kernel/sysctl.c

   1 /*
   2  * sysctl.c: General linux system control interface
   3  *
   4  * Begun 24 March 1995, Stephen Tweedie
   5  * Added /proc support, Dec 1995
   6  * Added bdflush entry and intvec min/max checking, 2/23/96, Tom Dyas.
   7  * Added hooks for /proc/sys/net (minor, minor patch), 96/4/1, Mike Shaver.
   8  * Added kernel/java-{interpreter,appletviewer}, 96/5/10, Mike Shaver.
   9  * Dynamic registration fixes, Stephen Tweedie.
  10  * Added kswapd-interval, ctrl-alt-del, printk stuff, 1/8/97, Chris Horn.
  11  * Made sysctl support optional via CONFIG_SYSCTL, 1/10/97, Chris
  12  *  Horn.
  13  * Added proc_doulongvec_ms_jiffies_minmax, 09/08/99, Carlos H. Bauer.
  14  * Added proc_doulongvec_minmax, 09/08/99, Carlos H. Bauer.
  15  * Changed linked lists to use list.h instead of lists.h, 02/24/00, Bill
  16  *  Wendling.
  17  * The list_for_each() macro wasn't appropriate for the sysctl loop.
  18  *  Removed it and replaced it with older style, 03/23/00, Bill Wendling
  19  */
  20
  21 #include <linux/module.h>
  22 #include <linux/mm.h>
  23 #include <linux/swap.h>
  24 #include <linux/slab.h>
  25 #include <linux/sysctl.h>
  26 #include <linux/proc_fs.h>
  27 #include <linux/capability.h>
  28 #include <linux/ctype.h>
  29 #include <linux/utsname.h>
  30 #include <linux/capability.h>
  31 #include <linux/smp_lock.h>
  32 #include <linux/init.h>
  33 #include <linux/kernel.h>
  34 #include <linux/kobject.h>
  35 #include <linux/net.h>
  36 #include <linux/sysrq.h>
  37 #include <linux/highuid.h>
  38 #include <linux/writeback.h>
  39 #include <linux/hugetlb.h>
  40 #include <linux/security.h>
  41 #include <linux/initrd.h>
  42 #include <linux/times.h>
  43 #include <linux/limits.h>
  44 #include <linux/dcache.h>
  45 #include <linux/syscalls.h>
  46 #include <linux/nfs_fs.h>
  47 #include <linux/acpi.h>
  48
  49 #include <asm/uaccess.h>
  50 #include <asm/processor.h>
  51
  52 extern int proc_nr_files(ctl_table *table, int write, struct file *filp,
  53                      void __user *buffer, size_t *lenp, loff_t *ppos);
  54
  55 #ifdef CONFIG_X86
  56 #include <asm/nmi.h>
  57 #include <asm/stacktrace.h>
  58 #endif
  59
  60 #if defined(CONFIG_SYSCTL)
  61
  62 /* External variables not in a header file. */
  63 extern int C_A_D;
  64 extern int sysctl_overcommit_memory;
  65 extern int sysctl_overcommit_ratio;
  66 extern int sysctl_panic_on_oom;
  67 extern int max_threads;
  68 extern int core_uses_pid;
  69 extern int suid_dumpable;
  70 extern char core_pattern[];
  71 extern int pid_max;
  72 extern int min_free_kbytes;
  73 extern int printk_ratelimit_jiffies;
  74 extern int printk_ratelimit_burst;
  75 extern int pid_max_min, pid_max_max;
  76 extern int sysctl_drop_caches;
  77 extern int percpu_pagelist_fraction;
  78 extern int compat_log;
  79
  80 /* this is needed for the proc_dointvec_minmax for [fs_]overflow UID and GID */
  81 static int maxolduid = 65535;
  82 static int minolduid;
  83 static int min_percpu_pagelist_fract = 8;
  84
  85 static int ngroups_max = NGROUPS_MAX;
  86
  87 #ifdef CONFIG_KMOD
  88 extern char modprobe_path[];
  89 #endif
  90 #ifdef CONFIG_CHR_DEV_SG
  91 extern int sg_big_buff;
  92 #endif
  93
  94 #ifdef __sparc__
  95 extern char reboot_command [];
  96 extern int stop_a_enabled;
  97 extern int scons_pwroff;
  98 #endif
  99
 100 #ifdef __hppa__
 101 extern int pwrsw_enabled;
 102 extern int unaligned_enabled;
 103 #endif
 104
 105 #ifdef CONFIG_S390
 106 #ifdef CONFIG_MATHEMU
 107 extern int sysctl_ieee_emulation_warnings;
 108 #endif
 109 extern int sysctl_userprocess_debug;
 110 extern int spin_retry;
 111 #endif
 112
 113 extern int sysctl_hz_timer;
 114
 115 #ifdef CONFIG_BSD_PROCESS_ACCT
 116 extern int acct_parm[];
 117 #endif
 118
 119 #ifdef CONFIG_IA64
 120 extern int no_unaligned_warning;
 121 #endif
 122
 123 #ifdef CONFIG_RT_MUTEXES
 124 extern int max_lock_depth;
 125 #endif
 126
 127 #ifdef CONFIG_SYSCTL_SYSCALL
 128 static int parse_table(int __user *, int, void __user *, size_t __user *,
 129                 void __user *, size_t, ctl_table *);
 130 #endif
 131
 132
 133 #ifdef CONFIG_PROC_SYSCTL
 134 static int proc_do_cad_pid(ctl_table *table, int write, struct file *filp,
 135                   void __user *buffer, size_t *lenp, loff_t *ppos);
 136 static int proc_dointvec_taint(ctl_table *table, int write, struct file *filp,
 137                                void __user *buffer, size_t *lenp, loff_t *ppos);
 138 #endif
 139
 140 static ctl_table root_table[];
 141 static struct ctl_table_header root_table_header =
 142         { root_table, LIST_HEAD_INIT(root_table_header.ctl_entry) };
 143
 144 static ctl_table kern_table[];
 145 static ctl_table vm_table[];
 146 static ctl_table fs_table[];
 147 static ctl_table debug_table[];
 148 static ctl_table dev_table[];
 149 extern ctl_table random_table[];
 150 #ifdef CONFIG_UNIX98_PTYS
 151 extern ctl_table pty_table[];
 152 #endif
 153 #ifdef CONFIG_INOTIFY_USER
 154 extern ctl_table inotify_table[];
 155 #endif
 156
 157 #ifdef HAVE_ARCH_PICK_MMAP_LAYOUT
 158 int sysctl_legacy_va_layout;
 159 #endif
 160
 161
 162 /* The default sysctl tables: */
 163
 164 static ctl_table root_table[] = {
 165         {
 166                 .ctl_name       = CTL_KERN,
 167                 .procname       = "kernel",
 168                 .mode           = 0555,
 169                 .child          = kern_table,
 170         },
 171         {
 172                 .ctl_name       = CTL_VM,
 173                 .procname       = "vm",
 174                 .mode           = 0555,
 175                 .child          = vm_table,
 176         },
 177 #ifdef CONFIG_NET
 178         {
 179                 .ctl_name       = CTL_NET,
 180                 .procname       = "net",
 181                 .mode           = 0555,
 182                 .child          = net_table,
 183         },
 184 #endif
 185         {
 186                 .ctl_name       = CTL_FS,
 187                 .procname       = "fs",
 188                 .mode           = 0555,
 189                 .child          = fs_table,
 190         },
 191         {
 192                 .ctl_name       = CTL_DEBUG,
 193                 .procname       = "debug",
 194                 .mode           = 0555,
 195                 .child          = debug_table,
 196         },
 197         {
 198                 .ctl_name       = CTL_DEV,
 199                 .procname       = "dev",
 200                 .mode           = 0555,
 201                 .child          = dev_table,
 202         },
 203
 204         { .ctl_name = 0 }
 205 };
 206
 207 static ctl_table kern_table[] = {
 208         {
 209                 .ctl_name       = KERN_PANIC,
 210                 .procname       = "panic",
 211                 .data           = &panic_timeout,
 212                 .maxlen         = sizeof(int),
 213                 .mode           = 0644,
 214                 .proc_handler   = &proc_dointvec,
 215         },
 216         {
 217                 .ctl_name       = KERN_CORE_USES_PID,
 218                 .procname       = "core_uses_pid",
 219                 .data           = &core_uses_pid,
 220                 .maxlen         = sizeof(int),
 221                 .mode           = 0644,
 222                 .proc_handler   = &proc_dointvec,
 223         },
 224         {
 225                 .ctl_name       = KERN_CORE_PATTERN,
 226                 .procname       = "core_pattern",
 227                 .data           = core_pattern,
 228                 .maxlen         = 128,
 229                 .mode           = 0644,
 230                 .proc_handler   = &proc_dostring,
 231                 .strategy       = &sysctl_string,
 232         },
 233 #ifdef CONFIG_PROC_SYSCTL
 234         {
 235                 .ctl_name       = KERN_TAINTED,
 236                 .procname       = "tainted",
 237                 .data           = &tainted,
 238                 .maxlen         = sizeof(int),
 239                 .mode           = 0644,
 240                 .proc_handler   = &proc_dointvec_taint,
 241         },
 242 #endif
 243         {
 244                 .ctl_name       = KERN_CAP_BSET,
 245                 .procname       = "cap-bound",
 246                 .data           = &cap_bset,
 247                 .maxlen         = sizeof(kernel_cap_t),
 248                 .mode           = 0600,
 249                 .proc_handler   = &proc_dointvec_bset,
 250         },
 251 #ifdef CONFIG_BLK_DEV_INITRD
 252         {
 253                 .ctl_name       = KERN_REALROOTDEV,
 254                 .procname       = "real-root-dev",
 255                 .data           = &real_root_dev,
 256                 .maxlen         = sizeof(int),
 257                 .mode           = 0644,
 258                 .proc_handler   = &proc_dointvec,
 259         },
 260 #endif
 261 #ifdef __sparc__
 262         {
 263                 .ctl_name       = KERN_SPARC_REBOOT,
 264                 .procname       = "reboot-cmd",
 265                 .data           = reboot_command,
 266                 .maxlen         = 256,
 267                 .mode           = 0644,
 268                 .proc_handler   = &proc_dostring,
 269                 .strategy       = &sysctl_string,
 270         },
 271         {
 272                 .ctl_name       = KERN_SPARC_STOP_A,
 273                 .procname       = "stop-a",
 274                 .data           = &stop_a_enabled,
 275                 .maxlen         = sizeof (int),
 276                 .mode           = 0644,
 277                 .proc_handler   = &proc_dointvec,
 278         },
 279         {
 280                 .ctl_name       = KERN_SPARC_SCONS_PWROFF,
 281                 .procname       = "scons-poweroff",
 282                 .data           = &scons_pwroff,
 283                 .maxlen         = sizeof (int),
 284                 .mode           = 0644,
 285                 .proc_handler   = &proc_dointvec,
 286         },
 287 #endif
 288 #ifdef __hppa__
 289         {
 290                 .ctl_name       = KERN_HPPA_PWRSW,
 291                 .procname       = "soft-power",
 292                 .data           = &pwrsw_enabled,
 293                 .maxlen         = sizeof (int),
 294                 .mode           = 0644,
 295                 .proc_handler   = &proc_dointvec,
 296         },
 297         {
 298                 .ctl_name       = KERN_HPPA_UNALIGNED,
 299                 .procname       = "unaligned-trap",
 300                 .data           = &unaligned_enabled,
 301                 .maxlen         = sizeof (int),
 302                 .mode           = 0644,
 303                 .proc_handler   = &proc_dointvec,
 304         },
 305 #endif
 306         {
 307                 .ctl_name       = KERN_CTLALTDEL,
 308                 .procname       = "ctrl-alt-del",
 309                 .data           = &C_A_D,
 310                 .maxlen         = sizeof(int),
 311                 .mode           = 0644,
 312                 .proc_handler   = &proc_dointvec,
 313         },
 314         {
 315                 .ctl_name       = KERN_PRINTK,
 316                 .procname       = "printk",
 317                 .data           = &console_loglevel,
 318                 .maxlen         = 4*sizeof(int),
 319                 .mode           = 0644,
 320                 .proc_handler   = &proc_dointvec,
 321         },
 322 #ifdef CONFIG_KMOD
 323         {
 324                 .ctl_name       = KERN_MODPROBE,
 325                 .procname       = "modprobe",
 326                 .data           = &modprobe_path,
 327                 .maxlen         = KMOD_PATH_LEN,
 328                 .mode           = 0644,
 329                 .proc_handler   = &proc_dostring,
 330                 .strategy       = &sysctl_string,
 331         },
 332 #endif
 333 #if defined(CONFIG_HOTPLUG) && defined(CONFIG_NET)
 334         {
 335                 .ctl_name       = KERN_HOTPLUG,
 336                 .procname       = "hotplug",
 337                 .data           = &uevent_helper,
 338                 .maxlen         = UEVENT_HELPER_PATH_LEN,
 339                 .mode           = 0644,
 340                 .proc_handler   = &proc_dostring,
 341                 .strategy       = &sysctl_string,
 342         },
 343 #endif
 344 #ifdef CONFIG_CHR_DEV_SG
 345         {
 346                 .ctl_name       = KERN_SG_BIG_BUFF,
 347                 .procname       = "sg-big-buff",
 348                 .data           = &sg_big_buff,
 349                 .maxlen         = sizeof (int),
 350                 .mode           = 0444,
 351                 .proc_handler   = &proc_dointvec,
 352         },
 353 #endif
 354 #ifdef CONFIG_BSD_PROCESS_ACCT
 355         {
 356                 .ctl_name       = KERN_ACCT,
 357                 .procname       = "acct",
 358                 .data           = &acct_parm,
 359                 .maxlen         = 3*sizeof(int),
 360                 .mode           = 0644,
 361                 .proc_handler   = &proc_dointvec,
 362         },
 363 #endif
 364 #ifdef CONFIG_MAGIC_SYSRQ
 365         {
 366                 .ctl_name       = KERN_SYSRQ,
 367                 .procname       = "sysrq",
 368                 .data           = &__sysrq_enabled,
 369                 .maxlen         = sizeof (int),
 370                 .mode           = 0644,
 371                 .proc_handler   = &proc_dointvec,
 372         },
 373 #endif
 374 #ifdef CONFIG_PROC_SYSCTL
 375         {
 376                 .ctl_name       = KERN_CADPID,
 377                 .procname       = "cad_pid",
 378                 .data           = NULL,
 379                 .maxlen         = sizeof (int),
 380                 .mode           = 0600,
 381                 .proc_handler   = &proc_do_cad_pid,
 382         },
 383 #endif
 384         {
 385                 .ctl_name       = KERN_MAX_THREADS,
 386                 .procname       = "threads-max",
 387                 .data           = &max_threads,
 388                 .maxlen         = sizeof(int),
 389                 .mode           = 0644,
 390                 .proc_handler   = &proc_dointvec,
 391         },
 392         {
 393                 .ctl_name       = KERN_RANDOM,
 394                 .procname       = "random",
 395                 .mode           = 0555,
 396                 .child          = random_table,
 397         },
 398 #ifdef CONFIG_UNIX98_PTYS
 399         {
 400                 .ctl_name       = KERN_PTY,
 401                 .procname       = "pty",
 402                 .mode           = 0555,
 403                 .child          = pty_table,
 404         },
 405 #endif
 406         {
 407                 .ctl_name       = KERN_OVERFLOWUID,
 408                 .procname       = "overflowuid",
 409                 .data           = &overflowuid,
 410                 .maxlen         = sizeof(int),
 411                 .mode           = 0644,
 412                 .proc_handler   = &proc_dointvec_minmax,
 413                 .strategy       = &sysctl_intvec,
 414                 .extra1         = &minolduid,
 415                 .extra2         = &maxolduid,
 416         },
 417         {
 418                 .ctl_name       = KERN_OVERFLOWGID,
 419                 .procname       = "overflowgid",
 420                 .data           = &overflowgid,
 421                 .maxlen         = sizeof(int),
 422                 .mode           = 0644,
 423                 .proc_handler   = &proc_dointvec_minmax,
 424                 .strategy       = &sysctl_intvec,
 425                 .extra1         = &minolduid,
 426                 .extra2         = &maxolduid,
 427         },
 428 #ifdef CONFIG_S390
 429 #ifdef CONFIG_MATHEMU
 430         {
 431                 .ctl_name       = KERN_IEEE_EMULATION_WARNINGS,
 432                 .procname       = "ieee_emulation_warnings",
 433                 .data           = &sysctl_ieee_emulation_warnings,
 434                 .maxlen         = sizeof(int),
 435                 .mode           = 0644,
 436                 .proc_handler   = &proc_dointvec,
 437         },
 438 #endif
 439 #ifdef CONFIG_NO_IDLE_HZ
 440         {
 441                 .ctl_name       = KERN_HZ_TIMER,
 442                 .procname       = "hz_timer",
 443                 .data           = &sysctl_hz_timer,
 444                 .maxlen         = sizeof(int),
 445                 .mode           = 0644,
 446                 .proc_handler   = &proc_dointvec,
 447         },
 448 #endif
 449         {
 450                 .ctl_name       = KERN_S390_USER_DEBUG_LOGGING,
 451                 .procname       = "userprocess_debug",
 452                 .data           = &sysctl_userprocess_debug,
 453                 .maxlen         = sizeof(int),
 454                 .mode           = 0644,
 455                 .proc_handler   = &proc_dointvec,
 456         },
 457 #endif
 458         {
 459                 .ctl_name       = KERN_PIDMAX,
 460                 .procname       = "pid_max",
 461                 .data           = &pid_max,
 462                 .maxlen         = sizeof (int),
 463                 .mode           = 0644,
 464                 .proc_handler   = &proc_dointvec_minmax,
 465                 .strategy       = sysctl_intvec,
 466                 .extra1         = &pid_max_min,
 467                 .extra2         = &pid_max_max,
 468         },
 469         {
 470                 .ctl_name       = KERN_PANIC_ON_OOPS,
 471                 .procname       = "panic_on_oops",
 472                 .data           = &panic_on_oops,
 473                 .maxlen         = sizeof(int),
 474                 .mode           = 0644,
 475                 .proc_handler   = &proc_dointvec,
 476         },
 477         {
 478                 .ctl_name       = KERN_PRINTK_RATELIMIT,
 479                 .procname       = "printk_ratelimit",
 480                 .data           = &printk_ratelimit_jiffies,
 481                 .maxlen         = sizeof(int),
 482                 .mode           = 0644,
 483                 .proc_handler   = &proc_dointvec_jiffies,
 484                 .strategy       = &sysctl_jiffies,
 485         },
 486         {
 487                 .ctl_name       = KERN_PRINTK_RATELIMIT_BURST,
 488                 .procname       = "printk_ratelimit_burst",
 489                 .data           = &printk_ratelimit_burst,
 490                 .maxlen         = sizeof(int),
 491                 .mode           = 0644,
 492                 .proc_handler   = &proc_dointvec,
 493         },
 494         {
 495                 .ctl_name       = KERN_NGROUPS_MAX,
 496                 .procname       = "ngroups_max",
 497                 .data           = &ngroups_max,
 498                 .maxlen         = sizeof (int),
 499                 .mode           = 0444,
 500                 .proc_handler   = &proc_dointvec,
 501         },
 502 #if defined(CONFIG_X86_LOCAL_APIC) && defined(CONFIG_X86)
 503         {
 504                 .ctl_name       = KERN_UNKNOWN_NMI_PANIC,
 505                 .procname       = "unknown_nmi_panic",
 506                 .data           = &unknown_nmi_panic,
 507                 .maxlen         = sizeof (int),
 508                 .mode           = 0644,
 509                 .proc_handler   = &proc_dointvec,
 510         },
 511         {
 512                 .ctl_name       = KERN_NMI_WATCHDOG,
 513                 .procname       = "nmi_watchdog",
 514                 .data           = &nmi_watchdog_enabled,
 515                 .maxlen         = sizeof (int),
 516                 .mode           = 0644,
 517                 .proc_handler   = &proc_nmi_enabled,
 518         },
 519 #endif
 520 #if defined(CONFIG_X86)
 521         {
 522                 .ctl_name       = KERN_PANIC_ON_NMI,
 523                 .procname       = "panic_on_unrecovered_nmi",
 524                 .data           = &panic_on_unrecovered_nmi,
 525                 .maxlen         = sizeof(int),
 526                 .mode           = 0644,
 527                 .proc_handler   = &proc_dointvec,
 528         },
 529         {
 530                 .ctl_name       = KERN_BOOTLOADER_TYPE,
 531                 .procname       = "bootloader_type",
 532                 .data           = &bootloader_type,
 533                 .maxlen         = sizeof (int),
 534                 .mode           = 0444,
 535                 .proc_handler   = &proc_dointvec,
 536         },
 537         {
 538                 .ctl_name       = CTL_UNNUMBERED,
 539                 .procname       = "kstack_depth_to_print",
 540                 .data           = &kstack_depth_to_print,
 541                 .maxlen         = sizeof(int),
 542                 .mode           = 0644,
 543                 .proc_handler   = &proc_dointvec,
 544         },
 545 #endif
 546 #if defined(CONFIG_MMU)
 547         {
 548                 .ctl_name       = KERN_RANDOMIZE,
 549                 .procname       = "randomize_va_space",
 550                 .data           = &randomize_va_space,
 551                 .maxlen         = sizeof(int),
 552                 .mode           = 0644,
 553                 .proc_handler   = &proc_dointvec,
 554         },
 555 #endif
 556 #if defined(CONFIG_S390) && defined(CONFIG_SMP)
 557         {
 558                 .ctl_name       = KERN_SPIN_RETRY,
 559                 .procname       = "spin_retry",
 560                 .data           = &spin_retry,
 561                 .maxlen         = sizeof (int),
 562                 .mode           = 0644,
 563                 .proc_handler   = &proc_dointvec,
 564         },
 565 #endif
 566 #ifdef CONFIG_ACPI_SLEEP
 567         {
 568                 .ctl_name       = KERN_ACPI_VIDEO_FLAGS,
 569                 .procname       = "acpi_video_flags",
 570                 .data           = &acpi_video_flags,
 571                 .maxlen         = sizeof (unsigned long),
 572                 .mode           = 0644,
 573                 .proc_handler   = &proc_doulongvec_minmax,
 574         },
 575 #endif
 576 #ifdef CONFIG_IA64
 577         {
 578                 .ctl_name       = KERN_IA64_UNALIGNED,
 579                 .procname       = "ignore-unaligned-usertrap",
 580                 .data           = &no_unaligned_warning,
 581                 .maxlen         = sizeof (int),
 582                 .mode           = 0644,
 583                 .proc_handler   = &proc_dointvec,
 584         },
 585 #endif
 586 #ifdef CONFIG_COMPAT
 587         {
 588                 .ctl_name       = KERN_COMPAT_LOG,
 589                 .procname       = "compat-log",
 590                 .data           = &compat_log,
 591                 .maxlen         = sizeof (int),
 592                 .mode           = 0644,
 593                 .proc_handler   = &proc_dointvec,
 594         },
 595 #endif
 596 #ifdef CONFIG_RT_MUTEXES
 597         {
 598                 .ctl_name       = KERN_MAX_LOCK_DEPTH,
 599                 .procname       = "max_lock_depth",
 600                 .data           = &max_lock_depth,
 601                 .maxlen         = sizeof(int),
 602                 .mode           = 0644,
 603                 .proc_handler   = &proc_dointvec,
 604         },
 605 #endif
 606
 607         { .ctl_name = 0 }
 608 };
 609
 610 /* Constants for minimum and maximum testing in vm_table.
 611    We use these as one-element integer vectors. */
 612 static int zero;
 613 static int one_hundred = 100;
 614
 615
 616 static ctl_table vm_table[] = {
 617         {
 618                 .ctl_name       = VM_OVERCOMMIT_MEMORY,
 619                 .procname       = "overcommit_memory",
 620                 .data           = &sysctl_overcommit_memory,
 621                 .maxlen         = sizeof(sysctl_overcommit_memory),
 622                 .mode           = 0644,
 623                 .proc_handler   = &proc_dointvec,
 624         },
 625         {
 626                 .ctl_name       = VM_PANIC_ON_OOM,
 627                 .procname       = "panic_on_oom",
 628                 .data           = &sysctl_panic_on_oom,
 629                 .maxlen         = sizeof(sysctl_panic_on_oom),
 630                 .mode           = 0644,
 631                 .proc_handler   = &proc_dointvec,
 632         },
 633         {
 634                 .ctl_name       = VM_OVERCOMMIT_RATIO,
 635                 .procname       = "overcommit_ratio",
 636                 .data           = &sysctl_overcommit_ratio,
 637                 .maxlen         = sizeof(sysctl_overcommit_ratio),
 638                 .mode           = 0644,
 639                 .proc_handler   = &proc_dointvec,
 640         },
 641         {
 642                 .ctl_name       = VM_PAGE_CLUSTER,
 643                 .procname       = "page-cluster",
 644                 .data           = &page_cluster,
 645                 .maxlen         = sizeof(int),
 646                 .mode           = 0644,
 647                 .proc_handler   = &proc_dointvec,
 648         },
 649         {
 650                 .ctl_name       = VM_DIRTY_BACKGROUND,
 651                 .procname       = "dirty_background_ratio",
 652                 .data           = &dirty_background_ratio,
 653                 .maxlen         = sizeof(dirty_background_ratio),
 654                 .mode           = 0644,
 655                 .proc_handler   = &proc_dointvec_minmax,
 656                 .strategy       = &sysctl_intvec,
 657                 .extra1         = &zero,
 658                 .extra2         = &one_hundred,
 659         },
 660         {
 661                 .ctl_name       = VM_DIRTY_RATIO,
 662                 .procname       = "dirty_ratio",
 663                 .data           = &vm_dirty_ratio,
 664                 .maxlen         = sizeof(vm_dirty_ratio),
 665                 .mode           = 0644,
 666                 .proc_handler   = &proc_dointvec_minmax,
 667                 .strategy       = &sysctl_intvec,
 668                 .extra1         = &zero,
 669                 .extra2         = &one_hundred,
 670         },
 671         {
 672                 .ctl_name       = VM_DIRTY_WB_CS,
 673                 .procname       = "dirty_writeback_centisecs",
 674                 .data           = &dirty_writeback_interval,
 675                 .maxlen         = sizeof(dirty_writeback_interval),
 676                 .mode           = 0644,
 677                 .proc_handler   = &dirty_writeback_centisecs_handler,
 678         },
 679         {
 680                 .ctl_name       = VM_DIRTY_EXPIRE_CS,
 681                 .procname       = "dirty_expire_centisecs",
 682                 .data           = &dirty_expire_interval,
 683                 .maxlen         = sizeof(dirty_expire_interval),
 684                 .mode           = 0644,
 685                 .proc_handler   = &proc_dointvec_userhz_jiffies,
 686         },
 687         {
 688                 .ctl_name       = VM_NR_PDFLUSH_THREADS,
 689                 .procname       = "nr_pdflush_threads",
 690                 .data           = &nr_pdflush_threads,
 691                 .maxlen         = sizeof nr_pdflush_threads,
 692                 .mode           = 0444 /* read-only*/,
 693                 .proc_handler   = &proc_dointvec,
 694         },
 695         {
 696                 .ctl_name       = VM_SWAPPINESS,
 697                 .procname       = "swappiness",
 698                 .data           = &vm_swappiness,
 699                 .maxlen         = sizeof(vm_swappiness),
 700                 .mode           = 0644,
 701                 .proc_handler   = &proc_dointvec_minmax,
 702                 .strategy       = &sysctl_intvec,
 703                 .extra1         = &zero,
 704                 .extra2         = &one_hundred,
 705         },
 706 #ifdef CONFIG_HUGETLB_PAGE
 707          {
 708                 .ctl_name       = VM_HUGETLB_PAGES,
 709                 .procname       = "nr_hugepages",
 710                 .data           = &max_huge_pages,
 711                 .maxlen         = sizeof(unsigned long),
 712                 .mode           = 0644,
 713                 .proc_handler   = &hugetlb_sysctl_handler,
 714                 .extra1         = (void *)&hugetlb_zero,
 715                 .extra2         = (void *)&hugetlb_infinity,
 716          },
 717          {
 718                 .ctl_name       = VM_HUGETLB_GROUP,
 719                 .procname       = "hugetlb_shm_group",
 720                 .data           = &sysctl_hugetlb_shm_group,
 721                 .maxlen         = sizeof(gid_t),
 722                 .mode           = 0644,
 723                 .proc_handler   = &proc_dointvec,
 724          },
 725 #endif
 726         {
 727                 .ctl_name       = VM_LOWMEM_RESERVE_RATIO,
 728                 .procname       = "lowmem_reserve_ratio",
 729                 .data           = &sysctl_lowmem_reserve_ratio,
 730                 .maxlen         = sizeof(sysctl_lowmem_reserve_ratio),
 731                 .mode           = 0644,
 732                 .proc_handler   = &lowmem_reserve_ratio_sysctl_handler,
 733                 .strategy       = &sysctl_intvec,
 734         },
 735         {
 736                 .ctl_name       = VM_DROP_PAGECACHE,
 737                 .procname       = "drop_caches",
 738                 .data           = &sysctl_drop_caches,
 739                 .maxlen         = sizeof(int),
 740                 .mode           = 0644,
 741                 .proc_handler   = drop_caches_sysctl_handler,
 742                 .strategy       = &sysctl_intvec,
 743         },
 744         {
 745                 .ctl_name       = VM_MIN_FREE_KBYTES,
 746                 .procname       = "min_free_kbytes",
 747                 .data           = &min_free_kbytes,
 748                 .maxlen         = sizeof(min_free_kbytes),
 749                 .mode           = 0644,
 750                 .proc_handler   = &min_free_kbytes_sysctl_handler,
 751                 .strategy       = &sysctl_intvec,
 752                 .extra1         = &zero,
 753         },
 754         {
 755                 .ctl_name       = VM_PERCPU_PAGELIST_FRACTION,
 756                 .procname       = "percpu_pagelist_fraction",
 757                 .data           = &percpu_pagelist_fraction,
 758                 .maxlen         = sizeof(percpu_pagelist_fraction),
 759                 .mode           = 0644,
 760                 .proc_handler   = &percpu_pagelist_fraction_sysctl_handler,
 761                 .strategy       = &sysctl_intvec,
 762                 .extra1         = &min_percpu_pagelist_fract,
 763         },
 764 #ifdef CONFIG_MMU
 765         {
 766                 .ctl_name       = VM_MAX_MAP_COUNT,
 767                 .procname       = "max_map_count",
 768                 .data           = &sysctl_max_map_count,
 769                 .maxlen         = sizeof(sysctl_max_map_count),
 770                 .mode           = 0644,
 771                 .proc_handler   = &proc_dointvec
 772         },
 773 #endif
 774         {
 775                 .ctl_name       = VM_LAPTOP_MODE,
 776                 .procname       = "laptop_mode",
 777                 .data           = &laptop_mode,
 778                 .maxlen         = sizeof(laptop_mode),
 779                 .mode           = 0644,
 780                 .proc_handler   = &proc_dointvec_jiffies,
 781                 .strategy       = &sysctl_jiffies,
 782         },
 783         {
 784                 .ctl_name       = VM_BLOCK_DUMP,
 785                 .procname       = "block_dump",
 786                 .data           = &block_dump,
 787                 .maxlen         = sizeof(block_dump),
 788                 .mode           = 0644,
 789                 .proc_handler   = &proc_dointvec,
 790                 .strategy       = &sysctl_intvec,
 791                 .extra1         = &zero,
 792         },
 793         {
 794                 .ctl_name       = VM_VFS_CACHE_PRESSURE,
 795                 .procname       = "vfs_cache_pressure",
 796                 .data           = &sysctl_vfs_cache_pressure,
 797                 .maxlen         = sizeof(sysctl_vfs_cache_pressure),
 798                 .mode           = 0644,
 799                 .proc_handler   = &proc_dointvec,
 800                 .strategy       = &sysctl_intvec,
 801                 .extra1         = &zero,
 802         },
 803 #ifdef HAVE_ARCH_PICK_MMAP_LAYOUT
 804         {
 805                 .ctl_name       = VM_LEGACY_VA_LAYOUT,
 806                 .procname       = "legacy_va_layout",
 807                 .data           = &sysctl_legacy_va_layout,
 808                 .maxlen         = sizeof(sysctl_legacy_va_layout),
 809                 .mode           = 0644,
 810                 .proc_handler   = &proc_dointvec,
 811                 .strategy       = &sysctl_intvec,
 812                 .extra1         = &zero,
 813         },
 814 #endif
 815 #ifdef CONFIG_NUMA
 816         {
 817                 .ctl_name       = VM_ZONE_RECLAIM_MODE,
 818                 .procname       = "zone_reclaim_mode",
 819                 .data           = &zone_reclaim_mode,
 820                 .maxlen         = sizeof(zone_reclaim_mode),
 821                 .mode           = 0644,
 822                 .proc_handler   = &proc_dointvec,
 823                 .strategy       = &sysctl_intvec,
 824                 .extra1         = &zero,
 825         },
 826         {
 827                 .ctl_name       = VM_MIN_UNMAPPED,
 828                 .procname       = "min_unmapped_ratio",
 829                 .data           = &sysctl_min_unmapped_ratio,
 830                 .maxlen         = sizeof(sysctl_min_unmapped_ratio),
 831                 .mode           = 0644,
 832                 .proc_handler   = &sysctl_min_unmapped_ratio_sysctl_handler,
 833                 .strategy       = &sysctl_intvec,
 834                 .extra1         = &zero,
 835                 .extra2         = &one_hundred,
 836         },
 837         {
 838                 .ctl_name       = VM_MIN_SLAB,
 839                 .procname       = "min_slab_ratio",
 840                 .data           = &sysctl_min_slab_ratio,
 841                 .maxlen         = sizeof(sysctl_min_slab_ratio),
 842                 .mode           = 0644,
 843                 .proc_handler   = &sysctl_min_slab_ratio_sysctl_handler,
 844                 .strategy       = &sysctl_intvec,
 845                 .extra1         = &zero,
 846                 .extra2         = &one_hundred,
 847         },
 848 #endif
 849 #ifdef CONFIG_X86_32
 850         {
 851                 .ctl_name       = VM_VDSO_ENABLED,
 852                 .procname       = "vdso_enabled",
 853                 .data           = &vdso_enabled,
 854                 .maxlen         = sizeof(vdso_enabled),
 855                 .mode           = 0644,
 856                 .proc_handler   = &proc_dointvec,
 857                 .strategy       = &sysctl_intvec,
 858                 .extra1         = &zero,
 859         },
 860 #endif
 861         { .ctl_name = 0 }
 862 };
 863
 864 #if defined(CONFIG_BINFMT_MISC) || defined(CONFIG_BINFMT_MISC_MODULE)
 865 static ctl_table binfmt_misc_table[] = {
 866         { .ctl_name = 0 }
 867 };
 868 #endif
 869
 870 static ctl_table fs_table[] = {
 871         {
 872                 .ctl_name       = FS_NRINODE,
 873                 .procname       = "inode-nr",
 874                 .data           = &inodes_stat,
 875                 .maxlen         = 2*sizeof(int),
 876                 .mode           = 0444,
 877                 .proc_handler   = &proc_dointvec,
 878         },
 879         {
 880                 .ctl_name       = FS_STATINODE,
 881                 .procname       = "inode-state",
 882                 .data           = &inodes_stat,
 883                 .maxlen         = 7*sizeof(int),
 884                 .mode           = 0444,
 885                 .proc_handler   = &proc_dointvec,
 886         },
 887         {
 888                 .ctl_name       = FS_NRFILE,
 889                 .procname       = "file-nr",
 890                 .data           = &files_stat,
 891                 .maxlen         = 3*sizeof(int),
 892                 .mode           = 0444,
 893                 .proc_handler   = &proc_nr_files,
 894         },
 895         {
 896                 .ctl_name       = FS_MAXFILE,
 897                 .procname       = "file-max",
 898                 .data           = &files_stat.max_files,
 899                 .maxlen         = sizeof(int),
 900                 .mode           = 0644,
 901                 .proc_handler   = &proc_dointvec,
 902         },
 903         {
 904                 .ctl_name       = FS_DENTRY,
 905                 .procname       = "dentry-state",
 906                 .data           = &dentry_stat,
 907                 .maxlen         = 6*sizeof(int),
 908                 .mode           = 0444,
 909                 .proc_handler   = &proc_dointvec,
 910         },
 911         {
 912                 .ctl_name       = FS_OVERFLOWUID,
 913                 .procname       = "overflowuid",
 914                 .data           = &fs_overflowuid,
 915                 .maxlen         = sizeof(int),
 916                 .mode           = 0644,
 917                 .proc_handler   = &proc_dointvec_minmax,
 918                 .strategy       = &sysctl_intvec,
 919                 .extra1         = &minolduid,
 920                 .extra2         = &maxolduid,
 921         },
 922         {
 923                 .ctl_name       = FS_OVERFLOWGID,
 924                 .procname       = "overflowgid",
 925                 .data           = &fs_overflowgid,
 926                 .maxlen         = sizeof(int),
 927                 .mode           = 0644,
 928                 .proc_handler   = &proc_dointvec_minmax,
 929                 .strategy       = &sysctl_intvec,
 930                 .extra1         = &minolduid,
 931                 .extra2         = &maxolduid,
 932         },
 933         {
 934                 .ctl_name       = FS_LEASES,
 935                 .procname       = "leases-enable",
 936                 .data           = &leases_enable,
 937                 .maxlen         = sizeof(int),
 938                 .mode           = 0644,
 939                 .proc_handler   = &proc_dointvec,
 940         },
 941 #ifdef CONFIG_DNOTIFY
 942         {
 943                 .ctl_name       = FS_DIR_NOTIFY,
 944                 .procname       = "dir-notify-enable",
 945                 .data           = &dir_notify_enable,
 946                 .maxlen         = sizeof(int),
 947                 .mode           = 0644,
 948                 .proc_handler   = &proc_dointvec,
 949         },
 950 #endif
 951 #ifdef CONFIG_MMU
 952         {
 953                 .ctl_name       = FS_LEASE_TIME,
 954                 .procname       = "lease-break-time",
 955                 .data           = &lease_break_time,
 956                 .maxlen         = sizeof(int),
 957                 .mode           = 0644,
 958                 .proc_handler   = &proc_dointvec,
 959         },
 960         {
 961                 .ctl_name       = FS_AIO_NR,
 962                 .procname       = "aio-nr",
 963                 .data           = &aio_nr,
 964                 .maxlen         = sizeof(aio_nr),
 965                 .mode           = 0444,
 966                 .proc_handler   = &proc_doulongvec_minmax,
 967         },
 968         {
 969                 .ctl_name       = FS_AIO_MAX_NR,
 970                 .procname       = "aio-max-nr",
 971                 .data           = &aio_max_nr,
 972                 .maxlen         = sizeof(aio_max_nr),
 973                 .mode           = 0644,
 974                 .proc_handler   = &proc_doulongvec_minmax,
 975         },
 976 #ifdef CONFIG_INOTIFY_USER
 977         {
 978                 .ctl_name       = FS_INOTIFY,
 979                 .procname       = "inotify",
 980                 .mode           = 0555,
 981                 .child          = inotify_table,
 982         },
 983 #endif
 984 #endif
 985         {
 986                 .ctl_name       = KERN_SETUID_DUMPABLE,
 987                 .procname       = "suid_dumpable",
 988                 .data           = &suid_dumpable,
 989                 .maxlen         = sizeof(int),
 990                 .mode           = 0644,
 991                 .proc_handler   = &proc_dointvec,
 992         },
 993 #if defined(CONFIG_BINFMT_MISC) || defined(CONFIG_BINFMT_MISC_MODULE)
 994         {
 995                 .ctl_name       = CTL_UNNUMBERED,
 996                 .procname       = "binfmt_misc",
 997                 .mode           = 0555,
 998                 .child          = binfmt_misc_table,
 999         },
1000 #endif
1001         { .ctl_name = 0 }
1002 };
1003
1004 static ctl_table debug_table[] = {
1005         { .ctl_name = 0 }
1006 };
1007
1008 static ctl_table dev_table[] = {
1009         { .ctl_name = 0 }
1010 };
1011
1012 static DEFINE_SPINLOCK(sysctl_lock);
1013
1014 /* called under sysctl_lock */
1015 static int use_table(struct ctl_table_header *p)
1016 {
1017         if (unlikely(p->unregistering))
1018                 return 0;
1019         p->used++;
1020         return 1;
1021 }
1022
1023 /* called under sysctl_lock */
1024 static void unuse_table(struct ctl_table_header *p)
1025 {
1026         if (!--p->used)
1027                 if (unlikely(p->unregistering))
1028                         complete(p->unregistering);
1029 }
1030
1031 /* called under sysctl_lock, will reacquire if has to wait */
1032 static void start_unregistering(struct ctl_table_header *p)
1033 {
1034         /*
1035          * if p->used is 0, nobody will ever touch that entry again;
1036          * we'll eliminate all paths to it before dropping sysctl_lock
1037          */
1038         if (unlikely(p->used)) {
1039                 struct completion wait;
1040                 init_completion(&wait);
1041                 p->unregistering = &wait;
1042                 spin_unlock(&sysctl_lock);
1043                 wait_for_completion(&wait);
1044                 spin_lock(&sysctl_lock);
1045         }
1046         /*
1047          * do not remove from the list until nobody holds it; walking the
1048          * list in do_sysctl() relies on that.
1049          */
1050         list_del_init(&p->ctl_entry);
1051 }
1052
1053 void sysctl_head_finish(struct ctl_table_header *head)
1054 {
1055         if (!head)
1056                 return;
1057         spin_lock(&sysctl_lock);
1058         unuse_table(head);
1059         spin_unlock(&sysctl_lock);
1060 }
1061
1062 struct ctl_table_header *sysctl_head_next(struct ctl_table_header *prev)
1063 {
1064         struct ctl_table_header *head;
1065         struct list_head *tmp;
1066         spin_lock(&sysctl_lock);
1067         if (prev) {
1068                 tmp = &prev->ctl_entry;
1069                 unuse_table(prev);
1070                 goto next;
1071         }
1072         tmp = &root_table_header.ctl_entry;
1073         for (;;) {
1074                 head = list_entry(tmp, struct ctl_table_header, ctl_entry);
1075
1076                 if (!use_table(head))
1077                         goto next;
1078                 spin_unlock(&sysctl_lock);
1079                 return head;
1080         next:
1081                 tmp = tmp->next;
1082                 if (tmp == &root_table_header.ctl_entry)
1083                         break;
1084         }
1085         spin_unlock(&sysctl_lock);
1086         return NULL;
1087 }
1088
1089 #ifdef CONFIG_SYSCTL_SYSCALL
1090 int do_sysctl(int __user *name, int nlen, void __user *oldval, size_t __user *oldlenp,
1091                void __user *newval, size_t newlen)
1092 {
1093         struct ctl_table_header *head;
1094         int error = -ENOTDIR;
1095
1096         if (nlen <= 0 || nlen >= CTL_MAXNAME)
1097                 return -ENOTDIR;
1098         if (oldval) {
1099                 int old_len;
1100                 if (!oldlenp || get_user(old_len, oldlenp))
1101                         return -EFAULT;
1102         }
1103
1104         for (head = sysctl_head_next(NULL); head;
1105                         head = sysctl_head_next(head)) {
1106                 error = parse_table(name, nlen, oldval, oldlenp,
1107                                         newval, newlen, head->ctl_table);
1108                 if (error != -ENOTDIR) {
1109                         sysctl_head_finish(head);
1110                         break;
1111                 }
1112         }
1113         return error;
1114 }
1115
1116 asmlinkage long sys_sysctl(struct __sysctl_args __user *args)
1117 {
1118         struct __sysctl_args tmp;
1119         int error;
1120
1121         if (copy_from_user(&tmp, args, sizeof(tmp)))
1122                 return -EFAULT;
1123
1124         lock_kernel();
1125         error = do_sysctl(tmp.name, tmp.nlen, tmp.oldval, tmp.oldlenp,
1126                           tmp.newval, tmp.newlen);
1127         unlock_kernel();
1128         return error;
1129 }
1130 #endif /* CONFIG_SYSCTL_SYSCALL */
1131
1132 /*
1133  * sysctl_perm does NOT grant the superuser all rights automatically, because
1134  * some sysctl variables are readonly even to root.
1135  */
1136
1137 static int test_perm(int mode, int op)
1138 {
1139         if (!current->euid)
1140                 mode >>= 6;
1141         else if (in_egroup_p(0))
1142                 mode >>= 3;
1143         if ((mode & op & 0007) == op)
1144                 return 0;
1145         return -EACCES;
1146 }
1147
1148 int sysctl_perm(ctl_table *table, int op)
1149 {
1150         int error;
1151         error = security_sysctl(table, op);
1152         if (error)
1153                 return error;
1154         return test_perm(table->mode, op);
1155 }
1156
1157 #ifdef CONFIG_SYSCTL_SYSCALL
1158 static int parse_table(int __user *name, int nlen,
1159                        void __user *oldval, size_t __user *oldlenp,
1160                        void __user *newval, size_t newlen,
1161                        ctl_table *table)
1162 {
1163         int n;
1164 repeat:
1165         if (!nlen)
1166                 return -ENOTDIR;
1167         if (get_user(n, name))
1168                 return -EFAULT;
1169         for ( ; table->ctl_name || table->procname; table++) {
1170                 if (!table->ctl_name)
1171                         continue;
1172                 if (n == table->ctl_name) {
1173                         int error;
1174                         if (table->child) {
1175                                 if (sysctl_perm(table, 001))
1176                                         return -EPERM;
1177                                 name++;
1178                                 nlen--;
1179                                 table = table->child;
1180                                 goto repeat;
1181                         }
1182                         error = do_sysctl_strategy(table, name, nlen,
1183                                                    oldval, oldlenp,
1184                                                    newval, newlen);
1185                         return error;
1186                 }
1187         }
1188         return -ENOTDIR;
1189 }
1190
1191 /* Perform the actual read/write of a sysctl table entry. */
1192 int do_sysctl_strategy (ctl_table *table,
1193                         int __user *name, int nlen,
1194                         void __user *oldval, size_t __user *oldlenp,
1195                         void __user *newval, size_t newlen)
1196 {
1197         int op = 0, rc;
1198         size_t len;
1199
1200         if (oldval)
1201                 op |= 004;
1202         if (newval)
1203                 op |= 002;
1204         if (sysctl_perm(table, op))
1205                 return -EPERM;
1206
1207         if (table->strategy) {
1208                 rc = table->strategy(table, name, nlen, oldval, oldlenp,
1209                                      newval, newlen);
1210                 if (rc < 0)
1211                         return rc;
1212                 if (rc > 0)
1213                         return 0;
1214         }
1215
1216         /* If there is no strategy routine, or if the strategy returns
1217          * zero, proceed with automatic r/w */
1218         if (table->data && table->maxlen) {
1219                 if (oldval && oldlenp) {
1220                         if (get_user(len, oldlenp))
1221                                 return -EFAULT;
1222                         if (len) {
1223                                 if (len > table->maxlen)
1224                                         len = table->maxlen;
1225                                 if(copy_to_user(oldval, table->data, len))
1226                                         return -EFAULT;
1227                                 if(put_user(len, oldlenp))
1228                                         return -EFAULT;
1229                         }
1230                 }
1231                 if (newval && newlen) {
1232                         len = newlen;
1233                         if (len > table->maxlen)
1234                                 len = table->maxlen;
1235                         if(copy_from_user(table->data, newval, len))
1236                                 return -EFAULT;
1237                 }
1238         }
1239         return 0;
1240 }
1241 #endif /* CONFIG_SYSCTL_SYSCALL */
1242
1243 /**
1244  * register_sysctl_table - register a sysctl hierarchy
1245  * @table: the top-level table structure
1246  *
1247  * Register a sysctl table hierarchy. @table should be a filled in ctl_table
1248  * array. An entry with a ctl_name of 0 terminates the table.
1249  *
1250  * The members of the &ctl_table structure are used as follows:
1251  *
1252  * ctl_name - This is the numeric sysctl value used by sysctl(2). The number
1253  *            must be unique within that level of sysctl
1254  *
1255  * procname - the name of the sysctl file under /proc/sys. Set to %NULL to not
1256  *            enter a sysctl file
1257  *
1258  * data - a pointer to data for use by proc_handler
1259  *
1260  * maxlen - the maximum size in bytes of the data
1261  *
1262  * mode - the file permissions for the /proc/sys file, and for sysctl(2)
1263  *
1264  * child - a pointer to the child sysctl table if this entry is a directory, or
1265  *         %NULL.
1266  *
1267  * proc_handler - the text handler routine (described below)
1268  *
1269  * strategy - the strategy routine (described below)
1270  *
1271  * de - for internal use by the sysctl routines
1272  *
1273  * extra1, extra2 - extra pointers usable by the proc handler routines
1274  *
1275  * Leaf nodes in the sysctl tree will be represented by a single file
1276  * under /proc; non-leaf nodes will be represented by directories.
1277  *
1278  * sysctl(2) can automatically manage read and write requests through
1279  * the sysctl table.  The data and maxlen fields of the ctl_table
1280  * struct enable minimal validation of the values being written to be
1281  * performed, and the mode field allows minimal authentication.
1282  *
1283  * More sophisticated management can be enabled by the provision of a
1284  * strategy routine with the table entry.  This will be called before
1285  * any automatic read or write of the data is performed.
1286  *
1287  * The strategy routine may return
1288  *
1289  * < 0 - Error occurred (error is passed to user process)
1290  *
1291  * 0   - OK - proceed with automatic read or write.
1292  *
1293  * > 0 - OK - read or write has been done by the strategy routine, so
1294  *       return immediately.
1295  *
1296  * There must be a proc_handler routine for any terminal nodes
1297  * mirrored under /proc/sys (non-terminals are handled by a built-in
1298  * directory handler).  Several default handlers are available to
1299  * cover common cases -
1300  *
1301  * proc_dostring(), proc_dointvec(), proc_dointvec_jiffies(),
1302  * proc_dointvec_userhz_jiffies(), proc_dointvec_minmax(),
1303  * proc_doulongvec_ms_jiffies_minmax(), proc_doulongvec_minmax()
1304  *
1305  * It is the handler's job to read the input buffer from user memory
1306  * and process it. The handler should return 0 on success.
1307  *
1308  * This routine returns %NULL on a failure to register, and a pointer
1309  * to the table header on success.
1310  */
1311 struct ctl_table_header *register_sysctl_table(ctl_table * table)
1312 {
1313         struct ctl_table_header *tmp;
1314         tmp = kmalloc(sizeof(struct ctl_table_header), GFP_KERNEL);
1315         if (!tmp)
1316                 return NULL;
1317         tmp->ctl_table = table;
1318         INIT_LIST_HEAD(&tmp->ctl_entry);
1319         tmp->used = 0;
1320         tmp->unregistering = NULL;
1321         spin_lock(&sysctl_lock);
1322         list_add_tail(&tmp->ctl_entry, &root_table_header.ctl_entry);
1323         spin_unlock(&sysctl_lock);
1324         return tmp;
1325 }
1326
1327 /**
1328  * unregister_sysctl_table - unregister a sysctl table hierarchy
1329  * @header: the header returned from register_sysctl_table
1330  *
1331  * Unregisters the sysctl table and all children. proc entries may not
1332  * actually be removed until they are no longer used by anyone.
1333  */
1334 void unregister_sysctl_table(struct ctl_table_header * header)
1335 {
1336         might_sleep();
1337         spin_lock(&sysctl_lock);
1338         start_unregistering(header);
1339         spin_unlock(&sysctl_lock);
1340         kfree(header);
1341 }
1342
1343 #else /* !CONFIG_SYSCTL */
1344 struct ctl_table_header * register_sysctl_table(ctl_table * table,
1345                                                 int insert_at_head)
1346 {
1347         return NULL;
1348 }
1349
1350 void unregister_sysctl_table(struct ctl_table_header * table)
1351 {
1352 }
1353
1354 #endif /* CONFIG_SYSCTL */
1355
1356 /*
1357  * /proc/sys support
1358  */
1359
1360 #ifdef CONFIG_PROC_SYSCTL
1361
1362 static int _proc_do_string(void* data, int maxlen, int write,
1363                            struct file *filp, void __user *buffer,
1364                            size_t *lenp, loff_t *ppos)
1365 {
1366         size_t len;
1367         char __user *p;
1368         char c;
1369
1370         if (!data || !maxlen || !*lenp) {
1371                 *lenp = 0;
1372                 return 0;
1373         }
1374
1375         if (write) {
1376                 len = 0;
1377                 p = buffer;
1378                 while (len < *lenp) {
1379                         if (get_user(c, p++))
1380                                 return -EFAULT;
1381                         if (c == 0 || c == '\n')
1382                                 break;
1383                         len++;
1384                 }
1385                 if (len >= maxlen)
1386                         len = maxlen-1;
1387                 if(copy_from_user(data, buffer, len))
1388                         return -EFAULT;
1389                 ((char *) data)[len] = 0;
1390                 *ppos += *lenp;
1391         } else {
1392                 len = strlen(data);
1393                 if (len > maxlen)
1394                         len = maxlen;
1395
1396                 if (*ppos > len) {
1397                         *lenp = 0;
1398                         return 0;
1399                 }
1400
1401                 data += *ppos;
1402                 len  -= *ppos;
1403
1404                 if (len > *lenp)
1405                         len = *lenp;
1406                 if (len)
1407                         if(copy_to_user(buffer, data, len))
1408                                 return -EFAULT;
1409                 if (len < *lenp) {
1410                         if(put_user('\n', ((char __user *) buffer) + len))
1411                                 return -EFAULT;
1412                         len++;
1413                 }
1414                 *lenp = len;
1415                 *ppos += len;
1416         }
1417         return 0;
1418 }
1419
1420 /**
1421  * proc_dostring - read a string sysctl
1422  * @table: the sysctl table
1423  * @write: %TRUE if this is a write to the sysctl file
1424  * @filp: the file structure
1425  * @buffer: the user buffer
1426  * @lenp: the size of the user buffer
1427  * @ppos: file position
1428  *
1429  * Reads/writes a string from/to the user buffer. If the kernel
1430  * buffer provided is not large enough to hold the string, the
1431  * string is truncated. The copied string is %NULL-terminated.
1432  * If the string is being read by the user process, it is copied
1433  * and a newline '\n' is added. It is truncated if the buffer is
1434  * not large enough.
1435  *
1436  * Returns 0 on success.
1437  */
1438 int proc_dostring(ctl_table *table, int write, struct file *filp,
1439                   void __user *buffer, size_t *lenp, loff_t *ppos)
1440 {
1441         return _proc_do_string(table->data, table->maxlen, write, filp,
1442                                buffer, lenp, ppos);
1443 }
1444
1445
1446 static int do_proc_dointvec_conv(int *negp, unsigned long *lvalp,
1447                                  int *valp,
1448                                  int write, void *data)
1449 {
1450         if (write) {
1451                 *valp = *negp ? -*lvalp : *lvalp;
1452         } else {
1453                 int val = *valp;
1454                 if (val < 0) {
1455                         *negp = -1;
1456                         *lvalp = (unsigned long)-val;
1457                 } else {
1458                         *negp = 0;
1459                         *lvalp = (unsigned long)val;
1460                 }
1461         }
1462         return 0;
1463 }
1464
1465 static int __do_proc_dointvec(void *tbl_data, ctl_table *table,
1466                   int write, struct file *filp, void __user *buffer,
1467                   size_t *lenp, loff_t *ppos,
1468                   int (*conv)(int *negp, unsigned long *lvalp, int *valp,
1469                               int write, void *data),
1470                   void *data)
1471 {
1472 #define TMPBUFLEN 21
1473         int *i, vleft, first=1, neg, val;
1474         unsigned long lval;
1475         size_t left, len;
1476
1477         char buf[TMPBUFLEN], *p;
1478         char __user *s = buffer;
1479
1480         if (!tbl_data || !table->maxlen || !*lenp ||
1481             (*ppos && !write)) {
1482                 *lenp = 0;
1483                 return 0;
1484         }
1485
1486         i = (int *) tbl_data;
1487         vleft = table->maxlen / sizeof(*i);
1488         left = *lenp;
1489
1490         if (!conv)
1491                 conv = do_proc_dointvec_conv;
1492
1493         for (; left && vleft--; i++, first=0) {
1494                 if (write) {
1495                         while (left) {
1496                                 char c;
1497                                 if (get_user(c, s))
1498                                         return -EFAULT;
1499                                 if (!isspace(c))
1500                                         break;
1501                                 left--;
1502                                 s++;
1503                         }
1504                         if (!left)
1505                                 break;
1506                         neg = 0;
1507                         len = left;
1508                         if (len > sizeof(buf) - 1)
1509                                 len = sizeof(buf) - 1;
1510                         if (copy_from_user(buf, s, len))
1511                                 return -EFAULT;
1512                         buf[len] = 0;
1513                         p = buf;
1514                         if (*p == '-' && left > 1) {
1515                                 neg = 1;
1516                                 p++;
1517                         }
1518                         if (*p < '0' || *p > '9')
1519                                 break;
1520
1521                         lval = simple_strtoul(p, &p, 0);
1522
1523                         len = p-buf;
1524                         if ((len < left) && *p && !isspace(*p))
1525                                 break;
1526                         if (neg)
1527                                 val = -val;
1528                         s += len;
1529                         left -= len;
1530
1531                         if (conv(&neg, &lval, i, 1, data))
1532                                 break;
1533                 } else {
1534                         p = buf;
1535                         if (!first)
1536                                 *p++ = '\t';
1537
1538                         if (conv(&neg, &lval, i, 0, data))
1539                                 break;
1540
1541                         sprintf(p, "%s%lu", neg ? "-" : "", lval);
1542                         len = strlen(buf);
1543                         if (len > left)
1544                                 len = left;
1545                         if(copy_to_user(s, buf, len))
1546                                 return -EFAULT;
1547                         left -= len;
1548                         s += len;
1549                 }
1550         }
1551
1552         if (!write && !first && left) {
1553                 if(put_user('\n', s))
1554                         return -EFAULT;
1555                 left--, s++;
1556         }
1557         if (write) {
1558                 while (left) {
1559                         char c;
1560                         if (get_user(c, s++))
1561                                 return -EFAULT;
1562                         if (!isspace(c))
1563                                 break;
1564                         left--;
1565                 }
1566         }
1567         if (write && first)
1568                 return -EINVAL;
1569         *lenp -= left;
1570         *ppos += *lenp;
1571         return 0;
1572 #undef TMPBUFLEN
1573 }
1574
1575 static int do_proc_dointvec(ctl_table *table, int write, struct file *filp,
1576                   void __user *buffer, size_t *lenp, loff_t *ppos,
1577                   int (*conv)(int *negp, unsigned long *lvalp, int *valp,
1578                               int write, void *data),
1579                   void *data)
1580 {
1581         return __do_proc_dointvec(table->data, table, write, filp,
1582                         buffer, lenp, ppos, conv, data);
1583 }
1584
1585 /**
1586  * proc_dointvec - read a vector of integers
1587  * @table: the sysctl table
1588  * @write: %TRUE if this is a write to the sysctl file
1589  * @filp: the file structure
1590  * @buffer: the user buffer
1591  * @lenp: the size of the user buffer
1592  * @ppos: file position
1593  *
1594  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
1595  * values from/to the user buffer, treated as an ASCII string.
1596  *
1597  * Returns 0 on success.
1598  */
1599 int proc_dointvec(ctl_table *table, int write, struct file *filp,
1600                      void __user *buffer, size_t *lenp, loff_t *ppos)
1601 {
1602     return do_proc_dointvec(table,write,filp,buffer,lenp,ppos,
1603                             NULL,NULL);
1604 }
1605
1606 #define OP_SET  0
1607 #define OP_AND  1
1608 #define OP_OR   2
1609
1610 static int do_proc_dointvec_bset_conv(int *negp, unsigned long *lvalp,
1611                                       int *valp,
1612                                       int write, void *data)
1613 {
1614         int op = *(int *)data;
1615         if (write) {
1616                 int val = *negp ? -*lvalp : *lvalp;
1617                 switch(op) {
1618                 case OP_SET:    *valp = val; break;
1619                 case OP_AND:    *valp &= val; break;
1620                 case OP_OR:     *valp |= val; break;
1621                 }
1622         } else {
1623                 int val = *valp;
1624                 if (val < 0) {
1625                         *negp = -1;
1626                         *lvalp = (unsigned long)-val;
1627                 } else {
1628                         *negp = 0;
1629                         *lvalp = (unsigned long)val;
1630                 }
1631         }
1632         return 0;
1633 }
1634
1635 /*
1636  *      init may raise the set.
1637  */
1638
1639 int proc_dointvec_bset(ctl_table *table, int write, struct file *filp,
1640                         void __user *buffer, size_t *lenp, loff_t *ppos)
1641 {
1642         int op;
1643
1644         if (write && !capable(CAP_SYS_MODULE)) {
1645                 return -EPERM;
1646         }
1647
1648         op = is_init(current) ? OP_SET : OP_AND;
1649         return do_proc_dointvec(table,write,filp,buffer,lenp,ppos,
1650                                 do_proc_dointvec_bset_conv,&op);
1651 }
1652
1653 /*
1654  *      Taint values can only be increased
1655  */
1656 static int proc_dointvec_taint(ctl_table *table, int write, struct file *filp,
1657                                void __user *buffer, size_t *lenp, loff_t *ppos)
1658 {
1659         int op;
1660
1661         if (!capable(CAP_SYS_ADMIN))
1662                 return -EPERM;
1663
1664         op = OP_OR;
1665         return do_proc_dointvec(table,write,filp,buffer,lenp,ppos,
1666                                 do_proc_dointvec_bset_conv,&op);
1667 }
1668
1669 struct do_proc_dointvec_minmax_conv_param {
1670         int *min;
1671         int *max;
1672 };
1673
1674 static int do_proc_dointvec_minmax_conv(int *negp, unsigned long *lvalp,
1675                                         int *valp,
1676                                         int write, void *data)
1677 {
1678         struct do_proc_dointvec_minmax_conv_param *param = data;
1679         if (write) {
1680                 int val = *negp ? -*lvalp : *lvalp;
1681                 if ((param->min && *param->min > val) ||
1682                     (param->max && *param->max < val))
1683                         return -EINVAL;
1684                 *valp = val;
1685         } else {
1686                 int val = *valp;
1687                 if (val < 0) {
1688                         *negp = -1;
1689                         *lvalp = (unsigned long)-val;
1690                 } else {
1691                         *negp = 0;
1692                         *lvalp = (unsigned long)val;
1693                 }
1694         }
1695         return 0;
1696 }
1697
1698 /**
1699  * proc_dointvec_minmax - read a vector of integers with min/max values
1700  * @table: the sysctl table
1701  * @write: %TRUE if this is a write to the sysctl file
1702  * @filp: the file structure
1703  * @buffer: the user buffer
1704  * @lenp: the size of the user buffer
1705  * @ppos: file position
1706  *
1707  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
1708  * values from/to the user buffer, treated as an ASCII string.
1709  *
1710  * This routine will ensure the values are within the range specified by
1711  * table->extra1 (min) and table->extra2 (max).
1712  *
1713  * Returns 0 on success.
1714  */
1715 int proc_dointvec_minmax(ctl_table *table, int write, struct file *filp,
1716                   void __user *buffer, size_t *lenp, loff_t *ppos)
1717 {
1718         struct do_proc_dointvec_minmax_conv_param param = {
1719                 .min = (int *) table->extra1,
1720                 .max = (int *) table->extra2,
1721         };
1722         return do_proc_dointvec(table, write, filp, buffer, lenp, ppos,
1723                                 do_proc_dointvec_minmax_conv, &param);
1724 }
1725
1726 static int __do_proc_doulongvec_minmax(void *data, ctl_table *table, int write,
1727                                      struct file *filp,
1728                                      void __user *buffer,
1729                                      size_t *lenp, loff_t *ppos,
1730                                      unsigned long convmul,
1731                                      unsigned long convdiv)
1732 {
1733 #define TMPBUFLEN 21
1734         unsigned long *i, *min, *max, val;
1735         int vleft, first=1, neg;
1736         size_t len, left;
1737         char buf[TMPBUFLEN], *p;
1738         char __user *s = buffer;
1739
1740         if (!data || !table->maxlen || !*lenp ||
1741             (*ppos && !write)) {
1742                 *lenp = 0;
1743                 return 0;
1744         }
1745
1746         i = (unsigned long *) data;
1747         min = (unsigned long *) table->extra1;
1748         max = (unsigned long *) table->extra2;
1749         vleft = table->maxlen / sizeof(unsigned long);
1750         left = *lenp;
1751
1752         for (; left && vleft--; i++, min++, max++, first=0) {
1753                 if (write) {
1754                         while (left) {
1755                                 char c;
1756                                 if (get_user(c, s))
1757                                         return -EFAULT;
1758                                 if (!isspace(c))
1759                                         break;
1760                                 left--;
1761                                 s++;
1762                         }
1763                         if (!left)
1764                                 break;
1765                         neg = 0;
1766                         len = left;
1767                         if (len > TMPBUFLEN-1)
1768                                 len = TMPBUFLEN-1;
1769                         if (copy_from_user(buf, s, len))
1770                                 return -EFAULT;
1771                         buf[len] = 0;
1772                         p = buf;
1773                         if (*p == '-' && left > 1) {
1774                                 neg = 1;
1775                                 p++;
1776                         }
1777                         if (*p < '0' || *p > '9')
1778                                 break;
1779                         val = simple_strtoul(p, &p, 0) * convmul / convdiv ;
1780                         len = p-buf;
1781                         if ((len < left) && *p && !isspace(*p))
1782                                 break;
1783                         if (neg)
1784                                 val = -val;
1785                         s += len;
1786                         left -= len;
1787
1788                         if(neg)
1789                                 continue;
1790                         if ((min && val < *min) || (max && val > *max))
1791                                 continue;
1792                         *i = val;
1793                 } else {
1794                         p = buf;
1795                         if (!first)
1796                                 *p++ = '\t';
1797                         sprintf(p, "%lu", convdiv * (*i) / convmul);
1798                         len = strlen(buf);
1799                         if (len > left)
1800                                 len = left;
1801                         if(copy_to_user(s, buf, len))
1802                                 return -EFAULT;
1803                         left -= len;
1804                         s += len;
1805                 }
1806         }
1807
1808         if (!write && !first && left) {
1809                 if(put_user('\n', s))
1810                         return -EFAULT;
1811                 left--, s++;
1812         }
1813         if (write) {
1814                 while (left) {
1815                         char c;
1816                         if (get_user(c, s++))
1817                                 return -EFAULT;
1818                         if (!isspace(c))
1819                                 break;
1820                         left--;
1821                 }
1822         }
1823         if (write && first)
1824                 return -EINVAL;
1825         *lenp -= left;
1826         *ppos += *lenp;
1827         return 0;
1828 #undef TMPBUFLEN
1829 }
1830
1831 static int do_proc_doulongvec_minmax(ctl_table *table, int write,
1832                                      struct file *filp,
1833                                      void __user *buffer,
1834                                      size_t *lenp, loff_t *ppos,
1835                                      unsigned long convmul,
1836                                      unsigned long convdiv)
1837 {
1838         return __do_proc_doulongvec_minmax(table->data, table, write,
1839                         filp, buffer, lenp, ppos, convmul, convdiv);
1840 }
1841
1842 /**
1843  * proc_doulongvec_minmax - read a vector of long integers with min/max values
1844  * @table: the sysctl table
1845  * @write: %TRUE if this is a write to the sysctl file
1846  * @filp: the file structure
1847  * @buffer: the user buffer
1848  * @lenp: the size of the user buffer
1849  * @ppos: file position
1850  *
1851  * Reads/writes up to table->maxlen/sizeof(unsigned long) unsigned long
1852  * values from/to the user buffer, treated as an ASCII string.
1853  *
1854  * This routine will ensure the values are within the range specified by
1855  * table->extra1 (min) and table->extra2 (max).
1856  *
1857  * Returns 0 on success.
1858  */
1859 int proc_doulongvec_minmax(ctl_table *table, int write, struct file *filp,
1860                            void __user *buffer, size_t *lenp, loff_t *ppos)
1861 {
1862     return do_proc_doulongvec_minmax(table, write, filp, buffer, lenp, ppos, 1l, 1l);
1863 }
1864
1865 /**
1866  * proc_doulongvec_ms_jiffies_minmax - read a vector of millisecond values with min/max values
1867  * @table: the sysctl table
1868  * @write: %TRUE if this is a write to the sysctl file
1869  * @filp: the file structure
1870  * @buffer: the user buffer
1871  * @lenp: the size of the user buffer
1872  * @ppos: file position
1873  *
1874  * Reads/writes up to table->maxlen/sizeof(unsigned long) unsigned long
1875  * values from/to the user buffer, treated as an ASCII string. The values
1876  * are treated as milliseconds, and converted to jiffies when they are stored.
1877  *
1878  * This routine will ensure the values are within the range specified by
1879  * table->extra1 (min) and table->extra2 (max).
1880  *
1881  * Returns 0 on success.
1882  */
1883 int proc_doulongvec_ms_jiffies_minmax(ctl_table *table, int write,
1884                                       struct file *filp,
1885                                       void __user *buffer,
1886                                       size_t *lenp, loff_t *ppos)
1887 {
1888     return do_proc_doulongvec_minmax(table, write, filp, buffer,
1889                                      lenp, ppos, HZ, 1000l);
1890 }
1891
1892
1893 static int do_proc_dointvec_jiffies_conv(int *negp, unsigned long *lvalp,
1894                                          int *valp,
1895                                          int write, void *data)
1896 {
1897         if (write) {
1898                 if (*lvalp > LONG_MAX / HZ)
1899                         return 1;
1900                 *valp = *negp ? -(*lvalp*HZ) : (*lvalp*HZ);
1901         } else {
1902                 int val = *valp;
1903                 unsigned long lval;
1904                 if (val < 0) {
1905                         *negp = -1;
1906                         lval = (unsigned long)-val;
1907                 } else {
1908                         *negp = 0;
1909                         lval = (unsigned long)val;
1910                 }
1911                 *lvalp = lval / HZ;
1912         }
1913         return 0;
1914 }
1915
1916 static int do_proc_dointvec_userhz_jiffies_conv(int *negp, unsigned long *lvalp,
1917                                                 int *valp,
1918                                                 int write, void *data)
1919 {
1920         if (write) {
1921                 if (USER_HZ < HZ && *lvalp > (LONG_MAX / HZ) * USER_HZ)
1922                         return 1;
1923                 *valp = clock_t_to_jiffies(*negp ? -*lvalp : *lvalp);
1924         } else {
1925                 int val = *valp;
1926                 unsigned long lval;
1927                 if (val < 0) {
1928                         *negp = -1;
1929                         lval = (unsigned long)-val;
1930                 } else {
1931                         *negp = 0;
1932                         lval = (unsigned long)val;
1933                 }
1934                 *lvalp = jiffies_to_clock_t(lval);
1935         }
1936         return 0;
1937 }
1938
1939 static int do_proc_dointvec_ms_jiffies_conv(int *negp, unsigned long *lvalp,
1940                                             int *valp,
1941                                             int write, void *data)
1942 {
1943         if (write) {
1944                 *valp = msecs_to_jiffies(*negp ? -*lvalp : *lvalp);
1945         } else {
1946                 int val = *valp;
1947                 unsigned long lval;
1948                 if (val < 0) {
1949                         *negp = -1;
1950                         lval = (unsigned long)-val;
1951                 } else {
1952                         *negp = 0;
1953                         lval = (unsigned long)val;
1954                 }
1955                 *lvalp = jiffies_to_msecs(lval);
1956         }
1957         return 0;
1958 }
1959
1960 /**
1961  * proc_dointvec_jiffies - read a vector of integers as seconds
1962  * @table: the sysctl table
1963  * @write: %TRUE if this is a write to the sysctl file
1964  * @filp: the file structure
1965  * @buffer: the user buffer
1966  * @lenp: the size of the user buffer
1967  * @ppos: file position
1968  *
1969  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
1970  * values from/to the user buffer, treated as an ASCII string.
1971  * The values read are assumed to be in seconds, and are converted into
1972  * jiffies.
1973  *
1974  * Returns 0 on success.
1975  */
1976 int proc_dointvec_jiffies(ctl_table *table, int write, struct file *filp,
1977                           void __user *buffer, size_t *lenp, loff_t *ppos)
1978 {
1979     return do_proc_dointvec(table,write,filp,buffer,lenp,ppos,
1980                             do_proc_dointvec_jiffies_conv,NULL);
1981 }
1982
1983 /**
1984  * proc_dointvec_userhz_jiffies - read a vector of integers as 1/USER_HZ seconds
1985  * @table: the sysctl table
1986  * @write: %TRUE if this is a write to the sysctl file
1987  * @filp: the file structure
1988  * @buffer: the user buffer
1989  * @lenp: the size of the user buffer
1990  * @ppos: pointer to the file position
1991  *
1992  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
1993  * values from/to the user buffer, treated as an ASCII string.
1994  * The values read are assumed to be in 1/USER_HZ seconds, and
1995  * are converted into jiffies.
1996  *
1997  * Returns 0 on success.
1998  */
1999 int proc_dointvec_userhz_jiffies(ctl_table *table, int write, struct file *filp,
2000                                  void __user *buffer, size_t *lenp, loff_t *ppos)
2001 {
2002     return do_proc_dointvec(table,write,filp,buffer,lenp,ppos,
2003                             do_proc_dointvec_userhz_jiffies_conv,NULL);
2004 }
2005
2006 /**
2007  * proc_dointvec_ms_jiffies - read a vector of integers as 1 milliseconds
2008  * @table: the sysctl table
2009  * @write: %TRUE if this is a write to the sysctl file
2010  * @filp: the file structure
2011  * @buffer: the user buffer
2012  * @lenp: the size of the user buffer
2013  * @ppos: file position
2014  * @ppos: the current position in the file
2015  *
2016  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2017  * values from/to the user buffer, treated as an ASCII string.
2018  * The values read are assumed to be in 1/1000 seconds, and
2019  * are converted into jiffies.
2020  *
2021  * Returns 0 on success.
2022  */
2023 int proc_dointvec_ms_jiffies(ctl_table *table, int write, struct file *filp,
2024                              void __user *buffer, size_t *lenp, loff_t *ppos)
2025 {
2026         return do_proc_dointvec(table, write, filp, buffer, lenp, ppos,
2027                                 do_proc_dointvec_ms_jiffies_conv, NULL);
2028 }
2029
2030 static int proc_do_cad_pid(ctl_table *table, int write, struct file *filp,
2031                            void __user *buffer, size_t *lenp, loff_t *ppos)
2032 {
2033         struct pid *new_pid;
2034         pid_t tmp;
2035         int r;
2036
2037         tmp = pid_nr(cad_pid);
2038
2039         r = __do_proc_dointvec(&tmp, table, write, filp, buffer,
2040                                lenp, ppos, NULL, NULL);
2041         if (r || !write)
2042                 return r;
2043
2044         new_pid = find_get_pid(tmp);
2045         if (!new_pid)
2046                 return -ESRCH;
2047
2048         put_pid(xchg(&cad_pid, new_pid));
2049         return 0;
2050 }
2051
2052 #else /* CONFIG_PROC_FS */
2053
2054 int proc_dostring(ctl_table *table, int write, struct file *filp,
2055                   void __user *buffer, size_t *lenp, loff_t *ppos)
2056 {
2057         return -ENOSYS;
2058 }
2059
2060 int proc_dointvec(ctl_table *table, int write, struct file *filp,
2061                   void __user *buffer, size_t *lenp, loff_t *ppos)
2062 {
2063         return -ENOSYS;
2064 }
2065
2066 int proc_dointvec_bset(ctl_table *table, int write, struct file *filp,
2067                         void __user *buffer, size_t *lenp, loff_t *ppos)
2068 {
2069         return -ENOSYS;
2070 }
2071
2072 int proc_dointvec_minmax(ctl_table *table, int write, struct file *filp,
2073                     void __user *buffer, size_t *lenp, loff_t *ppos)
2074 {
2075         return -ENOSYS;
2076 }
2077
2078 int proc_dointvec_jiffies(ctl_table *table, int write, struct file *filp,
2079                     void __user *buffer, size_t *lenp, loff_t *ppos)
2080 {
2081         return -ENOSYS;
2082 }
2083
2084 int proc_dointvec_userhz_jiffies(ctl_table *table, int write, struct file *filp,
2085                     void __user *buffer, size_t *lenp, loff_t *ppos)
2086 {
2087         return -ENOSYS;
2088 }
2089
2090 int proc_dointvec_ms_jiffies(ctl_table *table, int write, struct file *filp,
2091                              void __user *buffer, size_t *lenp, loff_t *ppos)
2092 {
2093         return -ENOSYS;
2094 }
2095
2096 int proc_doulongvec_minmax(ctl_table *table, int write, struct file *filp,
2097                     void __user *buffer, size_t *lenp, loff_t *ppos)
2098 {
2099         return -ENOSYS;
2100 }
2101
2102 int proc_doulongvec_ms_jiffies_minmax(ctl_table *table, int write,
2103                                       struct file *filp,
2104                                       void __user *buffer,
2105                                       size_t *lenp, loff_t *ppos)
2106 {
2107     return -ENOSYS;
2108 }
2109
2110
2111 #endif /* CONFIG_PROC_FS */
2112
2113
2114 #ifdef CONFIG_SYSCTL_SYSCALL
2115 /*
2116  * General sysctl support routines
2117  */
2118
2119 /* The generic string strategy routine: */
2120 int sysctl_string(ctl_table *table, int __user *name, int nlen,
2121                   void __user *oldval, size_t __user *oldlenp,
2122                   void __user *newval, size_t newlen)
2123 {
2124         if (!table->data || !table->maxlen)
2125                 return -ENOTDIR;
2126
2127         if (oldval && oldlenp) {
2128                 size_t bufsize;
2129                 if (get_user(bufsize, oldlenp))
2130                         return -EFAULT;
2131                 if (bufsize) {
2132                         size_t len = strlen(table->data), copied;
2133
2134                         /* This shouldn't trigger for a well-formed sysctl */
2135                         if (len > table->maxlen)
2136                                 len = table->maxlen;
2137
2138                         /* Copy up to a max of bufsize-1 bytes of the string */
2139                         copied = (len >= bufsize) ? bufsize - 1 : len;
2140
2141                         if (copy_to_user(oldval, table->data, copied) ||
2142                             put_user(0, (char __user *)(oldval + copied)))
2143                                 return -EFAULT;
2144                         if (put_user(len, oldlenp))
2145                                 return -EFAULT;
2146                 }
2147         }
2148         if (newval && newlen) {
2149                 size_t len = newlen;
2150                 if (len > table->maxlen)
2151                         len = table->maxlen;
2152                 if(copy_from_user(table->data, newval, len))
2153                         return -EFAULT;
2154                 if (len == table->maxlen)
2155                         len--;
2156                 ((char *) table->data)[len] = 0;
2157         }
2158         return 1;
2159 }
2160
2161 /*
2162  * This function makes sure that all of the integers in the vector
2163  * are between the minimum and maximum values given in the arrays
2164  * table->extra1 and table->extra2, respectively.
2165  */
2166 int sysctl_intvec(ctl_table *table, int __user *name, int nlen,
2167                 void __user *oldval, size_t __user *oldlenp,
2168                 void __user *newval, size_t newlen)
2169 {
2170
2171         if (newval && newlen) {
2172                 int __user *vec = (int __user *) newval;
2173                 int *min = (int *) table->extra1;
2174                 int *max = (int *) table->extra2;
2175                 size_t length;
2176                 int i;
2177
2178                 if (newlen % sizeof(int) != 0)
2179                         return -EINVAL;
2180
2181                 if (!table->extra1 && !table->extra2)
2182                         return 0;
2183
2184                 if (newlen > table->maxlen)
2185                         newlen = table->maxlen;
2186                 length = newlen / sizeof(int);
2187
2188                 for (i = 0; i < length; i++) {
2189                         int value;
2190                         if (get_user(value, vec + i))
2191                                 return -EFAULT;
2192                         if (min && value < min[i])
2193                                 return -EINVAL;
2194                         if (max && value > max[i])
2195                                 return -EINVAL;
2196                 }
2197         }
2198         return 0;
2199 }
2200
2201 /* Strategy function to convert jiffies to seconds */
2202 int sysctl_jiffies(ctl_table *table, int __user *name, int nlen,
2203                 void __user *oldval, size_t __user *oldlenp,
2204                 void __user *newval, size_t newlen)
2205 {
2206         if (oldval && oldlenp) {
2207                 size_t olen;
2208
2209                 if (get_user(olen, oldlenp))
2210                         return -EFAULT;
2211                 if (olen) {
2212                         int val;
2213
2214                         if (olen < sizeof(int))
2215                                 return -EINVAL;
2216
2217                         val = *(int *)(table->data) / HZ;
2218                         if (put_user(val, (int __user *)oldval))
2219                                 return -EFAULT;
2220                         if (put_user(sizeof(int), oldlenp))
2221                                 return -EFAULT;
2222                 }
2223         }
2224         if (newval && newlen) {
2225                 int new;
2226                 if (newlen != sizeof(int))
2227                         return -EINVAL;
2228                 if (get_user(new, (int __user *)newval))
2229                         return -EFAULT;
2230                 *(int *)(table->data) = new*HZ;
2231         }
2232         return 1;
2233 }
2234
2235 /* Strategy function to convert jiffies to seconds */
2236 int sysctl_ms_jiffies(ctl_table *table, int __user *name, int nlen,
2237                 void __user *oldval, size_t __user *oldlenp,
2238                 void __user *newval, size_t newlen)
2239 {
2240         if (oldval && oldlenp) {
2241                 size_t olen;
2242
2243                 if (get_user(olen, oldlenp))
2244                         return -EFAULT;
2245                 if (olen) {
2246                         int val;
2247
2248                         if (olen < sizeof(int))
2249                                 return -EINVAL;
2250
2251                         val = jiffies_to_msecs(*(int *)(table->data));
2252                         if (put_user(val, (int __user *)oldval))
2253                                 return -EFAULT;
2254                         if (put_user(sizeof(int), oldlenp))
2255                                 return -EFAULT;
2256                 }
2257         }
2258         if (newval && newlen) {
2259                 int new;
2260                 if (newlen != sizeof(int))
2261                         return -EINVAL;
2262                 if (get_user(new, (int __user *)newval))
2263                         return -EFAULT;
2264                 *(int *)(table->data) = msecs_to_jiffies(new);
2265         }
2266         return 1;
2267 }
2268
2269
2270
2271 #else /* CONFIG_SYSCTL_SYSCALL */
2272
2273
2274 asmlinkage long sys_sysctl(struct __sysctl_args __user *args)
2275 {
2276         static int msg_count;
2277         struct __sysctl_args tmp;
2278         int name[CTL_MAXNAME];
2279         int i;
2280
2281         /* Read in the sysctl name for better debug message logging */
2282         if (copy_from_user(&tmp, args, sizeof(tmp)))
2283                 return -EFAULT;
2284         if (tmp.nlen <= 0 || tmp.nlen >= CTL_MAXNAME)
2285                 return -ENOTDIR;
2286         for (i = 0; i < tmp.nlen; i++)
2287                 if (get_user(name[i], tmp.name + i))
2288                         return -EFAULT;
2289
2290         /* Ignore accesses to kernel.version */
2291         if ((tmp.nlen == 2) && (name[0] == CTL_KERN) && (name[1] == KERN_VERSION))
2292                 goto out;
2293
2294         if (msg_count < 5) {
2295                 msg_count++;
2296                 printk(KERN_INFO
2297                         "warning: process `%s' used the removed sysctl "
2298                         "system call with ", current->comm);
2299                 for (i = 0; i < tmp.nlen; i++)
2300                         printk("%d.", name[i]);
2301                 printk("\n");
2302         }
2303 out:
2304         return -ENOSYS;
2305 }
2306
2307 int sysctl_string(ctl_table *table, int __user *name, int nlen,
2308                   void __user *oldval, size_t __user *oldlenp,
2309                   void __user *newval, size_t newlen)
2310 {
2311         return -ENOSYS;
2312 }
2313
2314 int sysctl_intvec(ctl_table *table, int __user *name, int nlen,
2315                 void __user *oldval, size_t __user *oldlenp,
2316                 void __user *newval, size_t newlen)
2317 {
2318         return -ENOSYS;
2319 }
2320
2321 int sysctl_jiffies(ctl_table *table, int __user *name, int nlen,
2322                 void __user *oldval, size_t __user *oldlenp,
2323                 void __user *newval, size_t newlen)
2324 {
2325         return -ENOSYS;
2326 }
2327
2328 int sysctl_ms_jiffies(ctl_table *table, int __user *name, int nlen,
2329                 void __user *oldval, size_t __user *oldlenp,
2330                 void __user *newval, size_t newlen)
2331 {
2332         return -ENOSYS;
2333 }
2334
2335 #endif /* CONFIG_SYSCTL_SYSCALL */
2336
2337 /*
2338  * No sense putting this after each symbol definition, twice,
2339  * exception granted :-)
2340  */
2341 EXPORT_SYMBOL(proc_dointvec);
2342 EXPORT_SYMBOL(proc_dointvec_jiffies);
2343 EXPORT_SYMBOL(proc_dointvec_minmax);
2344 EXPORT_SYMBOL(proc_dointvec_userhz_jiffies);
2345 EXPORT_SYMBOL(proc_dointvec_ms_jiffies);
2346 EXPORT_SYMBOL(proc_dostring);
2347 EXPORT_SYMBOL(proc_doulongvec_minmax);
2348 EXPORT_SYMBOL(proc_doulongvec_ms_jiffies_minmax);
2349 EXPORT_SYMBOL(register_sysctl_table);
2350 EXPORT_SYMBOL(sysctl_intvec);
2351 EXPORT_SYMBOL(sysctl_jiffies);
2352 EXPORT_SYMBOL(sysctl_ms_jiffies);
2353 EXPORT_SYMBOL(sysctl_string);
2354 EXPORT_SYMBOL(unregister_sysctl_table);