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/config.h>
  22 #include <linux/module.h>
  23 #include <linux/mm.h>
  24 #include <linux/swap.h>
  25 #include <linux/slab.h>
  26 #include <linux/sysctl.h>
  27 #include <linux/proc_fs.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/sysrq.h>
  35 #include <linux/highuid.h>
  36 #include <linux/writeback.h>
  37 #include <linux/hugetlb.h>
  38 #include <linux/security.h>
  39 #include <linux/initrd.h>
  40 #include <linux/times.h>
  41 #include <linux/limits.h>
  42 #include <linux/dcache.h>
  43 #include <linux/syscalls.h>
  44
  45 #include <asm/uaccess.h>
  46 #include <asm/processor.h>
  47
  48 #ifdef CONFIG_ROOT_NFS
  49 #include <linux/nfs_fs.h>
  50 #endif
  51
  52 #if defined(CONFIG_SYSCTL)
  53
  54 /* External variables not in a header file. */
  55 extern int C_A_D;
  56 extern int sysctl_overcommit_memory;
  57 extern int sysctl_overcommit_ratio;
  58 extern int max_threads;
  59 extern int sysrq_enabled;
  60 extern int core_uses_pid;
  61 extern int suid_dumpable;
  62 extern char core_pattern[];
  63 extern int cad_pid;
  64 extern int pid_max;
  65 extern int min_free_kbytes;
  66 extern int printk_ratelimit_jiffies;
  67 extern int printk_ratelimit_burst;
  68 extern int pid_max_min, pid_max_max;
  69
  70 #if defined(CONFIG_X86_LOCAL_APIC) && defined(CONFIG_X86)
  71 int unknown_nmi_panic;
  72 extern int proc_unknown_nmi_panic(ctl_table *, int, struct file *,
  73                                   void __user *, size_t *, loff_t *);
  74 #endif
  75
  76 /* this is needed for the proc_dointvec_minmax for [fs_]overflow UID and GID */
  77 static int maxolduid = 65535;
  78 static int minolduid;
  79
  80 static int ngroups_max = NGROUPS_MAX;
  81
  82 #ifdef CONFIG_KMOD
  83 extern char modprobe_path[];
  84 #endif
  85 #ifdef CONFIG_HOTPLUG
  86 extern char hotplug_path[];
  87 #endif
  88 #ifdef CONFIG_CHR_DEV_SG
  89 extern int sg_big_buff;
  90 #endif
  91 #ifdef CONFIG_SYSVIPC
  92 extern size_t shm_ctlmax;
  93 extern size_t shm_ctlall;
  94 extern int shm_ctlmni;
  95 extern int msg_ctlmax;
  96 extern int msg_ctlmnb;
  97 extern int msg_ctlmni;
  98 extern int sem_ctls[];
  99 #endif
 100
 101 #ifdef __sparc__
 102 extern char reboot_command [];
 103 extern int stop_a_enabled;
 104 extern int scons_pwroff;
 105 #endif
 106
 107 #ifdef __hppa__
 108 extern int pwrsw_enabled;
 109 extern int unaligned_enabled;
 110 #endif
 111
 112 #ifdef CONFIG_ARCH_S390
 113 #ifdef CONFIG_MATHEMU
 114 extern int sysctl_ieee_emulation_warnings;
 115 #endif
 116 extern int sysctl_userprocess_debug;
 117 extern int spin_retry;
 118 #endif
 119
 120 extern int sysctl_hz_timer;
 121
 122 #ifdef CONFIG_BSD_PROCESS_ACCT
 123 extern int acct_parm[];
 124 #endif
 125
 126 int randomize_va_space = 1;
 127
 128 static int parse_table(int __user *, int, void __user *, size_t __user *, void __user *, size_t,
 129                        ctl_table *, void **);
 130 static int proc_doutsstring(ctl_table *table, int write, struct file *filp,
 131                   void __user *buffer, size_t *lenp, loff_t *ppos);
 132
 133 static ctl_table root_table[];
 134 static struct ctl_table_header root_table_header =
 135         { root_table, LIST_HEAD_INIT(root_table_header.ctl_entry) };
 136
 137 static ctl_table kern_table[];
 138 static ctl_table vm_table[];
 139 #ifdef CONFIG_NET
 140 extern ctl_table net_table[];
 141 #endif
 142 static ctl_table proc_table[];
 143 static ctl_table fs_table[];
 144 static ctl_table debug_table[];
 145 static ctl_table dev_table[];
 146 extern ctl_table random_table[];
 147 #ifdef CONFIG_UNIX98_PTYS
 148 extern ctl_table pty_table[];
 149 #endif
 150 #ifdef CONFIG_INOTIFY
 151 extern ctl_table inotify_table[];
 152 #endif
 153
 154 #ifdef HAVE_ARCH_PICK_MMAP_LAYOUT
 155 int sysctl_legacy_va_layout;
 156 #endif
 157
 158 /* /proc declarations: */
 159
 160 #ifdef CONFIG_PROC_FS
 161
 162 static ssize_t proc_readsys(struct file *, char __user *, size_t, loff_t *);
 163 static ssize_t proc_writesys(struct file *, const char __user *, size_t, loff_t *);
 164 static int proc_opensys(struct inode *, struct file *);
 165
 166 struct file_operations proc_sys_file_operations = {
 167         .open           = proc_opensys,
 168         .read           = proc_readsys,
 169         .write          = proc_writesys,
 170 };
 171
 172 extern struct proc_dir_entry *proc_sys_root;
 173
 174 static void register_proc_table(ctl_table *, struct proc_dir_entry *);
 175 static void unregister_proc_table(ctl_table *, struct proc_dir_entry *);
 176 #endif
 177
 178 /* The default sysctl tables: */
 179
 180 static ctl_table root_table[] = {
 181         {
 182                 .ctl_name       = CTL_KERN,
 183                 .procname       = "kernel",
 184                 .mode           = 0555,
 185                 .child          = kern_table,
 186         },
 187         {
 188                 .ctl_name       = CTL_VM,
 189                 .procname       = "vm",
 190                 .mode           = 0555,
 191                 .child          = vm_table,
 192         },
 193 #ifdef CONFIG_NET
 194         {
 195                 .ctl_name       = CTL_NET,
 196                 .procname       = "net",
 197                 .mode           = 0555,
 198                 .child          = net_table,
 199         },
 200 #endif
 201         {
 202                 .ctl_name       = CTL_PROC,
 203                 .procname       = "proc",
 204                 .mode           = 0555,
 205                 .child          = proc_table,
 206         },
 207         {
 208                 .ctl_name       = CTL_FS,
 209                 .procname       = "fs",
 210                 .mode           = 0555,
 211                 .child          = fs_table,
 212         },
 213         {
 214                 .ctl_name       = CTL_DEBUG,
 215                 .procname       = "debug",
 216                 .mode           = 0555,
 217                 .child          = debug_table,
 218         },
 219         {
 220                 .ctl_name       = CTL_DEV,
 221                 .procname       = "dev",
 222                 .mode           = 0555,
 223                 .child          = dev_table,
 224         },
 225
 226         { .ctl_name = 0 }
 227 };
 228
 229 static ctl_table kern_table[] = {
 230         {
 231                 .ctl_name       = KERN_OSTYPE,
 232                 .procname       = "ostype",
 233                 .data           = system_utsname.sysname,
 234                 .maxlen         = sizeof(system_utsname.sysname),
 235                 .mode           = 0444,
 236                 .proc_handler   = &proc_doutsstring,
 237                 .strategy       = &sysctl_string,
 238         },
 239         {
 240                 .ctl_name       = KERN_OSRELEASE,
 241                 .procname       = "osrelease",
 242                 .data           = system_utsname.release,
 243                 .maxlen         = sizeof(system_utsname.release),
 244                 .mode           = 0444,
 245                 .proc_handler   = &proc_doutsstring,
 246                 .strategy       = &sysctl_string,
 247         },
 248         {
 249                 .ctl_name       = KERN_VERSION,
 250                 .procname       = "version",
 251                 .data           = system_utsname.version,
 252                 .maxlen         = sizeof(system_utsname.version),
 253                 .mode           = 0444,
 254                 .proc_handler   = &proc_doutsstring,
 255                 .strategy       = &sysctl_string,
 256         },
 257         {
 258                 .ctl_name       = KERN_NODENAME,
 259                 .procname       = "hostname",
 260                 .data           = system_utsname.nodename,
 261                 .maxlen         = sizeof(system_utsname.nodename),
 262                 .mode           = 0644,
 263                 .proc_handler   = &proc_doutsstring,
 264                 .strategy       = &sysctl_string,
 265         },
 266         {
 267                 .ctl_name       = KERN_DOMAINNAME,
 268                 .procname       = "domainname",
 269                 .data           = system_utsname.domainname,
 270                 .maxlen         = sizeof(system_utsname.domainname),
 271                 .mode           = 0644,
 272                 .proc_handler   = &proc_doutsstring,
 273                 .strategy       = &sysctl_string,
 274         },
 275         {
 276                 .ctl_name       = KERN_PANIC,
 277                 .procname       = "panic",
 278                 .data           = &panic_timeout,
 279                 .maxlen         = sizeof(int),
 280                 .mode           = 0644,
 281                 .proc_handler   = &proc_dointvec,
 282         },
 283         {
 284                 .ctl_name       = KERN_CORE_USES_PID,
 285                 .procname       = "core_uses_pid",
 286                 .data           = &core_uses_pid,
 287                 .maxlen         = sizeof(int),
 288                 .mode           = 0644,
 289                 .proc_handler   = &proc_dointvec,
 290         },
 291         {
 292                 .ctl_name       = KERN_CORE_PATTERN,
 293                 .procname       = "core_pattern",
 294                 .data           = core_pattern,
 295                 .maxlen         = 64,
 296                 .mode           = 0644,
 297                 .proc_handler   = &proc_dostring,
 298                 .strategy       = &sysctl_string,
 299         },
 300         {
 301                 .ctl_name       = KERN_TAINTED,
 302                 .procname       = "tainted",
 303                 .data           = &tainted,
 304                 .maxlen         = sizeof(int),
 305                 .mode           = 0444,
 306                 .proc_handler   = &proc_dointvec,
 307         },
 308         {
 309                 .ctl_name       = KERN_CAP_BSET,
 310                 .procname       = "cap-bound",
 311                 .data           = &cap_bset,
 312                 .maxlen         = sizeof(kernel_cap_t),
 313                 .mode           = 0600,
 314                 .proc_handler   = &proc_dointvec_bset,
 315         },
 316 #ifdef CONFIG_BLK_DEV_INITRD
 317         {
 318                 .ctl_name       = KERN_REALROOTDEV,
 319                 .procname       = "real-root-dev",
 320                 .data           = &real_root_dev,
 321                 .maxlen         = sizeof(int),
 322                 .mode           = 0644,
 323                 .proc_handler   = &proc_dointvec,
 324         },
 325 #endif
 326 #ifdef __sparc__
 327         {
 328                 .ctl_name       = KERN_SPARC_REBOOT,
 329                 .procname       = "reboot-cmd",
 330                 .data           = reboot_command,
 331                 .maxlen         = 256,
 332                 .mode           = 0644,
 333                 .proc_handler   = &proc_dostring,
 334                 .strategy       = &sysctl_string,
 335         },
 336         {
 337                 .ctl_name       = KERN_SPARC_STOP_A,
 338                 .procname       = "stop-a",
 339                 .data           = &stop_a_enabled,
 340                 .maxlen         = sizeof (int),
 341                 .mode           = 0644,
 342                 .proc_handler   = &proc_dointvec,
 343         },
 344         {
 345                 .ctl_name       = KERN_SPARC_SCONS_PWROFF,
 346                 .procname       = "scons-poweroff",
 347                 .data           = &scons_pwroff,
 348                 .maxlen         = sizeof (int),
 349                 .mode           = 0644,
 350                 .proc_handler   = &proc_dointvec,
 351         },
 352 #endif
 353 #ifdef __hppa__
 354         {
 355                 .ctl_name       = KERN_HPPA_PWRSW,
 356                 .procname       = "soft-power",
 357                 .data           = &pwrsw_enabled,
 358                 .maxlen         = sizeof (int),
 359                 .mode           = 0644,
 360                 .proc_handler   = &proc_dointvec,
 361         },
 362         {
 363                 .ctl_name       = KERN_HPPA_UNALIGNED,
 364                 .procname       = "unaligned-trap",
 365                 .data           = &unaligned_enabled,
 366                 .maxlen         = sizeof (int),
 367                 .mode           = 0644,
 368                 .proc_handler   = &proc_dointvec,
 369         },
 370 #endif
 371         {
 372                 .ctl_name       = KERN_CTLALTDEL,
 373                 .procname       = "ctrl-alt-del",
 374                 .data           = &C_A_D,
 375                 .maxlen         = sizeof(int),
 376                 .mode           = 0644,
 377                 .proc_handler   = &proc_dointvec,
 378         },
 379         {
 380                 .ctl_name       = KERN_PRINTK,
 381                 .procname       = "printk",
 382                 .data           = &console_loglevel,
 383                 .maxlen         = 4*sizeof(int),
 384                 .mode           = 0644,
 385                 .proc_handler   = &proc_dointvec,
 386         },
 387 #ifdef CONFIG_KMOD
 388         {
 389                 .ctl_name       = KERN_MODPROBE,
 390                 .procname       = "modprobe",
 391                 .data           = &modprobe_path,
 392                 .maxlen         = KMOD_PATH_LEN,
 393                 .mode           = 0644,
 394                 .proc_handler   = &proc_dostring,
 395                 .strategy       = &sysctl_string,
 396         },
 397 #endif
 398 #ifdef CONFIG_HOTPLUG
 399         {
 400                 .ctl_name       = KERN_HOTPLUG,
 401                 .procname       = "hotplug",
 402                 .data           = &hotplug_path,
 403                 .maxlen         = HOTPLUG_PATH_LEN,
 404                 .mode           = 0644,
 405                 .proc_handler   = &proc_dostring,
 406                 .strategy       = &sysctl_string,
 407         },
 408 #endif
 409 #ifdef CONFIG_CHR_DEV_SG
 410         {
 411                 .ctl_name       = KERN_SG_BIG_BUFF,
 412                 .procname       = "sg-big-buff",
 413                 .data           = &sg_big_buff,
 414                 .maxlen         = sizeof (int),
 415                 .mode           = 0444,
 416                 .proc_handler   = &proc_dointvec,
 417         },
 418 #endif
 419 #ifdef CONFIG_BSD_PROCESS_ACCT
 420         {
 421                 .ctl_name       = KERN_ACCT,
 422                 .procname       = "acct",
 423                 .data           = &acct_parm,
 424                 .maxlen         = 3*sizeof(int),
 425                 .mode           = 0644,
 426                 .proc_handler   = &proc_dointvec,
 427         },
 428 #endif
 429 #ifdef CONFIG_SYSVIPC
 430         {
 431                 .ctl_name       = KERN_SHMMAX,
 432                 .procname       = "shmmax",
 433                 .data           = &shm_ctlmax,
 434                 .maxlen         = sizeof (size_t),
 435                 .mode           = 0644,
 436                 .proc_handler   = &proc_doulongvec_minmax,
 437         },
 438         {
 439                 .ctl_name       = KERN_SHMALL,
 440                 .procname       = "shmall",
 441                 .data           = &shm_ctlall,
 442                 .maxlen         = sizeof (size_t),
 443                 .mode           = 0644,
 444                 .proc_handler   = &proc_doulongvec_minmax,
 445         },
 446         {
 447                 .ctl_name       = KERN_SHMMNI,
 448                 .procname       = "shmmni",
 449                 .data           = &shm_ctlmni,
 450                 .maxlen         = sizeof (int),
 451                 .mode           = 0644,
 452                 .proc_handler   = &proc_dointvec,
 453         },
 454         {
 455                 .ctl_name       = KERN_MSGMAX,
 456                 .procname       = "msgmax",
 457                 .data           = &msg_ctlmax,
 458                 .maxlen         = sizeof (int),
 459                 .mode           = 0644,
 460                 .proc_handler   = &proc_dointvec,
 461         },
 462         {
 463                 .ctl_name       = KERN_MSGMNI,
 464                 .procname       = "msgmni",
 465                 .data           = &msg_ctlmni,
 466                 .maxlen         = sizeof (int),
 467                 .mode           = 0644,
 468                 .proc_handler   = &proc_dointvec,
 469         },
 470         {
 471                 .ctl_name       = KERN_MSGMNB,
 472                 .procname       =  "msgmnb",
 473                 .data           = &msg_ctlmnb,
 474                 .maxlen         = sizeof (int),
 475                 .mode           = 0644,
 476                 .proc_handler   = &proc_dointvec,
 477         },
 478         {
 479                 .ctl_name       = KERN_SEM,
 480                 .procname       = "sem",
 481                 .data           = &sem_ctls,
 482                 .maxlen         = 4*sizeof (int),
 483                 .mode           = 0644,
 484                 .proc_handler   = &proc_dointvec,
 485         },
 486 #endif
 487 #ifdef CONFIG_MAGIC_SYSRQ
 488         {
 489                 .ctl_name       = KERN_SYSRQ,
 490                 .procname       = "sysrq",
 491                 .data           = &sysrq_enabled,
 492                 .maxlen         = sizeof (int),
 493                 .mode           = 0644,
 494                 .proc_handler   = &proc_dointvec,
 495         },
 496 #endif
 497         {
 498                 .ctl_name       = KERN_CADPID,
 499                 .procname       = "cad_pid",
 500                 .data           = &cad_pid,
 501                 .maxlen         = sizeof (int),
 502                 .mode           = 0600,
 503                 .proc_handler   = &proc_dointvec,
 504         },
 505         {
 506                 .ctl_name       = KERN_MAX_THREADS,
 507                 .procname       = "threads-max",
 508                 .data           = &max_threads,
 509                 .maxlen         = sizeof(int),
 510                 .mode           = 0644,
 511                 .proc_handler   = &proc_dointvec,
 512         },
 513         {
 514                 .ctl_name       = KERN_RANDOM,
 515                 .procname       = "random",
 516                 .mode           = 0555,
 517                 .child          = random_table,
 518         },
 519 #ifdef CONFIG_UNIX98_PTYS
 520         {
 521                 .ctl_name       = KERN_PTY,
 522                 .procname       = "pty",
 523                 .mode           = 0555,
 524                 .child          = pty_table,
 525         },
 526 #endif
 527         {
 528                 .ctl_name       = KERN_OVERFLOWUID,
 529                 .procname       = "overflowuid",
 530                 .data           = &overflowuid,
 531                 .maxlen         = sizeof(int),
 532                 .mode           = 0644,
 533                 .proc_handler   = &proc_dointvec_minmax,
 534                 .strategy       = &sysctl_intvec,
 535                 .extra1         = &minolduid,
 536                 .extra2         = &maxolduid,
 537         },
 538         {
 539                 .ctl_name       = KERN_OVERFLOWGID,
 540                 .procname       = "overflowgid",
 541                 .data           = &overflowgid,
 542                 .maxlen         = sizeof(int),
 543                 .mode           = 0644,
 544                 .proc_handler   = &proc_dointvec_minmax,
 545                 .strategy       = &sysctl_intvec,
 546                 .extra1         = &minolduid,
 547                 .extra2         = &maxolduid,
 548         },
 549 #ifdef CONFIG_ARCH_S390
 550 #ifdef CONFIG_MATHEMU
 551         {
 552                 .ctl_name       = KERN_IEEE_EMULATION_WARNINGS,
 553                 .procname       = "ieee_emulation_warnings",
 554                 .data           = &sysctl_ieee_emulation_warnings,
 555                 .maxlen         = sizeof(int),
 556                 .mode           = 0644,
 557                 .proc_handler   = &proc_dointvec,
 558         },
 559 #endif
 560 #ifdef CONFIG_NO_IDLE_HZ
 561         {
 562                 .ctl_name       = KERN_HZ_TIMER,
 563                 .procname       = "hz_timer",
 564                 .data           = &sysctl_hz_timer,
 565                 .maxlen         = sizeof(int),
 566                 .mode           = 0644,
 567                 .proc_handler   = &proc_dointvec,
 568         },
 569 #endif
 570         {
 571                 .ctl_name       = KERN_S390_USER_DEBUG_LOGGING,
 572                 .procname       = "userprocess_debug",
 573                 .data           = &sysctl_userprocess_debug,
 574                 .maxlen         = sizeof(int),
 575                 .mode           = 0644,
 576                 .proc_handler   = &proc_dointvec,
 577         },
 578 #endif
 579         {
 580                 .ctl_name       = KERN_PIDMAX,
 581                 .procname       = "pid_max",
 582                 .data           = &pid_max,
 583                 .maxlen         = sizeof (int),
 584                 .mode           = 0644,
 585                 .proc_handler   = &proc_dointvec_minmax,
 586                 .strategy       = sysctl_intvec,
 587                 .extra1         = &pid_max_min,
 588                 .extra2         = &pid_max_max,
 589         },
 590         {
 591                 .ctl_name       = KERN_PANIC_ON_OOPS,
 592                 .procname       = "panic_on_oops",
 593                 .data           = &panic_on_oops,
 594                 .maxlen         = sizeof(int),
 595                 .mode           = 0644,
 596                 .proc_handler   = &proc_dointvec,
 597         },
 598         {
 599                 .ctl_name       = KERN_PRINTK_RATELIMIT,
 600                 .procname       = "printk_ratelimit",
 601                 .data           = &printk_ratelimit_jiffies,
 602                 .maxlen         = sizeof(int),
 603                 .mode           = 0644,
 604                 .proc_handler   = &proc_dointvec_jiffies,
 605                 .strategy       = &sysctl_jiffies,
 606         },
 607         {
 608                 .ctl_name       = KERN_PRINTK_RATELIMIT_BURST,
 609                 .procname       = "printk_ratelimit_burst",
 610                 .data           = &printk_ratelimit_burst,
 611                 .maxlen         = sizeof(int),
 612                 .mode           = 0644,
 613                 .proc_handler   = &proc_dointvec,
 614         },
 615         {
 616                 .ctl_name       = KERN_NGROUPS_MAX,
 617                 .procname       = "ngroups_max",
 618                 .data           = &ngroups_max,
 619                 .maxlen         = sizeof (int),
 620                 .mode           = 0444,
 621                 .proc_handler   = &proc_dointvec,
 622         },
 623 #if defined(CONFIG_X86_LOCAL_APIC) && defined(CONFIG_X86)
 624         {
 625                 .ctl_name       = KERN_UNKNOWN_NMI_PANIC,
 626                 .procname       = "unknown_nmi_panic",
 627                 .data           = &unknown_nmi_panic,
 628                 .maxlen         = sizeof (int),
 629                 .mode           = 0644,
 630                 .proc_handler   = &proc_unknown_nmi_panic,
 631         },
 632 #endif
 633 #if defined(CONFIG_X86)
 634         {
 635                 .ctl_name       = KERN_BOOTLOADER_TYPE,
 636                 .procname       = "bootloader_type",
 637                 .data           = &bootloader_type,
 638                 .maxlen         = sizeof (int),
 639                 .mode           = 0444,
 640                 .proc_handler   = &proc_dointvec,
 641         },
 642 #endif
 643         {
 644                 .ctl_name       = KERN_RANDOMIZE,
 645                 .procname       = "randomize_va_space",
 646                 .data           = &randomize_va_space,
 647                 .maxlen         = sizeof(int),
 648                 .mode           = 0644,
 649                 .proc_handler   = &proc_dointvec,
 650         },
 651 #if defined(CONFIG_ARCH_S390)
 652         {
 653                 .ctl_name       = KERN_SPIN_RETRY,
 654                 .procname       = "spin_retry",
 655                 .data           = &spin_retry,
 656                 .maxlen         = sizeof (int),
 657                 .mode           = 0644,
 658                 .proc_handler   = &proc_dointvec,
 659         },
 660 #endif
 661         { .ctl_name = 0 }
 662 };
 663
 664 /* Constants for minimum and maximum testing in vm_table.
 665    We use these as one-element integer vectors. */
 666 static int zero;
 667 static int one_hundred = 100;
 668
 669
 670 static ctl_table vm_table[] = {
 671         {
 672                 .ctl_name       = VM_OVERCOMMIT_MEMORY,
 673                 .procname       = "overcommit_memory",
 674                 .data           = &sysctl_overcommit_memory,
 675                 .maxlen         = sizeof(sysctl_overcommit_memory),
 676                 .mode           = 0644,
 677                 .proc_handler   = &proc_dointvec,
 678         },
 679         {
 680                 .ctl_name       = VM_OVERCOMMIT_RATIO,
 681                 .procname       = "overcommit_ratio",
 682                 .data           = &sysctl_overcommit_ratio,
 683                 .maxlen         = sizeof(sysctl_overcommit_ratio),
 684                 .mode           = 0644,
 685                 .proc_handler   = &proc_dointvec,
 686         },
 687         {
 688                 .ctl_name       = VM_PAGE_CLUSTER,
 689                 .procname       = "page-cluster",
 690                 .data           = &page_cluster,
 691                 .maxlen         = sizeof(int),
 692                 .mode           = 0644,
 693                 .proc_handler   = &proc_dointvec,
 694         },
 695         {
 696                 .ctl_name       = VM_DIRTY_BACKGROUND,
 697                 .procname       = "dirty_background_ratio",
 698                 .data           = &dirty_background_ratio,
 699                 .maxlen         = sizeof(dirty_background_ratio),
 700                 .mode           = 0644,
 701                 .proc_handler   = &proc_dointvec_minmax,
 702                 .strategy       = &sysctl_intvec,
 703                 .extra1         = &zero,
 704                 .extra2         = &one_hundred,
 705         },
 706         {
 707                 .ctl_name       = VM_DIRTY_RATIO,
 708                 .procname       = "dirty_ratio",
 709                 .data           = &vm_dirty_ratio,
 710                 .maxlen         = sizeof(vm_dirty_ratio),
 711                 .mode           = 0644,
 712                 .proc_handler   = &proc_dointvec_minmax,
 713                 .strategy       = &sysctl_intvec,
 714                 .extra1         = &zero,
 715                 .extra2         = &one_hundred,
 716         },
 717         {
 718                 .ctl_name       = VM_DIRTY_WB_CS,
 719                 .procname       = "dirty_writeback_centisecs",
 720                 .data           = &dirty_writeback_centisecs,
 721                 .maxlen         = sizeof(dirty_writeback_centisecs),
 722                 .mode           = 0644,
 723                 .proc_handler   = &dirty_writeback_centisecs_handler,
 724         },
 725         {
 726                 .ctl_name       = VM_DIRTY_EXPIRE_CS,
 727                 .procname       = "dirty_expire_centisecs",
 728                 .data           = &dirty_expire_centisecs,
 729                 .maxlen         = sizeof(dirty_expire_centisecs),
 730                 .mode           = 0644,
 731                 .proc_handler   = &proc_dointvec,
 732         },
 733         {
 734                 .ctl_name       = VM_NR_PDFLUSH_THREADS,
 735                 .procname       = "nr_pdflush_threads",
 736                 .data           = &nr_pdflush_threads,
 737                 .maxlen         = sizeof nr_pdflush_threads,
 738                 .mode           = 0444 /* read-only*/,
 739                 .proc_handler   = &proc_dointvec,
 740         },
 741         {
 742                 .ctl_name       = VM_SWAPPINESS,
 743                 .procname       = "swappiness",
 744                 .data           = &vm_swappiness,
 745                 .maxlen         = sizeof(vm_swappiness),
 746                 .mode           = 0644,
 747                 .proc_handler   = &proc_dointvec_minmax,
 748                 .strategy       = &sysctl_intvec,
 749                 .extra1         = &zero,
 750                 .extra2         = &one_hundred,
 751         },
 752 #ifdef CONFIG_HUGETLB_PAGE
 753          {
 754                 .ctl_name       = VM_HUGETLB_PAGES,
 755                 .procname       = "nr_hugepages",
 756                 .data           = &max_huge_pages,
 757                 .maxlen         = sizeof(unsigned long),
 758                 .mode           = 0644,
 759                 .proc_handler   = &hugetlb_sysctl_handler,
 760                 .extra1         = (void *)&hugetlb_zero,
 761                 .extra2         = (void *)&hugetlb_infinity,
 762          },
 763          {
 764                 .ctl_name       = VM_HUGETLB_GROUP,
 765                 .procname       = "hugetlb_shm_group",
 766                 .data           = &sysctl_hugetlb_shm_group,
 767                 .maxlen         = sizeof(gid_t),
 768                 .mode           = 0644,
 769                 .proc_handler   = &proc_dointvec,
 770          },
 771 #endif
 772         {
 773                 .ctl_name       = VM_LOWMEM_RESERVE_RATIO,
 774                 .procname       = "lowmem_reserve_ratio",
 775                 .data           = &sysctl_lowmem_reserve_ratio,
 776                 .maxlen         = sizeof(sysctl_lowmem_reserve_ratio),
 777                 .mode           = 0644,
 778                 .proc_handler   = &lowmem_reserve_ratio_sysctl_handler,
 779                 .strategy       = &sysctl_intvec,
 780         },
 781         {
 782                 .ctl_name       = VM_MIN_FREE_KBYTES,
 783                 .procname       = "min_free_kbytes",
 784                 .data           = &min_free_kbytes,
 785                 .maxlen         = sizeof(min_free_kbytes),
 786                 .mode           = 0644,
 787                 .proc_handler   = &min_free_kbytes_sysctl_handler,
 788                 .strategy       = &sysctl_intvec,
 789                 .extra1         = &zero,
 790         },
 791 #ifdef CONFIG_MMU
 792         {
 793                 .ctl_name       = VM_MAX_MAP_COUNT,
 794                 .procname       = "max_map_count",
 795                 .data           = &sysctl_max_map_count,
 796                 .maxlen         = sizeof(sysctl_max_map_count),
 797                 .mode           = 0644,
 798                 .proc_handler   = &proc_dointvec
 799         },
 800 #endif
 801         {
 802                 .ctl_name       = VM_LAPTOP_MODE,
 803                 .procname       = "laptop_mode",
 804                 .data           = &laptop_mode,
 805                 .maxlen         = sizeof(laptop_mode),
 806                 .mode           = 0644,
 807                 .proc_handler   = &proc_dointvec,
 808                 .strategy       = &sysctl_intvec,
 809                 .extra1         = &zero,
 810         },
 811         {
 812                 .ctl_name       = VM_BLOCK_DUMP,
 813                 .procname       = "block_dump",
 814                 .data           = &block_dump,
 815                 .maxlen         = sizeof(block_dump),
 816                 .mode           = 0644,
 817                 .proc_handler   = &proc_dointvec,
 818                 .strategy       = &sysctl_intvec,
 819                 .extra1         = &zero,
 820         },
 821         {
 822                 .ctl_name       = VM_VFS_CACHE_PRESSURE,
 823                 .procname       = "vfs_cache_pressure",
 824                 .data           = &sysctl_vfs_cache_pressure,
 825                 .maxlen         = sizeof(sysctl_vfs_cache_pressure),
 826                 .mode           = 0644,
 827                 .proc_handler   = &proc_dointvec,
 828                 .strategy       = &sysctl_intvec,
 829                 .extra1         = &zero,
 830         },
 831 #ifdef HAVE_ARCH_PICK_MMAP_LAYOUT
 832         {
 833                 .ctl_name       = VM_LEGACY_VA_LAYOUT,
 834                 .procname       = "legacy_va_layout",
 835                 .data           = &sysctl_legacy_va_layout,
 836                 .maxlen         = sizeof(sysctl_legacy_va_layout),
 837                 .mode           = 0644,
 838                 .proc_handler   = &proc_dointvec,
 839                 .strategy       = &sysctl_intvec,
 840                 .extra1         = &zero,
 841         },
 842 #endif
 843 #ifdef CONFIG_SWAP
 844         {
 845                 .ctl_name       = VM_SWAP_TOKEN_TIMEOUT,
 846                 .procname       = "swap_token_timeout",
 847                 .data           = &swap_token_default_timeout,
 848                 .maxlen         = sizeof(swap_token_default_timeout),
 849                 .mode           = 0644,
 850                 .proc_handler   = &proc_dointvec_jiffies,
 851                 .strategy       = &sysctl_jiffies,
 852         },
 853 #endif
 854         { .ctl_name = 0 }
 855 };
 856
 857 static ctl_table proc_table[] = {
 858         { .ctl_name = 0 }
 859 };
 860
 861 static ctl_table fs_table[] = {
 862         {
 863                 .ctl_name       = FS_NRINODE,
 864                 .procname       = "inode-nr",
 865                 .data           = &inodes_stat,
 866                 .maxlen         = 2*sizeof(int),
 867                 .mode           = 0444,
 868                 .proc_handler   = &proc_dointvec,
 869         },
 870         {
 871                 .ctl_name       = FS_STATINODE,
 872                 .procname       = "inode-state",
 873                 .data           = &inodes_stat,
 874                 .maxlen         = 7*sizeof(int),
 875                 .mode           = 0444,
 876                 .proc_handler   = &proc_dointvec,
 877         },
 878         {
 879                 .ctl_name       = FS_NRFILE,
 880                 .procname       = "file-nr",
 881                 .data           = &files_stat,
 882                 .maxlen         = 3*sizeof(int),
 883                 .mode           = 0444,
 884                 .proc_handler   = &proc_dointvec,
 885         },
 886         {
 887                 .ctl_name       = FS_MAXFILE,
 888                 .procname       = "file-max",
 889                 .data           = &files_stat.max_files,
 890                 .maxlen         = sizeof(int),
 891                 .mode           = 0644,
 892                 .proc_handler   = &proc_dointvec,
 893         },
 894         {
 895                 .ctl_name       = FS_DENTRY,
 896                 .procname       = "dentry-state",
 897                 .data           = &dentry_stat,
 898                 .maxlen         = 6*sizeof(int),
 899                 .mode           = 0444,
 900                 .proc_handler   = &proc_dointvec,
 901         },
 902         {
 903                 .ctl_name       = FS_OVERFLOWUID,
 904                 .procname       = "overflowuid",
 905                 .data           = &fs_overflowuid,
 906                 .maxlen         = sizeof(int),
 907                 .mode           = 0644,
 908                 .proc_handler   = &proc_dointvec_minmax,
 909                 .strategy       = &sysctl_intvec,
 910                 .extra1         = &minolduid,
 911                 .extra2         = &maxolduid,
 912         },
 913         {
 914                 .ctl_name       = FS_OVERFLOWGID,
 915                 .procname       = "overflowgid",
 916                 .data           = &fs_overflowgid,
 917                 .maxlen         = sizeof(int),
 918                 .mode           = 0644,
 919                 .proc_handler   = &proc_dointvec_minmax,
 920                 .strategy       = &sysctl_intvec,
 921                 .extra1         = &minolduid,
 922                 .extra2         = &maxolduid,
 923         },
 924         {
 925                 .ctl_name       = FS_LEASES,
 926                 .procname       = "leases-enable",
 927                 .data           = &leases_enable,
 928                 .maxlen         = sizeof(int),
 929                 .mode           = 0644,
 930                 .proc_handler   = &proc_dointvec,
 931         },
 932 #ifdef CONFIG_DNOTIFY
 933         {
 934                 .ctl_name       = FS_DIR_NOTIFY,
 935                 .procname       = "dir-notify-enable",
 936                 .data           = &dir_notify_enable,
 937                 .maxlen         = sizeof(int),
 938                 .mode           = 0644,
 939                 .proc_handler   = &proc_dointvec,
 940         },
 941 #endif
 942 #ifdef CONFIG_MMU
 943         {
 944                 .ctl_name       = FS_LEASE_TIME,
 945                 .procname       = "lease-break-time",
 946                 .data           = &lease_break_time,
 947                 .maxlen         = sizeof(int),
 948                 .mode           = 0644,
 949                 .proc_handler   = &proc_dointvec,
 950         },
 951         {
 952                 .ctl_name       = FS_AIO_NR,
 953                 .procname       = "aio-nr",
 954                 .data           = &aio_nr,
 955                 .maxlen         = sizeof(aio_nr),
 956                 .mode           = 0444,
 957                 .proc_handler   = &proc_dointvec,
 958         },
 959         {
 960                 .ctl_name       = FS_AIO_MAX_NR,
 961                 .procname       = "aio-max-nr",
 962                 .data           = &aio_max_nr,
 963                 .maxlen         = sizeof(aio_max_nr),
 964                 .mode           = 0644,
 965                 .proc_handler   = &proc_dointvec,
 966         },
 967 #ifdef CONFIG_INOTIFY
 968         {
 969                 .ctl_name       = FS_INOTIFY,
 970                 .procname       = "inotify",
 971                 .mode           = 0555,
 972                 .child          = inotify_table,
 973         },
 974 #endif
 975 #endif
 976         {
 977                 .ctl_name       = KERN_SETUID_DUMPABLE,
 978                 .procname       = "suid_dumpable",
 979                 .data           = &suid_dumpable,
 980                 .maxlen         = sizeof(int),
 981                 .mode           = 0644,
 982                 .proc_handler   = &proc_dointvec,
 983         },
 984         { .ctl_name = 0 }
 985 };
 986
 987 static ctl_table debug_table[] = {
 988         { .ctl_name = 0 }
 989 };
 990
 991 static ctl_table dev_table[] = {
 992         { .ctl_name = 0 }
 993 };
 994
 995 extern void init_irq_proc (void);
 996
 997 void __init sysctl_init(void)
 998 {
 999 #ifdef CONFIG_PROC_FS
1000         register_proc_table(root_table, proc_sys_root);
1001         init_irq_proc();
1002 #endif
1003 }
1004
1005 int do_sysctl(int __user *name, int nlen, void __user *oldval, size_t __user *oldlenp,
1006                void __user *newval, size_t newlen)
1007 {
1008         struct list_head *tmp;
1009
1010         if (nlen <= 0 || nlen >= CTL_MAXNAME)
1011                 return -ENOTDIR;
1012         if (oldval) {
1013                 int old_len;
1014                 if (!oldlenp || get_user(old_len, oldlenp))
1015                         return -EFAULT;
1016         }
1017         tmp = &root_table_header.ctl_entry;
1018         do {
1019                 struct ctl_table_header *head =
1020                         list_entry(tmp, struct ctl_table_header, ctl_entry);
1021                 void *context = NULL;
1022                 int error = parse_table(name, nlen, oldval, oldlenp,
1023                                         newval, newlen, head->ctl_table,
1024                                         &context);
1025                 kfree(context);
1026                 if (error != -ENOTDIR)
1027                         return error;
1028                 tmp = tmp->next;
1029         } while (tmp != &root_table_header.ctl_entry);
1030         return -ENOTDIR;
1031 }
1032
1033 asmlinkage long sys_sysctl(struct __sysctl_args __user *args)
1034 {
1035         struct __sysctl_args tmp;
1036         int error;
1037
1038         if (copy_from_user(&tmp, args, sizeof(tmp)))
1039                 return -EFAULT;
1040
1041         lock_kernel();
1042         error = do_sysctl(tmp.name, tmp.nlen, tmp.oldval, tmp.oldlenp,
1043                           tmp.newval, tmp.newlen);
1044         unlock_kernel();
1045         return error;
1046 }
1047
1048 /*
1049  * ctl_perm does NOT grant the superuser all rights automatically, because
1050  * some sysctl variables are readonly even to root.
1051  */
1052
1053 static int test_perm(int mode, int op)
1054 {
1055         if (!current->euid)
1056                 mode >>= 6;
1057         else if (in_egroup_p(0))
1058                 mode >>= 3;
1059         if ((mode & op & 0007) == op)
1060                 return 0;
1061         return -EACCES;
1062 }
1063
1064 static inline int ctl_perm(ctl_table *table, int op)
1065 {
1066         int error;
1067         error = security_sysctl(table, op);
1068         if (error)
1069                 return error;
1070         return test_perm(table->mode, op);
1071 }
1072
1073 static int parse_table(int __user *name, int nlen,
1074                        void __user *oldval, size_t __user *oldlenp,
1075                        void __user *newval, size_t newlen,
1076                        ctl_table *table, void **context)
1077 {
1078         int n;
1079 repeat:
1080         if (!nlen)
1081                 return -ENOTDIR;
1082         if (get_user(n, name))
1083                 return -EFAULT;
1084         for ( ; table->ctl_name; table++) {
1085                 if (n == table->ctl_name || table->ctl_name == CTL_ANY) {
1086                         int error;
1087                         if (table->child) {
1088                                 if (ctl_perm(table, 001))
1089                                         return -EPERM;
1090                                 if (table->strategy) {
1091                                         error = table->strategy(
1092                                                 table, name, nlen,
1093                                                 oldval, oldlenp,
1094                                                 newval, newlen, context);
1095                                         if (error)
1096                                                 return error;
1097                                 }
1098                                 name++;
1099                                 nlen--;
1100                                 table = table->child;
1101                                 goto repeat;
1102                         }
1103                         error = do_sysctl_strategy(table, name, nlen,
1104                                                    oldval, oldlenp,
1105                                                    newval, newlen, context);
1106                         return error;
1107                 }
1108         }
1109         return -ENOTDIR;
1110 }
1111
1112 /* Perform the actual read/write of a sysctl table entry. */
1113 int do_sysctl_strategy (ctl_table *table,
1114                         int __user *name, int nlen,
1115                         void __user *oldval, size_t __user *oldlenp,
1116                         void __user *newval, size_t newlen, void **context)
1117 {
1118         int op = 0, rc;
1119         size_t len;
1120
1121         if (oldval)
1122                 op |= 004;
1123         if (newval)
1124                 op |= 002;
1125         if (ctl_perm(table, op))
1126                 return -EPERM;
1127
1128         if (table->strategy) {
1129                 rc = table->strategy(table, name, nlen, oldval, oldlenp,
1130                                      newval, newlen, context);
1131                 if (rc < 0)
1132                         return rc;
1133                 if (rc > 0)
1134                         return 0;
1135         }
1136
1137         /* If there is no strategy routine, or if the strategy returns
1138          * zero, proceed with automatic r/w */
1139         if (table->data && table->maxlen) {
1140                 if (oldval && oldlenp) {
1141                         if (get_user(len, oldlenp))
1142                                 return -EFAULT;
1143                         if (len) {
1144                                 if (len > table->maxlen)
1145                                         len = table->maxlen;
1146                                 if(copy_to_user(oldval, table->data, len))
1147                                         return -EFAULT;
1148                                 if(put_user(len, oldlenp))
1149                                         return -EFAULT;
1150                         }
1151                 }
1152                 if (newval && newlen) {
1153                         len = newlen;
1154                         if (len > table->maxlen)
1155                                 len = table->maxlen;
1156                         if(copy_from_user(table->data, newval, len))
1157                                 return -EFAULT;
1158                 }
1159         }
1160         return 0;
1161 }
1162
1163 /**
1164  * register_sysctl_table - register a sysctl hierarchy
1165  * @table: the top-level table structure
1166  * @insert_at_head: whether the entry should be inserted in front or at the end
1167  *
1168  * Register a sysctl table hierarchy. @table should be a filled in ctl_table
1169  * array. An entry with a ctl_name of 0 terminates the table.
1170  *
1171  * The members of the &ctl_table structure are used as follows:
1172  *
1173  * ctl_name - This is the numeric sysctl value used by sysctl(2). The number
1174  *            must be unique within that level of sysctl
1175  *
1176  * procname - the name of the sysctl file under /proc/sys. Set to %NULL to not
1177  *            enter a sysctl file
1178  *
1179  * data - a pointer to data for use by proc_handler
1180  *
1181  * maxlen - the maximum size in bytes of the data
1182  *
1183  * mode - the file permissions for the /proc/sys file, and for sysctl(2)
1184  *
1185  * child - a pointer to the child sysctl table if this entry is a directory, or
1186  *         %NULL.
1187  *
1188  * proc_handler - the text handler routine (described below)
1189  *
1190  * strategy - the strategy routine (described below)
1191  *
1192  * de - for internal use by the sysctl routines
1193  *
1194  * extra1, extra2 - extra pointers usable by the proc handler routines
1195  *
1196  * Leaf nodes in the sysctl tree will be represented by a single file
1197  * under /proc; non-leaf nodes will be represented by directories.
1198  *
1199  * sysctl(2) can automatically manage read and write requests through
1200  * the sysctl table.  The data and maxlen fields of the ctl_table
1201  * struct enable minimal validation of the values being written to be
1202  * performed, and the mode field allows minimal authentication.
1203  *
1204  * More sophisticated management can be enabled by the provision of a
1205  * strategy routine with the table entry.  This will be called before
1206  * any automatic read or write of the data is performed.
1207  *
1208  * The strategy routine may return
1209  *
1210  * < 0 - Error occurred (error is passed to user process)
1211  *
1212  * 0   - OK - proceed with automatic read or write.
1213  *
1214  * > 0 - OK - read or write has been done by the strategy routine, so
1215  *       return immediately.
1216  *
1217  * There must be a proc_handler routine for any terminal nodes
1218  * mirrored under /proc/sys (non-terminals are handled by a built-in
1219  * directory handler).  Several default handlers are available to
1220  * cover common cases -
1221  *
1222  * proc_dostring(), proc_dointvec(), proc_dointvec_jiffies(),
1223  * proc_dointvec_userhz_jiffies(), proc_dointvec_minmax(),
1224  * proc_doulongvec_ms_jiffies_minmax(), proc_doulongvec_minmax()
1225  *
1226  * It is the handler's job to read the input buffer from user memory
1227  * and process it. The handler should return 0 on success.
1228  *
1229  * This routine returns %NULL on a failure to register, and a pointer
1230  * to the table header on success.
1231  */
1232 struct ctl_table_header *register_sysctl_table(ctl_table * table,
1233                                                int insert_at_head)
1234 {
1235         struct ctl_table_header *tmp;
1236         tmp = kmalloc(sizeof(struct ctl_table_header), GFP_KERNEL);
1237         if (!tmp)
1238                 return NULL;
1239         tmp->ctl_table = table;
1240         INIT_LIST_HEAD(&tmp->ctl_entry);
1241         if (insert_at_head)
1242                 list_add(&tmp->ctl_entry, &root_table_header.ctl_entry);
1243         else
1244                 list_add_tail(&tmp->ctl_entry, &root_table_header.ctl_entry);
1245 #ifdef CONFIG_PROC_FS
1246         register_proc_table(table, proc_sys_root);
1247 #endif
1248         return tmp;
1249 }
1250
1251 /**
1252  * unregister_sysctl_table - unregister a sysctl table hierarchy
1253  * @header: the header returned from register_sysctl_table
1254  *
1255  * Unregisters the sysctl table and all children. proc entries may not
1256  * actually be removed until they are no longer used by anyone.
1257  */
1258 void unregister_sysctl_table(struct ctl_table_header * header)
1259 {
1260         list_del(&header->ctl_entry);
1261 #ifdef CONFIG_PROC_FS
1262         unregister_proc_table(header->ctl_table, proc_sys_root);
1263 #endif
1264         kfree(header);
1265 }
1266
1267 /*
1268  * /proc/sys support
1269  */
1270
1271 #ifdef CONFIG_PROC_FS
1272
1273 /* Scan the sysctl entries in table and add them all into /proc */
1274 static void register_proc_table(ctl_table * table, struct proc_dir_entry *root)
1275 {
1276         struct proc_dir_entry *de;
1277         int len;
1278         mode_t mode;
1279
1280         for (; table->ctl_name; table++) {
1281                 /* Can't do anything without a proc name. */
1282                 if (!table->procname)
1283                         continue;
1284                 /* Maybe we can't do anything with it... */
1285                 if (!table->proc_handler && !table->child) {
1286                         printk(KERN_WARNING "SYSCTL: Can't register %s\n",
1287                                 table->procname);
1288                         continue;
1289                 }
1290
1291                 len = strlen(table->procname);
1292                 mode = table->mode;
1293
1294                 de = NULL;
1295                 if (table->proc_handler)
1296                         mode |= S_IFREG;
1297                 else {
1298                         mode |= S_IFDIR;
1299                         for (de = root->subdir; de; de = de->next) {
1300                                 if (proc_match(len, table->procname, de))
1301                                         break;
1302                         }
1303                         /* If the subdir exists already, de is non-NULL */
1304                 }
1305
1306                 if (!de) {
1307                         de = create_proc_entry(table->procname, mode, root);
1308                         if (!de)
1309                                 continue;
1310                         de->data = (void *) table;
1311                         if (table->proc_handler)
1312                                 de->proc_fops = &proc_sys_file_operations;
1313                 }
1314                 table->de = de;
1315                 if (de->mode & S_IFDIR)
1316                         register_proc_table(table->child, de);
1317         }
1318 }
1319
1320 /*
1321  * Unregister a /proc sysctl table and any subdirectories.
1322  */
1323 static void unregister_proc_table(ctl_table * table, struct proc_dir_entry *root)
1324 {
1325         struct proc_dir_entry *de;
1326         for (; table->ctl_name; table++) {
1327                 if (!(de = table->de))
1328                         continue;
1329                 if (de->mode & S_IFDIR) {
1330                         if (!table->child) {
1331                                 printk (KERN_ALERT "Help - malformed sysctl tree on free\n");
1332                                 continue;
1333                         }
1334                         unregister_proc_table(table->child, de);
1335
1336                         /* Don't unregister directories which still have entries.. */
1337                         if (de->subdir)
1338                                 continue;
1339                 }
1340
1341                 /* Don't unregister proc entries that are still being used.. */
1342                 if (atomic_read(&de->count))
1343                         continue;
1344
1345                 table->de = NULL;
1346                 remove_proc_entry(table->procname, root);
1347         }
1348 }
1349
1350 static ssize_t do_rw_proc(int write, struct file * file, char __user * buf,
1351                           size_t count, loff_t *ppos)
1352 {
1353         int op;
1354         struct proc_dir_entry *de;
1355         struct ctl_table *table;
1356         size_t res;
1357         ssize_t error;
1358
1359         de = PDE(file->f_dentry->d_inode);
1360         if (!de || !de->data)
1361                 return -ENOTDIR;
1362         table = (struct ctl_table *) de->data;
1363         if (!table || !table->proc_handler)
1364                 return -ENOTDIR;
1365         op = (write ? 002 : 004);
1366         if (ctl_perm(table, op))
1367                 return -EPERM;
1368
1369         res = count;
1370
1371         error = (*table->proc_handler) (table, write, file, buf, &res, ppos);
1372         if (error)
1373                 return error;
1374         return res;
1375 }
1376
1377 static int proc_opensys(struct inode *inode, struct file *file)
1378 {
1379         if (file->f_mode & FMODE_WRITE) {
1380                 /*
1381                  * sysctl entries that are not writable,
1382                  * are _NOT_ writable, capabilities or not.
1383                  */
1384                 if (!(inode->i_mode & S_IWUSR))
1385                         return -EPERM;
1386         }
1387
1388         return 0;
1389 }
1390
1391 static ssize_t proc_readsys(struct file * file, char __user * buf,
1392                             size_t count, loff_t *ppos)
1393 {
1394         return do_rw_proc(0, file, buf, count, ppos);
1395 }
1396
1397 static ssize_t proc_writesys(struct file * file, const char __user * buf,
1398                              size_t count, loff_t *ppos)
1399 {
1400         return do_rw_proc(1, file, (char __user *) buf, count, ppos);
1401 }
1402
1403 /**
1404  * proc_dostring - read a string sysctl
1405  * @table: the sysctl table
1406  * @write: %TRUE if this is a write to the sysctl file
1407  * @filp: the file structure
1408  * @buffer: the user buffer
1409  * @lenp: the size of the user buffer
1410  * @ppos: file position
1411  *
1412  * Reads/writes a string from/to the user buffer. If the kernel
1413  * buffer provided is not large enough to hold the string, the
1414  * string is truncated. The copied string is %NULL-terminated.
1415  * If the string is being read by the user process, it is copied
1416  * and a newline '\n' is added. It is truncated if the buffer is
1417  * not large enough.
1418  *
1419  * Returns 0 on success.
1420  */
1421 int proc_dostring(ctl_table *table, int write, struct file *filp,
1422                   void __user *buffer, size_t *lenp, loff_t *ppos)
1423 {
1424         size_t len;
1425         char __user *p;
1426         char c;
1427
1428         if (!table->data || !table->maxlen || !*lenp ||
1429             (*ppos && !write)) {
1430                 *lenp = 0;
1431                 return 0;
1432         }
1433
1434         if (write) {
1435                 len = 0;
1436                 p = buffer;
1437                 while (len < *lenp) {
1438                         if (get_user(c, p++))
1439                                 return -EFAULT;
1440                         if (c == 0 || c == '\n')
1441                                 break;
1442                         len++;
1443                 }
1444                 if (len >= table->maxlen)
1445                         len = table->maxlen-1;
1446                 if(copy_from_user(table->data, buffer, len))
1447                         return -EFAULT;
1448                 ((char *) table->data)[len] = 0;
1449                 *ppos += *lenp;
1450         } else {
1451                 len = strlen(table->data);
1452                 if (len > table->maxlen)
1453                         len = table->maxlen;
1454                 if (len > *lenp)
1455                         len = *lenp;
1456                 if (len)
1457                         if(copy_to_user(buffer, table->data, len))
1458                                 return -EFAULT;
1459                 if (len < *lenp) {
1460                         if(put_user('\n', ((char __user *) buffer) + len))
1461                                 return -EFAULT;
1462                         len++;
1463                 }
1464                 *lenp = len;
1465                 *ppos += len;
1466         }
1467         return 0;
1468 }
1469
1470 /*
1471  *      Special case of dostring for the UTS structure. This has locks
1472  *      to observe. Should this be in kernel/sys.c ????
1473  */
1474
1475 static int proc_doutsstring(ctl_table *table, int write, struct file *filp,
1476                   void __user *buffer, size_t *lenp, loff_t *ppos)
1477 {
1478         int r;
1479
1480         if (!write) {
1481                 down_read(&uts_sem);
1482                 r=proc_dostring(table,0,filp,buffer,lenp, ppos);
1483                 up_read(&uts_sem);
1484         } else {
1485                 down_write(&uts_sem);
1486                 r=proc_dostring(table,1,filp,buffer,lenp, ppos);
1487                 up_write(&uts_sem);
1488         }
1489         return r;
1490 }
1491
1492 static int do_proc_dointvec_conv(int *negp, unsigned long *lvalp,
1493                                  int *valp,
1494                                  int write, void *data)
1495 {
1496         if (write) {
1497                 *valp = *negp ? -*lvalp : *lvalp;
1498         } else {
1499                 int val = *valp;
1500                 if (val < 0) {
1501                         *negp = -1;
1502                         *lvalp = (unsigned long)-val;
1503                 } else {
1504                         *negp = 0;
1505                         *lvalp = (unsigned long)val;
1506                 }
1507         }
1508         return 0;
1509 }
1510
1511 static int do_proc_dointvec(ctl_table *table, int write, struct file *filp,
1512                   void __user *buffer, size_t *lenp, loff_t *ppos,
1513                   int (*conv)(int *negp, unsigned long *lvalp, int *valp,
1514                               int write, void *data),
1515                   void *data)
1516 {
1517 #define TMPBUFLEN 21
1518         int *i, vleft, first=1, neg, val;
1519         unsigned long lval;
1520         size_t left, len;
1521
1522         char buf[TMPBUFLEN], *p;
1523         char __user *s = buffer;
1524
1525         if (!table->data || !table->maxlen || !*lenp ||
1526             (*ppos && !write)) {
1527                 *lenp = 0;
1528                 return 0;
1529         }
1530
1531         i = (int *) table->data;
1532         vleft = table->maxlen / sizeof(*i);
1533         left = *lenp;
1534
1535         if (!conv)
1536                 conv = do_proc_dointvec_conv;
1537
1538         for (; left && vleft--; i++, first=0) {
1539                 if (write) {
1540                         while (left) {
1541                                 char c;
1542                                 if (get_user(c, s))
1543                                         return -EFAULT;
1544                                 if (!isspace(c))
1545                                         break;
1546                                 left--;
1547                                 s++;
1548                         }
1549                         if (!left)
1550                                 break;
1551                         neg = 0;
1552                         len = left;
1553                         if (len > sizeof(buf) - 1)
1554                                 len = sizeof(buf) - 1;
1555                         if (copy_from_user(buf, s, len))
1556                                 return -EFAULT;
1557                         buf[len] = 0;
1558                         p = buf;
1559                         if (*p == '-' && left > 1) {
1560                                 neg = 1;
1561                                 left--, p++;
1562                         }
1563                         if (*p < '0' || *p > '9')
1564                                 break;
1565
1566                         lval = simple_strtoul(p, &p, 0);
1567
1568                         len = p-buf;
1569                         if ((len < left) && *p && !isspace(*p))
1570                                 break;
1571                         if (neg)
1572                                 val = -val;
1573                         s += len;
1574                         left -= len;
1575
1576                         if (conv(&neg, &lval, i, 1, data))
1577                                 break;
1578                 } else {
1579                         p = buf;
1580                         if (!first)
1581                                 *p++ = '\t';
1582
1583                         if (conv(&neg, &lval, i, 0, data))
1584                                 break;
1585
1586                         sprintf(p, "%s%lu", neg ? "-" : "", lval);
1587                         len = strlen(buf);
1588                         if (len > left)
1589                                 len = left;
1590                         if(copy_to_user(s, buf, len))
1591                                 return -EFAULT;
1592                         left -= len;
1593                         s += len;
1594                 }
1595         }
1596
1597         if (!write && !first && left) {
1598                 if(put_user('\n', s))
1599                         return -EFAULT;
1600                 left--, s++;
1601         }
1602         if (write) {
1603                 while (left) {
1604                         char c;
1605                         if (get_user(c, s++))
1606                                 return -EFAULT;
1607                         if (!isspace(c))
1608                                 break;
1609                         left--;
1610                 }
1611         }
1612         if (write && first)
1613                 return -EINVAL;
1614         *lenp -= left;
1615         *ppos += *lenp;
1616         return 0;
1617 #undef TMPBUFLEN
1618 }
1619
1620 /**
1621  * proc_dointvec - read a vector of integers
1622  * @table: the sysctl table
1623  * @write: %TRUE if this is a write to the sysctl file
1624  * @filp: the file structure
1625  * @buffer: the user buffer
1626  * @lenp: the size of the user buffer
1627  * @ppos: file position
1628  *
1629  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
1630  * values from/to the user buffer, treated as an ASCII string.
1631  *
1632  * Returns 0 on success.
1633  */
1634 int proc_dointvec(ctl_table *table, int write, struct file *filp,
1635                      void __user *buffer, size_t *lenp, loff_t *ppos)
1636 {
1637     return do_proc_dointvec(table,write,filp,buffer,lenp,ppos,
1638                             NULL,NULL);
1639 }
1640
1641 #define OP_SET  0
1642 #define OP_AND  1
1643 #define OP_OR   2
1644 #define OP_MAX  3
1645 #define OP_MIN  4
1646
1647 static int do_proc_dointvec_bset_conv(int *negp, unsigned long *lvalp,
1648                                       int *valp,
1649                                       int write, void *data)
1650 {
1651         int op = *(int *)data;
1652         if (write) {
1653                 int val = *negp ? -*lvalp : *lvalp;
1654                 switch(op) {
1655                 case OP_SET:    *valp = val; break;
1656                 case OP_AND:    *valp &= val; break;
1657                 case OP_OR:     *valp |= val; break;
1658                 case OP_MAX:    if(*valp < val)
1659                                         *valp = val;
1660                                 break;
1661                 case OP_MIN:    if(*valp > val)
1662                                 *valp = val;
1663                                 break;
1664                 }
1665         } else {
1666                 int val = *valp;
1667                 if (val < 0) {
1668                         *negp = -1;
1669                         *lvalp = (unsigned long)-val;
1670                 } else {
1671                         *negp = 0;
1672                         *lvalp = (unsigned long)val;
1673                 }
1674         }
1675         return 0;
1676 }
1677
1678 /*
1679  *      init may raise the set.
1680  */
1681
1682 int proc_dointvec_bset(ctl_table *table, int write, struct file *filp,
1683                         void __user *buffer, size_t *lenp, loff_t *ppos)
1684 {
1685         int op;
1686
1687         if (!capable(CAP_SYS_MODULE)) {
1688                 return -EPERM;
1689         }
1690
1691         op = (current->pid == 1) ? OP_SET : OP_AND;
1692         return do_proc_dointvec(table,write,filp,buffer,lenp,ppos,
1693                                 do_proc_dointvec_bset_conv,&op);
1694 }
1695
1696 struct do_proc_dointvec_minmax_conv_param {
1697         int *min;
1698         int *max;
1699 };
1700
1701 static int do_proc_dointvec_minmax_conv(int *negp, unsigned long *lvalp,
1702                                         int *valp,
1703                                         int write, void *data)
1704 {
1705         struct do_proc_dointvec_minmax_conv_param *param = data;
1706         if (write) {
1707                 int val = *negp ? -*lvalp : *lvalp;
1708                 if ((param->min && *param->min > val) ||
1709                     (param->max && *param->max < val))
1710                         return -EINVAL;
1711                 *valp = val;
1712         } else {
1713                 int val = *valp;
1714                 if (val < 0) {
1715                         *negp = -1;
1716                         *lvalp = (unsigned long)-val;
1717                 } else {
1718                         *negp = 0;
1719                         *lvalp = (unsigned long)val;
1720                 }
1721         }
1722         return 0;
1723 }
1724
1725 /**
1726  * proc_dointvec_minmax - read a vector of integers with min/max values
1727  * @table: the sysctl table
1728  * @write: %TRUE if this is a write to the sysctl file
1729  * @filp: the file structure
1730  * @buffer: the user buffer
1731  * @lenp: the size of the user buffer
1732  * @ppos: file position
1733  *
1734  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
1735  * values from/to the user buffer, treated as an ASCII string.
1736  *
1737  * This routine will ensure the values are within the range specified by
1738  * table->extra1 (min) and table->extra2 (max).
1739  *
1740  * Returns 0 on success.
1741  */
1742 int proc_dointvec_minmax(ctl_table *table, int write, struct file *filp,
1743                   void __user *buffer, size_t *lenp, loff_t *ppos)
1744 {
1745         struct do_proc_dointvec_minmax_conv_param param = {
1746                 .min = (int *) table->extra1,
1747                 .max = (int *) table->extra2,
1748         };
1749         return do_proc_dointvec(table, write, filp, buffer, lenp, ppos,
1750                                 do_proc_dointvec_minmax_conv, &param);
1751 }
1752
1753 static int do_proc_doulongvec_minmax(ctl_table *table, int write,
1754                                      struct file *filp,
1755                                      void __user *buffer,
1756                                      size_t *lenp, loff_t *ppos,
1757                                      unsigned long convmul,
1758                                      unsigned long convdiv)
1759 {
1760 #define TMPBUFLEN 21
1761         unsigned long *i, *min, *max, val;
1762         int vleft, first=1, neg;
1763         size_t len, left;
1764         char buf[TMPBUFLEN], *p;
1765         char __user *s = buffer;
1766
1767         if (!table->data || !table->maxlen || !*lenp ||
1768             (*ppos && !write)) {
1769                 *lenp = 0;
1770                 return 0;
1771         }
1772
1773         i = (unsigned long *) table->data;
1774         min = (unsigned long *) table->extra1;
1775         max = (unsigned long *) table->extra2;
1776         vleft = table->maxlen / sizeof(unsigned long);
1777         left = *lenp;
1778
1779         for (; left && vleft--; i++, min++, max++, first=0) {
1780                 if (write) {
1781                         while (left) {
1782                                 char c;
1783                                 if (get_user(c, s))
1784                                         return -EFAULT;
1785                                 if (!isspace(c))
1786                                         break;
1787                                 left--;
1788                                 s++;
1789                         }
1790                         if (!left)
1791                                 break;
1792                         neg = 0;
1793                         len = left;
1794                         if (len > TMPBUFLEN-1)
1795                                 len = TMPBUFLEN-1;
1796                         if (copy_from_user(buf, s, len))
1797                                 return -EFAULT;
1798                         buf[len] = 0;
1799                         p = buf;
1800                         if (*p == '-' && left > 1) {
1801                                 neg = 1;
1802                                 left--, p++;
1803                         }
1804                         if (*p < '0' || *p > '9')
1805                                 break;
1806                         val = simple_strtoul(p, &p, 0) * convmul / convdiv ;
1807                         len = p-buf;
1808                         if ((len < left) && *p && !isspace(*p))
1809                                 break;
1810                         if (neg)
1811                                 val = -val;
1812                         s += len;
1813                         left -= len;
1814
1815                         if(neg)
1816                                 continue;
1817                         if ((min && val < *min) || (max && val > *max))
1818                                 continue;
1819                         *i = val;
1820                 } else {
1821                         p = buf;
1822                         if (!first)
1823                                 *p++ = '\t';
1824                         sprintf(p, "%lu", convdiv * (*i) / convmul);
1825                         len = strlen(buf);
1826                         if (len > left)
1827                                 len = left;
1828                         if(copy_to_user(s, buf, len))
1829                                 return -EFAULT;
1830                         left -= len;
1831                         s += len;
1832                 }
1833         }
1834
1835         if (!write && !first && left) {
1836                 if(put_user('\n', s))
1837                         return -EFAULT;
1838                 left--, s++;
1839         }
1840         if (write) {
1841                 while (left) {
1842                         char c;
1843                         if (get_user(c, s++))
1844                                 return -EFAULT;
1845                         if (!isspace(c))
1846                                 break;
1847                         left--;
1848                 }
1849         }
1850         if (write && first)
1851                 return -EINVAL;
1852         *lenp -= left;
1853         *ppos += *lenp;
1854         return 0;
1855 #undef TMPBUFLEN
1856 }
1857
1858 /**
1859  * proc_doulongvec_minmax - read a vector of long integers with min/max values
1860  * @table: the sysctl table
1861  * @write: %TRUE if this is a write to the sysctl file
1862  * @filp: the file structure
1863  * @buffer: the user buffer
1864  * @lenp: the size of the user buffer
1865  * @ppos: file position
1866  *
1867  * Reads/writes up to table->maxlen/sizeof(unsigned long) unsigned long
1868  * values from/to the user buffer, treated as an ASCII string.
1869  *
1870  * This routine will ensure the values are within the range specified by
1871  * table->extra1 (min) and table->extra2 (max).
1872  *
1873  * Returns 0 on success.
1874  */
1875 int proc_doulongvec_minmax(ctl_table *table, int write, struct file *filp,
1876                            void __user *buffer, size_t *lenp, loff_t *ppos)
1877 {
1878     return do_proc_doulongvec_minmax(table, write, filp, buffer, lenp, ppos, 1l, 1l);
1879 }
1880
1881 /**
1882  * proc_doulongvec_ms_jiffies_minmax - read a vector of millisecond values with min/max values
1883  * @table: the sysctl table
1884  * @write: %TRUE if this is a write to the sysctl file
1885  * @filp: the file structure
1886  * @buffer: the user buffer
1887  * @lenp: the size of the user buffer
1888  * @ppos: file position
1889  *
1890  * Reads/writes up to table->maxlen/sizeof(unsigned long) unsigned long
1891  * values from/to the user buffer, treated as an ASCII string. The values
1892  * are treated as milliseconds, and converted to jiffies when they are stored.
1893  *
1894  * This routine will ensure the values are within the range specified by
1895  * table->extra1 (min) and table->extra2 (max).
1896  *
1897  * Returns 0 on success.
1898  */
1899 int proc_doulongvec_ms_jiffies_minmax(ctl_table *table, int write,
1900                                       struct file *filp,
1901                                       void __user *buffer,
1902                                       size_t *lenp, loff_t *ppos)
1903 {
1904     return do_proc_doulongvec_minmax(table, write, filp, buffer,
1905                                      lenp, ppos, HZ, 1000l);
1906 }
1907
1908
1909 static int do_proc_dointvec_jiffies_conv(int *negp, unsigned long *lvalp,
1910                                          int *valp,
1911                                          int write, void *data)
1912 {
1913         if (write) {
1914                 *valp = *negp ? -(*lvalp*HZ) : (*lvalp*HZ);
1915         } else {
1916                 int val = *valp;
1917                 unsigned long lval;
1918                 if (val < 0) {
1919                         *negp = -1;
1920                         lval = (unsigned long)-val;
1921                 } else {
1922                         *negp = 0;
1923                         lval = (unsigned long)val;
1924                 }
1925                 *lvalp = lval / HZ;
1926         }
1927         return 0;
1928 }
1929
1930 static int do_proc_dointvec_userhz_jiffies_conv(int *negp, unsigned long *lvalp,
1931                                                 int *valp,
1932                                                 int write, void *data)
1933 {
1934         if (write) {
1935                 *valp = clock_t_to_jiffies(*negp ? -*lvalp : *lvalp);
1936         } else {
1937                 int val = *valp;
1938                 unsigned long lval;
1939                 if (val < 0) {
1940                         *negp = -1;
1941                         lval = (unsigned long)-val;
1942                 } else {
1943                         *negp = 0;
1944                         lval = (unsigned long)val;
1945                 }
1946                 *lvalp = jiffies_to_clock_t(lval);
1947         }
1948         return 0;
1949 }
1950
1951 static int do_proc_dointvec_ms_jiffies_conv(int *negp, unsigned long *lvalp,
1952                                             int *valp,
1953                                             int write, void *data)
1954 {
1955         if (write) {
1956                 *valp = msecs_to_jiffies(*negp ? -*lvalp : *lvalp);
1957         } else {
1958                 int val = *valp;
1959                 unsigned long lval;
1960                 if (val < 0) {
1961                         *negp = -1;
1962                         lval = (unsigned long)-val;
1963                 } else {
1964                         *negp = 0;
1965                         lval = (unsigned long)val;
1966                 }
1967                 *lvalp = jiffies_to_msecs(lval);
1968         }
1969         return 0;
1970 }
1971
1972 /**
1973  * proc_dointvec_jiffies - read a vector of integers as seconds
1974  * @table: the sysctl table
1975  * @write: %TRUE if this is a write to the sysctl file
1976  * @filp: the file structure
1977  * @buffer: the user buffer
1978  * @lenp: the size of the user buffer
1979  * @ppos: file position
1980  *
1981  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
1982  * values from/to the user buffer, treated as an ASCII string.
1983  * The values read are assumed to be in seconds, and are converted into
1984  * jiffies.
1985  *
1986  * Returns 0 on success.
1987  */
1988 int proc_dointvec_jiffies(ctl_table *table, int write, struct file *filp,
1989                           void __user *buffer, size_t *lenp, loff_t *ppos)
1990 {
1991     return do_proc_dointvec(table,write,filp,buffer,lenp,ppos,
1992                             do_proc_dointvec_jiffies_conv,NULL);
1993 }
1994
1995 /**
1996  * proc_dointvec_userhz_jiffies - read a vector of integers as 1/USER_HZ seconds
1997  * @table: the sysctl table
1998  * @write: %TRUE if this is a write to the sysctl file
1999  * @filp: the file structure
2000  * @buffer: the user buffer
2001  * @lenp: the size of the user buffer
2002  *
2003  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2004  * values from/to the user buffer, treated as an ASCII string.
2005  * The values read are assumed to be in 1/USER_HZ seconds, and
2006  * are converted into jiffies.
2007  *
2008  * Returns 0 on success.
2009  */
2010 int proc_dointvec_userhz_jiffies(ctl_table *table, int write, struct file *filp,
2011                                  void __user *buffer, size_t *lenp, loff_t *ppos)
2012 {
2013     return do_proc_dointvec(table,write,filp,buffer,lenp,ppos,
2014                             do_proc_dointvec_userhz_jiffies_conv,NULL);
2015 }
2016
2017 /**
2018  * proc_dointvec_ms_jiffies - read a vector of integers as 1 milliseconds
2019  * @table: the sysctl table
2020  * @write: %TRUE if this is a write to the sysctl file
2021  * @filp: the file structure
2022  * @buffer: the user buffer
2023  * @lenp: the size of the user buffer
2024  * @ppos: file position
2025  * @ppos: the current position in the file
2026  *
2027  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2028  * values from/to the user buffer, treated as an ASCII string.
2029  * The values read are assumed to be in 1/1000 seconds, and
2030  * are converted into jiffies.
2031  *
2032  * Returns 0 on success.
2033  */
2034 int proc_dointvec_ms_jiffies(ctl_table *table, int write, struct file *filp,
2035                              void __user *buffer, size_t *lenp, loff_t *ppos)
2036 {
2037         return do_proc_dointvec(table, write, filp, buffer, lenp, ppos,
2038                                 do_proc_dointvec_ms_jiffies_conv, NULL);
2039 }
2040
2041 #else /* CONFIG_PROC_FS */
2042
2043 int proc_dostring(ctl_table *table, int write, struct file *filp,
2044                   void __user *buffer, size_t *lenp, loff_t *ppos)
2045 {
2046         return -ENOSYS;
2047 }
2048
2049 static int proc_doutsstring(ctl_table *table, int write, struct file *filp,
2050                             void __user *buffer, size_t *lenp, loff_t *ppos)
2051 {
2052         return -ENOSYS;
2053 }
2054
2055 int proc_dointvec(ctl_table *table, int write, struct file *filp,
2056                   void __user *buffer, size_t *lenp, loff_t *ppos)
2057 {
2058         return -ENOSYS;
2059 }
2060
2061 int proc_dointvec_bset(ctl_table *table, int write, struct file *filp,
2062                         void __user *buffer, size_t *lenp, loff_t *ppos)
2063 {
2064         return -ENOSYS;
2065 }
2066
2067 int proc_dointvec_minmax(ctl_table *table, int write, struct file *filp,
2068                     void __user *buffer, size_t *lenp, loff_t *ppos)
2069 {
2070         return -ENOSYS;
2071 }
2072
2073 int proc_dointvec_jiffies(ctl_table *table, int write, struct file *filp,
2074                     void __user *buffer, size_t *lenp, loff_t *ppos)
2075 {
2076         return -ENOSYS;
2077 }
2078
2079 int proc_dointvec_userhz_jiffies(ctl_table *table, int write, struct file *filp,
2080                     void __user *buffer, size_t *lenp, loff_t *ppos)
2081 {
2082         return -ENOSYS;
2083 }
2084
2085 int proc_dointvec_ms_jiffies(ctl_table *table, int write, struct file *filp,
2086                              void __user *buffer, size_t *lenp, loff_t *ppos)
2087 {
2088         return -ENOSYS;
2089 }
2090
2091 int proc_doulongvec_minmax(ctl_table *table, int write, struct file *filp,
2092                     void __user *buffer, size_t *lenp, loff_t *ppos)
2093 {
2094         return -ENOSYS;
2095 }
2096
2097 int proc_doulongvec_ms_jiffies_minmax(ctl_table *table, int write,
2098                                       struct file *filp,
2099                                       void __user *buffer,
2100                                       size_t *lenp, loff_t *ppos)
2101 {
2102     return -ENOSYS;
2103 }
2104
2105
2106 #endif /* CONFIG_PROC_FS */
2107
2108
2109 /*
2110  * General sysctl support routines
2111  */
2112
2113 /* The generic string strategy routine: */
2114 int sysctl_string(ctl_table *table, int __user *name, int nlen,
2115                   void __user *oldval, size_t __user *oldlenp,
2116                   void __user *newval, size_t newlen, void **context)
2117 {
2118         size_t l, len;
2119
2120         if (!table->data || !table->maxlen)
2121                 return -ENOTDIR;
2122
2123         if (oldval && oldlenp) {
2124                 if (get_user(len, oldlenp))
2125                         return -EFAULT;
2126                 if (len) {
2127                         l = strlen(table->data);
2128                         if (len > l) len = l;
2129                         if (len >= table->maxlen)
2130                                 len = table->maxlen;
2131                         if(copy_to_user(oldval, table->data, len))
2132                                 return -EFAULT;
2133                         if(put_user(0, ((char __user *) oldval) + len))
2134                                 return -EFAULT;
2135                         if(put_user(len, oldlenp))
2136                                 return -EFAULT;
2137                 }
2138         }
2139         if (newval && newlen) {
2140                 len = newlen;
2141                 if (len > table->maxlen)
2142                         len = table->maxlen;
2143                 if(copy_from_user(table->data, newval, len))
2144                         return -EFAULT;
2145                 if (len == table->maxlen)
2146                         len--;
2147                 ((char *) table->data)[len] = 0;
2148         }
2149         return 0;
2150 }
2151
2152 /*
2153  * This function makes sure that all of the integers in the vector
2154  * are between the minimum and maximum values given in the arrays
2155  * table->extra1 and table->extra2, respectively.
2156  */
2157 int sysctl_intvec(ctl_table *table, int __user *name, int nlen,
2158                 void __user *oldval, size_t __user *oldlenp,
2159                 void __user *newval, size_t newlen, void **context)
2160 {
2161
2162         if (newval && newlen) {
2163                 int __user *vec = (int __user *) newval;
2164                 int *min = (int *) table->extra1;
2165                 int *max = (int *) table->extra2;
2166                 size_t length;
2167                 int i;
2168
2169                 if (newlen % sizeof(int) != 0)
2170                         return -EINVAL;
2171
2172                 if (!table->extra1 && !table->extra2)
2173                         return 0;
2174
2175                 if (newlen > table->maxlen)
2176                         newlen = table->maxlen;
2177                 length = newlen / sizeof(int);
2178
2179                 for (i = 0; i < length; i++) {
2180                         int value;
2181                         if (get_user(value, vec + i))
2182                                 return -EFAULT;
2183                         if (min && value < min[i])
2184                                 return -EINVAL;
2185                         if (max && value > max[i])
2186                                 return -EINVAL;
2187                 }
2188         }
2189         return 0;
2190 }
2191
2192 /* Strategy function to convert jiffies to seconds */
2193 int sysctl_jiffies(ctl_table *table, int __user *name, int nlen,
2194                 void __user *oldval, size_t __user *oldlenp,
2195                 void __user *newval, size_t newlen, void **context)
2196 {
2197         if (oldval) {
2198                 size_t olen;
2199                 if (oldlenp) {
2200                         if (get_user(olen, oldlenp))
2201                                 return -EFAULT;
2202                         if (olen!=sizeof(int))
2203                                 return -EINVAL;
2204                 }
2205                 if (put_user(*(int *)(table->data)/HZ, (int __user *)oldval) ||
2206                     (oldlenp && put_user(sizeof(int),oldlenp)))
2207                         return -EFAULT;
2208         }
2209         if (newval && newlen) {
2210                 int new;
2211                 if (newlen != sizeof(int))
2212                         return -EINVAL;
2213                 if (get_user(new, (int __user *)newval))
2214                         return -EFAULT;
2215                 *(int *)(table->data) = new*HZ;
2216         }
2217         return 1;
2218 }
2219
2220 /* Strategy function to convert jiffies to seconds */
2221 int sysctl_ms_jiffies(ctl_table *table, int __user *name, int nlen,
2222                 void __user *oldval, size_t __user *oldlenp,
2223                 void __user *newval, size_t newlen, void **context)
2224 {
2225         if (oldval) {
2226                 size_t olen;
2227                 if (oldlenp) {
2228                         if (get_user(olen, oldlenp))
2229                                 return -EFAULT;
2230                         if (olen!=sizeof(int))
2231                                 return -EINVAL;
2232                 }
2233                 if (put_user(jiffies_to_msecs(*(int *)(table->data)), (int __user *)oldval) ||
2234                     (oldlenp && put_user(sizeof(int),oldlenp)))
2235                         return -EFAULT;
2236         }
2237         if (newval && newlen) {
2238                 int new;
2239                 if (newlen != sizeof(int))
2240                         return -EINVAL;
2241                 if (get_user(new, (int __user *)newval))
2242                         return -EFAULT;
2243                 *(int *)(table->data) = msecs_to_jiffies(new);
2244         }
2245         return 1;
2246 }
2247
2248 #else /* CONFIG_SYSCTL */
2249
2250
2251 asmlinkage long sys_sysctl(struct __sysctl_args __user *args)
2252 {
2253         return -ENOSYS;
2254 }
2255
2256 int sysctl_string(ctl_table *table, int __user *name, int nlen,
2257                   void __user *oldval, size_t __user *oldlenp,
2258                   void __user *newval, size_t newlen, void **context)
2259 {
2260         return -ENOSYS;
2261 }
2262
2263 int sysctl_intvec(ctl_table *table, int __user *name, int nlen,
2264                 void __user *oldval, size_t __user *oldlenp,
2265                 void __user *newval, size_t newlen, void **context)
2266 {
2267         return -ENOSYS;
2268 }
2269
2270 int sysctl_jiffies(ctl_table *table, int __user *name, int nlen,
2271                 void __user *oldval, size_t __user *oldlenp,
2272                 void __user *newval, size_t newlen, void **context)
2273 {
2274         return -ENOSYS;
2275 }
2276
2277 int sysctl_ms_jiffies(ctl_table *table, int __user *name, int nlen,
2278                 void __user *oldval, size_t __user *oldlenp,
2279                 void __user *newval, size_t newlen, void **context)
2280 {
2281         return -ENOSYS;
2282 }
2283
2284 int proc_dostring(ctl_table *table, int write, struct file *filp,
2285                   void __user *buffer, size_t *lenp, loff_t *ppos)
2286 {
2287         return -ENOSYS;
2288 }
2289
2290 int proc_dointvec(ctl_table *table, int write, struct file *filp,
2291                   void __user *buffer, size_t *lenp, loff_t *ppos)
2292 {
2293         return -ENOSYS;
2294 }
2295
2296 int proc_dointvec_bset(ctl_table *table, int write, struct file *filp,
2297                         void __user *buffer, size_t *lenp, loff_t *ppos)
2298 {
2299         return -ENOSYS;
2300 }
2301
2302 int proc_dointvec_minmax(ctl_table *table, int write, struct file *filp,
2303                     void __user *buffer, size_t *lenp, loff_t *ppos)
2304 {
2305         return -ENOSYS;
2306 }
2307
2308 int proc_dointvec_jiffies(ctl_table *table, int write, struct file *filp,
2309                           void __user *buffer, size_t *lenp, loff_t *ppos)
2310 {
2311         return -ENOSYS;
2312 }
2313
2314 int proc_dointvec_userhz_jiffies(ctl_table *table, int write, struct file *filp,
2315                           void __user *buffer, size_t *lenp, loff_t *ppos)
2316 {
2317         return -ENOSYS;
2318 }
2319
2320 int proc_dointvec_ms_jiffies(ctl_table *table, int write, struct file *filp,
2321                              void __user *buffer, size_t *lenp, loff_t *ppos)
2322 {
2323         return -ENOSYS;
2324 }
2325
2326 int proc_doulongvec_minmax(ctl_table *table, int write, struct file *filp,
2327                     void __user *buffer, size_t *lenp, loff_t *ppos)
2328 {
2329         return -ENOSYS;
2330 }
2331
2332 int proc_doulongvec_ms_jiffies_minmax(ctl_table *table, int write,
2333                                       struct file *filp,
2334                                       void __user *buffer,
2335                                       size_t *lenp, loff_t *ppos)
2336 {
2337     return -ENOSYS;
2338 }
2339
2340 struct ctl_table_header * register_sysctl_table(ctl_table * table,
2341                                                 int insert_at_head)
2342 {
2343         return NULL;
2344 }
2345
2346 void unregister_sysctl_table(struct ctl_table_header * table)
2347 {
2348 }
2349
2350 #endif /* CONFIG_SYSCTL */
2351
2352 /*
2353  * No sense putting this after each symbol definition, twice,
2354  * exception granted :-)
2355  */
2356 EXPORT_SYMBOL(proc_dointvec);
2357 EXPORT_SYMBOL(proc_dointvec_jiffies);
2358 EXPORT_SYMBOL(proc_dointvec_minmax);
2359 EXPORT_SYMBOL(proc_dointvec_userhz_jiffies);
2360 EXPORT_SYMBOL(proc_dointvec_ms_jiffies);
2361 EXPORT_SYMBOL(proc_dostring);
2362 EXPORT_SYMBOL(proc_doulongvec_minmax);
2363 EXPORT_SYMBOL(proc_doulongvec_ms_jiffies_minmax);
2364 EXPORT_SYMBOL(register_sysctl_table);
2365 EXPORT_SYMBOL(sysctl_intvec);
2366 EXPORT_SYMBOL(sysctl_jiffies);
2367 EXPORT_SYMBOL(sysctl_ms_jiffies);
2368 EXPORT_SYMBOL(sysctl_string);
2369 EXPORT_SYMBOL(unregister_sysctl_table);