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/capability.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/security.h>
  41 #include <linux/initrd.h>
  42 #include <linux/times.h>
  43 #include <linux/limits.h>
  44 #include <linux/dcache.h>
  45 #include <linux/syscalls.h>
  46 #include <linux/nfs_fs.h>
  47 #include <linux/acpi.h>
  48 #include <linux/reboot.h>
  49
  50 #include <asm/uaccess.h>
  51 #include <asm/processor.h>
  52
  53 #ifdef CONFIG_X86
  54 #include <asm/nmi.h>
  55 #include <asm/stacktrace.h>
  56 #endif
  57
  58 #if defined(CONFIG_SYSCTL)
  59
  60 /* External variables not in a header file. */
  61 extern int C_A_D;
  62 extern int print_fatal_signals;
  63 extern int sysctl_overcommit_memory;
  64 extern int sysctl_overcommit_ratio;
  65 extern int sysctl_panic_on_oom;
  66 extern int max_threads;
  67 extern int core_uses_pid;
  68 extern int suid_dumpable;
  69 extern char core_pattern[];
  70 extern int pid_max;
  71 extern int min_free_kbytes;
  72 extern int printk_ratelimit_jiffies;
  73 extern int printk_ratelimit_burst;
  74 extern int pid_max_min, pid_max_max;
  75 extern int sysctl_drop_caches;
  76 extern int percpu_pagelist_fraction;
  77 extern int compat_log;
  78 extern int maps_protect;
  79 extern int sysctl_stat_interval;
  80 extern int audit_argv_kb;
  81
  82 /* this is needed for the proc_dointvec_minmax for [fs_]overflow UID and GID */
  83 static int maxolduid = 65535;
  84 static int minolduid;
  85 static int min_percpu_pagelist_fract = 8;
  86
  87 static int ngroups_max = NGROUPS_MAX;
  88
  89 #ifdef CONFIG_KMOD
  90 extern char modprobe_path[];
  91 #endif
  92 #ifdef CONFIG_CHR_DEV_SG
  93 extern int sg_big_buff;
  94 #endif
  95
  96 #ifdef __sparc__
  97 extern char reboot_command [];
  98 extern int stop_a_enabled;
  99 extern int scons_pwroff;
 100 #endif
 101
 102 #ifdef __hppa__
 103 extern int pwrsw_enabled;
 104 extern int unaligned_enabled;
 105 #endif
 106
 107 #ifdef CONFIG_S390
 108 #ifdef CONFIG_MATHEMU
 109 extern int sysctl_ieee_emulation_warnings;
 110 #endif
 111 extern int sysctl_userprocess_debug;
 112 extern int spin_retry;
 113 #endif
 114
 115 extern int sysctl_hz_timer;
 116
 117 #ifdef CONFIG_BSD_PROCESS_ACCT
 118 extern int acct_parm[];
 119 #endif
 120
 121 #ifdef CONFIG_IA64
 122 extern int no_unaligned_warning;
 123 #endif
 124
 125 #ifdef CONFIG_RT_MUTEXES
 126 extern int max_lock_depth;
 127 #endif
 128
 129 #ifdef CONFIG_SYSCTL_SYSCALL
 130 static int parse_table(int __user *, int, void __user *, size_t __user *,
 131                 void __user *, size_t, ctl_table *);
 132 #endif
 133
 134
 135 #ifdef CONFIG_PROC_SYSCTL
 136 static int proc_do_cad_pid(ctl_table *table, int write, struct file *filp,
 137                   void __user *buffer, size_t *lenp, loff_t *ppos);
 138 static int proc_dointvec_taint(ctl_table *table, int write, struct file *filp,
 139                                void __user *buffer, size_t *lenp, loff_t *ppos);
 140 #endif
 141
 142 static ctl_table root_table[];
 143 static struct ctl_table_header root_table_header =
 144         { root_table, LIST_HEAD_INIT(root_table_header.ctl_entry) };
 145
 146 static ctl_table kern_table[];
 147 static ctl_table vm_table[];
 148 static ctl_table fs_table[];
 149 static ctl_table debug_table[];
 150 static ctl_table dev_table[];
 151 extern ctl_table random_table[];
 152 #ifdef CONFIG_UNIX98_PTYS
 153 extern ctl_table pty_table[];
 154 #endif
 155 #ifdef CONFIG_INOTIFY_USER
 156 extern ctl_table inotify_table[];
 157 #endif
 158
 159 #ifdef HAVE_ARCH_PICK_MMAP_LAYOUT
 160 int sysctl_legacy_va_layout;
 161 #endif
 162
 163 extern int prove_locking;
 164 extern int lock_stat;
 165
 166 /* The default sysctl tables: */
 167
 168 static ctl_table root_table[] = {
 169         {
 170                 .ctl_name       = CTL_KERN,
 171                 .procname       = "kernel",
 172                 .mode           = 0555,
 173                 .child          = kern_table,
 174         },
 175         {
 176                 .ctl_name       = CTL_VM,
 177                 .procname       = "vm",
 178                 .mode           = 0555,
 179                 .child          = vm_table,
 180         },
 181 #ifdef CONFIG_NET
 182         {
 183                 .ctl_name       = CTL_NET,
 184                 .procname       = "net",
 185                 .mode           = 0555,
 186                 .child          = net_table,
 187         },
 188 #endif
 189         {
 190                 .ctl_name       = CTL_FS,
 191                 .procname       = "fs",
 192                 .mode           = 0555,
 193                 .child          = fs_table,
 194         },
 195         {
 196                 .ctl_name       = CTL_DEBUG,
 197                 .procname       = "debug",
 198                 .mode           = 0555,
 199                 .child          = debug_table,
 200         },
 201         {
 202                 .ctl_name       = CTL_DEV,
 203                 .procname       = "dev",
 204                 .mode           = 0555,
 205                 .child          = dev_table,
 206         },
 207 /*
 208  * NOTE: do not add new entries to this table unless you have read
 209  * Documentation/sysctl/ctl_unnumbered.txt
 210  */
 211         { .ctl_name = 0 }
 212 };
 213
 214 #ifdef CONFIG_SCHED_DEBUG
 215 static unsigned long min_sched_granularity_ns = 100000;         /* 100 usecs */
 216 static unsigned long max_sched_granularity_ns = 1000000000;     /* 1 second */
 217 static unsigned long min_wakeup_granularity_ns;                 /* 0 usecs */
 218 static unsigned long max_wakeup_granularity_ns = 1000000000;    /* 1 second */
 219 #endif
 220
 221 static ctl_table kern_table[] = {
 222 #ifdef CONFIG_SCHED_DEBUG
 223         {
 224                 .ctl_name       = CTL_UNNUMBERED,
 225                 .procname       = "sched_nr_latency",
 226                 .data           = &sysctl_sched_nr_latency,
 227                 .maxlen         = sizeof(unsigned int),
 228                 .mode           = 0644,
 229                 .proc_handler   = &proc_dointvec,
 230         },
 231         {
 232                 .ctl_name       = CTL_UNNUMBERED,
 233                 .procname       = "sched_latency_ns",
 234                 .data           = &sysctl_sched_latency,
 235                 .maxlen         = sizeof(unsigned int),
 236                 .mode           = 0644,
 237                 .proc_handler   = &proc_dointvec_minmax,
 238                 .strategy       = &sysctl_intvec,
 239                 .extra1         = &min_sched_granularity_ns,
 240                 .extra2         = &max_sched_granularity_ns,
 241         },
 242         {
 243                 .ctl_name       = CTL_UNNUMBERED,
 244                 .procname       = "sched_wakeup_granularity_ns",
 245                 .data           = &sysctl_sched_wakeup_granularity,
 246                 .maxlen         = sizeof(unsigned int),
 247                 .mode           = 0644,
 248                 .proc_handler   = &proc_dointvec_minmax,
 249                 .strategy       = &sysctl_intvec,
 250                 .extra1         = &min_wakeup_granularity_ns,
 251                 .extra2         = &max_wakeup_granularity_ns,
 252         },
 253         {
 254                 .ctl_name       = CTL_UNNUMBERED,
 255                 .procname       = "sched_batch_wakeup_granularity_ns",
 256                 .data           = &sysctl_sched_batch_wakeup_granularity,
 257                 .maxlen         = sizeof(unsigned int),
 258                 .mode           = 0644,
 259                 .proc_handler   = &proc_dointvec_minmax,
 260                 .strategy       = &sysctl_intvec,
 261                 .extra1         = &min_wakeup_granularity_ns,
 262                 .extra2         = &max_wakeup_granularity_ns,
 263         },
 264         {
 265                 .ctl_name       = CTL_UNNUMBERED,
 266                 .procname       = "sched_child_runs_first",
 267                 .data           = &sysctl_sched_child_runs_first,
 268                 .maxlen         = sizeof(unsigned int),
 269                 .mode           = 0644,
 270                 .proc_handler   = &proc_dointvec,
 271         },
 272         {
 273                 .ctl_name       = CTL_UNNUMBERED,
 274                 .procname       = "sched_features",
 275                 .data           = &sysctl_sched_features,
 276                 .maxlen         = sizeof(unsigned int),
 277                 .mode           = 0644,
 278                 .proc_handler   = &proc_dointvec,
 279         },
 280 #endif
 281         {
 282                 .ctl_name       = CTL_UNNUMBERED,
 283                 .procname       = "sched_compat_yield",
 284                 .data           = &sysctl_sched_compat_yield,
 285                 .maxlen         = sizeof(unsigned int),
 286                 .mode           = 0644,
 287                 .proc_handler   = &proc_dointvec,
 288         },
 289 #ifdef CONFIG_PROVE_LOCKING
 290         {
 291                 .ctl_name       = CTL_UNNUMBERED,
 292                 .procname       = "prove_locking",
 293                 .data           = &prove_locking,
 294                 .maxlen         = sizeof(int),
 295                 .mode           = 0644,
 296                 .proc_handler   = &proc_dointvec,
 297         },
 298 #endif
 299 #ifdef CONFIG_LOCK_STAT
 300         {
 301                 .ctl_name       = CTL_UNNUMBERED,
 302                 .procname       = "lock_stat",
 303                 .data           = &lock_stat,
 304                 .maxlen         = sizeof(int),
 305                 .mode           = 0644,
 306                 .proc_handler   = &proc_dointvec,
 307         },
 308 #endif
 309         {
 310                 .ctl_name       = KERN_PANIC,
 311                 .procname       = "panic",
 312                 .data           = &panic_timeout,
 313                 .maxlen         = sizeof(int),
 314                 .mode           = 0644,
 315                 .proc_handler   = &proc_dointvec,
 316         },
 317         {
 318                 .ctl_name       = KERN_CORE_USES_PID,
 319                 .procname       = "core_uses_pid",
 320                 .data           = &core_uses_pid,
 321                 .maxlen         = sizeof(int),
 322                 .mode           = 0644,
 323                 .proc_handler   = &proc_dointvec,
 324         },
 325 #ifdef CONFIG_AUDITSYSCALL
 326         {
 327                 .ctl_name       = CTL_UNNUMBERED,
 328                 .procname       = "audit_argv_kb",
 329                 .data           = &audit_argv_kb,
 330                 .maxlen         = sizeof(int),
 331                 .mode           = 0644,
 332                 .proc_handler   = &proc_dointvec,
 333         },
 334 #endif
 335         {
 336                 .ctl_name       = KERN_CORE_PATTERN,
 337                 .procname       = "core_pattern",
 338                 .data           = core_pattern,
 339                 .maxlen         = CORENAME_MAX_SIZE,
 340                 .mode           = 0644,
 341                 .proc_handler   = &proc_dostring,
 342                 .strategy       = &sysctl_string,
 343         },
 344 #ifdef CONFIG_PROC_SYSCTL
 345         {
 346                 .ctl_name       = KERN_TAINTED,
 347                 .procname       = "tainted",
 348                 .data           = &tainted,
 349                 .maxlen         = sizeof(int),
 350                 .mode           = 0644,
 351                 .proc_handler   = &proc_dointvec_taint,
 352         },
 353 #endif
 354         {
 355                 .ctl_name       = KERN_CAP_BSET,
 356                 .procname       = "cap-bound",
 357                 .data           = &cap_bset,
 358                 .maxlen         = sizeof(kernel_cap_t),
 359                 .mode           = 0600,
 360                 .proc_handler   = &proc_dointvec_bset,
 361         },
 362 #ifdef CONFIG_BLK_DEV_INITRD
 363         {
 364                 .ctl_name       = KERN_REALROOTDEV,
 365                 .procname       = "real-root-dev",
 366                 .data           = &real_root_dev,
 367                 .maxlen         = sizeof(int),
 368                 .mode           = 0644,
 369                 .proc_handler   = &proc_dointvec,
 370         },
 371 #endif
 372         {
 373                 .ctl_name       = CTL_UNNUMBERED,
 374                 .procname       = "print-fatal-signals",
 375                 .data           = &print_fatal_signals,
 376                 .maxlen         = sizeof(int),
 377                 .mode           = 0644,
 378                 .proc_handler   = &proc_dointvec,
 379         },
 380 #ifdef __sparc__
 381         {
 382                 .ctl_name       = KERN_SPARC_REBOOT,
 383                 .procname       = "reboot-cmd",
 384                 .data           = reboot_command,
 385                 .maxlen         = 256,
 386                 .mode           = 0644,
 387                 .proc_handler   = &proc_dostring,
 388                 .strategy       = &sysctl_string,
 389         },
 390         {
 391                 .ctl_name       = KERN_SPARC_STOP_A,
 392                 .procname       = "stop-a",
 393                 .data           = &stop_a_enabled,
 394                 .maxlen         = sizeof (int),
 395                 .mode           = 0644,
 396                 .proc_handler   = &proc_dointvec,
 397         },
 398         {
 399                 .ctl_name       = KERN_SPARC_SCONS_PWROFF,
 400                 .procname       = "scons-poweroff",
 401                 .data           = &scons_pwroff,
 402                 .maxlen         = sizeof (int),
 403                 .mode           = 0644,
 404                 .proc_handler   = &proc_dointvec,
 405         },
 406 #endif
 407 #ifdef __hppa__
 408         {
 409                 .ctl_name       = KERN_HPPA_PWRSW,
 410                 .procname       = "soft-power",
 411                 .data           = &pwrsw_enabled,
 412                 .maxlen         = sizeof (int),
 413                 .mode           = 0644,
 414                 .proc_handler   = &proc_dointvec,
 415         },
 416         {
 417                 .ctl_name       = KERN_HPPA_UNALIGNED,
 418                 .procname       = "unaligned-trap",
 419                 .data           = &unaligned_enabled,
 420                 .maxlen         = sizeof (int),
 421                 .mode           = 0644,
 422                 .proc_handler   = &proc_dointvec,
 423         },
 424 #endif
 425         {
 426                 .ctl_name       = KERN_CTLALTDEL,
 427                 .procname       = "ctrl-alt-del",
 428                 .data           = &C_A_D,
 429                 .maxlen         = sizeof(int),
 430                 .mode           = 0644,
 431                 .proc_handler   = &proc_dointvec,
 432         },
 433         {
 434                 .ctl_name       = KERN_PRINTK,
 435                 .procname       = "printk",
 436                 .data           = &console_loglevel,
 437                 .maxlen         = 4*sizeof(int),
 438                 .mode           = 0644,
 439                 .proc_handler   = &proc_dointvec,
 440         },
 441 #ifdef CONFIG_KMOD
 442         {
 443                 .ctl_name       = KERN_MODPROBE,
 444                 .procname       = "modprobe",
 445                 .data           = &modprobe_path,
 446                 .maxlen         = KMOD_PATH_LEN,
 447                 .mode           = 0644,
 448                 .proc_handler   = &proc_dostring,
 449                 .strategy       = &sysctl_string,
 450         },
 451 #endif
 452 #if defined(CONFIG_HOTPLUG) && defined(CONFIG_NET)
 453         {
 454                 .ctl_name       = KERN_HOTPLUG,
 455                 .procname       = "hotplug",
 456                 .data           = &uevent_helper,
 457                 .maxlen         = UEVENT_HELPER_PATH_LEN,
 458                 .mode           = 0644,
 459                 .proc_handler   = &proc_dostring,
 460                 .strategy       = &sysctl_string,
 461         },
 462 #endif
 463 #ifdef CONFIG_CHR_DEV_SG
 464         {
 465                 .ctl_name       = KERN_SG_BIG_BUFF,
 466                 .procname       = "sg-big-buff",
 467                 .data           = &sg_big_buff,
 468                 .maxlen         = sizeof (int),
 469                 .mode           = 0444,
 470                 .proc_handler   = &proc_dointvec,
 471         },
 472 #endif
 473 #ifdef CONFIG_BSD_PROCESS_ACCT
 474         {
 475                 .ctl_name       = KERN_ACCT,
 476                 .procname       = "acct",
 477                 .data           = &acct_parm,
 478                 .maxlen         = 3*sizeof(int),
 479                 .mode           = 0644,
 480                 .proc_handler   = &proc_dointvec,
 481         },
 482 #endif
 483 #ifdef CONFIG_MAGIC_SYSRQ
 484         {
 485                 .ctl_name       = KERN_SYSRQ,
 486                 .procname       = "sysrq",
 487                 .data           = &__sysrq_enabled,
 488                 .maxlen         = sizeof (int),
 489                 .mode           = 0644,
 490                 .proc_handler   = &proc_dointvec,
 491         },
 492 #endif
 493 #ifdef CONFIG_PROC_SYSCTL
 494         {
 495                 .ctl_name       = KERN_CADPID,
 496                 .procname       = "cad_pid",
 497                 .data           = NULL,
 498                 .maxlen         = sizeof (int),
 499                 .mode           = 0600,
 500                 .proc_handler   = &proc_do_cad_pid,
 501         },
 502 #endif
 503         {
 504                 .ctl_name       = KERN_MAX_THREADS,
 505                 .procname       = "threads-max",
 506                 .data           = &max_threads,
 507                 .maxlen         = sizeof(int),
 508                 .mode           = 0644,
 509                 .proc_handler   = &proc_dointvec,
 510         },
 511         {
 512                 .ctl_name       = KERN_RANDOM,
 513                 .procname       = "random",
 514                 .mode           = 0555,
 515                 .child          = random_table,
 516         },
 517 #ifdef CONFIG_UNIX98_PTYS
 518         {
 519                 .ctl_name       = KERN_PTY,
 520                 .procname       = "pty",
 521                 .mode           = 0555,
 522                 .child          = pty_table,
 523         },
 524 #endif
 525         {
 526                 .ctl_name       = KERN_OVERFLOWUID,
 527                 .procname       = "overflowuid",
 528                 .data           = &overflowuid,
 529                 .maxlen         = sizeof(int),
 530                 .mode           = 0644,
 531                 .proc_handler   = &proc_dointvec_minmax,
 532                 .strategy       = &sysctl_intvec,
 533                 .extra1         = &minolduid,
 534                 .extra2         = &maxolduid,
 535         },
 536         {
 537                 .ctl_name       = KERN_OVERFLOWGID,
 538                 .procname       = "overflowgid",
 539                 .data           = &overflowgid,
 540                 .maxlen         = sizeof(int),
 541                 .mode           = 0644,
 542                 .proc_handler   = &proc_dointvec_minmax,
 543                 .strategy       = &sysctl_intvec,
 544                 .extra1         = &minolduid,
 545                 .extra2         = &maxolduid,
 546         },
 547 #ifdef CONFIG_S390
 548 #ifdef CONFIG_MATHEMU
 549         {
 550                 .ctl_name       = KERN_IEEE_EMULATION_WARNINGS,
 551                 .procname       = "ieee_emulation_warnings",
 552                 .data           = &sysctl_ieee_emulation_warnings,
 553                 .maxlen         = sizeof(int),
 554                 .mode           = 0644,
 555                 .proc_handler   = &proc_dointvec,
 556         },
 557 #endif
 558 #ifdef CONFIG_NO_IDLE_HZ
 559         {
 560                 .ctl_name       = KERN_HZ_TIMER,
 561                 .procname       = "hz_timer",
 562                 .data           = &sysctl_hz_timer,
 563                 .maxlen         = sizeof(int),
 564                 .mode           = 0644,
 565                 .proc_handler   = &proc_dointvec,
 566         },
 567 #endif
 568         {
 569                 .ctl_name       = KERN_S390_USER_DEBUG_LOGGING,
 570                 .procname       = "userprocess_debug",
 571                 .data           = &sysctl_userprocess_debug,
 572                 .maxlen         = sizeof(int),
 573                 .mode           = 0644,
 574                 .proc_handler   = &proc_dointvec,
 575         },
 576 #endif
 577         {
 578                 .ctl_name       = KERN_PIDMAX,
 579                 .procname       = "pid_max",
 580                 .data           = &pid_max,
 581                 .maxlen         = sizeof (int),
 582                 .mode           = 0644,
 583                 .proc_handler   = &proc_dointvec_minmax,
 584                 .strategy       = sysctl_intvec,
 585                 .extra1         = &pid_max_min,
 586                 .extra2         = &pid_max_max,
 587         },
 588         {
 589                 .ctl_name       = KERN_PANIC_ON_OOPS,
 590                 .procname       = "panic_on_oops",
 591                 .data           = &panic_on_oops,
 592                 .maxlen         = sizeof(int),
 593                 .mode           = 0644,
 594                 .proc_handler   = &proc_dointvec,
 595         },
 596         {
 597                 .ctl_name       = KERN_PRINTK_RATELIMIT,
 598                 .procname       = "printk_ratelimit",
 599                 .data           = &printk_ratelimit_jiffies,
 600                 .maxlen         = sizeof(int),
 601                 .mode           = 0644,
 602                 .proc_handler   = &proc_dointvec_jiffies,
 603                 .strategy       = &sysctl_jiffies,
 604         },
 605         {
 606                 .ctl_name       = KERN_PRINTK_RATELIMIT_BURST,
 607                 .procname       = "printk_ratelimit_burst",
 608                 .data           = &printk_ratelimit_burst,
 609                 .maxlen         = sizeof(int),
 610                 .mode           = 0644,
 611                 .proc_handler   = &proc_dointvec,
 612         },
 613         {
 614                 .ctl_name       = KERN_NGROUPS_MAX,
 615                 .procname       = "ngroups_max",
 616                 .data           = &ngroups_max,
 617                 .maxlen         = sizeof (int),
 618                 .mode           = 0444,
 619                 .proc_handler   = &proc_dointvec,
 620         },
 621 #if defined(CONFIG_X86_LOCAL_APIC) && defined(CONFIG_X86)
 622         {
 623                 .ctl_name       = KERN_UNKNOWN_NMI_PANIC,
 624                 .procname       = "unknown_nmi_panic",
 625                 .data           = &unknown_nmi_panic,
 626                 .maxlen         = sizeof (int),
 627                 .mode           = 0644,
 628                 .proc_handler   = &proc_dointvec,
 629         },
 630         {
 631                 .ctl_name       = KERN_NMI_WATCHDOG,
 632                 .procname       = "nmi_watchdog",
 633                 .data           = &nmi_watchdog_enabled,
 634                 .maxlen         = sizeof (int),
 635                 .mode           = 0644,
 636                 .proc_handler   = &proc_nmi_enabled,
 637         },
 638 #endif
 639 #if defined(CONFIG_X86)
 640         {
 641                 .ctl_name       = KERN_PANIC_ON_NMI,
 642                 .procname       = "panic_on_unrecovered_nmi",
 643                 .data           = &panic_on_unrecovered_nmi,
 644                 .maxlen         = sizeof(int),
 645                 .mode           = 0644,
 646                 .proc_handler   = &proc_dointvec,
 647         },
 648         {
 649                 .ctl_name       = KERN_BOOTLOADER_TYPE,
 650                 .procname       = "bootloader_type",
 651                 .data           = &bootloader_type,
 652                 .maxlen         = sizeof (int),
 653                 .mode           = 0444,
 654                 .proc_handler   = &proc_dointvec,
 655         },
 656         {
 657                 .ctl_name       = CTL_UNNUMBERED,
 658                 .procname       = "kstack_depth_to_print",
 659                 .data           = &kstack_depth_to_print,
 660                 .maxlen         = sizeof(int),
 661                 .mode           = 0644,
 662                 .proc_handler   = &proc_dointvec,
 663         },
 664 #endif
 665 #if defined(CONFIG_MMU)
 666         {
 667                 .ctl_name       = KERN_RANDOMIZE,
 668                 .procname       = "randomize_va_space",
 669                 .data           = &randomize_va_space,
 670                 .maxlen         = sizeof(int),
 671                 .mode           = 0644,
 672                 .proc_handler   = &proc_dointvec,
 673         },
 674 #endif
 675 #if defined(CONFIG_S390) && defined(CONFIG_SMP)
 676         {
 677                 .ctl_name       = KERN_SPIN_RETRY,
 678                 .procname       = "spin_retry",
 679                 .data           = &spin_retry,
 680                 .maxlen         = sizeof (int),
 681                 .mode           = 0644,
 682                 .proc_handler   = &proc_dointvec,
 683         },
 684 #endif
 685 #if     defined(CONFIG_ACPI_SLEEP) && defined(CONFIG_X86)
 686         {
 687                 .ctl_name       = KERN_ACPI_VIDEO_FLAGS,
 688                 .procname       = "acpi_video_flags",
 689                 .data           = &acpi_realmode_flags,
 690                 .maxlen         = sizeof (unsigned long),
 691                 .mode           = 0644,
 692                 .proc_handler   = &proc_doulongvec_minmax,
 693         },
 694 #endif
 695 #ifdef CONFIG_IA64
 696         {
 697                 .ctl_name       = KERN_IA64_UNALIGNED,
 698                 .procname       = "ignore-unaligned-usertrap",
 699                 .data           = &no_unaligned_warning,
 700                 .maxlen         = sizeof (int),
 701                 .mode           = 0644,
 702                 .proc_handler   = &proc_dointvec,
 703         },
 704 #endif
 705 #ifdef CONFIG_COMPAT
 706         {
 707                 .ctl_name       = KERN_COMPAT_LOG,
 708                 .procname       = "compat-log",
 709                 .data           = &compat_log,
 710                 .maxlen         = sizeof (int),
 711                 .mode           = 0644,
 712                 .proc_handler   = &proc_dointvec,
 713         },
 714 #endif
 715 #ifdef CONFIG_RT_MUTEXES
 716         {
 717                 .ctl_name       = KERN_MAX_LOCK_DEPTH,
 718                 .procname       = "max_lock_depth",
 719                 .data           = &max_lock_depth,
 720                 .maxlen         = sizeof(int),
 721                 .mode           = 0644,
 722                 .proc_handler   = &proc_dointvec,
 723         },
 724 #endif
 725 #ifdef CONFIG_PROC_FS
 726         {
 727                 .ctl_name       = CTL_UNNUMBERED,
 728                 .procname       = "maps_protect",
 729                 .data           = &maps_protect,
 730                 .maxlen         = sizeof(int),
 731                 .mode           = 0644,
 732                 .proc_handler   = &proc_dointvec,
 733         },
 734 #endif
 735         {
 736                 .ctl_name       = CTL_UNNUMBERED,
 737                 .procname       = "poweroff_cmd",
 738                 .data           = &poweroff_cmd,
 739                 .maxlen         = POWEROFF_CMD_PATH_LEN,
 740                 .mode           = 0644,
 741                 .proc_handler   = &proc_dostring,
 742                 .strategy       = &sysctl_string,
 743         },
 744 /*
 745  * NOTE: do not add new entries to this table unless you have read
 746  * Documentation/sysctl/ctl_unnumbered.txt
 747  */
 748         { .ctl_name = 0 }
 749 };
 750
 751 /* Constants for minimum and maximum testing in vm_table.
 752    We use these as one-element integer vectors. */
 753 static int zero;
 754 static int two = 2;
 755 static int one_hundred = 100;
 756
 757
 758 static ctl_table vm_table[] = {
 759         {
 760                 .ctl_name       = VM_OVERCOMMIT_MEMORY,
 761                 .procname       = "overcommit_memory",
 762                 .data           = &sysctl_overcommit_memory,
 763                 .maxlen         = sizeof(sysctl_overcommit_memory),
 764                 .mode           = 0644,
 765                 .proc_handler   = &proc_dointvec,
 766         },
 767         {
 768                 .ctl_name       = VM_PANIC_ON_OOM,
 769                 .procname       = "panic_on_oom",
 770                 .data           = &sysctl_panic_on_oom,
 771                 .maxlen         = sizeof(sysctl_panic_on_oom),
 772                 .mode           = 0644,
 773                 .proc_handler   = &proc_dointvec,
 774         },
 775         {
 776                 .ctl_name       = VM_OVERCOMMIT_RATIO,
 777                 .procname       = "overcommit_ratio",
 778                 .data           = &sysctl_overcommit_ratio,
 779                 .maxlen         = sizeof(sysctl_overcommit_ratio),
 780                 .mode           = 0644,
 781                 .proc_handler   = &proc_dointvec,
 782         },
 783         {
 784                 .ctl_name       = VM_PAGE_CLUSTER,
 785                 .procname       = "page-cluster",
 786                 .data           = &page_cluster,
 787                 .maxlen         = sizeof(int),
 788                 .mode           = 0644,
 789                 .proc_handler   = &proc_dointvec,
 790         },
 791         {
 792                 .ctl_name       = VM_DIRTY_BACKGROUND,
 793                 .procname       = "dirty_background_ratio",
 794                 .data           = &dirty_background_ratio,
 795                 .maxlen         = sizeof(dirty_background_ratio),
 796                 .mode           = 0644,
 797                 .proc_handler   = &proc_dointvec_minmax,
 798                 .strategy       = &sysctl_intvec,
 799                 .extra1         = &zero,
 800                 .extra2         = &one_hundred,
 801         },
 802         {
 803                 .ctl_name       = VM_DIRTY_RATIO,
 804                 .procname       = "dirty_ratio",
 805                 .data           = &vm_dirty_ratio,
 806                 .maxlen         = sizeof(vm_dirty_ratio),
 807                 .mode           = 0644,
 808                 .proc_handler   = &proc_dointvec_minmax,
 809                 .strategy       = &sysctl_intvec,
 810                 .extra1         = &zero,
 811                 .extra2         = &one_hundred,
 812         },
 813         {
 814                 .ctl_name       = VM_DIRTY_WB_CS,
 815                 .procname       = "dirty_writeback_centisecs",
 816                 .data           = &dirty_writeback_interval,
 817                 .maxlen         = sizeof(dirty_writeback_interval),
 818                 .mode           = 0644,
 819                 .proc_handler   = &dirty_writeback_centisecs_handler,
 820         },
 821         {
 822                 .ctl_name       = VM_DIRTY_EXPIRE_CS,
 823                 .procname       = "dirty_expire_centisecs",
 824                 .data           = &dirty_expire_interval,
 825                 .maxlen         = sizeof(dirty_expire_interval),
 826                 .mode           = 0644,
 827                 .proc_handler   = &proc_dointvec_userhz_jiffies,
 828         },
 829         {
 830                 .ctl_name       = VM_NR_PDFLUSH_THREADS,
 831                 .procname       = "nr_pdflush_threads",
 832                 .data           = &nr_pdflush_threads,
 833                 .maxlen         = sizeof nr_pdflush_threads,
 834                 .mode           = 0444 /* read-only*/,
 835                 .proc_handler   = &proc_dointvec,
 836         },
 837         {
 838                 .ctl_name       = VM_SWAPPINESS,
 839                 .procname       = "swappiness",
 840                 .data           = &vm_swappiness,
 841                 .maxlen         = sizeof(vm_swappiness),
 842                 .mode           = 0644,
 843                 .proc_handler   = &proc_dointvec_minmax,
 844                 .strategy       = &sysctl_intvec,
 845                 .extra1         = &zero,
 846                 .extra2         = &one_hundred,
 847         },
 848 #ifdef CONFIG_HUGETLB_PAGE
 849          {
 850                 .ctl_name       = VM_HUGETLB_PAGES,
 851                 .procname       = "nr_hugepages",
 852                 .data           = &max_huge_pages,
 853                 .maxlen         = sizeof(unsigned long),
 854                 .mode           = 0644,
 855                 .proc_handler   = &hugetlb_sysctl_handler,
 856                 .extra1         = (void *)&hugetlb_zero,
 857                 .extra2         = (void *)&hugetlb_infinity,
 858          },
 859          {
 860                 .ctl_name       = VM_HUGETLB_GROUP,
 861                 .procname       = "hugetlb_shm_group",
 862                 .data           = &sysctl_hugetlb_shm_group,
 863                 .maxlen         = sizeof(gid_t),
 864                 .mode           = 0644,
 865                 .proc_handler   = &proc_dointvec,
 866          },
 867          {
 868                 .ctl_name       = CTL_UNNUMBERED,
 869                 .procname       = "hugepages_treat_as_movable",
 870                 .data           = &hugepages_treat_as_movable,
 871                 .maxlen         = sizeof(int),
 872                 .mode           = 0644,
 873                 .proc_handler   = &hugetlb_treat_movable_handler,
 874         },
 875 #endif
 876         {
 877                 .ctl_name       = VM_LOWMEM_RESERVE_RATIO,
 878                 .procname       = "lowmem_reserve_ratio",
 879                 .data           = &sysctl_lowmem_reserve_ratio,
 880                 .maxlen         = sizeof(sysctl_lowmem_reserve_ratio),
 881                 .mode           = 0644,
 882                 .proc_handler   = &lowmem_reserve_ratio_sysctl_handler,
 883                 .strategy       = &sysctl_intvec,
 884         },
 885         {
 886                 .ctl_name       = VM_DROP_PAGECACHE,
 887                 .procname       = "drop_caches",
 888                 .data           = &sysctl_drop_caches,
 889                 .maxlen         = sizeof(int),
 890                 .mode           = 0644,
 891                 .proc_handler   = drop_caches_sysctl_handler,
 892                 .strategy       = &sysctl_intvec,
 893         },
 894         {
 895                 .ctl_name       = VM_MIN_FREE_KBYTES,
 896                 .procname       = "min_free_kbytes",
 897                 .data           = &min_free_kbytes,
 898                 .maxlen         = sizeof(min_free_kbytes),
 899                 .mode           = 0644,
 900                 .proc_handler   = &min_free_kbytes_sysctl_handler,
 901                 .strategy       = &sysctl_intvec,
 902                 .extra1         = &zero,
 903         },
 904         {
 905                 .ctl_name       = VM_PERCPU_PAGELIST_FRACTION,
 906                 .procname       = "percpu_pagelist_fraction",
 907                 .data           = &percpu_pagelist_fraction,
 908                 .maxlen         = sizeof(percpu_pagelist_fraction),
 909                 .mode           = 0644,
 910                 .proc_handler   = &percpu_pagelist_fraction_sysctl_handler,
 911                 .strategy       = &sysctl_intvec,
 912                 .extra1         = &min_percpu_pagelist_fract,
 913         },
 914 #ifdef CONFIG_MMU
 915         {
 916                 .ctl_name       = VM_MAX_MAP_COUNT,
 917                 .procname       = "max_map_count",
 918                 .data           = &sysctl_max_map_count,
 919                 .maxlen         = sizeof(sysctl_max_map_count),
 920                 .mode           = 0644,
 921                 .proc_handler   = &proc_dointvec
 922         },
 923 #endif
 924         {
 925                 .ctl_name       = VM_LAPTOP_MODE,
 926                 .procname       = "laptop_mode",
 927                 .data           = &laptop_mode,
 928                 .maxlen         = sizeof(laptop_mode),
 929                 .mode           = 0644,
 930                 .proc_handler   = &proc_dointvec_jiffies,
 931                 .strategy       = &sysctl_jiffies,
 932         },
 933         {
 934                 .ctl_name       = VM_BLOCK_DUMP,
 935                 .procname       = "block_dump",
 936                 .data           = &block_dump,
 937                 .maxlen         = sizeof(block_dump),
 938                 .mode           = 0644,
 939                 .proc_handler   = &proc_dointvec,
 940                 .strategy       = &sysctl_intvec,
 941                 .extra1         = &zero,
 942         },
 943         {
 944                 .ctl_name       = VM_VFS_CACHE_PRESSURE,
 945                 .procname       = "vfs_cache_pressure",
 946                 .data           = &sysctl_vfs_cache_pressure,
 947                 .maxlen         = sizeof(sysctl_vfs_cache_pressure),
 948                 .mode           = 0644,
 949                 .proc_handler   = &proc_dointvec,
 950                 .strategy       = &sysctl_intvec,
 951                 .extra1         = &zero,
 952         },
 953 #ifdef HAVE_ARCH_PICK_MMAP_LAYOUT
 954         {
 955                 .ctl_name       = VM_LEGACY_VA_LAYOUT,
 956                 .procname       = "legacy_va_layout",
 957                 .data           = &sysctl_legacy_va_layout,
 958                 .maxlen         = sizeof(sysctl_legacy_va_layout),
 959                 .mode           = 0644,
 960                 .proc_handler   = &proc_dointvec,
 961                 .strategy       = &sysctl_intvec,
 962                 .extra1         = &zero,
 963         },
 964 #endif
 965 #ifdef CONFIG_NUMA
 966         {
 967                 .ctl_name       = VM_ZONE_RECLAIM_MODE,
 968                 .procname       = "zone_reclaim_mode",
 969                 .data           = &zone_reclaim_mode,
 970                 .maxlen         = sizeof(zone_reclaim_mode),
 971                 .mode           = 0644,
 972                 .proc_handler   = &proc_dointvec,
 973                 .strategy       = &sysctl_intvec,
 974                 .extra1         = &zero,
 975         },
 976         {
 977                 .ctl_name       = VM_MIN_UNMAPPED,
 978                 .procname       = "min_unmapped_ratio",
 979                 .data           = &sysctl_min_unmapped_ratio,
 980                 .maxlen         = sizeof(sysctl_min_unmapped_ratio),
 981                 .mode           = 0644,
 982                 .proc_handler   = &sysctl_min_unmapped_ratio_sysctl_handler,
 983                 .strategy       = &sysctl_intvec,
 984                 .extra1         = &zero,
 985                 .extra2         = &one_hundred,
 986         },
 987         {
 988                 .ctl_name       = VM_MIN_SLAB,
 989                 .procname       = "min_slab_ratio",
 990                 .data           = &sysctl_min_slab_ratio,
 991                 .maxlen         = sizeof(sysctl_min_slab_ratio),
 992                 .mode           = 0644,
 993                 .proc_handler   = &sysctl_min_slab_ratio_sysctl_handler,
 994                 .strategy       = &sysctl_intvec,
 995                 .extra1         = &zero,
 996                 .extra2         = &one_hundred,
 997         },
 998 #endif
 999 #ifdef CONFIG_SMP
1000         {
1001                 .ctl_name       = CTL_UNNUMBERED,
1002                 .procname       = "stat_interval",
1003                 .data           = &sysctl_stat_interval,
1004                 .maxlen         = sizeof(sysctl_stat_interval),
1005                 .mode           = 0644,
1006                 .proc_handler   = &proc_dointvec_jiffies,
1007                 .strategy       = &sysctl_jiffies,
1008         },
1009 #endif
1010 #ifdef CONFIG_SECURITY
1011         {
1012                 .ctl_name       = CTL_UNNUMBERED,
1013                 .procname       = "mmap_min_addr",
1014                 .data           = &mmap_min_addr,
1015                 .maxlen         = sizeof(unsigned long),
1016                 .mode           = 0644,
1017                 .proc_handler   = &proc_doulongvec_minmax,
1018         },
1019 #endif
1020 #ifdef CONFIG_NUMA
1021         {
1022                 .ctl_name       = CTL_UNNUMBERED,
1023                 .procname       = "numa_zonelist_order",
1024                 .data           = &numa_zonelist_order,
1025                 .maxlen         = NUMA_ZONELIST_ORDER_LEN,
1026                 .mode           = 0644,
1027                 .proc_handler   = &numa_zonelist_order_handler,
1028                 .strategy       = &sysctl_string,
1029         },
1030 #endif
1031 #if (defined(CONFIG_X86_32) && !defined(CONFIG_UML))|| \
1032    (defined(CONFIG_SUPERH) && defined(CONFIG_VSYSCALL))
1033         {
1034                 .ctl_name       = VM_VDSO_ENABLED,
1035                 .procname       = "vdso_enabled",
1036                 .data           = &vdso_enabled,
1037                 .maxlen         = sizeof(vdso_enabled),
1038                 .mode           = 0644,
1039                 .proc_handler   = &proc_dointvec,
1040                 .strategy       = &sysctl_intvec,
1041                 .extra1         = &zero,
1042         },
1043 #endif
1044 /*
1045  * NOTE: do not add new entries to this table unless you have read
1046  * Documentation/sysctl/ctl_unnumbered.txt
1047  */
1048         { .ctl_name = 0 }
1049 };
1050
1051 #if defined(CONFIG_BINFMT_MISC) || defined(CONFIG_BINFMT_MISC_MODULE)
1052 static ctl_table binfmt_misc_table[] = {
1053         { .ctl_name = 0 }
1054 };
1055 #endif
1056
1057 static ctl_table fs_table[] = {
1058         {
1059                 .ctl_name       = FS_NRINODE,
1060                 .procname       = "inode-nr",
1061                 .data           = &inodes_stat,
1062                 .maxlen         = 2*sizeof(int),
1063                 .mode           = 0444,
1064                 .proc_handler   = &proc_dointvec,
1065         },
1066         {
1067                 .ctl_name       = FS_STATINODE,
1068                 .procname       = "inode-state",
1069                 .data           = &inodes_stat,
1070                 .maxlen         = 7*sizeof(int),
1071                 .mode           = 0444,
1072                 .proc_handler   = &proc_dointvec,
1073         },
1074         {
1075                 .ctl_name       = FS_NRFILE,
1076                 .procname       = "file-nr",
1077                 .data           = &files_stat,
1078                 .maxlen         = 3*sizeof(int),
1079                 .mode           = 0444,
1080                 .proc_handler   = &proc_nr_files,
1081         },
1082         {
1083                 .ctl_name       = FS_MAXFILE,
1084                 .procname       = "file-max",
1085                 .data           = &files_stat.max_files,
1086                 .maxlen         = sizeof(int),
1087                 .mode           = 0644,
1088                 .proc_handler   = &proc_dointvec,
1089         },
1090         {
1091                 .ctl_name       = FS_DENTRY,
1092                 .procname       = "dentry-state",
1093                 .data           = &dentry_stat,
1094                 .maxlen         = 6*sizeof(int),
1095                 .mode           = 0444,
1096                 .proc_handler   = &proc_dointvec,
1097         },
1098         {
1099                 .ctl_name       = FS_OVERFLOWUID,
1100                 .procname       = "overflowuid",
1101                 .data           = &fs_overflowuid,
1102                 .maxlen         = sizeof(int),
1103                 .mode           = 0644,
1104                 .proc_handler   = &proc_dointvec_minmax,
1105                 .strategy       = &sysctl_intvec,
1106                 .extra1         = &minolduid,
1107                 .extra2         = &maxolduid,
1108         },
1109         {
1110                 .ctl_name       = FS_OVERFLOWGID,
1111                 .procname       = "overflowgid",
1112                 .data           = &fs_overflowgid,
1113                 .maxlen         = sizeof(int),
1114                 .mode           = 0644,
1115                 .proc_handler   = &proc_dointvec_minmax,
1116                 .strategy       = &sysctl_intvec,
1117                 .extra1         = &minolduid,
1118                 .extra2         = &maxolduid,
1119         },
1120         {
1121                 .ctl_name       = FS_LEASES,
1122                 .procname       = "leases-enable",
1123                 .data           = &leases_enable,
1124                 .maxlen         = sizeof(int),
1125                 .mode           = 0644,
1126                 .proc_handler   = &proc_dointvec,
1127         },
1128 #ifdef CONFIG_DNOTIFY
1129         {
1130                 .ctl_name       = FS_DIR_NOTIFY,
1131                 .procname       = "dir-notify-enable",
1132                 .data           = &dir_notify_enable,
1133                 .maxlen         = sizeof(int),
1134                 .mode           = 0644,
1135                 .proc_handler   = &proc_dointvec,
1136         },
1137 #endif
1138 #ifdef CONFIG_MMU
1139         {
1140                 .ctl_name       = FS_LEASE_TIME,
1141                 .procname       = "lease-break-time",
1142                 .data           = &lease_break_time,
1143                 .maxlen         = sizeof(int),
1144                 .mode           = 0644,
1145                 .proc_handler   = &proc_dointvec_minmax,
1146                 .strategy       = &sysctl_intvec,
1147                 .extra1         = &zero,
1148                 .extra2         = &two,
1149         },
1150         {
1151                 .ctl_name       = FS_AIO_NR,
1152                 .procname       = "aio-nr",
1153                 .data           = &aio_nr,
1154                 .maxlen         = sizeof(aio_nr),
1155                 .mode           = 0444,
1156                 .proc_handler   = &proc_doulongvec_minmax,
1157         },
1158         {
1159                 .ctl_name       = FS_AIO_MAX_NR,
1160                 .procname       = "aio-max-nr",
1161                 .data           = &aio_max_nr,
1162                 .maxlen         = sizeof(aio_max_nr),
1163                 .mode           = 0644,
1164                 .proc_handler   = &proc_doulongvec_minmax,
1165         },
1166 #ifdef CONFIG_INOTIFY_USER
1167         {
1168                 .ctl_name       = FS_INOTIFY,
1169                 .procname       = "inotify",
1170                 .mode           = 0555,
1171                 .child          = inotify_table,
1172         },
1173 #endif
1174 #endif
1175         {
1176                 .ctl_name       = KERN_SETUID_DUMPABLE,
1177                 .procname       = "suid_dumpable",
1178                 .data           = &suid_dumpable,
1179                 .maxlen         = sizeof(int),
1180                 .mode           = 0644,
1181                 .proc_handler   = &proc_dointvec,
1182         },
1183 #if defined(CONFIG_BINFMT_MISC) || defined(CONFIG_BINFMT_MISC_MODULE)
1184         {
1185                 .ctl_name       = CTL_UNNUMBERED,
1186                 .procname       = "binfmt_misc",
1187                 .mode           = 0555,
1188                 .child          = binfmt_misc_table,
1189         },
1190 #endif
1191 /*
1192  * NOTE: do not add new entries to this table unless you have read
1193  * Documentation/sysctl/ctl_unnumbered.txt
1194  */
1195         { .ctl_name = 0 }
1196 };
1197
1198 static ctl_table debug_table[] = {
1199 #if defined(CONFIG_X86) || defined(CONFIG_PPC)
1200         {
1201                 .ctl_name       = CTL_UNNUMBERED,
1202                 .procname       = "exception-trace",
1203                 .data           = &show_unhandled_signals,
1204                 .maxlen         = sizeof(int),
1205                 .mode           = 0644,
1206                 .proc_handler   = proc_dointvec
1207         },
1208 #endif
1209         { .ctl_name = 0 }
1210 };
1211
1212 static ctl_table dev_table[] = {
1213         { .ctl_name = 0 }
1214 };
1215
1216 static DEFINE_SPINLOCK(sysctl_lock);
1217
1218 /* called under sysctl_lock */
1219 static int use_table(struct ctl_table_header *p)
1220 {
1221         if (unlikely(p->unregistering))
1222                 return 0;
1223         p->used++;
1224         return 1;
1225 }
1226
1227 /* called under sysctl_lock */
1228 static void unuse_table(struct ctl_table_header *p)
1229 {
1230         if (!--p->used)
1231                 if (unlikely(p->unregistering))
1232                         complete(p->unregistering);
1233 }
1234
1235 /* called under sysctl_lock, will reacquire if has to wait */
1236 static void start_unregistering(struct ctl_table_header *p)
1237 {
1238         /*
1239          * if p->used is 0, nobody will ever touch that entry again;
1240          * we'll eliminate all paths to it before dropping sysctl_lock
1241          */
1242         if (unlikely(p->used)) {
1243                 struct completion wait;
1244                 init_completion(&wait);
1245                 p->unregistering = &wait;
1246                 spin_unlock(&sysctl_lock);
1247                 wait_for_completion(&wait);
1248                 spin_lock(&sysctl_lock);
1249         }
1250         /*
1251          * do not remove from the list until nobody holds it; walking the
1252          * list in do_sysctl() relies on that.
1253          */
1254         list_del_init(&p->ctl_entry);
1255 }
1256
1257 void sysctl_head_finish(struct ctl_table_header *head)
1258 {
1259         if (!head)
1260                 return;
1261         spin_lock(&sysctl_lock);
1262         unuse_table(head);
1263         spin_unlock(&sysctl_lock);
1264 }
1265
1266 struct ctl_table_header *sysctl_head_next(struct ctl_table_header *prev)
1267 {
1268         struct ctl_table_header *head;
1269         struct list_head *tmp;
1270         spin_lock(&sysctl_lock);
1271         if (prev) {
1272                 tmp = &prev->ctl_entry;
1273                 unuse_table(prev);
1274                 goto next;
1275         }
1276         tmp = &root_table_header.ctl_entry;
1277         for (;;) {
1278                 head = list_entry(tmp, struct ctl_table_header, ctl_entry);
1279
1280                 if (!use_table(head))
1281                         goto next;
1282                 spin_unlock(&sysctl_lock);
1283                 return head;
1284         next:
1285                 tmp = tmp->next;
1286                 if (tmp == &root_table_header.ctl_entry)
1287                         break;
1288         }
1289         spin_unlock(&sysctl_lock);
1290         return NULL;
1291 }
1292
1293 #ifdef CONFIG_SYSCTL_SYSCALL
1294 int do_sysctl(int __user *name, int nlen, void __user *oldval, size_t __user *oldlenp,
1295                void __user *newval, size_t newlen)
1296 {
1297         struct ctl_table_header *head;
1298         int error = -ENOTDIR;
1299
1300         if (nlen <= 0 || nlen >= CTL_MAXNAME)
1301                 return -ENOTDIR;
1302         if (oldval) {
1303                 int old_len;
1304                 if (!oldlenp || get_user(old_len, oldlenp))
1305                         return -EFAULT;
1306         }
1307
1308         for (head = sysctl_head_next(NULL); head;
1309                         head = sysctl_head_next(head)) {
1310                 error = parse_table(name, nlen, oldval, oldlenp,
1311                                         newval, newlen, head->ctl_table);
1312                 if (error != -ENOTDIR) {
1313                         sysctl_head_finish(head);
1314                         break;
1315                 }
1316         }
1317         return error;
1318 }
1319
1320 asmlinkage long sys_sysctl(struct __sysctl_args __user *args)
1321 {
1322         struct __sysctl_args tmp;
1323         int error;
1324
1325         if (copy_from_user(&tmp, args, sizeof(tmp)))
1326                 return -EFAULT;
1327
1328         lock_kernel();
1329         error = do_sysctl(tmp.name, tmp.nlen, tmp.oldval, tmp.oldlenp,
1330                           tmp.newval, tmp.newlen);
1331         unlock_kernel();
1332         return error;
1333 }
1334 #endif /* CONFIG_SYSCTL_SYSCALL */
1335
1336 /*
1337  * sysctl_perm does NOT grant the superuser all rights automatically, because
1338  * some sysctl variables are readonly even to root.
1339  */
1340
1341 static int test_perm(int mode, int op)
1342 {
1343         if (!current->euid)
1344                 mode >>= 6;
1345         else if (in_egroup_p(0))
1346                 mode >>= 3;
1347         if ((mode & op & 0007) == op)
1348                 return 0;
1349         return -EACCES;
1350 }
1351
1352 int sysctl_perm(ctl_table *table, int op)
1353 {
1354         int error;
1355         error = security_sysctl(table, op);
1356         if (error)
1357                 return error;
1358         return test_perm(table->mode, op);
1359 }
1360
1361 #ifdef CONFIG_SYSCTL_SYSCALL
1362 static int parse_table(int __user *name, int nlen,
1363                        void __user *oldval, size_t __user *oldlenp,
1364                        void __user *newval, size_t newlen,
1365                        ctl_table *table)
1366 {
1367         int n;
1368 repeat:
1369         if (!nlen)
1370                 return -ENOTDIR;
1371         if (get_user(n, name))
1372                 return -EFAULT;
1373         for ( ; table->ctl_name || table->procname; table++) {
1374                 if (!table->ctl_name)
1375                         continue;
1376                 if (n == table->ctl_name) {
1377                         int error;
1378                         if (table->child) {
1379                                 if (sysctl_perm(table, 001))
1380                                         return -EPERM;
1381                                 name++;
1382                                 nlen--;
1383                                 table = table->child;
1384                                 goto repeat;
1385                         }
1386                         error = do_sysctl_strategy(table, name, nlen,
1387                                                    oldval, oldlenp,
1388                                                    newval, newlen);
1389                         return error;
1390                 }
1391         }
1392         return -ENOTDIR;
1393 }
1394
1395 /* Perform the actual read/write of a sysctl table entry. */
1396 int do_sysctl_strategy (ctl_table *table,
1397                         int __user *name, int nlen,
1398                         void __user *oldval, size_t __user *oldlenp,
1399                         void __user *newval, size_t newlen)
1400 {
1401         int op = 0, rc;
1402         size_t len;
1403
1404         if (oldval)
1405                 op |= 004;
1406         if (newval)
1407                 op |= 002;
1408         if (sysctl_perm(table, op))
1409                 return -EPERM;
1410
1411         if (table->strategy) {
1412                 rc = table->strategy(table, name, nlen, oldval, oldlenp,
1413                                      newval, newlen);
1414                 if (rc < 0)
1415                         return rc;
1416                 if (rc > 0)
1417                         return 0;
1418         }
1419
1420         /* If there is no strategy routine, or if the strategy returns
1421          * zero, proceed with automatic r/w */
1422         if (table->data && table->maxlen) {
1423                 if (oldval && oldlenp) {
1424                         if (get_user(len, oldlenp))
1425                                 return -EFAULT;
1426                         if (len) {
1427                                 if (len > table->maxlen)
1428                                         len = table->maxlen;
1429                                 if(copy_to_user(oldval, table->data, len))
1430                                         return -EFAULT;
1431                                 if(put_user(len, oldlenp))
1432                                         return -EFAULT;
1433                         }
1434                 }
1435                 if (newval && newlen) {
1436                         len = newlen;
1437                         if (len > table->maxlen)
1438                                 len = table->maxlen;
1439                         if(copy_from_user(table->data, newval, len))
1440                                 return -EFAULT;
1441                 }
1442         }
1443         return 0;
1444 }
1445 #endif /* CONFIG_SYSCTL_SYSCALL */
1446
1447 static void sysctl_set_parent(struct ctl_table *parent, struct ctl_table *table)
1448 {
1449         for (; table->ctl_name || table->procname; table++) {
1450                 table->parent = parent;
1451                 if (table->child)
1452                         sysctl_set_parent(table, table->child);
1453         }
1454 }
1455
1456 static __init int sysctl_init(void)
1457 {
1458         sysctl_set_parent(NULL, root_table);
1459         return 0;
1460 }
1461
1462 core_initcall(sysctl_init);
1463
1464 /**
1465  * register_sysctl_table - register a sysctl hierarchy
1466  * @table: the top-level table structure
1467  *
1468  * Register a sysctl table hierarchy. @table should be a filled in ctl_table
1469  * array. An entry with a ctl_name of 0 terminates the table.
1470  *
1471  * The members of the &ctl_table structure are used as follows:
1472  *
1473  * ctl_name - This is the numeric sysctl value used by sysctl(2). The number
1474  *            must be unique within that level of sysctl
1475  *
1476  * procname - the name of the sysctl file under /proc/sys. Set to %NULL to not
1477  *            enter a sysctl file
1478  *
1479  * data - a pointer to data for use by proc_handler
1480  *
1481  * maxlen - the maximum size in bytes of the data
1482  *
1483  * mode - the file permissions for the /proc/sys file, and for sysctl(2)
1484  *
1485  * child - a pointer to the child sysctl table if this entry is a directory, or
1486  *         %NULL.
1487  *
1488  * proc_handler - the text handler routine (described below)
1489  *
1490  * strategy - the strategy routine (described below)
1491  *
1492  * de - for internal use by the sysctl routines
1493  *
1494  * extra1, extra2 - extra pointers usable by the proc handler routines
1495  *
1496  * Leaf nodes in the sysctl tree will be represented by a single file
1497  * under /proc; non-leaf nodes will be represented by directories.
1498  *
1499  * sysctl(2) can automatically manage read and write requests through
1500  * the sysctl table.  The data and maxlen fields of the ctl_table
1501  * struct enable minimal validation of the values being written to be
1502  * performed, and the mode field allows minimal authentication.
1503  *
1504  * More sophisticated management can be enabled by the provision of a
1505  * strategy routine with the table entry.  This will be called before
1506  * any automatic read or write of the data is performed.
1507  *
1508  * The strategy routine may return
1509  *
1510  * < 0 - Error occurred (error is passed to user process)
1511  *
1512  * 0   - OK - proceed with automatic read or write.
1513  *
1514  * > 0 - OK - read or write has been done by the strategy routine, so
1515  *       return immediately.
1516  *
1517  * There must be a proc_handler routine for any terminal nodes
1518  * mirrored under /proc/sys (non-terminals are handled by a built-in
1519  * directory handler).  Several default handlers are available to
1520  * cover common cases -
1521  *
1522  * proc_dostring(), proc_dointvec(), proc_dointvec_jiffies(),
1523  * proc_dointvec_userhz_jiffies(), proc_dointvec_minmax(),
1524  * proc_doulongvec_ms_jiffies_minmax(), proc_doulongvec_minmax()
1525  *
1526  * It is the handler's job to read the input buffer from user memory
1527  * and process it. The handler should return 0 on success.
1528  *
1529  * This routine returns %NULL on a failure to register, and a pointer
1530  * to the table header on success.
1531  */
1532 struct ctl_table_header *register_sysctl_table(ctl_table * table)
1533 {
1534         struct ctl_table_header *tmp;
1535         tmp = kmalloc(sizeof(struct ctl_table_header), GFP_KERNEL);
1536         if (!tmp)
1537                 return NULL;
1538         tmp->ctl_table = table;
1539         INIT_LIST_HEAD(&tmp->ctl_entry);
1540         tmp->used = 0;
1541         tmp->unregistering = NULL;
1542         sysctl_set_parent(NULL, table);
1543         spin_lock(&sysctl_lock);
1544         list_add_tail(&tmp->ctl_entry, &root_table_header.ctl_entry);
1545         spin_unlock(&sysctl_lock);
1546         return tmp;
1547 }
1548
1549 /**
1550  * unregister_sysctl_table - unregister a sysctl table hierarchy
1551  * @header: the header returned from register_sysctl_table
1552  *
1553  * Unregisters the sysctl table and all children. proc entries may not
1554  * actually be removed until they are no longer used by anyone.
1555  */
1556 void unregister_sysctl_table(struct ctl_table_header * header)
1557 {
1558         might_sleep();
1559         spin_lock(&sysctl_lock);
1560         start_unregistering(header);
1561         spin_unlock(&sysctl_lock);
1562         kfree(header);
1563 }
1564
1565 #else /* !CONFIG_SYSCTL */
1566 struct ctl_table_header *register_sysctl_table(ctl_table * table)
1567 {
1568         return NULL;
1569 }
1570
1571 void unregister_sysctl_table(struct ctl_table_header * table)
1572 {
1573 }
1574
1575 #endif /* CONFIG_SYSCTL */
1576
1577 /*
1578  * /proc/sys support
1579  */
1580
1581 #ifdef CONFIG_PROC_SYSCTL
1582
1583 static int _proc_do_string(void* data, int maxlen, int write,
1584                            struct file *filp, void __user *buffer,
1585                            size_t *lenp, loff_t *ppos)
1586 {
1587         size_t len;
1588         char __user *p;
1589         char c;
1590
1591         if (!data || !maxlen || !*lenp) {
1592                 *lenp = 0;
1593                 return 0;
1594         }
1595
1596         if (write) {
1597                 len = 0;
1598                 p = buffer;
1599                 while (len < *lenp) {
1600                         if (get_user(c, p++))
1601                                 return -EFAULT;
1602                         if (c == 0 || c == '\n')
1603                                 break;
1604                         len++;
1605                 }
1606                 if (len >= maxlen)
1607                         len = maxlen-1;
1608                 if(copy_from_user(data, buffer, len))
1609                         return -EFAULT;
1610                 ((char *) data)[len] = 0;
1611                 *ppos += *lenp;
1612         } else {
1613                 len = strlen(data);
1614                 if (len > maxlen)
1615                         len = maxlen;
1616
1617                 if (*ppos > len) {
1618                         *lenp = 0;
1619                         return 0;
1620                 }
1621
1622                 data += *ppos;
1623                 len  -= *ppos;
1624
1625                 if (len > *lenp)
1626                         len = *lenp;
1627                 if (len)
1628                         if(copy_to_user(buffer, data, len))
1629                                 return -EFAULT;
1630                 if (len < *lenp) {
1631                         if(put_user('\n', ((char __user *) buffer) + len))
1632                                 return -EFAULT;
1633                         len++;
1634                 }
1635                 *lenp = len;
1636                 *ppos += len;
1637         }
1638         return 0;
1639 }
1640
1641 /**
1642  * proc_dostring - read a string sysctl
1643  * @table: the sysctl table
1644  * @write: %TRUE if this is a write to the sysctl file
1645  * @filp: the file structure
1646  * @buffer: the user buffer
1647  * @lenp: the size of the user buffer
1648  * @ppos: file position
1649  *
1650  * Reads/writes a string from/to the user buffer. If the kernel
1651  * buffer provided is not large enough to hold the string, the
1652  * string is truncated. The copied string is %NULL-terminated.
1653  * If the string is being read by the user process, it is copied
1654  * and a newline '\n' is added. It is truncated if the buffer is
1655  * not large enough.
1656  *
1657  * Returns 0 on success.
1658  */
1659 int proc_dostring(ctl_table *table, int write, struct file *filp,
1660                   void __user *buffer, size_t *lenp, loff_t *ppos)
1661 {
1662         return _proc_do_string(table->data, table->maxlen, write, filp,
1663                                buffer, lenp, ppos);
1664 }
1665
1666
1667 static int do_proc_dointvec_conv(int *negp, unsigned long *lvalp,
1668                                  int *valp,
1669                                  int write, void *data)
1670 {
1671         if (write) {
1672                 *valp = *negp ? -*lvalp : *lvalp;
1673         } else {
1674                 int val = *valp;
1675                 if (val < 0) {
1676                         *negp = -1;
1677                         *lvalp = (unsigned long)-val;
1678                 } else {
1679                         *negp = 0;
1680                         *lvalp = (unsigned long)val;
1681                 }
1682         }
1683         return 0;
1684 }
1685
1686 static int __do_proc_dointvec(void *tbl_data, ctl_table *table,
1687                   int write, struct file *filp, void __user *buffer,
1688                   size_t *lenp, loff_t *ppos,
1689                   int (*conv)(int *negp, unsigned long *lvalp, int *valp,
1690                               int write, void *data),
1691                   void *data)
1692 {
1693 #define TMPBUFLEN 21
1694         int *i, vleft, first=1, neg, val;
1695         unsigned long lval;
1696         size_t left, len;
1697
1698         char buf[TMPBUFLEN], *p;
1699         char __user *s = buffer;
1700
1701         if (!tbl_data || !table->maxlen || !*lenp ||
1702             (*ppos && !write)) {
1703                 *lenp = 0;
1704                 return 0;
1705         }
1706
1707         i = (int *) tbl_data;
1708         vleft = table->maxlen / sizeof(*i);
1709         left = *lenp;
1710
1711         if (!conv)
1712                 conv = do_proc_dointvec_conv;
1713
1714         for (; left && vleft--; i++, first=0) {
1715                 if (write) {
1716                         while (left) {
1717                                 char c;
1718                                 if (get_user(c, s))
1719                                         return -EFAULT;
1720                                 if (!isspace(c))
1721                                         break;
1722                                 left--;
1723                                 s++;
1724                         }
1725                         if (!left)
1726                                 break;
1727                         neg = 0;
1728                         len = left;
1729                         if (len > sizeof(buf) - 1)
1730                                 len = sizeof(buf) - 1;
1731                         if (copy_from_user(buf, s, len))
1732                                 return -EFAULT;
1733                         buf[len] = 0;
1734                         p = buf;
1735                         if (*p == '-' && left > 1) {
1736                                 neg = 1;
1737                                 p++;
1738                         }
1739                         if (*p < '0' || *p > '9')
1740                                 break;
1741
1742                         lval = simple_strtoul(p, &p, 0);
1743
1744                         len = p-buf;
1745                         if ((len < left) && *p && !isspace(*p))
1746                                 break;
1747                         if (neg)
1748                                 val = -val;
1749                         s += len;
1750                         left -= len;
1751
1752                         if (conv(&neg, &lval, i, 1, data))
1753                                 break;
1754                 } else {
1755                         p = buf;
1756                         if (!first)
1757                                 *p++ = '\t';
1758
1759                         if (conv(&neg, &lval, i, 0, data))
1760                                 break;
1761
1762                         sprintf(p, "%s%lu", neg ? "-" : "", lval);
1763                         len = strlen(buf);
1764                         if (len > left)
1765                                 len = left;
1766                         if(copy_to_user(s, buf, len))
1767                                 return -EFAULT;
1768                         left -= len;
1769                         s += len;
1770                 }
1771         }
1772
1773         if (!write && !first && left) {
1774                 if(put_user('\n', s))
1775                         return -EFAULT;
1776                 left--, s++;
1777         }
1778         if (write) {
1779                 while (left) {
1780                         char c;
1781                         if (get_user(c, s++))
1782                                 return -EFAULT;
1783                         if (!isspace(c))
1784                                 break;
1785                         left--;
1786                 }
1787         }
1788         if (write && first)
1789                 return -EINVAL;
1790         *lenp -= left;
1791         *ppos += *lenp;
1792         return 0;
1793 #undef TMPBUFLEN
1794 }
1795
1796 static int do_proc_dointvec(ctl_table *table, int write, struct file *filp,
1797                   void __user *buffer, size_t *lenp, loff_t *ppos,
1798                   int (*conv)(int *negp, unsigned long *lvalp, int *valp,
1799                               int write, void *data),
1800                   void *data)
1801 {
1802         return __do_proc_dointvec(table->data, table, write, filp,
1803                         buffer, lenp, ppos, conv, data);
1804 }
1805
1806 /**
1807  * proc_dointvec - read a vector of integers
1808  * @table: the sysctl table
1809  * @write: %TRUE if this is a write to the sysctl file
1810  * @filp: the file structure
1811  * @buffer: the user buffer
1812  * @lenp: the size of the user buffer
1813  * @ppos: file position
1814  *
1815  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
1816  * values from/to the user buffer, treated as an ASCII string.
1817  *
1818  * Returns 0 on success.
1819  */
1820 int proc_dointvec(ctl_table *table, int write, struct file *filp,
1821                      void __user *buffer, size_t *lenp, loff_t *ppos)
1822 {
1823     return do_proc_dointvec(table,write,filp,buffer,lenp,ppos,
1824                             NULL,NULL);
1825 }
1826
1827 #define OP_SET  0
1828 #define OP_AND  1
1829 #define OP_OR   2
1830
1831 static int do_proc_dointvec_bset_conv(int *negp, unsigned long *lvalp,
1832                                       int *valp,
1833                                       int write, void *data)
1834 {
1835         int op = *(int *)data;
1836         if (write) {
1837                 int val = *negp ? -*lvalp : *lvalp;
1838                 switch(op) {
1839                 case OP_SET:    *valp = val; break;
1840                 case OP_AND:    *valp &= val; break;
1841                 case OP_OR:     *valp |= val; break;
1842                 }
1843         } else {
1844                 int val = *valp;
1845                 if (val < 0) {
1846                         *negp = -1;
1847                         *lvalp = (unsigned long)-val;
1848                 } else {
1849                         *negp = 0;
1850                         *lvalp = (unsigned long)val;
1851                 }
1852         }
1853         return 0;
1854 }
1855
1856 /*
1857  *      init may raise the set.
1858  */
1859
1860 int proc_dointvec_bset(ctl_table *table, int write, struct file *filp,
1861                         void __user *buffer, size_t *lenp, loff_t *ppos)
1862 {
1863         int op;
1864
1865         if (write && !capable(CAP_SYS_MODULE)) {
1866                 return -EPERM;
1867         }
1868
1869         op = is_init(current) ? OP_SET : OP_AND;
1870         return do_proc_dointvec(table,write,filp,buffer,lenp,ppos,
1871                                 do_proc_dointvec_bset_conv,&op);
1872 }
1873
1874 /*
1875  *      Taint values can only be increased
1876  */
1877 static int proc_dointvec_taint(ctl_table *table, int write, struct file *filp,
1878                                void __user *buffer, size_t *lenp, loff_t *ppos)
1879 {
1880         int op;
1881
1882         if (write && !capable(CAP_SYS_ADMIN))
1883                 return -EPERM;
1884
1885         op = OP_OR;
1886         return do_proc_dointvec(table,write,filp,buffer,lenp,ppos,
1887                                 do_proc_dointvec_bset_conv,&op);
1888 }
1889
1890 struct do_proc_dointvec_minmax_conv_param {
1891         int *min;
1892         int *max;
1893 };
1894
1895 static int do_proc_dointvec_minmax_conv(int *negp, unsigned long *lvalp,
1896                                         int *valp,
1897                                         int write, void *data)
1898 {
1899         struct do_proc_dointvec_minmax_conv_param *param = data;
1900         if (write) {
1901                 int val = *negp ? -*lvalp : *lvalp;
1902                 if ((param->min && *param->min > val) ||
1903                     (param->max && *param->max < val))
1904                         return -EINVAL;
1905                 *valp = val;
1906         } else {
1907                 int val = *valp;
1908                 if (val < 0) {
1909                         *negp = -1;
1910                         *lvalp = (unsigned long)-val;
1911                 } else {
1912                         *negp = 0;
1913                         *lvalp = (unsigned long)val;
1914                 }
1915         }
1916         return 0;
1917 }
1918
1919 /**
1920  * proc_dointvec_minmax - read a vector of integers with min/max values
1921  * @table: the sysctl table
1922  * @write: %TRUE if this is a write to the sysctl file
1923  * @filp: the file structure
1924  * @buffer: the user buffer
1925  * @lenp: the size of the user buffer
1926  * @ppos: file position
1927  *
1928  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
1929  * values from/to the user buffer, treated as an ASCII string.
1930  *
1931  * This routine will ensure the values are within the range specified by
1932  * table->extra1 (min) and table->extra2 (max).
1933  *
1934  * Returns 0 on success.
1935  */
1936 int proc_dointvec_minmax(ctl_table *table, int write, struct file *filp,
1937                   void __user *buffer, size_t *lenp, loff_t *ppos)
1938 {
1939         struct do_proc_dointvec_minmax_conv_param param = {
1940                 .min = (int *) table->extra1,
1941                 .max = (int *) table->extra2,
1942         };
1943         return do_proc_dointvec(table, write, filp, buffer, lenp, ppos,
1944                                 do_proc_dointvec_minmax_conv, &param);
1945 }
1946
1947 static int __do_proc_doulongvec_minmax(void *data, ctl_table *table, int write,
1948                                      struct file *filp,
1949                                      void __user *buffer,
1950                                      size_t *lenp, loff_t *ppos,
1951                                      unsigned long convmul,
1952                                      unsigned long convdiv)
1953 {
1954 #define TMPBUFLEN 21
1955         unsigned long *i, *min, *max, val;
1956         int vleft, first=1, neg;
1957         size_t len, left;
1958         char buf[TMPBUFLEN], *p;
1959         char __user *s = buffer;
1960
1961         if (!data || !table->maxlen || !*lenp ||
1962             (*ppos && !write)) {
1963                 *lenp = 0;
1964                 return 0;
1965         }
1966
1967         i = (unsigned long *) data;
1968         min = (unsigned long *) table->extra1;
1969         max = (unsigned long *) table->extra2;
1970         vleft = table->maxlen / sizeof(unsigned long);
1971         left = *lenp;
1972
1973         for (; left && vleft--; i++, min++, max++, first=0) {
1974                 if (write) {
1975                         while (left) {
1976                                 char c;
1977                                 if (get_user(c, s))
1978                                         return -EFAULT;
1979                                 if (!isspace(c))
1980                                         break;
1981                                 left--;
1982                                 s++;
1983                         }
1984                         if (!left)
1985                                 break;
1986                         neg = 0;
1987                         len = left;
1988                         if (len > TMPBUFLEN-1)
1989                                 len = TMPBUFLEN-1;
1990                         if (copy_from_user(buf, s, len))
1991                                 return -EFAULT;
1992                         buf[len] = 0;
1993                         p = buf;
1994                         if (*p == '-' && left > 1) {
1995                                 neg = 1;
1996                                 p++;
1997                         }
1998                         if (*p < '0' || *p > '9')
1999                                 break;
2000                         val = simple_strtoul(p, &p, 0) * convmul / convdiv ;
2001                         len = p-buf;
2002                         if ((len < left) && *p && !isspace(*p))
2003                                 break;
2004                         if (neg)
2005                                 val = -val;
2006                         s += len;
2007                         left -= len;
2008
2009                         if(neg)
2010                                 continue;
2011                         if ((min && val < *min) || (max && val > *max))
2012                                 continue;
2013                         *i = val;
2014                 } else {
2015                         p = buf;
2016                         if (!first)
2017                                 *p++ = '\t';
2018                         sprintf(p, "%lu", convdiv * (*i) / convmul);
2019                         len = strlen(buf);
2020                         if (len > left)
2021                                 len = left;
2022                         if(copy_to_user(s, buf, len))
2023                                 return -EFAULT;
2024                         left -= len;
2025                         s += len;
2026                 }
2027         }
2028
2029         if (!write && !first && left) {
2030                 if(put_user('\n', s))
2031                         return -EFAULT;
2032                 left--, s++;
2033         }
2034         if (write) {
2035                 while (left) {
2036                         char c;
2037                         if (get_user(c, s++))
2038                                 return -EFAULT;
2039                         if (!isspace(c))
2040                                 break;
2041                         left--;
2042                 }
2043         }
2044         if (write && first)
2045                 return -EINVAL;
2046         *lenp -= left;
2047         *ppos += *lenp;
2048         return 0;
2049 #undef TMPBUFLEN
2050 }
2051
2052 static int do_proc_doulongvec_minmax(ctl_table *table, int write,
2053                                      struct file *filp,
2054                                      void __user *buffer,
2055                                      size_t *lenp, loff_t *ppos,
2056                                      unsigned long convmul,
2057                                      unsigned long convdiv)
2058 {
2059         return __do_proc_doulongvec_minmax(table->data, table, write,
2060                         filp, buffer, lenp, ppos, convmul, convdiv);
2061 }
2062
2063 /**
2064  * proc_doulongvec_minmax - read a vector of long integers with min/max values
2065  * @table: the sysctl table
2066  * @write: %TRUE if this is a write to the sysctl file
2067  * @filp: the file structure
2068  * @buffer: the user buffer
2069  * @lenp: the size of the user buffer
2070  * @ppos: file position
2071  *
2072  * Reads/writes up to table->maxlen/sizeof(unsigned long) unsigned long
2073  * values from/to the user buffer, treated as an ASCII string.
2074  *
2075  * This routine will ensure the values are within the range specified by
2076  * table->extra1 (min) and table->extra2 (max).
2077  *
2078  * Returns 0 on success.
2079  */
2080 int proc_doulongvec_minmax(ctl_table *table, int write, struct file *filp,
2081                            void __user *buffer, size_t *lenp, loff_t *ppos)
2082 {
2083     return do_proc_doulongvec_minmax(table, write, filp, buffer, lenp, ppos, 1l, 1l);
2084 }
2085
2086 /**
2087  * proc_doulongvec_ms_jiffies_minmax - read a vector of millisecond values with min/max values
2088  * @table: the sysctl table
2089  * @write: %TRUE if this is a write to the sysctl file
2090  * @filp: the file structure
2091  * @buffer: the user buffer
2092  * @lenp: the size of the user buffer
2093  * @ppos: file position
2094  *
2095  * Reads/writes up to table->maxlen/sizeof(unsigned long) unsigned long
2096  * values from/to the user buffer, treated as an ASCII string. The values
2097  * are treated as milliseconds, and converted to jiffies when they are stored.
2098  *
2099  * This routine will ensure the values are within the range specified by
2100  * table->extra1 (min) and table->extra2 (max).
2101  *
2102  * Returns 0 on success.
2103  */
2104 int proc_doulongvec_ms_jiffies_minmax(ctl_table *table, int write,
2105                                       struct file *filp,
2106                                       void __user *buffer,
2107                                       size_t *lenp, loff_t *ppos)
2108 {
2109     return do_proc_doulongvec_minmax(table, write, filp, buffer,
2110                                      lenp, ppos, HZ, 1000l);
2111 }
2112
2113
2114 static int do_proc_dointvec_jiffies_conv(int *negp, unsigned long *lvalp,
2115                                          int *valp,
2116                                          int write, void *data)
2117 {
2118         if (write) {
2119                 if (*lvalp > LONG_MAX / HZ)
2120                         return 1;
2121                 *valp = *negp ? -(*lvalp*HZ) : (*lvalp*HZ);
2122         } else {
2123                 int val = *valp;
2124                 unsigned long lval;
2125                 if (val < 0) {
2126                         *negp = -1;
2127                         lval = (unsigned long)-val;
2128                 } else {
2129                         *negp = 0;
2130                         lval = (unsigned long)val;
2131                 }
2132                 *lvalp = lval / HZ;
2133         }
2134         return 0;
2135 }
2136
2137 static int do_proc_dointvec_userhz_jiffies_conv(int *negp, unsigned long *lvalp,
2138                                                 int *valp,
2139                                                 int write, void *data)
2140 {
2141         if (write) {
2142                 if (USER_HZ < HZ && *lvalp > (LONG_MAX / HZ) * USER_HZ)
2143                         return 1;
2144                 *valp = clock_t_to_jiffies(*negp ? -*lvalp : *lvalp);
2145         } else {
2146                 int val = *valp;
2147                 unsigned long lval;
2148                 if (val < 0) {
2149                         *negp = -1;
2150                         lval = (unsigned long)-val;
2151                 } else {
2152                         *negp = 0;
2153                         lval = (unsigned long)val;
2154                 }
2155                 *lvalp = jiffies_to_clock_t(lval);
2156         }
2157         return 0;
2158 }
2159
2160 static int do_proc_dointvec_ms_jiffies_conv(int *negp, unsigned long *lvalp,
2161                                             int *valp,
2162                                             int write, void *data)
2163 {
2164         if (write) {
2165                 *valp = msecs_to_jiffies(*negp ? -*lvalp : *lvalp);
2166         } else {
2167                 int val = *valp;
2168                 unsigned long lval;
2169                 if (val < 0) {
2170                         *negp = -1;
2171                         lval = (unsigned long)-val;
2172                 } else {
2173                         *negp = 0;
2174                         lval = (unsigned long)val;
2175                 }
2176                 *lvalp = jiffies_to_msecs(lval);
2177         }
2178         return 0;
2179 }
2180
2181 /**
2182  * proc_dointvec_jiffies - read a vector of integers as seconds
2183  * @table: the sysctl table
2184  * @write: %TRUE if this is a write to the sysctl file
2185  * @filp: the file structure
2186  * @buffer: the user buffer
2187  * @lenp: the size of the user buffer
2188  * @ppos: file position
2189  *
2190  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2191  * values from/to the user buffer, treated as an ASCII string.
2192  * The values read are assumed to be in seconds, and are converted into
2193  * jiffies.
2194  *
2195  * Returns 0 on success.
2196  */
2197 int proc_dointvec_jiffies(ctl_table *table, int write, struct file *filp,
2198                           void __user *buffer, size_t *lenp, loff_t *ppos)
2199 {
2200     return do_proc_dointvec(table,write,filp,buffer,lenp,ppos,
2201                             do_proc_dointvec_jiffies_conv,NULL);
2202 }
2203
2204 /**
2205  * proc_dointvec_userhz_jiffies - read a vector of integers as 1/USER_HZ seconds
2206  * @table: the sysctl table
2207  * @write: %TRUE if this is a write to the sysctl file
2208  * @filp: the file structure
2209  * @buffer: the user buffer
2210  * @lenp: the size of the user buffer
2211  * @ppos: pointer to the file position
2212  *
2213  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2214  * values from/to the user buffer, treated as an ASCII string.
2215  * The values read are assumed to be in 1/USER_HZ seconds, and
2216  * are converted into jiffies.
2217  *
2218  * Returns 0 on success.
2219  */
2220 int proc_dointvec_userhz_jiffies(ctl_table *table, int write, struct file *filp,
2221                                  void __user *buffer, size_t *lenp, loff_t *ppos)
2222 {
2223     return do_proc_dointvec(table,write,filp,buffer,lenp,ppos,
2224                             do_proc_dointvec_userhz_jiffies_conv,NULL);
2225 }
2226
2227 /**
2228  * proc_dointvec_ms_jiffies - read a vector of integers as 1 milliseconds
2229  * @table: the sysctl table
2230  * @write: %TRUE if this is a write to the sysctl file
2231  * @filp: the file structure
2232  * @buffer: the user buffer
2233  * @lenp: the size of the user buffer
2234  * @ppos: file position
2235  * @ppos: the current position in the file
2236  *
2237  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2238  * values from/to the user buffer, treated as an ASCII string.
2239  * The values read are assumed to be in 1/1000 seconds, and
2240  * are converted into jiffies.
2241  *
2242  * Returns 0 on success.
2243  */
2244 int proc_dointvec_ms_jiffies(ctl_table *table, int write, struct file *filp,
2245                              void __user *buffer, size_t *lenp, loff_t *ppos)
2246 {
2247         return do_proc_dointvec(table, write, filp, buffer, lenp, ppos,
2248                                 do_proc_dointvec_ms_jiffies_conv, NULL);
2249 }
2250
2251 static int proc_do_cad_pid(ctl_table *table, int write, struct file *filp,
2252                            void __user *buffer, size_t *lenp, loff_t *ppos)
2253 {
2254         struct pid *new_pid;
2255         pid_t tmp;
2256         int r;
2257
2258         tmp = pid_nr(cad_pid);
2259
2260         r = __do_proc_dointvec(&tmp, table, write, filp, buffer,
2261                                lenp, ppos, NULL, NULL);
2262         if (r || !write)
2263                 return r;
2264
2265         new_pid = find_get_pid(tmp);
2266         if (!new_pid)
2267                 return -ESRCH;
2268
2269         put_pid(xchg(&cad_pid, new_pid));
2270         return 0;
2271 }
2272
2273 #else /* CONFIG_PROC_FS */
2274
2275 int proc_dostring(ctl_table *table, int write, struct file *filp,
2276                   void __user *buffer, size_t *lenp, loff_t *ppos)
2277 {
2278         return -ENOSYS;
2279 }
2280
2281 int proc_dointvec(ctl_table *table, int write, struct file *filp,
2282                   void __user *buffer, size_t *lenp, loff_t *ppos)
2283 {
2284         return -ENOSYS;
2285 }
2286
2287 int proc_dointvec_bset(ctl_table *table, int write, struct file *filp,
2288                         void __user *buffer, size_t *lenp, loff_t *ppos)
2289 {
2290         return -ENOSYS;
2291 }
2292
2293 int proc_dointvec_minmax(ctl_table *table, int write, struct file *filp,
2294                     void __user *buffer, size_t *lenp, loff_t *ppos)
2295 {
2296         return -ENOSYS;
2297 }
2298
2299 int proc_dointvec_jiffies(ctl_table *table, int write, struct file *filp,
2300                     void __user *buffer, size_t *lenp, loff_t *ppos)
2301 {
2302         return -ENOSYS;
2303 }
2304
2305 int proc_dointvec_userhz_jiffies(ctl_table *table, int write, struct file *filp,
2306                     void __user *buffer, size_t *lenp, loff_t *ppos)
2307 {
2308         return -ENOSYS;
2309 }
2310
2311 int proc_dointvec_ms_jiffies(ctl_table *table, int write, struct file *filp,
2312                              void __user *buffer, size_t *lenp, loff_t *ppos)
2313 {
2314         return -ENOSYS;
2315 }
2316
2317 int proc_doulongvec_minmax(ctl_table *table, int write, struct file *filp,
2318                     void __user *buffer, size_t *lenp, loff_t *ppos)
2319 {
2320         return -ENOSYS;
2321 }
2322
2323 int proc_doulongvec_ms_jiffies_minmax(ctl_table *table, int write,
2324                                       struct file *filp,
2325                                       void __user *buffer,
2326                                       size_t *lenp, loff_t *ppos)
2327 {
2328     return -ENOSYS;
2329 }
2330
2331
2332 #endif /* CONFIG_PROC_FS */
2333
2334
2335 #ifdef CONFIG_SYSCTL_SYSCALL
2336 /*
2337  * General sysctl support routines
2338  */
2339
2340 /* The generic string strategy routine: */
2341 int sysctl_string(ctl_table *table, int __user *name, int nlen,
2342                   void __user *oldval, size_t __user *oldlenp,
2343                   void __user *newval, size_t newlen)
2344 {
2345         if (!table->data || !table->maxlen)
2346                 return -ENOTDIR;
2347
2348         if (oldval && oldlenp) {
2349                 size_t bufsize;
2350                 if (get_user(bufsize, oldlenp))
2351                         return -EFAULT;
2352                 if (bufsize) {
2353                         size_t len = strlen(table->data), copied;
2354
2355                         /* This shouldn't trigger for a well-formed sysctl */
2356                         if (len > table->maxlen)
2357                                 len = table->maxlen;
2358
2359                         /* Copy up to a max of bufsize-1 bytes of the string */
2360                         copied = (len >= bufsize) ? bufsize - 1 : len;
2361
2362                         if (copy_to_user(oldval, table->data, copied) ||
2363                             put_user(0, (char __user *)(oldval + copied)))
2364                                 return -EFAULT;
2365                         if (put_user(len, oldlenp))
2366                                 return -EFAULT;
2367                 }
2368         }
2369         if (newval && newlen) {
2370                 size_t len = newlen;
2371                 if (len > table->maxlen)
2372                         len = table->maxlen;
2373                 if(copy_from_user(table->data, newval, len))
2374                         return -EFAULT;
2375                 if (len == table->maxlen)
2376                         len--;
2377                 ((char *) table->data)[len] = 0;
2378         }
2379         return 1;
2380 }
2381
2382 /*
2383  * This function makes sure that all of the integers in the vector
2384  * are between the minimum and maximum values given in the arrays
2385  * table->extra1 and table->extra2, respectively.
2386  */
2387 int sysctl_intvec(ctl_table *table, int __user *name, int nlen,
2388                 void __user *oldval, size_t __user *oldlenp,
2389                 void __user *newval, size_t newlen)
2390 {
2391
2392         if (newval && newlen) {
2393                 int __user *vec = (int __user *) newval;
2394                 int *min = (int *) table->extra1;
2395                 int *max = (int *) table->extra2;
2396                 size_t length;
2397                 int i;
2398
2399                 if (newlen % sizeof(int) != 0)
2400                         return -EINVAL;
2401
2402                 if (!table->extra1 && !table->extra2)
2403                         return 0;
2404
2405                 if (newlen > table->maxlen)
2406                         newlen = table->maxlen;
2407                 length = newlen / sizeof(int);
2408
2409                 for (i = 0; i < length; i++) {
2410                         int value;
2411                         if (get_user(value, vec + i))
2412                                 return -EFAULT;
2413                         if (min && value < min[i])
2414                                 return -EINVAL;
2415                         if (max && value > max[i])
2416                                 return -EINVAL;
2417                 }
2418         }
2419         return 0;
2420 }
2421
2422 /* Strategy function to convert jiffies to seconds */
2423 int sysctl_jiffies(ctl_table *table, int __user *name, int nlen,
2424                 void __user *oldval, size_t __user *oldlenp,
2425                 void __user *newval, size_t newlen)
2426 {
2427         if (oldval && oldlenp) {
2428                 size_t olen;
2429
2430                 if (get_user(olen, oldlenp))
2431                         return -EFAULT;
2432                 if (olen) {
2433                         int val;
2434
2435                         if (olen < sizeof(int))
2436                                 return -EINVAL;
2437
2438                         val = *(int *)(table->data) / HZ;
2439                         if (put_user(val, (int __user *)oldval))
2440                                 return -EFAULT;
2441                         if (put_user(sizeof(int), oldlenp))
2442                                 return -EFAULT;
2443                 }
2444         }
2445         if (newval && newlen) {
2446                 int new;
2447                 if (newlen != sizeof(int))
2448                         return -EINVAL;
2449                 if (get_user(new, (int __user *)newval))
2450                         return -EFAULT;
2451                 *(int *)(table->data) = new*HZ;
2452         }
2453         return 1;
2454 }
2455
2456 /* Strategy function to convert jiffies to seconds */
2457 int sysctl_ms_jiffies(ctl_table *table, int __user *name, int nlen,
2458                 void __user *oldval, size_t __user *oldlenp,
2459                 void __user *newval, size_t newlen)
2460 {
2461         if (oldval && oldlenp) {
2462                 size_t olen;
2463
2464                 if (get_user(olen, oldlenp))
2465                         return -EFAULT;
2466                 if (olen) {
2467                         int val;
2468
2469                         if (olen < sizeof(int))
2470                                 return -EINVAL;
2471
2472                         val = jiffies_to_msecs(*(int *)(table->data));
2473                         if (put_user(val, (int __user *)oldval))
2474                                 return -EFAULT;
2475                         if (put_user(sizeof(int), oldlenp))
2476                                 return -EFAULT;
2477                 }
2478         }
2479         if (newval && newlen) {
2480                 int new;
2481                 if (newlen != sizeof(int))
2482                         return -EINVAL;
2483                 if (get_user(new, (int __user *)newval))
2484                         return -EFAULT;
2485                 *(int *)(table->data) = msecs_to_jiffies(new);
2486         }
2487         return 1;
2488 }
2489
2490
2491
2492 #else /* CONFIG_SYSCTL_SYSCALL */
2493
2494
2495 asmlinkage long sys_sysctl(struct __sysctl_args __user *args)
2496 {
2497         static int msg_count;
2498         struct __sysctl_args tmp;
2499         int name[CTL_MAXNAME];
2500         int i;
2501
2502         /* Read in the sysctl name for better debug message logging */
2503         if (copy_from_user(&tmp, args, sizeof(tmp)))
2504                 return -EFAULT;
2505         if (tmp.nlen <= 0 || tmp.nlen >= CTL_MAXNAME)
2506                 return -ENOTDIR;
2507         for (i = 0; i < tmp.nlen; i++)
2508                 if (get_user(name[i], tmp.name + i))
2509                         return -EFAULT;
2510
2511         /* Ignore accesses to kernel.version */
2512         if ((tmp.nlen == 2) && (name[0] == CTL_KERN) && (name[1] == KERN_VERSION))
2513                 goto out;
2514
2515         if (msg_count < 5) {
2516                 msg_count++;
2517                 printk(KERN_INFO
2518                         "warning: process `%s' used the removed sysctl "
2519                         "system call with ", current->comm);
2520                 for (i = 0; i < tmp.nlen; i++)
2521                         printk("%d.", name[i]);
2522                 printk("\n");
2523         }
2524 out:
2525         return -ENOSYS;
2526 }
2527
2528 int sysctl_string(ctl_table *table, int __user *name, int nlen,
2529                   void __user *oldval, size_t __user *oldlenp,
2530                   void __user *newval, size_t newlen)
2531 {
2532         return -ENOSYS;
2533 }
2534
2535 int sysctl_intvec(ctl_table *table, int __user *name, int nlen,
2536                 void __user *oldval, size_t __user *oldlenp,
2537                 void __user *newval, size_t newlen)
2538 {
2539         return -ENOSYS;
2540 }
2541
2542 int sysctl_jiffies(ctl_table *table, int __user *name, int nlen,
2543                 void __user *oldval, size_t __user *oldlenp,
2544                 void __user *newval, size_t newlen)
2545 {
2546         return -ENOSYS;
2547 }
2548
2549 int sysctl_ms_jiffies(ctl_table *table, int __user *name, int nlen,
2550                 void __user *oldval, size_t __user *oldlenp,
2551                 void __user *newval, size_t newlen)
2552 {
2553         return -ENOSYS;
2554 }
2555
2556 #endif /* CONFIG_SYSCTL_SYSCALL */
2557
2558 /*
2559  * No sense putting this after each symbol definition, twice,
2560  * exception granted :-)
2561  */
2562 EXPORT_SYMBOL(proc_dointvec);
2563 EXPORT_SYMBOL(proc_dointvec_jiffies);
2564 EXPORT_SYMBOL(proc_dointvec_minmax);
2565 EXPORT_SYMBOL(proc_dointvec_userhz_jiffies);
2566 EXPORT_SYMBOL(proc_dointvec_ms_jiffies);
2567 EXPORT_SYMBOL(proc_dostring);
2568 EXPORT_SYMBOL(proc_doulongvec_minmax);
2569 EXPORT_SYMBOL(proc_doulongvec_ms_jiffies_minmax);
2570 EXPORT_SYMBOL(register_sysctl_table);
2571 EXPORT_SYMBOL(sysctl_intvec);
2572 EXPORT_SYMBOL(sysctl_jiffies);
2573 EXPORT_SYMBOL(sysctl_ms_jiffies);
2574 EXPORT_SYMBOL(sysctl_string);
2575 EXPORT_SYMBOL(unregister_sysctl_table);