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
  52 #include <asm/uaccess.h>
  53 #include <asm/processor.h>
  54
  55 #ifdef CONFIG_X86
  56 #include <asm/nmi.h>
  57 #include <asm/stacktrace.h>
  58 #include <asm/io.h>
  59 #endif
  60
  61 static int deprecated_sysctl_warning(struct __sysctl_args *args);
  62
  63 #if defined(CONFIG_SYSCTL)
  64
  65 /* External variables not in a header file. */
  66 extern int C_A_D;
  67 extern int print_fatal_signals;
  68 extern int sysctl_overcommit_memory;
  69 extern int sysctl_overcommit_ratio;
  70 extern int sysctl_panic_on_oom;
  71 extern int sysctl_oom_kill_allocating_task;
  72 extern int sysctl_oom_dump_tasks;
  73 extern int max_threads;
  74 extern int core_uses_pid;
  75 extern int suid_dumpable;
  76 extern char core_pattern[];
  77 extern int pid_max;
  78 extern int min_free_kbytes;
  79 extern int pid_max_min, pid_max_max;
  80 extern int sysctl_drop_caches;
  81 extern int percpu_pagelist_fraction;
  82 extern int compat_log;
  83 extern int latencytop_enabled;
  84 extern int sysctl_nr_open_min, sysctl_nr_open_max;
  85 #ifdef CONFIG_RCU_TORTURE_TEST
  86 extern int rcutorture_runnable;
  87 #endif /* #ifdef CONFIG_RCU_TORTURE_TEST */
  88
  89 /* Constants used for minimum and  maximum */
  90 #if defined(CONFIG_HIGHMEM) || defined(CONFIG_DETECT_SOFTLOCKUP)
  91 static int one = 1;
  92 #endif
  93
  94 #ifdef CONFIG_DETECT_SOFTLOCKUP
  95 static int sixty = 60;
  96 static int neg_one = -1;
  97 #endif
  98
  99 #if defined(CONFIG_MMU) && defined(CONFIG_FILE_LOCKING)
 100 static int two = 2;
 101 #endif
 102
 103 static int zero;
 104 static int one_hundred = 100;
 105
 106 /* this is needed for the proc_dointvec_minmax for [fs_]overflow UID and GID */
 107 static int maxolduid = 65535;
 108 static int minolduid;
 109 static int min_percpu_pagelist_fract = 8;
 110
 111 static int ngroups_max = NGROUPS_MAX;
 112
 113 #ifdef CONFIG_MODULES
 114 extern char modprobe_path[];
 115 #endif
 116 #ifdef CONFIG_CHR_DEV_SG
 117 extern int sg_big_buff;
 118 #endif
 119
 120 #ifdef CONFIG_SPARC
 121 #include <asm/system.h>
 122 #endif
 123
 124 #ifdef __hppa__
 125 extern int pwrsw_enabled;
 126 extern int unaligned_enabled;
 127 #endif
 128
 129 #ifdef CONFIG_S390
 130 #ifdef CONFIG_MATHEMU
 131 extern int sysctl_ieee_emulation_warnings;
 132 #endif
 133 extern int sysctl_userprocess_debug;
 134 extern int spin_retry;
 135 #endif
 136
 137 #ifdef CONFIG_BSD_PROCESS_ACCT
 138 extern int acct_parm[];
 139 #endif
 140
 141 #ifdef CONFIG_IA64
 142 extern int no_unaligned_warning;
 143 #endif
 144
 145 #ifdef CONFIG_RT_MUTEXES
 146 extern int max_lock_depth;
 147 #endif
 148
 149 #ifdef CONFIG_PROC_SYSCTL
 150 static int proc_do_cad_pid(struct ctl_table *table, int write, struct file *filp,
 151                   void __user *buffer, size_t *lenp, loff_t *ppos);
 152 static int proc_taint(struct ctl_table *table, int write, struct file *filp,
 153                                void __user *buffer, size_t *lenp, loff_t *ppos);
 154 #endif
 155
 156 static struct ctl_table root_table[];
 157 static struct ctl_table_root sysctl_table_root;
 158 static struct ctl_table_header root_table_header = {
 159         .count = 1,
 160         .ctl_table = root_table,
 161         .ctl_entry = LIST_HEAD_INIT(sysctl_table_root.default_set.list),
 162         .root = &sysctl_table_root,
 163         .set = &sysctl_table_root.default_set,
 164 };
 165 static struct ctl_table_root sysctl_table_root = {
 166         .root_list = LIST_HEAD_INIT(sysctl_table_root.root_list),
 167         .default_set.list = LIST_HEAD_INIT(root_table_header.ctl_entry),
 168 };
 169
 170 static struct ctl_table kern_table[];
 171 static struct ctl_table vm_table[];
 172 static struct ctl_table fs_table[];
 173 static struct ctl_table debug_table[];
 174 static struct ctl_table dev_table[];
 175 extern struct ctl_table random_table[];
 176 #ifdef CONFIG_INOTIFY_USER
 177 extern struct ctl_table inotify_table[];
 178 #endif
 179
 180 #ifdef HAVE_ARCH_PICK_MMAP_LAYOUT
 181 int sysctl_legacy_va_layout;
 182 #endif
 183
 184 extern int prove_locking;
 185 extern int lock_stat;
 186
 187 /* The default sysctl tables: */
 188
 189 static struct ctl_table root_table[] = {
 190         {
 191                 .ctl_name       = CTL_KERN,
 192                 .procname       = "kernel",
 193                 .mode           = 0555,
 194                 .child          = kern_table,
 195         },
 196         {
 197                 .ctl_name       = CTL_VM,
 198                 .procname       = "vm",
 199                 .mode           = 0555,
 200                 .child          = vm_table,
 201         },
 202         {
 203                 .ctl_name       = CTL_FS,
 204                 .procname       = "fs",
 205                 .mode           = 0555,
 206                 .child          = fs_table,
 207         },
 208         {
 209                 .ctl_name       = CTL_DEBUG,
 210                 .procname       = "debug",
 211                 .mode           = 0555,
 212                 .child          = debug_table,
 213         },
 214         {
 215                 .ctl_name       = CTL_DEV,
 216                 .procname       = "dev",
 217                 .mode           = 0555,
 218                 .child          = dev_table,
 219         },
 220 /*
 221  * NOTE: do not add new entries to this table unless you have read
 222  * Documentation/sysctl/ctl_unnumbered.txt
 223  */
 224         { .ctl_name = 0 }
 225 };
 226
 227 #ifdef CONFIG_SCHED_DEBUG
 228 static int min_sched_granularity_ns = 100000;           /* 100 usecs */
 229 static int max_sched_granularity_ns = NSEC_PER_SEC;     /* 1 second */
 230 static int min_wakeup_granularity_ns;                   /* 0 usecs */
 231 static int max_wakeup_granularity_ns = NSEC_PER_SEC;    /* 1 second */
 232 #endif
 233
 234 static struct ctl_table kern_table[] = {
 235 #ifdef CONFIG_SCHED_DEBUG
 236         {
 237                 .ctl_name       = CTL_UNNUMBERED,
 238                 .procname       = "sched_min_granularity_ns",
 239                 .data           = &sysctl_sched_min_granularity,
 240                 .maxlen         = sizeof(unsigned int),
 241                 .mode           = 0644,
 242                 .proc_handler   = &sched_nr_latency_handler,
 243                 .strategy       = &sysctl_intvec,
 244                 .extra1         = &min_sched_granularity_ns,
 245                 .extra2         = &max_sched_granularity_ns,
 246         },
 247         {
 248                 .ctl_name       = CTL_UNNUMBERED,
 249                 .procname       = "sched_latency_ns",
 250                 .data           = &sysctl_sched_latency,
 251                 .maxlen         = sizeof(unsigned int),
 252                 .mode           = 0644,
 253                 .proc_handler   = &sched_nr_latency_handler,
 254                 .strategy       = &sysctl_intvec,
 255                 .extra1         = &min_sched_granularity_ns,
 256                 .extra2         = &max_sched_granularity_ns,
 257         },
 258         {
 259                 .ctl_name       = CTL_UNNUMBERED,
 260                 .procname       = "sched_wakeup_granularity_ns",
 261                 .data           = &sysctl_sched_wakeup_granularity,
 262                 .maxlen         = sizeof(unsigned int),
 263                 .mode           = 0644,
 264                 .proc_handler   = &proc_dointvec_minmax,
 265                 .strategy       = &sysctl_intvec,
 266                 .extra1         = &min_wakeup_granularity_ns,
 267                 .extra2         = &max_wakeup_granularity_ns,
 268         },
 269         {
 270                 .ctl_name       = CTL_UNNUMBERED,
 271                 .procname       = "sched_shares_ratelimit",
 272                 .data           = &sysctl_sched_shares_ratelimit,
 273                 .maxlen         = sizeof(unsigned int),
 274                 .mode           = 0644,
 275                 .proc_handler   = &proc_dointvec,
 276         },
 277         {
 278                 .ctl_name       = CTL_UNNUMBERED,
 279                 .procname       = "sched_shares_thresh",
 280                 .data           = &sysctl_sched_shares_thresh,
 281                 .maxlen         = sizeof(unsigned int),
 282                 .mode           = 0644,
 283                 .proc_handler   = &proc_dointvec_minmax,
 284                 .strategy       = &sysctl_intvec,
 285                 .extra1         = &zero,
 286         },
 287         {
 288                 .ctl_name       = CTL_UNNUMBERED,
 289                 .procname       = "sched_child_runs_first",
 290                 .data           = &sysctl_sched_child_runs_first,
 291                 .maxlen         = sizeof(unsigned int),
 292                 .mode           = 0644,
 293                 .proc_handler   = &proc_dointvec,
 294         },
 295         {
 296                 .ctl_name       = CTL_UNNUMBERED,
 297                 .procname       = "sched_features",
 298                 .data           = &sysctl_sched_features,
 299                 .maxlen         = sizeof(unsigned int),
 300                 .mode           = 0644,
 301                 .proc_handler   = &proc_dointvec,
 302         },
 303         {
 304                 .ctl_name       = CTL_UNNUMBERED,
 305                 .procname       = "sched_migration_cost",
 306                 .data           = &sysctl_sched_migration_cost,
 307                 .maxlen         = sizeof(unsigned int),
 308                 .mode           = 0644,
 309                 .proc_handler   = &proc_dointvec,
 310         },
 311         {
 312                 .ctl_name       = CTL_UNNUMBERED,
 313                 .procname       = "sched_nr_migrate",
 314                 .data           = &sysctl_sched_nr_migrate,
 315                 .maxlen         = sizeof(unsigned int),
 316                 .mode           = 0644,
 317                 .proc_handler   = &proc_dointvec,
 318         },
 319 #endif
 320         {
 321                 .ctl_name       = CTL_UNNUMBERED,
 322                 .procname       = "sched_rt_period_us",
 323                 .data           = &sysctl_sched_rt_period,
 324                 .maxlen         = sizeof(unsigned int),
 325                 .mode           = 0644,
 326                 .proc_handler   = &sched_rt_handler,
 327         },
 328         {
 329                 .ctl_name       = CTL_UNNUMBERED,
 330                 .procname       = "sched_rt_runtime_us",
 331                 .data           = &sysctl_sched_rt_runtime,
 332                 .maxlen         = sizeof(int),
 333                 .mode           = 0644,
 334                 .proc_handler   = &sched_rt_handler,
 335         },
 336         {
 337                 .ctl_name       = CTL_UNNUMBERED,
 338                 .procname       = "sched_compat_yield",
 339                 .data           = &sysctl_sched_compat_yield,
 340                 .maxlen         = sizeof(unsigned int),
 341                 .mode           = 0644,
 342                 .proc_handler   = &proc_dointvec,
 343         },
 344 #ifdef CONFIG_PROVE_LOCKING
 345         {
 346                 .ctl_name       = CTL_UNNUMBERED,
 347                 .procname       = "prove_locking",
 348                 .data           = &prove_locking,
 349                 .maxlen         = sizeof(int),
 350                 .mode           = 0644,
 351                 .proc_handler   = &proc_dointvec,
 352         },
 353 #endif
 354 #ifdef CONFIG_LOCK_STAT
 355         {
 356                 .ctl_name       = CTL_UNNUMBERED,
 357                 .procname       = "lock_stat",
 358                 .data           = &lock_stat,
 359                 .maxlen         = sizeof(int),
 360                 .mode           = 0644,
 361                 .proc_handler   = &proc_dointvec,
 362         },
 363 #endif
 364         {
 365                 .ctl_name       = KERN_PANIC,
 366                 .procname       = "panic",
 367                 .data           = &panic_timeout,
 368                 .maxlen         = sizeof(int),
 369                 .mode           = 0644,
 370                 .proc_handler   = &proc_dointvec,
 371         },
 372         {
 373                 .ctl_name       = KERN_CORE_USES_PID,
 374                 .procname       = "core_uses_pid",
 375                 .data           = &core_uses_pid,
 376                 .maxlen         = sizeof(int),
 377                 .mode           = 0644,
 378                 .proc_handler   = &proc_dointvec,
 379         },
 380         {
 381                 .ctl_name       = KERN_CORE_PATTERN,
 382                 .procname       = "core_pattern",
 383                 .data           = core_pattern,
 384                 .maxlen         = CORENAME_MAX_SIZE,
 385                 .mode           = 0644,
 386                 .proc_handler   = &proc_dostring,
 387                 .strategy       = &sysctl_string,
 388         },
 389 #ifdef CONFIG_PROC_SYSCTL
 390         {
 391                 .procname       = "tainted",
 392                 .maxlen         = sizeof(long),
 393                 .mode           = 0644,
 394                 .proc_handler   = &proc_taint,
 395         },
 396 #endif
 397 #ifdef CONFIG_LATENCYTOP
 398         {
 399                 .procname       = "latencytop",
 400                 .data           = &latencytop_enabled,
 401                 .maxlen         = sizeof(int),
 402                 .mode           = 0644,
 403                 .proc_handler   = &proc_dointvec,
 404         },
 405 #endif
 406 #ifdef CONFIG_BLK_DEV_INITRD
 407         {
 408                 .ctl_name       = KERN_REALROOTDEV,
 409                 .procname       = "real-root-dev",
 410                 .data           = &real_root_dev,
 411                 .maxlen         = sizeof(int),
 412                 .mode           = 0644,
 413                 .proc_handler   = &proc_dointvec,
 414         },
 415 #endif
 416         {
 417                 .ctl_name       = CTL_UNNUMBERED,
 418                 .procname       = "print-fatal-signals",
 419                 .data           = &print_fatal_signals,
 420                 .maxlen         = sizeof(int),
 421                 .mode           = 0644,
 422                 .proc_handler   = &proc_dointvec,
 423         },
 424 #ifdef CONFIG_SPARC
 425         {
 426                 .ctl_name       = KERN_SPARC_REBOOT,
 427                 .procname       = "reboot-cmd",
 428                 .data           = reboot_command,
 429                 .maxlen         = 256,
 430                 .mode           = 0644,
 431                 .proc_handler   = &proc_dostring,
 432                 .strategy       = &sysctl_string,
 433         },
 434         {
 435                 .ctl_name       = KERN_SPARC_STOP_A,
 436                 .procname       = "stop-a",
 437                 .data           = &stop_a_enabled,
 438                 .maxlen         = sizeof (int),
 439                 .mode           = 0644,
 440                 .proc_handler   = &proc_dointvec,
 441         },
 442         {
 443                 .ctl_name       = KERN_SPARC_SCONS_PWROFF,
 444                 .procname       = "scons-poweroff",
 445                 .data           = &scons_pwroff,
 446                 .maxlen         = sizeof (int),
 447                 .mode           = 0644,
 448                 .proc_handler   = &proc_dointvec,
 449         },
 450 #endif
 451 #ifdef __hppa__
 452         {
 453                 .ctl_name       = KERN_HPPA_PWRSW,
 454                 .procname       = "soft-power",
 455                 .data           = &pwrsw_enabled,
 456                 .maxlen         = sizeof (int),
 457                 .mode           = 0644,
 458                 .proc_handler   = &proc_dointvec,
 459         },
 460         {
 461                 .ctl_name       = KERN_HPPA_UNALIGNED,
 462                 .procname       = "unaligned-trap",
 463                 .data           = &unaligned_enabled,
 464                 .maxlen         = sizeof (int),
 465                 .mode           = 0644,
 466                 .proc_handler   = &proc_dointvec,
 467         },
 468 #endif
 469         {
 470                 .ctl_name       = KERN_CTLALTDEL,
 471                 .procname       = "ctrl-alt-del",
 472                 .data           = &C_A_D,
 473                 .maxlen         = sizeof(int),
 474                 .mode           = 0644,
 475                 .proc_handler   = &proc_dointvec,
 476         },
 477 #ifdef CONFIG_FUNCTION_TRACER
 478         {
 479                 .ctl_name       = CTL_UNNUMBERED,
 480                 .procname       = "ftrace_enabled",
 481                 .data           = &ftrace_enabled,
 482                 .maxlen         = sizeof(int),
 483                 .mode           = 0644,
 484                 .proc_handler   = &ftrace_enable_sysctl,
 485         },
 486 #endif
 487 #ifdef CONFIG_TRACING
 488         {
 489                 .ctl_name       = CTL_UNNUMBERED,
 490                 .procname       = "ftrace_dump_on_opps",
 491                 .data           = &ftrace_dump_on_oops,
 492                 .maxlen         = sizeof(int),
 493                 .mode           = 0644,
 494                 .proc_handler   = &proc_dointvec,
 495         },
 496 #endif
 497 #ifdef CONFIG_MODULES
 498         {
 499                 .ctl_name       = KERN_MODPROBE,
 500                 .procname       = "modprobe",
 501                 .data           = &modprobe_path,
 502                 .maxlen         = KMOD_PATH_LEN,
 503                 .mode           = 0644,
 504                 .proc_handler   = &proc_dostring,
 505                 .strategy       = &sysctl_string,
 506         },
 507 #endif
 508 #if defined(CONFIG_HOTPLUG) && defined(CONFIG_NET)
 509         {
 510                 .ctl_name       = KERN_HOTPLUG,
 511                 .procname       = "hotplug",
 512                 .data           = &uevent_helper,
 513                 .maxlen         = UEVENT_HELPER_PATH_LEN,
 514                 .mode           = 0644,
 515                 .proc_handler   = &proc_dostring,
 516                 .strategy       = &sysctl_string,
 517         },
 518 #endif
 519 #ifdef CONFIG_CHR_DEV_SG
 520         {
 521                 .ctl_name       = KERN_SG_BIG_BUFF,
 522                 .procname       = "sg-big-buff",
 523                 .data           = &sg_big_buff,
 524                 .maxlen         = sizeof (int),
 525                 .mode           = 0444,
 526                 .proc_handler   = &proc_dointvec,
 527         },
 528 #endif
 529 #ifdef CONFIG_BSD_PROCESS_ACCT
 530         {
 531                 .ctl_name       = KERN_ACCT,
 532                 .procname       = "acct",
 533                 .data           = &acct_parm,
 534                 .maxlen         = 3*sizeof(int),
 535                 .mode           = 0644,
 536                 .proc_handler   = &proc_dointvec,
 537         },
 538 #endif
 539 #ifdef CONFIG_MAGIC_SYSRQ
 540         {
 541                 .ctl_name       = KERN_SYSRQ,
 542                 .procname       = "sysrq",
 543                 .data           = &__sysrq_enabled,
 544                 .maxlen         = sizeof (int),
 545                 .mode           = 0644,
 546                 .proc_handler   = &proc_dointvec,
 547         },
 548 #endif
 549 #ifdef CONFIG_PROC_SYSCTL
 550         {
 551                 .procname       = "cad_pid",
 552                 .data           = NULL,
 553                 .maxlen         = sizeof (int),
 554                 .mode           = 0600,
 555                 .proc_handler   = &proc_do_cad_pid,
 556         },
 557 #endif
 558         {
 559                 .ctl_name       = KERN_MAX_THREADS,
 560                 .procname       = "threads-max",
 561                 .data           = &max_threads,
 562                 .maxlen         = sizeof(int),
 563                 .mode           = 0644,
 564                 .proc_handler   = &proc_dointvec,
 565         },
 566         {
 567                 .ctl_name       = KERN_RANDOM,
 568                 .procname       = "random",
 569                 .mode           = 0555,
 570                 .child          = random_table,
 571         },
 572         {
 573                 .ctl_name       = KERN_OVERFLOWUID,
 574                 .procname       = "overflowuid",
 575                 .data           = &overflowuid,
 576                 .maxlen         = sizeof(int),
 577                 .mode           = 0644,
 578                 .proc_handler   = &proc_dointvec_minmax,
 579                 .strategy       = &sysctl_intvec,
 580                 .extra1         = &minolduid,
 581                 .extra2         = &maxolduid,
 582         },
 583         {
 584                 .ctl_name       = KERN_OVERFLOWGID,
 585                 .procname       = "overflowgid",
 586                 .data           = &overflowgid,
 587                 .maxlen         = sizeof(int),
 588                 .mode           = 0644,
 589                 .proc_handler   = &proc_dointvec_minmax,
 590                 .strategy       = &sysctl_intvec,
 591                 .extra1         = &minolduid,
 592                 .extra2         = &maxolduid,
 593         },
 594 #ifdef CONFIG_S390
 595 #ifdef CONFIG_MATHEMU
 596         {
 597                 .ctl_name       = KERN_IEEE_EMULATION_WARNINGS,
 598                 .procname       = "ieee_emulation_warnings",
 599                 .data           = &sysctl_ieee_emulation_warnings,
 600                 .maxlen         = sizeof(int),
 601                 .mode           = 0644,
 602                 .proc_handler   = &proc_dointvec,
 603         },
 604 #endif
 605         {
 606                 .ctl_name       = KERN_S390_USER_DEBUG_LOGGING,
 607                 .procname       = "userprocess_debug",
 608                 .data           = &sysctl_userprocess_debug,
 609                 .maxlen         = sizeof(int),
 610                 .mode           = 0644,
 611                 .proc_handler   = &proc_dointvec,
 612         },
 613 #endif
 614         {
 615                 .ctl_name       = KERN_PIDMAX,
 616                 .procname       = "pid_max",
 617                 .data           = &pid_max,
 618                 .maxlen         = sizeof (int),
 619                 .mode           = 0644,
 620                 .proc_handler   = &proc_dointvec_minmax,
 621                 .strategy       = sysctl_intvec,
 622                 .extra1         = &pid_max_min,
 623                 .extra2         = &pid_max_max,
 624         },
 625         {
 626                 .ctl_name       = KERN_PANIC_ON_OOPS,
 627                 .procname       = "panic_on_oops",
 628                 .data           = &panic_on_oops,
 629                 .maxlen         = sizeof(int),
 630                 .mode           = 0644,
 631                 .proc_handler   = &proc_dointvec,
 632         },
 633 #if defined CONFIG_PRINTK
 634         {
 635                 .ctl_name       = KERN_PRINTK,
 636                 .procname       = "printk",
 637                 .data           = &console_loglevel,
 638                 .maxlen         = 4*sizeof(int),
 639                 .mode           = 0644,
 640                 .proc_handler   = &proc_dointvec,
 641         },
 642         {
 643                 .ctl_name       = KERN_PRINTK_RATELIMIT,
 644                 .procname       = "printk_ratelimit",
 645                 .data           = &printk_ratelimit_state.interval,
 646                 .maxlen         = sizeof(int),
 647                 .mode           = 0644,
 648                 .proc_handler   = &proc_dointvec_jiffies,
 649                 .strategy       = &sysctl_jiffies,
 650         },
 651         {
 652                 .ctl_name       = KERN_PRINTK_RATELIMIT_BURST,
 653                 .procname       = "printk_ratelimit_burst",
 654                 .data           = &printk_ratelimit_state.burst,
 655                 .maxlen         = sizeof(int),
 656                 .mode           = 0644,
 657                 .proc_handler   = &proc_dointvec,
 658         },
 659 #endif
 660         {
 661                 .ctl_name       = KERN_NGROUPS_MAX,
 662                 .procname       = "ngroups_max",
 663                 .data           = &ngroups_max,
 664                 .maxlen         = sizeof (int),
 665                 .mode           = 0444,
 666                 .proc_handler   = &proc_dointvec,
 667         },
 668 #if defined(CONFIG_X86_LOCAL_APIC) && defined(CONFIG_X86)
 669         {
 670                 .ctl_name       = KERN_UNKNOWN_NMI_PANIC,
 671                 .procname       = "unknown_nmi_panic",
 672                 .data           = &unknown_nmi_panic,
 673                 .maxlen         = sizeof (int),
 674                 .mode           = 0644,
 675                 .proc_handler   = &proc_dointvec,
 676         },
 677         {
 678                 .procname       = "nmi_watchdog",
 679                 .data           = &nmi_watchdog_enabled,
 680                 .maxlen         = sizeof (int),
 681                 .mode           = 0644,
 682                 .proc_handler   = &proc_nmi_enabled,
 683         },
 684 #endif
 685 #if defined(CONFIG_X86)
 686         {
 687                 .ctl_name       = KERN_PANIC_ON_NMI,
 688                 .procname       = "panic_on_unrecovered_nmi",
 689                 .data           = &panic_on_unrecovered_nmi,
 690                 .maxlen         = sizeof(int),
 691                 .mode           = 0644,
 692                 .proc_handler   = &proc_dointvec,
 693         },
 694         {
 695                 .ctl_name       = KERN_BOOTLOADER_TYPE,
 696                 .procname       = "bootloader_type",
 697                 .data           = &bootloader_type,
 698                 .maxlen         = sizeof (int),
 699                 .mode           = 0444,
 700                 .proc_handler   = &proc_dointvec,
 701         },
 702         {
 703                 .ctl_name       = CTL_UNNUMBERED,
 704                 .procname       = "kstack_depth_to_print",
 705                 .data           = &kstack_depth_to_print,
 706                 .maxlen         = sizeof(int),
 707                 .mode           = 0644,
 708                 .proc_handler   = &proc_dointvec,
 709         },
 710         {
 711                 .ctl_name       = CTL_UNNUMBERED,
 712                 .procname       = "io_delay_type",
 713                 .data           = &io_delay_type,
 714                 .maxlen         = sizeof(int),
 715                 .mode           = 0644,
 716                 .proc_handler   = &proc_dointvec,
 717         },
 718 #endif
 719 #if defined(CONFIG_MMU)
 720         {
 721                 .ctl_name       = KERN_RANDOMIZE,
 722                 .procname       = "randomize_va_space",
 723                 .data           = &randomize_va_space,
 724                 .maxlen         = sizeof(int),
 725                 .mode           = 0644,
 726                 .proc_handler   = &proc_dointvec,
 727         },
 728 #endif
 729 #if defined(CONFIG_S390) && defined(CONFIG_SMP)
 730         {
 731                 .ctl_name       = KERN_SPIN_RETRY,
 732                 .procname       = "spin_retry",
 733                 .data           = &spin_retry,
 734                 .maxlen         = sizeof (int),
 735                 .mode           = 0644,
 736                 .proc_handler   = &proc_dointvec,
 737         },
 738 #endif
 739 #if     defined(CONFIG_ACPI_SLEEP) && defined(CONFIG_X86)
 740         {
 741                 .procname       = "acpi_video_flags",
 742                 .data           = &acpi_realmode_flags,
 743                 .maxlen         = sizeof (unsigned long),
 744                 .mode           = 0644,
 745                 .proc_handler   = &proc_doulongvec_minmax,
 746         },
 747 #endif
 748 #ifdef CONFIG_IA64
 749         {
 750                 .ctl_name       = KERN_IA64_UNALIGNED,
 751                 .procname       = "ignore-unaligned-usertrap",
 752                 .data           = &no_unaligned_warning,
 753                 .maxlen         = sizeof (int),
 754                 .mode           = 0644,
 755                 .proc_handler   = &proc_dointvec,
 756         },
 757 #endif
 758 #ifdef CONFIG_DETECT_SOFTLOCKUP
 759         {
 760                 .ctl_name       = CTL_UNNUMBERED,
 761                 .procname       = "softlockup_panic",
 762                 .data           = &softlockup_panic,
 763                 .maxlen         = sizeof(int),
 764                 .mode           = 0644,
 765                 .proc_handler   = &proc_dointvec_minmax,
 766                 .strategy       = &sysctl_intvec,
 767                 .extra1         = &zero,
 768                 .extra2         = &one,
 769         },
 770         {
 771                 .ctl_name       = CTL_UNNUMBERED,
 772                 .procname       = "softlockup_thresh",
 773                 .data           = &softlockup_thresh,
 774                 .maxlen         = sizeof(int),
 775                 .mode           = 0644,
 776                 .proc_handler   = &proc_dointvec_minmax,
 777                 .strategy       = &sysctl_intvec,
 778                 .extra1         = &neg_one,
 779                 .extra2         = &sixty,
 780         },
 781         {
 782                 .ctl_name       = CTL_UNNUMBERED,
 783                 .procname       = "hung_task_check_count",
 784                 .data           = &sysctl_hung_task_check_count,
 785                 .maxlen         = sizeof(unsigned long),
 786                 .mode           = 0644,
 787                 .proc_handler   = &proc_doulongvec_minmax,
 788                 .strategy       = &sysctl_intvec,
 789         },
 790         {
 791                 .ctl_name       = CTL_UNNUMBERED,
 792                 .procname       = "hung_task_timeout_secs",
 793                 .data           = &sysctl_hung_task_timeout_secs,
 794                 .maxlen         = sizeof(unsigned long),
 795                 .mode           = 0644,
 796                 .proc_handler   = &proc_doulongvec_minmax,
 797                 .strategy       = &sysctl_intvec,
 798         },
 799         {
 800                 .ctl_name       = CTL_UNNUMBERED,
 801                 .procname       = "hung_task_warnings",
 802                 .data           = &sysctl_hung_task_warnings,
 803                 .maxlen         = sizeof(unsigned long),
 804                 .mode           = 0644,
 805                 .proc_handler   = &proc_doulongvec_minmax,
 806                 .strategy       = &sysctl_intvec,
 807         },
 808 #endif
 809 #ifdef CONFIG_COMPAT
 810         {
 811                 .ctl_name       = KERN_COMPAT_LOG,
 812                 .procname       = "compat-log",
 813                 .data           = &compat_log,
 814                 .maxlen         = sizeof (int),
 815                 .mode           = 0644,
 816                 .proc_handler   = &proc_dointvec,
 817         },
 818 #endif
 819 #ifdef CONFIG_RT_MUTEXES
 820         {
 821                 .ctl_name       = KERN_MAX_LOCK_DEPTH,
 822                 .procname       = "max_lock_depth",
 823                 .data           = &max_lock_depth,
 824                 .maxlen         = sizeof(int),
 825                 .mode           = 0644,
 826                 .proc_handler   = &proc_dointvec,
 827         },
 828 #endif
 829         {
 830                 .ctl_name       = CTL_UNNUMBERED,
 831                 .procname       = "poweroff_cmd",
 832                 .data           = &poweroff_cmd,
 833                 .maxlen         = POWEROFF_CMD_PATH_LEN,
 834                 .mode           = 0644,
 835                 .proc_handler   = &proc_dostring,
 836                 .strategy       = &sysctl_string,
 837         },
 838 #ifdef CONFIG_KEYS
 839         {
 840                 .ctl_name       = CTL_UNNUMBERED,
 841                 .procname       = "keys",
 842                 .mode           = 0555,
 843                 .child          = key_sysctls,
 844         },
 845 #endif
 846 #ifdef CONFIG_RCU_TORTURE_TEST
 847         {
 848                 .ctl_name       = CTL_UNNUMBERED,
 849                 .procname       = "rcutorture_runnable",
 850                 .data           = &rcutorture_runnable,
 851                 .maxlen         = sizeof(int),
 852                 .mode           = 0644,
 853                 .proc_handler   = &proc_dointvec,
 854         },
 855 #endif
 856 #ifdef CONFIG_UNEVICTABLE_LRU
 857         {
 858                 .ctl_name       = CTL_UNNUMBERED,
 859                 .procname       = "scan_unevictable_pages",
 860                 .data           = &scan_unevictable_pages,
 861                 .maxlen         = sizeof(scan_unevictable_pages),
 862                 .mode           = 0644,
 863                 .proc_handler   = &scan_unevictable_handler,
 864         },
 865 #endif
 866 /*
 867  * NOTE: do not add new entries to this table unless you have read
 868  * Documentation/sysctl/ctl_unnumbered.txt
 869  */
 870         { .ctl_name = 0 }
 871 };
 872
 873 static struct ctl_table vm_table[] = {
 874         {
 875                 .ctl_name       = VM_OVERCOMMIT_MEMORY,
 876                 .procname       = "overcommit_memory",
 877                 .data           = &sysctl_overcommit_memory,
 878                 .maxlen         = sizeof(sysctl_overcommit_memory),
 879                 .mode           = 0644,
 880                 .proc_handler   = &proc_dointvec,
 881         },
 882         {
 883                 .ctl_name       = VM_PANIC_ON_OOM,
 884                 .procname       = "panic_on_oom",
 885                 .data           = &sysctl_panic_on_oom,
 886                 .maxlen         = sizeof(sysctl_panic_on_oom),
 887                 .mode           = 0644,
 888                 .proc_handler   = &proc_dointvec,
 889         },
 890         {
 891                 .ctl_name       = CTL_UNNUMBERED,
 892                 .procname       = "oom_kill_allocating_task",
 893                 .data           = &sysctl_oom_kill_allocating_task,
 894                 .maxlen         = sizeof(sysctl_oom_kill_allocating_task),
 895                 .mode           = 0644,
 896                 .proc_handler   = &proc_dointvec,
 897         },
 898         {
 899                 .ctl_name       = CTL_UNNUMBERED,
 900                 .procname       = "oom_dump_tasks",
 901                 .data           = &sysctl_oom_dump_tasks,
 902                 .maxlen         = sizeof(sysctl_oom_dump_tasks),
 903                 .mode           = 0644,
 904                 .proc_handler   = &proc_dointvec,
 905         },
 906         {
 907                 .ctl_name       = VM_OVERCOMMIT_RATIO,
 908                 .procname       = "overcommit_ratio",
 909                 .data           = &sysctl_overcommit_ratio,
 910                 .maxlen         = sizeof(sysctl_overcommit_ratio),
 911                 .mode           = 0644,
 912                 .proc_handler   = &proc_dointvec,
 913         },
 914         {
 915                 .ctl_name       = VM_PAGE_CLUSTER,
 916                 .procname       = "page-cluster",
 917                 .data           = &page_cluster,
 918                 .maxlen         = sizeof(int),
 919                 .mode           = 0644,
 920                 .proc_handler   = &proc_dointvec,
 921         },
 922         {
 923                 .ctl_name       = VM_DIRTY_BACKGROUND,
 924                 .procname       = "dirty_background_ratio",
 925                 .data           = &dirty_background_ratio,
 926                 .maxlen         = sizeof(dirty_background_ratio),
 927                 .mode           = 0644,
 928                 .proc_handler   = &proc_dointvec_minmax,
 929                 .strategy       = &sysctl_intvec,
 930                 .extra1         = &zero,
 931                 .extra2         = &one_hundred,
 932         },
 933         {
 934                 .ctl_name       = VM_DIRTY_RATIO,
 935                 .procname       = "dirty_ratio",
 936                 .data           = &vm_dirty_ratio,
 937                 .maxlen         = sizeof(vm_dirty_ratio),
 938                 .mode           = 0644,
 939                 .proc_handler   = &dirty_ratio_handler,
 940                 .strategy       = &sysctl_intvec,
 941                 .extra1         = &zero,
 942                 .extra2         = &one_hundred,
 943         },
 944         {
 945                 .procname       = "dirty_writeback_centisecs",
 946                 .data           = &dirty_writeback_interval,
 947                 .maxlen         = sizeof(dirty_writeback_interval),
 948                 .mode           = 0644,
 949                 .proc_handler   = &dirty_writeback_centisecs_handler,
 950         },
 951         {
 952                 .procname       = "dirty_expire_centisecs",
 953                 .data           = &dirty_expire_interval,
 954                 .maxlen         = sizeof(dirty_expire_interval),
 955                 .mode           = 0644,
 956                 .proc_handler   = &proc_dointvec_userhz_jiffies,
 957         },
 958         {
 959                 .ctl_name       = VM_NR_PDFLUSH_THREADS,
 960                 .procname       = "nr_pdflush_threads",
 961                 .data           = &nr_pdflush_threads,
 962                 .maxlen         = sizeof nr_pdflush_threads,
 963                 .mode           = 0444 /* read-only*/,
 964                 .proc_handler   = &proc_dointvec,
 965         },
 966         {
 967                 .ctl_name       = VM_SWAPPINESS,
 968                 .procname       = "swappiness",
 969                 .data           = &vm_swappiness,
 970                 .maxlen         = sizeof(vm_swappiness),
 971                 .mode           = 0644,
 972                 .proc_handler   = &proc_dointvec_minmax,
 973                 .strategy       = &sysctl_intvec,
 974                 .extra1         = &zero,
 975                 .extra2         = &one_hundred,
 976         },
 977 #ifdef CONFIG_HUGETLB_PAGE
 978          {
 979                 .procname       = "nr_hugepages",
 980                 .data           = NULL,
 981                 .maxlen         = sizeof(unsigned long),
 982                 .mode           = 0644,
 983                 .proc_handler   = &hugetlb_sysctl_handler,
 984                 .extra1         = (void *)&hugetlb_zero,
 985                 .extra2         = (void *)&hugetlb_infinity,
 986          },
 987          {
 988                 .ctl_name       = VM_HUGETLB_GROUP,
 989                 .procname       = "hugetlb_shm_group",
 990                 .data           = &sysctl_hugetlb_shm_group,
 991                 .maxlen         = sizeof(gid_t),
 992                 .mode           = 0644,
 993                 .proc_handler   = &proc_dointvec,
 994          },
 995          {
 996                 .ctl_name       = CTL_UNNUMBERED,
 997                 .procname       = "hugepages_treat_as_movable",
 998                 .data           = &hugepages_treat_as_movable,
 999                 .maxlen         = sizeof(int),
1000                 .mode           = 0644,
1001                 .proc_handler   = &hugetlb_treat_movable_handler,
1002         },
1003         {
1004                 .ctl_name       = CTL_UNNUMBERED,
1005                 .procname       = "nr_overcommit_hugepages",
1006                 .data           = NULL,
1007                 .maxlen         = sizeof(unsigned long),
1008                 .mode           = 0644,
1009                 .proc_handler   = &hugetlb_overcommit_handler,
1010                 .extra1         = (void *)&hugetlb_zero,
1011                 .extra2         = (void *)&hugetlb_infinity,
1012         },
1013 #endif
1014         {
1015                 .ctl_name       = VM_LOWMEM_RESERVE_RATIO,
1016                 .procname       = "lowmem_reserve_ratio",
1017                 .data           = &sysctl_lowmem_reserve_ratio,
1018                 .maxlen         = sizeof(sysctl_lowmem_reserve_ratio),
1019                 .mode           = 0644,
1020                 .proc_handler   = &lowmem_reserve_ratio_sysctl_handler,
1021                 .strategy       = &sysctl_intvec,
1022         },
1023         {
1024                 .ctl_name       = VM_DROP_PAGECACHE,
1025                 .procname       = "drop_caches",
1026                 .data           = &sysctl_drop_caches,
1027                 .maxlen         = sizeof(int),
1028                 .mode           = 0644,
1029                 .proc_handler   = drop_caches_sysctl_handler,
1030                 .strategy       = &sysctl_intvec,
1031         },
1032         {
1033                 .ctl_name       = VM_MIN_FREE_KBYTES,
1034                 .procname       = "min_free_kbytes",
1035                 .data           = &min_free_kbytes,
1036                 .maxlen         = sizeof(min_free_kbytes),
1037                 .mode           = 0644,
1038                 .proc_handler   = &min_free_kbytes_sysctl_handler,
1039                 .strategy       = &sysctl_intvec,
1040                 .extra1         = &zero,
1041         },
1042         {
1043                 .ctl_name       = VM_PERCPU_PAGELIST_FRACTION,
1044                 .procname       = "percpu_pagelist_fraction",
1045                 .data           = &percpu_pagelist_fraction,
1046                 .maxlen         = sizeof(percpu_pagelist_fraction),
1047                 .mode           = 0644,
1048                 .proc_handler   = &percpu_pagelist_fraction_sysctl_handler,
1049                 .strategy       = &sysctl_intvec,
1050                 .extra1         = &min_percpu_pagelist_fract,
1051         },
1052 #ifdef CONFIG_MMU
1053         {
1054                 .ctl_name       = VM_MAX_MAP_COUNT,
1055                 .procname       = "max_map_count",
1056                 .data           = &sysctl_max_map_count,
1057                 .maxlen         = sizeof(sysctl_max_map_count),
1058                 .mode           = 0644,
1059                 .proc_handler   = &proc_dointvec
1060         },
1061 #endif
1062         {
1063                 .ctl_name       = VM_LAPTOP_MODE,
1064                 .procname       = "laptop_mode",
1065                 .data           = &laptop_mode,
1066                 .maxlen         = sizeof(laptop_mode),
1067                 .mode           = 0644,
1068                 .proc_handler   = &proc_dointvec_jiffies,
1069                 .strategy       = &sysctl_jiffies,
1070         },
1071         {
1072                 .ctl_name       = VM_BLOCK_DUMP,
1073                 .procname       = "block_dump",
1074                 .data           = &block_dump,
1075                 .maxlen         = sizeof(block_dump),
1076                 .mode           = 0644,
1077                 .proc_handler   = &proc_dointvec,
1078                 .strategy       = &sysctl_intvec,
1079                 .extra1         = &zero,
1080         },
1081         {
1082                 .ctl_name       = VM_VFS_CACHE_PRESSURE,
1083                 .procname       = "vfs_cache_pressure",
1084                 .data           = &sysctl_vfs_cache_pressure,
1085                 .maxlen         = sizeof(sysctl_vfs_cache_pressure),
1086                 .mode           = 0644,
1087                 .proc_handler   = &proc_dointvec,
1088                 .strategy       = &sysctl_intvec,
1089                 .extra1         = &zero,
1090         },
1091 #ifdef HAVE_ARCH_PICK_MMAP_LAYOUT
1092         {
1093                 .ctl_name       = VM_LEGACY_VA_LAYOUT,
1094                 .procname       = "legacy_va_layout",
1095                 .data           = &sysctl_legacy_va_layout,
1096                 .maxlen         = sizeof(sysctl_legacy_va_layout),
1097                 .mode           = 0644,
1098                 .proc_handler   = &proc_dointvec,
1099                 .strategy       = &sysctl_intvec,
1100                 .extra1         = &zero,
1101         },
1102 #endif
1103 #ifdef CONFIG_NUMA
1104         {
1105                 .ctl_name       = VM_ZONE_RECLAIM_MODE,
1106                 .procname       = "zone_reclaim_mode",
1107                 .data           = &zone_reclaim_mode,
1108                 .maxlen         = sizeof(zone_reclaim_mode),
1109                 .mode           = 0644,
1110                 .proc_handler   = &proc_dointvec,
1111                 .strategy       = &sysctl_intvec,
1112                 .extra1         = &zero,
1113         },
1114         {
1115                 .ctl_name       = VM_MIN_UNMAPPED,
1116                 .procname       = "min_unmapped_ratio",
1117                 .data           = &sysctl_min_unmapped_ratio,
1118                 .maxlen         = sizeof(sysctl_min_unmapped_ratio),
1119                 .mode           = 0644,
1120                 .proc_handler   = &sysctl_min_unmapped_ratio_sysctl_handler,
1121                 .strategy       = &sysctl_intvec,
1122                 .extra1         = &zero,
1123                 .extra2         = &one_hundred,
1124         },
1125         {
1126                 .ctl_name       = VM_MIN_SLAB,
1127                 .procname       = "min_slab_ratio",
1128                 .data           = &sysctl_min_slab_ratio,
1129                 .maxlen         = sizeof(sysctl_min_slab_ratio),
1130                 .mode           = 0644,
1131                 .proc_handler   = &sysctl_min_slab_ratio_sysctl_handler,
1132                 .strategy       = &sysctl_intvec,
1133                 .extra1         = &zero,
1134                 .extra2         = &one_hundred,
1135         },
1136 #endif
1137 #ifdef CONFIG_SMP
1138         {
1139                 .ctl_name       = CTL_UNNUMBERED,
1140                 .procname       = "stat_interval",
1141                 .data           = &sysctl_stat_interval,
1142                 .maxlen         = sizeof(sysctl_stat_interval),
1143                 .mode           = 0644,
1144                 .proc_handler   = &proc_dointvec_jiffies,
1145                 .strategy       = &sysctl_jiffies,
1146         },
1147 #endif
1148 #ifdef CONFIG_SECURITY
1149         {
1150                 .ctl_name       = CTL_UNNUMBERED,
1151                 .procname       = "mmap_min_addr",
1152                 .data           = &mmap_min_addr,
1153                 .maxlen         = sizeof(unsigned long),
1154                 .mode           = 0644,
1155                 .proc_handler   = &proc_doulongvec_minmax,
1156         },
1157 #endif
1158 #ifdef CONFIG_NUMA
1159         {
1160                 .ctl_name       = CTL_UNNUMBERED,
1161                 .procname       = "numa_zonelist_order",
1162                 .data           = &numa_zonelist_order,
1163                 .maxlen         = NUMA_ZONELIST_ORDER_LEN,
1164                 .mode           = 0644,
1165                 .proc_handler   = &numa_zonelist_order_handler,
1166                 .strategy       = &sysctl_string,
1167         },
1168 #endif
1169 #if (defined(CONFIG_X86_32) && !defined(CONFIG_UML))|| \
1170    (defined(CONFIG_SUPERH) && defined(CONFIG_VSYSCALL))
1171         {
1172                 .ctl_name       = VM_VDSO_ENABLED,
1173                 .procname       = "vdso_enabled",
1174                 .data           = &vdso_enabled,
1175                 .maxlen         = sizeof(vdso_enabled),
1176                 .mode           = 0644,
1177                 .proc_handler   = &proc_dointvec,
1178                 .strategy       = &sysctl_intvec,
1179                 .extra1         = &zero,
1180         },
1181 #endif
1182 #ifdef CONFIG_HIGHMEM
1183         {
1184                 .ctl_name       = CTL_UNNUMBERED,
1185                 .procname       = "highmem_is_dirtyable",
1186                 .data           = &vm_highmem_is_dirtyable,
1187                 .maxlen         = sizeof(vm_highmem_is_dirtyable),
1188                 .mode           = 0644,
1189                 .proc_handler   = &proc_dointvec_minmax,
1190                 .strategy       = &sysctl_intvec,
1191                 .extra1         = &zero,
1192                 .extra2         = &one,
1193         },
1194 #endif
1195 /*
1196  * NOTE: do not add new entries to this table unless you have read
1197  * Documentation/sysctl/ctl_unnumbered.txt
1198  */
1199         { .ctl_name = 0 }
1200 };
1201
1202 #if defined(CONFIG_BINFMT_MISC) || defined(CONFIG_BINFMT_MISC_MODULE)
1203 static struct ctl_table binfmt_misc_table[] = {
1204         { .ctl_name = 0 }
1205 };
1206 #endif
1207
1208 static struct ctl_table fs_table[] = {
1209         {
1210                 .ctl_name       = FS_NRINODE,
1211                 .procname       = "inode-nr",
1212                 .data           = &inodes_stat,
1213                 .maxlen         = 2*sizeof(int),
1214                 .mode           = 0444,
1215                 .proc_handler   = &proc_dointvec,
1216         },
1217         {
1218                 .ctl_name       = FS_STATINODE,
1219                 .procname       = "inode-state",
1220                 .data           = &inodes_stat,
1221                 .maxlen         = 7*sizeof(int),
1222                 .mode           = 0444,
1223                 .proc_handler   = &proc_dointvec,
1224         },
1225         {
1226                 .procname       = "file-nr",
1227                 .data           = &files_stat,
1228                 .maxlen         = 3*sizeof(int),
1229                 .mode           = 0444,
1230                 .proc_handler   = &proc_nr_files,
1231         },
1232         {
1233                 .ctl_name       = FS_MAXFILE,
1234                 .procname       = "file-max",
1235                 .data           = &files_stat.max_files,
1236                 .maxlen         = sizeof(int),
1237                 .mode           = 0644,
1238                 .proc_handler   = &proc_dointvec,
1239         },
1240         {
1241                 .ctl_name       = CTL_UNNUMBERED,
1242                 .procname       = "nr_open",
1243                 .data           = &sysctl_nr_open,
1244                 .maxlen         = sizeof(int),
1245                 .mode           = 0644,
1246                 .proc_handler   = &proc_dointvec_minmax,
1247                 .extra1         = &sysctl_nr_open_min,
1248                 .extra2         = &sysctl_nr_open_max,
1249         },
1250         {
1251                 .ctl_name       = FS_DENTRY,
1252                 .procname       = "dentry-state",
1253                 .data           = &dentry_stat,
1254                 .maxlen         = 6*sizeof(int),
1255                 .mode           = 0444,
1256                 .proc_handler   = &proc_dointvec,
1257         },
1258         {
1259                 .ctl_name       = FS_OVERFLOWUID,
1260                 .procname       = "overflowuid",
1261                 .data           = &fs_overflowuid,
1262                 .maxlen         = sizeof(int),
1263                 .mode           = 0644,
1264                 .proc_handler   = &proc_dointvec_minmax,
1265                 .strategy       = &sysctl_intvec,
1266                 .extra1         = &minolduid,
1267                 .extra2         = &maxolduid,
1268         },
1269         {
1270                 .ctl_name       = FS_OVERFLOWGID,
1271                 .procname       = "overflowgid",
1272                 .data           = &fs_overflowgid,
1273                 .maxlen         = sizeof(int),
1274                 .mode           = 0644,
1275                 .proc_handler   = &proc_dointvec_minmax,
1276                 .strategy       = &sysctl_intvec,
1277                 .extra1         = &minolduid,
1278                 .extra2         = &maxolduid,
1279         },
1280 #ifdef CONFIG_FILE_LOCKING
1281         {
1282                 .ctl_name       = FS_LEASES,
1283                 .procname       = "leases-enable",
1284                 .data           = &leases_enable,
1285                 .maxlen         = sizeof(int),
1286                 .mode           = 0644,
1287                 .proc_handler   = &proc_dointvec,
1288         },
1289 #endif
1290 #ifdef CONFIG_DNOTIFY
1291         {
1292                 .ctl_name       = FS_DIR_NOTIFY,
1293                 .procname       = "dir-notify-enable",
1294                 .data           = &dir_notify_enable,
1295                 .maxlen         = sizeof(int),
1296                 .mode           = 0644,
1297                 .proc_handler   = &proc_dointvec,
1298         },
1299 #endif
1300 #ifdef CONFIG_MMU
1301 #ifdef CONFIG_FILE_LOCKING
1302         {
1303                 .ctl_name       = FS_LEASE_TIME,
1304                 .procname       = "lease-break-time",
1305                 .data           = &lease_break_time,
1306                 .maxlen         = sizeof(int),
1307                 .mode           = 0644,
1308                 .proc_handler   = &proc_dointvec_minmax,
1309                 .strategy       = &sysctl_intvec,
1310                 .extra1         = &zero,
1311                 .extra2         = &two,
1312         },
1313 #endif
1314 #ifdef CONFIG_AIO
1315         {
1316                 .procname       = "aio-nr",
1317                 .data           = &aio_nr,
1318                 .maxlen         = sizeof(aio_nr),
1319                 .mode           = 0444,
1320                 .proc_handler   = &proc_doulongvec_minmax,
1321         },
1322         {
1323                 .procname       = "aio-max-nr",
1324                 .data           = &aio_max_nr,
1325                 .maxlen         = sizeof(aio_max_nr),
1326                 .mode           = 0644,
1327                 .proc_handler   = &proc_doulongvec_minmax,
1328         },
1329 #endif /* CONFIG_AIO */
1330 #ifdef CONFIG_INOTIFY_USER
1331         {
1332                 .ctl_name       = FS_INOTIFY,
1333                 .procname       = "inotify",
1334                 .mode           = 0555,
1335                 .child          = inotify_table,
1336         },
1337 #endif
1338 #endif
1339         {
1340                 .ctl_name       = KERN_SETUID_DUMPABLE,
1341                 .procname       = "suid_dumpable",
1342                 .data           = &suid_dumpable,
1343                 .maxlen         = sizeof(int),
1344                 .mode           = 0644,
1345                 .proc_handler   = &proc_dointvec,
1346         },
1347 #if defined(CONFIG_BINFMT_MISC) || defined(CONFIG_BINFMT_MISC_MODULE)
1348         {
1349                 .ctl_name       = CTL_UNNUMBERED,
1350                 .procname       = "binfmt_misc",
1351                 .mode           = 0555,
1352                 .child          = binfmt_misc_table,
1353         },
1354 #endif
1355 /*
1356  * NOTE: do not add new entries to this table unless you have read
1357  * Documentation/sysctl/ctl_unnumbered.txt
1358  */
1359         { .ctl_name = 0 }
1360 };
1361
1362 static struct ctl_table debug_table[] = {
1363 #if defined(CONFIG_X86) || defined(CONFIG_PPC)
1364         {
1365                 .ctl_name       = CTL_UNNUMBERED,
1366                 .procname       = "exception-trace",
1367                 .data           = &show_unhandled_signals,
1368                 .maxlen         = sizeof(int),
1369                 .mode           = 0644,
1370                 .proc_handler   = proc_dointvec
1371         },
1372 #endif
1373         { .ctl_name = 0 }
1374 };
1375
1376 static struct ctl_table dev_table[] = {
1377         { .ctl_name = 0 }
1378 };
1379
1380 static DEFINE_SPINLOCK(sysctl_lock);
1381
1382 /* called under sysctl_lock */
1383 static int use_table(struct ctl_table_header *p)
1384 {
1385         if (unlikely(p->unregistering))
1386                 return 0;
1387         p->used++;
1388         return 1;
1389 }
1390
1391 /* called under sysctl_lock */
1392 static void unuse_table(struct ctl_table_header *p)
1393 {
1394         if (!--p->used)
1395                 if (unlikely(p->unregistering))
1396                         complete(p->unregistering);
1397 }
1398
1399 /* called under sysctl_lock, will reacquire if has to wait */
1400 static void start_unregistering(struct ctl_table_header *p)
1401 {
1402         /*
1403          * if p->used is 0, nobody will ever touch that entry again;
1404          * we'll eliminate all paths to it before dropping sysctl_lock
1405          */
1406         if (unlikely(p->used)) {
1407                 struct completion wait;
1408                 init_completion(&wait);
1409                 p->unregistering = &wait;
1410                 spin_unlock(&sysctl_lock);
1411                 wait_for_completion(&wait);
1412                 spin_lock(&sysctl_lock);
1413         } else {
1414                 /* anything non-NULL; we'll never dereference it */
1415                 p->unregistering = ERR_PTR(-EINVAL);
1416         }
1417         /*
1418          * do not remove from the list until nobody holds it; walking the
1419          * list in do_sysctl() relies on that.
1420          */
1421         list_del_init(&p->ctl_entry);
1422 }
1423
1424 void sysctl_head_get(struct ctl_table_header *head)
1425 {
1426         spin_lock(&sysctl_lock);
1427         head->count++;
1428         spin_unlock(&sysctl_lock);
1429 }
1430
1431 void sysctl_head_put(struct ctl_table_header *head)
1432 {
1433         spin_lock(&sysctl_lock);
1434         if (!--head->count)
1435                 kfree(head);
1436         spin_unlock(&sysctl_lock);
1437 }
1438
1439 struct ctl_table_header *sysctl_head_grab(struct ctl_table_header *head)
1440 {
1441         if (!head)
1442                 BUG();
1443         spin_lock(&sysctl_lock);
1444         if (!use_table(head))
1445                 head = ERR_PTR(-ENOENT);
1446         spin_unlock(&sysctl_lock);
1447         return head;
1448 }
1449
1450 void sysctl_head_finish(struct ctl_table_header *head)
1451 {
1452         if (!head)
1453                 return;
1454         spin_lock(&sysctl_lock);
1455         unuse_table(head);
1456         spin_unlock(&sysctl_lock);
1457 }
1458
1459 static struct ctl_table_set *
1460 lookup_header_set(struct ctl_table_root *root, struct nsproxy *namespaces)
1461 {
1462         struct ctl_table_set *set = &root->default_set;
1463         if (root->lookup)
1464                 set = root->lookup(root, namespaces);
1465         return set;
1466 }
1467
1468 static struct list_head *
1469 lookup_header_list(struct ctl_table_root *root, struct nsproxy *namespaces)
1470 {
1471         struct ctl_table_set *set = lookup_header_set(root, namespaces);
1472         return &set->list;
1473 }
1474
1475 struct ctl_table_header *__sysctl_head_next(struct nsproxy *namespaces,
1476                                             struct ctl_table_header *prev)
1477 {
1478         struct ctl_table_root *root;
1479         struct list_head *header_list;
1480         struct ctl_table_header *head;
1481         struct list_head *tmp;
1482
1483         spin_lock(&sysctl_lock);
1484         if (prev) {
1485                 head = prev;
1486                 tmp = &prev->ctl_entry;
1487                 unuse_table(prev);
1488                 goto next;
1489         }
1490         tmp = &root_table_header.ctl_entry;
1491         for (;;) {
1492                 head = list_entry(tmp, struct ctl_table_header, ctl_entry);
1493
1494                 if (!use_table(head))
1495                         goto next;
1496                 spin_unlock(&sysctl_lock);
1497                 return head;
1498         next:
1499                 root = head->root;
1500                 tmp = tmp->next;
1501                 header_list = lookup_header_list(root, namespaces);
1502                 if (tmp != header_list)
1503                         continue;
1504
1505                 do {
1506                         root = list_entry(root->root_list.next,
1507                                         struct ctl_table_root, root_list);
1508                         if (root == &sysctl_table_root)
1509                                 goto out;
1510                         header_list = lookup_header_list(root, namespaces);
1511                 } while (list_empty(header_list));
1512                 tmp = header_list->next;
1513         }
1514 out:
1515         spin_unlock(&sysctl_lock);
1516         return NULL;
1517 }
1518
1519 struct ctl_table_header *sysctl_head_next(struct ctl_table_header *prev)
1520 {
1521         return __sysctl_head_next(current->nsproxy, prev);
1522 }
1523
1524 void register_sysctl_root(struct ctl_table_root *root)
1525 {
1526         spin_lock(&sysctl_lock);
1527         list_add_tail(&root->root_list, &sysctl_table_root.root_list);
1528         spin_unlock(&sysctl_lock);
1529 }
1530
1531 #ifdef CONFIG_SYSCTL_SYSCALL
1532 /* Perform the actual read/write of a sysctl table entry. */
1533 static int do_sysctl_strategy(struct ctl_table_root *root,
1534                         struct ctl_table *table,
1535                         void __user *oldval, size_t __user *oldlenp,
1536                         void __user *newval, size_t newlen)
1537 {
1538         int op = 0, rc;
1539
1540         if (oldval)
1541                 op |= MAY_READ;
1542         if (newval)
1543                 op |= MAY_WRITE;
1544         if (sysctl_perm(root, table, op))
1545                 return -EPERM;
1546
1547         if (table->strategy) {
1548                 rc = table->strategy(table, oldval, oldlenp, newval, newlen);
1549                 if (rc < 0)
1550                         return rc;
1551                 if (rc > 0)
1552                         return 0;
1553         }
1554
1555         /* If there is no strategy routine, or if the strategy returns
1556          * zero, proceed with automatic r/w */
1557         if (table->data && table->maxlen) {
1558                 rc = sysctl_data(table, oldval, oldlenp, newval, newlen);
1559                 if (rc < 0)
1560                         return rc;
1561         }
1562         return 0;
1563 }
1564
1565 static int parse_table(int __user *name, int nlen,
1566                        void __user *oldval, size_t __user *oldlenp,
1567                        void __user *newval, size_t newlen,
1568                        struct ctl_table_root *root,
1569                        struct ctl_table *table)
1570 {
1571         int n;
1572 repeat:
1573         if (!nlen)
1574                 return -ENOTDIR;
1575         if (get_user(n, name))
1576                 return -EFAULT;
1577         for ( ; table->ctl_name || table->procname; table++) {
1578                 if (!table->ctl_name)
1579                         continue;
1580                 if (n == table->ctl_name) {
1581                         int error;
1582                         if (table->child) {
1583                                 if (sysctl_perm(root, table, MAY_EXEC))
1584                                         return -EPERM;
1585                                 name++;
1586                                 nlen--;
1587                                 table = table->child;
1588                                 goto repeat;
1589                         }
1590                         error = do_sysctl_strategy(root, table,
1591                                                    oldval, oldlenp,
1592                                                    newval, newlen);
1593                         return error;
1594                 }
1595         }
1596         return -ENOTDIR;
1597 }
1598
1599 int do_sysctl(int __user *name, int nlen, void __user *oldval, size_t __user *oldlenp,
1600                void __user *newval, size_t newlen)
1601 {
1602         struct ctl_table_header *head;
1603         int error = -ENOTDIR;
1604
1605         if (nlen <= 0 || nlen >= CTL_MAXNAME)
1606                 return -ENOTDIR;
1607         if (oldval) {
1608                 int old_len;
1609                 if (!oldlenp || get_user(old_len, oldlenp))
1610                         return -EFAULT;
1611         }
1612
1613         for (head = sysctl_head_next(NULL); head;
1614                         head = sysctl_head_next(head)) {
1615                 error = parse_table(name, nlen, oldval, oldlenp,
1616                                         newval, newlen,
1617                                         head->root, head->ctl_table);
1618                 if (error != -ENOTDIR) {
1619                         sysctl_head_finish(head);
1620                         break;
1621                 }
1622         }
1623         return error;
1624 }
1625
1626 asmlinkage long sys_sysctl(struct __sysctl_args __user *args)
1627 {
1628         struct __sysctl_args tmp;
1629         int error;
1630
1631         if (copy_from_user(&tmp, args, sizeof(tmp)))
1632                 return -EFAULT;
1633
1634         error = deprecated_sysctl_warning(&tmp);
1635         if (error)
1636                 goto out;
1637
1638         lock_kernel();
1639         error = do_sysctl(tmp.name, tmp.nlen, tmp.oldval, tmp.oldlenp,
1640                           tmp.newval, tmp.newlen);
1641         unlock_kernel();
1642 out:
1643         return error;
1644 }
1645 #endif /* CONFIG_SYSCTL_SYSCALL */
1646
1647 /*
1648  * sysctl_perm does NOT grant the superuser all rights automatically, because
1649  * some sysctl variables are readonly even to root.
1650  */
1651
1652 static int test_perm(int mode, int op)
1653 {
1654         if (!current->euid)
1655                 mode >>= 6;
1656         else if (in_egroup_p(0))
1657                 mode >>= 3;
1658         if ((op & ~mode & (MAY_READ|MAY_WRITE|MAY_EXEC)) == 0)
1659                 return 0;
1660         return -EACCES;
1661 }
1662
1663 int sysctl_perm(struct ctl_table_root *root, struct ctl_table *table, int op)
1664 {
1665         int error;
1666         int mode;
1667
1668         error = security_sysctl(table, op & (MAY_READ | MAY_WRITE | MAY_EXEC));
1669         if (error)
1670                 return error;
1671
1672         if (root->permissions)
1673                 mode = root->permissions(root, current->nsproxy, table);
1674         else
1675                 mode = table->mode;
1676
1677         return test_perm(mode, op);
1678 }
1679
1680 static void sysctl_set_parent(struct ctl_table *parent, struct ctl_table *table)
1681 {
1682         for (; table->ctl_name || table->procname; table++) {
1683                 table->parent = parent;
1684                 if (table->child)
1685                         sysctl_set_parent(table, table->child);
1686         }
1687 }
1688
1689 static __init int sysctl_init(void)
1690 {
1691         sysctl_set_parent(NULL, root_table);
1692 #ifdef CONFIG_SYSCTL_SYSCALL_CHECK
1693         {
1694                 int err;
1695                 err = sysctl_check_table(current->nsproxy, root_table);
1696         }
1697 #endif
1698         return 0;
1699 }
1700
1701 core_initcall(sysctl_init);
1702
1703 static struct ctl_table *is_branch_in(struct ctl_table *branch,
1704                                       struct ctl_table *table)
1705 {
1706         struct ctl_table *p;
1707         const char *s = branch->procname;
1708
1709         /* branch should have named subdirectory as its first element */
1710         if (!s || !branch->child)
1711                 return NULL;
1712
1713         /* ... and nothing else */
1714         if (branch[1].procname || branch[1].ctl_name)
1715                 return NULL;
1716
1717         /* table should contain subdirectory with the same name */
1718         for (p = table; p->procname || p->ctl_name; p++) {
1719                 if (!p->child)
1720                         continue;
1721                 if (p->procname && strcmp(p->procname, s) == 0)
1722                         return p;
1723         }
1724         return NULL;
1725 }
1726
1727 /* see if attaching q to p would be an improvement */
1728 static void try_attach(struct ctl_table_header *p, struct ctl_table_header *q)
1729 {
1730         struct ctl_table *to = p->ctl_table, *by = q->ctl_table;
1731         struct ctl_table *next;
1732         int is_better = 0;
1733         int not_in_parent = !p->attached_by;
1734
1735         while ((next = is_branch_in(by, to)) != NULL) {
1736                 if (by == q->attached_by)
1737                         is_better = 1;
1738                 if (to == p->attached_by)
1739                         not_in_parent = 1;
1740                 by = by->child;
1741                 to = next->child;
1742         }
1743
1744         if (is_better && not_in_parent) {
1745                 q->attached_by = by;
1746                 q->attached_to = to;
1747                 q->parent = p;
1748         }
1749 }
1750
1751 /**
1752  * __register_sysctl_paths - register a sysctl hierarchy
1753  * @root: List of sysctl headers to register on
1754  * @namespaces: Data to compute which lists of sysctl entries are visible
1755  * @path: The path to the directory the sysctl table is in.
1756  * @table: the top-level table structure
1757  *
1758  * Register a sysctl table hierarchy. @table should be a filled in ctl_table
1759  * array. A completely 0 filled entry terminates the table.
1760  *
1761  * The members of the &struct ctl_table structure are used as follows:
1762  *
1763  * ctl_name - This is the numeric sysctl value used by sysctl(2). The number
1764  *            must be unique within that level of sysctl
1765  *
1766  * procname - the name of the sysctl file under /proc/sys. Set to %NULL to not
1767  *            enter a sysctl file
1768  *
1769  * data - a pointer to data for use by proc_handler
1770  *
1771  * maxlen - the maximum size in bytes of the data
1772  *
1773  * mode - the file permissions for the /proc/sys file, and for sysctl(2)
1774  *
1775  * child - a pointer to the child sysctl table if this entry is a directory, or
1776  *         %NULL.
1777  *
1778  * proc_handler - the text handler routine (described below)
1779  *
1780  * strategy - the strategy routine (described below)
1781  *
1782  * de - for internal use by the sysctl routines
1783  *
1784  * extra1, extra2 - extra pointers usable by the proc handler routines
1785  *
1786  * Leaf nodes in the sysctl tree will be represented by a single file
1787  * under /proc; non-leaf nodes will be represented by directories.
1788  *
1789  * sysctl(2) can automatically manage read and write requests through
1790  * the sysctl table.  The data and maxlen fields of the ctl_table
1791  * struct enable minimal validation of the values being written to be
1792  * performed, and the mode field allows minimal authentication.
1793  *
1794  * More sophisticated management can be enabled by the provision of a
1795  * strategy routine with the table entry.  This will be called before
1796  * any automatic read or write of the data is performed.
1797  *
1798  * The strategy routine may return
1799  *
1800  * < 0 - Error occurred (error is passed to user process)
1801  *
1802  * 0   - OK - proceed with automatic read or write.
1803  *
1804  * > 0 - OK - read or write has been done by the strategy routine, so
1805  *       return immediately.
1806  *
1807  * There must be a proc_handler routine for any terminal nodes
1808  * mirrored under /proc/sys (non-terminals are handled by a built-in
1809  * directory handler).  Several default handlers are available to
1810  * cover common cases -
1811  *
1812  * proc_dostring(), proc_dointvec(), proc_dointvec_jiffies(),
1813  * proc_dointvec_userhz_jiffies(), proc_dointvec_minmax(),
1814  * proc_doulongvec_ms_jiffies_minmax(), proc_doulongvec_minmax()
1815  *
1816  * It is the handler's job to read the input buffer from user memory
1817  * and process it. The handler should return 0 on success.
1818  *
1819  * This routine returns %NULL on a failure to register, and a pointer
1820  * to the table header on success.
1821  */
1822 struct ctl_table_header *__register_sysctl_paths(
1823         struct ctl_table_root *root,
1824         struct nsproxy *namespaces,
1825         const struct ctl_path *path, struct ctl_table *table)
1826 {
1827         struct ctl_table_header *header;
1828         struct ctl_table *new, **prevp;
1829         unsigned int n, npath;
1830         struct ctl_table_set *set;
1831
1832         /* Count the path components */
1833         for (npath = 0; path[npath].ctl_name || path[npath].procname; ++npath)
1834                 ;
1835
1836         /*
1837          * For each path component, allocate a 2-element ctl_table array.
1838          * The first array element will be filled with the sysctl entry
1839          * for this, the second will be the sentinel (ctl_name == 0).
1840          *
1841          * We allocate everything in one go so that we don't have to
1842          * worry about freeing additional memory in unregister_sysctl_table.
1843          */
1844         header = kzalloc(sizeof(struct ctl_table_header) +
1845                          (2 * npath * sizeof(struct ctl_table)), GFP_KERNEL);
1846         if (!header)
1847                 return NULL;
1848
1849         new = (struct ctl_table *) (header + 1);
1850
1851         /* Now connect the dots */
1852         prevp = &header->ctl_table;
1853         for (n = 0; n < npath; ++n, ++path) {
1854                 /* Copy the procname */
1855                 new->procname = path->procname;
1856                 new->ctl_name = path->ctl_name;
1857                 new->mode     = 0555;
1858
1859                 *prevp = new;
1860                 prevp = &new->child;
1861
1862                 new += 2;
1863         }
1864         *prevp = table;
1865         header->ctl_table_arg = table;
1866
1867         INIT_LIST_HEAD(&header->ctl_entry);
1868         header->used = 0;
1869         header->unregistering = NULL;
1870         header->root = root;
1871         sysctl_set_parent(NULL, header->ctl_table);
1872         header->count = 1;
1873 #ifdef CONFIG_SYSCTL_SYSCALL_CHECK
1874         if (sysctl_check_table(namespaces, header->ctl_table)) {
1875                 kfree(header);
1876                 return NULL;
1877         }
1878 #endif
1879         spin_lock(&sysctl_lock);
1880         header->set = lookup_header_set(root, namespaces);
1881         header->attached_by = header->ctl_table;
1882         header->attached_to = root_table;
1883         header->parent = &root_table_header;
1884         for (set = header->set; set; set = set->parent) {
1885                 struct ctl_table_header *p;
1886                 list_for_each_entry(p, &set->list, ctl_entry) {
1887                         if (p->unregistering)
1888                                 continue;
1889                         try_attach(p, header);
1890                 }
1891         }
1892         header->parent->count++;
1893         list_add_tail(&header->ctl_entry, &header->set->list);
1894         spin_unlock(&sysctl_lock);
1895
1896         return header;
1897 }
1898
1899 /**
1900  * register_sysctl_table_path - register a sysctl table hierarchy
1901  * @path: The path to the directory the sysctl table is in.
1902  * @table: the top-level table structure
1903  *
1904  * Register a sysctl table hierarchy. @table should be a filled in ctl_table
1905  * array. A completely 0 filled entry terminates the table.
1906  *
1907  * See __register_sysctl_paths for more details.
1908  */
1909 struct ctl_table_header *register_sysctl_paths(const struct ctl_path *path,
1910                                                 struct ctl_table *table)
1911 {
1912         return __register_sysctl_paths(&sysctl_table_root, current->nsproxy,
1913                                         path, table);
1914 }
1915
1916 /**
1917  * register_sysctl_table - register a sysctl table hierarchy
1918  * @table: the top-level table structure
1919  *
1920  * Register a sysctl table hierarchy. @table should be a filled in ctl_table
1921  * array. A completely 0 filled entry terminates the table.
1922  *
1923  * See register_sysctl_paths for more details.
1924  */
1925 struct ctl_table_header *register_sysctl_table(struct ctl_table *table)
1926 {
1927         static const struct ctl_path null_path[] = { {} };
1928
1929         return register_sysctl_paths(null_path, table);
1930 }
1931
1932 /**
1933  * unregister_sysctl_table - unregister a sysctl table hierarchy
1934  * @header: the header returned from register_sysctl_table
1935  *
1936  * Unregisters the sysctl table and all children. proc entries may not
1937  * actually be removed until they are no longer used by anyone.
1938  */
1939 void unregister_sysctl_table(struct ctl_table_header * header)
1940 {
1941         might_sleep();
1942
1943         if (header == NULL)
1944                 return;
1945
1946         spin_lock(&sysctl_lock);
1947         start_unregistering(header);
1948         if (!--header->parent->count) {
1949                 WARN_ON(1);
1950                 kfree(header->parent);
1951         }
1952         if (!--header->count)
1953                 kfree(header);
1954         spin_unlock(&sysctl_lock);
1955 }
1956
1957 int sysctl_is_seen(struct ctl_table_header *p)
1958 {
1959         struct ctl_table_set *set = p->set;
1960         int res;
1961         spin_lock(&sysctl_lock);
1962         if (p->unregistering)
1963                 res = 0;
1964         else if (!set->is_seen)
1965                 res = 1;
1966         else
1967                 res = set->is_seen(set);
1968         spin_unlock(&sysctl_lock);
1969         return res;
1970 }
1971
1972 void setup_sysctl_set(struct ctl_table_set *p,
1973         struct ctl_table_set *parent,
1974         int (*is_seen)(struct ctl_table_set *))
1975 {
1976         INIT_LIST_HEAD(&p->list);
1977         p->parent = parent ? parent : &sysctl_table_root.default_set;
1978         p->is_seen = is_seen;
1979 }
1980
1981 #else /* !CONFIG_SYSCTL */
1982 struct ctl_table_header *register_sysctl_table(struct ctl_table * table)
1983 {
1984         return NULL;
1985 }
1986
1987 struct ctl_table_header *register_sysctl_paths(const struct ctl_path *path,
1988                                                     struct ctl_table *table)
1989 {
1990         return NULL;
1991 }
1992
1993 void unregister_sysctl_table(struct ctl_table_header * table)
1994 {
1995 }
1996
1997 void setup_sysctl_set(struct ctl_table_set *p,
1998         struct ctl_table_set *parent,
1999         int (*is_seen)(struct ctl_table_set *))
2000 {
2001 }
2002
2003 void sysctl_head_put(struct ctl_table_header *head)
2004 {
2005 }
2006
2007 #endif /* CONFIG_SYSCTL */
2008
2009 /*
2010  * /proc/sys support
2011  */
2012
2013 #ifdef CONFIG_PROC_SYSCTL
2014
2015 static int _proc_do_string(void* data, int maxlen, int write,
2016                            struct file *filp, void __user *buffer,
2017                            size_t *lenp, loff_t *ppos)
2018 {
2019         size_t len;
2020         char __user *p;
2021         char c;
2022
2023         if (!data || !maxlen || !*lenp) {
2024                 *lenp = 0;
2025                 return 0;
2026         }
2027
2028         if (write) {
2029                 len = 0;
2030                 p = buffer;
2031                 while (len < *lenp) {
2032                         if (get_user(c, p++))
2033                                 return -EFAULT;
2034                         if (c == 0 || c == '\n')
2035                                 break;
2036                         len++;
2037                 }
2038                 if (len >= maxlen)
2039                         len = maxlen-1;
2040                 if(copy_from_user(data, buffer, len))
2041                         return -EFAULT;
2042                 ((char *) data)[len] = 0;
2043                 *ppos += *lenp;
2044         } else {
2045                 len = strlen(data);
2046                 if (len > maxlen)
2047                         len = maxlen;
2048
2049                 if (*ppos > len) {
2050                         *lenp = 0;
2051                         return 0;
2052                 }
2053
2054                 data += *ppos;
2055                 len  -= *ppos;
2056
2057                 if (len > *lenp)
2058                         len = *lenp;
2059                 if (len)
2060                         if(copy_to_user(buffer, data, len))
2061                                 return -EFAULT;
2062                 if (len < *lenp) {
2063                         if(put_user('\n', ((char __user *) buffer) + len))
2064                                 return -EFAULT;
2065                         len++;
2066                 }
2067                 *lenp = len;
2068                 *ppos += len;
2069         }
2070         return 0;
2071 }
2072
2073 /**
2074  * proc_dostring - read a string sysctl
2075  * @table: the sysctl table
2076  * @write: %TRUE if this is a write to the sysctl file
2077  * @filp: the file structure
2078  * @buffer: the user buffer
2079  * @lenp: the size of the user buffer
2080  * @ppos: file position
2081  *
2082  * Reads/writes a string from/to the user buffer. If the kernel
2083  * buffer provided is not large enough to hold the string, the
2084  * string is truncated. The copied string is %NULL-terminated.
2085  * If the string is being read by the user process, it is copied
2086  * and a newline '\n' is added. It is truncated if the buffer is
2087  * not large enough.
2088  *
2089  * Returns 0 on success.
2090  */
2091 int proc_dostring(struct ctl_table *table, int write, struct file *filp,
2092                   void __user *buffer, size_t *lenp, loff_t *ppos)
2093 {
2094         return _proc_do_string(table->data, table->maxlen, write, filp,
2095                                buffer, lenp, ppos);
2096 }
2097
2098
2099 static int do_proc_dointvec_conv(int *negp, unsigned long *lvalp,
2100                                  int *valp,
2101                                  int write, void *data)
2102 {
2103         if (write) {
2104                 *valp = *negp ? -*lvalp : *lvalp;
2105         } else {
2106                 int val = *valp;
2107                 if (val < 0) {
2108                         *negp = -1;
2109                         *lvalp = (unsigned long)-val;
2110                 } else {
2111                         *negp = 0;
2112                         *lvalp = (unsigned long)val;
2113                 }
2114         }
2115         return 0;
2116 }
2117
2118 static int __do_proc_dointvec(void *tbl_data, struct ctl_table *table,
2119                   int write, struct file *filp, void __user *buffer,
2120                   size_t *lenp, loff_t *ppos,
2121                   int (*conv)(int *negp, unsigned long *lvalp, int *valp,
2122                               int write, void *data),
2123                   void *data)
2124 {
2125 #define TMPBUFLEN 21
2126         int *i, vleft, first=1, neg, val;
2127         unsigned long lval;
2128         size_t left, len;
2129
2130         char buf[TMPBUFLEN], *p;
2131         char __user *s = buffer;
2132
2133         if (!tbl_data || !table->maxlen || !*lenp ||
2134             (*ppos && !write)) {
2135                 *lenp = 0;
2136                 return 0;
2137         }
2138
2139         i = (int *) tbl_data;
2140         vleft = table->maxlen / sizeof(*i);
2141         left = *lenp;
2142
2143         if (!conv)
2144                 conv = do_proc_dointvec_conv;
2145
2146         for (; left && vleft--; i++, first=0) {
2147                 if (write) {
2148                         while (left) {
2149                                 char c;
2150                                 if (get_user(c, s))
2151                                         return -EFAULT;
2152                                 if (!isspace(c))
2153                                         break;
2154                                 left--;
2155                                 s++;
2156                         }
2157                         if (!left)
2158                                 break;
2159                         neg = 0;
2160                         len = left;
2161                         if (len > sizeof(buf) - 1)
2162                                 len = sizeof(buf) - 1;
2163                         if (copy_from_user(buf, s, len))
2164                                 return -EFAULT;
2165                         buf[len] = 0;
2166                         p = buf;
2167                         if (*p == '-' && left > 1) {
2168                                 neg = 1;
2169                                 p++;
2170                         }
2171                         if (*p < '0' || *p > '9')
2172                                 break;
2173
2174                         lval = simple_strtoul(p, &p, 0);
2175
2176                         len = p-buf;
2177                         if ((len < left) && *p && !isspace(*p))
2178                                 break;
2179                         if (neg)
2180                                 val = -val;
2181                         s += len;
2182                         left -= len;
2183
2184                         if (conv(&neg, &lval, i, 1, data))
2185                                 break;
2186                 } else {
2187                         p = buf;
2188                         if (!first)
2189                                 *p++ = '\t';
2190
2191                         if (conv(&neg, &lval, i, 0, data))
2192                                 break;
2193
2194                         sprintf(p, "%s%lu", neg ? "-" : "", lval);
2195                         len = strlen(buf);
2196                         if (len > left)
2197                                 len = left;
2198                         if(copy_to_user(s, buf, len))
2199                                 return -EFAULT;
2200                         left -= len;
2201                         s += len;
2202                 }
2203         }
2204
2205         if (!write && !first && left) {
2206                 if(put_user('\n', s))
2207                         return -EFAULT;
2208                 left--, s++;
2209         }
2210         if (write) {
2211                 while (left) {
2212                         char c;
2213                         if (get_user(c, s++))
2214                                 return -EFAULT;
2215                         if (!isspace(c))
2216                                 break;
2217                         left--;
2218                 }
2219         }
2220         if (write && first)
2221                 return -EINVAL;
2222         *lenp -= left;
2223         *ppos += *lenp;
2224         return 0;
2225 #undef TMPBUFLEN
2226 }
2227
2228 static int do_proc_dointvec(struct ctl_table *table, int write, struct file *filp,
2229                   void __user *buffer, size_t *lenp, loff_t *ppos,
2230                   int (*conv)(int *negp, unsigned long *lvalp, int *valp,
2231                               int write, void *data),
2232                   void *data)
2233 {
2234         return __do_proc_dointvec(table->data, table, write, filp,
2235                         buffer, lenp, ppos, conv, data);
2236 }
2237
2238 /**
2239  * proc_dointvec - read a vector of integers
2240  * @table: the sysctl table
2241  * @write: %TRUE if this is a write to the sysctl file
2242  * @filp: the file structure
2243  * @buffer: the user buffer
2244  * @lenp: the size of the user buffer
2245  * @ppos: file position
2246  *
2247  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2248  * values from/to the user buffer, treated as an ASCII string.
2249  *
2250  * Returns 0 on success.
2251  */
2252 int proc_dointvec(struct ctl_table *table, int write, struct file *filp,
2253                      void __user *buffer, size_t *lenp, loff_t *ppos)
2254 {
2255     return do_proc_dointvec(table,write,filp,buffer,lenp,ppos,
2256                             NULL,NULL);
2257 }
2258
2259 /*
2260  * Taint values can only be increased
2261  * This means we can safely use a temporary.
2262  */
2263 static int proc_taint(struct ctl_table *table, int write, struct file *filp,
2264                                void __user *buffer, size_t *lenp, loff_t *ppos)
2265 {
2266         struct ctl_table t;
2267         unsigned long tmptaint = get_taint();
2268         int err;
2269
2270         if (write && !capable(CAP_SYS_ADMIN))
2271                 return -EPERM;
2272
2273         t = *table;
2274         t.data = &tmptaint;
2275         err = proc_doulongvec_minmax(&t, write, filp, buffer, lenp, ppos);
2276         if (err < 0)
2277                 return err;
2278
2279         if (write) {
2280                 /*
2281                  * Poor man's atomic or. Not worth adding a primitive
2282                  * to everyone's atomic.h for this
2283                  */
2284                 int i;
2285                 for (i = 0; i < BITS_PER_LONG && tmptaint >> i; i++) {
2286                         if ((tmptaint >> i) & 1)
2287                                 add_taint(i);
2288                 }
2289         }
2290
2291         return err;
2292 }
2293
2294 struct do_proc_dointvec_minmax_conv_param {
2295         int *min;
2296         int *max;
2297 };
2298
2299 static int do_proc_dointvec_minmax_conv(int *negp, unsigned long *lvalp,
2300                                         int *valp,
2301                                         int write, void *data)
2302 {
2303         struct do_proc_dointvec_minmax_conv_param *param = data;
2304         if (write) {
2305                 int val = *negp ? -*lvalp : *lvalp;
2306                 if ((param->min && *param->min > val) ||
2307                     (param->max && *param->max < val))
2308                         return -EINVAL;
2309                 *valp = val;
2310         } else {
2311                 int val = *valp;
2312                 if (val < 0) {
2313                         *negp = -1;
2314                         *lvalp = (unsigned long)-val;
2315                 } else {
2316                         *negp = 0;
2317                         *lvalp = (unsigned long)val;
2318                 }
2319         }
2320         return 0;
2321 }
2322
2323 /**
2324  * proc_dointvec_minmax - read a vector of integers with min/max values
2325  * @table: the sysctl table
2326  * @write: %TRUE if this is a write to the sysctl file
2327  * @filp: the file structure
2328  * @buffer: the user buffer
2329  * @lenp: the size of the user buffer
2330  * @ppos: file position
2331  *
2332  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2333  * values from/to the user buffer, treated as an ASCII string.
2334  *
2335  * This routine will ensure the values are within the range specified by
2336  * table->extra1 (min) and table->extra2 (max).
2337  *
2338  * Returns 0 on success.
2339  */
2340 int proc_dointvec_minmax(struct ctl_table *table, int write, struct file *filp,
2341                   void __user *buffer, size_t *lenp, loff_t *ppos)
2342 {
2343         struct do_proc_dointvec_minmax_conv_param param = {
2344                 .min = (int *) table->extra1,
2345                 .max = (int *) table->extra2,
2346         };
2347         return do_proc_dointvec(table, write, filp, buffer, lenp, ppos,
2348                                 do_proc_dointvec_minmax_conv, &param);
2349 }
2350
2351 static int __do_proc_doulongvec_minmax(void *data, struct ctl_table *table, int write,
2352                                      struct file *filp,
2353                                      void __user *buffer,
2354                                      size_t *lenp, loff_t *ppos,
2355                                      unsigned long convmul,
2356                                      unsigned long convdiv)
2357 {
2358 #define TMPBUFLEN 21
2359         unsigned long *i, *min, *max, val;
2360         int vleft, first=1, neg;
2361         size_t len, left;
2362         char buf[TMPBUFLEN], *p;
2363         char __user *s = buffer;
2364
2365         if (!data || !table->maxlen || !*lenp ||
2366             (*ppos && !write)) {
2367                 *lenp = 0;
2368                 return 0;
2369         }
2370
2371         i = (unsigned long *) data;
2372         min = (unsigned long *) table->extra1;
2373         max = (unsigned long *) table->extra2;
2374         vleft = table->maxlen / sizeof(unsigned long);
2375         left = *lenp;
2376
2377         for (; left && vleft--; i++, min++, max++, first=0) {
2378                 if (write) {
2379                         while (left) {
2380                                 char c;
2381                                 if (get_user(c, s))
2382                                         return -EFAULT;
2383                                 if (!isspace(c))
2384                                         break;
2385                                 left--;
2386                                 s++;
2387                         }
2388                         if (!left)
2389                                 break;
2390                         neg = 0;
2391                         len = left;
2392                         if (len > TMPBUFLEN-1)
2393                                 len = TMPBUFLEN-1;
2394                         if (copy_from_user(buf, s, len))
2395                                 return -EFAULT;
2396                         buf[len] = 0;
2397                         p = buf;
2398                         if (*p == '-' && left > 1) {
2399                                 neg = 1;
2400                                 p++;
2401                         }
2402                         if (*p < '0' || *p > '9')
2403                                 break;
2404                         val = simple_strtoul(p, &p, 0) * convmul / convdiv ;
2405                         len = p-buf;
2406                         if ((len < left) && *p && !isspace(*p))
2407                                 break;
2408                         if (neg)
2409                                 val = -val;
2410                         s += len;
2411                         left -= len;
2412
2413                         if(neg)
2414                                 continue;
2415                         if ((min && val < *min) || (max && val > *max))
2416                                 continue;
2417                         *i = val;
2418                 } else {
2419                         p = buf;
2420                         if (!first)
2421                                 *p++ = '\t';
2422                         sprintf(p, "%lu", convdiv * (*i) / convmul);
2423                         len = strlen(buf);
2424                         if (len > left)
2425                                 len = left;
2426                         if(copy_to_user(s, buf, len))
2427                                 return -EFAULT;
2428                         left -= len;
2429                         s += len;
2430                 }
2431         }
2432
2433         if (!write && !first && left) {
2434                 if(put_user('\n', s))
2435                         return -EFAULT;
2436                 left--, s++;
2437         }
2438         if (write) {
2439                 while (left) {
2440                         char c;
2441                         if (get_user(c, s++))
2442                                 return -EFAULT;
2443                         if (!isspace(c))
2444                                 break;
2445                         left--;
2446                 }
2447         }
2448         if (write && first)
2449                 return -EINVAL;
2450         *lenp -= left;
2451         *ppos += *lenp;
2452         return 0;
2453 #undef TMPBUFLEN
2454 }
2455
2456 static int do_proc_doulongvec_minmax(struct ctl_table *table, int write,
2457                                      struct file *filp,
2458                                      void __user *buffer,
2459                                      size_t *lenp, loff_t *ppos,
2460                                      unsigned long convmul,
2461                                      unsigned long convdiv)
2462 {
2463         return __do_proc_doulongvec_minmax(table->data, table, write,
2464                         filp, buffer, lenp, ppos, convmul, convdiv);
2465 }
2466
2467 /**
2468  * proc_doulongvec_minmax - read a vector of long integers with min/max values
2469  * @table: the sysctl table
2470  * @write: %TRUE if this is a write to the sysctl file
2471  * @filp: the file structure
2472  * @buffer: the user buffer
2473  * @lenp: the size of the user buffer
2474  * @ppos: file position
2475  *
2476  * Reads/writes up to table->maxlen/sizeof(unsigned long) unsigned long
2477  * values from/to the user buffer, treated as an ASCII string.
2478  *
2479  * This routine will ensure the values are within the range specified by
2480  * table->extra1 (min) and table->extra2 (max).
2481  *
2482  * Returns 0 on success.
2483  */
2484 int proc_doulongvec_minmax(struct ctl_table *table, int write, struct file *filp,
2485                            void __user *buffer, size_t *lenp, loff_t *ppos)
2486 {
2487     return do_proc_doulongvec_minmax(table, write, filp, buffer, lenp, ppos, 1l, 1l);
2488 }
2489
2490 /**
2491  * proc_doulongvec_ms_jiffies_minmax - read a vector of millisecond values with min/max values
2492  * @table: the sysctl table
2493  * @write: %TRUE if this is a write to the sysctl file
2494  * @filp: the file structure
2495  * @buffer: the user buffer
2496  * @lenp: the size of the user buffer
2497  * @ppos: file position
2498  *
2499  * Reads/writes up to table->maxlen/sizeof(unsigned long) unsigned long
2500  * values from/to the user buffer, treated as an ASCII string. The values
2501  * are treated as milliseconds, and converted to jiffies when they are stored.
2502  *
2503  * This routine will ensure the values are within the range specified by
2504  * table->extra1 (min) and table->extra2 (max).
2505  *
2506  * Returns 0 on success.
2507  */
2508 int proc_doulongvec_ms_jiffies_minmax(struct ctl_table *table, int write,
2509                                       struct file *filp,
2510                                       void __user *buffer,
2511                                       size_t *lenp, loff_t *ppos)
2512 {
2513     return do_proc_doulongvec_minmax(table, write, filp, buffer,
2514                                      lenp, ppos, HZ, 1000l);
2515 }
2516
2517
2518 static int do_proc_dointvec_jiffies_conv(int *negp, unsigned long *lvalp,
2519                                          int *valp,
2520                                          int write, void *data)
2521 {
2522         if (write) {
2523                 if (*lvalp > LONG_MAX / HZ)
2524                         return 1;
2525                 *valp = *negp ? -(*lvalp*HZ) : (*lvalp*HZ);
2526         } else {
2527                 int val = *valp;
2528                 unsigned long lval;
2529                 if (val < 0) {
2530                         *negp = -1;
2531                         lval = (unsigned long)-val;
2532                 } else {
2533                         *negp = 0;
2534                         lval = (unsigned long)val;
2535                 }
2536                 *lvalp = lval / HZ;
2537         }
2538         return 0;
2539 }
2540
2541 static int do_proc_dointvec_userhz_jiffies_conv(int *negp, unsigned long *lvalp,
2542                                                 int *valp,
2543                                                 int write, void *data)
2544 {
2545         if (write) {
2546                 if (USER_HZ < HZ && *lvalp > (LONG_MAX / HZ) * USER_HZ)
2547                         return 1;
2548                 *valp = clock_t_to_jiffies(*negp ? -*lvalp : *lvalp);
2549         } else {
2550                 int val = *valp;
2551                 unsigned long lval;
2552                 if (val < 0) {
2553                         *negp = -1;
2554                         lval = (unsigned long)-val;
2555                 } else {
2556                         *negp = 0;
2557                         lval = (unsigned long)val;
2558                 }
2559                 *lvalp = jiffies_to_clock_t(lval);
2560         }
2561         return 0;
2562 }
2563
2564 static int do_proc_dointvec_ms_jiffies_conv(int *negp, unsigned long *lvalp,
2565                                             int *valp,
2566                                             int write, void *data)
2567 {
2568         if (write) {
2569                 *valp = msecs_to_jiffies(*negp ? -*lvalp : *lvalp);
2570         } else {
2571                 int val = *valp;
2572                 unsigned long lval;
2573                 if (val < 0) {
2574                         *negp = -1;
2575                         lval = (unsigned long)-val;
2576                 } else {
2577                         *negp = 0;
2578                         lval = (unsigned long)val;
2579                 }
2580                 *lvalp = jiffies_to_msecs(lval);
2581         }
2582         return 0;
2583 }
2584
2585 /**
2586  * proc_dointvec_jiffies - read a vector of integers as seconds
2587  * @table: the sysctl table
2588  * @write: %TRUE if this is a write to the sysctl file
2589  * @filp: the file structure
2590  * @buffer: the user buffer
2591  * @lenp: the size of the user buffer
2592  * @ppos: file position
2593  *
2594  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2595  * values from/to the user buffer, treated as an ASCII string.
2596  * The values read are assumed to be in seconds, and are converted into
2597  * jiffies.
2598  *
2599  * Returns 0 on success.
2600  */
2601 int proc_dointvec_jiffies(struct ctl_table *table, int write, struct file *filp,
2602                           void __user *buffer, size_t *lenp, loff_t *ppos)
2603 {
2604     return do_proc_dointvec(table,write,filp,buffer,lenp,ppos,
2605                             do_proc_dointvec_jiffies_conv,NULL);
2606 }
2607
2608 /**
2609  * proc_dointvec_userhz_jiffies - read a vector of integers as 1/USER_HZ seconds
2610  * @table: the sysctl table
2611  * @write: %TRUE if this is a write to the sysctl file
2612  * @filp: the file structure
2613  * @buffer: the user buffer
2614  * @lenp: the size of the user buffer
2615  * @ppos: pointer to the file position
2616  *
2617  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2618  * values from/to the user buffer, treated as an ASCII string.
2619  * The values read are assumed to be in 1/USER_HZ seconds, and
2620  * are converted into jiffies.
2621  *
2622  * Returns 0 on success.
2623  */
2624 int proc_dointvec_userhz_jiffies(struct ctl_table *table, int write, struct file *filp,
2625                                  void __user *buffer, size_t *lenp, loff_t *ppos)
2626 {
2627     return do_proc_dointvec(table,write,filp,buffer,lenp,ppos,
2628                             do_proc_dointvec_userhz_jiffies_conv,NULL);
2629 }
2630
2631 /**
2632  * proc_dointvec_ms_jiffies - read a vector of integers as 1 milliseconds
2633  * @table: the sysctl table
2634  * @write: %TRUE if this is a write to the sysctl file
2635  * @filp: the file structure
2636  * @buffer: the user buffer
2637  * @lenp: the size of the user buffer
2638  * @ppos: file position
2639  * @ppos: the current position in the file
2640  *
2641  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2642  * values from/to the user buffer, treated as an ASCII string.
2643  * The values read are assumed to be in 1/1000 seconds, and
2644  * are converted into jiffies.
2645  *
2646  * Returns 0 on success.
2647  */
2648 int proc_dointvec_ms_jiffies(struct ctl_table *table, int write, struct file *filp,
2649                              void __user *buffer, size_t *lenp, loff_t *ppos)
2650 {
2651         return do_proc_dointvec(table, write, filp, buffer, lenp, ppos,
2652                                 do_proc_dointvec_ms_jiffies_conv, NULL);
2653 }
2654
2655 static int proc_do_cad_pid(struct ctl_table *table, int write, struct file *filp,
2656                            void __user *buffer, size_t *lenp, loff_t *ppos)
2657 {
2658         struct pid *new_pid;
2659         pid_t tmp;
2660         int r;
2661
2662         tmp = pid_vnr(cad_pid);
2663
2664         r = __do_proc_dointvec(&tmp, table, write, filp, buffer,
2665                                lenp, ppos, NULL, NULL);
2666         if (r || !write)
2667                 return r;
2668
2669         new_pid = find_get_pid(tmp);
2670         if (!new_pid)
2671                 return -ESRCH;
2672
2673         put_pid(xchg(&cad_pid, new_pid));
2674         return 0;
2675 }
2676
2677 #else /* CONFIG_PROC_FS */
2678
2679 int proc_dostring(struct ctl_table *table, int write, struct file *filp,
2680                   void __user *buffer, size_t *lenp, loff_t *ppos)
2681 {
2682         return -ENOSYS;
2683 }
2684
2685 int proc_dointvec(struct ctl_table *table, int write, struct file *filp,
2686                   void __user *buffer, size_t *lenp, loff_t *ppos)
2687 {
2688         return -ENOSYS;
2689 }
2690
2691 int proc_dointvec_minmax(struct ctl_table *table, int write, struct file *filp,
2692                     void __user *buffer, size_t *lenp, loff_t *ppos)
2693 {
2694         return -ENOSYS;
2695 }
2696
2697 int proc_dointvec_jiffies(struct ctl_table *table, int write, struct file *filp,
2698                     void __user *buffer, size_t *lenp, loff_t *ppos)
2699 {
2700         return -ENOSYS;
2701 }
2702
2703 int proc_dointvec_userhz_jiffies(struct ctl_table *table, int write, struct file *filp,
2704                     void __user *buffer, size_t *lenp, loff_t *ppos)
2705 {
2706         return -ENOSYS;
2707 }
2708
2709 int proc_dointvec_ms_jiffies(struct ctl_table *table, int write, struct file *filp,
2710                              void __user *buffer, size_t *lenp, loff_t *ppos)
2711 {
2712         return -ENOSYS;
2713 }
2714
2715 int proc_doulongvec_minmax(struct ctl_table *table, int write, struct file *filp,
2716                     void __user *buffer, size_t *lenp, loff_t *ppos)
2717 {
2718         return -ENOSYS;
2719 }
2720
2721 int proc_doulongvec_ms_jiffies_minmax(struct ctl_table *table, int write,
2722                                       struct file *filp,
2723                                       void __user *buffer,
2724                                       size_t *lenp, loff_t *ppos)
2725 {
2726     return -ENOSYS;
2727 }
2728
2729
2730 #endif /* CONFIG_PROC_FS */
2731
2732
2733 #ifdef CONFIG_SYSCTL_SYSCALL
2734 /*
2735  * General sysctl support routines
2736  */
2737
2738 /* The generic sysctl data routine (used if no strategy routine supplied) */
2739 int sysctl_data(struct ctl_table *table,
2740                 void __user *oldval, size_t __user *oldlenp,
2741                 void __user *newval, size_t newlen)
2742 {
2743         size_t len;
2744
2745         /* Get out of I don't have a variable */
2746         if (!table->data || !table->maxlen)
2747                 return -ENOTDIR;
2748
2749         if (oldval && oldlenp) {
2750                 if (get_user(len, oldlenp))
2751                         return -EFAULT;
2752                 if (len) {
2753                         if (len > table->maxlen)
2754                                 len = table->maxlen;
2755                         if (copy_to_user(oldval, table->data, len))
2756                                 return -EFAULT;
2757                         if (put_user(len, oldlenp))
2758                                 return -EFAULT;
2759                 }
2760         }
2761
2762         if (newval && newlen) {
2763                 if (newlen > table->maxlen)
2764                         newlen = table->maxlen;
2765
2766                 if (copy_from_user(table->data, newval, newlen))
2767                         return -EFAULT;
2768         }
2769         return 1;
2770 }
2771
2772 /* The generic string strategy routine: */
2773 int sysctl_string(struct ctl_table *table,
2774                   void __user *oldval, size_t __user *oldlenp,
2775                   void __user *newval, size_t newlen)
2776 {
2777         if (!table->data || !table->maxlen)
2778                 return -ENOTDIR;
2779
2780         if (oldval && oldlenp) {
2781                 size_t bufsize;
2782                 if (get_user(bufsize, oldlenp))
2783                         return -EFAULT;
2784                 if (bufsize) {
2785                         size_t len = strlen(table->data), copied;
2786
2787                         /* This shouldn't trigger for a well-formed sysctl */
2788                         if (len > table->maxlen)
2789                                 len = table->maxlen;
2790
2791                         /* Copy up to a max of bufsize-1 bytes of the string */
2792                         copied = (len >= bufsize) ? bufsize - 1 : len;
2793
2794                         if (copy_to_user(oldval, table->data, copied) ||
2795                             put_user(0, (char __user *)(oldval + copied)))
2796                                 return -EFAULT;
2797                         if (put_user(len, oldlenp))
2798                                 return -EFAULT;
2799                 }
2800         }
2801         if (newval && newlen) {
2802                 size_t len = newlen;
2803                 if (len > table->maxlen)
2804                         len = table->maxlen;
2805                 if(copy_from_user(table->data, newval, len))
2806                         return -EFAULT;
2807                 if (len == table->maxlen)
2808                         len--;
2809                 ((char *) table->data)[len] = 0;
2810         }
2811         return 1;
2812 }
2813
2814 /*
2815  * This function makes sure that all of the integers in the vector
2816  * are between the minimum and maximum values given in the arrays
2817  * table->extra1 and table->extra2, respectively.
2818  */
2819 int sysctl_intvec(struct ctl_table *table,
2820                 void __user *oldval, size_t __user *oldlenp,
2821                 void __user *newval, size_t newlen)
2822 {
2823
2824         if (newval && newlen) {
2825                 int __user *vec = (int __user *) newval;
2826                 int *min = (int *) table->extra1;
2827                 int *max = (int *) table->extra2;
2828                 size_t length;
2829                 int i;
2830
2831                 if (newlen % sizeof(int) != 0)
2832                         return -EINVAL;
2833
2834                 if (!table->extra1 && !table->extra2)
2835                         return 0;
2836
2837                 if (newlen > table->maxlen)
2838                         newlen = table->maxlen;
2839                 length = newlen / sizeof(int);
2840
2841                 for (i = 0; i < length; i++) {
2842                         int value;
2843                         if (get_user(value, vec + i))
2844                                 return -EFAULT;
2845                         if (min && value < min[i])
2846                                 return -EINVAL;
2847                         if (max && value > max[i])
2848                                 return -EINVAL;
2849                 }
2850         }
2851         return 0;
2852 }
2853
2854 /* Strategy function to convert jiffies to seconds */
2855 int sysctl_jiffies(struct ctl_table *table,
2856                 void __user *oldval, size_t __user *oldlenp,
2857                 void __user *newval, size_t newlen)
2858 {
2859         if (oldval && oldlenp) {
2860                 size_t olen;
2861
2862                 if (get_user(olen, oldlenp))
2863                         return -EFAULT;
2864                 if (olen) {
2865                         int val;
2866
2867                         if (olen < sizeof(int))
2868                                 return -EINVAL;
2869
2870                         val = *(int *)(table->data) / HZ;
2871                         if (put_user(val, (int __user *)oldval))
2872                                 return -EFAULT;
2873                         if (put_user(sizeof(int), oldlenp))
2874                                 return -EFAULT;
2875                 }
2876         }
2877         if (newval && newlen) {
2878                 int new;
2879                 if (newlen != sizeof(int))
2880                         return -EINVAL;
2881                 if (get_user(new, (int __user *)newval))
2882                         return -EFAULT;
2883                 *(int *)(table->data) = new*HZ;
2884         }
2885         return 1;
2886 }
2887
2888 /* Strategy function to convert jiffies to seconds */
2889 int sysctl_ms_jiffies(struct ctl_table *table,
2890                 void __user *oldval, size_t __user *oldlenp,
2891                 void __user *newval, size_t newlen)
2892 {
2893         if (oldval && oldlenp) {
2894                 size_t olen;
2895
2896                 if (get_user(olen, oldlenp))
2897                         return -EFAULT;
2898                 if (olen) {
2899                         int val;
2900
2901                         if (olen < sizeof(int))
2902                                 return -EINVAL;
2903
2904                         val = jiffies_to_msecs(*(int *)(table->data));
2905                         if (put_user(val, (int __user *)oldval))
2906                                 return -EFAULT;
2907                         if (put_user(sizeof(int), oldlenp))
2908                                 return -EFAULT;
2909                 }
2910         }
2911         if (newval && newlen) {
2912                 int new;
2913                 if (newlen != sizeof(int))
2914                         return -EINVAL;
2915                 if (get_user(new, (int __user *)newval))
2916                         return -EFAULT;
2917                 *(int *)(table->data) = msecs_to_jiffies(new);
2918         }
2919         return 1;
2920 }
2921
2922
2923
2924 #else /* CONFIG_SYSCTL_SYSCALL */
2925
2926
2927 asmlinkage long sys_sysctl(struct __sysctl_args __user *args)
2928 {
2929         struct __sysctl_args tmp;
2930         int error;
2931
2932         if (copy_from_user(&tmp, args, sizeof(tmp)))
2933                 return -EFAULT;
2934
2935         error = deprecated_sysctl_warning(&tmp);
2936
2937         /* If no error reading the parameters then just -ENOSYS ... */
2938         if (!error)
2939                 error = -ENOSYS;
2940
2941         return error;
2942 }
2943
2944 int sysctl_data(struct ctl_table *table,
2945                   void __user *oldval, size_t __user *oldlenp,
2946                   void __user *newval, size_t newlen)
2947 {
2948         return -ENOSYS;
2949 }
2950
2951 int sysctl_string(struct ctl_table *table,
2952                   void __user *oldval, size_t __user *oldlenp,
2953                   void __user *newval, size_t newlen)
2954 {
2955         return -ENOSYS;
2956 }
2957
2958 int sysctl_intvec(struct ctl_table *table,
2959                 void __user *oldval, size_t __user *oldlenp,
2960                 void __user *newval, size_t newlen)
2961 {
2962         return -ENOSYS;
2963 }
2964
2965 int sysctl_jiffies(struct ctl_table *table,
2966                 void __user *oldval, size_t __user *oldlenp,
2967                 void __user *newval, size_t newlen)
2968 {
2969         return -ENOSYS;
2970 }
2971
2972 int sysctl_ms_jiffies(struct ctl_table *table,
2973                 void __user *oldval, size_t __user *oldlenp,
2974                 void __user *newval, size_t newlen)
2975 {
2976         return -ENOSYS;
2977 }
2978
2979 #endif /* CONFIG_SYSCTL_SYSCALL */
2980
2981 static int deprecated_sysctl_warning(struct __sysctl_args *args)
2982 {
2983         static int msg_count;
2984         int name[CTL_MAXNAME];
2985         int i;
2986
2987         /* Check args->nlen. */
2988         if (args->nlen < 0 || args->nlen > CTL_MAXNAME)
2989                 return -ENOTDIR;
2990
2991         /* Read in the sysctl name for better debug message logging */
2992         for (i = 0; i < args->nlen; i++)
2993                 if (get_user(name[i], args->name + i))
2994                         return -EFAULT;
2995
2996         /* Ignore accesses to kernel.version */
2997         if ((args->nlen == 2) && (name[0] == CTL_KERN) && (name[1] == KERN_VERSION))
2998                 return 0;
2999
3000         if (msg_count < 5) {
3001                 msg_count++;
3002                 printk(KERN_INFO
3003                         "warning: process `%s' used the deprecated sysctl "
3004                         "system call with ", current->comm);
3005                 for (i = 0; i < args->nlen; i++)
3006                         printk("%d.", name[i]);
3007                 printk("\n");
3008         }
3009         return 0;
3010 }
3011
3012 /*
3013  * No sense putting this after each symbol definition, twice,
3014  * exception granted :-)
3015  */
3016 EXPORT_SYMBOL(proc_dointvec);
3017 EXPORT_SYMBOL(proc_dointvec_jiffies);
3018 EXPORT_SYMBOL(proc_dointvec_minmax);
3019 EXPORT_SYMBOL(proc_dointvec_userhz_jiffies);
3020 EXPORT_SYMBOL(proc_dointvec_ms_jiffies);
3021 EXPORT_SYMBOL(proc_dostring);
3022 EXPORT_SYMBOL(proc_doulongvec_minmax);
3023 EXPORT_SYMBOL(proc_doulongvec_ms_jiffies_minmax);
3024 EXPORT_SYMBOL(register_sysctl_table);
3025 EXPORT_SYMBOL(register_sysctl_paths);
3026 EXPORT_SYMBOL(sysctl_intvec);
3027 EXPORT_SYMBOL(sysctl_jiffies);
3028 EXPORT_SYMBOL(sysctl_ms_jiffies);
3029 EXPORT_SYMBOL(sysctl_string);
3030 EXPORT_SYMBOL(sysctl_data);
3031 EXPORT_SYMBOL(unregister_sysctl_table);