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/utsname.h>
  30 #include <linux/smp_lock.h>
  31 #include <linux/fs.h>
  32 #include <linux/init.h>
  33 #include <linux/kernel.h>
  34 #include <linux/kobject.h>
  35 #include <linux/net.h>
  36 #include <linux/sysrq.h>
  37 #include <linux/highuid.h>
  38 #include <linux/writeback.h>
  39 #include <linux/hugetlb.h>
  40 #include <linux/initrd.h>
  41 #include <linux/key.h>
  42 #include <linux/times.h>
  43 #include <linux/limits.h>
  44 #include <linux/dcache.h>
  45 #include <linux/syscalls.h>
  46 #include <linux/vmstat.h>
  47 #include <linux/nfs_fs.h>
  48 #include <linux/acpi.h>
  49 #include <linux/reboot.h>
  50 #include <linux/ftrace.h>
  51
  52 #include <asm/uaccess.h>
  53 #include <asm/processor.h>
  54
  55 #ifdef CONFIG_X86
  56 #include <asm/nmi.h>
  57 #include <asm/stacktrace.h>
  58 #include <asm/io.h>
  59 #endif
  60
  61 static int deprecated_sysctl_warning(struct __sysctl_args *args);
  62
  63 #if defined(CONFIG_SYSCTL)
  64
  65 /* External variables not in a header file. */
  66 extern int C_A_D;
  67 extern int print_fatal_signals;
  68 extern int sysctl_overcommit_memory;
  69 extern int sysctl_overcommit_ratio;
  70 extern int sysctl_panic_on_oom;
  71 extern int sysctl_oom_kill_allocating_task;
  72 extern int sysctl_oom_dump_tasks;
  73 extern int max_threads;
  74 extern int core_uses_pid;
  75 extern int suid_dumpable;
  76 extern char core_pattern[];
  77 extern int pid_max;
  78 extern int min_free_kbytes;
  79 extern int pid_max_min, pid_max_max;
  80 extern int sysctl_drop_caches;
  81 extern int percpu_pagelist_fraction;
  82 extern int compat_log;
  83 extern int latencytop_enabled;
  84 extern int sysctl_nr_open_min, sysctl_nr_open_max;
  85 #ifdef CONFIG_RCU_TORTURE_TEST
  86 extern int rcutorture_runnable;
  87 #endif /* #ifdef CONFIG_RCU_TORTURE_TEST */
  88
  89 /* Constants used for minimum and  maximum */
  90 #if defined(CONFIG_HIGHMEM) || defined(CONFIG_DETECT_SOFTLOCKUP)
  91 static int one = 1;
  92 #endif
  93
  94 #ifdef CONFIG_DETECT_SOFTLOCKUP
  95 static int sixty = 60;
  96 static int neg_one = -1;
  97 #endif
  98
  99 #if defined(CONFIG_MMU) && defined(CONFIG_FILE_LOCKING)
 100 static int two = 2;
 101 #endif
 102
 103 static int zero;
 104 static int one_hundred = 100;
 105
 106 /* this is needed for the proc_dointvec_minmax for [fs_]overflow UID and GID */
 107 static int maxolduid = 65535;
 108 static int minolduid;
 109 static int min_percpu_pagelist_fract = 8;
 110
 111 static int ngroups_max = NGROUPS_MAX;
 112
 113 #ifdef CONFIG_MODULES
 114 extern char modprobe_path[];
 115 #endif
 116 #ifdef CONFIG_CHR_DEV_SG
 117 extern int sg_big_buff;
 118 #endif
 119
 120 #ifdef CONFIG_SPARC
 121 #include <asm/system.h>
 122 #endif
 123
 124 #ifdef __hppa__
 125 extern int pwrsw_enabled;
 126 extern int unaligned_enabled;
 127 #endif
 128
 129 #ifdef CONFIG_S390
 130 #ifdef CONFIG_MATHEMU
 131 extern int sysctl_ieee_emulation_warnings;
 132 #endif
 133 extern int sysctl_userprocess_debug;
 134 extern int spin_retry;
 135 #endif
 136
 137 #ifdef CONFIG_BSD_PROCESS_ACCT
 138 extern int acct_parm[];
 139 #endif
 140
 141 #ifdef CONFIG_IA64
 142 extern int no_unaligned_warning;
 143 #endif
 144
 145 #ifdef CONFIG_RT_MUTEXES
 146 extern int max_lock_depth;
 147 #endif
 148
 149 #ifdef CONFIG_PROC_SYSCTL
 150 static int proc_do_cad_pid(struct ctl_table *table, int write, struct file *filp,
 151                   void __user *buffer, size_t *lenp, loff_t *ppos);
 152 static int proc_taint(struct ctl_table *table, int write, struct file *filp,
 153                                void __user *buffer, size_t *lenp, loff_t *ppos);
 154 #endif
 155
 156 static struct ctl_table root_table[];
 157 static struct ctl_table_root sysctl_table_root;
 158 static struct ctl_table_header root_table_header = {
 159         .count = 1,
 160         .ctl_table = root_table,
 161         .ctl_entry = LIST_HEAD_INIT(sysctl_table_root.default_set.list),
 162         .root = &sysctl_table_root,
 163         .set = &sysctl_table_root.default_set,
 164 };
 165 static struct ctl_table_root sysctl_table_root = {
 166         .root_list = LIST_HEAD_INIT(sysctl_table_root.root_list),
 167         .default_set.list = LIST_HEAD_INIT(root_table_header.ctl_entry),
 168 };
 169
 170 static struct ctl_table kern_table[];
 171 static struct ctl_table vm_table[];
 172 static struct ctl_table fs_table[];
 173 static struct ctl_table debug_table[];
 174 static struct ctl_table dev_table[];
 175 extern struct ctl_table random_table[];
 176 #ifdef CONFIG_INOTIFY_USER
 177 extern struct ctl_table inotify_table[];
 178 #endif
 179 #ifdef CONFIG_EPOLL
 180 extern struct ctl_table epoll_table[];
 181 #endif
 182
 183 #ifdef HAVE_ARCH_PICK_MMAP_LAYOUT
 184 int sysctl_legacy_va_layout;
 185 #endif
 186
 187 extern int prove_locking;
 188 extern int lock_stat;
 189
 190 /* The default sysctl tables: */
 191
 192 static struct ctl_table root_table[] = {
 193         {
 194                 .ctl_name       = CTL_KERN,
 195                 .procname       = "kernel",
 196                 .mode           = 0555,
 197                 .child          = kern_table,
 198         },
 199         {
 200                 .ctl_name       = CTL_VM,
 201                 .procname       = "vm",
 202                 .mode           = 0555,
 203                 .child          = vm_table,
 204         },
 205         {
 206                 .ctl_name       = CTL_FS,
 207                 .procname       = "fs",
 208                 .mode           = 0555,
 209                 .child          = fs_table,
 210         },
 211         {
 212                 .ctl_name       = CTL_DEBUG,
 213                 .procname       = "debug",
 214                 .mode           = 0555,
 215                 .child          = debug_table,
 216         },
 217         {
 218                 .ctl_name       = CTL_DEV,
 219                 .procname       = "dev",
 220                 .mode           = 0555,
 221                 .child          = dev_table,
 222         },
 223 /*
 224  * NOTE: do not add new entries to this table unless you have read
 225  * Documentation/sysctl/ctl_unnumbered.txt
 226  */
 227         { .ctl_name = 0 }
 228 };
 229
 230 #ifdef CONFIG_SCHED_DEBUG
 231 static int min_sched_granularity_ns = 100000;           /* 100 usecs */
 232 static int max_sched_granularity_ns = NSEC_PER_SEC;     /* 1 second */
 233 static int min_wakeup_granularity_ns;                   /* 0 usecs */
 234 static int max_wakeup_granularity_ns = NSEC_PER_SEC;    /* 1 second */
 235 #endif
 236
 237 static struct ctl_table kern_table[] = {
 238 #ifdef CONFIG_SCHED_DEBUG
 239         {
 240                 .ctl_name       = CTL_UNNUMBERED,
 241                 .procname       = "sched_min_granularity_ns",
 242                 .data           = &sysctl_sched_min_granularity,
 243                 .maxlen         = sizeof(unsigned int),
 244                 .mode           = 0644,
 245                 .proc_handler   = &sched_nr_latency_handler,
 246                 .strategy       = &sysctl_intvec,
 247                 .extra1         = &min_sched_granularity_ns,
 248                 .extra2         = &max_sched_granularity_ns,
 249         },
 250         {
 251                 .ctl_name       = CTL_UNNUMBERED,
 252                 .procname       = "sched_latency_ns",
 253                 .data           = &sysctl_sched_latency,
 254                 .maxlen         = sizeof(unsigned int),
 255                 .mode           = 0644,
 256                 .proc_handler   = &sched_nr_latency_handler,
 257                 .strategy       = &sysctl_intvec,
 258                 .extra1         = &min_sched_granularity_ns,
 259                 .extra2         = &max_sched_granularity_ns,
 260         },
 261         {
 262                 .ctl_name       = CTL_UNNUMBERED,
 263                 .procname       = "sched_wakeup_granularity_ns",
 264                 .data           = &sysctl_sched_wakeup_granularity,
 265                 .maxlen         = sizeof(unsigned int),
 266                 .mode           = 0644,
 267                 .proc_handler   = &proc_dointvec_minmax,
 268                 .strategy       = &sysctl_intvec,
 269                 .extra1         = &min_wakeup_granularity_ns,
 270                 .extra2         = &max_wakeup_granularity_ns,
 271         },
 272         {
 273                 .ctl_name       = CTL_UNNUMBERED,
 274                 .procname       = "sched_shares_ratelimit",
 275                 .data           = &sysctl_sched_shares_ratelimit,
 276                 .maxlen         = sizeof(unsigned int),
 277                 .mode           = 0644,
 278                 .proc_handler   = &proc_dointvec,
 279         },
 280         {
 281                 .ctl_name       = CTL_UNNUMBERED,
 282                 .procname       = "sched_shares_thresh",
 283                 .data           = &sysctl_sched_shares_thresh,
 284                 .maxlen         = sizeof(unsigned int),
 285                 .mode           = 0644,
 286                 .proc_handler   = &proc_dointvec_minmax,
 287                 .strategy       = &sysctl_intvec,
 288                 .extra1         = &zero,
 289         },
 290         {
 291                 .ctl_name       = CTL_UNNUMBERED,
 292                 .procname       = "sched_child_runs_first",
 293                 .data           = &sysctl_sched_child_runs_first,
 294                 .maxlen         = sizeof(unsigned int),
 295                 .mode           = 0644,
 296                 .proc_handler   = &proc_dointvec,
 297         },
 298         {
 299                 .ctl_name       = CTL_UNNUMBERED,
 300                 .procname       = "sched_features",
 301                 .data           = &sysctl_sched_features,
 302                 .maxlen         = sizeof(unsigned int),
 303                 .mode           = 0644,
 304                 .proc_handler   = &proc_dointvec,
 305         },
 306         {
 307                 .ctl_name       = CTL_UNNUMBERED,
 308                 .procname       = "sched_migration_cost",
 309                 .data           = &sysctl_sched_migration_cost,
 310                 .maxlen         = sizeof(unsigned int),
 311                 .mode           = 0644,
 312                 .proc_handler   = &proc_dointvec,
 313         },
 314         {
 315                 .ctl_name       = CTL_UNNUMBERED,
 316                 .procname       = "sched_nr_migrate",
 317                 .data           = &sysctl_sched_nr_migrate,
 318                 .maxlen         = sizeof(unsigned int),
 319                 .mode           = 0644,
 320                 .proc_handler   = &proc_dointvec,
 321         },
 322 #endif
 323         {
 324                 .ctl_name       = CTL_UNNUMBERED,
 325                 .procname       = "sched_rt_period_us",
 326                 .data           = &sysctl_sched_rt_period,
 327                 .maxlen         = sizeof(unsigned int),
 328                 .mode           = 0644,
 329                 .proc_handler   = &sched_rt_handler,
 330         },
 331         {
 332                 .ctl_name       = CTL_UNNUMBERED,
 333                 .procname       = "sched_rt_runtime_us",
 334                 .data           = &sysctl_sched_rt_runtime,
 335                 .maxlen         = sizeof(int),
 336                 .mode           = 0644,
 337                 .proc_handler   = &sched_rt_handler,
 338         },
 339         {
 340                 .ctl_name       = CTL_UNNUMBERED,
 341                 .procname       = "sched_compat_yield",
 342                 .data           = &sysctl_sched_compat_yield,
 343                 .maxlen         = sizeof(unsigned int),
 344                 .mode           = 0644,
 345                 .proc_handler   = &proc_dointvec,
 346         },
 347 #ifdef CONFIG_PROVE_LOCKING
 348         {
 349                 .ctl_name       = CTL_UNNUMBERED,
 350                 .procname       = "prove_locking",
 351                 .data           = &prove_locking,
 352                 .maxlen         = sizeof(int),
 353                 .mode           = 0644,
 354                 .proc_handler   = &proc_dointvec,
 355         },
 356 #endif
 357 #ifdef CONFIG_LOCK_STAT
 358         {
 359                 .ctl_name       = CTL_UNNUMBERED,
 360                 .procname       = "lock_stat",
 361                 .data           = &lock_stat,
 362                 .maxlen         = sizeof(int),
 363                 .mode           = 0644,
 364                 .proc_handler   = &proc_dointvec,
 365         },
 366 #endif
 367         {
 368                 .ctl_name       = KERN_PANIC,
 369                 .procname       = "panic",
 370                 .data           = &panic_timeout,
 371                 .maxlen         = sizeof(int),
 372                 .mode           = 0644,
 373                 .proc_handler   = &proc_dointvec,
 374         },
 375         {
 376                 .ctl_name       = KERN_CORE_USES_PID,
 377                 .procname       = "core_uses_pid",
 378                 .data           = &core_uses_pid,
 379                 .maxlen         = sizeof(int),
 380                 .mode           = 0644,
 381                 .proc_handler   = &proc_dointvec,
 382         },
 383         {
 384                 .ctl_name       = KERN_CORE_PATTERN,
 385                 .procname       = "core_pattern",
 386                 .data           = core_pattern,
 387                 .maxlen         = CORENAME_MAX_SIZE,
 388                 .mode           = 0644,
 389                 .proc_handler   = &proc_dostring,
 390                 .strategy       = &sysctl_string,
 391         },
 392 #ifdef CONFIG_PROC_SYSCTL
 393         {
 394                 .procname       = "tainted",
 395                 .maxlen         = sizeof(long),
 396                 .mode           = 0644,
 397                 .proc_handler   = &proc_taint,
 398         },
 399 #endif
 400 #ifdef CONFIG_LATENCYTOP
 401         {
 402                 .procname       = "latencytop",
 403                 .data           = &latencytop_enabled,
 404                 .maxlen         = sizeof(int),
 405                 .mode           = 0644,
 406                 .proc_handler   = &proc_dointvec,
 407         },
 408 #endif
 409 #ifdef CONFIG_BLK_DEV_INITRD
 410         {
 411                 .ctl_name       = KERN_REALROOTDEV,
 412                 .procname       = "real-root-dev",
 413                 .data           = &real_root_dev,
 414                 .maxlen         = sizeof(int),
 415                 .mode           = 0644,
 416                 .proc_handler   = &proc_dointvec,
 417         },
 418 #endif
 419         {
 420                 .ctl_name       = CTL_UNNUMBERED,
 421                 .procname       = "print-fatal-signals",
 422                 .data           = &print_fatal_signals,
 423                 .maxlen         = sizeof(int),
 424                 .mode           = 0644,
 425                 .proc_handler   = &proc_dointvec,
 426         },
 427 #ifdef CONFIG_SPARC
 428         {
 429                 .ctl_name       = KERN_SPARC_REBOOT,
 430                 .procname       = "reboot-cmd",
 431                 .data           = reboot_command,
 432                 .maxlen         = 256,
 433                 .mode           = 0644,
 434                 .proc_handler   = &proc_dostring,
 435                 .strategy       = &sysctl_string,
 436         },
 437         {
 438                 .ctl_name       = KERN_SPARC_STOP_A,
 439                 .procname       = "stop-a",
 440                 .data           = &stop_a_enabled,
 441                 .maxlen         = sizeof (int),
 442                 .mode           = 0644,
 443                 .proc_handler   = &proc_dointvec,
 444         },
 445         {
 446                 .ctl_name       = KERN_SPARC_SCONS_PWROFF,
 447                 .procname       = "scons-poweroff",
 448                 .data           = &scons_pwroff,
 449                 .maxlen         = sizeof (int),
 450                 .mode           = 0644,
 451                 .proc_handler   = &proc_dointvec,
 452         },
 453 #endif
 454 #ifdef __hppa__
 455         {
 456                 .ctl_name       = KERN_HPPA_PWRSW,
 457                 .procname       = "soft-power",
 458                 .data           = &pwrsw_enabled,
 459                 .maxlen         = sizeof (int),
 460                 .mode           = 0644,
 461                 .proc_handler   = &proc_dointvec,
 462         },
 463         {
 464                 .ctl_name       = KERN_HPPA_UNALIGNED,
 465                 .procname       = "unaligned-trap",
 466                 .data           = &unaligned_enabled,
 467                 .maxlen         = sizeof (int),
 468                 .mode           = 0644,
 469                 .proc_handler   = &proc_dointvec,
 470         },
 471 #endif
 472         {
 473                 .ctl_name       = KERN_CTLALTDEL,
 474                 .procname       = "ctrl-alt-del",
 475                 .data           = &C_A_D,
 476                 .maxlen         = sizeof(int),
 477                 .mode           = 0644,
 478                 .proc_handler   = &proc_dointvec,
 479         },
 480 #ifdef CONFIG_FUNCTION_TRACER
 481         {
 482                 .ctl_name       = CTL_UNNUMBERED,
 483                 .procname       = "ftrace_enabled",
 484                 .data           = &ftrace_enabled,
 485                 .maxlen         = sizeof(int),
 486                 .mode           = 0644,
 487                 .proc_handler   = &ftrace_enable_sysctl,
 488         },
 489 #endif
 490 #ifdef CONFIG_TRACING
 491         {
 492                 .ctl_name       = CTL_UNNUMBERED,
 493                 .procname       = "ftrace_dump_on_oops",
 494                 .data           = &ftrace_dump_on_oops,
 495                 .maxlen         = sizeof(int),
 496                 .mode           = 0644,
 497                 .proc_handler   = &proc_dointvec,
 498         },
 499 #endif
 500 #ifdef CONFIG_MODULES
 501         {
 502                 .ctl_name       = KERN_MODPROBE,
 503                 .procname       = "modprobe",
 504                 .data           = &modprobe_path,
 505                 .maxlen         = KMOD_PATH_LEN,
 506                 .mode           = 0644,
 507                 .proc_handler   = &proc_dostring,
 508                 .strategy       = &sysctl_string,
 509         },
 510 #endif
 511 #if defined(CONFIG_HOTPLUG) && defined(CONFIG_NET)
 512         {
 513                 .ctl_name       = KERN_HOTPLUG,
 514                 .procname       = "hotplug",
 515                 .data           = &uevent_helper,
 516                 .maxlen         = UEVENT_HELPER_PATH_LEN,
 517                 .mode           = 0644,
 518                 .proc_handler   = &proc_dostring,
 519                 .strategy       = &sysctl_string,
 520         },
 521 #endif
 522 #ifdef CONFIG_CHR_DEV_SG
 523         {
 524                 .ctl_name       = KERN_SG_BIG_BUFF,
 525                 .procname       = "sg-big-buff",
 526                 .data           = &sg_big_buff,
 527                 .maxlen         = sizeof (int),
 528                 .mode           = 0444,
 529                 .proc_handler   = &proc_dointvec,
 530         },
 531 #endif
 532 #ifdef CONFIG_BSD_PROCESS_ACCT
 533         {
 534                 .ctl_name       = KERN_ACCT,
 535                 .procname       = "acct",
 536                 .data           = &acct_parm,
 537                 .maxlen         = 3*sizeof(int),
 538                 .mode           = 0644,
 539                 .proc_handler   = &proc_dointvec,
 540         },
 541 #endif
 542 #ifdef CONFIG_MAGIC_SYSRQ
 543         {
 544                 .ctl_name       = KERN_SYSRQ,
 545                 .procname       = "sysrq",
 546                 .data           = &__sysrq_enabled,
 547                 .maxlen         = sizeof (int),
 548                 .mode           = 0644,
 549                 .proc_handler   = &proc_dointvec,
 550         },
 551 #endif
 552 #ifdef CONFIG_PROC_SYSCTL
 553         {
 554                 .procname       = "cad_pid",
 555                 .data           = NULL,
 556                 .maxlen         = sizeof (int),
 557                 .mode           = 0600,
 558                 .proc_handler   = &proc_do_cad_pid,
 559         },
 560 #endif
 561         {
 562                 .ctl_name       = KERN_MAX_THREADS,
 563                 .procname       = "threads-max",
 564                 .data           = &max_threads,
 565                 .maxlen         = sizeof(int),
 566                 .mode           = 0644,
 567                 .proc_handler   = &proc_dointvec,
 568         },
 569         {
 570                 .ctl_name       = KERN_RANDOM,
 571                 .procname       = "random",
 572                 .mode           = 0555,
 573                 .child          = random_table,
 574         },
 575         {
 576                 .ctl_name       = KERN_OVERFLOWUID,
 577                 .procname       = "overflowuid",
 578                 .data           = &overflowuid,
 579                 .maxlen         = sizeof(int),
 580                 .mode           = 0644,
 581                 .proc_handler   = &proc_dointvec_minmax,
 582                 .strategy       = &sysctl_intvec,
 583                 .extra1         = &minolduid,
 584                 .extra2         = &maxolduid,
 585         },
 586         {
 587                 .ctl_name       = KERN_OVERFLOWGID,
 588                 .procname       = "overflowgid",
 589                 .data           = &overflowgid,
 590                 .maxlen         = sizeof(int),
 591                 .mode           = 0644,
 592                 .proc_handler   = &proc_dointvec_minmax,
 593                 .strategy       = &sysctl_intvec,
 594                 .extra1         = &minolduid,
 595                 .extra2         = &maxolduid,
 596         },
 597 #ifdef CONFIG_S390
 598 #ifdef CONFIG_MATHEMU
 599         {
 600                 .ctl_name       = KERN_IEEE_EMULATION_WARNINGS,
 601                 .procname       = "ieee_emulation_warnings",
 602                 .data           = &sysctl_ieee_emulation_warnings,
 603                 .maxlen         = sizeof(int),
 604                 .mode           = 0644,
 605                 .proc_handler   = &proc_dointvec,
 606         },
 607 #endif
 608         {
 609                 .ctl_name       = KERN_S390_USER_DEBUG_LOGGING,
 610                 .procname       = "userprocess_debug",
 611                 .data           = &sysctl_userprocess_debug,
 612                 .maxlen         = sizeof(int),
 613                 .mode           = 0644,
 614                 .proc_handler   = &proc_dointvec,
 615         },
 616 #endif
 617         {
 618                 .ctl_name       = KERN_PIDMAX,
 619                 .procname       = "pid_max",
 620                 .data           = &pid_max,
 621                 .maxlen         = sizeof (int),
 622                 .mode           = 0644,
 623                 .proc_handler   = &proc_dointvec_minmax,
 624                 .strategy       = sysctl_intvec,
 625                 .extra1         = &pid_max_min,
 626                 .extra2         = &pid_max_max,
 627         },
 628         {
 629                 .ctl_name       = KERN_PANIC_ON_OOPS,
 630                 .procname       = "panic_on_oops",
 631                 .data           = &panic_on_oops,
 632                 .maxlen         = sizeof(int),
 633                 .mode           = 0644,
 634                 .proc_handler   = &proc_dointvec,
 635         },
 636 #if defined CONFIG_PRINTK
 637         {
 638                 .ctl_name       = KERN_PRINTK,
 639                 .procname       = "printk",
 640                 .data           = &console_loglevel,
 641                 .maxlen         = 4*sizeof(int),
 642                 .mode           = 0644,
 643                 .proc_handler   = &proc_dointvec,
 644         },
 645         {
 646                 .ctl_name       = KERN_PRINTK_RATELIMIT,
 647                 .procname       = "printk_ratelimit",
 648                 .data           = &printk_ratelimit_state.interval,
 649                 .maxlen         = sizeof(int),
 650                 .mode           = 0644,
 651                 .proc_handler   = &proc_dointvec_jiffies,
 652                 .strategy       = &sysctl_jiffies,
 653         },
 654         {
 655                 .ctl_name       = KERN_PRINTK_RATELIMIT_BURST,
 656                 .procname       = "printk_ratelimit_burst",
 657                 .data           = &printk_ratelimit_state.burst,
 658                 .maxlen         = sizeof(int),
 659                 .mode           = 0644,
 660                 .proc_handler   = &proc_dointvec,
 661         },
 662 #endif
 663         {
 664                 .ctl_name       = KERN_NGROUPS_MAX,
 665                 .procname       = "ngroups_max",
 666                 .data           = &ngroups_max,
 667                 .maxlen         = sizeof (int),
 668                 .mode           = 0444,
 669                 .proc_handler   = &proc_dointvec,
 670         },
 671 #if defined(CONFIG_X86_LOCAL_APIC) && defined(CONFIG_X86)
 672         {
 673                 .ctl_name       = KERN_UNKNOWN_NMI_PANIC,
 674                 .procname       = "unknown_nmi_panic",
 675                 .data           = &unknown_nmi_panic,
 676                 .maxlen         = sizeof (int),
 677                 .mode           = 0644,
 678                 .proc_handler   = &proc_dointvec,
 679         },
 680         {
 681                 .procname       = "nmi_watchdog",
 682                 .data           = &nmi_watchdog_enabled,
 683                 .maxlen         = sizeof (int),
 684                 .mode           = 0644,
 685                 .proc_handler   = &proc_nmi_enabled,
 686         },
 687 #endif
 688 #if defined(CONFIG_X86)
 689         {
 690                 .ctl_name       = KERN_PANIC_ON_NMI,
 691                 .procname       = "panic_on_unrecovered_nmi",
 692                 .data           = &panic_on_unrecovered_nmi,
 693                 .maxlen         = sizeof(int),
 694                 .mode           = 0644,
 695                 .proc_handler   = &proc_dointvec,
 696         },
 697         {
 698                 .ctl_name       = KERN_BOOTLOADER_TYPE,
 699                 .procname       = "bootloader_type",
 700                 .data           = &bootloader_type,
 701                 .maxlen         = sizeof (int),
 702                 .mode           = 0444,
 703                 .proc_handler   = &proc_dointvec,
 704         },
 705         {
 706                 .ctl_name       = CTL_UNNUMBERED,
 707                 .procname       = "kstack_depth_to_print",
 708                 .data           = &kstack_depth_to_print,
 709                 .maxlen         = sizeof(int),
 710                 .mode           = 0644,
 711                 .proc_handler   = &proc_dointvec,
 712         },
 713         {
 714                 .ctl_name       = CTL_UNNUMBERED,
 715                 .procname       = "io_delay_type",
 716                 .data           = &io_delay_type,
 717                 .maxlen         = sizeof(int),
 718                 .mode           = 0644,
 719                 .proc_handler   = &proc_dointvec,
 720         },
 721 #endif
 722 #if defined(CONFIG_MMU)
 723         {
 724                 .ctl_name       = KERN_RANDOMIZE,
 725                 .procname       = "randomize_va_space",
 726                 .data           = &randomize_va_space,
 727                 .maxlen         = sizeof(int),
 728                 .mode           = 0644,
 729                 .proc_handler   = &proc_dointvec,
 730         },
 731 #endif
 732 #if defined(CONFIG_S390) && defined(CONFIG_SMP)
 733         {
 734                 .ctl_name       = KERN_SPIN_RETRY,
 735                 .procname       = "spin_retry",
 736                 .data           = &spin_retry,
 737                 .maxlen         = sizeof (int),
 738                 .mode           = 0644,
 739                 .proc_handler   = &proc_dointvec,
 740         },
 741 #endif
 742 #if     defined(CONFIG_ACPI_SLEEP) && defined(CONFIG_X86)
 743         {
 744                 .procname       = "acpi_video_flags",
 745                 .data           = &acpi_realmode_flags,
 746                 .maxlen         = sizeof (unsigned long),
 747                 .mode           = 0644,
 748                 .proc_handler   = &proc_doulongvec_minmax,
 749         },
 750 #endif
 751 #ifdef CONFIG_IA64
 752         {
 753                 .ctl_name       = KERN_IA64_UNALIGNED,
 754                 .procname       = "ignore-unaligned-usertrap",
 755                 .data           = &no_unaligned_warning,
 756                 .maxlen         = sizeof (int),
 757                 .mode           = 0644,
 758                 .proc_handler   = &proc_dointvec,
 759         },
 760 #endif
 761 #ifdef CONFIG_DETECT_SOFTLOCKUP
 762         {
 763                 .ctl_name       = CTL_UNNUMBERED,
 764                 .procname       = "softlockup_panic",
 765                 .data           = &softlockup_panic,
 766                 .maxlen         = sizeof(int),
 767                 .mode           = 0644,
 768                 .proc_handler   = &proc_dointvec_minmax,
 769                 .strategy       = &sysctl_intvec,
 770                 .extra1         = &zero,
 771                 .extra2         = &one,
 772         },
 773         {
 774                 .ctl_name       = CTL_UNNUMBERED,
 775                 .procname       = "softlockup_thresh",
 776                 .data           = &softlockup_thresh,
 777                 .maxlen         = sizeof(int),
 778                 .mode           = 0644,
 779                 .proc_handler   = &proc_dointvec_minmax,
 780                 .strategy       = &sysctl_intvec,
 781                 .extra1         = &neg_one,
 782                 .extra2         = &sixty,
 783         },
 784         {
 785                 .ctl_name       = CTL_UNNUMBERED,
 786                 .procname       = "hung_task_check_count",
 787                 .data           = &sysctl_hung_task_check_count,
 788                 .maxlen         = sizeof(unsigned long),
 789                 .mode           = 0644,
 790                 .proc_handler   = &proc_doulongvec_minmax,
 791                 .strategy       = &sysctl_intvec,
 792         },
 793         {
 794                 .ctl_name       = CTL_UNNUMBERED,
 795                 .procname       = "hung_task_timeout_secs",
 796                 .data           = &sysctl_hung_task_timeout_secs,
 797                 .maxlen         = sizeof(unsigned long),
 798                 .mode           = 0644,
 799                 .proc_handler   = &proc_doulongvec_minmax,
 800                 .strategy       = &sysctl_intvec,
 801         },
 802         {
 803                 .ctl_name       = CTL_UNNUMBERED,
 804                 .procname       = "hung_task_warnings",
 805                 .data           = &sysctl_hung_task_warnings,
 806                 .maxlen         = sizeof(unsigned long),
 807                 .mode           = 0644,
 808                 .proc_handler   = &proc_doulongvec_minmax,
 809                 .strategy       = &sysctl_intvec,
 810         },
 811 #endif
 812 #ifdef CONFIG_COMPAT
 813         {
 814                 .ctl_name       = KERN_COMPAT_LOG,
 815                 .procname       = "compat-log",
 816                 .data           = &compat_log,
 817                 .maxlen         = sizeof (int),
 818                 .mode           = 0644,
 819                 .proc_handler   = &proc_dointvec,
 820         },
 821 #endif
 822 #ifdef CONFIG_RT_MUTEXES
 823         {
 824                 .ctl_name       = KERN_MAX_LOCK_DEPTH,
 825                 .procname       = "max_lock_depth",
 826                 .data           = &max_lock_depth,
 827                 .maxlen         = sizeof(int),
 828                 .mode           = 0644,
 829                 .proc_handler   = &proc_dointvec,
 830         },
 831 #endif
 832         {
 833                 .ctl_name       = CTL_UNNUMBERED,
 834                 .procname       = "poweroff_cmd",
 835                 .data           = &poweroff_cmd,
 836                 .maxlen         = POWEROFF_CMD_PATH_LEN,
 837                 .mode           = 0644,
 838                 .proc_handler   = &proc_dostring,
 839                 .strategy       = &sysctl_string,
 840         },
 841 #ifdef CONFIG_KEYS
 842         {
 843                 .ctl_name       = CTL_UNNUMBERED,
 844                 .procname       = "keys",
 845                 .mode           = 0555,
 846                 .child          = key_sysctls,
 847         },
 848 #endif
 849 #ifdef CONFIG_RCU_TORTURE_TEST
 850         {
 851                 .ctl_name       = CTL_UNNUMBERED,
 852                 .procname       = "rcutorture_runnable",
 853                 .data           = &rcutorture_runnable,
 854                 .maxlen         = sizeof(int),
 855                 .mode           = 0644,
 856                 .proc_handler   = &proc_dointvec,
 857         },
 858 #endif
 859 #ifdef CONFIG_UNEVICTABLE_LRU
 860         {
 861                 .ctl_name       = CTL_UNNUMBERED,
 862                 .procname       = "scan_unevictable_pages",
 863                 .data           = &scan_unevictable_pages,
 864                 .maxlen         = sizeof(scan_unevictable_pages),
 865                 .mode           = 0644,
 866                 .proc_handler   = &scan_unevictable_handler,
 867         },
 868 #endif
 869 /*
 870  * NOTE: do not add new entries to this table unless you have read
 871  * Documentation/sysctl/ctl_unnumbered.txt
 872  */
 873         { .ctl_name = 0 }
 874 };
 875
 876 static struct ctl_table vm_table[] = {
 877         {
 878                 .ctl_name       = VM_OVERCOMMIT_MEMORY,
 879                 .procname       = "overcommit_memory",
 880                 .data           = &sysctl_overcommit_memory,
 881                 .maxlen         = sizeof(sysctl_overcommit_memory),
 882                 .mode           = 0644,
 883                 .proc_handler   = &proc_dointvec,
 884         },
 885         {
 886                 .ctl_name       = VM_PANIC_ON_OOM,
 887                 .procname       = "panic_on_oom",
 888                 .data           = &sysctl_panic_on_oom,
 889                 .maxlen         = sizeof(sysctl_panic_on_oom),
 890                 .mode           = 0644,
 891                 .proc_handler   = &proc_dointvec,
 892         },
 893         {
 894                 .ctl_name       = CTL_UNNUMBERED,
 895                 .procname       = "oom_kill_allocating_task",
 896                 .data           = &sysctl_oom_kill_allocating_task,
 897                 .maxlen         = sizeof(sysctl_oom_kill_allocating_task),
 898                 .mode           = 0644,
 899                 .proc_handler   = &proc_dointvec,
 900         },
 901         {
 902                 .ctl_name       = CTL_UNNUMBERED,
 903                 .procname       = "oom_dump_tasks",
 904                 .data           = &sysctl_oom_dump_tasks,
 905                 .maxlen         = sizeof(sysctl_oom_dump_tasks),
 906                 .mode           = 0644,
 907                 .proc_handler   = &proc_dointvec,
 908         },
 909         {
 910                 .ctl_name       = VM_OVERCOMMIT_RATIO,
 911                 .procname       = "overcommit_ratio",
 912                 .data           = &sysctl_overcommit_ratio,
 913                 .maxlen         = sizeof(sysctl_overcommit_ratio),
 914                 .mode           = 0644,
 915                 .proc_handler   = &proc_dointvec,
 916         },
 917         {
 918                 .ctl_name       = VM_PAGE_CLUSTER,
 919                 .procname       = "page-cluster",
 920                 .data           = &page_cluster,
 921                 .maxlen         = sizeof(int),
 922                 .mode           = 0644,
 923                 .proc_handler   = &proc_dointvec,
 924         },
 925         {
 926                 .ctl_name       = VM_DIRTY_BACKGROUND,
 927                 .procname       = "dirty_background_ratio",
 928                 .data           = &dirty_background_ratio,
 929                 .maxlen         = sizeof(dirty_background_ratio),
 930                 .mode           = 0644,
 931                 .proc_handler   = &proc_dointvec_minmax,
 932                 .strategy       = &sysctl_intvec,
 933                 .extra1         = &zero,
 934                 .extra2         = &one_hundred,
 935         },
 936         {
 937                 .ctl_name       = VM_DIRTY_RATIO,
 938                 .procname       = "dirty_ratio",
 939                 .data           = &vm_dirty_ratio,
 940                 .maxlen         = sizeof(vm_dirty_ratio),
 941                 .mode           = 0644,
 942                 .proc_handler   = &dirty_ratio_handler,
 943                 .strategy       = &sysctl_intvec,
 944                 .extra1         = &zero,
 945                 .extra2         = &one_hundred,
 946         },
 947         {
 948                 .procname       = "dirty_writeback_centisecs",
 949                 .data           = &dirty_writeback_interval,
 950                 .maxlen         = sizeof(dirty_writeback_interval),
 951                 .mode           = 0644,
 952                 .proc_handler   = &dirty_writeback_centisecs_handler,
 953         },
 954         {
 955                 .procname       = "dirty_expire_centisecs",
 956                 .data           = &dirty_expire_interval,
 957                 .maxlen         = sizeof(dirty_expire_interval),
 958                 .mode           = 0644,
 959                 .proc_handler   = &proc_dointvec_userhz_jiffies,
 960         },
 961         {
 962                 .ctl_name       = VM_NR_PDFLUSH_THREADS,
 963                 .procname       = "nr_pdflush_threads",
 964                 .data           = &nr_pdflush_threads,
 965                 .maxlen         = sizeof nr_pdflush_threads,
 966                 .mode           = 0444 /* read-only*/,
 967                 .proc_handler   = &proc_dointvec,
 968         },
 969         {
 970                 .ctl_name       = VM_SWAPPINESS,
 971                 .procname       = "swappiness",
 972                 .data           = &vm_swappiness,
 973                 .maxlen         = sizeof(vm_swappiness),
 974                 .mode           = 0644,
 975                 .proc_handler   = &proc_dointvec_minmax,
 976                 .strategy       = &sysctl_intvec,
 977                 .extra1         = &zero,
 978                 .extra2         = &one_hundred,
 979         },
 980 #ifdef CONFIG_HUGETLB_PAGE
 981          {
 982                 .procname       = "nr_hugepages",
 983                 .data           = NULL,
 984                 .maxlen         = sizeof(unsigned long),
 985                 .mode           = 0644,
 986                 .proc_handler   = &hugetlb_sysctl_handler,
 987                 .extra1         = (void *)&hugetlb_zero,
 988                 .extra2         = (void *)&hugetlb_infinity,
 989          },
 990          {
 991                 .ctl_name       = VM_HUGETLB_GROUP,
 992                 .procname       = "hugetlb_shm_group",
 993                 .data           = &sysctl_hugetlb_shm_group,
 994                 .maxlen         = sizeof(gid_t),
 995                 .mode           = 0644,
 996                 .proc_handler   = &proc_dointvec,
 997          },
 998          {
 999                 .ctl_name       = CTL_UNNUMBERED,
1000                 .procname       = "hugepages_treat_as_movable",
1001                 .data           = &hugepages_treat_as_movable,
1002                 .maxlen         = sizeof(int),
1003                 .mode           = 0644,
1004                 .proc_handler   = &hugetlb_treat_movable_handler,
1005         },
1006         {
1007                 .ctl_name       = CTL_UNNUMBERED,
1008                 .procname       = "nr_overcommit_hugepages",
1009                 .data           = NULL,
1010                 .maxlen         = sizeof(unsigned long),
1011                 .mode           = 0644,
1012                 .proc_handler   = &hugetlb_overcommit_handler,
1013                 .extra1         = (void *)&hugetlb_zero,
1014                 .extra2         = (void *)&hugetlb_infinity,
1015         },
1016 #endif
1017         {
1018                 .ctl_name       = VM_LOWMEM_RESERVE_RATIO,
1019                 .procname       = "lowmem_reserve_ratio",
1020                 .data           = &sysctl_lowmem_reserve_ratio,
1021                 .maxlen         = sizeof(sysctl_lowmem_reserve_ratio),
1022                 .mode           = 0644,
1023                 .proc_handler   = &lowmem_reserve_ratio_sysctl_handler,
1024                 .strategy       = &sysctl_intvec,
1025         },
1026         {
1027                 .ctl_name       = VM_DROP_PAGECACHE,
1028                 .procname       = "drop_caches",
1029                 .data           = &sysctl_drop_caches,
1030                 .maxlen         = sizeof(int),
1031                 .mode           = 0644,
1032                 .proc_handler   = drop_caches_sysctl_handler,
1033                 .strategy       = &sysctl_intvec,
1034         },
1035         {
1036                 .ctl_name       = VM_MIN_FREE_KBYTES,
1037                 .procname       = "min_free_kbytes",
1038                 .data           = &min_free_kbytes,
1039                 .maxlen         = sizeof(min_free_kbytes),
1040                 .mode           = 0644,
1041                 .proc_handler   = &min_free_kbytes_sysctl_handler,
1042                 .strategy       = &sysctl_intvec,
1043                 .extra1         = &zero,
1044         },
1045         {
1046                 .ctl_name       = VM_PERCPU_PAGELIST_FRACTION,
1047                 .procname       = "percpu_pagelist_fraction",
1048                 .data           = &percpu_pagelist_fraction,
1049                 .maxlen         = sizeof(percpu_pagelist_fraction),
1050                 .mode           = 0644,
1051                 .proc_handler   = &percpu_pagelist_fraction_sysctl_handler,
1052                 .strategy       = &sysctl_intvec,
1053                 .extra1         = &min_percpu_pagelist_fract,
1054         },
1055 #ifdef CONFIG_MMU
1056         {
1057                 .ctl_name       = VM_MAX_MAP_COUNT,
1058                 .procname       = "max_map_count",
1059                 .data           = &sysctl_max_map_count,
1060                 .maxlen         = sizeof(sysctl_max_map_count),
1061                 .mode           = 0644,
1062                 .proc_handler   = &proc_dointvec
1063         },
1064 #endif
1065         {
1066                 .ctl_name       = VM_LAPTOP_MODE,
1067                 .procname       = "laptop_mode",
1068                 .data           = &laptop_mode,
1069                 .maxlen         = sizeof(laptop_mode),
1070                 .mode           = 0644,
1071                 .proc_handler   = &proc_dointvec_jiffies,
1072                 .strategy       = &sysctl_jiffies,
1073         },
1074         {
1075                 .ctl_name       = VM_BLOCK_DUMP,
1076                 .procname       = "block_dump",
1077                 .data           = &block_dump,
1078                 .maxlen         = sizeof(block_dump),
1079                 .mode           = 0644,
1080                 .proc_handler   = &proc_dointvec,
1081                 .strategy       = &sysctl_intvec,
1082                 .extra1         = &zero,
1083         },
1084         {
1085                 .ctl_name       = VM_VFS_CACHE_PRESSURE,
1086                 .procname       = "vfs_cache_pressure",
1087                 .data           = &sysctl_vfs_cache_pressure,
1088                 .maxlen         = sizeof(sysctl_vfs_cache_pressure),
1089                 .mode           = 0644,
1090                 .proc_handler   = &proc_dointvec,
1091                 .strategy       = &sysctl_intvec,
1092                 .extra1         = &zero,
1093         },
1094 #ifdef HAVE_ARCH_PICK_MMAP_LAYOUT
1095         {
1096                 .ctl_name       = VM_LEGACY_VA_LAYOUT,
1097                 .procname       = "legacy_va_layout",
1098                 .data           = &sysctl_legacy_va_layout,
1099                 .maxlen         = sizeof(sysctl_legacy_va_layout),
1100                 .mode           = 0644,
1101                 .proc_handler   = &proc_dointvec,
1102                 .strategy       = &sysctl_intvec,
1103                 .extra1         = &zero,
1104         },
1105 #endif
1106 #ifdef CONFIG_NUMA
1107         {
1108                 .ctl_name       = VM_ZONE_RECLAIM_MODE,
1109                 .procname       = "zone_reclaim_mode",
1110                 .data           = &zone_reclaim_mode,
1111                 .maxlen         = sizeof(zone_reclaim_mode),
1112                 .mode           = 0644,
1113                 .proc_handler   = &proc_dointvec,
1114                 .strategy       = &sysctl_intvec,
1115                 .extra1         = &zero,
1116         },
1117         {
1118                 .ctl_name       = VM_MIN_UNMAPPED,
1119                 .procname       = "min_unmapped_ratio",
1120                 .data           = &sysctl_min_unmapped_ratio,
1121                 .maxlen         = sizeof(sysctl_min_unmapped_ratio),
1122                 .mode           = 0644,
1123                 .proc_handler   = &sysctl_min_unmapped_ratio_sysctl_handler,
1124                 .strategy       = &sysctl_intvec,
1125                 .extra1         = &zero,
1126                 .extra2         = &one_hundred,
1127         },
1128         {
1129                 .ctl_name       = VM_MIN_SLAB,
1130                 .procname       = "min_slab_ratio",
1131                 .data           = &sysctl_min_slab_ratio,
1132                 .maxlen         = sizeof(sysctl_min_slab_ratio),
1133                 .mode           = 0644,
1134                 .proc_handler   = &sysctl_min_slab_ratio_sysctl_handler,
1135                 .strategy       = &sysctl_intvec,
1136                 .extra1         = &zero,
1137                 .extra2         = &one_hundred,
1138         },
1139 #endif
1140 #ifdef CONFIG_SMP
1141         {
1142                 .ctl_name       = CTL_UNNUMBERED,
1143                 .procname       = "stat_interval",
1144                 .data           = &sysctl_stat_interval,
1145                 .maxlen         = sizeof(sysctl_stat_interval),
1146                 .mode           = 0644,
1147                 .proc_handler   = &proc_dointvec_jiffies,
1148                 .strategy       = &sysctl_jiffies,
1149         },
1150 #endif
1151 #ifdef CONFIG_SECURITY
1152         {
1153                 .ctl_name       = CTL_UNNUMBERED,
1154                 .procname       = "mmap_min_addr",
1155                 .data           = &mmap_min_addr,
1156                 .maxlen         = sizeof(unsigned long),
1157                 .mode           = 0644,
1158                 .proc_handler   = &proc_doulongvec_minmax,
1159         },
1160 #endif
1161 #ifdef CONFIG_NUMA
1162         {
1163                 .ctl_name       = CTL_UNNUMBERED,
1164                 .procname       = "numa_zonelist_order",
1165                 .data           = &numa_zonelist_order,
1166                 .maxlen         = NUMA_ZONELIST_ORDER_LEN,
1167                 .mode           = 0644,
1168                 .proc_handler   = &numa_zonelist_order_handler,
1169                 .strategy       = &sysctl_string,
1170         },
1171 #endif
1172 #if (defined(CONFIG_X86_32) && !defined(CONFIG_UML))|| \
1173    (defined(CONFIG_SUPERH) && defined(CONFIG_VSYSCALL))
1174         {
1175                 .ctl_name       = VM_VDSO_ENABLED,
1176                 .procname       = "vdso_enabled",
1177                 .data           = &vdso_enabled,
1178                 .maxlen         = sizeof(vdso_enabled),
1179                 .mode           = 0644,
1180                 .proc_handler   = &proc_dointvec,
1181                 .strategy       = &sysctl_intvec,
1182                 .extra1         = &zero,
1183         },
1184 #endif
1185 #ifdef CONFIG_HIGHMEM
1186         {
1187                 .ctl_name       = CTL_UNNUMBERED,
1188                 .procname       = "highmem_is_dirtyable",
1189                 .data           = &vm_highmem_is_dirtyable,
1190                 .maxlen         = sizeof(vm_highmem_is_dirtyable),
1191                 .mode           = 0644,
1192                 .proc_handler   = &proc_dointvec_minmax,
1193                 .strategy       = &sysctl_intvec,
1194                 .extra1         = &zero,
1195                 .extra2         = &one,
1196         },
1197 #endif
1198 /*
1199  * NOTE: do not add new entries to this table unless you have read
1200  * Documentation/sysctl/ctl_unnumbered.txt
1201  */
1202         { .ctl_name = 0 }
1203 };
1204
1205 #if defined(CONFIG_BINFMT_MISC) || defined(CONFIG_BINFMT_MISC_MODULE)
1206 static struct ctl_table binfmt_misc_table[] = {
1207         { .ctl_name = 0 }
1208 };
1209 #endif
1210
1211 static struct ctl_table fs_table[] = {
1212         {
1213                 .ctl_name       = FS_NRINODE,
1214                 .procname       = "inode-nr",
1215                 .data           = &inodes_stat,
1216                 .maxlen         = 2*sizeof(int),
1217                 .mode           = 0444,
1218                 .proc_handler   = &proc_dointvec,
1219         },
1220         {
1221                 .ctl_name       = FS_STATINODE,
1222                 .procname       = "inode-state",
1223                 .data           = &inodes_stat,
1224                 .maxlen         = 7*sizeof(int),
1225                 .mode           = 0444,
1226                 .proc_handler   = &proc_dointvec,
1227         },
1228         {
1229                 .procname       = "file-nr",
1230                 .data           = &files_stat,
1231                 .maxlen         = 3*sizeof(int),
1232                 .mode           = 0444,
1233                 .proc_handler   = &proc_nr_files,
1234         },
1235         {
1236                 .ctl_name       = FS_MAXFILE,
1237                 .procname       = "file-max",
1238                 .data           = &files_stat.max_files,
1239                 .maxlen         = sizeof(int),
1240                 .mode           = 0644,
1241                 .proc_handler   = &proc_dointvec,
1242         },
1243         {
1244                 .ctl_name       = CTL_UNNUMBERED,
1245                 .procname       = "nr_open",
1246                 .data           = &sysctl_nr_open,
1247                 .maxlen         = sizeof(int),
1248                 .mode           = 0644,
1249                 .proc_handler   = &proc_dointvec_minmax,
1250                 .extra1         = &sysctl_nr_open_min,
1251                 .extra2         = &sysctl_nr_open_max,
1252         },
1253         {
1254                 .ctl_name       = FS_DENTRY,
1255                 .procname       = "dentry-state",
1256                 .data           = &dentry_stat,
1257                 .maxlen         = 6*sizeof(int),
1258                 .mode           = 0444,
1259                 .proc_handler   = &proc_dointvec,
1260         },
1261         {
1262                 .ctl_name       = FS_OVERFLOWUID,
1263                 .procname       = "overflowuid",
1264                 .data           = &fs_overflowuid,
1265                 .maxlen         = sizeof(int),
1266                 .mode           = 0644,
1267                 .proc_handler   = &proc_dointvec_minmax,
1268                 .strategy       = &sysctl_intvec,
1269                 .extra1         = &minolduid,
1270                 .extra2         = &maxolduid,
1271         },
1272         {
1273                 .ctl_name       = FS_OVERFLOWGID,
1274                 .procname       = "overflowgid",
1275                 .data           = &fs_overflowgid,
1276                 .maxlen         = sizeof(int),
1277                 .mode           = 0644,
1278                 .proc_handler   = &proc_dointvec_minmax,
1279                 .strategy       = &sysctl_intvec,
1280                 .extra1         = &minolduid,
1281                 .extra2         = &maxolduid,
1282         },
1283 #ifdef CONFIG_FILE_LOCKING
1284         {
1285                 .ctl_name       = FS_LEASES,
1286                 .procname       = "leases-enable",
1287                 .data           = &leases_enable,
1288                 .maxlen         = sizeof(int),
1289                 .mode           = 0644,
1290                 .proc_handler   = &proc_dointvec,
1291         },
1292 #endif
1293 #ifdef CONFIG_DNOTIFY
1294         {
1295                 .ctl_name       = FS_DIR_NOTIFY,
1296                 .procname       = "dir-notify-enable",
1297                 .data           = &dir_notify_enable,
1298                 .maxlen         = sizeof(int),
1299                 .mode           = 0644,
1300                 .proc_handler   = &proc_dointvec,
1301         },
1302 #endif
1303 #ifdef CONFIG_MMU
1304 #ifdef CONFIG_FILE_LOCKING
1305         {
1306                 .ctl_name       = FS_LEASE_TIME,
1307                 .procname       = "lease-break-time",
1308                 .data           = &lease_break_time,
1309                 .maxlen         = sizeof(int),
1310                 .mode           = 0644,
1311                 .proc_handler   = &proc_dointvec_minmax,
1312                 .strategy       = &sysctl_intvec,
1313                 .extra1         = &zero,
1314                 .extra2         = &two,
1315         },
1316 #endif
1317 #ifdef CONFIG_AIO
1318         {
1319                 .procname       = "aio-nr",
1320                 .data           = &aio_nr,
1321                 .maxlen         = sizeof(aio_nr),
1322                 .mode           = 0444,
1323                 .proc_handler   = &proc_doulongvec_minmax,
1324         },
1325         {
1326                 .procname       = "aio-max-nr",
1327                 .data           = &aio_max_nr,
1328                 .maxlen         = sizeof(aio_max_nr),
1329                 .mode           = 0644,
1330                 .proc_handler   = &proc_doulongvec_minmax,
1331         },
1332 #endif /* CONFIG_AIO */
1333 #ifdef CONFIG_INOTIFY_USER
1334         {
1335                 .ctl_name       = FS_INOTIFY,
1336                 .procname       = "inotify",
1337                 .mode           = 0555,
1338                 .child          = inotify_table,
1339         },
1340 #endif
1341 #ifdef CONFIG_EPOLL
1342         {
1343                 .procname       = "epoll",
1344                 .mode           = 0555,
1345                 .child          = epoll_table,
1346         },
1347 #endif
1348 #endif
1349         {
1350                 .ctl_name       = KERN_SETUID_DUMPABLE,
1351                 .procname       = "suid_dumpable",
1352                 .data           = &suid_dumpable,
1353                 .maxlen         = sizeof(int),
1354                 .mode           = 0644,
1355                 .proc_handler   = &proc_dointvec,
1356         },
1357 #if defined(CONFIG_BINFMT_MISC) || defined(CONFIG_BINFMT_MISC_MODULE)
1358         {
1359                 .ctl_name       = CTL_UNNUMBERED,
1360                 .procname       = "binfmt_misc",
1361                 .mode           = 0555,
1362                 .child          = binfmt_misc_table,
1363         },
1364 #endif
1365 /*
1366  * NOTE: do not add new entries to this table unless you have read
1367  * Documentation/sysctl/ctl_unnumbered.txt
1368  */
1369         { .ctl_name = 0 }
1370 };
1371
1372 static struct ctl_table debug_table[] = {
1373 #if defined(CONFIG_X86) || defined(CONFIG_PPC)
1374         {
1375                 .ctl_name       = CTL_UNNUMBERED,
1376                 .procname       = "exception-trace",
1377                 .data           = &show_unhandled_signals,
1378                 .maxlen         = sizeof(int),
1379                 .mode           = 0644,
1380                 .proc_handler   = proc_dointvec
1381         },
1382 #endif
1383         { .ctl_name = 0 }
1384 };
1385
1386 static struct ctl_table dev_table[] = {
1387         { .ctl_name = 0 }
1388 };
1389
1390 static DEFINE_SPINLOCK(sysctl_lock);
1391
1392 /* called under sysctl_lock */
1393 static int use_table(struct ctl_table_header *p)
1394 {
1395         if (unlikely(p->unregistering))
1396                 return 0;
1397         p->used++;
1398         return 1;
1399 }
1400
1401 /* called under sysctl_lock */
1402 static void unuse_table(struct ctl_table_header *p)
1403 {
1404         if (!--p->used)
1405                 if (unlikely(p->unregistering))
1406                         complete(p->unregistering);
1407 }
1408
1409 /* called under sysctl_lock, will reacquire if has to wait */
1410 static void start_unregistering(struct ctl_table_header *p)
1411 {
1412         /*
1413          * if p->used is 0, nobody will ever touch that entry again;
1414          * we'll eliminate all paths to it before dropping sysctl_lock
1415          */
1416         if (unlikely(p->used)) {
1417                 struct completion wait;
1418                 init_completion(&wait);
1419                 p->unregistering = &wait;
1420                 spin_unlock(&sysctl_lock);
1421                 wait_for_completion(&wait);
1422                 spin_lock(&sysctl_lock);
1423         } else {
1424                 /* anything non-NULL; we'll never dereference it */
1425                 p->unregistering = ERR_PTR(-EINVAL);
1426         }
1427         /*
1428          * do not remove from the list until nobody holds it; walking the
1429          * list in do_sysctl() relies on that.
1430          */
1431         list_del_init(&p->ctl_entry);
1432 }
1433
1434 void sysctl_head_get(struct ctl_table_header *head)
1435 {
1436         spin_lock(&sysctl_lock);
1437         head->count++;
1438         spin_unlock(&sysctl_lock);
1439 }
1440
1441 void sysctl_head_put(struct ctl_table_header *head)
1442 {
1443         spin_lock(&sysctl_lock);
1444         if (!--head->count)
1445                 kfree(head);
1446         spin_unlock(&sysctl_lock);
1447 }
1448
1449 struct ctl_table_header *sysctl_head_grab(struct ctl_table_header *head)
1450 {
1451         if (!head)
1452                 BUG();
1453         spin_lock(&sysctl_lock);
1454         if (!use_table(head))
1455                 head = ERR_PTR(-ENOENT);
1456         spin_unlock(&sysctl_lock);
1457         return head;
1458 }
1459
1460 void sysctl_head_finish(struct ctl_table_header *head)
1461 {
1462         if (!head)
1463                 return;
1464         spin_lock(&sysctl_lock);
1465         unuse_table(head);
1466         spin_unlock(&sysctl_lock);
1467 }
1468
1469 static struct ctl_table_set *
1470 lookup_header_set(struct ctl_table_root *root, struct nsproxy *namespaces)
1471 {
1472         struct ctl_table_set *set = &root->default_set;
1473         if (root->lookup)
1474                 set = root->lookup(root, namespaces);
1475         return set;
1476 }
1477
1478 static struct list_head *
1479 lookup_header_list(struct ctl_table_root *root, struct nsproxy *namespaces)
1480 {
1481         struct ctl_table_set *set = lookup_header_set(root, namespaces);
1482         return &set->list;
1483 }
1484
1485 struct ctl_table_header *__sysctl_head_next(struct nsproxy *namespaces,
1486                                             struct ctl_table_header *prev)
1487 {
1488         struct ctl_table_root *root;
1489         struct list_head *header_list;
1490         struct ctl_table_header *head;
1491         struct list_head *tmp;
1492
1493         spin_lock(&sysctl_lock);
1494         if (prev) {
1495                 head = prev;
1496                 tmp = &prev->ctl_entry;
1497                 unuse_table(prev);
1498                 goto next;
1499         }
1500         tmp = &root_table_header.ctl_entry;
1501         for (;;) {
1502                 head = list_entry(tmp, struct ctl_table_header, ctl_entry);
1503
1504                 if (!use_table(head))
1505                         goto next;
1506                 spin_unlock(&sysctl_lock);
1507                 return head;
1508         next:
1509                 root = head->root;
1510                 tmp = tmp->next;
1511                 header_list = lookup_header_list(root, namespaces);
1512                 if (tmp != header_list)
1513                         continue;
1514
1515                 do {
1516                         root = list_entry(root->root_list.next,
1517                                         struct ctl_table_root, root_list);
1518                         if (root == &sysctl_table_root)
1519                                 goto out;
1520                         header_list = lookup_header_list(root, namespaces);
1521                 } while (list_empty(header_list));
1522                 tmp = header_list->next;
1523         }
1524 out:
1525         spin_unlock(&sysctl_lock);
1526         return NULL;
1527 }
1528
1529 struct ctl_table_header *sysctl_head_next(struct ctl_table_header *prev)
1530 {
1531         return __sysctl_head_next(current->nsproxy, prev);
1532 }
1533
1534 void register_sysctl_root(struct ctl_table_root *root)
1535 {
1536         spin_lock(&sysctl_lock);
1537         list_add_tail(&root->root_list, &sysctl_table_root.root_list);
1538         spin_unlock(&sysctl_lock);
1539 }
1540
1541 #ifdef CONFIG_SYSCTL_SYSCALL
1542 /* Perform the actual read/write of a sysctl table entry. */
1543 static int do_sysctl_strategy(struct ctl_table_root *root,
1544                         struct ctl_table *table,
1545                         void __user *oldval, size_t __user *oldlenp,
1546                         void __user *newval, size_t newlen)
1547 {
1548         int op = 0, rc;
1549
1550         if (oldval)
1551                 op |= MAY_READ;
1552         if (newval)
1553                 op |= MAY_WRITE;
1554         if (sysctl_perm(root, table, op))
1555                 return -EPERM;
1556
1557         if (table->strategy) {
1558                 rc = table->strategy(table, oldval, oldlenp, newval, newlen);
1559                 if (rc < 0)
1560                         return rc;
1561                 if (rc > 0)
1562                         return 0;
1563         }
1564
1565         /* If there is no strategy routine, or if the strategy returns
1566          * zero, proceed with automatic r/w */
1567         if (table->data && table->maxlen) {
1568                 rc = sysctl_data(table, oldval, oldlenp, newval, newlen);
1569                 if (rc < 0)
1570                         return rc;
1571         }
1572         return 0;
1573 }
1574
1575 static int parse_table(int __user *name, int nlen,
1576                        void __user *oldval, size_t __user *oldlenp,
1577                        void __user *newval, size_t newlen,
1578                        struct ctl_table_root *root,
1579                        struct ctl_table *table)
1580 {
1581         int n;
1582 repeat:
1583         if (!nlen)
1584                 return -ENOTDIR;
1585         if (get_user(n, name))
1586                 return -EFAULT;
1587         for ( ; table->ctl_name || table->procname; table++) {
1588                 if (!table->ctl_name)
1589                         continue;
1590                 if (n == table->ctl_name) {
1591                         int error;
1592                         if (table->child) {
1593                                 if (sysctl_perm(root, table, MAY_EXEC))
1594                                         return -EPERM;
1595                                 name++;
1596                                 nlen--;
1597                                 table = table->child;
1598                                 goto repeat;
1599                         }
1600                         error = do_sysctl_strategy(root, table,
1601                                                    oldval, oldlenp,
1602                                                    newval, newlen);
1603                         return error;
1604                 }
1605         }
1606         return -ENOTDIR;
1607 }
1608
1609 int do_sysctl(int __user *name, int nlen, void __user *oldval, size_t __user *oldlenp,
1610                void __user *newval, size_t newlen)
1611 {
1612         struct ctl_table_header *head;
1613         int error = -ENOTDIR;
1614
1615         if (nlen <= 0 || nlen >= CTL_MAXNAME)
1616                 return -ENOTDIR;
1617         if (oldval) {
1618                 int old_len;
1619                 if (!oldlenp || get_user(old_len, oldlenp))
1620                         return -EFAULT;
1621         }
1622
1623         for (head = sysctl_head_next(NULL); head;
1624                         head = sysctl_head_next(head)) {
1625                 error = parse_table(name, nlen, oldval, oldlenp,
1626                                         newval, newlen,
1627                                         head->root, head->ctl_table);
1628                 if (error != -ENOTDIR) {
1629                         sysctl_head_finish(head);
1630                         break;
1631                 }
1632         }
1633         return error;
1634 }
1635
1636 asmlinkage long sys_sysctl(struct __sysctl_args __user *args)
1637 {
1638         struct __sysctl_args tmp;
1639         int error;
1640
1641         if (copy_from_user(&tmp, args, sizeof(tmp)))
1642                 return -EFAULT;
1643
1644         error = deprecated_sysctl_warning(&tmp);
1645         if (error)
1646                 goto out;
1647
1648         lock_kernel();
1649         error = do_sysctl(tmp.name, tmp.nlen, tmp.oldval, tmp.oldlenp,
1650                           tmp.newval, tmp.newlen);
1651         unlock_kernel();
1652 out:
1653         return error;
1654 }
1655 #endif /* CONFIG_SYSCTL_SYSCALL */
1656
1657 /*
1658  * sysctl_perm does NOT grant the superuser all rights automatically, because
1659  * some sysctl variables are readonly even to root.
1660  */
1661
1662 static int test_perm(int mode, int op)
1663 {
1664         if (!current->euid)
1665                 mode >>= 6;
1666         else if (in_egroup_p(0))
1667                 mode >>= 3;
1668         if ((op & ~mode & (MAY_READ|MAY_WRITE|MAY_EXEC)) == 0)
1669                 return 0;
1670         return -EACCES;
1671 }
1672
1673 int sysctl_perm(struct ctl_table_root *root, struct ctl_table *table, int op)
1674 {
1675         int error;
1676         int mode;
1677
1678         error = security_sysctl(table, op & (MAY_READ | MAY_WRITE | MAY_EXEC));
1679         if (error)
1680                 return error;
1681
1682         if (root->permissions)
1683                 mode = root->permissions(root, current->nsproxy, table);
1684         else
1685                 mode = table->mode;
1686
1687         return test_perm(mode, op);
1688 }
1689
1690 static void sysctl_set_parent(struct ctl_table *parent, struct ctl_table *table)
1691 {
1692         for (; table->ctl_name || table->procname; table++) {
1693                 table->parent = parent;
1694                 if (table->child)
1695                         sysctl_set_parent(table, table->child);
1696         }
1697 }
1698
1699 static __init int sysctl_init(void)
1700 {
1701         sysctl_set_parent(NULL, root_table);
1702 #ifdef CONFIG_SYSCTL_SYSCALL_CHECK
1703         {
1704                 int err;
1705                 err = sysctl_check_table(current->nsproxy, root_table);
1706         }
1707 #endif
1708         return 0;
1709 }
1710
1711 core_initcall(sysctl_init);
1712
1713 static struct ctl_table *is_branch_in(struct ctl_table *branch,
1714                                       struct ctl_table *table)
1715 {
1716         struct ctl_table *p;
1717         const char *s = branch->procname;
1718
1719         /* branch should have named subdirectory as its first element */
1720         if (!s || !branch->child)
1721                 return NULL;
1722
1723         /* ... and nothing else */
1724         if (branch[1].procname || branch[1].ctl_name)
1725                 return NULL;
1726
1727         /* table should contain subdirectory with the same name */
1728         for (p = table; p->procname || p->ctl_name; p++) {
1729                 if (!p->child)
1730                         continue;
1731                 if (p->procname && strcmp(p->procname, s) == 0)
1732                         return p;
1733         }
1734         return NULL;
1735 }
1736
1737 /* see if attaching q to p would be an improvement */
1738 static void try_attach(struct ctl_table_header *p, struct ctl_table_header *q)
1739 {
1740         struct ctl_table *to = p->ctl_table, *by = q->ctl_table;
1741         struct ctl_table *next;
1742         int is_better = 0;
1743         int not_in_parent = !p->attached_by;
1744
1745         while ((next = is_branch_in(by, to)) != NULL) {
1746                 if (by == q->attached_by)
1747                         is_better = 1;
1748                 if (to == p->attached_by)
1749                         not_in_parent = 1;
1750                 by = by->child;
1751                 to = next->child;
1752         }
1753
1754         if (is_better && not_in_parent) {
1755                 q->attached_by = by;
1756                 q->attached_to = to;
1757                 q->parent = p;
1758         }
1759 }
1760
1761 /**
1762  * __register_sysctl_paths - register a sysctl hierarchy
1763  * @root: List of sysctl headers to register on
1764  * @namespaces: Data to compute which lists of sysctl entries are visible
1765  * @path: The path to the directory the sysctl table is in.
1766  * @table: the top-level table structure
1767  *
1768  * Register a sysctl table hierarchy. @table should be a filled in ctl_table
1769  * array. A completely 0 filled entry terminates the table.
1770  *
1771  * The members of the &struct ctl_table structure are used as follows:
1772  *
1773  * ctl_name - This is the numeric sysctl value used by sysctl(2). The number
1774  *            must be unique within that level of sysctl
1775  *
1776  * procname - the name of the sysctl file under /proc/sys. Set to %NULL to not
1777  *            enter a sysctl file
1778  *
1779  * data - a pointer to data for use by proc_handler
1780  *
1781  * maxlen - the maximum size in bytes of the data
1782  *
1783  * mode - the file permissions for the /proc/sys file, and for sysctl(2)
1784  *
1785  * child - a pointer to the child sysctl table if this entry is a directory, or
1786  *         %NULL.
1787  *
1788  * proc_handler - the text handler routine (described below)
1789  *
1790  * strategy - the strategy routine (described below)
1791  *
1792  * de - for internal use by the sysctl routines
1793  *
1794  * extra1, extra2 - extra pointers usable by the proc handler routines
1795  *
1796  * Leaf nodes in the sysctl tree will be represented by a single file
1797  * under /proc; non-leaf nodes will be represented by directories.
1798  *
1799  * sysctl(2) can automatically manage read and write requests through
1800  * the sysctl table.  The data and maxlen fields of the ctl_table
1801  * struct enable minimal validation of the values being written to be
1802  * performed, and the mode field allows minimal authentication.
1803  *
1804  * More sophisticated management can be enabled by the provision of a
1805  * strategy routine with the table entry.  This will be called before
1806  * any automatic read or write of the data is performed.
1807  *
1808  * The strategy routine may return
1809  *
1810  * < 0 - Error occurred (error is passed to user process)
1811  *
1812  * 0   - OK - proceed with automatic read or write.
1813  *
1814  * > 0 - OK - read or write has been done by the strategy routine, so
1815  *       return immediately.
1816  *
1817  * There must be a proc_handler routine for any terminal nodes
1818  * mirrored under /proc/sys (non-terminals are handled by a built-in
1819  * directory handler).  Several default handlers are available to
1820  * cover common cases -
1821  *
1822  * proc_dostring(), proc_dointvec(), proc_dointvec_jiffies(),
1823  * proc_dointvec_userhz_jiffies(), proc_dointvec_minmax(),
1824  * proc_doulongvec_ms_jiffies_minmax(), proc_doulongvec_minmax()
1825  *
1826  * It is the handler's job to read the input buffer from user memory
1827  * and process it. The handler should return 0 on success.
1828  *
1829  * This routine returns %NULL on a failure to register, and a pointer
1830  * to the table header on success.
1831  */
1832 struct ctl_table_header *__register_sysctl_paths(
1833         struct ctl_table_root *root,
1834         struct nsproxy *namespaces,
1835         const struct ctl_path *path, struct ctl_table *table)
1836 {
1837         struct ctl_table_header *header;
1838         struct ctl_table *new, **prevp;
1839         unsigned int n, npath;
1840         struct ctl_table_set *set;
1841
1842         /* Count the path components */
1843         for (npath = 0; path[npath].ctl_name || path[npath].procname; ++npath)
1844                 ;
1845
1846         /*
1847          * For each path component, allocate a 2-element ctl_table array.
1848          * The first array element will be filled with the sysctl entry
1849          * for this, the second will be the sentinel (ctl_name == 0).
1850          *
1851          * We allocate everything in one go so that we don't have to
1852          * worry about freeing additional memory in unregister_sysctl_table.
1853          */
1854         header = kzalloc(sizeof(struct ctl_table_header) +
1855                          (2 * npath * sizeof(struct ctl_table)), GFP_KERNEL);
1856         if (!header)
1857                 return NULL;
1858
1859         new = (struct ctl_table *) (header + 1);
1860
1861         /* Now connect the dots */
1862         prevp = &header->ctl_table;
1863         for (n = 0; n < npath; ++n, ++path) {
1864                 /* Copy the procname */
1865                 new->procname = path->procname;
1866                 new->ctl_name = path->ctl_name;
1867                 new->mode     = 0555;
1868
1869                 *prevp = new;
1870                 prevp = &new->child;
1871
1872                 new += 2;
1873         }
1874         *prevp = table;
1875         header->ctl_table_arg = table;
1876
1877         INIT_LIST_HEAD(&header->ctl_entry);
1878         header->used = 0;
1879         header->unregistering = NULL;
1880         header->root = root;
1881         sysctl_set_parent(NULL, header->ctl_table);
1882         header->count = 1;
1883 #ifdef CONFIG_SYSCTL_SYSCALL_CHECK
1884         if (sysctl_check_table(namespaces, header->ctl_table)) {
1885                 kfree(header);
1886                 return NULL;
1887         }
1888 #endif
1889         spin_lock(&sysctl_lock);
1890         header->set = lookup_header_set(root, namespaces);
1891         header->attached_by = header->ctl_table;
1892         header->attached_to = root_table;
1893         header->parent = &root_table_header;
1894         for (set = header->set; set; set = set->parent) {
1895                 struct ctl_table_header *p;
1896                 list_for_each_entry(p, &set->list, ctl_entry) {
1897                         if (p->unregistering)
1898                                 continue;
1899                         try_attach(p, header);
1900                 }
1901         }
1902         header->parent->count++;
1903         list_add_tail(&header->ctl_entry, &header->set->list);
1904         spin_unlock(&sysctl_lock);
1905
1906         return header;
1907 }
1908
1909 /**
1910  * register_sysctl_table_path - register a sysctl table hierarchy
1911  * @path: The path to the directory the sysctl table is in.
1912  * @table: the top-level table structure
1913  *
1914  * Register a sysctl table hierarchy. @table should be a filled in ctl_table
1915  * array. A completely 0 filled entry terminates the table.
1916  *
1917  * See __register_sysctl_paths for more details.
1918  */
1919 struct ctl_table_header *register_sysctl_paths(const struct ctl_path *path,
1920                                                 struct ctl_table *table)
1921 {
1922         return __register_sysctl_paths(&sysctl_table_root, current->nsproxy,
1923                                         path, table);
1924 }
1925
1926 /**
1927  * register_sysctl_table - register a sysctl table hierarchy
1928  * @table: the top-level table structure
1929  *
1930  * Register a sysctl table hierarchy. @table should be a filled in ctl_table
1931  * array. A completely 0 filled entry terminates the table.
1932  *
1933  * See register_sysctl_paths for more details.
1934  */
1935 struct ctl_table_header *register_sysctl_table(struct ctl_table *table)
1936 {
1937         static const struct ctl_path null_path[] = { {} };
1938
1939         return register_sysctl_paths(null_path, table);
1940 }
1941
1942 /**
1943  * unregister_sysctl_table - unregister a sysctl table hierarchy
1944  * @header: the header returned from register_sysctl_table
1945  *
1946  * Unregisters the sysctl table and all children. proc entries may not
1947  * actually be removed until they are no longer used by anyone.
1948  */
1949 void unregister_sysctl_table(struct ctl_table_header * header)
1950 {
1951         might_sleep();
1952
1953         if (header == NULL)
1954                 return;
1955
1956         spin_lock(&sysctl_lock);
1957         start_unregistering(header);
1958         if (!--header->parent->count) {
1959                 WARN_ON(1);
1960                 kfree(header->parent);
1961         }
1962         if (!--header->count)
1963                 kfree(header);
1964         spin_unlock(&sysctl_lock);
1965 }
1966
1967 int sysctl_is_seen(struct ctl_table_header *p)
1968 {
1969         struct ctl_table_set *set = p->set;
1970         int res;
1971         spin_lock(&sysctl_lock);
1972         if (p->unregistering)
1973                 res = 0;
1974         else if (!set->is_seen)
1975                 res = 1;
1976         else
1977                 res = set->is_seen(set);
1978         spin_unlock(&sysctl_lock);
1979         return res;
1980 }
1981
1982 void setup_sysctl_set(struct ctl_table_set *p,
1983         struct ctl_table_set *parent,
1984         int (*is_seen)(struct ctl_table_set *))
1985 {
1986         INIT_LIST_HEAD(&p->list);
1987         p->parent = parent ? parent : &sysctl_table_root.default_set;
1988         p->is_seen = is_seen;
1989 }
1990
1991 #else /* !CONFIG_SYSCTL */
1992 struct ctl_table_header *register_sysctl_table(struct ctl_table * table)
1993 {
1994         return NULL;
1995 }
1996
1997 struct ctl_table_header *register_sysctl_paths(const struct ctl_path *path,
1998                                                     struct ctl_table *table)
1999 {
2000         return NULL;
2001 }
2002
2003 void unregister_sysctl_table(struct ctl_table_header * table)
2004 {
2005 }
2006
2007 void setup_sysctl_set(struct ctl_table_set *p,
2008         struct ctl_table_set *parent,
2009         int (*is_seen)(struct ctl_table_set *))
2010 {
2011 }
2012
2013 void sysctl_head_put(struct ctl_table_header *head)
2014 {
2015 }
2016
2017 #endif /* CONFIG_SYSCTL */
2018
2019 /*
2020  * /proc/sys support
2021  */
2022
2023 #ifdef CONFIG_PROC_SYSCTL
2024
2025 static int _proc_do_string(void* data, int maxlen, int write,
2026                            struct file *filp, void __user *buffer,
2027                            size_t *lenp, loff_t *ppos)
2028 {
2029         size_t len;
2030         char __user *p;
2031         char c;
2032
2033         if (!data || !maxlen || !*lenp) {
2034                 *lenp = 0;
2035                 return 0;
2036         }
2037
2038         if (write) {
2039                 len = 0;
2040                 p = buffer;
2041                 while (len < *lenp) {
2042                         if (get_user(c, p++))
2043                                 return -EFAULT;
2044                         if (c == 0 || c == '\n')
2045                                 break;
2046                         len++;
2047                 }
2048                 if (len >= maxlen)
2049                         len = maxlen-1;
2050                 if(copy_from_user(data, buffer, len))
2051                         return -EFAULT;
2052                 ((char *) data)[len] = 0;
2053                 *ppos += *lenp;
2054         } else {
2055                 len = strlen(data);
2056                 if (len > maxlen)
2057                         len = maxlen;
2058
2059                 if (*ppos > len) {
2060                         *lenp = 0;
2061                         return 0;
2062                 }
2063
2064                 data += *ppos;
2065                 len  -= *ppos;
2066
2067                 if (len > *lenp)
2068                         len = *lenp;
2069                 if (len)
2070                         if(copy_to_user(buffer, data, len))
2071                                 return -EFAULT;
2072                 if (len < *lenp) {
2073                         if(put_user('\n', ((char __user *) buffer) + len))
2074                                 return -EFAULT;
2075                         len++;
2076                 }
2077                 *lenp = len;
2078                 *ppos += len;
2079         }
2080         return 0;
2081 }
2082
2083 /**
2084  * proc_dostring - read a string sysctl
2085  * @table: the sysctl table
2086  * @write: %TRUE if this is a write to the sysctl file
2087  * @filp: the file structure
2088  * @buffer: the user buffer
2089  * @lenp: the size of the user buffer
2090  * @ppos: file position
2091  *
2092  * Reads/writes a string from/to the user buffer. If the kernel
2093  * buffer provided is not large enough to hold the string, the
2094  * string is truncated. The copied string is %NULL-terminated.
2095  * If the string is being read by the user process, it is copied
2096  * and a newline '\n' is added. It is truncated if the buffer is
2097  * not large enough.
2098  *
2099  * Returns 0 on success.
2100  */
2101 int proc_dostring(struct ctl_table *table, int write, struct file *filp,
2102                   void __user *buffer, size_t *lenp, loff_t *ppos)
2103 {
2104         return _proc_do_string(table->data, table->maxlen, write, filp,
2105                                buffer, lenp, ppos);
2106 }
2107
2108
2109 static int do_proc_dointvec_conv(int *negp, unsigned long *lvalp,
2110                                  int *valp,
2111                                  int write, void *data)
2112 {
2113         if (write) {
2114                 *valp = *negp ? -*lvalp : *lvalp;
2115         } else {
2116                 int val = *valp;
2117                 if (val < 0) {
2118                         *negp = -1;
2119                         *lvalp = (unsigned long)-val;
2120                 } else {
2121                         *negp = 0;
2122                         *lvalp = (unsigned long)val;
2123                 }
2124         }
2125         return 0;
2126 }
2127
2128 static int __do_proc_dointvec(void *tbl_data, struct ctl_table *table,
2129                   int write, struct file *filp, void __user *buffer,
2130                   size_t *lenp, loff_t *ppos,
2131                   int (*conv)(int *negp, unsigned long *lvalp, int *valp,
2132                               int write, void *data),
2133                   void *data)
2134 {
2135 #define TMPBUFLEN 21
2136         int *i, vleft, first=1, neg, val;
2137         unsigned long lval;
2138         size_t left, len;
2139
2140         char buf[TMPBUFLEN], *p;
2141         char __user *s = buffer;
2142
2143         if (!tbl_data || !table->maxlen || !*lenp ||
2144             (*ppos && !write)) {
2145                 *lenp = 0;
2146                 return 0;
2147         }
2148
2149         i = (int *) tbl_data;
2150         vleft = table->maxlen / sizeof(*i);
2151         left = *lenp;
2152
2153         if (!conv)
2154                 conv = do_proc_dointvec_conv;
2155
2156         for (; left && vleft--; i++, first=0) {
2157                 if (write) {
2158                         while (left) {
2159                                 char c;
2160                                 if (get_user(c, s))
2161                                         return -EFAULT;
2162                                 if (!isspace(c))
2163                                         break;
2164                                 left--;
2165                                 s++;
2166                         }
2167                         if (!left)
2168                                 break;
2169                         neg = 0;
2170                         len = left;
2171                         if (len > sizeof(buf) - 1)
2172                                 len = sizeof(buf) - 1;
2173                         if (copy_from_user(buf, s, len))
2174                                 return -EFAULT;
2175                         buf[len] = 0;
2176                         p = buf;
2177                         if (*p == '-' && left > 1) {
2178                                 neg = 1;
2179                                 p++;
2180                         }
2181                         if (*p < '0' || *p > '9')
2182                                 break;
2183
2184                         lval = simple_strtoul(p, &p, 0);
2185
2186                         len = p-buf;
2187                         if ((len < left) && *p && !isspace(*p))
2188                                 break;
2189                         if (neg)
2190                                 val = -val;
2191                         s += len;
2192                         left -= len;
2193
2194                         if (conv(&neg, &lval, i, 1, data))
2195                                 break;
2196                 } else {
2197                         p = buf;
2198                         if (!first)
2199                                 *p++ = '\t';
2200
2201                         if (conv(&neg, &lval, i, 0, data))
2202                                 break;
2203
2204                         sprintf(p, "%s%lu", neg ? "-" : "", lval);
2205                         len = strlen(buf);
2206                         if (len > left)
2207                                 len = left;
2208                         if(copy_to_user(s, buf, len))
2209                                 return -EFAULT;
2210                         left -= len;
2211                         s += len;
2212                 }
2213         }
2214
2215         if (!write && !first && left) {
2216                 if(put_user('\n', s))
2217                         return -EFAULT;
2218                 left--, s++;
2219         }
2220         if (write) {
2221                 while (left) {
2222                         char c;
2223                         if (get_user(c, s++))
2224                                 return -EFAULT;
2225                         if (!isspace(c))
2226                                 break;
2227                         left--;
2228                 }
2229         }
2230         if (write && first)
2231                 return -EINVAL;
2232         *lenp -= left;
2233         *ppos += *lenp;
2234         return 0;
2235 #undef TMPBUFLEN
2236 }
2237
2238 static int do_proc_dointvec(struct ctl_table *table, int write, struct file *filp,
2239                   void __user *buffer, 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         return __do_proc_dointvec(table->data, table, write, filp,
2245                         buffer, lenp, ppos, conv, data);
2246 }
2247
2248 /**
2249  * proc_dointvec - read a vector of integers
2250  * @table: the sysctl table
2251  * @write: %TRUE if this is a write to the sysctl file
2252  * @filp: the file structure
2253  * @buffer: the user buffer
2254  * @lenp: the size of the user buffer
2255  * @ppos: file position
2256  *
2257  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2258  * values from/to the user buffer, treated as an ASCII string.
2259  *
2260  * Returns 0 on success.
2261  */
2262 int proc_dointvec(struct ctl_table *table, int write, struct file *filp,
2263                      void __user *buffer, size_t *lenp, loff_t *ppos)
2264 {
2265     return do_proc_dointvec(table,write,filp,buffer,lenp,ppos,
2266                             NULL,NULL);
2267 }
2268
2269 /*
2270  * Taint values can only be increased
2271  * This means we can safely use a temporary.
2272  */
2273 static int proc_taint(struct ctl_table *table, int write, struct file *filp,
2274                                void __user *buffer, size_t *lenp, loff_t *ppos)
2275 {
2276         struct ctl_table t;
2277         unsigned long tmptaint = get_taint();
2278         int err;
2279
2280         if (write && !capable(CAP_SYS_ADMIN))
2281                 return -EPERM;
2282
2283         t = *table;
2284         t.data = &tmptaint;
2285         err = proc_doulongvec_minmax(&t, write, filp, buffer, lenp, ppos);
2286         if (err < 0)
2287                 return err;
2288
2289         if (write) {
2290                 /*
2291                  * Poor man's atomic or. Not worth adding a primitive
2292                  * to everyone's atomic.h for this
2293                  */
2294                 int i;
2295                 for (i = 0; i < BITS_PER_LONG && tmptaint >> i; i++) {
2296                         if ((tmptaint >> i) & 1)
2297                                 add_taint(i);
2298                 }
2299         }
2300
2301         return err;
2302 }
2303
2304 struct do_proc_dointvec_minmax_conv_param {
2305         int *min;
2306         int *max;
2307 };
2308
2309 static int do_proc_dointvec_minmax_conv(int *negp, unsigned long *lvalp,
2310                                         int *valp,
2311                                         int write, void *data)
2312 {
2313         struct do_proc_dointvec_minmax_conv_param *param = data;
2314         if (write) {
2315                 int val = *negp ? -*lvalp : *lvalp;
2316                 if ((param->min && *param->min > val) ||
2317                     (param->max && *param->max < val))
2318                         return -EINVAL;
2319                 *valp = val;
2320         } else {
2321                 int val = *valp;
2322                 if (val < 0) {
2323                         *negp = -1;
2324                         *lvalp = (unsigned long)-val;
2325                 } else {
2326                         *negp = 0;
2327                         *lvalp = (unsigned long)val;
2328                 }
2329         }
2330         return 0;
2331 }
2332
2333 /**
2334  * proc_dointvec_minmax - read a vector of integers with min/max values
2335  * @table: the sysctl table
2336  * @write: %TRUE if this is a write to the sysctl file
2337  * @filp: the file structure
2338  * @buffer: the user buffer
2339  * @lenp: the size of the user buffer
2340  * @ppos: file position
2341  *
2342  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2343  * values from/to the user buffer, treated as an ASCII string.
2344  *
2345  * This routine will ensure the values are within the range specified by
2346  * table->extra1 (min) and table->extra2 (max).
2347  *
2348  * Returns 0 on success.
2349  */
2350 int proc_dointvec_minmax(struct ctl_table *table, int write, struct file *filp,
2351                   void __user *buffer, size_t *lenp, loff_t *ppos)
2352 {
2353         struct do_proc_dointvec_minmax_conv_param param = {
2354                 .min = (int *) table->extra1,
2355                 .max = (int *) table->extra2,
2356         };
2357         return do_proc_dointvec(table, write, filp, buffer, lenp, ppos,
2358                                 do_proc_dointvec_minmax_conv, &param);
2359 }
2360
2361 static int __do_proc_doulongvec_minmax(void *data, struct ctl_table *table, int write,
2362                                      struct file *filp,
2363                                      void __user *buffer,
2364                                      size_t *lenp, loff_t *ppos,
2365                                      unsigned long convmul,
2366                                      unsigned long convdiv)
2367 {
2368 #define TMPBUFLEN 21
2369         unsigned long *i, *min, *max, val;
2370         int vleft, first=1, neg;
2371         size_t len, left;
2372         char buf[TMPBUFLEN], *p;
2373         char __user *s = buffer;
2374
2375         if (!data || !table->maxlen || !*lenp ||
2376             (*ppos && !write)) {
2377                 *lenp = 0;
2378                 return 0;
2379         }
2380
2381         i = (unsigned long *) data;
2382         min = (unsigned long *) table->extra1;
2383         max = (unsigned long *) table->extra2;
2384         vleft = table->maxlen / sizeof(unsigned long);
2385         left = *lenp;
2386
2387         for (; left && vleft--; i++, min++, max++, first=0) {
2388                 if (write) {
2389                         while (left) {
2390                                 char c;
2391                                 if (get_user(c, s))
2392                                         return -EFAULT;
2393                                 if (!isspace(c))
2394                                         break;
2395                                 left--;
2396                                 s++;
2397                         }
2398                         if (!left)
2399                                 break;
2400                         neg = 0;
2401                         len = left;
2402                         if (len > TMPBUFLEN-1)
2403                                 len = TMPBUFLEN-1;
2404                         if (copy_from_user(buf, s, len))
2405                                 return -EFAULT;
2406                         buf[len] = 0;
2407                         p = buf;
2408                         if (*p == '-' && left > 1) {
2409                                 neg = 1;
2410                                 p++;
2411                         }
2412                         if (*p < '0' || *p > '9')
2413                                 break;
2414                         val = simple_strtoul(p, &p, 0) * convmul / convdiv ;
2415                         len = p-buf;
2416                         if ((len < left) && *p && !isspace(*p))
2417                                 break;
2418                         if (neg)
2419                                 val = -val;
2420                         s += len;
2421                         left -= len;
2422
2423                         if(neg)
2424                                 continue;
2425                         if ((min && val < *min) || (max && val > *max))
2426                                 continue;
2427                         *i = val;
2428                 } else {
2429                         p = buf;
2430                         if (!first)
2431                                 *p++ = '\t';
2432                         sprintf(p, "%lu", convdiv * (*i) / convmul);
2433                         len = strlen(buf);
2434                         if (len > left)
2435                                 len = left;
2436                         if(copy_to_user(s, buf, len))
2437                                 return -EFAULT;
2438                         left -= len;
2439                         s += len;
2440                 }
2441         }
2442
2443         if (!write && !first && left) {
2444                 if(put_user('\n', s))
2445                         return -EFAULT;
2446                 left--, s++;
2447         }
2448         if (write) {
2449                 while (left) {
2450                         char c;
2451                         if (get_user(c, s++))
2452                                 return -EFAULT;
2453                         if (!isspace(c))
2454                                 break;
2455                         left--;
2456                 }
2457         }
2458         if (write && first)
2459                 return -EINVAL;
2460         *lenp -= left;
2461         *ppos += *lenp;
2462         return 0;
2463 #undef TMPBUFLEN
2464 }
2465
2466 static int do_proc_doulongvec_minmax(struct ctl_table *table, int write,
2467                                      struct file *filp,
2468                                      void __user *buffer,
2469                                      size_t *lenp, loff_t *ppos,
2470                                      unsigned long convmul,
2471                                      unsigned long convdiv)
2472 {
2473         return __do_proc_doulongvec_minmax(table->data, table, write,
2474                         filp, buffer, lenp, ppos, convmul, convdiv);
2475 }
2476
2477 /**
2478  * proc_doulongvec_minmax - read a vector of long integers with min/max values
2479  * @table: the sysctl table
2480  * @write: %TRUE if this is a write to the sysctl file
2481  * @filp: the file structure
2482  * @buffer: the user buffer
2483  * @lenp: the size of the user buffer
2484  * @ppos: file position
2485  *
2486  * Reads/writes up to table->maxlen/sizeof(unsigned long) unsigned long
2487  * values from/to the user buffer, treated as an ASCII string.
2488  *
2489  * This routine will ensure the values are within the range specified by
2490  * table->extra1 (min) and table->extra2 (max).
2491  *
2492  * Returns 0 on success.
2493  */
2494 int proc_doulongvec_minmax(struct ctl_table *table, int write, struct file *filp,
2495                            void __user *buffer, size_t *lenp, loff_t *ppos)
2496 {
2497     return do_proc_doulongvec_minmax(table, write, filp, buffer, lenp, ppos, 1l, 1l);
2498 }
2499
2500 /**
2501  * proc_doulongvec_ms_jiffies_minmax - read a vector of millisecond values with min/max values
2502  * @table: the sysctl table
2503  * @write: %TRUE if this is a write to the sysctl file
2504  * @filp: the file structure
2505  * @buffer: the user buffer
2506  * @lenp: the size of the user buffer
2507  * @ppos: file position
2508  *
2509  * Reads/writes up to table->maxlen/sizeof(unsigned long) unsigned long
2510  * values from/to the user buffer, treated as an ASCII string. The values
2511  * are treated as milliseconds, and converted to jiffies when they are stored.
2512  *
2513  * This routine will ensure the values are within the range specified by
2514  * table->extra1 (min) and table->extra2 (max).
2515  *
2516  * Returns 0 on success.
2517  */
2518 int proc_doulongvec_ms_jiffies_minmax(struct ctl_table *table, int write,
2519                                       struct file *filp,
2520                                       void __user *buffer,
2521                                       size_t *lenp, loff_t *ppos)
2522 {
2523     return do_proc_doulongvec_minmax(table, write, filp, buffer,
2524                                      lenp, ppos, HZ, 1000l);
2525 }
2526
2527
2528 static int do_proc_dointvec_jiffies_conv(int *negp, unsigned long *lvalp,
2529                                          int *valp,
2530                                          int write, void *data)
2531 {
2532         if (write) {
2533                 if (*lvalp > LONG_MAX / HZ)
2534                         return 1;
2535                 *valp = *negp ? -(*lvalp*HZ) : (*lvalp*HZ);
2536         } else {
2537                 int val = *valp;
2538                 unsigned long lval;
2539                 if (val < 0) {
2540                         *negp = -1;
2541                         lval = (unsigned long)-val;
2542                 } else {
2543                         *negp = 0;
2544                         lval = (unsigned long)val;
2545                 }
2546                 *lvalp = lval / HZ;
2547         }
2548         return 0;
2549 }
2550
2551 static int do_proc_dointvec_userhz_jiffies_conv(int *negp, unsigned long *lvalp,
2552                                                 int *valp,
2553                                                 int write, void *data)
2554 {
2555         if (write) {
2556                 if (USER_HZ < HZ && *lvalp > (LONG_MAX / HZ) * USER_HZ)
2557                         return 1;
2558                 *valp = clock_t_to_jiffies(*negp ? -*lvalp : *lvalp);
2559         } else {
2560                 int val = *valp;
2561                 unsigned long lval;
2562                 if (val < 0) {
2563                         *negp = -1;
2564                         lval = (unsigned long)-val;
2565                 } else {
2566                         *negp = 0;
2567                         lval = (unsigned long)val;
2568                 }
2569                 *lvalp = jiffies_to_clock_t(lval);
2570         }
2571         return 0;
2572 }
2573
2574 static int do_proc_dointvec_ms_jiffies_conv(int *negp, unsigned long *lvalp,
2575                                             int *valp,
2576                                             int write, void *data)
2577 {
2578         if (write) {
2579                 *valp = msecs_to_jiffies(*negp ? -*lvalp : *lvalp);
2580         } else {
2581                 int val = *valp;
2582                 unsigned long lval;
2583                 if (val < 0) {
2584                         *negp = -1;
2585                         lval = (unsigned long)-val;
2586                 } else {
2587                         *negp = 0;
2588                         lval = (unsigned long)val;
2589                 }
2590                 *lvalp = jiffies_to_msecs(lval);
2591         }
2592         return 0;
2593 }
2594
2595 /**
2596  * proc_dointvec_jiffies - read a vector of integers as seconds
2597  * @table: the sysctl table
2598  * @write: %TRUE if this is a write to the sysctl file
2599  * @filp: the file structure
2600  * @buffer: the user buffer
2601  * @lenp: the size of the user buffer
2602  * @ppos: file position
2603  *
2604  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2605  * values from/to the user buffer, treated as an ASCII string.
2606  * The values read are assumed to be in seconds, and are converted into
2607  * jiffies.
2608  *
2609  * Returns 0 on success.
2610  */
2611 int proc_dointvec_jiffies(struct ctl_table *table, int write, struct file *filp,
2612                           void __user *buffer, size_t *lenp, loff_t *ppos)
2613 {
2614     return do_proc_dointvec(table,write,filp,buffer,lenp,ppos,
2615                             do_proc_dointvec_jiffies_conv,NULL);
2616 }
2617
2618 /**
2619  * proc_dointvec_userhz_jiffies - read a vector of integers as 1/USER_HZ seconds
2620  * @table: the sysctl table
2621  * @write: %TRUE if this is a write to the sysctl file
2622  * @filp: the file structure
2623  * @buffer: the user buffer
2624  * @lenp: the size of the user buffer
2625  * @ppos: pointer to the file position
2626  *
2627  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2628  * values from/to the user buffer, treated as an ASCII string.
2629  * The values read are assumed to be in 1/USER_HZ seconds, and
2630  * are converted into jiffies.
2631  *
2632  * Returns 0 on success.
2633  */
2634 int proc_dointvec_userhz_jiffies(struct ctl_table *table, int write, struct file *filp,
2635                                  void __user *buffer, size_t *lenp, loff_t *ppos)
2636 {
2637     return do_proc_dointvec(table,write,filp,buffer,lenp,ppos,
2638                             do_proc_dointvec_userhz_jiffies_conv,NULL);
2639 }
2640
2641 /**
2642  * proc_dointvec_ms_jiffies - read a vector of integers as 1 milliseconds
2643  * @table: the sysctl table
2644  * @write: %TRUE if this is a write to the sysctl file
2645  * @filp: the file structure
2646  * @buffer: the user buffer
2647  * @lenp: the size of the user buffer
2648  * @ppos: file position
2649  * @ppos: the current position in the file
2650  *
2651  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2652  * values from/to the user buffer, treated as an ASCII string.
2653  * The values read are assumed to be in 1/1000 seconds, and
2654  * are converted into jiffies.
2655  *
2656  * Returns 0 on success.
2657  */
2658 int proc_dointvec_ms_jiffies(struct ctl_table *table, int write, struct file *filp,
2659                              void __user *buffer, size_t *lenp, loff_t *ppos)
2660 {
2661         return do_proc_dointvec(table, write, filp, buffer, lenp, ppos,
2662                                 do_proc_dointvec_ms_jiffies_conv, NULL);
2663 }
2664
2665 static int proc_do_cad_pid(struct ctl_table *table, int write, struct file *filp,
2666                            void __user *buffer, size_t *lenp, loff_t *ppos)
2667 {
2668         struct pid *new_pid;
2669         pid_t tmp;
2670         int r;
2671
2672         tmp = pid_vnr(cad_pid);
2673
2674         r = __do_proc_dointvec(&tmp, table, write, filp, buffer,
2675                                lenp, ppos, NULL, NULL);
2676         if (r || !write)
2677                 return r;
2678
2679         new_pid = find_get_pid(tmp);
2680         if (!new_pid)
2681                 return -ESRCH;
2682
2683         put_pid(xchg(&cad_pid, new_pid));
2684         return 0;
2685 }
2686
2687 #else /* CONFIG_PROC_FS */
2688
2689 int proc_dostring(struct ctl_table *table, int write, struct file *filp,
2690                   void __user *buffer, size_t *lenp, loff_t *ppos)
2691 {
2692         return -ENOSYS;
2693 }
2694
2695 int proc_dointvec(struct ctl_table *table, int write, struct file *filp,
2696                   void __user *buffer, size_t *lenp, loff_t *ppos)
2697 {
2698         return -ENOSYS;
2699 }
2700
2701 int proc_dointvec_minmax(struct ctl_table *table, int write, struct file *filp,
2702                     void __user *buffer, size_t *lenp, loff_t *ppos)
2703 {
2704         return -ENOSYS;
2705 }
2706
2707 int proc_dointvec_jiffies(struct ctl_table *table, int write, struct file *filp,
2708                     void __user *buffer, size_t *lenp, loff_t *ppos)
2709 {
2710         return -ENOSYS;
2711 }
2712
2713 int proc_dointvec_userhz_jiffies(struct ctl_table *table, int write, struct file *filp,
2714                     void __user *buffer, size_t *lenp, loff_t *ppos)
2715 {
2716         return -ENOSYS;
2717 }
2718
2719 int proc_dointvec_ms_jiffies(struct ctl_table *table, int write, struct file *filp,
2720                              void __user *buffer, size_t *lenp, loff_t *ppos)
2721 {
2722         return -ENOSYS;
2723 }
2724
2725 int proc_doulongvec_minmax(struct ctl_table *table, int write, struct file *filp,
2726                     void __user *buffer, size_t *lenp, loff_t *ppos)
2727 {
2728         return -ENOSYS;
2729 }
2730
2731 int proc_doulongvec_ms_jiffies_minmax(struct ctl_table *table, int write,
2732                                       struct file *filp,
2733                                       void __user *buffer,
2734                                       size_t *lenp, loff_t *ppos)
2735 {
2736     return -ENOSYS;
2737 }
2738
2739
2740 #endif /* CONFIG_PROC_FS */
2741
2742
2743 #ifdef CONFIG_SYSCTL_SYSCALL
2744 /*
2745  * General sysctl support routines
2746  */
2747
2748 /* The generic sysctl data routine (used if no strategy routine supplied) */
2749 int sysctl_data(struct ctl_table *table,
2750                 void __user *oldval, size_t __user *oldlenp,
2751                 void __user *newval, size_t newlen)
2752 {
2753         size_t len;
2754
2755         /* Get out of I don't have a variable */
2756         if (!table->data || !table->maxlen)
2757                 return -ENOTDIR;
2758
2759         if (oldval && oldlenp) {
2760                 if (get_user(len, oldlenp))
2761                         return -EFAULT;
2762                 if (len) {
2763                         if (len > table->maxlen)
2764                                 len = table->maxlen;
2765                         if (copy_to_user(oldval, table->data, len))
2766                                 return -EFAULT;
2767                         if (put_user(len, oldlenp))
2768                                 return -EFAULT;
2769                 }
2770         }
2771
2772         if (newval && newlen) {
2773                 if (newlen > table->maxlen)
2774                         newlen = table->maxlen;
2775
2776                 if (copy_from_user(table->data, newval, newlen))
2777                         return -EFAULT;
2778         }
2779         return 1;
2780 }
2781
2782 /* The generic string strategy routine: */
2783 int sysctl_string(struct ctl_table *table,
2784                   void __user *oldval, size_t __user *oldlenp,
2785                   void __user *newval, size_t newlen)
2786 {
2787         if (!table->data || !table->maxlen)
2788                 return -ENOTDIR;
2789
2790         if (oldval && oldlenp) {
2791                 size_t bufsize;
2792                 if (get_user(bufsize, oldlenp))
2793                         return -EFAULT;
2794                 if (bufsize) {
2795                         size_t len = strlen(table->data), copied;
2796
2797                         /* This shouldn't trigger for a well-formed sysctl */
2798                         if (len > table->maxlen)
2799                                 len = table->maxlen;
2800
2801                         /* Copy up to a max of bufsize-1 bytes of the string */
2802                         copied = (len >= bufsize) ? bufsize - 1 : len;
2803
2804                         if (copy_to_user(oldval, table->data, copied) ||
2805                             put_user(0, (char __user *)(oldval + copied)))
2806                                 return -EFAULT;
2807                         if (put_user(len, oldlenp))
2808                                 return -EFAULT;
2809                 }
2810         }
2811         if (newval && newlen) {
2812                 size_t len = newlen;
2813                 if (len > table->maxlen)
2814                         len = table->maxlen;
2815                 if(copy_from_user(table->data, newval, len))
2816                         return -EFAULT;
2817                 if (len == table->maxlen)
2818                         len--;
2819                 ((char *) table->data)[len] = 0;
2820         }
2821         return 1;
2822 }
2823
2824 /*
2825  * This function makes sure that all of the integers in the vector
2826  * are between the minimum and maximum values given in the arrays
2827  * table->extra1 and table->extra2, respectively.
2828  */
2829 int sysctl_intvec(struct ctl_table *table,
2830                 void __user *oldval, size_t __user *oldlenp,
2831                 void __user *newval, size_t newlen)
2832 {
2833
2834         if (newval && newlen) {
2835                 int __user *vec = (int __user *) newval;
2836                 int *min = (int *) table->extra1;
2837                 int *max = (int *) table->extra2;
2838                 size_t length;
2839                 int i;
2840
2841                 if (newlen % sizeof(int) != 0)
2842                         return -EINVAL;
2843
2844                 if (!table->extra1 && !table->extra2)
2845                         return 0;
2846
2847                 if (newlen > table->maxlen)
2848                         newlen = table->maxlen;
2849                 length = newlen / sizeof(int);
2850
2851                 for (i = 0; i < length; i++) {
2852                         int value;
2853                         if (get_user(value, vec + i))
2854                                 return -EFAULT;
2855                         if (min && value < min[i])
2856                                 return -EINVAL;
2857                         if (max && value > max[i])
2858                                 return -EINVAL;
2859                 }
2860         }
2861         return 0;
2862 }
2863
2864 /* Strategy function to convert jiffies to seconds */
2865 int sysctl_jiffies(struct ctl_table *table,
2866                 void __user *oldval, size_t __user *oldlenp,
2867                 void __user *newval, size_t newlen)
2868 {
2869         if (oldval && oldlenp) {
2870                 size_t olen;
2871
2872                 if (get_user(olen, oldlenp))
2873                         return -EFAULT;
2874                 if (olen) {
2875                         int val;
2876
2877                         if (olen < sizeof(int))
2878                                 return -EINVAL;
2879
2880                         val = *(int *)(table->data) / HZ;
2881                         if (put_user(val, (int __user *)oldval))
2882                                 return -EFAULT;
2883                         if (put_user(sizeof(int), oldlenp))
2884                                 return -EFAULT;
2885                 }
2886         }
2887         if (newval && newlen) {
2888                 int new;
2889                 if (newlen != sizeof(int))
2890                         return -EINVAL;
2891                 if (get_user(new, (int __user *)newval))
2892                         return -EFAULT;
2893                 *(int *)(table->data) = new*HZ;
2894         }
2895         return 1;
2896 }
2897
2898 /* Strategy function to convert jiffies to seconds */
2899 int sysctl_ms_jiffies(struct ctl_table *table,
2900                 void __user *oldval, size_t __user *oldlenp,
2901                 void __user *newval, size_t newlen)
2902 {
2903         if (oldval && oldlenp) {
2904                 size_t olen;
2905
2906                 if (get_user(olen, oldlenp))
2907                         return -EFAULT;
2908                 if (olen) {
2909                         int val;
2910
2911                         if (olen < sizeof(int))
2912                                 return -EINVAL;
2913
2914                         val = jiffies_to_msecs(*(int *)(table->data));
2915                         if (put_user(val, (int __user *)oldval))
2916                                 return -EFAULT;
2917                         if (put_user(sizeof(int), oldlenp))
2918                                 return -EFAULT;
2919                 }
2920         }
2921         if (newval && newlen) {
2922                 int new;
2923                 if (newlen != sizeof(int))
2924                         return -EINVAL;
2925                 if (get_user(new, (int __user *)newval))
2926                         return -EFAULT;
2927                 *(int *)(table->data) = msecs_to_jiffies(new);
2928         }
2929         return 1;
2930 }
2931
2932
2933
2934 #else /* CONFIG_SYSCTL_SYSCALL */
2935
2936
2937 asmlinkage long sys_sysctl(struct __sysctl_args __user *args)
2938 {
2939         struct __sysctl_args tmp;
2940         int error;
2941
2942         if (copy_from_user(&tmp, args, sizeof(tmp)))
2943                 return -EFAULT;
2944
2945         error = deprecated_sysctl_warning(&tmp);
2946
2947         /* If no error reading the parameters then just -ENOSYS ... */
2948         if (!error)
2949                 error = -ENOSYS;
2950
2951         return error;
2952 }
2953
2954 int sysctl_data(struct ctl_table *table,
2955                   void __user *oldval, size_t __user *oldlenp,
2956                   void __user *newval, size_t newlen)
2957 {
2958         return -ENOSYS;
2959 }
2960
2961 int sysctl_string(struct ctl_table *table,
2962                   void __user *oldval, size_t __user *oldlenp,
2963                   void __user *newval, size_t newlen)
2964 {
2965         return -ENOSYS;
2966 }
2967
2968 int sysctl_intvec(struct ctl_table *table,
2969                 void __user *oldval, size_t __user *oldlenp,
2970                 void __user *newval, size_t newlen)
2971 {
2972         return -ENOSYS;
2973 }
2974
2975 int sysctl_jiffies(struct ctl_table *table,
2976                 void __user *oldval, size_t __user *oldlenp,
2977                 void __user *newval, size_t newlen)
2978 {
2979         return -ENOSYS;
2980 }
2981
2982 int sysctl_ms_jiffies(struct ctl_table *table,
2983                 void __user *oldval, size_t __user *oldlenp,
2984                 void __user *newval, size_t newlen)
2985 {
2986         return -ENOSYS;
2987 }
2988
2989 #endif /* CONFIG_SYSCTL_SYSCALL */
2990
2991 static int deprecated_sysctl_warning(struct __sysctl_args *args)
2992 {
2993         static int msg_count;
2994         int name[CTL_MAXNAME];
2995         int i;
2996
2997         /* Check args->nlen. */
2998         if (args->nlen < 0 || args->nlen > CTL_MAXNAME)
2999                 return -ENOTDIR;
3000
3001         /* Read in the sysctl name for better debug message logging */
3002         for (i = 0; i < args->nlen; i++)
3003                 if (get_user(name[i], args->name + i))
3004                         return -EFAULT;
3005
3006         /* Ignore accesses to kernel.version */
3007         if ((args->nlen == 2) && (name[0] == CTL_KERN) && (name[1] == KERN_VERSION))
3008                 return 0;
3009
3010         if (msg_count < 5) {
3011                 msg_count++;
3012                 printk(KERN_INFO
3013                         "warning: process `%s' used the deprecated sysctl "
3014                         "system call with ", current->comm);
3015                 for (i = 0; i < args->nlen; i++)
3016                         printk("%d.", name[i]);
3017                 printk("\n");
3018         }
3019         return 0;
3020 }
3021
3022 /*
3023  * No sense putting this after each symbol definition, twice,
3024  * exception granted :-)
3025  */
3026 EXPORT_SYMBOL(proc_dointvec);
3027 EXPORT_SYMBOL(proc_dointvec_jiffies);
3028 EXPORT_SYMBOL(proc_dointvec_minmax);
3029 EXPORT_SYMBOL(proc_dointvec_userhz_jiffies);
3030 EXPORT_SYMBOL(proc_dointvec_ms_jiffies);
3031 EXPORT_SYMBOL(proc_dostring);
3032 EXPORT_SYMBOL(proc_doulongvec_minmax);
3033 EXPORT_SYMBOL(proc_doulongvec_ms_jiffies_minmax);
3034 EXPORT_SYMBOL(register_sysctl_table);
3035 EXPORT_SYMBOL(register_sysctl_paths);
3036 EXPORT_SYMBOL(sysctl_intvec);
3037 EXPORT_SYMBOL(sysctl_jiffies);
3038 EXPORT_SYMBOL(sysctl_ms_jiffies);
3039 EXPORT_SYMBOL(sysctl_string);
3040 EXPORT_SYMBOL(sysctl_data);
3041 EXPORT_SYMBOL(unregister_sysctl_table);