SAFE public projects git trees. - safe/jmp/linux-2.6/blob - 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/kmemcheck.h>
  30 #include <linux/fs.h>
  31 #include <linux/init.h>
  32 #include <linux/kernel.h>
  33 #include <linux/kobject.h>
  34 #include <linux/net.h>
  35 #include <linux/sysrq.h>
  36 #include <linux/highuid.h>
  37 #include <linux/writeback.h>
  38 #include <linux/hugetlb.h>
  39 #include <linux/initrd.h>
  40 #include <linux/key.h>
  41 #include <linux/times.h>
  42 #include <linux/limits.h>
  43 #include <linux/dcache.h>
  44 #include <linux/syscalls.h>
  45 #include <linux/vmstat.h>
  46 #include <linux/nfs_fs.h>
  47 #include <linux/acpi.h>
  48 #include <linux/reboot.h>
  49 #include <linux/ftrace.h>
  50 #include <linux/slow-work.h>
  51 #include <linux/perf_event.h>
  52
  53 #include <asm/uaccess.h>
  54 #include <asm/processor.h>
  55
  56 #ifdef CONFIG_X86
  57 #include <asm/nmi.h>
  58 #include <asm/stacktrace.h>
  59 #include <asm/io.h>
  60 #endif
  61
  62
  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 unsigned int core_pipe_limit;
  78 extern int pid_max;
  79 extern int min_free_kbytes;
  80 extern int pid_max_min, pid_max_max;
  81 extern int sysctl_drop_caches;
  82 extern int percpu_pagelist_fraction;
  83 extern int compat_log;
  84 extern int latencytop_enabled;
  85 extern int sysctl_nr_open_min, sysctl_nr_open_max;
  86 #ifndef CONFIG_MMU
  87 extern int sysctl_nr_trim_pages;
  88 #endif
  89 #ifdef CONFIG_RCU_TORTURE_TEST
  90 extern int rcutorture_runnable;
  91 #endif /* #ifdef CONFIG_RCU_TORTURE_TEST */
  92 #ifdef CONFIG_BLOCK
  93 extern int blk_iopoll_enabled;
  94 #endif
  95
  96 /* Constants used for minimum and  maximum */
  97 #ifdef CONFIG_DETECT_SOFTLOCKUP
  98 static int sixty = 60;
  99 static int neg_one = -1;
 100 #endif
 101
 102 static int zero;
 103 static int __maybe_unused one = 1;
 104 static int __maybe_unused two = 2;
 105 static unsigned long one_ul = 1;
 106 static int one_hundred = 100;
 107 #ifdef CONFIG_PRINTK
 108 static int ten_thousand = 10000;
 109 #endif
 110
 111 /* this is needed for the proc_doulongvec_minmax of vm_dirty_bytes */
 112 static unsigned long dirty_bytes_min = 2 * PAGE_SIZE;
 113
 114 /* this is needed for the proc_dointvec_minmax for [fs_]overflow UID and GID */
 115 static int maxolduid = 65535;
 116 static int minolduid;
 117 static int min_percpu_pagelist_fract = 8;
 118
 119 static int ngroups_max = NGROUPS_MAX;
 120
 121 #ifdef CONFIG_MODULES
 122 extern char modprobe_path[];
 123 extern int modules_disabled;
 124 #endif
 125 #ifdef CONFIG_CHR_DEV_SG
 126 extern int sg_big_buff;
 127 #endif
 128
 129 #ifdef CONFIG_SPARC
 130 #include <asm/system.h>
 131 #endif
 132
 133 #ifdef CONFIG_SPARC64
 134 extern int sysctl_tsb_ratio;
 135 #endif
 136
 137 #ifdef __hppa__
 138 extern int pwrsw_enabled;
 139 extern int unaligned_enabled;
 140 #endif
 141
 142 #ifdef CONFIG_S390
 143 #ifdef CONFIG_MATHEMU
 144 extern int sysctl_ieee_emulation_warnings;
 145 #endif
 146 extern int sysctl_userprocess_debug;
 147 extern int spin_retry;
 148 #endif
 149
 150 #ifdef CONFIG_BSD_PROCESS_ACCT
 151 extern int acct_parm[];
 152 #endif
 153
 154 #ifdef CONFIG_IA64
 155 extern int no_unaligned_warning;
 156 extern int unaligned_dump_stack;
 157 #endif
 158
 159 #ifdef CONFIG_RT_MUTEXES
 160 extern int max_lock_depth;
 161 #endif
 162
 163 #ifdef CONFIG_PROC_SYSCTL
 164 static int proc_do_cad_pid(struct ctl_table *table, int write,
 165                   void __user *buffer, size_t *lenp, loff_t *ppos);
 166 static int proc_taint(struct ctl_table *table, int write,
 167                                void __user *buffer, size_t *lenp, loff_t *ppos);
 168 #endif
 169
 170 static struct ctl_table root_table[];
 171 static struct ctl_table_root sysctl_table_root;
 172 static struct ctl_table_header root_table_header = {
 173         .count = 1,
 174         .ctl_table = root_table,
 175         .ctl_entry = LIST_HEAD_INIT(sysctl_table_root.default_set.list),
 176         .root = &sysctl_table_root,
 177         .set = &sysctl_table_root.default_set,
 178 };
 179 static struct ctl_table_root sysctl_table_root = {
 180         .root_list = LIST_HEAD_INIT(sysctl_table_root.root_list),
 181         .default_set.list = LIST_HEAD_INIT(root_table_header.ctl_entry),
 182 };
 183
 184 static struct ctl_table kern_table[];
 185 static struct ctl_table vm_table[];
 186 static struct ctl_table fs_table[];
 187 static struct ctl_table debug_table[];
 188 static struct ctl_table dev_table[];
 189 extern struct ctl_table random_table[];
 190 #ifdef CONFIG_INOTIFY_USER
 191 extern struct ctl_table inotify_table[];
 192 #endif
 193 #ifdef CONFIG_EPOLL
 194 extern struct ctl_table epoll_table[];
 195 #endif
 196
 197 #ifdef HAVE_ARCH_PICK_MMAP_LAYOUT
 198 int sysctl_legacy_va_layout;
 199 #endif
 200
 201 extern int prove_locking;
 202 extern int lock_stat;
 203
 204 /* The default sysctl tables: */
 205
 206 static struct ctl_table root_table[] = {
 207         {
 208                 .ctl_name       = CTL_KERN,
 209                 .procname       = "kernel",
 210                 .mode           = 0555,
 211                 .child          = kern_table,
 212         },
 213         {
 214                 .ctl_name       = CTL_VM,
 215                 .procname       = "vm",
 216                 .mode           = 0555,
 217                 .child          = vm_table,
 218         },
 219         {
 220                 .ctl_name       = CTL_FS,
 221                 .procname       = "fs",
 222                 .mode           = 0555,
 223                 .child          = fs_table,
 224         },
 225         {
 226                 .ctl_name       = CTL_DEBUG,
 227                 .procname       = "debug",
 228                 .mode           = 0555,
 229                 .child          = debug_table,
 230         },
 231         {
 232                 .ctl_name       = CTL_DEV,
 233                 .procname       = "dev",
 234                 .mode           = 0555,
 235                 .child          = dev_table,
 236         },
 237 /*
 238  * NOTE: do not add new entries to this table unless you have read
 239  * Documentation/sysctl/ctl_unnumbered.txt
 240  */
 241         { .ctl_name = 0 }
 242 };
 243
 244 #ifdef CONFIG_SCHED_DEBUG
 245 static int min_sched_granularity_ns = 100000;           /* 100 usecs */
 246 static int max_sched_granularity_ns = NSEC_PER_SEC;     /* 1 second */
 247 static int min_wakeup_granularity_ns;                   /* 0 usecs */
 248 static int max_wakeup_granularity_ns = NSEC_PER_SEC;    /* 1 second */
 249 #endif
 250
 251 static struct ctl_table kern_table[] = {
 252         {
 253                 .ctl_name       = CTL_UNNUMBERED,
 254                 .procname       = "sched_child_runs_first",
 255                 .data           = &sysctl_sched_child_runs_first,
 256                 .maxlen         = sizeof(unsigned int),
 257                 .mode           = 0644,
 258                 .proc_handler   = &proc_dointvec,
 259         },
 260 #ifdef CONFIG_SCHED_DEBUG
 261         {
 262                 .ctl_name       = CTL_UNNUMBERED,
 263                 .procname       = "sched_min_granularity_ns",
 264                 .data           = &sysctl_sched_min_granularity,
 265                 .maxlen         = sizeof(unsigned int),
 266                 .mode           = 0644,
 267                 .proc_handler   = &sched_nr_latency_handler,
 268                 .strategy       = &sysctl_intvec,
 269                 .extra1         = &min_sched_granularity_ns,
 270                 .extra2         = &max_sched_granularity_ns,
 271         },
 272         {
 273                 .ctl_name       = CTL_UNNUMBERED,
 274                 .procname       = "sched_latency_ns",
 275                 .data           = &sysctl_sched_latency,
 276                 .maxlen         = sizeof(unsigned int),
 277                 .mode           = 0644,
 278                 .proc_handler   = &sched_nr_latency_handler,
 279                 .strategy       = &sysctl_intvec,
 280                 .extra1         = &min_sched_granularity_ns,
 281                 .extra2         = &max_sched_granularity_ns,
 282         },
 283         {
 284                 .ctl_name       = CTL_UNNUMBERED,
 285                 .procname       = "sched_wakeup_granularity_ns",
 286                 .data           = &sysctl_sched_wakeup_granularity,
 287                 .maxlen         = sizeof(unsigned int),
 288                 .mode           = 0644,
 289                 .proc_handler   = &proc_dointvec_minmax,
 290                 .strategy       = &sysctl_intvec,
 291                 .extra1         = &min_wakeup_granularity_ns,
 292                 .extra2         = &max_wakeup_granularity_ns,
 293         },
 294         {
 295                 .ctl_name       = CTL_UNNUMBERED,
 296                 .procname       = "sched_shares_ratelimit",
 297                 .data           = &sysctl_sched_shares_ratelimit,
 298                 .maxlen         = sizeof(unsigned int),
 299                 .mode           = 0644,
 300                 .proc_handler   = &proc_dointvec,
 301         },
 302         {
 303                 .ctl_name       = CTL_UNNUMBERED,
 304                 .procname       = "sched_shares_thresh",
 305                 .data           = &sysctl_sched_shares_thresh,
 306                 .maxlen         = sizeof(unsigned int),
 307                 .mode           = 0644,
 308                 .proc_handler   = &proc_dointvec_minmax,
 309                 .strategy       = &sysctl_intvec,
 310                 .extra1         = &zero,
 311         },
 312         {
 313                 .ctl_name       = CTL_UNNUMBERED,
 314                 .procname       = "sched_features",
 315                 .data           = &sysctl_sched_features,
 316                 .maxlen         = sizeof(unsigned int),
 317                 .mode           = 0644,
 318                 .proc_handler   = &proc_dointvec,
 319         },
 320         {
 321                 .ctl_name       = CTL_UNNUMBERED,
 322                 .procname       = "sched_migration_cost",
 323                 .data           = &sysctl_sched_migration_cost,
 324                 .maxlen         = sizeof(unsigned int),
 325                 .mode           = 0644,
 326                 .proc_handler   = &proc_dointvec,
 327         },
 328         {
 329                 .ctl_name       = CTL_UNNUMBERED,
 330                 .procname       = "sched_nr_migrate",
 331                 .data           = &sysctl_sched_nr_migrate,
 332                 .maxlen         = sizeof(unsigned int),
 333                 .mode           = 0644,
 334                 .proc_handler   = &proc_dointvec,
 335         },
 336         {
 337                 .ctl_name       = CTL_UNNUMBERED,
 338                 .procname       = "sched_time_avg",
 339                 .data           = &sysctl_sched_time_avg,
 340                 .maxlen         = sizeof(unsigned int),
 341                 .mode           = 0644,
 342                 .proc_handler   = &proc_dointvec,
 343         },
 344         {
 345                 .ctl_name       = CTL_UNNUMBERED,
 346                 .procname       = "timer_migration",
 347                 .data           = &sysctl_timer_migration,
 348                 .maxlen         = sizeof(unsigned int),
 349                 .mode           = 0644,
 350                 .proc_handler   = &proc_dointvec_minmax,
 351                 .strategy       = &sysctl_intvec,
 352                 .extra1         = &zero,
 353                 .extra2         = &one,
 354         },
 355 #endif
 356         {
 357                 .ctl_name       = CTL_UNNUMBERED,
 358                 .procname       = "sched_rt_period_us",
 359                 .data           = &sysctl_sched_rt_period,
 360                 .maxlen         = sizeof(unsigned int),
 361                 .mode           = 0644,
 362                 .proc_handler   = &sched_rt_handler,
 363         },
 364         {
 365                 .ctl_name       = CTL_UNNUMBERED,
 366                 .procname       = "sched_rt_runtime_us",
 367                 .data           = &sysctl_sched_rt_runtime,
 368                 .maxlen         = sizeof(int),
 369                 .mode           = 0644,
 370                 .proc_handler   = &sched_rt_handler,
 371         },
 372         {
 373                 .ctl_name       = CTL_UNNUMBERED,
 374                 .procname       = "sched_compat_yield",
 375                 .data           = &sysctl_sched_compat_yield,
 376                 .maxlen         = sizeof(unsigned int),
 377                 .mode           = 0644,
 378                 .proc_handler   = &proc_dointvec,
 379         },
 380 #ifdef CONFIG_PROVE_LOCKING
 381         {
 382                 .ctl_name       = CTL_UNNUMBERED,
 383                 .procname       = "prove_locking",
 384                 .data           = &prove_locking,
 385                 .maxlen         = sizeof(int),
 386                 .mode           = 0644,
 387                 .proc_handler   = &proc_dointvec,
 388         },
 389 #endif
 390 #ifdef CONFIG_LOCK_STAT
 391         {
 392                 .ctl_name       = CTL_UNNUMBERED,
 393                 .procname       = "lock_stat",
 394                 .data           = &lock_stat,
 395                 .maxlen         = sizeof(int),
 396                 .mode           = 0644,
 397                 .proc_handler   = &proc_dointvec,
 398         },
 399 #endif
 400         {
 401                 .ctl_name       = KERN_PANIC,
 402                 .procname       = "panic",
 403                 .data           = &panic_timeout,
 404                 .maxlen         = sizeof(int),
 405                 .mode           = 0644,
 406                 .proc_handler   = &proc_dointvec,
 407         },
 408         {
 409                 .ctl_name       = KERN_CORE_USES_PID,
 410                 .procname       = "core_uses_pid",
 411                 .data           = &core_uses_pid,
 412                 .maxlen         = sizeof(int),
 413                 .mode           = 0644,
 414                 .proc_handler   = &proc_dointvec,
 415         },
 416         {
 417                 .ctl_name       = KERN_CORE_PATTERN,
 418                 .procname       = "core_pattern",
 419                 .data           = core_pattern,
 420                 .maxlen         = CORENAME_MAX_SIZE,
 421                 .mode           = 0644,
 422                 .proc_handler   = &proc_dostring,
 423                 .strategy       = &sysctl_string,
 424         },
 425         {
 426                 .ctl_name       = CTL_UNNUMBERED,
 427                 .procname       = "core_pipe_limit",
 428                 .data           = &core_pipe_limit,
 429                 .maxlen         = sizeof(unsigned int),
 430                 .mode           = 0644,
 431                 .proc_handler   = &proc_dointvec,
 432         },
 433 #ifdef CONFIG_PROC_SYSCTL
 434         {
 435                 .procname       = "tainted",
 436                 .maxlen         = sizeof(long),
 437                 .mode           = 0644,
 438                 .proc_handler   = &proc_taint,
 439         },
 440 #endif
 441 #ifdef CONFIG_LATENCYTOP
 442         {
 443                 .procname       = "latencytop",
 444                 .data           = &latencytop_enabled,
 445                 .maxlen         = sizeof(int),
 446                 .mode           = 0644,
 447                 .proc_handler   = &proc_dointvec,
 448         },
 449 #endif
 450 #ifdef CONFIG_BLK_DEV_INITRD
 451         {
 452                 .ctl_name       = KERN_REALROOTDEV,
 453                 .procname       = "real-root-dev",
 454                 .data           = &real_root_dev,
 455                 .maxlen         = sizeof(int),
 456                 .mode           = 0644,
 457                 .proc_handler   = &proc_dointvec,
 458         },
 459 #endif
 460         {
 461                 .ctl_name       = CTL_UNNUMBERED,
 462                 .procname       = "print-fatal-signals",
 463                 .data           = &print_fatal_signals,
 464                 .maxlen         = sizeof(int),
 465                 .mode           = 0644,
 466                 .proc_handler   = &proc_dointvec,
 467         },
 468 #ifdef CONFIG_SPARC
 469         {
 470                 .ctl_name       = KERN_SPARC_REBOOT,
 471                 .procname       = "reboot-cmd",
 472                 .data           = reboot_command,
 473                 .maxlen         = 256,
 474                 .mode           = 0644,
 475                 .proc_handler   = &proc_dostring,
 476                 .strategy       = &sysctl_string,
 477         },
 478         {
 479                 .ctl_name       = KERN_SPARC_STOP_A,
 480                 .procname       = "stop-a",
 481                 .data           = &stop_a_enabled,
 482                 .maxlen         = sizeof (int),
 483                 .mode           = 0644,
 484                 .proc_handler   = &proc_dointvec,
 485         },
 486         {
 487                 .ctl_name       = KERN_SPARC_SCONS_PWROFF,
 488                 .procname       = "scons-poweroff",
 489                 .data           = &scons_pwroff,
 490                 .maxlen         = sizeof (int),
 491                 .mode           = 0644,
 492                 .proc_handler   = &proc_dointvec,
 493         },
 494 #endif
 495 #ifdef CONFIG_SPARC64
 496         {
 497                 .ctl_name       = CTL_UNNUMBERED,
 498                 .procname       = "tsb-ratio",
 499                 .data           = &sysctl_tsb_ratio,
 500                 .maxlen         = sizeof (int),
 501                 .mode           = 0644,
 502                 .proc_handler   = &proc_dointvec,
 503         },
 504 #endif
 505 #ifdef __hppa__
 506         {
 507                 .ctl_name       = KERN_HPPA_PWRSW,
 508                 .procname       = "soft-power",
 509                 .data           = &pwrsw_enabled,
 510                 .maxlen         = sizeof (int),
 511                 .mode           = 0644,
 512                 .proc_handler   = &proc_dointvec,
 513         },
 514         {
 515                 .ctl_name       = KERN_HPPA_UNALIGNED,
 516                 .procname       = "unaligned-trap",
 517                 .data           = &unaligned_enabled,
 518                 .maxlen         = sizeof (int),
 519                 .mode           = 0644,
 520                 .proc_handler   = &proc_dointvec,
 521         },
 522 #endif
 523         {
 524                 .ctl_name       = KERN_CTLALTDEL,
 525                 .procname       = "ctrl-alt-del",
 526                 .data           = &C_A_D,
 527                 .maxlen         = sizeof(int),
 528                 .mode           = 0644,
 529                 .proc_handler   = &proc_dointvec,
 530         },
 531 #ifdef CONFIG_FUNCTION_TRACER
 532         {
 533                 .ctl_name       = CTL_UNNUMBERED,
 534                 .procname       = "ftrace_enabled",
 535                 .data           = &ftrace_enabled,
 536                 .maxlen         = sizeof(int),
 537                 .mode           = 0644,
 538                 .proc_handler   = &ftrace_enable_sysctl,
 539         },
 540 #endif
 541 #ifdef CONFIG_STACK_TRACER
 542         {
 543                 .ctl_name       = CTL_UNNUMBERED,
 544                 .procname       = "stack_tracer_enabled",
 545                 .data           = &stack_tracer_enabled,
 546                 .maxlen         = sizeof(int),
 547                 .mode           = 0644,
 548                 .proc_handler   = &stack_trace_sysctl,
 549         },
 550 #endif
 551 #ifdef CONFIG_TRACING
 552         {
 553                 .ctl_name       = CTL_UNNUMBERED,
 554                 .procname       = "ftrace_dump_on_oops",
 555                 .data           = &ftrace_dump_on_oops,
 556                 .maxlen         = sizeof(int),
 557                 .mode           = 0644,
 558                 .proc_handler   = &proc_dointvec,
 559         },
 560 #endif
 561 #ifdef CONFIG_MODULES
 562         {
 563                 .ctl_name       = KERN_MODPROBE,
 564                 .procname       = "modprobe",
 565                 .data           = &modprobe_path,
 566                 .maxlen         = KMOD_PATH_LEN,
 567                 .mode           = 0644,
 568                 .proc_handler   = &proc_dostring,
 569                 .strategy       = &sysctl_string,
 570         },
 571         {
 572                 .ctl_name       = CTL_UNNUMBERED,
 573                 .procname       = "modules_disabled",
 574                 .data           = &modules_disabled,
 575                 .maxlen         = sizeof(int),
 576                 .mode           = 0644,
 577                 /* only handle a transition from default "0" to "1" */
 578                 .proc_handler   = &proc_dointvec_minmax,
 579                 .extra1         = &one,
 580                 .extra2         = &one,
 581         },
 582 #endif
 583 #if defined(CONFIG_HOTPLUG) && defined(CONFIG_NET)
 584         {
 585                 .ctl_name       = KERN_HOTPLUG,
 586                 .procname       = "hotplug",
 587                 .data           = &uevent_helper,
 588                 .maxlen         = UEVENT_HELPER_PATH_LEN,
 589                 .mode           = 0644,
 590                 .proc_handler   = &proc_dostring,
 591                 .strategy       = &sysctl_string,
 592         },
 593 #endif
 594 #ifdef CONFIG_CHR_DEV_SG
 595         {
 596                 .ctl_name       = KERN_SG_BIG_BUFF,
 597                 .procname       = "sg-big-buff",
 598                 .data           = &sg_big_buff,
 599                 .maxlen         = sizeof (int),
 600                 .mode           = 0444,
 601                 .proc_handler   = &proc_dointvec,
 602         },
 603 #endif
 604 #ifdef CONFIG_BSD_PROCESS_ACCT
 605         {
 606                 .ctl_name       = KERN_ACCT,
 607                 .procname       = "acct",
 608                 .data           = &acct_parm,
 609                 .maxlen         = 3*sizeof(int),
 610                 .mode           = 0644,
 611                 .proc_handler   = &proc_dointvec,
 612         },
 613 #endif
 614 #ifdef CONFIG_MAGIC_SYSRQ
 615         {
 616                 .ctl_name       = KERN_SYSRQ,
 617                 .procname       = "sysrq",
 618                 .data           = &__sysrq_enabled,
 619                 .maxlen         = sizeof (int),
 620                 .mode           = 0644,
 621                 .proc_handler   = &proc_dointvec,
 622         },
 623 #endif
 624 #ifdef CONFIG_PROC_SYSCTL
 625         {
 626                 .procname       = "cad_pid",
 627                 .data           = NULL,
 628                 .maxlen         = sizeof (int),
 629                 .mode           = 0600,
 630                 .proc_handler   = &proc_do_cad_pid,
 631         },
 632 #endif
 633         {
 634                 .ctl_name       = KERN_MAX_THREADS,
 635                 .procname       = "threads-max",
 636                 .data           = &max_threads,
 637                 .maxlen         = sizeof(int),
 638                 .mode           = 0644,
 639                 .proc_handler   = &proc_dointvec,
 640         },
 641         {
 642                 .ctl_name       = KERN_RANDOM,
 643                 .procname       = "random",
 644                 .mode           = 0555,
 645                 .child          = random_table,
 646         },
 647         {
 648                 .ctl_name       = KERN_OVERFLOWUID,
 649                 .procname       = "overflowuid",
 650                 .data           = &overflowuid,
 651                 .maxlen         = sizeof(int),
 652                 .mode           = 0644,
 653                 .proc_handler   = &proc_dointvec_minmax,
 654                 .strategy       = &sysctl_intvec,
 655                 .extra1         = &minolduid,
 656                 .extra2         = &maxolduid,
 657         },
 658         {
 659                 .ctl_name       = KERN_OVERFLOWGID,
 660                 .procname       = "overflowgid",
 661                 .data           = &overflowgid,
 662                 .maxlen         = sizeof(int),
 663                 .mode           = 0644,
 664                 .proc_handler   = &proc_dointvec_minmax,
 665                 .strategy       = &sysctl_intvec,
 666                 .extra1         = &minolduid,
 667                 .extra2         = &maxolduid,
 668         },
 669 #ifdef CONFIG_S390
 670 #ifdef CONFIG_MATHEMU
 671         {
 672                 .ctl_name       = KERN_IEEE_EMULATION_WARNINGS,
 673                 .procname       = "ieee_emulation_warnings",
 674                 .data           = &sysctl_ieee_emulation_warnings,
 675                 .maxlen         = sizeof(int),
 676                 .mode           = 0644,
 677                 .proc_handler   = &proc_dointvec,
 678         },
 679 #endif
 680         {
 681                 .ctl_name       = KERN_S390_USER_DEBUG_LOGGING,
 682                 .procname       = "userprocess_debug",
 683                 .data           = &sysctl_userprocess_debug,
 684                 .maxlen         = sizeof(int),
 685                 .mode           = 0644,
 686                 .proc_handler   = &proc_dointvec,
 687         },
 688 #endif
 689         {
 690                 .ctl_name       = KERN_PIDMAX,
 691                 .procname       = "pid_max",
 692                 .data           = &pid_max,
 693                 .maxlen         = sizeof (int),
 694                 .mode           = 0644,
 695                 .proc_handler   = &proc_dointvec_minmax,
 696                 .strategy       = sysctl_intvec,
 697                 .extra1         = &pid_max_min,
 698                 .extra2         = &pid_max_max,
 699         },
 700         {
 701                 .ctl_name       = KERN_PANIC_ON_OOPS,
 702                 .procname       = "panic_on_oops",
 703                 .data           = &panic_on_oops,
 704                 .maxlen         = sizeof(int),
 705                 .mode           = 0644,
 706                 .proc_handler   = &proc_dointvec,
 707         },
 708 #if defined CONFIG_PRINTK
 709         {
 710                 .ctl_name       = KERN_PRINTK,
 711                 .procname       = "printk",
 712                 .data           = &console_loglevel,
 713                 .maxlen         = 4*sizeof(int),
 714                 .mode           = 0644,
 715                 .proc_handler   = &proc_dointvec,
 716         },
 717         {
 718                 .ctl_name       = KERN_PRINTK_RATELIMIT,
 719                 .procname       = "printk_ratelimit",
 720                 .data           = &printk_ratelimit_state.interval,
 721                 .maxlen         = sizeof(int),
 722                 .mode           = 0644,
 723                 .proc_handler   = &proc_dointvec_jiffies,
 724                 .strategy       = &sysctl_jiffies,
 725         },
 726         {
 727                 .ctl_name       = KERN_PRINTK_RATELIMIT_BURST,
 728                 .procname       = "printk_ratelimit_burst",
 729                 .data           = &printk_ratelimit_state.burst,
 730                 .maxlen         = sizeof(int),
 731                 .mode           = 0644,
 732                 .proc_handler   = &proc_dointvec,
 733         },
 734         {
 735                 .ctl_name       = CTL_UNNUMBERED,
 736                 .procname       = "printk_delay",
 737                 .data           = &printk_delay_msec,
 738                 .maxlen         = sizeof(int),
 739                 .mode           = 0644,
 740                 .proc_handler   = &proc_dointvec_minmax,
 741                 .strategy       = &sysctl_intvec,
 742                 .extra1         = &zero,
 743                 .extra2         = &ten_thousand,
 744         },
 745 #endif
 746         {
 747                 .ctl_name       = KERN_NGROUPS_MAX,
 748                 .procname       = "ngroups_max",
 749                 .data           = &ngroups_max,
 750                 .maxlen         = sizeof (int),
 751                 .mode           = 0444,
 752                 .proc_handler   = &proc_dointvec,
 753         },
 754 #if defined(CONFIG_X86_LOCAL_APIC) && defined(CONFIG_X86)
 755         {
 756                 .ctl_name       = KERN_UNKNOWN_NMI_PANIC,
 757                 .procname       = "unknown_nmi_panic",
 758                 .data           = &unknown_nmi_panic,
 759                 .maxlen         = sizeof (int),
 760                 .mode           = 0644,
 761                 .proc_handler   = &proc_dointvec,
 762         },
 763         {
 764                 .procname       = "nmi_watchdog",
 765                 .data           = &nmi_watchdog_enabled,
 766                 .maxlen         = sizeof (int),
 767                 .mode           = 0644,
 768                 .proc_handler   = &proc_nmi_enabled,
 769         },
 770 #endif
 771 #if defined(CONFIG_X86)
 772         {
 773                 .ctl_name       = KERN_PANIC_ON_NMI,
 774                 .procname       = "panic_on_unrecovered_nmi",
 775                 .data           = &panic_on_unrecovered_nmi,
 776                 .maxlen         = sizeof(int),
 777                 .mode           = 0644,
 778                 .proc_handler   = &proc_dointvec,
 779         },
 780         {
 781                 .ctl_name       = CTL_UNNUMBERED,
 782                 .procname       = "panic_on_io_nmi",
 783                 .data           = &panic_on_io_nmi,
 784                 .maxlen         = sizeof(int),
 785                 .mode           = 0644,
 786                 .proc_handler   = &proc_dointvec,
 787         },
 788         {
 789                 .ctl_name       = KERN_BOOTLOADER_TYPE,
 790                 .procname       = "bootloader_type",
 791                 .data           = &bootloader_type,
 792                 .maxlen         = sizeof (int),
 793                 .mode           = 0444,
 794                 .proc_handler   = &proc_dointvec,
 795         },
 796         {
 797                 .ctl_name       = CTL_UNNUMBERED,
 798                 .procname       = "bootloader_version",
 799                 .data           = &bootloader_version,
 800                 .maxlen         = sizeof (int),
 801                 .mode           = 0444,
 802                 .proc_handler   = &proc_dointvec,
 803         },
 804         {
 805                 .ctl_name       = CTL_UNNUMBERED,
 806                 .procname       = "kstack_depth_to_print",
 807                 .data           = &kstack_depth_to_print,
 808                 .maxlen         = sizeof(int),
 809                 .mode           = 0644,
 810                 .proc_handler   = &proc_dointvec,
 811         },
 812         {
 813                 .ctl_name       = CTL_UNNUMBERED,
 814                 .procname       = "io_delay_type",
 815                 .data           = &io_delay_type,
 816                 .maxlen         = sizeof(int),
 817                 .mode           = 0644,
 818                 .proc_handler   = &proc_dointvec,
 819         },
 820 #endif
 821 #if defined(CONFIG_MMU)
 822         {
 823                 .ctl_name       = KERN_RANDOMIZE,
 824                 .procname       = "randomize_va_space",
 825                 .data           = &randomize_va_space,
 826                 .maxlen         = sizeof(int),
 827                 .mode           = 0644,
 828                 .proc_handler   = &proc_dointvec,
 829         },
 830 #endif
 831 #if defined(CONFIG_S390) && defined(CONFIG_SMP)
 832         {
 833                 .ctl_name       = KERN_SPIN_RETRY,
 834                 .procname       = "spin_retry",
 835                 .data           = &spin_retry,
 836                 .maxlen         = sizeof (int),
 837                 .mode           = 0644,
 838                 .proc_handler   = &proc_dointvec,
 839         },
 840 #endif
 841 #if     defined(CONFIG_ACPI_SLEEP) && defined(CONFIG_X86)
 842         {
 843                 .procname       = "acpi_video_flags",
 844                 .data           = &acpi_realmode_flags,
 845                 .maxlen         = sizeof (unsigned long),
 846                 .mode           = 0644,
 847                 .proc_handler   = &proc_doulongvec_minmax,
 848         },
 849 #endif
 850 #ifdef CONFIG_IA64
 851         {
 852                 .ctl_name       = KERN_IA64_UNALIGNED,
 853                 .procname       = "ignore-unaligned-usertrap",
 854                 .data           = &no_unaligned_warning,
 855                 .maxlen         = sizeof (int),
 856                 .mode           = 0644,
 857                 .proc_handler   = &proc_dointvec,
 858         },
 859         {
 860                 .ctl_name       = CTL_UNNUMBERED,
 861                 .procname       = "unaligned-dump-stack",
 862                 .data           = &unaligned_dump_stack,
 863                 .maxlen         = sizeof (int),
 864                 .mode           = 0644,
 865                 .proc_handler   = &proc_dointvec,
 866         },
 867 #endif
 868 #ifdef CONFIG_DETECT_SOFTLOCKUP
 869         {
 870                 .ctl_name       = CTL_UNNUMBERED,
 871                 .procname       = "softlockup_panic",
 872                 .data           = &softlockup_panic,
 873                 .maxlen         = sizeof(int),
 874                 .mode           = 0644,
 875                 .proc_handler   = &proc_dointvec_minmax,
 876                 .strategy       = &sysctl_intvec,
 877                 .extra1         = &zero,
 878                 .extra2         = &one,
 879         },
 880         {
 881                 .ctl_name       = CTL_UNNUMBERED,
 882                 .procname       = "softlockup_thresh",
 883                 .data           = &softlockup_thresh,
 884                 .maxlen         = sizeof(int),
 885                 .mode           = 0644,
 886                 .proc_handler   = &proc_dosoftlockup_thresh,
 887                 .strategy       = &sysctl_intvec,
 888                 .extra1         = &neg_one,
 889                 .extra2         = &sixty,
 890         },
 891 #endif
 892 #ifdef CONFIG_DETECT_HUNG_TASK
 893         {
 894                 .ctl_name       = CTL_UNNUMBERED,
 895                 .procname       = "hung_task_panic",
 896                 .data           = &sysctl_hung_task_panic,
 897                 .maxlen         = sizeof(int),
 898                 .mode           = 0644,
 899                 .proc_handler   = &proc_dointvec_minmax,
 900                 .strategy       = &sysctl_intvec,
 901                 .extra1         = &zero,
 902                 .extra2         = &one,
 903         },
 904         {
 905                 .ctl_name       = CTL_UNNUMBERED,
 906                 .procname       = "hung_task_check_count",
 907                 .data           = &sysctl_hung_task_check_count,
 908                 .maxlen         = sizeof(unsigned long),
 909                 .mode           = 0644,
 910                 .proc_handler   = &proc_doulongvec_minmax,
 911                 .strategy       = &sysctl_intvec,
 912         },
 913         {
 914                 .ctl_name       = CTL_UNNUMBERED,
 915                 .procname       = "hung_task_timeout_secs",
 916                 .data           = &sysctl_hung_task_timeout_secs,
 917                 .maxlen         = sizeof(unsigned long),
 918                 .mode           = 0644,
 919                 .proc_handler   = &proc_dohung_task_timeout_secs,
 920                 .strategy       = &sysctl_intvec,
 921         },
 922         {
 923                 .ctl_name       = CTL_UNNUMBERED,
 924                 .procname       = "hung_task_warnings",
 925                 .data           = &sysctl_hung_task_warnings,
 926                 .maxlen         = sizeof(unsigned long),
 927                 .mode           = 0644,
 928                 .proc_handler   = &proc_doulongvec_minmax,
 929                 .strategy       = &sysctl_intvec,
 930         },
 931 #endif
 932 #ifdef CONFIG_COMPAT
 933         {
 934                 .ctl_name       = KERN_COMPAT_LOG,
 935                 .procname       = "compat-log",
 936                 .data           = &compat_log,
 937                 .maxlen         = sizeof (int),
 938                 .mode           = 0644,
 939                 .proc_handler   = &proc_dointvec,
 940         },
 941 #endif
 942 #ifdef CONFIG_RT_MUTEXES
 943         {
 944                 .ctl_name       = KERN_MAX_LOCK_DEPTH,
 945                 .procname       = "max_lock_depth",
 946                 .data           = &max_lock_depth,
 947                 .maxlen         = sizeof(int),
 948                 .mode           = 0644,
 949                 .proc_handler   = &proc_dointvec,
 950         },
 951 #endif
 952         {
 953                 .ctl_name       = CTL_UNNUMBERED,
 954                 .procname       = "poweroff_cmd",
 955                 .data           = &poweroff_cmd,
 956                 .maxlen         = POWEROFF_CMD_PATH_LEN,
 957                 .mode           = 0644,
 958                 .proc_handler   = &proc_dostring,
 959                 .strategy       = &sysctl_string,
 960         },
 961 #ifdef CONFIG_KEYS
 962         {
 963                 .ctl_name       = CTL_UNNUMBERED,
 964                 .procname       = "keys",
 965                 .mode           = 0555,
 966                 .child          = key_sysctls,
 967         },
 968 #endif
 969 #ifdef CONFIG_RCU_TORTURE_TEST
 970         {
 971                 .ctl_name       = CTL_UNNUMBERED,
 972                 .procname       = "rcutorture_runnable",
 973                 .data           = &rcutorture_runnable,
 974                 .maxlen         = sizeof(int),
 975                 .mode           = 0644,
 976                 .proc_handler   = &proc_dointvec,
 977         },
 978 #endif
 979 #ifdef CONFIG_SLOW_WORK
 980         {
 981                 .ctl_name       = CTL_UNNUMBERED,
 982                 .procname       = "slow-work",
 983                 .mode           = 0555,
 984                 .child          = slow_work_sysctls,
 985         },
 986 #endif
 987 #ifdef CONFIG_PERF_EVENTS
 988         {
 989                 .ctl_name       = CTL_UNNUMBERED,
 990                 .procname       = "perf_event_paranoid",
 991                 .data           = &sysctl_perf_event_paranoid,
 992                 .maxlen         = sizeof(sysctl_perf_event_paranoid),
 993                 .mode           = 0644,
 994                 .proc_handler   = &proc_dointvec,
 995         },
 996         {
 997                 .ctl_name       = CTL_UNNUMBERED,
 998                 .procname       = "perf_event_mlock_kb",
 999                 .data           = &sysctl_perf_event_mlock,
1000                 .maxlen         = sizeof(sysctl_perf_event_mlock),
1001                 .mode           = 0644,
1002                 .proc_handler   = &proc_dointvec,
1003         },
1004         {
1005                 .ctl_name       = CTL_UNNUMBERED,
1006                 .procname       = "perf_event_max_sample_rate",
1007                 .data           = &sysctl_perf_event_sample_rate,
1008                 .maxlen         = sizeof(sysctl_perf_event_sample_rate),
1009                 .mode           = 0644,
1010                 .proc_handler   = &proc_dointvec,
1011         },
1012 #endif
1013 #ifdef CONFIG_KMEMCHECK
1014         {
1015                 .ctl_name       = CTL_UNNUMBERED,
1016                 .procname       = "kmemcheck",
1017                 .data           = &kmemcheck_enabled,
1018                 .maxlen         = sizeof(int),
1019                 .mode           = 0644,
1020                 .proc_handler   = &proc_dointvec,
1021         },
1022 #endif
1023 #ifdef CONFIG_BLOCK
1024         {
1025                 .ctl_name       = CTL_UNNUMBERED,
1026                 .procname       = "blk_iopoll",
1027                 .data           = &blk_iopoll_enabled,
1028                 .maxlen         = sizeof(int),
1029                 .mode           = 0644,
1030                 .proc_handler   = &proc_dointvec,
1031         },
1032 #endif
1033 /*
1034  * NOTE: do not add new entries to this table unless you have read
1035  * Documentation/sysctl/ctl_unnumbered.txt
1036  */
1037         { .ctl_name = 0 }
1038 };
1039
1040 static struct ctl_table vm_table[] = {
1041         {
1042                 .ctl_name       = VM_OVERCOMMIT_MEMORY,
1043                 .procname       = "overcommit_memory",
1044                 .data           = &sysctl_overcommit_memory,
1045                 .maxlen         = sizeof(sysctl_overcommit_memory),
1046                 .mode           = 0644,
1047                 .proc_handler   = &proc_dointvec,
1048         },
1049         {
1050                 .ctl_name       = VM_PANIC_ON_OOM,
1051                 .procname       = "panic_on_oom",
1052                 .data           = &sysctl_panic_on_oom,
1053                 .maxlen         = sizeof(sysctl_panic_on_oom),
1054                 .mode           = 0644,
1055                 .proc_handler   = &proc_dointvec,
1056         },
1057         {
1058                 .ctl_name       = CTL_UNNUMBERED,
1059                 .procname       = "oom_kill_allocating_task",
1060                 .data           = &sysctl_oom_kill_allocating_task,
1061                 .maxlen         = sizeof(sysctl_oom_kill_allocating_task),
1062                 .mode           = 0644,
1063                 .proc_handler   = &proc_dointvec,
1064         },
1065         {
1066                 .ctl_name       = CTL_UNNUMBERED,
1067                 .procname       = "oom_dump_tasks",
1068                 .data           = &sysctl_oom_dump_tasks,
1069                 .maxlen         = sizeof(sysctl_oom_dump_tasks),
1070                 .mode           = 0644,
1071                 .proc_handler   = &proc_dointvec,
1072         },
1073         {
1074                 .ctl_name       = VM_OVERCOMMIT_RATIO,
1075                 .procname       = "overcommit_ratio",
1076                 .data           = &sysctl_overcommit_ratio,
1077                 .maxlen         = sizeof(sysctl_overcommit_ratio),
1078                 .mode           = 0644,
1079                 .proc_handler   = &proc_dointvec,
1080         },
1081         {
1082                 .ctl_name       = VM_PAGE_CLUSTER,
1083                 .procname       = "page-cluster",
1084                 .data           = &page_cluster,
1085                 .maxlen         = sizeof(int),
1086                 .mode           = 0644,
1087                 .proc_handler   = &proc_dointvec,
1088         },
1089         {
1090                 .ctl_name       = VM_DIRTY_BACKGROUND,
1091                 .procname       = "dirty_background_ratio",
1092                 .data           = &dirty_background_ratio,
1093                 .maxlen         = sizeof(dirty_background_ratio),
1094                 .mode           = 0644,
1095                 .proc_handler   = &dirty_background_ratio_handler,
1096                 .strategy       = &sysctl_intvec,
1097                 .extra1         = &zero,
1098                 .extra2         = &one_hundred,
1099         },
1100         {
1101                 .ctl_name       = CTL_UNNUMBERED,
1102                 .procname       = "dirty_background_bytes",
1103                 .data           = &dirty_background_bytes,
1104                 .maxlen         = sizeof(dirty_background_bytes),
1105                 .mode           = 0644,
1106                 .proc_handler   = &dirty_background_bytes_handler,
1107                 .strategy       = &sysctl_intvec,
1108                 .extra1         = &one_ul,
1109         },
1110         {
1111                 .ctl_name       = VM_DIRTY_RATIO,
1112                 .procname       = "dirty_ratio",
1113                 .data           = &vm_dirty_ratio,
1114                 .maxlen         = sizeof(vm_dirty_ratio),
1115                 .mode           = 0644,
1116                 .proc_handler   = &dirty_ratio_handler,
1117                 .strategy       = &sysctl_intvec,
1118                 .extra1         = &zero,
1119                 .extra2         = &one_hundred,
1120         },
1121         {
1122                 .ctl_name       = CTL_UNNUMBERED,
1123                 .procname       = "dirty_bytes",
1124                 .data           = &vm_dirty_bytes,
1125                 .maxlen         = sizeof(vm_dirty_bytes),
1126                 .mode           = 0644,
1127                 .proc_handler   = &dirty_bytes_handler,
1128                 .strategy       = &sysctl_intvec,
1129                 .extra1         = &dirty_bytes_min,
1130         },
1131         {
1132                 .procname       = "dirty_writeback_centisecs",
1133                 .data           = &dirty_writeback_interval,
1134                 .maxlen         = sizeof(dirty_writeback_interval),
1135                 .mode           = 0644,
1136                 .proc_handler   = &dirty_writeback_centisecs_handler,
1137         },
1138         {
1139                 .procname       = "dirty_expire_centisecs",
1140                 .data           = &dirty_expire_interval,
1141                 .maxlen         = sizeof(dirty_expire_interval),
1142                 .mode           = 0644,
1143                 .proc_handler   = &proc_dointvec,
1144         },
1145         {
1146                 .ctl_name       = VM_NR_PDFLUSH_THREADS,
1147                 .procname       = "nr_pdflush_threads",
1148                 .data           = &nr_pdflush_threads,
1149                 .maxlen         = sizeof nr_pdflush_threads,
1150                 .mode           = 0444 /* read-only*/,
1151                 .proc_handler   = &proc_dointvec,
1152         },
1153         {
1154                 .ctl_name       = VM_SWAPPINESS,
1155                 .procname       = "swappiness",
1156                 .data           = &vm_swappiness,
1157                 .maxlen         = sizeof(vm_swappiness),
1158                 .mode           = 0644,
1159                 .proc_handler   = &proc_dointvec_minmax,
1160                 .strategy       = &sysctl_intvec,
1161                 .extra1         = &zero,
1162                 .extra2         = &one_hundred,
1163         },
1164 #ifdef CONFIG_HUGETLB_PAGE
1165          {
1166                 .procname       = "nr_hugepages",
1167                 .data           = NULL,
1168                 .maxlen         = sizeof(unsigned long),
1169                 .mode           = 0644,
1170                 .proc_handler   = &hugetlb_sysctl_handler,
1171                 .extra1         = (void *)&hugetlb_zero,
1172                 .extra2         = (void *)&hugetlb_infinity,
1173          },
1174          {
1175                 .ctl_name       = VM_HUGETLB_GROUP,
1176                 .procname       = "hugetlb_shm_group",
1177                 .data           = &sysctl_hugetlb_shm_group,
1178                 .maxlen         = sizeof(gid_t),
1179                 .mode           = 0644,
1180                 .proc_handler   = &proc_dointvec,
1181          },
1182          {
1183                 .ctl_name       = CTL_UNNUMBERED,
1184                 .procname       = "hugepages_treat_as_movable",
1185                 .data           = &hugepages_treat_as_movable,
1186                 .maxlen         = sizeof(int),
1187                 .mode           = 0644,
1188                 .proc_handler   = &hugetlb_treat_movable_handler,
1189         },
1190         {
1191                 .ctl_name       = CTL_UNNUMBERED,
1192                 .procname       = "nr_overcommit_hugepages",
1193                 .data           = NULL,
1194                 .maxlen         = sizeof(unsigned long),
1195                 .mode           = 0644,
1196                 .proc_handler   = &hugetlb_overcommit_handler,
1197                 .extra1         = (void *)&hugetlb_zero,
1198                 .extra2         = (void *)&hugetlb_infinity,
1199         },
1200 #endif
1201         {
1202                 .ctl_name       = VM_LOWMEM_RESERVE_RATIO,
1203                 .procname       = "lowmem_reserve_ratio",
1204                 .data           = &sysctl_lowmem_reserve_ratio,
1205                 .maxlen         = sizeof(sysctl_lowmem_reserve_ratio),
1206                 .mode           = 0644,
1207                 .proc_handler   = &lowmem_reserve_ratio_sysctl_handler,
1208                 .strategy       = &sysctl_intvec,
1209         },
1210         {
1211                 .ctl_name       = VM_DROP_PAGECACHE,
1212                 .procname       = "drop_caches",
1213                 .data           = &sysctl_drop_caches,
1214                 .maxlen         = sizeof(int),
1215                 .mode           = 0644,
1216                 .proc_handler   = drop_caches_sysctl_handler,
1217                 .strategy       = &sysctl_intvec,
1218         },
1219         {
1220                 .ctl_name       = VM_MIN_FREE_KBYTES,
1221                 .procname       = "min_free_kbytes",
1222                 .data           = &min_free_kbytes,
1223                 .maxlen         = sizeof(min_free_kbytes),
1224                 .mode           = 0644,
1225                 .proc_handler   = &min_free_kbytes_sysctl_handler,
1226                 .strategy       = &sysctl_intvec,
1227                 .extra1         = &zero,
1228         },
1229         {
1230                 .ctl_name       = VM_PERCPU_PAGELIST_FRACTION,
1231                 .procname       = "percpu_pagelist_fraction",
1232                 .data           = &percpu_pagelist_fraction,
1233                 .maxlen         = sizeof(percpu_pagelist_fraction),
1234                 .mode           = 0644,
1235                 .proc_handler   = &percpu_pagelist_fraction_sysctl_handler,
1236                 .strategy       = &sysctl_intvec,
1237                 .extra1         = &min_percpu_pagelist_fract,
1238         },
1239 #ifdef CONFIG_MMU
1240         {
1241                 .ctl_name       = VM_MAX_MAP_COUNT,
1242                 .procname       = "max_map_count",
1243                 .data           = &sysctl_max_map_count,
1244                 .maxlen         = sizeof(sysctl_max_map_count),
1245                 .mode           = 0644,
1246                 .proc_handler   = &proc_dointvec
1247         },
1248 #else
1249         {
1250                 .ctl_name       = CTL_UNNUMBERED,
1251                 .procname       = "nr_trim_pages",
1252                 .data           = &sysctl_nr_trim_pages,
1253                 .maxlen         = sizeof(sysctl_nr_trim_pages),
1254                 .mode           = 0644,
1255                 .proc_handler   = &proc_dointvec_minmax,
1256                 .strategy       = &sysctl_intvec,
1257                 .extra1         = &zero,
1258         },
1259 #endif
1260         {
1261                 .ctl_name       = VM_LAPTOP_MODE,
1262                 .procname       = "laptop_mode",
1263                 .data           = &laptop_mode,
1264                 .maxlen         = sizeof(laptop_mode),
1265                 .mode           = 0644,
1266                 .proc_handler   = &proc_dointvec_jiffies,
1267                 .strategy       = &sysctl_jiffies,
1268         },
1269         {
1270                 .ctl_name       = VM_BLOCK_DUMP,
1271                 .procname       = "block_dump",
1272                 .data           = &block_dump,
1273                 .maxlen         = sizeof(block_dump),
1274                 .mode           = 0644,
1275                 .proc_handler   = &proc_dointvec,
1276                 .strategy       = &sysctl_intvec,
1277                 .extra1         = &zero,
1278         },
1279         {
1280                 .ctl_name       = VM_VFS_CACHE_PRESSURE,
1281                 .procname       = "vfs_cache_pressure",
1282                 .data           = &sysctl_vfs_cache_pressure,
1283                 .maxlen         = sizeof(sysctl_vfs_cache_pressure),
1284                 .mode           = 0644,
1285                 .proc_handler   = &proc_dointvec,
1286                 .strategy       = &sysctl_intvec,
1287                 .extra1         = &zero,
1288         },
1289 #ifdef HAVE_ARCH_PICK_MMAP_LAYOUT
1290         {
1291                 .ctl_name       = VM_LEGACY_VA_LAYOUT,
1292                 .procname       = "legacy_va_layout",
1293                 .data           = &sysctl_legacy_va_layout,
1294                 .maxlen         = sizeof(sysctl_legacy_va_layout),
1295                 .mode           = 0644,
1296                 .proc_handler   = &proc_dointvec,
1297                 .strategy       = &sysctl_intvec,
1298                 .extra1         = &zero,
1299         },
1300 #endif
1301 #ifdef CONFIG_NUMA
1302         {
1303                 .ctl_name       = VM_ZONE_RECLAIM_MODE,
1304                 .procname       = "zone_reclaim_mode",
1305                 .data           = &zone_reclaim_mode,
1306                 .maxlen         = sizeof(zone_reclaim_mode),
1307                 .mode           = 0644,
1308                 .proc_handler   = &proc_dointvec,
1309                 .strategy       = &sysctl_intvec,
1310                 .extra1         = &zero,
1311         },
1312         {
1313                 .ctl_name       = VM_MIN_UNMAPPED,
1314                 .procname       = "min_unmapped_ratio",
1315                 .data           = &sysctl_min_unmapped_ratio,
1316                 .maxlen         = sizeof(sysctl_min_unmapped_ratio),
1317                 .mode           = 0644,
1318                 .proc_handler   = &sysctl_min_unmapped_ratio_sysctl_handler,
1319                 .strategy       = &sysctl_intvec,
1320                 .extra1         = &zero,
1321                 .extra2         = &one_hundred,
1322         },
1323         {
1324                 .ctl_name       = VM_MIN_SLAB,
1325                 .procname       = "min_slab_ratio",
1326                 .data           = &sysctl_min_slab_ratio,
1327                 .maxlen         = sizeof(sysctl_min_slab_ratio),
1328                 .mode           = 0644,
1329                 .proc_handler   = &sysctl_min_slab_ratio_sysctl_handler,
1330                 .strategy       = &sysctl_intvec,
1331                 .extra1         = &zero,
1332                 .extra2         = &one_hundred,
1333         },
1334 #endif
1335 #ifdef CONFIG_SMP
1336         {
1337                 .ctl_name       = CTL_UNNUMBERED,
1338                 .procname       = "stat_interval",
1339                 .data           = &sysctl_stat_interval,
1340                 .maxlen         = sizeof(sysctl_stat_interval),
1341                 .mode           = 0644,
1342                 .proc_handler   = &proc_dointvec_jiffies,
1343                 .strategy       = &sysctl_jiffies,
1344         },
1345 #endif
1346         {
1347                 .ctl_name       = CTL_UNNUMBERED,
1348                 .procname       = "mmap_min_addr",
1349                 .data           = &dac_mmap_min_addr,
1350                 .maxlen         = sizeof(unsigned long),
1351                 .mode           = 0644,
1352                 .proc_handler   = &mmap_min_addr_handler,
1353         },
1354 #ifdef CONFIG_NUMA
1355         {
1356                 .ctl_name       = CTL_UNNUMBERED,
1357                 .procname       = "numa_zonelist_order",
1358                 .data           = &numa_zonelist_order,
1359                 .maxlen         = NUMA_ZONELIST_ORDER_LEN,
1360                 .mode           = 0644,
1361                 .proc_handler   = &numa_zonelist_order_handler,
1362                 .strategy       = &sysctl_string,
1363         },
1364 #endif
1365 #if (defined(CONFIG_X86_32) && !defined(CONFIG_UML))|| \
1366    (defined(CONFIG_SUPERH) && defined(CONFIG_VSYSCALL))
1367         {
1368                 .ctl_name       = VM_VDSO_ENABLED,
1369                 .procname       = "vdso_enabled",
1370                 .data           = &vdso_enabled,
1371                 .maxlen         = sizeof(vdso_enabled),
1372                 .mode           = 0644,
1373                 .proc_handler   = &proc_dointvec,
1374                 .strategy       = &sysctl_intvec,
1375                 .extra1         = &zero,
1376         },
1377 #endif
1378 #ifdef CONFIG_HIGHMEM
1379         {
1380                 .ctl_name       = CTL_UNNUMBERED,
1381                 .procname       = "highmem_is_dirtyable",
1382                 .data           = &vm_highmem_is_dirtyable,
1383                 .maxlen         = sizeof(vm_highmem_is_dirtyable),
1384                 .mode           = 0644,
1385                 .proc_handler   = &proc_dointvec_minmax,
1386                 .strategy       = &sysctl_intvec,
1387                 .extra1         = &zero,
1388                 .extra2         = &one,
1389         },
1390 #endif
1391         {
1392                 .ctl_name       = CTL_UNNUMBERED,
1393                 .procname       = "scan_unevictable_pages",
1394                 .data           = &scan_unevictable_pages,
1395                 .maxlen         = sizeof(scan_unevictable_pages),
1396                 .mode           = 0644,
1397                 .proc_handler   = &scan_unevictable_handler,
1398         },
1399 #ifdef CONFIG_MEMORY_FAILURE
1400         {
1401                 .ctl_name       = CTL_UNNUMBERED,
1402                 .procname       = "memory_failure_early_kill",
1403                 .data           = &sysctl_memory_failure_early_kill,
1404                 .maxlen         = sizeof(sysctl_memory_failure_early_kill),
1405                 .mode           = 0644,
1406                 .proc_handler   = &proc_dointvec_minmax,
1407                 .strategy       = &sysctl_intvec,
1408                 .extra1         = &zero,
1409                 .extra2         = &one,
1410         },
1411         {
1412                 .ctl_name       = CTL_UNNUMBERED,
1413                 .procname       = "memory_failure_recovery",
1414                 .data           = &sysctl_memory_failure_recovery,
1415                 .maxlen         = sizeof(sysctl_memory_failure_recovery),
1416                 .mode           = 0644,
1417                 .proc_handler   = &proc_dointvec_minmax,
1418                 .strategy       = &sysctl_intvec,
1419                 .extra1         = &zero,
1420                 .extra2         = &one,
1421         },
1422 #endif
1423
1424 /*
1425  * NOTE: do not add new entries to this table unless you have read
1426  * Documentation/sysctl/ctl_unnumbered.txt
1427  */
1428         { .ctl_name = 0 }
1429 };
1430
1431 #if defined(CONFIG_BINFMT_MISC) || defined(CONFIG_BINFMT_MISC_MODULE)
1432 static struct ctl_table binfmt_misc_table[] = {
1433         { .ctl_name = 0 }
1434 };
1435 #endif
1436
1437 static struct ctl_table fs_table[] = {
1438         {
1439                 .ctl_name       = FS_NRINODE,
1440                 .procname       = "inode-nr",
1441                 .data           = &inodes_stat,
1442                 .maxlen         = 2*sizeof(int),
1443                 .mode           = 0444,
1444                 .proc_handler   = &proc_dointvec,
1445         },
1446         {
1447                 .ctl_name       = FS_STATINODE,
1448                 .procname       = "inode-state",
1449                 .data           = &inodes_stat,
1450                 .maxlen         = 7*sizeof(int),
1451                 .mode           = 0444,
1452                 .proc_handler   = &proc_dointvec,
1453         },
1454         {
1455                 .procname       = "file-nr",
1456                 .data           = &files_stat,
1457                 .maxlen         = 3*sizeof(int),
1458                 .mode           = 0444,
1459                 .proc_handler   = &proc_nr_files,
1460         },
1461         {
1462                 .ctl_name       = FS_MAXFILE,
1463                 .procname       = "file-max",
1464                 .data           = &files_stat.max_files,
1465                 .maxlen         = sizeof(int),
1466                 .mode           = 0644,
1467                 .proc_handler   = &proc_dointvec,
1468         },
1469         {
1470                 .ctl_name       = CTL_UNNUMBERED,
1471                 .procname       = "nr_open",
1472                 .data           = &sysctl_nr_open,
1473                 .maxlen         = sizeof(int),
1474                 .mode           = 0644,
1475                 .proc_handler   = &proc_dointvec_minmax,
1476                 .extra1         = &sysctl_nr_open_min,
1477                 .extra2         = &sysctl_nr_open_max,
1478         },
1479         {
1480                 .ctl_name       = FS_DENTRY,
1481                 .procname       = "dentry-state",
1482                 .data           = &dentry_stat,
1483                 .maxlen         = 6*sizeof(int),
1484                 .mode           = 0444,
1485                 .proc_handler   = &proc_dointvec,
1486         },
1487         {
1488                 .ctl_name       = FS_OVERFLOWUID,
1489                 .procname       = "overflowuid",
1490                 .data           = &fs_overflowuid,
1491                 .maxlen         = sizeof(int),
1492                 .mode           = 0644,
1493                 .proc_handler   = &proc_dointvec_minmax,
1494                 .strategy       = &sysctl_intvec,
1495                 .extra1         = &minolduid,
1496                 .extra2         = &maxolduid,
1497         },
1498         {
1499                 .ctl_name       = FS_OVERFLOWGID,
1500                 .procname       = "overflowgid",
1501                 .data           = &fs_overflowgid,
1502                 .maxlen         = sizeof(int),
1503                 .mode           = 0644,
1504                 .proc_handler   = &proc_dointvec_minmax,
1505                 .strategy       = &sysctl_intvec,
1506                 .extra1         = &minolduid,
1507                 .extra2         = &maxolduid,
1508         },
1509 #ifdef CONFIG_FILE_LOCKING
1510         {
1511                 .ctl_name       = FS_LEASES,
1512                 .procname       = "leases-enable",
1513                 .data           = &leases_enable,
1514                 .maxlen         = sizeof(int),
1515                 .mode           = 0644,
1516                 .proc_handler   = &proc_dointvec,
1517         },
1518 #endif
1519 #ifdef CONFIG_DNOTIFY
1520         {
1521                 .ctl_name       = FS_DIR_NOTIFY,
1522                 .procname       = "dir-notify-enable",
1523                 .data           = &dir_notify_enable,
1524                 .maxlen         = sizeof(int),
1525                 .mode           = 0644,
1526                 .proc_handler   = &proc_dointvec,
1527         },
1528 #endif
1529 #ifdef CONFIG_MMU
1530 #ifdef CONFIG_FILE_LOCKING
1531         {
1532                 .ctl_name       = FS_LEASE_TIME,
1533                 .procname       = "lease-break-time",
1534                 .data           = &lease_break_time,
1535                 .maxlen         = sizeof(int),
1536                 .mode           = 0644,
1537                 .proc_handler   = &proc_dointvec,
1538         },
1539 #endif
1540 #ifdef CONFIG_AIO
1541         {
1542                 .procname       = "aio-nr",
1543                 .data           = &aio_nr,
1544                 .maxlen         = sizeof(aio_nr),
1545                 .mode           = 0444,
1546                 .proc_handler   = &proc_doulongvec_minmax,
1547         },
1548         {
1549                 .procname       = "aio-max-nr",
1550                 .data           = &aio_max_nr,
1551                 .maxlen         = sizeof(aio_max_nr),
1552                 .mode           = 0644,
1553                 .proc_handler   = &proc_doulongvec_minmax,
1554         },
1555 #endif /* CONFIG_AIO */
1556 #ifdef CONFIG_INOTIFY_USER
1557         {
1558                 .ctl_name       = FS_INOTIFY,
1559                 .procname       = "inotify",
1560                 .mode           = 0555,
1561                 .child          = inotify_table,
1562         },
1563 #endif
1564 #ifdef CONFIG_EPOLL
1565         {
1566                 .procname       = "epoll",
1567                 .mode           = 0555,
1568                 .child          = epoll_table,
1569         },
1570 #endif
1571 #endif
1572         {
1573                 .ctl_name       = KERN_SETUID_DUMPABLE,
1574                 .procname       = "suid_dumpable",
1575                 .data           = &suid_dumpable,
1576                 .maxlen         = sizeof(int),
1577                 .mode           = 0644,
1578                 .proc_handler   = &proc_dointvec_minmax,
1579                 .strategy       = &sysctl_intvec,
1580                 .extra1         = &zero,
1581                 .extra2         = &two,
1582         },
1583 #if defined(CONFIG_BINFMT_MISC) || defined(CONFIG_BINFMT_MISC_MODULE)
1584         {
1585                 .ctl_name       = CTL_UNNUMBERED,
1586                 .procname       = "binfmt_misc",
1587                 .mode           = 0555,
1588                 .child          = binfmt_misc_table,
1589         },
1590 #endif
1591 /*
1592  * NOTE: do not add new entries to this table unless you have read
1593  * Documentation/sysctl/ctl_unnumbered.txt
1594  */
1595         { .ctl_name = 0 }
1596 };
1597
1598 static struct ctl_table debug_table[] = {
1599 #if defined(CONFIG_X86) || defined(CONFIG_PPC)
1600         {
1601                 .ctl_name       = CTL_UNNUMBERED,
1602                 .procname       = "exception-trace",
1603                 .data           = &show_unhandled_signals,
1604                 .maxlen         = sizeof(int),
1605                 .mode           = 0644,
1606                 .proc_handler   = proc_dointvec
1607         },
1608 #endif
1609         { .ctl_name = 0 }
1610 };
1611
1612 static struct ctl_table dev_table[] = {
1613         { .ctl_name = 0 }
1614 };
1615
1616 static DEFINE_SPINLOCK(sysctl_lock);
1617
1618 /* called under sysctl_lock */
1619 static int use_table(struct ctl_table_header *p)
1620 {
1621         if (unlikely(p->unregistering))
1622                 return 0;
1623         p->used++;
1624         return 1;
1625 }
1626
1627 /* called under sysctl_lock */
1628 static void unuse_table(struct ctl_table_header *p)
1629 {
1630         if (!--p->used)
1631                 if (unlikely(p->unregistering))
1632                         complete(p->unregistering);
1633 }
1634
1635 /* called under sysctl_lock, will reacquire if has to wait */
1636 static void start_unregistering(struct ctl_table_header *p)
1637 {
1638         /*
1639          * if p->used is 0, nobody will ever touch that entry again;
1640          * we'll eliminate all paths to it before dropping sysctl_lock
1641          */
1642         if (unlikely(p->used)) {
1643                 struct completion wait;
1644                 init_completion(&wait);
1645                 p->unregistering = &wait;
1646                 spin_unlock(&sysctl_lock);
1647                 wait_for_completion(&wait);
1648                 spin_lock(&sysctl_lock);
1649         } else {
1650                 /* anything non-NULL; we'll never dereference it */
1651                 p->unregistering = ERR_PTR(-EINVAL);
1652         }
1653         /*
1654          * do not remove from the list until nobody holds it; walking the
1655          * list in do_sysctl() relies on that.
1656          */
1657         list_del_init(&p->ctl_entry);
1658 }
1659
1660 void sysctl_head_get(struct ctl_table_header *head)
1661 {
1662         spin_lock(&sysctl_lock);
1663         head->count++;
1664         spin_unlock(&sysctl_lock);
1665 }
1666
1667 void sysctl_head_put(struct ctl_table_header *head)
1668 {
1669         spin_lock(&sysctl_lock);
1670         if (!--head->count)
1671                 kfree(head);
1672         spin_unlock(&sysctl_lock);
1673 }
1674
1675 struct ctl_table_header *sysctl_head_grab(struct ctl_table_header *head)
1676 {
1677         if (!head)
1678                 BUG();
1679         spin_lock(&sysctl_lock);
1680         if (!use_table(head))
1681                 head = ERR_PTR(-ENOENT);
1682         spin_unlock(&sysctl_lock);
1683         return head;
1684 }
1685
1686 void sysctl_head_finish(struct ctl_table_header *head)
1687 {
1688         if (!head)
1689                 return;
1690         spin_lock(&sysctl_lock);
1691         unuse_table(head);
1692         spin_unlock(&sysctl_lock);
1693 }
1694
1695 static struct ctl_table_set *
1696 lookup_header_set(struct ctl_table_root *root, struct nsproxy *namespaces)
1697 {
1698         struct ctl_table_set *set = &root->default_set;
1699         if (root->lookup)
1700                 set = root->lookup(root, namespaces);
1701         return set;
1702 }
1703
1704 static struct list_head *
1705 lookup_header_list(struct ctl_table_root *root, struct nsproxy *namespaces)
1706 {
1707         struct ctl_table_set *set = lookup_header_set(root, namespaces);
1708         return &set->list;
1709 }
1710
1711 struct ctl_table_header *__sysctl_head_next(struct nsproxy *namespaces,
1712                                             struct ctl_table_header *prev)
1713 {
1714         struct ctl_table_root *root;
1715         struct list_head *header_list;
1716         struct ctl_table_header *head;
1717         struct list_head *tmp;
1718
1719         spin_lock(&sysctl_lock);
1720         if (prev) {
1721                 head = prev;
1722                 tmp = &prev->ctl_entry;
1723                 unuse_table(prev);
1724                 goto next;
1725         }
1726         tmp = &root_table_header.ctl_entry;
1727         for (;;) {
1728                 head = list_entry(tmp, struct ctl_table_header, ctl_entry);
1729
1730                 if (!use_table(head))
1731                         goto next;
1732                 spin_unlock(&sysctl_lock);
1733                 return head;
1734         next:
1735                 root = head->root;
1736                 tmp = tmp->next;
1737                 header_list = lookup_header_list(root, namespaces);
1738                 if (tmp != header_list)
1739                         continue;
1740
1741                 do {
1742                         root = list_entry(root->root_list.next,
1743                                         struct ctl_table_root, root_list);
1744                         if (root == &sysctl_table_root)
1745                                 goto out;
1746                         header_list = lookup_header_list(root, namespaces);
1747                 } while (list_empty(header_list));
1748                 tmp = header_list->next;
1749         }
1750 out:
1751         spin_unlock(&sysctl_lock);
1752         return NULL;
1753 }
1754
1755 struct ctl_table_header *sysctl_head_next(struct ctl_table_header *prev)
1756 {
1757         return __sysctl_head_next(current->nsproxy, prev);
1758 }
1759
1760 void register_sysctl_root(struct ctl_table_root *root)
1761 {
1762         spin_lock(&sysctl_lock);
1763         list_add_tail(&root->root_list, &sysctl_table_root.root_list);
1764         spin_unlock(&sysctl_lock);
1765 }
1766
1767 /*
1768  * sysctl_perm does NOT grant the superuser all rights automatically, because
1769  * some sysctl variables are readonly even to root.
1770  */
1771
1772 static int test_perm(int mode, int op)
1773 {
1774         if (!current_euid())
1775                 mode >>= 6;
1776         else if (in_egroup_p(0))
1777                 mode >>= 3;
1778         if ((op & ~mode & (MAY_READ|MAY_WRITE|MAY_EXEC)) == 0)
1779                 return 0;
1780         return -EACCES;
1781 }
1782
1783 int sysctl_perm(struct ctl_table_root *root, struct ctl_table *table, int op)
1784 {
1785         int error;
1786         int mode;
1787
1788         error = security_sysctl(table, op & (MAY_READ | MAY_WRITE | MAY_EXEC));
1789         if (error)
1790                 return error;
1791
1792         if (root->permissions)
1793                 mode = root->permissions(root, current->nsproxy, table);
1794         else
1795                 mode = table->mode;
1796
1797         return test_perm(mode, op);
1798 }
1799
1800 static void sysctl_set_parent(struct ctl_table *parent, struct ctl_table *table)
1801 {
1802         for (; table->ctl_name || table->procname; table++) {
1803                 table->parent = parent;
1804                 if (table->child)
1805                         sysctl_set_parent(table, table->child);
1806         }
1807 }
1808
1809 static __init int sysctl_init(void)
1810 {
1811         sysctl_set_parent(NULL, root_table);
1812 #ifdef CONFIG_SYSCTL_SYSCALL_CHECK
1813         {
1814                 int err;
1815                 err = sysctl_check_table(current->nsproxy, root_table);
1816         }
1817 #endif
1818         return 0;
1819 }
1820
1821 core_initcall(sysctl_init);
1822
1823 static struct ctl_table *is_branch_in(struct ctl_table *branch,
1824                                       struct ctl_table *table)
1825 {
1826         struct ctl_table *p;
1827         const char *s = branch->procname;
1828
1829         /* branch should have named subdirectory as its first element */
1830         if (!s || !branch->child)
1831                 return NULL;
1832
1833         /* ... and nothing else */
1834         if (branch[1].procname || branch[1].ctl_name)
1835                 return NULL;
1836
1837         /* table should contain subdirectory with the same name */
1838         for (p = table; p->procname || p->ctl_name; p++) {
1839                 if (!p->child)
1840                         continue;
1841                 if (p->procname && strcmp(p->procname, s) == 0)
1842                         return p;
1843         }
1844         return NULL;
1845 }
1846
1847 /* see if attaching q to p would be an improvement */
1848 static void try_attach(struct ctl_table_header *p, struct ctl_table_header *q)
1849 {
1850         struct ctl_table *to = p->ctl_table, *by = q->ctl_table;
1851         struct ctl_table *next;
1852         int is_better = 0;
1853         int not_in_parent = !p->attached_by;
1854
1855         while ((next = is_branch_in(by, to)) != NULL) {
1856                 if (by == q->attached_by)
1857                         is_better = 1;
1858                 if (to == p->attached_by)
1859                         not_in_parent = 1;
1860                 by = by->child;
1861                 to = next->child;
1862         }
1863
1864         if (is_better && not_in_parent) {
1865                 q->attached_by = by;
1866                 q->attached_to = to;
1867                 q->parent = p;
1868         }
1869 }
1870
1871 /**
1872  * __register_sysctl_paths - register a sysctl hierarchy
1873  * @root: List of sysctl headers to register on
1874  * @namespaces: Data to compute which lists of sysctl entries are visible
1875  * @path: The path to the directory the sysctl table is in.
1876  * @table: the top-level table structure
1877  *
1878  * Register a sysctl table hierarchy. @table should be a filled in ctl_table
1879  * array. A completely 0 filled entry terminates the table.
1880  *
1881  * The members of the &struct ctl_table structure are used as follows:
1882  *
1883  * ctl_name - This is the numeric sysctl value used by sysctl(2). The number
1884  *            must be unique within that level of sysctl
1885  *
1886  * procname - the name of the sysctl file under /proc/sys. Set to %NULL to not
1887  *            enter a sysctl file
1888  *
1889  * data - a pointer to data for use by proc_handler
1890  *
1891  * maxlen - the maximum size in bytes of the data
1892  *
1893  * mode - the file permissions for the /proc/sys file, and for sysctl(2)
1894  *
1895  * child - a pointer to the child sysctl table if this entry is a directory, or
1896  *         %NULL.
1897  *
1898  * proc_handler - the text handler routine (described below)
1899  *
1900  * strategy - the strategy routine (described below)
1901  *
1902  * de - for internal use by the sysctl routines
1903  *
1904  * extra1, extra2 - extra pointers usable by the proc handler routines
1905  *
1906  * Leaf nodes in the sysctl tree will be represented by a single file
1907  * under /proc; non-leaf nodes will be represented by directories.
1908  *
1909  * sysctl(2) can automatically manage read and write requests through
1910  * the sysctl table.  The data and maxlen fields of the ctl_table
1911  * struct enable minimal validation of the values being written to be
1912  * performed, and the mode field allows minimal authentication.
1913  *
1914  * More sophisticated management can be enabled by the provision of a
1915  * strategy routine with the table entry.  This will be called before
1916  * any automatic read or write of the data is performed.
1917  *
1918  * The strategy routine may return
1919  *
1920  * < 0 - Error occurred (error is passed to user process)
1921  *
1922  * 0   - OK - proceed with automatic read or write.
1923  *
1924  * > 0 - OK - read or write has been done by the strategy routine, so
1925  *       return immediately.
1926  *
1927  * There must be a proc_handler routine for any terminal nodes
1928  * mirrored under /proc/sys (non-terminals are handled by a built-in
1929  * directory handler).  Several default handlers are available to
1930  * cover common cases -
1931  *
1932  * proc_dostring(), proc_dointvec(), proc_dointvec_jiffies(),
1933  * proc_dointvec_userhz_jiffies(), proc_dointvec_minmax(),
1934  * proc_doulongvec_ms_jiffies_minmax(), proc_doulongvec_minmax()
1935  *
1936  * It is the handler's job to read the input buffer from user memory
1937  * and process it. The handler should return 0 on success.
1938  *
1939  * This routine returns %NULL on a failure to register, and a pointer
1940  * to the table header on success.
1941  */
1942 struct ctl_table_header *__register_sysctl_paths(
1943         struct ctl_table_root *root,
1944         struct nsproxy *namespaces,
1945         const struct ctl_path *path, struct ctl_table *table)
1946 {
1947         struct ctl_table_header *header;
1948         struct ctl_table *new, **prevp;
1949         unsigned int n, npath;
1950         struct ctl_table_set *set;
1951
1952         /* Count the path components */
1953         for (npath = 0; path[npath].ctl_name || path[npath].procname; ++npath)
1954                 ;
1955
1956         /*
1957          * For each path component, allocate a 2-element ctl_table array.
1958          * The first array element will be filled with the sysctl entry
1959          * for this, the second will be the sentinel (ctl_name == 0).
1960          *
1961          * We allocate everything in one go so that we don't have to
1962          * worry about freeing additional memory in unregister_sysctl_table.
1963          */
1964         header = kzalloc(sizeof(struct ctl_table_header) +
1965                          (2 * npath * sizeof(struct ctl_table)), GFP_KERNEL);
1966         if (!header)
1967                 return NULL;
1968
1969         new = (struct ctl_table *) (header + 1);
1970
1971         /* Now connect the dots */
1972         prevp = &header->ctl_table;
1973         for (n = 0; n < npath; ++n, ++path) {
1974                 /* Copy the procname */
1975                 new->procname = path->procname;
1976                 new->ctl_name = path->ctl_name;
1977                 new->mode     = 0555;
1978
1979                 *prevp = new;
1980                 prevp = &new->child;
1981
1982                 new += 2;
1983         }
1984         *prevp = table;
1985         header->ctl_table_arg = table;
1986
1987         INIT_LIST_HEAD(&header->ctl_entry);
1988         header->used = 0;
1989         header->unregistering = NULL;
1990         header->root = root;
1991         sysctl_set_parent(NULL, header->ctl_table);
1992         header->count = 1;
1993 #ifdef CONFIG_SYSCTL_SYSCALL_CHECK
1994         if (sysctl_check_table(namespaces, header->ctl_table)) {
1995                 kfree(header);
1996                 return NULL;
1997         }
1998 #endif
1999         spin_lock(&sysctl_lock);
2000         header->set = lookup_header_set(root, namespaces);
2001         header->attached_by = header->ctl_table;
2002         header->attached_to = root_table;
2003         header->parent = &root_table_header;
2004         for (set = header->set; set; set = set->parent) {
2005                 struct ctl_table_header *p;
2006                 list_for_each_entry(p, &set->list, ctl_entry) {
2007                         if (p->unregistering)
2008                                 continue;
2009                         try_attach(p, header);
2010                 }
2011         }
2012         header->parent->count++;
2013         list_add_tail(&header->ctl_entry, &header->set->list);
2014         spin_unlock(&sysctl_lock);
2015
2016         return header;
2017 }
2018
2019 /**
2020  * register_sysctl_table_path - register a sysctl table hierarchy
2021  * @path: The path to the directory the sysctl table is in.
2022  * @table: the top-level table structure
2023  *
2024  * Register a sysctl table hierarchy. @table should be a filled in ctl_table
2025  * array. A completely 0 filled entry terminates the table.
2026  *
2027  * See __register_sysctl_paths for more details.
2028  */
2029 struct ctl_table_header *register_sysctl_paths(const struct ctl_path *path,
2030                                                 struct ctl_table *table)
2031 {
2032         return __register_sysctl_paths(&sysctl_table_root, current->nsproxy,
2033                                         path, table);
2034 }
2035
2036 /**
2037  * register_sysctl_table - register a sysctl table hierarchy
2038  * @table: the top-level table structure
2039  *
2040  * Register a sysctl table hierarchy. @table should be a filled in ctl_table
2041  * array. A completely 0 filled entry terminates the table.
2042  *
2043  * See register_sysctl_paths for more details.
2044  */
2045 struct ctl_table_header *register_sysctl_table(struct ctl_table *table)
2046 {
2047         static const struct ctl_path null_path[] = { {} };
2048
2049         return register_sysctl_paths(null_path, table);
2050 }
2051
2052 /**
2053  * unregister_sysctl_table - unregister a sysctl table hierarchy
2054  * @header: the header returned from register_sysctl_table
2055  *
2056  * Unregisters the sysctl table and all children. proc entries may not
2057  * actually be removed until they are no longer used by anyone.
2058  */
2059 void unregister_sysctl_table(struct ctl_table_header * header)
2060 {
2061         might_sleep();
2062
2063         if (header == NULL)
2064                 return;
2065
2066         spin_lock(&sysctl_lock);
2067         start_unregistering(header);
2068         if (!--header->parent->count) {
2069                 WARN_ON(1);
2070                 kfree(header->parent);
2071         }
2072         if (!--header->count)
2073                 kfree(header);
2074         spin_unlock(&sysctl_lock);
2075 }
2076
2077 int sysctl_is_seen(struct ctl_table_header *p)
2078 {
2079         struct ctl_table_set *set = p->set;
2080         int res;
2081         spin_lock(&sysctl_lock);
2082         if (p->unregistering)
2083                 res = 0;
2084         else if (!set->is_seen)
2085                 res = 1;
2086         else
2087                 res = set->is_seen(set);
2088         spin_unlock(&sysctl_lock);
2089         return res;
2090 }
2091
2092 void setup_sysctl_set(struct ctl_table_set *p,
2093         struct ctl_table_set *parent,
2094         int (*is_seen)(struct ctl_table_set *))
2095 {
2096         INIT_LIST_HEAD(&p->list);
2097         p->parent = parent ? parent : &sysctl_table_root.default_set;
2098         p->is_seen = is_seen;
2099 }
2100
2101 #else /* !CONFIG_SYSCTL */
2102 struct ctl_table_header *register_sysctl_table(struct ctl_table * table)
2103 {
2104         return NULL;
2105 }
2106
2107 struct ctl_table_header *register_sysctl_paths(const struct ctl_path *path,
2108                                                     struct ctl_table *table)
2109 {
2110         return NULL;
2111 }
2112
2113 void unregister_sysctl_table(struct ctl_table_header * table)
2114 {
2115 }
2116
2117 void setup_sysctl_set(struct ctl_table_set *p,
2118         struct ctl_table_set *parent,
2119         int (*is_seen)(struct ctl_table_set *))
2120 {
2121 }
2122
2123 void sysctl_head_put(struct ctl_table_header *head)
2124 {
2125 }
2126
2127 #endif /* CONFIG_SYSCTL */
2128
2129 /*
2130  * /proc/sys support
2131  */
2132
2133 #ifdef CONFIG_PROC_SYSCTL
2134
2135 static int _proc_do_string(void* data, int maxlen, int write,
2136                            void __user *buffer,
2137                            size_t *lenp, loff_t *ppos)
2138 {
2139         size_t len;
2140         char __user *p;
2141         char c;
2142
2143         if (!data || !maxlen || !*lenp) {
2144                 *lenp = 0;
2145                 return 0;
2146         }
2147
2148         if (write) {
2149                 len = 0;
2150                 p = buffer;
2151                 while (len < *lenp) {
2152                         if (get_user(c, p++))
2153                                 return -EFAULT;
2154                         if (c == 0 || c == '\n')
2155                                 break;
2156                         len++;
2157                 }
2158                 if (len >= maxlen)
2159                         len = maxlen-1;
2160                 if(copy_from_user(data, buffer, len))
2161                         return -EFAULT;
2162                 ((char *) data)[len] = 0;
2163                 *ppos += *lenp;
2164         } else {
2165                 len = strlen(data);
2166                 if (len > maxlen)
2167                         len = maxlen;
2168
2169                 if (*ppos > len) {
2170                         *lenp = 0;
2171                         return 0;
2172                 }
2173
2174                 data += *ppos;
2175                 len  -= *ppos;
2176
2177                 if (len > *lenp)
2178                         len = *lenp;
2179                 if (len)
2180                         if(copy_to_user(buffer, data, len))
2181                                 return -EFAULT;
2182                 if (len < *lenp) {
2183                         if(put_user('\n', ((char __user *) buffer) + len))
2184                                 return -EFAULT;
2185                         len++;
2186                 }
2187                 *lenp = len;
2188                 *ppos += len;
2189         }
2190         return 0;
2191 }
2192
2193 /**
2194  * proc_dostring - read a string sysctl
2195  * @table: the sysctl table
2196  * @write: %TRUE if this is a write to the sysctl file
2197  * @buffer: the user buffer
2198  * @lenp: the size of the user buffer
2199  * @ppos: file position
2200  *
2201  * Reads/writes a string from/to the user buffer. If the kernel
2202  * buffer provided is not large enough to hold the string, the
2203  * string is truncated. The copied string is %NULL-terminated.
2204  * If the string is being read by the user process, it is copied
2205  * and a newline '\n' is added. It is truncated if the buffer is
2206  * not large enough.
2207  *
2208  * Returns 0 on success.
2209  */
2210 int proc_dostring(struct ctl_table *table, int write,
2211                   void __user *buffer, size_t *lenp, loff_t *ppos)
2212 {
2213         return _proc_do_string(table->data, table->maxlen, write,
2214                                buffer, lenp, ppos);
2215 }
2216
2217
2218 static int do_proc_dointvec_conv(int *negp, unsigned long *lvalp,
2219                                  int *valp,
2220                                  int write, void *data)
2221 {
2222         if (write) {
2223                 *valp = *negp ? -*lvalp : *lvalp;
2224         } else {
2225                 int val = *valp;
2226                 if (val < 0) {
2227                         *negp = -1;
2228                         *lvalp = (unsigned long)-val;
2229                 } else {
2230                         *negp = 0;
2231                         *lvalp = (unsigned long)val;
2232                 }
2233         }
2234         return 0;
2235 }
2236
2237 static int __do_proc_dointvec(void *tbl_data, struct ctl_table *table,
2238                   int write, void __user *buffer,
2239                   size_t *lenp, loff_t *ppos,
2240                   int (*conv)(int *negp, unsigned long *lvalp, int *valp,
2241                               int write, void *data),
2242                   void *data)
2243 {
2244 #define TMPBUFLEN 21
2245         int *i, vleft, first = 1, neg;
2246         unsigned long lval;
2247         size_t left, len;
2248
2249         char buf[TMPBUFLEN], *p;
2250         char __user *s = buffer;
2251
2252         if (!tbl_data || !table->maxlen || !*lenp ||
2253             (*ppos && !write)) {
2254                 *lenp = 0;
2255                 return 0;
2256         }
2257
2258         i = (int *) tbl_data;
2259         vleft = table->maxlen / sizeof(*i);
2260         left = *lenp;
2261
2262         if (!conv)
2263                 conv = do_proc_dointvec_conv;
2264
2265         for (; left && vleft--; i++, first=0) {
2266                 if (write) {
2267                         while (left) {
2268                                 char c;
2269                                 if (get_user(c, s))
2270                                         return -EFAULT;
2271                                 if (!isspace(c))
2272                                         break;
2273                                 left--;
2274                                 s++;
2275                         }
2276                         if (!left)
2277                                 break;
2278                         neg = 0;
2279                         len = left;
2280                         if (len > sizeof(buf) - 1)
2281                                 len = sizeof(buf) - 1;
2282                         if (copy_from_user(buf, s, len))
2283                                 return -EFAULT;
2284                         buf[len] = 0;
2285                         p = buf;
2286                         if (*p == '-' && left > 1) {
2287                                 neg = 1;
2288                                 p++;
2289                         }
2290                         if (*p < '0' || *p > '9')
2291                                 break;
2292
2293                         lval = simple_strtoul(p, &p, 0);
2294
2295                         len = p-buf;
2296                         if ((len < left) && *p && !isspace(*p))
2297                                 break;
2298                         s += len;
2299                         left -= len;
2300
2301                         if (conv(&neg, &lval, i, 1, data))
2302                                 break;
2303                 } else {
2304                         p = buf;
2305                         if (!first)
2306                                 *p++ = '\t';
2307
2308                         if (conv(&neg, &lval, i, 0, data))
2309                                 break;
2310
2311                         sprintf(p, "%s%lu", neg ? "-" : "", lval);
2312                         len = strlen(buf);
2313                         if (len > left)
2314                                 len = left;
2315                         if(copy_to_user(s, buf, len))
2316                                 return -EFAULT;
2317                         left -= len;
2318                         s += len;
2319                 }
2320         }
2321
2322         if (!write && !first && left) {
2323                 if(put_user('\n', s))
2324                         return -EFAULT;
2325                 left--, s++;
2326         }
2327         if (write) {
2328                 while (left) {
2329                         char c;
2330                         if (get_user(c, s++))
2331                                 return -EFAULT;
2332                         if (!isspace(c))
2333                                 break;
2334                         left--;
2335                 }
2336         }
2337         if (write && first)
2338                 return -EINVAL;
2339         *lenp -= left;
2340         *ppos += *lenp;
2341         return 0;
2342 #undef TMPBUFLEN
2343 }
2344
2345 static int do_proc_dointvec(struct ctl_table *table, int write,
2346                   void __user *buffer, size_t *lenp, loff_t *ppos,
2347                   int (*conv)(int *negp, unsigned long *lvalp, int *valp,
2348                               int write, void *data),
2349                   void *data)
2350 {
2351         return __do_proc_dointvec(table->data, table, write,
2352                         buffer, lenp, ppos, conv, data);
2353 }
2354
2355 /**
2356  * proc_dointvec - read a vector of integers
2357  * @table: the sysctl table
2358  * @write: %TRUE if this is a write to the sysctl file
2359  * @buffer: the user buffer
2360  * @lenp: the size of the user buffer
2361  * @ppos: file position
2362  *
2363  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2364  * values from/to the user buffer, treated as an ASCII string.
2365  *
2366  * Returns 0 on success.
2367  */
2368 int proc_dointvec(struct ctl_table *table, int write,
2369                      void __user *buffer, size_t *lenp, loff_t *ppos)
2370 {
2371     return do_proc_dointvec(table,write,buffer,lenp,ppos,
2372                             NULL,NULL);
2373 }
2374
2375 /*
2376  * Taint values can only be increased
2377  * This means we can safely use a temporary.
2378  */
2379 static int proc_taint(struct ctl_table *table, int write,
2380                                void __user *buffer, size_t *lenp, loff_t *ppos)
2381 {
2382         struct ctl_table t;
2383         unsigned long tmptaint = get_taint();
2384         int err;
2385
2386         if (write && !capable(CAP_SYS_ADMIN))
2387                 return -EPERM;
2388
2389         t = *table;
2390         t.data = &tmptaint;
2391         err = proc_doulongvec_minmax(&t, write, buffer, lenp, ppos);
2392         if (err < 0)
2393                 return err;
2394
2395         if (write) {
2396                 /*
2397                  * Poor man's atomic or. Not worth adding a primitive
2398                  * to everyone's atomic.h for this
2399                  */
2400                 int i;
2401                 for (i = 0; i < BITS_PER_LONG && tmptaint >> i; i++) {
2402                         if ((tmptaint >> i) & 1)
2403                                 add_taint(i);
2404                 }
2405         }
2406
2407         return err;
2408 }
2409
2410 struct do_proc_dointvec_minmax_conv_param {
2411         int *min;
2412         int *max;
2413 };
2414
2415 static int do_proc_dointvec_minmax_conv(int *negp, unsigned long *lvalp,
2416                                         int *valp,
2417                                         int write, void *data)
2418 {
2419         struct do_proc_dointvec_minmax_conv_param *param = data;
2420         if (write) {
2421                 int val = *negp ? -*lvalp : *lvalp;
2422                 if ((param->min && *param->min > val) ||
2423                     (param->max && *param->max < val))
2424                         return -EINVAL;
2425                 *valp = val;
2426         } else {
2427                 int val = *valp;
2428                 if (val < 0) {
2429                         *negp = -1;
2430                         *lvalp = (unsigned long)-val;
2431                 } else {
2432                         *negp = 0;
2433                         *lvalp = (unsigned long)val;
2434                 }
2435         }
2436         return 0;
2437 }
2438
2439 /**
2440  * proc_dointvec_minmax - read a vector of integers with min/max values
2441  * @table: the sysctl table
2442  * @write: %TRUE if this is a write to the sysctl file
2443  * @buffer: the user buffer
2444  * @lenp: the size of the user buffer
2445  * @ppos: file position
2446  *
2447  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2448  * values from/to the user buffer, treated as an ASCII string.
2449  *
2450  * This routine will ensure the values are within the range specified by
2451  * table->extra1 (min) and table->extra2 (max).
2452  *
2453  * Returns 0 on success.
2454  */
2455 int proc_dointvec_minmax(struct ctl_table *table, int write,
2456                   void __user *buffer, size_t *lenp, loff_t *ppos)
2457 {
2458         struct do_proc_dointvec_minmax_conv_param param = {
2459                 .min = (int *) table->extra1,
2460                 .max = (int *) table->extra2,
2461         };
2462         return do_proc_dointvec(table, write, buffer, lenp, ppos,
2463                                 do_proc_dointvec_minmax_conv, &param);
2464 }
2465
2466 static int __do_proc_doulongvec_minmax(void *data, struct ctl_table *table, int write,
2467                                      void __user *buffer,
2468                                      size_t *lenp, loff_t *ppos,
2469                                      unsigned long convmul,
2470                                      unsigned long convdiv)
2471 {
2472 #define TMPBUFLEN 21
2473         unsigned long *i, *min, *max, val;
2474         int vleft, first=1, neg;
2475         size_t len, left;
2476         char buf[TMPBUFLEN], *p;
2477         char __user *s = buffer;
2478
2479         if (!data || !table->maxlen || !*lenp ||
2480             (*ppos && !write)) {
2481                 *lenp = 0;
2482                 return 0;
2483         }
2484
2485         i = (unsigned long *) data;
2486         min = (unsigned long *) table->extra1;
2487         max = (unsigned long *) table->extra2;
2488         vleft = table->maxlen / sizeof(unsigned long);
2489         left = *lenp;
2490
2491         for (; left && vleft--; i++, min++, max++, first=0) {
2492                 if (write) {
2493                         while (left) {
2494                                 char c;
2495                                 if (get_user(c, s))
2496                                         return -EFAULT;
2497                                 if (!isspace(c))
2498                                         break;
2499                                 left--;
2500                                 s++;
2501                         }
2502                         if (!left)
2503                                 break;
2504                         neg = 0;
2505                         len = left;
2506                         if (len > TMPBUFLEN-1)
2507                                 len = TMPBUFLEN-1;
2508                         if (copy_from_user(buf, s, len))
2509                                 return -EFAULT;
2510                         buf[len] = 0;
2511                         p = buf;
2512                         if (*p == '-' && left > 1) {
2513                                 neg = 1;
2514                                 p++;
2515                         }
2516                         if (*p < '0' || *p > '9')
2517                                 break;
2518                         val = simple_strtoul(p, &p, 0) * convmul / convdiv ;
2519                         len = p-buf;
2520                         if ((len < left) && *p && !isspace(*p))
2521                                 break;
2522                         if (neg)
2523                                 val = -val;
2524                         s += len;
2525                         left -= len;
2526
2527                         if(neg)
2528                                 continue;
2529                         if ((min && val < *min) || (max && val > *max))
2530                                 continue;
2531                         *i = val;
2532                 } else {
2533                         p = buf;
2534                         if (!first)
2535                                 *p++ = '\t';
2536                         sprintf(p, "%lu", convdiv * (*i) / convmul);
2537                         len = strlen(buf);
2538                         if (len > left)
2539                                 len = left;
2540                         if(copy_to_user(s, buf, len))
2541                                 return -EFAULT;
2542                         left -= len;
2543                         s += len;
2544                 }
2545         }
2546
2547         if (!write && !first && left) {
2548                 if(put_user('\n', s))
2549                         return -EFAULT;
2550                 left--, s++;
2551         }
2552         if (write) {
2553                 while (left) {
2554                         char c;
2555                         if (get_user(c, s++))
2556                                 return -EFAULT;
2557                         if (!isspace(c))
2558                                 break;
2559                         left--;
2560                 }
2561         }
2562         if (write && first)
2563                 return -EINVAL;
2564         *lenp -= left;
2565         *ppos += *lenp;
2566         return 0;
2567 #undef TMPBUFLEN
2568 }
2569
2570 static int do_proc_doulongvec_minmax(struct ctl_table *table, int write,
2571                                      void __user *buffer,
2572                                      size_t *lenp, loff_t *ppos,
2573                                      unsigned long convmul,
2574                                      unsigned long convdiv)
2575 {
2576         return __do_proc_doulongvec_minmax(table->data, table, write,
2577                         buffer, lenp, ppos, convmul, convdiv);
2578 }
2579
2580 /**
2581  * proc_doulongvec_minmax - read a vector of long integers with min/max values
2582  * @table: the sysctl table
2583  * @write: %TRUE if this is a write to the sysctl file
2584  * @buffer: the user buffer
2585  * @lenp: the size of the user buffer
2586  * @ppos: file position
2587  *
2588  * Reads/writes up to table->maxlen/sizeof(unsigned long) unsigned long
2589  * values from/to the user buffer, treated as an ASCII string.
2590  *
2591  * This routine will ensure the values are within the range specified by
2592  * table->extra1 (min) and table->extra2 (max).
2593  *
2594  * Returns 0 on success.
2595  */
2596 int proc_doulongvec_minmax(struct ctl_table *table, int write,
2597                            void __user *buffer, size_t *lenp, loff_t *ppos)
2598 {
2599     return do_proc_doulongvec_minmax(table, write, buffer, lenp, ppos, 1l, 1l);
2600 }
2601
2602 /**
2603  * proc_doulongvec_ms_jiffies_minmax - read a vector of millisecond values with min/max values
2604  * @table: the sysctl table
2605  * @write: %TRUE if this is a write to the sysctl file
2606  * @buffer: the user buffer
2607  * @lenp: the size of the user buffer
2608  * @ppos: file position
2609  *
2610  * Reads/writes up to table->maxlen/sizeof(unsigned long) unsigned long
2611  * values from/to the user buffer, treated as an ASCII string. The values
2612  * are treated as milliseconds, and converted to jiffies when they are stored.
2613  *
2614  * This routine will ensure the values are within the range specified by
2615  * table->extra1 (min) and table->extra2 (max).
2616  *
2617  * Returns 0 on success.
2618  */
2619 int proc_doulongvec_ms_jiffies_minmax(struct ctl_table *table, int write,
2620                                       void __user *buffer,
2621                                       size_t *lenp, loff_t *ppos)
2622 {
2623     return do_proc_doulongvec_minmax(table, write, buffer,
2624                                      lenp, ppos, HZ, 1000l);
2625 }
2626
2627
2628 static int do_proc_dointvec_jiffies_conv(int *negp, unsigned long *lvalp,
2629                                          int *valp,
2630                                          int write, void *data)
2631 {
2632         if (write) {
2633                 if (*lvalp > LONG_MAX / HZ)
2634                         return 1;
2635                 *valp = *negp ? -(*lvalp*HZ) : (*lvalp*HZ);
2636         } else {
2637                 int val = *valp;
2638                 unsigned long lval;
2639                 if (val < 0) {
2640                         *negp = -1;
2641                         lval = (unsigned long)-val;
2642                 } else {
2643                         *negp = 0;
2644                         lval = (unsigned long)val;
2645                 }
2646                 *lvalp = lval / HZ;
2647         }
2648         return 0;
2649 }
2650
2651 static int do_proc_dointvec_userhz_jiffies_conv(int *negp, unsigned long *lvalp,
2652                                                 int *valp,
2653                                                 int write, void *data)
2654 {
2655         if (write) {
2656                 if (USER_HZ < HZ && *lvalp > (LONG_MAX / HZ) * USER_HZ)
2657                         return 1;
2658                 *valp = clock_t_to_jiffies(*negp ? -*lvalp : *lvalp);
2659         } else {
2660                 int val = *valp;
2661                 unsigned long lval;
2662                 if (val < 0) {
2663                         *negp = -1;
2664                         lval = (unsigned long)-val;
2665                 } else {
2666                         *negp = 0;
2667                         lval = (unsigned long)val;
2668                 }
2669                 *lvalp = jiffies_to_clock_t(lval);
2670         }
2671         return 0;
2672 }
2673
2674 static int do_proc_dointvec_ms_jiffies_conv(int *negp, unsigned long *lvalp,
2675                                             int *valp,
2676                                             int write, void *data)
2677 {
2678         if (write) {
2679                 *valp = msecs_to_jiffies(*negp ? -*lvalp : *lvalp);
2680         } else {
2681                 int val = *valp;
2682                 unsigned long lval;
2683                 if (val < 0) {
2684                         *negp = -1;
2685                         lval = (unsigned long)-val;
2686                 } else {
2687                         *negp = 0;
2688                         lval = (unsigned long)val;
2689                 }
2690                 *lvalp = jiffies_to_msecs(lval);
2691         }
2692         return 0;
2693 }
2694
2695 /**
2696  * proc_dointvec_jiffies - read a vector of integers as seconds
2697  * @table: the sysctl table
2698  * @write: %TRUE if this is a write to the sysctl file
2699  * @buffer: the user buffer
2700  * @lenp: the size of the user buffer
2701  * @ppos: file position
2702  *
2703  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2704  * values from/to the user buffer, treated as an ASCII string.
2705  * The values read are assumed to be in seconds, and are converted into
2706  * jiffies.
2707  *
2708  * Returns 0 on success.
2709  */
2710 int proc_dointvec_jiffies(struct ctl_table *table, int write,
2711                           void __user *buffer, size_t *lenp, loff_t *ppos)
2712 {
2713     return do_proc_dointvec(table,write,buffer,lenp,ppos,
2714                             do_proc_dointvec_jiffies_conv,NULL);
2715 }
2716
2717 /**
2718  * proc_dointvec_userhz_jiffies - read a vector of integers as 1/USER_HZ seconds
2719  * @table: the sysctl table
2720  * @write: %TRUE if this is a write to the sysctl file
2721  * @buffer: the user buffer
2722  * @lenp: the size of the user buffer
2723  * @ppos: pointer to the file position
2724  *
2725  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2726  * values from/to the user buffer, treated as an ASCII string.
2727  * The values read are assumed to be in 1/USER_HZ seconds, and
2728  * are converted into jiffies.
2729  *
2730  * Returns 0 on success.
2731  */
2732 int proc_dointvec_userhz_jiffies(struct ctl_table *table, int write,
2733                                  void __user *buffer, size_t *lenp, loff_t *ppos)
2734 {
2735     return do_proc_dointvec(table,write,buffer,lenp,ppos,
2736                             do_proc_dointvec_userhz_jiffies_conv,NULL);
2737 }
2738
2739 /**
2740  * proc_dointvec_ms_jiffies - read a vector of integers as 1 milliseconds
2741  * @table: the sysctl table
2742  * @write: %TRUE if this is a write to the sysctl file
2743  * @buffer: the user buffer
2744  * @lenp: the size of the user buffer
2745  * @ppos: file position
2746  * @ppos: the current position in the file
2747  *
2748  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2749  * values from/to the user buffer, treated as an ASCII string.
2750  * The values read are assumed to be in 1/1000 seconds, and
2751  * are converted into jiffies.
2752  *
2753  * Returns 0 on success.
2754  */
2755 int proc_dointvec_ms_jiffies(struct ctl_table *table, int write,
2756                              void __user *buffer, size_t *lenp, loff_t *ppos)
2757 {
2758         return do_proc_dointvec(table, write, buffer, lenp, ppos,
2759                                 do_proc_dointvec_ms_jiffies_conv, NULL);
2760 }
2761
2762 static int proc_do_cad_pid(struct ctl_table *table, int write,
2763                            void __user *buffer, size_t *lenp, loff_t *ppos)
2764 {
2765         struct pid *new_pid;
2766         pid_t tmp;
2767         int r;
2768
2769         tmp = pid_vnr(cad_pid);
2770
2771         r = __do_proc_dointvec(&tmp, table, write, buffer,
2772                                lenp, ppos, NULL, NULL);
2773         if (r || !write)
2774                 return r;
2775
2776         new_pid = find_get_pid(tmp);
2777         if (!new_pid)
2778                 return -ESRCH;
2779
2780         put_pid(xchg(&cad_pid, new_pid));
2781         return 0;
2782 }
2783
2784 #else /* CONFIG_PROC_FS */
2785
2786 int proc_dostring(struct ctl_table *table, int write,
2787                   void __user *buffer, size_t *lenp, loff_t *ppos)
2788 {
2789         return -ENOSYS;
2790 }
2791
2792 int proc_dointvec(struct ctl_table *table, int write,
2793                   void __user *buffer, size_t *lenp, loff_t *ppos)
2794 {
2795         return -ENOSYS;
2796 }
2797
2798 int proc_dointvec_minmax(struct ctl_table *table, int write,
2799                     void __user *buffer, size_t *lenp, loff_t *ppos)
2800 {
2801         return -ENOSYS;
2802 }
2803
2804 int proc_dointvec_jiffies(struct ctl_table *table, int write,
2805                     void __user *buffer, size_t *lenp, loff_t *ppos)
2806 {
2807         return -ENOSYS;
2808 }
2809
2810 int proc_dointvec_userhz_jiffies(struct ctl_table *table, int write,
2811                     void __user *buffer, size_t *lenp, loff_t *ppos)
2812 {
2813         return -ENOSYS;
2814 }
2815
2816 int proc_dointvec_ms_jiffies(struct ctl_table *table, int write,
2817                              void __user *buffer, size_t *lenp, loff_t *ppos)
2818 {
2819         return -ENOSYS;
2820 }
2821
2822 int proc_doulongvec_minmax(struct ctl_table *table, int write,
2823                     void __user *buffer, size_t *lenp, loff_t *ppos)
2824 {
2825         return -ENOSYS;
2826 }
2827
2828 int proc_doulongvec_ms_jiffies_minmax(struct ctl_table *table, int write,
2829                                       void __user *buffer,
2830                                       size_t *lenp, loff_t *ppos)
2831 {
2832     return -ENOSYS;
2833 }
2834
2835
2836 #endif /* CONFIG_PROC_FS */
2837
2838
2839 #ifdef CONFIG_SYSCTL_SYSCALL
2840 /*
2841  * General sysctl support routines
2842  */
2843
2844 /* The generic sysctl data routine (used if no strategy routine supplied) */
2845 int sysctl_data(struct ctl_table *table,
2846                 void __user *oldval, size_t __user *oldlenp,
2847                 void __user *newval, size_t newlen)
2848 {
2849         size_t len;
2850
2851         /* Get out of I don't have a variable */
2852         if (!table->data || !table->maxlen)
2853                 return -ENOTDIR;
2854
2855         if (oldval && oldlenp) {
2856                 if (get_user(len, oldlenp))
2857                         return -EFAULT;
2858                 if (len) {
2859                         if (len > table->maxlen)
2860                                 len = table->maxlen;
2861                         if (copy_to_user(oldval, table->data, len))
2862                                 return -EFAULT;
2863                         if (put_user(len, oldlenp))
2864                                 return -EFAULT;
2865                 }
2866         }
2867
2868         if (newval && newlen) {
2869                 if (newlen > table->maxlen)
2870                         newlen = table->maxlen;
2871
2872                 if (copy_from_user(table->data, newval, newlen))
2873                         return -EFAULT;
2874         }
2875         return 1;
2876 }
2877
2878 /* The generic string strategy routine: */
2879 int sysctl_string(struct ctl_table *table,
2880                   void __user *oldval, size_t __user *oldlenp,
2881                   void __user *newval, size_t newlen)
2882 {
2883         if (!table->data || !table->maxlen)
2884                 return -ENOTDIR;
2885
2886         if (oldval && oldlenp) {
2887                 size_t bufsize;
2888                 if (get_user(bufsize, oldlenp))
2889                         return -EFAULT;
2890                 if (bufsize) {
2891                         size_t len = strlen(table->data), copied;
2892
2893                         /* This shouldn't trigger for a well-formed sysctl */
2894                         if (len > table->maxlen)
2895                                 len = table->maxlen;
2896
2897                         /* Copy up to a max of bufsize-1 bytes of the string */
2898                         copied = (len >= bufsize) ? bufsize - 1 : len;
2899
2900                         if (copy_to_user(oldval, table->data, copied) ||
2901                             put_user(0, (char __user *)(oldval + copied)))
2902                                 return -EFAULT;
2903                         if (put_user(len, oldlenp))
2904                                 return -EFAULT;
2905                 }
2906         }
2907         if (newval && newlen) {
2908                 size_t len = newlen;
2909                 if (len > table->maxlen)
2910                         len = table->maxlen;
2911                 if(copy_from_user(table->data, newval, len))
2912                         return -EFAULT;
2913                 if (len == table->maxlen)
2914                         len--;
2915                 ((char *) table->data)[len] = 0;
2916         }
2917         return 1;
2918 }
2919
2920 /*
2921  * This function makes sure that all of the integers in the vector
2922  * are between the minimum and maximum values given in the arrays
2923  * table->extra1 and table->extra2, respectively.
2924  */
2925 int sysctl_intvec(struct ctl_table *table,
2926                 void __user *oldval, size_t __user *oldlenp,
2927                 void __user *newval, size_t newlen)
2928 {
2929
2930         if (newval && newlen) {
2931                 int __user *vec = (int __user *) newval;
2932                 int *min = (int *) table->extra1;
2933                 int *max = (int *) table->extra2;
2934                 size_t length;
2935                 int i;
2936
2937                 if (newlen % sizeof(int) != 0)
2938                         return -EINVAL;
2939
2940                 if (!table->extra1 && !table->extra2)
2941                         return 0;
2942
2943                 if (newlen > table->maxlen)
2944                         newlen = table->maxlen;
2945                 length = newlen / sizeof(int);
2946
2947                 for (i = 0; i < length; i++) {
2948                         int value;
2949                         if (get_user(value, vec + i))
2950                                 return -EFAULT;
2951                         if (min && value < min[i])
2952                                 return -EINVAL;
2953                         if (max && value > max[i])
2954                                 return -EINVAL;
2955                 }
2956         }
2957         return 0;
2958 }
2959
2960 /* Strategy function to convert jiffies to seconds */
2961 int sysctl_jiffies(struct ctl_table *table,
2962                 void __user *oldval, size_t __user *oldlenp,
2963                 void __user *newval, size_t newlen)
2964 {
2965         if (oldval && oldlenp) {
2966                 size_t olen;
2967
2968                 if (get_user(olen, oldlenp))
2969                         return -EFAULT;
2970                 if (olen) {
2971                         int val;
2972
2973                         if (olen < sizeof(int))
2974                                 return -EINVAL;
2975
2976                         val = *(int *)(table->data) / HZ;
2977                         if (put_user(val, (int __user *)oldval))
2978                                 return -EFAULT;
2979                         if (put_user(sizeof(int), oldlenp))
2980                                 return -EFAULT;
2981                 }
2982         }
2983         if (newval && newlen) {
2984                 int new;
2985                 if (newlen != sizeof(int))
2986                         return -EINVAL;
2987                 if (get_user(new, (int __user *)newval))
2988                         return -EFAULT;
2989                 *(int *)(table->data) = new*HZ;
2990         }
2991         return 1;
2992 }
2993
2994 /* Strategy function to convert jiffies to seconds */
2995 int sysctl_ms_jiffies(struct ctl_table *table,
2996                 void __user *oldval, size_t __user *oldlenp,
2997                 void __user *newval, size_t newlen)
2998 {
2999         if (oldval && oldlenp) {
3000                 size_t olen;
3001
3002                 if (get_user(olen, oldlenp))
3003                         return -EFAULT;
3004                 if (olen) {
3005                         int val;
3006
3007                         if (olen < sizeof(int))
3008                                 return -EINVAL;
3009
3010                         val = jiffies_to_msecs(*(int *)(table->data));
3011                         if (put_user(val, (int __user *)oldval))
3012                                 return -EFAULT;
3013                         if (put_user(sizeof(int), oldlenp))
3014                                 return -EFAULT;
3015                 }
3016         }
3017         if (newval && newlen) {
3018                 int new;
3019                 if (newlen != sizeof(int))
3020                         return -EINVAL;
3021                 if (get_user(new, (int __user *)newval))
3022                         return -EFAULT;
3023                 *(int *)(table->data) = msecs_to_jiffies(new);
3024         }
3025         return 1;
3026 }
3027
3028
3029
3030 #else /* CONFIG_SYSCTL_SYSCALL */
3031
3032
3033 int sysctl_data(struct ctl_table *table,
3034                   void __user *oldval, size_t __user *oldlenp,
3035                   void __user *newval, size_t newlen)
3036 {
3037         return -ENOSYS;
3038 }
3039
3040 int sysctl_string(struct ctl_table *table,
3041                   void __user *oldval, size_t __user *oldlenp,
3042                   void __user *newval, size_t newlen)
3043 {
3044         return -ENOSYS;
3045 }
3046
3047 int sysctl_intvec(struct ctl_table *table,
3048                 void __user *oldval, size_t __user *oldlenp,
3049                 void __user *newval, size_t newlen)
3050 {
3051         return -ENOSYS;
3052 }
3053
3054 int sysctl_jiffies(struct ctl_table *table,
3055                 void __user *oldval, size_t __user *oldlenp,
3056                 void __user *newval, size_t newlen)
3057 {
3058         return -ENOSYS;
3059 }
3060
3061 int sysctl_ms_jiffies(struct ctl_table *table,
3062                 void __user *oldval, size_t __user *oldlenp,
3063                 void __user *newval, size_t newlen)
3064 {
3065         return -ENOSYS;
3066 }
3067
3068 #endif /* CONFIG_SYSCTL_SYSCALL */
3069
3070
3071 /*
3072  * No sense putting this after each symbol definition, twice,
3073  * exception granted :-)
3074  */
3075 EXPORT_SYMBOL(proc_dointvec);
3076 EXPORT_SYMBOL(proc_dointvec_jiffies);
3077 EXPORT_SYMBOL(proc_dointvec_minmax);
3078 EXPORT_SYMBOL(proc_dointvec_userhz_jiffies);
3079 EXPORT_SYMBOL(proc_dointvec_ms_jiffies);
3080 EXPORT_SYMBOL(proc_dostring);
3081 EXPORT_SYMBOL(proc_doulongvec_minmax);
3082 EXPORT_SYMBOL(proc_doulongvec_ms_jiffies_minmax);
3083 EXPORT_SYMBOL(register_sysctl_table);
3084 EXPORT_SYMBOL(register_sysctl_paths);
3085 EXPORT_SYMBOL(sysctl_intvec);
3086 EXPORT_SYMBOL(sysctl_jiffies);
3087 EXPORT_SYMBOL(sysctl_ms_jiffies);
3088 EXPORT_SYMBOL(sysctl_string);
3089 EXPORT_SYMBOL(sysctl_data);
3090 EXPORT_SYMBOL(unregister_sysctl_table);