557         if (root->lookup)
1558                 set = root->lookup(root, namespaces);
1559         return set;
1560 }
1561
1562 static struct list_head *
1563 lookup_header_list(struct ctl_table_root *root, struct nsproxy *namespaces)
1564 {
1565         struct ctl_table_set *set = lookup_header_set(root, namespaces);
1566         return &set->list;
1567 }
1568
1569 struct ctl_table_header *__sysctl_head_next(struct nsproxy *namespaces,
1570                                             struct ctl_table_header *prev)
1571 {
1572         struct ctl_table_root *root;
1573         struct list_head *header_list;
1574         struct ctl_table_header *head;
1575         struct list_head *tmp;
1576
1577         spin_lock(&sysctl_lock);
1578         if (prev) {
1579                 head = prev;
1580                 tmp = &prev->ctl_entry;
1581                 unuse_table(prev);
1582                 goto next;
1583         }
1584         tmp = &root_table_header.ctl_entry;
1585         for (;;) {
1586                 head = list_entry(tmp, struct ctl_table_header, ctl_entry);
1587
1588                 if (!use_table(head))
1589                         goto next;
1590                 spin_unlock(&sysctl_lock);
1591                 return head;
1592         next:
1593                 root = head->root;
1594                 tmp = tmp->next;
1595                 header_list = lookup_header_list(root, namespaces);
1596                 if (tmp != header_list)
1597                         continue;
1598
1599                 do {
1600                         root = list_entry(root->root_list.next,
1601                                         struct ctl_table_root, root_list);
1602                         if (root == &sysctl_table_root)
1603                                 goto out;
1604                         header_list = lookup_header_list(root, namespaces);
1605                 } while (list_empty(header_list));
1606                 tmp = header_list->next;
1607         }
1608 out:
1609         spin_unlock(&sysctl_lock);
1610         return NULL;
1611 }
1612
1613 struct ctl_table_header *sysctl_head_next(struct ctl_table_header *prev)
1614 {
1615         return __sysctl_head_next(current->nsproxy, prev);
1616 }
1617
1618 void register_sysctl_root(struct ctl_table_root *root)
1619 {
1620         spin_lock(&sysctl_lock);
1621         list_add_tail(&root->root_list, &sysctl_table_root.root_list);
1622         spin_unlock(&sysctl_lock);
1623 }
1624
1625 #ifdef CONFIG_SYSCTL_SYSCALL
1626 /* Perform the actual read/write of a sysctl table entry. */
1627 static int do_sysctl_strategy(struct ctl_table_root *root,
1628                         struct ctl_table *table,
1629                         void __user *oldval, size_t __user *oldlenp,
1630                         void __user *newval, size_t newlen)
1631 {
1632         int op = 0, rc;
1633
1634         if (oldval)
1635                 op |= MAY_READ;
1636         if (newval)
1637                 op |= MAY_WRITE;
1638         if (sysctl_perm(root, table, op))
1639                 return -EPERM;
1640
1641         if (table->strategy) {
1642                 rc = table->strategy(table, oldval, oldlenp, newval, newlen);
1643                 if (rc < 0)
1644                         return rc;
1645                 if (rc > 0)
1646                         return 0;
1647         }
1648
1649         /* If there is no strategy routine, or if the strategy returns
1650          * zero, proceed with automatic r/w */
1651         if (table->data && table->maxlen) {
1652                 rc = sysctl_data(table, oldval, oldlenp, newval, newlen);
1653                 if (rc < 0)
1654                         return rc;
1655         }
1656         return 0;
1657 }
1658
1659 static int parse_table(int __user *name, int nlen,
1660                        void __user *oldval, size_t __user *oldlenp,
1661                        void __user *newval, size_t newlen,
1662                        struct ctl_table_root *root,
1663                        struct ctl_table *table)
1664 {
1665         int n;
1666 repeat:
1667         if (!nlen)
1668                 return -ENOTDIR;
1669         if (get_user(n, name))
1670                 return -EFAULT;
1671         for ( ; table->ctl_name || table->procname; table++) {
1672                 if (!table->ctl_name)
1673                         continue;
1674                 if (n == table->ctl_name) {
1675                         int error;
1676                         if (table->child) {
1677                                 if (sysctl_perm(root, table, MAY_EXEC))
1678                                         return -EPERM;
1679                                 name++;
1680                                 nlen--;
1681                                 table = table->child;
1682                                 goto repeat;
1683                         }
1684                         error = do_sysctl_strategy(root, table,
1685                                                    oldval, oldlenp,
1686                                                    newval, newlen);
1687                         return error;
1688                 }
1689         }
1690         return -ENOTDIR;
1691 }
1692
1693 int do_sysctl(int __user *name, int nlen, void __user *oldval, size_t __user *oldlenp,
1694                void __user *newval, size_t newlen)
1695 {
1696         struct ctl_table_header *head;
1697         int error = -ENOTDIR;
1698
1699         if (nlen <= 0 || nlen >= CTL_MAXNAME)
1700                 return -ENOTDIR;
1701         if (oldval) {
1702                 int old_len;
1703                 if (!oldlenp || get_user(old_len, oldlenp))
1704                         return -EFAULT;
1705         }
1706
1707         for (head = sysctl_head_next(NULL); head;
1708                         head = sysctl_head_next(head)) {
1709                 error = parse_table(name, nlen, oldval, oldlenp,
1710                                         newval, newlen,
1711                                         head->root, head->ctl_table);
1712                 if (error != -ENOTDIR) {
1713                         sysctl_head_finish(head);
1714                         break;
1715                 }
1716         }
1717         return error;
1718 }
1719
1720 SYSCALL_DEFINE1(sysctl, struct __sysctl_args __user *, args)
1721 {
1722         struct __sysctl_args tmp;
1723         int error;
1724
1725         if (copy_from_user(&tmp, args, sizeof(tmp)))
1726                 return -EFAULT;
1727
1728         error = deprecated_sysctl_warning(&tmp);
1729         if (error)
1730                 goto out;
1731
1732         lock_kernel();
1733         error = do_sysctl(tmp.name, tmp.nlen, tmp.oldval, tmp.oldlenp,
1734                           tmp.newval, tmp.newlen);
1735         unlock_kernel();
1736 out:
1737         return error;
1738 }
1739 #endif /* CONFIG_SYSCTL_SYSCALL */
1740
1741 /*
1742  * sysctl_perm does NOT grant the superuser all rights automatically, because
1743  * some sysctl variables are readonly even to root.
1744  */
1745
1746 static int test_perm(int mode, int op)
1747 {
1748         if (!current_euid())
1749                 mode >>= 6;
1750         else if (in_egroup_p(0))
1751                 mode >>= 3;
1752         if ((op & ~mode & (MAY_READ|MAY_WRITE|MAY_EXEC)) == 0)
1753                 return 0;
1754         return -EACCES;
1755 }
1756
1757 int sysctl_perm(struct ctl_table_root *root, struct ctl_table *table, int op)
1758 {
1759         int error;
1760         int mode;
1761
1762         error = security_sysctl(table, op & (MAY_READ | MAY_WRITE | MAY_EXEC));
1763         if (error)
1764                 return error;
1765
1766         if (root->permissions)
1767                 mode = root->permissions(root, current->nsproxy, table);
1768         else
1769                 mode = table->mode;
1770
1771         return test_perm(mode, op);
1772 }
1773
1774 static void sysctl_set_parent(struct ctl_table *parent, struct ctl_table *table)
1775 {
1776         for (; table->ctl_name || table->procname; table++) {
1777                 table->parent = parent;
1778                 if (table->child)
1779                         sysctl_set_parent(table, table->child);
1780         }
1781 }
1782
1783 static __init int sysctl_init(void)
1784 {
1785         sysctl_set_parent(NULL, root_table);
1786 #ifdef CONFIG_SYSCTL_SYSCALL_CHECK
1787         {
1788                 int err;
1789                 err = sysctl_check_table(current->nsproxy, root_table);
1790         }
1791 #endif
1792         return 0;
1793 }
1794
1795 core_initcall(sysctl_init);
1796
1797 static struct ctl_table *is_branch_in(struct ctl_table *branch,
1798                                       struct ctl_table *table)
1799 {
1800         struct ctl_table *p;
1801         const char *s = branch->procname;
1802
1803         /* branch should have named subdirectory as its first element */
1804         if (!s || !branch->child)
1805                 return NULL;
1806
1807         /* ... and nothing else */
1808         if (branch[1].procname || branch[1].ctl_name)
1809                 return NULL;
1810
1811         /* table should contain subdirectory with the same name */
1812         for (p = table; p->procname || p->ctl_name; p++) {
1813                 if (!p->child)
1814                         continue;
1815                 if (p->procname && strcmp(p->procname, s) == 0)
1816                         return p;
1817         }
1818         return NULL;
1819 }
1820
1821 /* see if attaching q to p would be an improvement */
1822 static void try_attach(struct ctl_table_header *p, struct ctl_table_header *q)
1823 {
1824         struct ctl_table *to = p->ctl_table, *by = q->ctl_table;
1825         struct ctl_table *next;
1826         int is_better = 0;
1827         int not_in_parent = !p->attached_by;
1828
1829         while ((next = is_branch_in(by, to)) != NULL) {
1830                 if (by == q->attached_by)
1831                         is_better = 1;
1832                 if (to == p->attached_by)
1833                         not_in_parent = 1;
1834                 by = by->child;
1835                 to = next->child;
1836         }
1837
1838         if (is_better && not_in_parent) {
1839                 q->attached_by = by;
1840                 q->attached_to = to;
1841                 q->parent = p;
1842         }
1843 }
1844
1845 /**
1846  * __register_sysctl_paths - register a sysctl hierarchy
1847  * @root: List of sysctl headers to register on
1848  * @namespaces: Data to compute which lists of sysctl entries are visible
1849  * @path: The path to the directory the sysctl table is in.
1850  * @table: the top-level table structure
1851  *
1852  * Register a sysctl table hierarchy. @table should be a filled in ctl_table
1853  * array. A completely 0 filled entry terminates the table.
1854  *
1855  * The members of the &struct ctl_table structure are used as follows:
1856  *
1857  * ctl_name - This is the numeric sysctl value used by sysctl(2). The number
1858  *            must be unique within that level of sysctl
1859  *
1860  * procname - the name of the sysctl file under /proc/sys. Set to %NULL to not
1861  *            enter a sysctl file
1862  *
1863  * data - a pointer to data for use by proc_handler
1864  *
1865  * maxlen - the maximum size in bytes of the data
1866  *
1867  * mode - the file permissions for the /proc/sys file, and for sysctl(2)
1868  *
1869  * child - a pointer to the child sysctl table if this entry is a directory, or
1870  *         %NULL.
1871  *
1872  * proc_handler - the text handler routine (described below)
1873  *
1874  * strategy - the strategy routine (described below)
1875  *
1876  * de - for internal use by the sysctl routines
1877  *
1878  * extra1, extra2 - extra pointers usable by the proc handler routines
1879  *
1880  * Leaf nodes in the sysctl tree will be represented by a single file
1881  * under /proc; non-leaf nodes will be represented by directories.
1882  *
1883  * sysctl(2) can automatically manage read and write requests through
1884  * the sysctl table.  The data and maxlen fields of the ctl_table
1885  * struct enable minimal validation of the values being written to be
1886  * performed, and the mode field allows minimal authentication.
1887  *
1888  * More sophisticated management can be enabled by the provision of a
1889  * strategy routine with the table entry.  This will be called before
1890  * any automatic read or write of the data is performed.
1891  *
1892  * The strategy routine may return
1893  *
1894  * < 0 - Error occurred (error is passed to user process)
1895  *
1896  * 0   - OK - proceed with automatic read or write.
1897  *
1898  * > 0 - OK - read or write has been done by the strategy routine, so
1899  *       return immediately.
1900  *
1901  * There must be a proc_handler routine for any terminal nodes
1902  * mirrored under /proc/sys (non-terminals are handled by a built-in
1903  * directory handler).  Several default handlers are available to
1904  * cover common cases -
1905  *
1906  * proc_dostring(), proc_dointvec(), proc_dointvec_jiffies(),
1907  * proc_dointvec_userhz_jiffies(), proc_dointvec_minmax(),
1908  * proc_doulongvec_ms_jiffies_minmax(), proc_doulongvec_minmax()
1909  *
1910  * It is the handler's job to read the input buffer from user memory
1911  * and process it. The handler should return 0 on success.
1912  *
1913  * This routine returns %NULL on a failure to register, and a pointer
1914  * to the table header on success.
1915  */
1916 struct ctl_table_header *__register_sysctl_paths(
1917         struct ctl_table_root *root,
1918         struct nsproxy *namespaces,
1919         const struct ctl_path *path, struct ctl_table *table)
1920 {
1921         struct ctl_table_header *header;
1922         struct ctl_table *new, **prevp;
1923         unsigned int n, npath;
1924         struct ctl_table_set *set;
1925
1926         /* Count the path components */
1927         for (npath = 0; path[npath].ctl_name || path[npath].procname; ++npath)
1928                 ;
1929
1930         /*
1931          * For each path component, allocate a 2-element ctl_table array.
1932          * The first array element will be filled with the sysctl entry
1933          * for this, the second will be the sentinel (ctl_name == 0).
1934          *
1935          * We allocate everything in one go so that we don't have to
1936          * worry about freeing additional memory in unregister_sysctl_table.
1937          */
1938         header = kzalloc(sizeof(struct ctl_table_header) +
1939                          (2 * npath * sizeof(struct ctl_table)), GFP_KERNEL);
1940         if (!header)
1941                 return NULL;
1942
1943         new = (struct ctl_table *) (header + 1);
1944
1945         /* Now connect the dots */
1946         prevp = &header->ctl_table;
1947         for (n = 0; n < npath; ++n, ++path) {
1948                 /* Copy the procname */
1949                 new->procname = path->procname;
1950                 new->ctl_name = path->ctl_name;
1951                 new->mode     = 0555;
1952
1953                 *prevp = new;
1954                 prevp = &new->child;
1955
1956                 new += 2;
1957         }
1958         *prevp = table;
1959         header->ctl_table_arg = table;
1960
1961         INIT_LIST_HEAD(&header->ctl_entry);
1962         header->used = 0;
1963         header->unregistering = NULL;
1964         header->root = root;
1965         sysctl_set_parent(NULL, header->ctl_table);
1966         header->count = 1;
1967 #ifdef CONFIG_SYSCTL_SYSCALL_CHECK
1968         if (sysctl_check_table(namespaces, header->ctl_table)) {
1969                 kfree(header);
1970                 return NULL;
1971         }
1972 #endif
1973         spin_lock(&sysctl_lock);
1974         header->set = lookup_header_set(root, namespaces);
1975         header->attached_by = header->ctl_table;
1976         header->attached_to = root_table;
1977         header->parent = &root_table_header;
1978         for (set = header->set; set; set = set->parent) {
1979                 struct ctl_table_header *p;
1980                 list_for_each_entry(p, &set->list, ctl_entry) {
1981                         if (p->unregistering)
1982                                 continue;
1983                         try_attach(p, header);
1984                 }
1985         }
1986         header->parent->count++;
1987         list_add_tail(&header->ctl_entry, &header->set->list);
1988         spin_unlock(&sysctl_lock);
1989
1990         return header;
1991 }
1992
1993 /**
1994  * register_sysctl_table_path - register a sysctl table hierarchy
1995  * @path: The path to the directory the sysctl table is in.
1996  * @table: the top-level table structure
1997  *
1998  * Register a sysctl table hierarchy. @table should be a filled in ctl_table
1999  * array. A completely 0 filled entry terminates the table.
2000  *
2001  * See __register_sysctl_paths for more details.
2002  */
2003 struct ctl_table_header *register_sysctl_paths(const struct ctl_path *path,
2004                                                 struct ctl_table *table)
2005 {
2006         return __register_sysctl_paths(&sysctl_table_root, current->nsproxy,
2007                                         path, table);
2008 }
2009
2010 /**
2011  * register_sysctl_table - register a sysctl table hierarchy
2012  * @table: the top-level table structure
2013  *
2014  * Register a sysctl table hierarchy. @table should be a filled in ctl_table
2015  * array. A completely 0 filled entry terminates the table.
2016  *
2017  * See register_sysctl_paths for more details.
2018  */
2019 struct ctl_table_header *register_sysctl_table(struct ctl_table *table)
2020 {
2021         static const struct ctl_path null_path[] = { {} };
2022
2023         return register_sysctl_paths(null_path, table);
2024 }
2025
2026 /**
2027  * unregister_sysctl_table - unregister a sysctl table hierarchy
2028  * @header: the header returned from register_sysctl_table
2029  *
2030  * Unregisters the sysctl table and all children. proc entries may not
2031  * actually be removed until they are no longer used by anyone.
2032  */
2033 void unregister_sysctl_table(struct ctl_table_header * header)
2034 {
2035         might_sleep();
2036
2037         if (header == NULL)
2038                 return;
2039
2040         spin_lock(&sysctl_lock);
2041         start_unregistering(header);
2042         if (!--header->parent->count) {
2043                 WARN_ON(1);
2044                 kfree(header->parent);
2045         }
2046         if (!--header->count)
2047                 kfree(header);
2048         spin_unlock(&sysctl_lock);
2049 }
2050
2051 int sysctl_is_seen(struct ctl_table_header *p)
2052 {
2053         struct ctl_table_set *set = p->set;
2054         int res;
2055         spin_lock(&sysctl_lock);
2056         if (p->unregistering)
2057                 res = 0;
2058         else if (!set->is_seen)
2059                 res = 1;
2060         else
2061                 res = set->is_seen(set);
2062         spin_unlock(&sysctl_lock);
2063         return res;
2064 }
2065
2066 void setup_sysctl_set(struct ctl_table_set *p,
2067         struct ctl_table_set *parent,
2068         int (*is_seen)(struct ctl_table_set *))
2069 {
2070         INIT_LIST_HEAD(&p->list);
2071         p->parent = parent ? parent : &sysctl_table_root.default_set;
2072         p->is_seen = is_seen;
2073 }
2074
2075 #else /* !CONFIG_SYSCTL */
2076 struct ctl_table_header *register_sysctl_table(struct ctl_table * table)
2077 {
2078         return NULL;
2079 }
2080
2081 struct ctl_table_header *register_sysctl_paths(const struct ctl_path *path,
2082                                                     struct ctl_table *table)
2083 {
2084         return NULL;
2085 }
2086
2087 void unregister_sysctl_table(struct ctl_table_header * table)
2088 {
2089 }
2090
2091 void setup_sysctl_set(struct ctl_table_set *p,
2092         struct ctl_table_set *parent,
2093         int (*is_seen)(struct ctl_table_set *))
2094 {
2095 }
2096
2097 void sysctl_head_put(struct ctl_table_header *head)
2098 {
2099 }
2100
2101 #endif /* CONFIG_SYSCTL */
2102
2103 /*
2104  * /proc/sys support
2105  */
2106
2107 #ifdef CONFIG_PROC_SYSCTL
2108
2109 static int _proc_do_string(void* data, int maxlen, int write,
2110                            struct file *filp, void __user *buffer,
2111                            size_t *lenp, loff_t *ppos)
2112 {
2113         size_t len;
2114         char __user *p;
2115         char c;
2116
2117         if (!data || !maxlen || !*lenp) {
2118                 *lenp = 0;
2119                 return 0;
2120         }
2121
2122         if (write) {
2123                 len = 0;
2124                 p = buffer;
2125                 while (len < *lenp) {
2126                         if (get_user(c, p++))
2127                                 return -EFAULT;
2128                         if (c == 0 || c == '\n')
2129                                 break;
2130                         len++;
2131                 }
2132                 if (len >= maxlen)
2133                         len = maxlen-1;
2134                 if(copy_from_user(data, buffer, len))
2135                         return -EFAULT;
2136                 ((char *) data)[len] = 0;
2137                 *ppos += *lenp;
2138         } else {
2139                 len = strlen(data);
2140                 if (len > maxlen)
2141                         len = maxlen;
2142
2143                 if (*ppos > len) {
2144                         *lenp = 0;
2145                         return 0;
2146                 }
2147
2148                 data += *ppos;
2149                 len  -= *ppos;
2150
2151                 if (len > *lenp)
2152                         len = *lenp;
2153                 if (len)
2154                         if(copy_to_user(buffer, data, len))
2155                                 return -EFAULT;
2156                 if (len < *lenp) {
2157                         if(put_user('\n', ((char __user *) buffer) + len))
2158                                 return -EFAULT;
2159                         len++;
2160                 }
2161                 *lenp = len;
2162                 *ppos += len;
2163         }
2164         return 0;
2165 }
2166
2167 /**
2168  * proc_dostring - read a string sysctl
2169  * @table: the sysctl table
2170  * @write: %TRUE if this is a write to the sysctl file
2171  * @filp: the file structure
2172  * @buffer: the user buffer
2173  * @lenp: the size of the user buffer
2174  * @ppos: file position
2175  *
2176  * Reads/writes a string from/to the user buffer. If the kernel
2177  * buffer provided is not large enough to hold the string, the
2178  * string is truncated. The copied string is %NULL-terminated.
2179  * If the string is being read by the user process, it is copied
2180  * and a newline '\n' is added. It is truncated if the buffer is
2181  * not large enough.
2182  *
2183  * Returns 0 on success.
2184  */
2185 int proc_dostring(struct ctl_table *table, int write, struct file *filp,
2186                   void __user *buffer, size_t *lenp, loff_t *ppos)
2187 {
2188         return _proc_do_string(table->data, table->maxlen, write, filp,
2189                                buffer, lenp, ppos);
2190 }
2191
2192
2193 static int do_proc_dointvec_conv(int *negp, unsigned long *lvalp,
2194                                  int *valp,
2195                                  int write, void *data)
2196 {
2197         if (write) {
2198                 *valp = *negp ? -*lvalp : *lvalp;
2199         } else {
2200                 int val = *valp;
2201                 if (val < 0) {
2202                         *negp = -1;
2203                         *lvalp = (unsigned long)-val;
2204                 } else {
2205                         *negp = 0;
2206                         *lvalp = (unsigned long)val;
2207                 }
2208         }
2209         return 0;
2210 }
2211
2212 static int __do_proc_dointvec(void *tbl_data, struct ctl_table *table,
2213                   int write, struct file *filp, void __user *buffer,
2214                   size_t *lenp, loff_t *ppos,
2215                   int (*conv)(int *negp, unsigned long *lvalp, int *valp,
2216                               int write, void *data),
2217                   void *data)
2218 {
2219 #define TMPBUFLEN 21
2220         int *i, vleft, first=1, neg, val;
2221         unsigned long lval;
2222         size_t left, len;
2223
2224         char buf[TMPBUFLEN], *p;
2225         char __user *s = buffer;
2226
2227         if (!tbl_data || !table->maxlen || !*lenp ||
2228             (*ppos && !write)) {
2229                 *lenp = 0;
2230                 return 0;
2231         }
2232
2233         i = (int *) tbl_data;
2234         vleft = table->maxlen / sizeof(*i);
2235         left = *lenp;
2236
2237         if (!conv)
2238                 conv = do_proc_dointvec_conv;
2239
2240         for (; left && vleft--; i++, first=0) {
2241                 if (write) {
2242                         while (left) {
2243                                 char c;
2244                                 if (get_user(c, s))
2245                                         return -EFAULT;
2246                                 if (!isspace(c))
2247                                         break;
2248                                 left--;
2249                                 s++;
2250                         }
2251                         if (!left)
2252                                 break;
2253                         neg = 0;
2254                         len = left;
2255                         if (len > sizeof(buf) - 1)
2256                                 len = sizeof(buf) - 1;
2257                         if (copy_from_user(buf, s, len))
2258                                 return -EFAULT;
2259                         buf[len] = 0;
2260                         p = buf;
2261                         if (*p == '-' && left > 1) {
2262                                 neg = 1;
2263                                 p++;
2264                         }
2265                         if (*p < '0' || *p > '9')
2266                                 break;
2267
2268                         lval = simple_strtoul(p, &p, 0);
2269
2270                         len = p-buf;
2271                         if ((len < left) && *p && !isspace(*p))
2272                                 break;
2273                         if (neg)
2274                                 val = -val;
2275                         s += len;
2276                         left -= len;
2277
2278                         if (conv(&neg, &lval, i, 1, data))
2279                                 break;
2280                 } else {
2281                         p = buf;
2282                         if (!first)
2283                                 *p++ = '\t';
2284
2285                         if (conv(&neg, &lval, i, 0, data))
2286                                 break;
2287
2288                         sprintf(p, "%s%lu", neg ? "-" : "", lval);
2289                         len = strlen(buf);
2290                         if (len > left)
2291                                 len = left;
2292                         if(copy_to_user(s, buf, len))
2293                                 return -EFAULT;
2294                         left -= len;
2295                         s += len;
2296                 }
2297         }
2298
2299         if (!write && !first && left) {
2300                 if(put_user('\n', s))
2301                         return -EFAULT;
2302                 left--, s++;
2303         }
2304         if (write) {
2305                 while (left) {
2306                         char c;
2307                         if (get_user(c, s++))
2308                                 return -EFAULT;
2309                         if (!isspace(c))
2310                                 break;
2311                         left--;
2312                 }
2313         }
2314         if (write && first)
2315                 return -EINVAL;
2316         *lenp -= left;
2317         *ppos += *lenp;
2318         return 0;
2319 #undef TMPBUFLEN
2320 }
2321
2322 static int do_proc_dointvec(struct ctl_table *table, int write, struct file *filp,
2323                   void __user *buffer, size_t *lenp, loff_t *ppos,
2324                   int (*conv)(int *negp, unsigned long *lvalp, int *valp,
2325                               int write, void *data),
2326                   void *data)
2327 {
2328         return __do_proc_dointvec(table->data, table, write, filp,
2329                         buffer, lenp, ppos, conv, data);
2330 }
2331
2332 /**
2333  * proc_dointvec - read a vector of integers
2334  * @table: the sysctl table
2335  * @write: %TRUE if this is a write to the sysctl file
2336  * @filp: the file structure
2337  * @buffer: the user buffer
2338  * @lenp: the size of the user buffer
2339  * @ppos: file position
2340  *
2341  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2342  * values from/to the user buffer, treated as an ASCII string.
2343  *
2344  * Returns 0 on success.
2345  */
2346 int proc_dointvec(struct ctl_table *table, int write, struct file *filp,
2347                      void __user *buffer, size_t *lenp, loff_t *ppos)
2348 {
2349     return do_proc_dointvec(table,write,filp,buffer,lenp,ppos,
2350                             NULL,NULL);
2351 }
2352
2353 /*
2354  * Taint values can only be increased
2355  * This means we can safely use a temporary.
2356  */
2357 static int proc_taint(struct ctl_table *table, int write, struct file *filp,
2358                                void __user *buffer, size_t *lenp, loff_t *ppos)
2359 {
2360         struct ctl_table t;
2361         unsigned long tmptaint = get_taint();
2362         int err;
2363
2364         if (write && !capable(CAP_SYS_ADMIN))
2365                 return -EPERM;
2366
2367         t = *table;
2368         t.data = &tmptaint;
2369         err = proc_doulongvec_minmax(&t, write, filp, buffer, lenp, ppos);
2370         if (err < 0)
2371                 return err;
2372
2373         if (write) {
2374                 /*
2375                  * Poor man's atomic or. Not worth adding a primitive
2376                  * to everyone's atomic.h for this
2377                  */
2378                 int i;
2379                 for (i = 0; i < BITS_PER_LONG && tmptaint >> i; i++) {
2380                         if ((tmptaint >> i) & 1)
2381                                 add_taint(i);
2382                 }
2383         }
2384
2385         return err;
2386 }
2387
2388 struct do_proc_dointvec_minmax_conv_param {
2389         int *min;
2390         int *max;
2391 };
2392
2393 static int do_proc_dointvec_minmax_conv(int *negp, unsigned long *lvalp,
2394                                         int *valp,
2395                                         int write, void *data)
2396 {
2397         struct do_proc_dointvec_minmax_conv_param *param = data;
2398         if (write) {
2399                 int val = *negp ? -*lvalp : *lvalp;
2400                 if ((param->min && *param->min > val) ||
2401                     (param->max && *param->max < val))
2402                         return -EINVAL;
2403                 *valp = val;
2404         } else {
2405                 int val = *valp;
2406                 if (val < 0) {
2407                         *negp = -1;
2408                         *lvalp = (unsigned long)-val;
2409                 } else {
2410                         *negp = 0;
2411                         *lvalp = (unsigned long)val;
2412                 }
2413         }
2414         return 0;
2415 }
2416
2417 /**
2418  * proc_dointvec_minmax - read a vector of integers with min/max values
2419  * @table: the sysctl table
2420  * @write: %TRUE if this is a write to the sysctl file
2421  * @filp: the file structure
2422  * @buffer: the user buffer
2423  * @lenp: the size of the user buffer
2424  * @ppos: file position
2425  *
2426  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2427  * values from/to the user buffer, treated as an ASCII string.
2428  *
2429  * This routine will ensure the values are within the range specified by
2430  * table->extra1 (min) and table->extra2 (max).
2431  *
2432  * Returns 0 on success.
2433  */
2434 int proc_dointvec_minmax(struct ctl_table *table, int write, struct file *filp,
2435                   void __user *buffer, size_t *lenp, loff_t *ppos)
2436 {
2437         struct do_proc_dointvec_minmax_conv_param param = {
2438                 .min = (int *) table->extra1,
2439                 .max = (int *) table->extra2,
2440         };
2441         return do_proc_dointvec(table, write, filp, buffer, lenp, ppos,
2442                                 do_proc_dointvec_minmax_conv, &param);
2443 }
2444
2445 static int __do_proc_doulongvec_minmax(void *data, struct ctl_table *table, int write,
2446                                      struct file *filp,
2447                                      void __user *buffer,
2448                                      size_t *lenp, loff_t *ppos,
2449                                      unsigned long convmul,
2450                                      unsigned long convdiv)
2451 {
2452 #define TMPBUFLEN 21
2453         unsigned long *i, *min, *max, val;
2454         int vleft, first=1, neg;
2455         size_t len, left;
2456         char buf[TMPBUFLEN], *p;
2457         char __user *s = buffer;
2458
2459         if (!data || !table->maxlen || !*lenp ||
2460             (*ppos && !write)) {
2461                 *lenp = 0;
2462                 return 0;
2463         }
2464
2465         i = (unsigned long *) data;
2466         min = (unsigned long *) table->extra1;
2467         max = (unsigned long *) table->extra2;
2468         vleft = table->maxlen / sizeof(unsigned long);
2469         left = *lenp;
2470
2471         for (; left && vleft--; i++, min++, max++, first=0) {
2472                 if (write) {
2473                         while (left) {
2474                                 char c;
2475                                 if (get_user(c, s))
2476                                         return -EFAULT;
2477                                 if (!isspace(c))
2478                                         break;
2479                                 left--;
2480                                 s++;
2481                         }
2482                         if (!left)
2483                                 break;
2484                         neg = 0;
2485                         len = left;
2486                         if (len > TMPBUFLEN-1)
2487                                 len = TMPBUFLEN-1;
2488                         if (copy_from_user(buf, s, len))
2489                                 return -EFAULT;
2490                         buf[len] = 0;
2491                         p = buf;
2492                         if (*p == '-' && left > 1) {
2493                                 neg = 1;
2494                                 p++;
2495                         }
2496                         if (*p < '0' || *p > '9')
2497                                 break;
2498                         val = simple_strtoul(p, &p, 0) * convmul / convdiv ;
2499                         len = p-buf;
2500                         if ((len < left) && *p && !isspace(*p))
2501                                 break;
2502                         if (neg)
2503                                 val = -val;
2504                         s += len;
2505                         left -= len;
2506
2507                         if(neg)
2508                                 continue;
2509                         if ((min && val < *min) || (max && val > *max))
2510                                 continue;
2511                         *i = val;
2512                 } else {
2513                         p = buf;
2514                         if (!first)
2515                                 *p++ = '\t';
2516                         sprintf(p, "%lu", convdiv * (*i) / convmul);
2517                         len = strlen(buf);
2518                         if (len > left)
2519                                 len = left;
2520                         if(copy_to_user(s, buf, len))
2521                                 return -EFAULT;
2522                         left -= len;
2523                         s += len;
2524                 }
2525         }
2526
2527         if (!write && !first && left) {
2528                 if(put_user('\n', s))
2529                         return -EFAULT;
2530                 left--, s++;
2531         }
2532         if (write) {
2533                 while (left) {
2534                         char c;
2535                         if (get_user(c, s++))
2536                                 return -EFAULT;
2537                         if (!isspace(c))
2538                                 break;
2539                         left--;
2540                 }
2541         }
2542         if (write && first)
2543                 return -EINVAL;
2544         *lenp -= left;
2545         *ppos += *lenp;
2546         return 0;
2547 #undef TMPBUFLEN
2548 }
2549
2550 static int do_proc_doulongvec_minmax(struct ctl_table *table, int write,
2551                                      struct file *filp,
2552                                      void __user *buffer,
2553                                      size_t *lenp, loff_t *ppos,
2554                                      unsigned long convmul,
2555                                      unsigned long convdiv)
2556 {
2557         return __do_proc_doulongvec_minmax(table->data, table, write,
2558                         filp, buffer, lenp, ppos, convmul, convdiv);
2559 }
2560
2561 /**
2562  * proc_doulongvec_minmax - read a vector of long integers with min/max values
2563  * @table: the sysctl table
2564  * @write: %TRUE if this is a write to the sysctl file
2565  * @filp: the file structure
2566  * @buffer: the user buffer
2567  * @lenp: the size of the user buffer
2568  * @ppos: file position
2569  *
2570  * Reads/writes up to table->maxlen/sizeof(unsigned long) unsigned long
2571  * values from/to the user buffer, treated as an ASCII string.
2572  *
2573  * This routine will ensure the values are within the range specified by
2574  * table->extra1 (min) and table->extra2 (max).
2575  *
2576  * Returns 0 on success.
2577  */
2578 int proc_doulongvec_minmax(struct ctl_table *table, int write, struct file *filp,
2579                            void __user *buffer, size_t *lenp, loff_t *ppos)
2580 {
2581     return do_proc_doulongvec_minmax(table, write, filp, buffer, lenp, ppos, 1l, 1l);
2582 }
2583
2584 /**
2585  * proc_doulongvec_ms_jiffies_minmax - read a vector of millisecond values with min/max values
2586  * @table: the sysctl table
2587  * @write: %TRUE if this is a write to the sysctl file
2588  * @filp: the file structure
2589  * @buffer: the user buffer
2590  * @lenp: the size of the user buffer
2591  * @ppos: file position
2592  *
2593  * Reads/writes up to table->maxlen/sizeof(unsigned long) unsigned long
2594  * values from/to the user buffer, treated as an ASCII string. The values
2595  * are treated as milliseconds, and converted to jiffies when they are stored.
2596  *
2597  * This routine will ensure the values are within the range specified by
2598  * table->extra1 (min) and table->extra2 (max).
2599  *
2600  * Returns 0 on success.
2601  */
2602 int proc_doulongvec_ms_jiffies_minmax(struct ctl_table *table, int write,
2603                                       struct file *filp,
2604                                       void __user *buffer,
2605                                       size_t *lenp, loff_t *ppos)
2606 {
2607     return do_proc_doulongvec_minmax(table, write, filp, buffer,
2608                                      lenp, ppos, HZ, 1000l);
2609 }
2610
2611
2612 static int do_proc_dointvec_jiffies_conv(int *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 = -1;
2625                         lval = (unsigned long)-val;
2626                 } else {
2627                         *negp = 0;
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(int *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 = -1;
2648                         lval = (unsigned long)-val;
2649                 } else {
2650                         *negp = 0;
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(int *negp, unsigned long *lvalp,
2659                                             int *valp,
2660                                             int write, void *data)
2661 {
2662         if (write) {
2663                 *valp = msecs_to_jiffies(*negp ? -*lvalp : *lvalp);
2664         } else {
2665                 int val = *valp;
2666                 unsigned long lval;
2667                 if (val < 0) {
2668                         *negp = -1;
2669                         lval = (unsigned long)-val;
2670                 } else {
2671                         *negp = 0;
2672                         lval = (unsigned long)val;
2673                 }
2674                 *lvalp = jiffies_to_msecs(lval);
2675         }
2676         return 0;
2677 }
2678
2679 /**
2680  * proc_dointvec_jiffies - read a vector of integers as seconds
2681  * @table: the sysctl table
2682  * @write: %TRUE if this is a write to the sysctl file
2683  * @filp: the file structure
2684  * @buffer: the user buffer
2685  * @lenp: the size of the user buffer
2686  * @ppos: file position
2687  *
2688  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2689  * values from/to the user buffer, treated as an ASCII string.
2690  * The values read are assumed to be in seconds, and are converted into
2691  * jiffies.
2692  *
2693  * Returns 0 on success.
2694  */
2695 int proc_dointvec_jiffies(struct ctl_table *table, int write, struct file *filp,
2696                           void __user *buffer, size_t *lenp, loff_t *ppos)
2697 {
2698     return do_proc_dointvec(table,write,filp,buffer,lenp,ppos,
2699                             do_proc_dointvec_jiffies_conv,NULL);
2700 }
2701
2702 /**
2703  * proc_dointvec_userhz_jiffies - read a vector of integers as 1/USER_HZ seconds
2704  * @table: the sysctl table
2705  * @write: %TRUE if this is a write to the sysctl file
2706  * @filp: the file structure
2707  * @buffer: the user buffer
2708  * @lenp: the size of the user buffer
2709  * @ppos: pointer to the file position
2710  *
2711  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2712  * values from/to the user buffer, treated as an ASCII string.
2713  * The values read are assumed to be in 1/USER_HZ seconds, and
2714  * are converted into jiffies.
2715  *
2716  * Returns 0 on success.
2717  */
2718 int proc_dointvec_userhz_jiffies(struct ctl_table *table, int write, struct file *filp,
2719                                  void __user *buffer, size_t *lenp, loff_t *ppos)
2720 {
2721     return do_proc_dointvec(table,write,filp,buffer,lenp,ppos,
2722                             do_proc_dointvec_userhz_jiffies_conv,NULL);
2723 }
2724
2725 /**
2726  * proc_dointvec_ms_jiffies - read a vector of integers as 1 milliseconds
2727  * @table: the sysctl table
2728  * @write: %TRUE if this is a write to the sysctl file
2729  * @filp: the file structure
2730  * @buffer: the user buffer
2731  * @lenp: the size of the user buffer
2732  * @ppos: file position
2733  * @ppos: the current position in the file
2734  *
2735  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2736  * values from/to the user buffer, treated as an ASCII string.
2737  * The values read are assumed to be in 1/1000 seconds, and
2738  * are converted into jiffies.
2739  *
2740  * Returns 0 on success.
2741  */
2742 int proc_dointvec_ms_jiffies(struct ctl_table *table, int write, struct file *filp,
2743                              void __user *buffer, size_t *lenp, loff_t *ppos)
2744 {
2745         return do_proc_dointvec(table, write, filp, buffer, lenp, ppos,
2746                                 do_proc_dointvec_ms_jiffies_conv, NULL);
2747 }
2748
2749 static int proc_do_cad_pid(struct ctl_table *table, int write, struct file *filp,
2750                            void __user *buffer, size_t *lenp, loff_t *ppos)
2751 {
2752         struct pid *new_pid;
2753         pid_t tmp;
2754         int r;
2755
2756         tmp = pid_vnr(cad_pid);
2757
2758         r = __do_proc_dointvec(&tmp, table, write, filp, buffer,
2759                                lenp, ppos, NULL, NULL);
2760         if (r || !write)
2761                 return r;
2762
2763         new_pid = find_get_pid(tmp);
2764         if (!new_pid)
2765                 return -ESRCH;
2766
2767         put_pid(xchg(&cad_pid, new_pid));
2768         return 0;
2769 }
2770
2771 #else /* CONFIG_PROC_FS */
2772
2773 int proc_dostring(struct ctl_table *table, int write, struct file *filp,
2774                   void __user *buffer, size_t *lenp, loff_t *ppos)
2775 {
2776         return -ENOSYS;
2777 }
2778
2779 int proc_dointvec(struct ctl_table *table, int write, struct file *filp,
2780                   void __user *buffer, size_t *lenp, loff_t *ppos)
2781 {
2782         return -ENOSYS;
2783 }
2784
2785 int proc_dointvec_minmax(struct ctl_table *table, int write, struct file *filp,
2786                     void __user *buffer, size_t *lenp, loff_t *ppos)
2787 {
2788         return -ENOSYS;
2789 }
2790
2791 int proc_dointvec_jiffies(struct ctl_table *table, int write, struct file *filp,
2792                     void __user *buffer, size_t *lenp, loff_t *ppos)
2793 {
2794         return -ENOSYS;
2795 }
2796
2797 int proc_dointvec_userhz_jiffies(struct ctl_table *table, int write, struct file *filp,
2798                     void __user *buffer, size_t *lenp, loff_t *ppos)
2799 {
2800         return -ENOSYS;
2801 }
2802
2803 int proc_dointvec_ms_jiffies(struct ctl_table *table, int write, struct file *filp,
2804                              void __user *buffer, size_t *lenp, loff_t *ppos)
2805 {
2806         return -ENOSYS;
2807 }
2808
2809 int proc_doulongvec_minmax(struct ctl_table *table, int write, struct file *filp,
2810                     void __user *buffer, size_t *lenp, loff_t *ppos)
2811 {
2812         return -ENOSYS;
2813 }
2814
2815 int proc_doulongvec_ms_jiffies_minmax(struct ctl_table *table, int write,
2816                                       struct file *filp,
2817                                       void __user *buffer,
2818                                       size_t *lenp, loff_t *ppos)
2819 {
2820     return -ENOSYS;
2821 }
2822
2823
2824 #endif /* CONFIG_PROC_FS */
2825
2826
2827 #ifdef CONFIG_SYSCTL_SYSCALL
2828 /*
2829  * General sysctl support routines
2830  */
2831
2832 /* The generic sysctl data routine (used if no strategy routine supplied) */
2833 int sysctl_data(struct ctl_table *table,
2834                 void __user *oldval, size_t __user *oldlenp,
2835                 void __user *newval, size_t newlen)
2836 {
2837         size_t len;
2838
2839         /* Get out of I don't have a variable */
2840         if (!table->data || !table->maxlen)
2841                 return -ENOTDIR;
2842
2843         if (oldval && oldlenp) {
2844                 if (get_user(len, oldlenp))
2845                         return -EFAULT;
2846                 if (len) {
2847                         if (len > table->maxlen)
2848                                 len = table->maxlen;
2849                         if (copy_to_user(oldval, table->data, len))
2850                                 return -EFAULT;
2851                         if (put_user(len, oldlenp))
2852                                 return -EFAULT;
2853                 }
2854         }
2855
2856         if (newval && newlen) {
2857                 if (newlen > table->maxlen)
2858                         newlen = table->maxlen;
2859
2860                 if (copy_from_user(table->data, newval, newlen))
2861                         return -EFAULT;
2862         }
2863         return 1;
2864 }
2865
2866 /* The generic string strategy routine: */
2867 int sysctl_string(struct ctl_table *table,
2868                   void __user *oldval, size_t __user *oldlenp,
2869                   void __user *newval, size_t newlen)
2870 {
2871         if (!table->data || !table->maxlen)
2872                 return -ENOTDIR;
2873
2874         if (oldval && oldlenp) {
2875                 size_t bufsize;
2876                 if (get_user(bufsize, oldlenp))
2877                         return -EFAULT;
2878                 if (bufsize) {
2879                         size_t len = strlen(table->data), copied;
2880
2881                         /* This shouldn't trigger for a well-formed sysctl */
2882                         if (len > table->maxlen)
2883                                 len = table->maxlen;
2884
2885                         /* Copy up to a max of bufsize-1 bytes of the string */
2886                         copied = (len >= bufsize) ? bufsize - 1 : len;
2887
2888                         if (copy_to_user(oldval, table->data, copied) ||
2889                             put_user(0, (char __user *)(oldval + copied)))
2890                                 return -EFAULT;
2891                         if (put_user(len, oldlenp))
2892                                 return -EFAULT;
2893                 }
2894         }
2895         if (newval && newlen) {
2896                 size_t len = newlen;
2897                 if (len > table->maxlen)
2898                         len = table->maxlen;
2899                 if(copy_from_user(table->data, newval, len))
2900                         return -EFAULT;
2901                 if (len == table->maxlen)
2902                         len--;
2903                 ((char *) table->data)[len] = 0;
2904         }
2905         return 1;
2906 }
2907
2908 /*
2909  * This function makes sure that all of the integers in the vector
2910  * are between the minimum and maximum values given in the arrays
2911  * table->extra1 and table->extra2, respectively.
2912  */
2913 int sysctl_intvec(struct ctl_table *table,
2914                 void __user *oldval, size_t __user *oldlenp,
2915                 void __user *newval, size_t newlen)
2916 {
2917
2918         if (newval && newlen) {
2919                 int __user *vec = (int __user *) newval;
2920                 int *min = (int *) table->extra1;
2921                 int *max = (int *) table->extra2;
2922                 size_t length;
2923                 int i;
2924
2925                 if (newlen % sizeof(int) != 0)
2926                         return -EINVAL;
2927
2928                 if (!table->extra1 && !table->extra2)
2929                         return 0;
2930
2931                 if (newlen > table->maxlen)
2932                         newlen = table->maxlen;
2933                 length = newlen / sizeof(int);
2934
2935                 for (i = 0; i < length; i++) {
2936                         int value;
2937                         if (get_user(value, vec + i))
2938                                 return -EFAULT;
2939                         if (min && value < min[i])
2940                                 return -EINVAL;
2941                         if (max && value > max[i])
2942                                 return -EINVAL;
2943                 }
2944         }
2945         return 0;
2946 }
2947
2948 /* Strategy function to convert jiffies to seconds */
2949 int sysctl_jiffies(struct ctl_table *table,
2950                 void __user *oldval, size_t __user *oldlenp,
2951                 void __user *newval, size_t newlen)
2952 {
2953         if (oldval && oldlenp) {
2954                 size_t olen;
2955
2956                 if (get_user(olen, oldlenp))
2957                         return -EFAULT;
2958                 if (olen) {
2959                         int val;
2960
2961                         if (olen < sizeof(int))
2962                                 return -EINVAL;
2963
2964                         val = *(int *)(table->data) / HZ;
2965                         if (put_user(val, (int __user *)oldval))
2966                                 return -EFAULT;
2967                         if (put_user(sizeof(int), oldlenp))
2968                                 return -EFAULT;
2969                 }
2970         }
2971         if (newval && newlen) {
2972                 int new;
2973                 if (newlen != sizeof(int))
2974                         return -EINVAL;
2975                 if (get_user(new, (int __user *)newval))
2976                         return -EFAULT;
2977                 *(int *)(table->data) = new*HZ;
2978         }
2979         return 1;
2980 }
2981
2982 /* Strategy function to convert jiffies to seconds */
2983 int sysctl_ms_jiffies(struct ctl_table *table,
2984                 void __user *oldval, size_t __user *oldlenp,
2985                 void __user *newval, size_t newlen)
2986 {
2987         if (oldval && oldlenp) {
2988                 size_t olen;
2989
2990                 if (get_user(olen, oldlenp))
2991                         return -EFAULT;
2992                 if (olen) {
2993                         int val;
2994
2995                         if (olen < sizeof(int))
2996                                 return -EINVAL;
2997
2998                         val = jiffies_to_msecs(*(int *)(table->data));
2999                         if (put_user(val, (int __user *)oldval))
3000                                 return -EFAULT;
3001                         if (put_user(sizeof(int), oldlenp))
3002                                 return -EFAULT;
3003                 }
3004         }
3005         if (newval && newlen) {
3006                 int new;
3007                 if (newlen != sizeof(int))
3008                         return -EINVAL;
3009                 if (get_user(new, (int __user *)newval))
3010                         return -EFAULT;
3011                 *(int *)(table->data) = msecs_to_jiffies(new);
3012         }
3013         return 1;
3014 }
3015
3016
3017
3018 #else /* CONFIG_SYSCTL_SYSCALL */
3019
3020
3021 SYSCALL_DEFINE1(sysctl, struct __sysctl_args __user *, args)
3022 {
3023         struct __sysctl_args tmp;
3024         int error;
3025
3026         if (copy_from_user(&tmp, args, sizeof(tmp)))
3027                 return -EFAULT;
3028
3029         error = deprecated_sysctl_warning(&tmp);
3030
3031         /* If no error reading the parameters then just -ENOSYS ... */
3032         if (!error)
3033                 error = -ENOSYS;
3034
3035         return error;
3036 }
3037
3038 int sysctl_data(struct ctl_table *table,
3039                   void __user *oldval, size_t __user *oldlenp,
3040                   void __user *newval, size_t newlen)
3041 {
3042         return -ENOSYS;
3043 }
3044
3045 int sysctl_string(struct ctl_table *table,
3046                   void __user *oldval, size_t __user *oldlenp,
3047                   void __user *newval, size_t newlen)
3048 {
3049         return -ENOSYS;
3050 }
3051
3052 int sysctl_intvec(struct ctl_table *table,
3053                 void __user *oldval, size_t __user *oldlenp,
3054                 void __user *newval, size_t newlen)
3055 {
3056         return -ENOSYS;
3057 }
3058
3059 int sysctl_jiffies(struct ctl_table *table,
3060                 void __user *oldval, size_t __user *oldlenp,
3061                 void __user *newval, size_t newlen)
3062 {
3063         return -ENOSYS;
3064 }
3065
3066 int sysctl_ms_jiffies(struct ctl_table *table,
3067                 void __user *oldval, size_t __user *oldlenp,
3068                 void __user *newval, size_t newlen)
3069 {
3070         return -ENOSYS;
3071 }
3072
3073 #endif /* CONFIG_SYSCTL_SYSCALL */
3074
3075 static int deprecated_sysctl_warning(struct __sysctl_args *args)
3076 {
3077         static int msg_count;
3078         int name[CTL_MAXNAME];
3079         int i;
3080
3081         /* Check args->nlen. */
3082         if (args->nlen < 0 || args->nlen > CTL_MAXNAME)
3083                 return -ENOTDIR;
3084
3085         /* Read in the sysctl name for better debug message logging */
3086         for (i = 0; i < args->nlen; i++)
3087                 if (get_user(name[i], args->name + i))
3088                         return -EFAULT;
3089
3090         /* Ignore accesses to kernel.version */
3091         if ((args->nlen == 2) && (name[0] == CTL_KERN) && (name[1] == KERN_VERSION))
3092                 return 0;
3093
3094         if (msg_count < 5) {
3095                 msg_count++;
3096                 printk(KERN_INFO
3097                         "warning: process `%s' used the deprecated sysctl "
3098                         "system call with ", current->comm);
3099                 for (i = 0; i < args->nlen; i++)
3100                         printk("%d.", name[i]);
3101                 printk("\n");
3102         }
3103         return 0;
3104 }
3105
3106 /*
3107  * No sense putting this after each symbol definition, twice,
3108  * exception granted :-)
3109  */
3110 EXPORT_SYMBOL(proc_dointvec);
3111 EXPORT_SYMBOL(proc_dointvec_jiffies);
3112 EXPORT_SYMBOL(proc_dointvec_minmax);
3113 EXPORT_SYMBOL(proc_dointvec_userhz_jiffies);
3114 EXPORT_SYMBOL(proc_dointvec_ms_jiffies);
3115 EXPORT_SYMBOL(proc_dostring);
3116 EXPORT_SYMBOL(proc_doulongvec_minmax);
3117 EXPORT_SYMBOL(proc_doulongvec_ms_jiffies_minmax);
3118 EXPORT_SYMBOL(register_sysctl_table);
3119 EXPORT_SYMBOL(register_sysctl_paths);
3120 EXPORT_SYMBOL(sysctl_intvec);
3121 EXPORT_SYMBOL(sysctl_jiffies);
3122 EXPORT_SYMBOL(sysctl_ms_jiffies);
3123 EXPORT_SYMBOL(sysctl_string);
3124 EXPORT_SYMBOL(sysctl_data);
3125 EXPORT_SYMBOL(unregister_sysctl_table);