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