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