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/kobject.h>
  35 #include <linux/net.h>
  36 #include <linux/sysrq.h>
  37 #include <linux/highuid.h>
  38 #include <linux/writeback.h>
  39 #include <linux/hugetlb.h>
  40 #include <linux/security.h>
  41 #include <linux/initrd.h>
  42 #include <linux/times.h>
  43 #include <linux/limits.h>
  44 #include <linux/dcache.h>
  45 #include <linux/syscalls.h>
  46
  47 #include <asm/uaccess.h>
  48 #include <asm/processor.h>
  49
  50 #ifdef CONFIG_ROOT_NFS
  51 #include <linux/nfs_fs.h>
  52 #endif
  53
  54 #if defined(CONFIG_SYSCTL)
  55
  56 /* External variables not in a header file. */
  57 extern int C_A_D;
  58 extern int sysctl_overcommit_memory;
  59 extern int sysctl_overcommit_ratio;
  60 extern int max_threads;
  61 extern int sysrq_enabled;
  62 extern int core_uses_pid;
  63 extern int suid_dumpable;
  64 extern char core_pattern[];
  65 extern int cad_pid;
  66 extern int pid_max;
  67 extern int min_free_kbytes;
  68 extern int printk_ratelimit_jiffies;
  69 extern int printk_ratelimit_burst;
  70 extern int pid_max_min, pid_max_max;
  71
  72 #if defined(CONFIG_X86_LOCAL_APIC) && defined(CONFIG_X86)
  73 int unknown_nmi_panic;
  74 extern int proc_unknown_nmi_panic(ctl_table *, int, struct file *,
  75                                   void __user *, size_t *, loff_t *);
  76 #endif
  77
  78 /* this is needed for the proc_dointvec_minmax for [fs_]overflow UID and GID */
  79 static int maxolduid = 65535;
  80 static int minolduid;
  81
  82 static int ngroups_max = NGROUPS_MAX;
  83
  84 #ifdef CONFIG_KMOD
  85 extern char modprobe_path[];
  86 #endif
  87 #ifdef CONFIG_CHR_DEV_SG
  88 extern int sg_big_buff;
  89 #endif
  90 #ifdef CONFIG_SYSVIPC
  91 extern size_t shm_ctlmax;
  92 extern size_t shm_ctlall;
  93 extern int shm_ctlmni;
  94 extern int msg_ctlmax;
  95 extern int msg_ctlmnb;
  96 extern int msg_ctlmni;
  97 extern int sem_ctls[];
  98 #endif
  99
 100 #ifdef __sparc__
 101 extern char reboot_command [];
 102 extern int stop_a_enabled;
 103 extern int scons_pwroff;
 104 #endif
 105
 106 #ifdef __hppa__
 107 extern int pwrsw_enabled;
 108 extern int unaligned_enabled;
 109 #endif
 110
 111 #ifdef CONFIG_ARCH_S390
 112 #ifdef CONFIG_MATHEMU
 113 extern int sysctl_ieee_emulation_warnings;
 114 #endif
 115 extern int sysctl_userprocess_debug;
 116 extern int spin_retry;
 117 #endif
 118
 119 extern int sysctl_hz_timer;
 120
 121 #ifdef CONFIG_BSD_PROCESS_ACCT
 122 extern int acct_parm[];
 123 #endif
 124
 125 int randomize_va_space = 1;
 126
 127 static int parse_table(int __user *, int, void __user *, size_t __user *, void __user *, size_t,
 128                        ctl_table *, void **);
 129 static int proc_doutsstring(ctl_table *table, int write, struct file *filp,
 130                   void __user *buffer, size_t *lenp, loff_t *ppos);
 131
 132 static ctl_table root_table[];
 133 static struct ctl_table_header root_table_header =
 134         { root_table, LIST_HEAD_INIT(root_table_header.ctl_entry) };
 135
 136 static ctl_table kern_table[];
 137 static ctl_table vm_table[];
 138 static ctl_table proc_table[];
 139 static ctl_table fs_table[];
 140 static ctl_table debug_table[];
 141 static ctl_table dev_table[];
 142 extern ctl_table random_table[];
 143 #ifdef CONFIG_UNIX98_PTYS
 144 extern ctl_table pty_table[];
 145 #endif
 146 #ifdef CONFIG_INOTIFY
 147 extern ctl_table inotify_table[];
 148 #endif
 149
 150 #ifdef HAVE_ARCH_PICK_MMAP_LAYOUT
 151 int sysctl_legacy_va_layout;
 152 #endif
 153
 154 /* /proc declarations: */
 155
 156 #ifdef CONFIG_PROC_FS
 157
 158 static ssize_t proc_readsys(struct file *, char __user *, size_t, loff_t *);
 159 static ssize_t proc_writesys(struct file *, const char __user *, size_t, loff_t *);
 160 static int proc_opensys(struct inode *, struct file *);
 161
 162 struct file_operations proc_sys_file_operations = {
 163         .open           = proc_opensys,
 164         .read           = proc_readsys,
 165         .write          = proc_writesys,
 166 };
 167
 168 extern struct proc_dir_entry *proc_sys_root;
 169
 170 static void register_proc_table(ctl_table *, struct proc_dir_entry *, void *);
 171 static void unregister_proc_table(ctl_table *, struct proc_dir_entry *);
 172 #endif
 173
 174 /* The default sysctl tables: */
 175
 176 static ctl_table root_table[] = {
 177         {
 178                 .ctl_name       = CTL_KERN,
 179                 .procname       = "kernel",
 180                 .mode           = 0555,
 181                 .child          = kern_table,
 182         },
 183         {
 184                 .ctl_name       = CTL_VM,
 185                 .procname       = "vm",
 186                 .mode           = 0555,
 187                 .child          = vm_table,
 188         },
 189 #ifdef CONFIG_NET
 190         {
 191                 .ctl_name       = CTL_NET,
 192                 .procname       = "net",
 193                 .mode           = 0555,
 194                 .child          = net_table,
 195         },
 196 #endif
 197         {
 198                 .ctl_name       = CTL_PROC,
 199                 .procname       = "proc",
 200                 .mode           = 0555,
 201                 .child          = proc_table,
 202         },
 203         {
 204                 .ctl_name       = CTL_FS,
 205                 .procname       = "fs",
 206                 .mode           = 0555,
 207                 .child          = fs_table,
 208         },
 209         {
 210                 .ctl_name       = CTL_DEBUG,
 211                 .procname       = "debug",
 212                 .mode           = 0555,
 213                 .child          = debug_table,
 214         },
 215         {
 216                 .ctl_name       = CTL_DEV,
 217                 .procname       = "dev",
 218                 .mode           = 0555,
 219                 .child          = dev_table,
 220         },
 221
 222         { .ctl_name = 0 }
 223 };
 224
 225 static ctl_table kern_table[] = {
 226         {
 227                 .ctl_name       = KERN_OSTYPE,
 228                 .procname       = "ostype",
 229                 .data           = system_utsname.sysname,
 230                 .maxlen         = sizeof(system_utsname.sysname),
 231                 .mode           = 0444,
 232                 .proc_handler   = &proc_doutsstring,
 233                 .strategy       = &sysctl_string,
 234         },
 235         {
 236                 .ctl_name       = KERN_OSRELEASE,
 237                 .procname       = "osrelease",
 238                 .data           = system_utsname.release,
 239                 .maxlen         = sizeof(system_utsname.release),
 240                 .mode           = 0444,
 241                 .proc_handler   = &proc_doutsstring,
 242                 .strategy       = &sysctl_string,
 243         },
 244         {
 245                 .ctl_name       = KERN_VERSION,
 246                 .procname       = "version",
 247                 .data           = system_utsname.version,
 248                 .maxlen         = sizeof(system_utsname.version),
 249                 .mode           = 0444,
 250                 .proc_handler   = &proc_doutsstring,
 251                 .strategy       = &sysctl_string,
 252         },
 253         {
 254                 .ctl_name       = KERN_NODENAME,
 255                 .procname       = "hostname",
 256                 .data           = system_utsname.nodename,
 257                 .maxlen         = sizeof(system_utsname.nodename),
 258                 .mode           = 0644,
 259                 .proc_handler   = &proc_doutsstring,
 260                 .strategy       = &sysctl_string,
 261         },
 262         {
 263                 .ctl_name       = KERN_DOMAINNAME,
 264                 .procname       = "domainname",
 265                 .data           = system_utsname.domainname,
 266                 .maxlen         = sizeof(system_utsname.domainname),
 267                 .mode           = 0644,
 268                 .proc_handler   = &proc_doutsstring,
 269                 .strategy       = &sysctl_string,
 270         },
 271         {
 272                 .ctl_name       = KERN_PANIC,
 273                 .procname       = "panic",
 274                 .data           = &panic_timeout,
 275                 .maxlen         = sizeof(int),
 276                 .mode           = 0644,
 277                 .proc_handler   = &proc_dointvec,
 278         },
 279         {
 280                 .ctl_name       = KERN_CORE_USES_PID,
 281                 .procname       = "core_uses_pid",
 282                 .data           = &core_uses_pid,
 283                 .maxlen         = sizeof(int),
 284                 .mode           = 0644,
 285                 .proc_handler   = &proc_dointvec,
 286         },
 287         {
 288                 .ctl_name       = KERN_CORE_PATTERN,
 289                 .procname       = "core_pattern",
 290                 .data           = core_pattern,
 291                 .maxlen         = 64,
 292                 .mode           = 0644,
 293                 .proc_handler   = &proc_dostring,
 294                 .strategy       = &sysctl_string,
 295         },
 296         {
 297                 .ctl_name       = KERN_TAINTED,
 298                 .procname       = "tainted",
 299                 .data           = &tainted,
 300                 .maxlen         = sizeof(int),
 301                 .mode           = 0444,
 302                 .proc_handler   = &proc_dointvec,
 303         },
 304         {
 305                 .ctl_name       = KERN_CAP_BSET,
 306                 .procname       = "cap-bound",
 307                 .data           = &cap_bset,
 308                 .maxlen         = sizeof(kernel_cap_t),
 309                 .mode           = 0600,
 310                 .proc_handler   = &proc_dointvec_bset,
 311         },
 312 #ifdef CONFIG_BLK_DEV_INITRD
 313         {
 314                 .ctl_name       = KERN_REALROOTDEV,
 315                 .procname       = "real-root-dev",
 316                 .data           = &real_root_dev,
 317                 .maxlen         = sizeof(int),
 318                 .mode           = 0644,
 319                 .proc_handler   = &proc_dointvec,
 320         },
 321 #endif
 322 #ifdef __sparc__
 323         {
 324                 .ctl_name       = KERN_SPARC_REBOOT,
 325                 .procname       = "reboot-cmd",
 326                 .data           = reboot_command,
 327                 .maxlen         = 256,
 328                 .mode           = 0644,
 329                 .proc_handler   = &proc_dostring,
 330                 .strategy       = &sysctl_string,
 331         },
 332         {
 333                 .ctl_name       = KERN_SPARC_STOP_A,
 334                 .procname       = "stop-a",
 335                 .data           = &stop_a_enabled,
 336                 .maxlen         = sizeof (int),
 337                 .mode           = 0644,
 338                 .proc_handler   = &proc_dointvec,
 339         },
 340         {
 341                 .ctl_name       = KERN_SPARC_SCONS_PWROFF,
 342                 .procname       = "scons-poweroff",
 343                 .data           = &scons_pwroff,
 344                 .maxlen         = sizeof (int),
 345                 .mode           = 0644,
 346                 .proc_handler   = &proc_dointvec,
 347         },
 348 #endif
 349 #ifdef __hppa__
 350         {
 351                 .ctl_name       = KERN_HPPA_PWRSW,
 352                 .procname       = "soft-power",
 353                 .data           = &pwrsw_enabled,
 354                 .maxlen         = sizeof (int),
 355                 .mode           = 0644,
 356                 .proc_handler   = &proc_dointvec,
 357         },
 358         {
 359                 .ctl_name       = KERN_HPPA_UNALIGNED,
 360                 .procname       = "unaligned-trap",
 361                 .data           = &unaligned_enabled,
 362                 .maxlen         = sizeof (int),
 363                 .mode           = 0644,
 364                 .proc_handler   = &proc_dointvec,
 365         },
 366 #endif
 367         {
 368                 .ctl_name       = KERN_CTLALTDEL,
 369                 .procname       = "ctrl-alt-del",
 370                 .data           = &C_A_D,
 371                 .maxlen         = sizeof(int),
 372                 .mode           = 0644,
 373                 .proc_handler   = &proc_dointvec,
 374         },
 375         {
 376                 .ctl_name       = KERN_PRINTK,
 377                 .procname       = "printk",
 378                 .data           = &console_loglevel,
 379                 .maxlen         = 4*sizeof(int),
 380                 .mode           = 0644,
 381                 .proc_handler   = &proc_dointvec,
 382         },
 383 #ifdef CONFIG_KMOD
 384         {
 385                 .ctl_name       = KERN_MODPROBE,
 386                 .procname       = "modprobe",
 387                 .data           = &modprobe_path,
 388                 .maxlen         = KMOD_PATH_LEN,
 389                 .mode           = 0644,
 390                 .proc_handler   = &proc_dostring,
 391                 .strategy       = &sysctl_string,
 392         },
 393 #endif
 394 #ifdef CONFIG_HOTPLUG
 395         {
 396                 .ctl_name       = KERN_HOTPLUG,
 397                 .procname       = "hotplug",
 398                 .data           = &uevent_helper,
 399                 .maxlen         = UEVENT_HELPER_PATH_LEN,
 400                 .mode           = 0644,
 401                 .proc_handler   = &proc_dostring,
 402                 .strategy       = &sysctl_string,
 403         },
 404 #endif
 405 #ifdef CONFIG_CHR_DEV_SG
 406         {
 407                 .ctl_name       = KERN_SG_BIG_BUFF,
 408                 .procname       = "sg-big-buff",
 409                 .data           = &sg_big_buff,
 410                 .maxlen         = sizeof (int),
 411                 .mode           = 0444,
 412                 .proc_handler   = &proc_dointvec,
 413         },
 414 #endif
 415 #ifdef CONFIG_BSD_PROCESS_ACCT
 416         {
 417                 .ctl_name       = KERN_ACCT,
 418                 .procname       = "acct",
 419                 .data           = &acct_parm,
 420                 .maxlen         = 3*sizeof(int),
 421                 .mode           = 0644,
 422                 .proc_handler   = &proc_dointvec,
 423         },
 424 #endif
 425 #ifdef CONFIG_SYSVIPC
 426         {
 427                 .ctl_name       = KERN_SHMMAX,
 428                 .procname       = "shmmax",
 429                 .data           = &shm_ctlmax,
 430                 .maxlen         = sizeof (size_t),
 431                 .mode           = 0644,
 432                 .proc_handler   = &proc_doulongvec_minmax,
 433         },
 434         {
 435                 .ctl_name       = KERN_SHMALL,
 436                 .procname       = "shmall",
 437                 .data           = &shm_ctlall,
 438                 .maxlen         = sizeof (size_t),
 439                 .mode           = 0644,
 440                 .proc_handler   = &proc_doulongvec_minmax,
 441         },
 442         {
 443                 .ctl_name       = KERN_SHMMNI,
 444                 .procname       = "shmmni",
 445                 .data           = &shm_ctlmni,
 446                 .maxlen         = sizeof (int),
 447                 .mode           = 0644,
 448                 .proc_handler   = &proc_dointvec,
 449         },
 450         {
 451                 .ctl_name       = KERN_MSGMAX,
 452                 .procname       = "msgmax",
 453                 .data           = &msg_ctlmax,
 454                 .maxlen         = sizeof (int),
 455                 .mode           = 0644,
 456                 .proc_handler   = &proc_dointvec,
 457         },
 458         {
 459                 .ctl_name       = KERN_MSGMNI,
 460                 .procname       = "msgmni",
 461                 .data           = &msg_ctlmni,
 462                 .maxlen         = sizeof (int),
 463                 .mode           = 0644,
 464                 .proc_handler   = &proc_dointvec,
 465         },
 466         {
 467                 .ctl_name       = KERN_MSGMNB,
 468                 .procname       =  "msgmnb",
 469                 .data           = &msg_ctlmnb,
 470                 .maxlen         = sizeof (int),
 471                 .mode           = 0644,
 472                 .proc_handler   = &proc_dointvec,
 473         },
 474         {
 475                 .ctl_name       = KERN_SEM,
 476                 .procname       = "sem",
 477                 .data           = &sem_ctls,
 478                 .maxlen         = 4*sizeof (int),
 479                 .mode           = 0644,
 480                 .proc_handler   = &proc_dointvec,
 481         },
 482 #endif
 483 #ifdef CONFIG_MAGIC_SYSRQ
 484         {
 485                 .ctl_name       = KERN_SYSRQ,
 486                 .procname       = "sysrq",
 487                 .data           = &sysrq_enabled,
 488                 .maxlen         = sizeof (int),
 489                 .mode           = 0644,
 490                 .proc_handler   = &proc_dointvec,
 491         },
 492 #endif
 493         {
 494                 .ctl_name       = KERN_CADPID,
 495                 .procname       = "cad_pid",
 496                 .data           = &cad_pid,
 497                 .maxlen         = sizeof (int),
 498                 .mode           = 0600,
 499                 .proc_handler   = &proc_dointvec,
 500         },
 501         {
 502                 .ctl_name       = KERN_MAX_THREADS,
 503                 .procname       = "threads-max",
 504                 .data           = &max_threads,
 505                 .maxlen         = sizeof(int),
 506                 .mode           = 0644,
 507                 .proc_handler   = &proc_dointvec,
 508         },
 509         {
 510                 .ctl_name       = KERN_RANDOM,
 511                 .procname       = "random",
 512                 .mode           = 0555,
 513                 .child          = random_table,
 514         },
 515 #ifdef CONFIG_UNIX98_PTYS
 516         {
 517                 .ctl_name       = KERN_PTY,
 518                 .procname       = "pty",
 519                 .mode           = 0555,
 520                 .child          = pty_table,
 521         },
 522 #endif
 523         {
 524                 .ctl_name       = KERN_OVERFLOWUID,
 525                 .procname       = "overflowuid",
 526                 .data           = &overflowuid,
 527                 .maxlen         = sizeof(int),
 528                 .mode           = 0644,
 529                 .proc_handler   = &proc_dointvec_minmax,
 530                 .strategy       = &sysctl_intvec,
 531                 .extra1         = &minolduid,
 532                 .extra2         = &maxolduid,
 533         },
 534         {
 535                 .ctl_name       = KERN_OVERFLOWGID,
 536                 .procname       = "overflowgid",
 537                 .data           = &overflowgid,
 538                 .maxlen         = sizeof(int),
 539                 .mode           = 0644,
 540                 .proc_handler   = &proc_dointvec_minmax,
 541                 .strategy       = &sysctl_intvec,
 542                 .extra1         = &minolduid,
 543                 .extra2         = &maxolduid,
 544         },
 545 #ifdef CONFIG_ARCH_S390
 546 #ifdef CONFIG_MATHEMU
 547         {
 548                 .ctl_name       = KERN_IEEE_EMULATION_WARNINGS,
 549                 .procname       = "ieee_emulation_warnings",
 550                 .data           = &sysctl_ieee_emulation_warnings,
 551                 .maxlen         = sizeof(int),
 552                 .mode           = 0644,
 553                 .proc_handler   = &proc_dointvec,
 554         },
 555 #endif
 556 #ifdef CONFIG_NO_IDLE_HZ
 557         {
 558                 .ctl_name       = KERN_HZ_TIMER,
 559                 .procname       = "hz_timer",
 560                 .data           = &sysctl_hz_timer,
 561                 .maxlen         = sizeof(int),
 562                 .mode           = 0644,
 563                 .proc_handler   = &proc_dointvec,
 564         },
 565 #endif
 566         {
 567                 .ctl_name       = KERN_S390_USER_DEBUG_LOGGING,
 568                 .procname       = "userprocess_debug",
 569                 .data           = &sysctl_userprocess_debug,
 570                 .maxlen         = sizeof(int),
 571                 .mode           = 0644,
 572                 .proc_handler   = &proc_dointvec,
 573         },
 574 #endif
 575         {
 576                 .ctl_name       = KERN_PIDMAX,
 577                 .procname       = "pid_max",
 578                 .data           = &pid_max,
 579                 .maxlen         = sizeof (int),
 580                 .mode           = 0644,
 581                 .proc_handler   = &proc_dointvec_minmax,
 582                 .strategy       = sysctl_intvec,
 583                 .extra1         = &pid_max_min,
 584                 .extra2         = &pid_max_max,
 585         },
 586         {
 587                 .ctl_name       = KERN_PANIC_ON_OOPS,
 588                 .procname       = "panic_on_oops",
 589                 .data           = &panic_on_oops,
 590                 .maxlen         = sizeof(int),
 591                 .mode           = 0644,
 592                 .proc_handler   = &proc_dointvec,
 593         },
 594         {
 595                 .ctl_name       = KERN_PRINTK_RATELIMIT,
 596                 .procname       = "printk_ratelimit",
 597                 .data           = &printk_ratelimit_jiffies,
 598                 .maxlen         = sizeof(int),
 599                 .mode           = 0644,
 600                 .proc_handler   = &proc_dointvec_jiffies,
 601                 .strategy       = &sysctl_jiffies,
 602         },
 603         {
 604                 .ctl_name       = KERN_PRINTK_RATELIMIT_BURST,
 605                 .procname       = "printk_ratelimit_burst",
 606                 .data           = &printk_ratelimit_burst,
 607                 .maxlen         = sizeof(int),
 608                 .mode           = 0644,
 609                 .proc_handler   = &proc_dointvec,
 610         },
 611         {
 612                 .ctl_name       = KERN_NGROUPS_MAX,
 613                 .procname       = "ngroups_max",
 614                 .data           = &ngroups_max,
 615                 .maxlen         = sizeof (int),
 616                 .mode           = 0444,
 617                 .proc_handler   = &proc_dointvec,
 618         },
 619 #if defined(CONFIG_X86_LOCAL_APIC) && defined(CONFIG_X86)
 620         {
 621                 .ctl_name       = KERN_UNKNOWN_NMI_PANIC,
 622                 .procname       = "unknown_nmi_panic",
 623                 .data           = &unknown_nmi_panic,
 624                 .maxlen         = sizeof (int),
 625                 .mode           = 0644,
 626                 .proc_handler   = &proc_unknown_nmi_panic,
 627         },
 628 #endif
 629 #if defined(CONFIG_X86)
 630         {
 631                 .ctl_name       = KERN_BOOTLOADER_TYPE,
 632                 .procname       = "bootloader_type",
 633                 .data           = &bootloader_type,
 634                 .maxlen         = sizeof (int),
 635                 .mode           = 0444,
 636                 .proc_handler   = &proc_dointvec,
 637         },
 638 #endif
 639         {
 640                 .ctl_name       = KERN_RANDOMIZE,
 641                 .procname       = "randomize_va_space",
 642                 .data           = &randomize_va_space,
 643                 .maxlen         = sizeof(int),
 644                 .mode           = 0644,
 645                 .proc_handler   = &proc_dointvec,
 646         },
 647 #if defined(CONFIG_ARCH_S390)
 648         {
 649                 .ctl_name       = KERN_SPIN_RETRY,
 650                 .procname       = "spin_retry",
 651                 .data           = &spin_retry,
 652                 .maxlen         = sizeof (int),
 653                 .mode           = 0644,
 654                 .proc_handler   = &proc_dointvec,
 655         },
 656 #endif
 657         { .ctl_name = 0 }
 658 };
 659
 660 /* Constants for minimum and maximum testing in vm_table.
 661    We use these as one-element integer vectors. */
 662 static int zero;
 663 static int one_hundred = 100;
 664
 665
 666 static ctl_table vm_table[] = {
 667         {
 668                 .ctl_name       = VM_OVERCOMMIT_MEMORY,
 669                 .procname       = "overcommit_memory",
 670                 .data           = &sysctl_overcommit_memory,
 671                 .maxlen         = sizeof(sysctl_overcommit_memory),
 672                 .mode           = 0644,
 673                 .proc_handler   = &proc_dointvec,
 674         },
 675         {
 676                 .ctl_name       = VM_OVERCOMMIT_RATIO,
 677                 .procname       = "overcommit_ratio",
 678                 .data           = &sysctl_overcommit_ratio,
 679                 .maxlen         = sizeof(sysctl_overcommit_ratio),
 680                 .mode           = 0644,
 681                 .proc_handler   = &proc_dointvec,
 682         },
 683         {
 684                 .ctl_name       = VM_PAGE_CLUSTER,
 685                 .procname       = "page-cluster",
 686                 .data           = &page_cluster,
 687                 .maxlen         = sizeof(int),
 688                 .mode           = 0644,
 689                 .proc_handler   = &proc_dointvec,
 690         },
 691         {
 692                 .ctl_name       = VM_DIRTY_BACKGROUND,
 693                 .procname       = "dirty_background_ratio",
 694                 .data           = &dirty_background_ratio,
 695                 .maxlen         = sizeof(dirty_background_ratio),
 696                 .mode           = 0644,
 697                 .proc_handler   = &proc_dointvec_minmax,
 698                 .strategy       = &sysctl_intvec,
 699                 .extra1         = &zero,
 700                 .extra2         = &one_hundred,
 701         },
 702         {
 703                 .ctl_name       = VM_DIRTY_RATIO,
 704                 .procname       = "dirty_ratio",
 705                 .data           = &vm_dirty_ratio,
 706                 .maxlen         = sizeof(vm_dirty_ratio),
 707                 .mode           = 0644,
 708                 .proc_handler   = &proc_dointvec_minmax,
 709                 .strategy       = &sysctl_intvec,
 710                 .extra1         = &zero,
 711                 .extra2         = &one_hundred,
 712         },
 713         {
 714                 .ctl_name       = VM_DIRTY_WB_CS,
 715                 .procname       = "dirty_writeback_centisecs",
 716                 .data           = &dirty_writeback_centisecs,
 717                 .maxlen         = sizeof(dirty_writeback_centisecs),
 718                 .mode           = 0644,
 719                 .proc_handler   = &dirty_writeback_centisecs_handler,
 720         },
 721         {
 722                 .ctl_name       = VM_DIRTY_EXPIRE_CS,
 723                 .procname       = "dirty_expire_centisecs",
 724                 .data           = &dirty_expire_centisecs,
 725                 .maxlen         = sizeof(dirty_expire_centisecs),
 726                 .mode           = 0644,
 727                 .proc_handler   = &proc_dointvec,
 728         },
 729         {
 730                 .ctl_name       = VM_NR_PDFLUSH_THREADS,
 731                 .procname       = "nr_pdflush_threads",
 732                 .data           = &nr_pdflush_threads,
 733                 .maxlen         = sizeof nr_pdflush_threads,
 734                 .mode           = 0444 /* read-only*/,
 735                 .proc_handler   = &proc_dointvec,
 736         },
 737         {
 738                 .ctl_name       = VM_SWAPPINESS,
 739                 .procname       = "swappiness",
 740                 .data           = &vm_swappiness,
 741                 .maxlen         = sizeof(vm_swappiness),
 742                 .mode           = 0644,
 743                 .proc_handler   = &proc_dointvec_minmax,
 744                 .strategy       = &sysctl_intvec,
 745                 .extra1         = &zero,
 746                 .extra2         = &one_hundred,
 747         },
 748 #ifdef CONFIG_HUGETLB_PAGE
 749          {
 750                 .ctl_name       = VM_HUGETLB_PAGES,
 751                 .procname       = "nr_hugepages",
 752                 .data           = &max_huge_pages,
 753                 .maxlen         = sizeof(unsigned long),
 754                 .mode           = 0644,
 755                 .proc_handler   = &hugetlb_sysctl_handler,
 756                 .extra1         = (void *)&hugetlb_zero,
 757                 .extra2         = (void *)&hugetlb_infinity,
 758          },
 759          {
 760                 .ctl_name       = VM_HUGETLB_GROUP,
 761                 .procname       = "hugetlb_shm_group",
 762                 .data           = &sysctl_hugetlb_shm_group,
 763                 .maxlen         = sizeof(gid_t),
 764                 .mode           = 0644,
 765                 .proc_handler   = &proc_dointvec,
 766          },
 767 #endif
 768         {
 769                 .ctl_name       = VM_LOWMEM_RESERVE_RATIO,
 770                 .procname       = "lowmem_reserve_ratio",
 771                 .data           = &sysctl_lowmem_reserve_ratio,
 772                 .maxlen         = sizeof(sysctl_lowmem_reserve_ratio),
 773                 .mode           = 0644,
 774                 .proc_handler   = &lowmem_reserve_ratio_sysctl_handler,
 775                 .strategy       = &sysctl_intvec,
 776         },
 777         {
 778                 .ctl_name       = VM_MIN_FREE_KBYTES,
 779                 .procname       = "min_free_kbytes",
 780                 .data           = &min_free_kbytes,
 781                 .maxlen         = sizeof(min_free_kbytes),
 782                 .mode           = 0644,
 783                 .proc_handler   = &min_free_kbytes_sysctl_handler,
 784                 .strategy       = &sysctl_intvec,
 785                 .extra1         = &zero,
 786         },
 787 #ifdef CONFIG_MMU
 788         {
 789                 .ctl_name       = VM_MAX_MAP_COUNT,
 790                 .procname       = "max_map_count",
 791                 .data           = &sysctl_max_map_count,
 792                 .maxlen         = sizeof(sysctl_max_map_count),
 793                 .mode           = 0644,
 794                 .proc_handler   = &proc_dointvec
 795         },
 796 #endif
 797         {
 798                 .ctl_name       = VM_LAPTOP_MODE,
 799                 .procname       = "laptop_mode",
 800                 .data           = &laptop_mode,
 801                 .maxlen         = sizeof(laptop_mode),
 802                 .mode           = 0644,
 803                 .proc_handler   = &proc_dointvec,
 804                 .strategy       = &sysctl_intvec,
 805                 .extra1         = &zero,
 806         },
 807         {
 808                 .ctl_name       = VM_BLOCK_DUMP,
 809                 .procname       = "block_dump",
 810                 .data           = &block_dump,
 811                 .maxlen         = sizeof(block_dump),
 812                 .mode           = 0644,
 813                 .proc_handler   = &proc_dointvec,
 814                 .strategy       = &sysctl_intvec,
 815                 .extra1         = &zero,
 816         },
 817         {
 818                 .ctl_name       = VM_VFS_CACHE_PRESSURE,
 819                 .procname       = "vfs_cache_pressure",
 820                 .data           = &sysctl_vfs_cache_pressure,
 821                 .maxlen         = sizeof(sysctl_vfs_cache_pressure),
 822                 .mode           = 0644,
 823                 .proc_handler   = &proc_dointvec,
 824                 .strategy       = &sysctl_intvec,
 825                 .extra1         = &zero,
 826         },
 827 #ifdef HAVE_ARCH_PICK_MMAP_LAYOUT
 828         {
 829                 .ctl_name       = VM_LEGACY_VA_LAYOUT,
 830                 .procname       = "legacy_va_layout",
 831                 .data           = &sysctl_legacy_va_layout,
 832                 .maxlen         = sizeof(sysctl_legacy_va_layout),
 833                 .mode           = 0644,
 834                 .proc_handler   = &proc_dointvec,
 835                 .strategy       = &sysctl_intvec,
 836                 .extra1         = &zero,
 837         },
 838 #endif
 839 #ifdef CONFIG_SWAP
 840         {
 841                 .ctl_name       = VM_SWAP_TOKEN_TIMEOUT,
 842                 .procname       = "swap_token_timeout",
 843                 .data           = &swap_token_default_timeout,
 844                 .maxlen         = sizeof(swap_token_default_timeout),
 845                 .mode           = 0644,
 846                 .proc_handler   = &proc_dointvec_jiffies,
 847                 .strategy       = &sysctl_jiffies,
 848         },
 849 #endif
 850         { .ctl_name = 0 }
 851 };
 852
 853 static ctl_table proc_table[] = {
 854         { .ctl_name = 0 }
 855 };
 856
 857 static ctl_table fs_table[] = {
 858         {
 859                 .ctl_name       = FS_NRINODE,
 860                 .procname       = "inode-nr",
 861                 .data           = &inodes_stat,
 862                 .maxlen         = 2*sizeof(int),
 863                 .mode           = 0444,
 864                 .proc_handler   = &proc_dointvec,
 865         },
 866         {
 867                 .ctl_name       = FS_STATINODE,
 868                 .procname       = "inode-state",
 869                 .data           = &inodes_stat,
 870                 .maxlen         = 7*sizeof(int),
 871                 .mode           = 0444,
 872                 .proc_handler   = &proc_dointvec,
 873         },
 874         {
 875                 .ctl_name       = FS_NRFILE,
 876                 .procname       = "file-nr",
 877                 .data           = &files_stat,
 878                 .maxlen         = 3*sizeof(int),
 879                 .mode           = 0444,
 880                 .proc_handler   = &proc_dointvec,
 881         },
 882         {
 883                 .ctl_name       = FS_MAXFILE,
 884                 .procname       = "file-max",
 885                 .data           = &files_stat.max_files,
 886                 .maxlen         = sizeof(int),
 887                 .mode           = 0644,
 888                 .proc_handler   = &proc_dointvec,
 889         },
 890         {
 891                 .ctl_name       = FS_DENTRY,
 892                 .procname       = "dentry-state",
 893                 .data           = &dentry_stat,
 894                 .maxlen         = 6*sizeof(int),
 895                 .mode           = 0444,
 896                 .proc_handler   = &proc_dointvec,
 897         },
 898         {
 899                 .ctl_name       = FS_OVERFLOWUID,
 900                 .procname       = "overflowuid",
 901                 .data           = &fs_overflowuid,
 902                 .maxlen         = sizeof(int),
 903                 .mode           = 0644,
 904                 .proc_handler   = &proc_dointvec_minmax,
 905                 .strategy       = &sysctl_intvec,
 906                 .extra1         = &minolduid,
 907                 .extra2         = &maxolduid,
 908         },
 909         {
 910                 .ctl_name       = FS_OVERFLOWGID,
 911                 .procname       = "overflowgid",
 912                 .data           = &fs_overflowgid,
 913                 .maxlen         = sizeof(int),
 914                 .mode           = 0644,
 915                 .proc_handler   = &proc_dointvec_minmax,
 916                 .strategy       = &sysctl_intvec,
 917                 .extra1         = &minolduid,
 918                 .extra2         = &maxolduid,
 919         },
 920         {
 921                 .ctl_name       = FS_LEASES,
 922                 .procname       = "leases-enable",
 923                 .data           = &leases_enable,
 924                 .maxlen         = sizeof(int),
 925                 .mode           = 0644,
 926                 .proc_handler   = &proc_dointvec,
 927         },
 928 #ifdef CONFIG_DNOTIFY
 929         {
 930                 .ctl_name       = FS_DIR_NOTIFY,
 931                 .procname       = "dir-notify-enable",
 932                 .data           = &dir_notify_enable,
 933                 .maxlen         = sizeof(int),
 934                 .mode           = 0644,
 935                 .proc_handler   = &proc_dointvec,
 936         },
 937 #endif
 938 #ifdef CONFIG_MMU
 939         {
 940                 .ctl_name       = FS_LEASE_TIME,
 941                 .procname       = "lease-break-time",
 942                 .data           = &lease_break_time,
 943                 .maxlen         = sizeof(int),
 944                 .mode           = 0644,
 945                 .proc_handler   = &proc_dointvec,
 946         },
 947         {
 948                 .ctl_name       = FS_AIO_NR,
 949                 .procname       = "aio-nr",
 950                 .data           = &aio_nr,
 951                 .maxlen         = sizeof(aio_nr),
 952                 .mode           = 0444,
 953                 .proc_handler   = &proc_doulongvec_minmax,
 954         },
 955         {
 956                 .ctl_name       = FS_AIO_MAX_NR,
 957                 .procname       = "aio-max-nr",
 958                 .data           = &aio_max_nr,
 959                 .maxlen         = sizeof(aio_max_nr),
 960                 .mode           = 0644,
 961                 .proc_handler   = &proc_doulongvec_minmax,
 962         },
 963 #ifdef CONFIG_INOTIFY
 964         {
 965                 .ctl_name       = FS_INOTIFY,
 966                 .procname       = "inotify",
 967                 .mode           = 0555,
 968                 .child          = inotify_table,
 969         },
 970 #endif
 971 #endif
 972         {
 973                 .ctl_name       = KERN_SETUID_DUMPABLE,
 974                 .procname       = "suid_dumpable",
 975                 .data           = &suid_dumpable,
 976                 .maxlen         = sizeof(int),
 977                 .mode           = 0644,
 978                 .proc_handler   = &proc_dointvec,
 979         },
 980         { .ctl_name = 0 }
 981 };
 982
 983 static ctl_table debug_table[] = {
 984         { .ctl_name = 0 }
 985 };
 986
 987 static ctl_table dev_table[] = {
 988         { .ctl_name = 0 }
 989 };
 990
 991 extern void init_irq_proc (void);
 992
 993 static DEFINE_SPINLOCK(sysctl_lock);
 994
 995 /* called under sysctl_lock */
 996 static int use_table(struct ctl_table_header *p)
 997 {
 998         if (unlikely(p->unregistering))
 999                 return 0;
1000         p->used++;
1001         return 1;
1002 }
1003
1004 /* called under sysctl_lock */
1005 static void unuse_table(struct ctl_table_header *p)
1006 {
1007         if (!--p->used)
1008                 if (unlikely(p->unregistering))
1009                         complete(p->unregistering);
1010 }
1011
1012 /* called under sysctl_lock, will reacquire if has to wait */
1013 static void start_unregistering(struct ctl_table_header *p)
1014 {
1015         /*
1016          * if p->used is 0, nobody will ever touch that entry again;
1017          * we'll eliminate all paths to it before dropping sysctl_lock
1018          */
1019         if (unlikely(p->used)) {
1020                 struct completion wait;
1021                 init_completion(&wait);
1022                 p->unregistering = &wait;
1023                 spin_unlock(&sysctl_lock);
1024                 wait_for_completion(&wait);
1025                 spin_lock(&sysctl_lock);
1026         }
1027         /*
1028          * do not remove from the list until nobody holds it; walking the
1029          * list in do_sysctl() relies on that.
1030          */
1031         list_del_init(&p->ctl_entry);
1032 }
1033
1034 void __init sysctl_init(void)
1035 {
1036 #ifdef CONFIG_PROC_FS
1037         register_proc_table(root_table, proc_sys_root, &root_table_header);
1038         init_irq_proc();
1039 #endif
1040 }
1041
1042 int do_sysctl(int __user *name, int nlen, void __user *oldval, size_t __user *oldlenp,
1043                void __user *newval, size_t newlen)
1044 {
1045         struct list_head *tmp;
1046         int error = -ENOTDIR;
1047
1048         if (nlen <= 0 || nlen >= CTL_MAXNAME)
1049                 return -ENOTDIR;
1050         if (oldval) {
1051                 int old_len;
1052                 if (!oldlenp || get_user(old_len, oldlenp))
1053                         return -EFAULT;
1054         }
1055         spin_lock(&sysctl_lock);
1056         tmp = &root_table_header.ctl_entry;
1057         do {
1058                 struct ctl_table_header *head =
1059                         list_entry(tmp, struct ctl_table_header, ctl_entry);
1060                 void *context = NULL;
1061
1062                 if (!use_table(head))
1063                         continue;
1064
1065                 spin_unlock(&sysctl_lock);
1066
1067                 error = parse_table(name, nlen, oldval, oldlenp,
1068                                         newval, newlen, head->ctl_table,
1069                                         &context);
1070                 kfree(context);
1071
1072                 spin_lock(&sysctl_lock);
1073                 unuse_table(head);
1074                 if (error != -ENOTDIR)
1075                         break;
1076         } while ((tmp = tmp->next) != &root_table_header.ctl_entry);
1077         spin_unlock(&sysctl_lock);
1078         return error;
1079 }
1080
1081 asmlinkage long sys_sysctl(struct __sysctl_args __user *args)
1082 {
1083         struct __sysctl_args tmp;
1084         int error;
1085
1086         if (copy_from_user(&tmp, args, sizeof(tmp)))
1087                 return -EFAULT;
1088
1089         lock_kernel();
1090         error = do_sysctl(tmp.name, tmp.nlen, tmp.oldval, tmp.oldlenp,
1091                           tmp.newval, tmp.newlen);
1092         unlock_kernel();
1093         return error;
1094 }
1095
1096 /*
1097  * ctl_perm does NOT grant the superuser all rights automatically, because
1098  * some sysctl variables are readonly even to root.
1099  */
1100
1101 static int test_perm(int mode, int op)
1102 {
1103         if (!current->euid)
1104                 mode >>= 6;
1105         else if (in_egroup_p(0))
1106                 mode >>= 3;
1107         if ((mode & op & 0007) == op)
1108                 return 0;
1109         return -EACCES;
1110 }
1111
1112 static inline int ctl_perm(ctl_table *table, int op)
1113 {
1114         int error;
1115         error = security_sysctl(table, op);
1116         if (error)
1117                 return error;
1118         return test_perm(table->mode, op);
1119 }
1120
1121 static int parse_table(int __user *name, int nlen,
1122                        void __user *oldval, size_t __user *oldlenp,
1123                        void __user *newval, size_t newlen,
1124                        ctl_table *table, void **context)
1125 {
1126         int n;
1127 repeat:
1128         if (!nlen)
1129                 return -ENOTDIR;
1130         if (get_user(n, name))
1131                 return -EFAULT;
1132         for ( ; table->ctl_name; table++) {
1133                 if (n == table->ctl_name || table->ctl_name == CTL_ANY) {
1134                         int error;
1135                         if (table->child) {
1136                                 if (ctl_perm(table, 001))
1137                                         return -EPERM;
1138                                 if (table->strategy) {
1139                                         error = table->strategy(
1140                                                 table, name, nlen,
1141                                                 oldval, oldlenp,
1142                                                 newval, newlen, context);
1143                                         if (error)
1144                                                 return error;
1145                                 }
1146                                 name++;
1147                                 nlen--;
1148                                 table = table->child;
1149                                 goto repeat;
1150                         }
1151                         error = do_sysctl_strategy(table, name, nlen,
1152                                                    oldval, oldlenp,
1153                                                    newval, newlen, context);
1154                         return error;
1155                 }
1156         }
1157         return -ENOTDIR;
1158 }
1159
1160 /* Perform the actual read/write of a sysctl table entry. */
1161 int do_sysctl_strategy (ctl_table *table,
1162                         int __user *name, int nlen,
1163                         void __user *oldval, size_t __user *oldlenp,
1164                         void __user *newval, size_t newlen, void **context)
1165 {
1166         int op = 0, rc;
1167         size_t len;
1168
1169         if (oldval)
1170                 op |= 004;
1171         if (newval)
1172                 op |= 002;
1173         if (ctl_perm(table, op))
1174                 return -EPERM;
1175
1176         if (table->strategy) {
1177                 rc = table->strategy(table, name, nlen, oldval, oldlenp,
1178                                      newval, newlen, context);
1179                 if (rc < 0)
1180                         return rc;
1181                 if (rc > 0)
1182                         return 0;
1183         }
1184
1185         /* If there is no strategy routine, or if the strategy returns
1186          * zero, proceed with automatic r/w */
1187         if (table->data && table->maxlen) {
1188                 if (oldval && oldlenp) {
1189                         if (get_user(len, oldlenp))
1190                                 return -EFAULT;
1191                         if (len) {
1192                                 if (len > table->maxlen)
1193                                         len = table->maxlen;
1194                                 if(copy_to_user(oldval, table->data, len))
1195                                         return -EFAULT;
1196                                 if(put_user(len, oldlenp))
1197                                         return -EFAULT;
1198                         }
1199                 }
1200                 if (newval && newlen) {
1201                         len = newlen;
1202                         if (len > table->maxlen)
1203                                 len = table->maxlen;
1204                         if(copy_from_user(table->data, newval, len))
1205                                 return -EFAULT;
1206                 }
1207         }
1208         return 0;
1209 }
1210
1211 /**
1212  * register_sysctl_table - register a sysctl hierarchy
1213  * @table: the top-level table structure
1214  * @insert_at_head: whether the entry should be inserted in front or at the end
1215  *
1216  * Register a sysctl table hierarchy. @table should be a filled in ctl_table
1217  * array. An entry with a ctl_name of 0 terminates the table.
1218  *
1219  * The members of the &ctl_table structure are used as follows:
1220  *
1221  * ctl_name - This is the numeric sysctl value used by sysctl(2). The number
1222  *            must be unique within that level of sysctl
1223  *
1224  * procname - the name of the sysctl file under /proc/sys. Set to %NULL to not
1225  *            enter a sysctl file
1226  *
1227  * data - a pointer to data for use by proc_handler
1228  *
1229  * maxlen - the maximum size in bytes of the data
1230  *
1231  * mode - the file permissions for the /proc/sys file, and for sysctl(2)
1232  *
1233  * child - a pointer to the child sysctl table if this entry is a directory, or
1234  *         %NULL.
1235  *
1236  * proc_handler - the text handler routine (described below)
1237  *
1238  * strategy - the strategy routine (described below)
1239  *
1240  * de - for internal use by the sysctl routines
1241  *
1242  * extra1, extra2 - extra pointers usable by the proc handler routines
1243  *
1244  * Leaf nodes in the sysctl tree will be represented by a single file
1245  * under /proc; non-leaf nodes will be represented by directories.
1246  *
1247  * sysctl(2) can automatically manage read and write requests through
1248  * the sysctl table.  The data and maxlen fields of the ctl_table
1249  * struct enable minimal validation of the values being written to be
1250  * performed, and the mode field allows minimal authentication.
1251  *
1252  * More sophisticated management can be enabled by the provision of a
1253  * strategy routine with the table entry.  This will be called before
1254  * any automatic read or write of the data is performed.
1255  *
1256  * The strategy routine may return
1257  *
1258  * < 0 - Error occurred (error is passed to user process)
1259  *
1260  * 0   - OK - proceed with automatic read or write.
1261  *
1262  * > 0 - OK - read or write has been done by the strategy routine, so
1263  *       return immediately.
1264  *
1265  * There must be a proc_handler routine for any terminal nodes
1266  * mirrored under /proc/sys (non-terminals are handled by a built-in
1267  * directory handler).  Several default handlers are available to
1268  * cover common cases -
1269  *
1270  * proc_dostring(), proc_dointvec(), proc_dointvec_jiffies(),
1271  * proc_dointvec_userhz_jiffies(), proc_dointvec_minmax(),
1272  * proc_doulongvec_ms_jiffies_minmax(), proc_doulongvec_minmax()
1273  *
1274  * It is the handler's job to read the input buffer from user memory
1275  * and process it. The handler should return 0 on success.
1276  *
1277  * This routine returns %NULL on a failure to register, and a pointer
1278  * to the table header on success.
1279  */
1280 struct ctl_table_header *register_sysctl_table(ctl_table * table,
1281                                                int insert_at_head)
1282 {
1283         struct ctl_table_header *tmp;
1284         tmp = kmalloc(sizeof(struct ctl_table_header), GFP_KERNEL);
1285         if (!tmp)
1286                 return NULL;
1287         tmp->ctl_table = table;
1288         INIT_LIST_HEAD(&tmp->ctl_entry);
1289         tmp->used = 0;
1290         tmp->unregistering = NULL;
1291         spin_lock(&sysctl_lock);
1292         if (insert_at_head)
1293                 list_add(&tmp->ctl_entry, &root_table_header.ctl_entry);
1294         else
1295                 list_add_tail(&tmp->ctl_entry, &root_table_header.ctl_entry);
1296         spin_unlock(&sysctl_lock);
1297 #ifdef CONFIG_PROC_FS
1298         register_proc_table(table, proc_sys_root, tmp);
1299 #endif
1300         return tmp;
1301 }
1302
1303 /**
1304  * unregister_sysctl_table - unregister a sysctl table hierarchy
1305  * @header: the header returned from register_sysctl_table
1306  *
1307  * Unregisters the sysctl table and all children. proc entries may not
1308  * actually be removed until they are no longer used by anyone.
1309  */
1310 void unregister_sysctl_table(struct ctl_table_header * header)
1311 {
1312         might_sleep();
1313         spin_lock(&sysctl_lock);
1314         start_unregistering(header);
1315 #ifdef CONFIG_PROC_FS
1316         unregister_proc_table(header->ctl_table, proc_sys_root);
1317 #endif
1318         spin_unlock(&sysctl_lock);
1319         kfree(header);
1320 }
1321
1322 /*
1323  * /proc/sys support
1324  */
1325
1326 #ifdef CONFIG_PROC_FS
1327
1328 /* Scan the sysctl entries in table and add them all into /proc */
1329 static void register_proc_table(ctl_table * table, struct proc_dir_entry *root, void *set)
1330 {
1331         struct proc_dir_entry *de;
1332         int len;
1333         mode_t mode;
1334
1335         for (; table->ctl_name; table++) {
1336                 /* Can't do anything without a proc name. */
1337                 if (!table->procname)
1338                         continue;
1339                 /* Maybe we can't do anything with it... */
1340                 if (!table->proc_handler && !table->child) {
1341                         printk(KERN_WARNING "SYSCTL: Can't register %s\n",
1342                                 table->procname);
1343                         continue;
1344                 }
1345
1346                 len = strlen(table->procname);
1347                 mode = table->mode;
1348
1349                 de = NULL;
1350                 if (table->proc_handler)
1351                         mode |= S_IFREG;
1352                 else {
1353                         mode |= S_IFDIR;
1354                         for (de = root->subdir; de; de = de->next) {
1355                                 if (proc_match(len, table->procname, de))
1356                                         break;
1357                         }
1358                         /* If the subdir exists already, de is non-NULL */
1359                 }
1360
1361                 if (!de) {
1362                         de = create_proc_entry(table->procname, mode, root);
1363                         if (!de)
1364                                 continue;
1365                         de->set = set;
1366                         de->data = (void *) table;
1367                         if (table->proc_handler)
1368                                 de->proc_fops = &proc_sys_file_operations;
1369                 }
1370                 table->de = de;
1371                 if (de->mode & S_IFDIR)
1372                         register_proc_table(table->child, de, set);
1373         }
1374 }
1375
1376 /*
1377  * Unregister a /proc sysctl table and any subdirectories.
1378  */
1379 static void unregister_proc_table(ctl_table * table, struct proc_dir_entry *root)
1380 {
1381         struct proc_dir_entry *de;
1382         for (; table->ctl_name; table++) {
1383                 if (!(de = table->de))
1384                         continue;
1385                 if (de->mode & S_IFDIR) {
1386                         if (!table->child) {
1387                                 printk (KERN_ALERT "Help - malformed sysctl tree on free\n");
1388                                 continue;
1389                         }
1390                         unregister_proc_table(table->child, de);
1391
1392                         /* Don't unregister directories which still have entries.. */
1393                         if (de->subdir)
1394                                 continue;
1395                 }
1396
1397                 /*
1398                  * In any case, mark the entry as goner; we'll keep it
1399                  * around if it's busy, but we'll know to do nothing with
1400                  * its fields.  We are under sysctl_lock here.
1401                  */
1402                 de->data = NULL;
1403
1404                 /* Don't unregister proc entries that are still being used.. */
1405                 if (atomic_read(&de->count))
1406                         continue;
1407
1408                 table->de = NULL;
1409                 remove_proc_entry(table->procname, root);
1410         }
1411 }
1412
1413 static ssize_t do_rw_proc(int write, struct file * file, char __user * buf,
1414                           size_t count, loff_t *ppos)
1415 {
1416         int op;
1417         struct proc_dir_entry *de = PDE(file->f_dentry->d_inode);
1418         struct ctl_table *table;
1419         size_t res;
1420         ssize_t error = -ENOTDIR;
1421
1422         spin_lock(&sysctl_lock);
1423         if (de && de->data && use_table(de->set)) {
1424                 /*
1425                  * at that point we know that sysctl was not unregistered
1426                  * and won't be until we finish
1427                  */
1428                 spin_unlock(&sysctl_lock);
1429                 table = (struct ctl_table *) de->data;
1430                 if (!table || !table->proc_handler)
1431                         goto out;
1432                 error = -EPERM;
1433                 op = (write ? 002 : 004);
1434                 if (ctl_perm(table, op))
1435                         goto out;
1436
1437                 /* careful: calling conventions are nasty here */
1438                 res = count;
1439                 error = (*table->proc_handler)(table, write, file,
1440                                                 buf, &res, ppos);
1441                 if (!error)
1442                         error = res;
1443         out:
1444                 spin_lock(&sysctl_lock);
1445                 unuse_table(de->set);
1446         }
1447         spin_unlock(&sysctl_lock);
1448         return error;
1449 }
1450
1451 static int proc_opensys(struct inode *inode, struct file *file)
1452 {
1453         if (file->f_mode & FMODE_WRITE) {
1454                 /*
1455                  * sysctl entries that are not writable,
1456                  * are _NOT_ writable, capabilities or not.
1457                  */
1458                 if (!(inode->i_mode & S_IWUSR))
1459                         return -EPERM;
1460         }
1461
1462         return 0;
1463 }
1464
1465 static ssize_t proc_readsys(struct file * file, char __user * buf,
1466                             size_t count, loff_t *ppos)
1467 {
1468         return do_rw_proc(0, file, buf, count, ppos);
1469 }
1470
1471 static ssize_t proc_writesys(struct file * file, const char __user * buf,
1472                              size_t count, loff_t *ppos)
1473 {
1474         return do_rw_proc(1, file, (char __user *) buf, count, ppos);
1475 }
1476
1477 /**
1478  * proc_dostring - read a string sysctl
1479  * @table: the sysctl table
1480  * @write: %TRUE if this is a write to the sysctl file
1481  * @filp: the file structure
1482  * @buffer: the user buffer
1483  * @lenp: the size of the user buffer
1484  * @ppos: file position
1485  *
1486  * Reads/writes a string from/to the user buffer. If the kernel
1487  * buffer provided is not large enough to hold the string, the
1488  * string is truncated. The copied string is %NULL-terminated.
1489  * If the string is being read by the user process, it is copied
1490  * and a newline '\n' is added. It is truncated if the buffer is
1491  * not large enough.
1492  *
1493  * Returns 0 on success.
1494  */
1495 int proc_dostring(ctl_table *table, int write, struct file *filp,
1496                   void __user *buffer, size_t *lenp, loff_t *ppos)
1497 {
1498         size_t len;
1499         char __user *p;
1500         char c;
1501
1502         if (!table->data || !table->maxlen || !*lenp ||
1503             (*ppos && !write)) {
1504                 *lenp = 0;
1505                 return 0;
1506         }
1507
1508         if (write) {
1509                 len = 0;
1510                 p = buffer;
1511                 while (len < *lenp) {
1512                         if (get_user(c, p++))
1513                                 return -EFAULT;
1514                         if (c == 0 || c == '\n')
1515                                 break;
1516                         len++;
1517                 }
1518                 if (len >= table->maxlen)
1519                         len = table->maxlen-1;
1520                 if(copy_from_user(table->data, buffer, len))
1521                         return -EFAULT;
1522                 ((char *) table->data)[len] = 0;
1523                 *ppos += *lenp;
1524         } else {
1525                 len = strlen(table->data);
1526                 if (len > table->maxlen)
1527                         len = table->maxlen;
1528                 if (len > *lenp)
1529                         len = *lenp;
1530                 if (len)
1531                         if(copy_to_user(buffer, table->data, len))
1532                                 return -EFAULT;
1533                 if (len < *lenp) {
1534                         if(put_user('\n', ((char __user *) buffer) + len))
1535                                 return -EFAULT;
1536                         len++;
1537                 }
1538                 *lenp = len;
1539                 *ppos += len;
1540         }
1541         return 0;
1542 }
1543
1544 /*
1545  *      Special case of dostring for the UTS structure. This has locks
1546  *      to observe. Should this be in kernel/sys.c ????
1547  */
1548
1549 static int proc_doutsstring(ctl_table *table, int write, struct file *filp,
1550                   void __user *buffer, size_t *lenp, loff_t *ppos)
1551 {
1552         int r;
1553
1554         if (!write) {
1555                 down_read(&uts_sem);
1556                 r=proc_dostring(table,0,filp,buffer,lenp, ppos);
1557                 up_read(&uts_sem);
1558         } else {
1559                 down_write(&uts_sem);
1560                 r=proc_dostring(table,1,filp,buffer,lenp, ppos);
1561                 up_write(&uts_sem);
1562         }
1563         return r;
1564 }
1565
1566 static int do_proc_dointvec_conv(int *negp, unsigned long *lvalp,
1567                                  int *valp,
1568                                  int write, void *data)
1569 {
1570         if (write) {
1571                 *valp = *negp ? -*lvalp : *lvalp;
1572         } else {
1573                 int val = *valp;
1574                 if (val < 0) {
1575                         *negp = -1;
1576                         *lvalp = (unsigned long)-val;
1577                 } else {
1578                         *negp = 0;
1579                         *lvalp = (unsigned long)val;
1580                 }
1581         }
1582         return 0;
1583 }
1584
1585 static int do_proc_dointvec(ctl_table *table, int write, struct file *filp,
1586                   void __user *buffer, size_t *lenp, loff_t *ppos,
1587                   int (*conv)(int *negp, unsigned long *lvalp, int *valp,
1588                               int write, void *data),
1589                   void *data)
1590 {
1591 #define TMPBUFLEN 21
1592         int *i, vleft, first=1, neg, val;
1593         unsigned long lval;
1594         size_t left, len;
1595
1596         char buf[TMPBUFLEN], *p;
1597         char __user *s = buffer;
1598
1599         if (!table->data || !table->maxlen || !*lenp ||
1600             (*ppos && !write)) {
1601                 *lenp = 0;
1602                 return 0;
1603         }
1604
1605         i = (int *) table->data;
1606         vleft = table->maxlen / sizeof(*i);
1607         left = *lenp;
1608
1609         if (!conv)
1610                 conv = do_proc_dointvec_conv;
1611
1612         for (; left && vleft--; i++, first=0) {
1613                 if (write) {
1614                         while (left) {
1615                                 char c;
1616                                 if (get_user(c, s))
1617                                         return -EFAULT;
1618                                 if (!isspace(c))
1619                                         break;
1620                                 left--;
1621                                 s++;
1622                         }
1623                         if (!left)
1624                                 break;
1625                         neg = 0;
1626                         len = left;
1627                         if (len > sizeof(buf) - 1)
1628                                 len = sizeof(buf) - 1;
1629                         if (copy_from_user(buf, s, len))
1630                                 return -EFAULT;
1631                         buf[len] = 0;
1632                         p = buf;
1633                         if (*p == '-' && left > 1) {
1634                                 neg = 1;
1635                                 left--, p++;
1636                         }
1637                         if (*p < '0' || *p > '9')
1638                                 break;
1639
1640                         lval = simple_strtoul(p, &p, 0);
1641
1642                         len = p-buf;
1643                         if ((len < left) && *p && !isspace(*p))
1644                                 break;
1645                         if (neg)
1646                                 val = -val;
1647                         s += len;
1648                         left -= len;
1649
1650                         if (conv(&neg, &lval, i, 1, data))
1651                                 break;
1652                 } else {
1653                         p = buf;
1654                         if (!first)
1655                                 *p++ = '\t';
1656
1657                         if (conv(&neg, &lval, i, 0, data))
1658                                 break;
1659
1660                         sprintf(p, "%s%lu", neg ? "-" : "", lval);
1661                         len = strlen(buf);
1662                         if (len > left)
1663                                 len = left;
1664                         if(copy_to_user(s, buf, len))
1665                                 return -EFAULT;
1666                         left -= len;
1667                         s += len;
1668                 }
1669         }
1670
1671         if (!write && !first && left) {
1672                 if(put_user('\n', s))
1673                         return -EFAULT;
1674                 left--, s++;
1675         }
1676         if (write) {
1677                 while (left) {
1678                         char c;
1679                         if (get_user(c, s++))
1680                                 return -EFAULT;
1681                         if (!isspace(c))
1682                                 break;
1683                         left--;
1684                 }
1685         }
1686         if (write && first)
1687                 return -EINVAL;
1688         *lenp -= left;
1689         *ppos += *lenp;
1690         return 0;
1691 #undef TMPBUFLEN
1692 }
1693
1694 /**
1695  * proc_dointvec - read a vector of integers
1696  * @table: the sysctl table
1697  * @write: %TRUE if this is a write to the sysctl file
1698  * @filp: the file structure
1699  * @buffer: the user buffer
1700  * @lenp: the size of the user buffer
1701  * @ppos: file position
1702  *
1703  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
1704  * values from/to the user buffer, treated as an ASCII string.
1705  *
1706  * Returns 0 on success.
1707  */
1708 int proc_dointvec(ctl_table *table, int write, struct file *filp,
1709                      void __user *buffer, size_t *lenp, loff_t *ppos)
1710 {
1711     return do_proc_dointvec(table,write,filp,buffer,lenp,ppos,
1712                             NULL,NULL);
1713 }
1714
1715 #define OP_SET  0
1716 #define OP_AND  1
1717 #define OP_OR   2
1718 #define OP_MAX  3
1719 #define OP_MIN  4
1720
1721 static int do_proc_dointvec_bset_conv(int *negp, unsigned long *lvalp,
1722                                       int *valp,
1723                                       int write, void *data)
1724 {
1725         int op = *(int *)data;
1726         if (write) {
1727                 int val = *negp ? -*lvalp : *lvalp;
1728                 switch(op) {
1729                 case OP_SET:    *valp = val; break;
1730                 case OP_AND:    *valp &= val; break;
1731                 case OP_OR:     *valp |= val; break;
1732                 case OP_MAX:    if(*valp < val)
1733                                         *valp = val;
1734                                 break;
1735                 case OP_MIN:    if(*valp > val)
1736                                 *valp = val;
1737                                 break;
1738                 }
1739         } else {
1740                 int val = *valp;
1741                 if (val < 0) {
1742                         *negp = -1;
1743                         *lvalp = (unsigned long)-val;
1744                 } else {
1745                         *negp = 0;
1746                         *lvalp = (unsigned long)val;
1747                 }
1748         }
1749         return 0;
1750 }
1751
1752 /*
1753  *      init may raise the set.
1754  */
1755
1756 int proc_dointvec_bset(ctl_table *table, int write, struct file *filp,
1757                         void __user *buffer, size_t *lenp, loff_t *ppos)
1758 {
1759         int op;
1760
1761         if (!capable(CAP_SYS_MODULE)) {
1762                 return -EPERM;
1763         }
1764
1765         op = (current->pid == 1) ? OP_SET : OP_AND;
1766         return do_proc_dointvec(table,write,filp,buffer,lenp,ppos,
1767                                 do_proc_dointvec_bset_conv,&op);
1768 }
1769
1770 struct do_proc_dointvec_minmax_conv_param {
1771         int *min;
1772         int *max;
1773 };
1774
1775 static int do_proc_dointvec_minmax_conv(int *negp, unsigned long *lvalp,
1776                                         int *valp,
1777                                         int write, void *data)
1778 {
1779         struct do_proc_dointvec_minmax_conv_param *param = data;
1780         if (write) {
1781                 int val = *negp ? -*lvalp : *lvalp;
1782                 if ((param->min && *param->min > val) ||
1783                     (param->max && *param->max < val))
1784                         return -EINVAL;
1785                 *valp = val;
1786         } else {
1787                 int val = *valp;
1788                 if (val < 0) {
1789                         *negp = -1;
1790                         *lvalp = (unsigned long)-val;
1791                 } else {
1792                         *negp = 0;
1793                         *lvalp = (unsigned long)val;
1794                 }
1795         }
1796         return 0;
1797 }
1798
1799 /**
1800  * proc_dointvec_minmax - read a vector of integers with min/max values
1801  * @table: the sysctl table
1802  * @write: %TRUE if this is a write to the sysctl file
1803  * @filp: the file structure
1804  * @buffer: the user buffer
1805  * @lenp: the size of the user buffer
1806  * @ppos: file position
1807  *
1808  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
1809  * values from/to the user buffer, treated as an ASCII string.
1810  *
1811  * This routine will ensure the values are within the range specified by
1812  * table->extra1 (min) and table->extra2 (max).
1813  *
1814  * Returns 0 on success.
1815  */
1816 int proc_dointvec_minmax(ctl_table *table, int write, struct file *filp,
1817                   void __user *buffer, size_t *lenp, loff_t *ppos)
1818 {
1819         struct do_proc_dointvec_minmax_conv_param param = {
1820                 .min = (int *) table->extra1,
1821                 .max = (int *) table->extra2,
1822         };
1823         return do_proc_dointvec(table, write, filp, buffer, lenp, ppos,
1824                                 do_proc_dointvec_minmax_conv, &param);
1825 }
1826
1827 static int do_proc_doulongvec_minmax(ctl_table *table, int write,
1828                                      struct file *filp,
1829                                      void __user *buffer,
1830                                      size_t *lenp, loff_t *ppos,
1831                                      unsigned long convmul,
1832                                      unsigned long convdiv)
1833 {
1834 #define TMPBUFLEN 21
1835         unsigned long *i, *min, *max, val;
1836         int vleft, first=1, neg;
1837         size_t len, left;
1838         char buf[TMPBUFLEN], *p;
1839         char __user *s = buffer;
1840
1841         if (!table->data || !table->maxlen || !*lenp ||
1842             (*ppos && !write)) {
1843                 *lenp = 0;
1844                 return 0;
1845         }
1846
1847         i = (unsigned long *) table->data;
1848         min = (unsigned long *) table->extra1;
1849         max = (unsigned long *) table->extra2;
1850         vleft = table->maxlen / sizeof(unsigned long);
1851         left = *lenp;
1852
1853         for (; left && vleft--; i++, min++, max++, first=0) {
1854                 if (write) {
1855                         while (left) {
1856                                 char c;
1857                                 if (get_user(c, s))
1858                                         return -EFAULT;
1859                                 if (!isspace(c))
1860                                         break;
1861                                 left--;
1862                                 s++;
1863                         }
1864                         if (!left)
1865                                 break;
1866                         neg = 0;
1867                         len = left;
1868                         if (len > TMPBUFLEN-1)
1869                                 len = TMPBUFLEN-1;
1870                         if (copy_from_user(buf, s, len))
1871                                 return -EFAULT;
1872                         buf[len] = 0;
1873                         p = buf;
1874                         if (*p == '-' && left > 1) {
1875                                 neg = 1;
1876                                 left--, p++;
1877                         }
1878                         if (*p < '0' || *p > '9')
1879                                 break;
1880                         val = simple_strtoul(p, &p, 0) * convmul / convdiv ;
1881                         len = p-buf;
1882                         if ((len < left) && *p && !isspace(*p))
1883                                 break;
1884                         if (neg)
1885                                 val = -val;
1886                         s += len;
1887                         left -= len;
1888
1889                         if(neg)
1890                                 continue;
1891                         if ((min && val < *min) || (max && val > *max))
1892                                 continue;
1893                         *i = val;
1894                 } else {
1895                         p = buf;
1896                         if (!first)
1897                                 *p++ = '\t';
1898                         sprintf(p, "%lu", convdiv * (*i) / convmul);
1899                         len = strlen(buf);
1900                         if (len > left)
1901                                 len = left;
1902                         if(copy_to_user(s, buf, len))
1903                                 return -EFAULT;
1904                         left -= len;
1905                         s += len;
1906                 }
1907         }
1908
1909         if (!write && !first && left) {
1910                 if(put_user('\n', s))
1911                         return -EFAULT;
1912                 left--, s++;
1913         }
1914         if (write) {
1915                 while (left) {
1916                         char c;
1917                         if (get_user(c, s++))
1918                                 return -EFAULT;
1919                         if (!isspace(c))
1920                                 break;
1921                         left--;
1922                 }
1923         }
1924         if (write && first)
1925                 return -EINVAL;
1926         *lenp -= left;
1927         *ppos += *lenp;
1928         return 0;
1929 #undef TMPBUFLEN
1930 }
1931
1932 /**
1933  * proc_doulongvec_minmax - read a vector of long integers with min/max values
1934  * @table: the sysctl table
1935  * @write: %TRUE if this is a write to the sysctl file
1936  * @filp: the file structure
1937  * @buffer: the user buffer
1938  * @lenp: the size of the user buffer
1939  * @ppos: file position
1940  *
1941  * Reads/writes up to table->maxlen/sizeof(unsigned long) unsigned long
1942  * values from/to the user buffer, treated as an ASCII string.
1943  *
1944  * This routine will ensure the values are within the range specified by
1945  * table->extra1 (min) and table->extra2 (max).
1946  *
1947  * Returns 0 on success.
1948  */
1949 int proc_doulongvec_minmax(ctl_table *table, int write, struct file *filp,
1950                            void __user *buffer, size_t *lenp, loff_t *ppos)
1951 {
1952     return do_proc_doulongvec_minmax(table, write, filp, buffer, lenp, ppos, 1l, 1l);
1953 }
1954
1955 /**
1956  * proc_doulongvec_ms_jiffies_minmax - read a vector of millisecond values with min/max values
1957  * @table: the sysctl table
1958  * @write: %TRUE if this is a write to the sysctl file
1959  * @filp: the file structure
1960  * @buffer: the user buffer
1961  * @lenp: the size of the user buffer
1962  * @ppos: file position
1963  *
1964  * Reads/writes up to table->maxlen/sizeof(unsigned long) unsigned long
1965  * values from/to the user buffer, treated as an ASCII string. The values
1966  * are treated as milliseconds, and converted to jiffies when they are stored.
1967  *
1968  * This routine will ensure the values are within the range specified by
1969  * table->extra1 (min) and table->extra2 (max).
1970  *
1971  * Returns 0 on success.
1972  */
1973 int proc_doulongvec_ms_jiffies_minmax(ctl_table *table, int write,
1974                                       struct file *filp,
1975                                       void __user *buffer,
1976                                       size_t *lenp, loff_t *ppos)
1977 {
1978     return do_proc_doulongvec_minmax(table, write, filp, buffer,
1979                                      lenp, ppos, HZ, 1000l);
1980 }
1981
1982
1983 static int do_proc_dointvec_jiffies_conv(int *negp, unsigned long *lvalp,
1984                                          int *valp,
1985                                          int write, void *data)
1986 {
1987         if (write) {
1988                 *valp = *negp ? -(*lvalp*HZ) : (*lvalp*HZ);
1989         } else {
1990                 int val = *valp;
1991                 unsigned long lval;
1992                 if (val < 0) {
1993                         *negp = -1;
1994                         lval = (unsigned long)-val;
1995                 } else {
1996                         *negp = 0;
1997                         lval = (unsigned long)val;
1998                 }
1999                 *lvalp = lval / HZ;
2000         }
2001         return 0;
2002 }
2003
2004 static int do_proc_dointvec_userhz_jiffies_conv(int *negp, unsigned long *lvalp,
2005                                                 int *valp,
2006                                                 int write, void *data)
2007 {
2008         if (write) {
2009                 *valp = clock_t_to_jiffies(*negp ? -*lvalp : *lvalp);
2010         } else {
2011                 int val = *valp;
2012                 unsigned long lval;
2013                 if (val < 0) {
2014                         *negp = -1;
2015                         lval = (unsigned long)-val;
2016                 } else {
2017                         *negp = 0;
2018                         lval = (unsigned long)val;
2019                 }
2020                 *lvalp = jiffies_to_clock_t(lval);
2021         }
2022         return 0;
2023 }
2024
2025 static int do_proc_dointvec_ms_jiffies_conv(int *negp, unsigned long *lvalp,
2026                                             int *valp,
2027                                             int write, void *data)
2028 {
2029         if (write) {
2030                 *valp = msecs_to_jiffies(*negp ? -*lvalp : *lvalp);
2031         } else {
2032                 int val = *valp;
2033                 unsigned long lval;
2034                 if (val < 0) {
2035                         *negp = -1;
2036                         lval = (unsigned long)-val;
2037                 } else {
2038                         *negp = 0;
2039                         lval = (unsigned long)val;
2040                 }
2041                 *lvalp = jiffies_to_msecs(lval);
2042         }
2043         return 0;
2044 }
2045
2046 /**
2047  * proc_dointvec_jiffies - read a vector of integers as seconds
2048  * @table: the sysctl table
2049  * @write: %TRUE if this is a write to the sysctl file
2050  * @filp: the file structure
2051  * @buffer: the user buffer
2052  * @lenp: the size of the user buffer
2053  * @ppos: file position
2054  *
2055  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2056  * values from/to the user buffer, treated as an ASCII string.
2057  * The values read are assumed to be in seconds, and are converted into
2058  * jiffies.
2059  *
2060  * Returns 0 on success.
2061  */
2062 int proc_dointvec_jiffies(ctl_table *table, int write, struct file *filp,
2063                           void __user *buffer, size_t *lenp, loff_t *ppos)
2064 {
2065     return do_proc_dointvec(table,write,filp,buffer,lenp,ppos,
2066                             do_proc_dointvec_jiffies_conv,NULL);
2067 }
2068
2069 /**
2070  * proc_dointvec_userhz_jiffies - read a vector of integers as 1/USER_HZ seconds
2071  * @table: the sysctl table
2072  * @write: %TRUE if this is a write to the sysctl file
2073  * @filp: the file structure
2074  * @buffer: the user buffer
2075  * @lenp: the size of the user buffer
2076  * @ppos: pointer to the file position
2077  *
2078  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2079  * values from/to the user buffer, treated as an ASCII string.
2080  * The values read are assumed to be in 1/USER_HZ seconds, and
2081  * are converted into jiffies.
2082  *
2083  * Returns 0 on success.
2084  */
2085 int proc_dointvec_userhz_jiffies(ctl_table *table, int write, struct file *filp,
2086                                  void __user *buffer, size_t *lenp, loff_t *ppos)
2087 {
2088     return do_proc_dointvec(table,write,filp,buffer,lenp,ppos,
2089                             do_proc_dointvec_userhz_jiffies_conv,NULL);
2090 }
2091
2092 /**
2093  * proc_dointvec_ms_jiffies - read a vector of integers as 1 milliseconds
2094  * @table: the sysctl table
2095  * @write: %TRUE if this is a write to the sysctl file
2096  * @filp: the file structure
2097  * @buffer: the user buffer
2098  * @lenp: the size of the user buffer
2099  * @ppos: file position
2100  * @ppos: the current position in the file
2101  *
2102  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2103  * values from/to the user buffer, treated as an ASCII string.
2104  * The values read are assumed to be in 1/1000 seconds, and
2105  * are converted into jiffies.
2106  *
2107  * Returns 0 on success.
2108  */
2109 int proc_dointvec_ms_jiffies(ctl_table *table, int write, struct file *filp,
2110                              void __user *buffer, size_t *lenp, loff_t *ppos)
2111 {
2112         return do_proc_dointvec(table, write, filp, buffer, lenp, ppos,
2113                                 do_proc_dointvec_ms_jiffies_conv, NULL);
2114 }
2115
2116 #else /* CONFIG_PROC_FS */
2117
2118 int proc_dostring(ctl_table *table, int write, struct file *filp,
2119                   void __user *buffer, size_t *lenp, loff_t *ppos)
2120 {
2121         return -ENOSYS;
2122 }
2123
2124 static int proc_doutsstring(ctl_table *table, int write, struct file *filp,
2125                             void __user *buffer, size_t *lenp, loff_t *ppos)
2126 {
2127         return -ENOSYS;
2128 }
2129
2130 int proc_dointvec(ctl_table *table, int write, struct file *filp,
2131                   void __user *buffer, size_t *lenp, loff_t *ppos)
2132 {
2133         return -ENOSYS;
2134 }
2135
2136 int proc_dointvec_bset(ctl_table *table, int write, struct file *filp,
2137                         void __user *buffer, size_t *lenp, loff_t *ppos)
2138 {
2139         return -ENOSYS;
2140 }
2141
2142 int proc_dointvec_minmax(ctl_table *table, int write, struct file *filp,
2143                     void __user *buffer, size_t *lenp, loff_t *ppos)
2144 {
2145         return -ENOSYS;
2146 }
2147
2148 int proc_dointvec_jiffies(ctl_table *table, int write, struct file *filp,
2149                     void __user *buffer, size_t *lenp, loff_t *ppos)
2150 {
2151         return -ENOSYS;
2152 }
2153
2154 int proc_dointvec_userhz_jiffies(ctl_table *table, int write, struct file *filp,
2155                     void __user *buffer, size_t *lenp, loff_t *ppos)
2156 {
2157         return -ENOSYS;
2158 }
2159
2160 int proc_dointvec_ms_jiffies(ctl_table *table, int write, struct file *filp,
2161                              void __user *buffer, size_t *lenp, loff_t *ppos)
2162 {
2163         return -ENOSYS;
2164 }
2165
2166 int proc_doulongvec_minmax(ctl_table *table, int write, struct file *filp,
2167                     void __user *buffer, size_t *lenp, loff_t *ppos)
2168 {
2169         return -ENOSYS;
2170 }
2171
2172 int proc_doulongvec_ms_jiffies_minmax(ctl_table *table, int write,
2173                                       struct file *filp,
2174                                       void __user *buffer,
2175                                       size_t *lenp, loff_t *ppos)
2176 {
2177     return -ENOSYS;
2178 }
2179
2180
2181 #endif /* CONFIG_PROC_FS */
2182
2183
2184 /*
2185  * General sysctl support routines
2186  */
2187
2188 /* The generic string strategy routine: */
2189 int sysctl_string(ctl_table *table, int __user *name, int nlen,
2190                   void __user *oldval, size_t __user *oldlenp,
2191                   void __user *newval, size_t newlen, void **context)
2192 {
2193         if (!table->data || !table->maxlen)
2194                 return -ENOTDIR;
2195
2196         if (oldval && oldlenp) {
2197                 size_t bufsize;
2198                 if (get_user(bufsize, oldlenp))
2199                         return -EFAULT;
2200                 if (bufsize) {
2201                         size_t len = strlen(table->data), copied;
2202
2203                         /* This shouldn't trigger for a well-formed sysctl */
2204                         if (len > table->maxlen)
2205                                 len = table->maxlen;
2206
2207                         /* Copy up to a max of bufsize-1 bytes of the string */
2208                         copied = (len >= bufsize) ? bufsize - 1 : len;
2209
2210                         if (copy_to_user(oldval, table->data, copied) ||
2211                             put_user(0, (char __user *)(oldval + copied)))
2212                                 return -EFAULT;
2213                         if (put_user(len, oldlenp))
2214                                 return -EFAULT;
2215                 }
2216         }
2217         if (newval && newlen) {
2218                 size_t len = newlen;
2219                 if (len > table->maxlen)
2220                         len = table->maxlen;
2221                 if(copy_from_user(table->data, newval, len))
2222                         return -EFAULT;
2223                 if (len == table->maxlen)
2224                         len--;
2225                 ((char *) table->data)[len] = 0;
2226         }
2227         return 1;
2228 }
2229
2230 /*
2231  * This function makes sure that all of the integers in the vector
2232  * are between the minimum and maximum values given in the arrays
2233  * table->extra1 and table->extra2, respectively.
2234  */
2235 int sysctl_intvec(ctl_table *table, int __user *name, int nlen,
2236                 void __user *oldval, size_t __user *oldlenp,
2237                 void __user *newval, size_t newlen, void **context)
2238 {
2239
2240         if (newval && newlen) {
2241                 int __user *vec = (int __user *) newval;
2242                 int *min = (int *) table->extra1;
2243                 int *max = (int *) table->extra2;
2244                 size_t length;
2245                 int i;
2246
2247                 if (newlen % sizeof(int) != 0)
2248                         return -EINVAL;
2249
2250                 if (!table->extra1 && !table->extra2)
2251                         return 0;
2252
2253                 if (newlen > table->maxlen)
2254                         newlen = table->maxlen;
2255                 length = newlen / sizeof(int);
2256
2257                 for (i = 0; i < length; i++) {
2258                         int value;
2259                         if (get_user(value, vec + i))
2260                                 return -EFAULT;
2261                         if (min && value < min[i])
2262                                 return -EINVAL;
2263                         if (max && value > max[i])
2264                                 return -EINVAL;
2265                 }
2266         }
2267         return 0;
2268 }
2269
2270 /* Strategy function to convert jiffies to seconds */
2271 int sysctl_jiffies(ctl_table *table, int __user *name, int nlen,
2272                 void __user *oldval, size_t __user *oldlenp,
2273                 void __user *newval, size_t newlen, void **context)
2274 {
2275         if (oldval) {
2276                 size_t olen;
2277                 if (oldlenp) {
2278                         if (get_user(olen, oldlenp))
2279                                 return -EFAULT;
2280                         if (olen!=sizeof(int))
2281                                 return -EINVAL;
2282                 }
2283                 if (put_user(*(int *)(table->data)/HZ, (int __user *)oldval) ||
2284                     (oldlenp && put_user(sizeof(int),oldlenp)))
2285                         return -EFAULT;
2286         }
2287         if (newval && newlen) {
2288                 int new;
2289                 if (newlen != sizeof(int))
2290                         return -EINVAL;
2291                 if (get_user(new, (int __user *)newval))
2292                         return -EFAULT;
2293                 *(int *)(table->data) = new*HZ;
2294         }
2295         return 1;
2296 }
2297
2298 /* Strategy function to convert jiffies to seconds */
2299 int sysctl_ms_jiffies(ctl_table *table, int __user *name, int nlen,
2300                 void __user *oldval, size_t __user *oldlenp,
2301                 void __user *newval, size_t newlen, void **context)
2302 {
2303         if (oldval) {
2304                 size_t olen;
2305                 if (oldlenp) {
2306                         if (get_user(olen, oldlenp))
2307                                 return -EFAULT;
2308                         if (olen!=sizeof(int))
2309                                 return -EINVAL;
2310                 }
2311                 if (put_user(jiffies_to_msecs(*(int *)(table->data)), (int __user *)oldval) ||
2312                     (oldlenp && put_user(sizeof(int),oldlenp)))
2313                         return -EFAULT;
2314         }
2315         if (newval && newlen) {
2316                 int new;
2317                 if (newlen != sizeof(int))
2318                         return -EINVAL;
2319                 if (get_user(new, (int __user *)newval))
2320                         return -EFAULT;
2321                 *(int *)(table->data) = msecs_to_jiffies(new);
2322         }
2323         return 1;
2324 }
2325
2326 #else /* CONFIG_SYSCTL */
2327
2328
2329 asmlinkage long sys_sysctl(struct __sysctl_args __user *args)
2330 {
2331         return -ENOSYS;
2332 }
2333
2334 int sysctl_string(ctl_table *table, int __user *name, int nlen,
2335                   void __user *oldval, size_t __user *oldlenp,
2336                   void __user *newval, size_t newlen, void **context)
2337 {
2338         return -ENOSYS;
2339 }
2340
2341 int sysctl_intvec(ctl_table *table, int __user *name, int nlen,
2342                 void __user *oldval, size_t __user *oldlenp,
2343                 void __user *newval, size_t newlen, void **context)
2344 {
2345         return -ENOSYS;
2346 }
2347
2348 int sysctl_jiffies(ctl_table *table, int __user *name, int nlen,
2349                 void __user *oldval, size_t __user *oldlenp,
2350                 void __user *newval, size_t newlen, void **context)
2351 {
2352         return -ENOSYS;
2353 }
2354
2355 int sysctl_ms_jiffies(ctl_table *table, int __user *name, int nlen,
2356                 void __user *oldval, size_t __user *oldlenp,
2357                 void __user *newval, size_t newlen, void **context)
2358 {
2359         return -ENOSYS;
2360 }
2361
2362 int proc_dostring(ctl_table *table, int write, struct file *filp,
2363                   void __user *buffer, size_t *lenp, loff_t *ppos)
2364 {
2365         return -ENOSYS;
2366 }
2367
2368 int proc_dointvec(ctl_table *table, int write, struct file *filp,
2369                   void __user *buffer, size_t *lenp, loff_t *ppos)
2370 {
2371         return -ENOSYS;
2372 }
2373
2374 int proc_dointvec_bset(ctl_table *table, int write, struct file *filp,
2375                         void __user *buffer, size_t *lenp, loff_t *ppos)
2376 {
2377         return -ENOSYS;
2378 }
2379
2380 int proc_dointvec_minmax(ctl_table *table, int write, struct file *filp,
2381                     void __user *buffer, size_t *lenp, loff_t *ppos)
2382 {
2383         return -ENOSYS;
2384 }
2385
2386 int proc_dointvec_jiffies(ctl_table *table, int write, struct file *filp,
2387                           void __user *buffer, size_t *lenp, loff_t *ppos)
2388 {
2389         return -ENOSYS;
2390 }
2391
2392 int proc_dointvec_userhz_jiffies(ctl_table *table, int write, struct file *filp,
2393                           void __user *buffer, size_t *lenp, loff_t *ppos)
2394 {
2395         return -ENOSYS;
2396 }
2397
2398 int proc_dointvec_ms_jiffies(ctl_table *table, int write, struct file *filp,
2399                              void __user *buffer, size_t *lenp, loff_t *ppos)
2400 {
2401         return -ENOSYS;
2402 }
2403
2404 int proc_doulongvec_minmax(ctl_table *table, int write, struct file *filp,
2405                     void __user *buffer, size_t *lenp, loff_t *ppos)
2406 {
2407         return -ENOSYS;
2408 }
2409
2410 int proc_doulongvec_ms_jiffies_minmax(ctl_table *table, int write,
2411                                       struct file *filp,
2412                                       void __user *buffer,
2413                                       size_t *lenp, loff_t *ppos)
2414 {
2415     return -ENOSYS;
2416 }
2417
2418 struct ctl_table_header * register_sysctl_table(ctl_table * table,
2419                                                 int insert_at_head)
2420 {
2421         return NULL;
2422 }
2423
2424 void unregister_sysctl_table(struct ctl_table_header * table)
2425 {
2426 }
2427
2428 #endif /* CONFIG_SYSCTL */
2429
2430 /*
2431  * No sense putting this after each symbol definition, twice,
2432  * exception granted :-)
2433  */
2434 EXPORT_SYMBOL(proc_dointvec);
2435 EXPORT_SYMBOL(proc_dointvec_jiffies);
2436 EXPORT_SYMBOL(proc_dointvec_minmax);
2437 EXPORT_SYMBOL(proc_dointvec_userhz_jiffies);
2438 EXPORT_SYMBOL(proc_dointvec_ms_jiffies);
2439 EXPORT_SYMBOL(proc_dostring);
2440 EXPORT_SYMBOL(proc_doulongvec_minmax);
2441 EXPORT_SYMBOL(proc_doulongvec_ms_jiffies_minmax);
2442 EXPORT_SYMBOL(register_sysctl_table);
2443 EXPORT_SYMBOL(sysctl_intvec);
2444 EXPORT_SYMBOL(sysctl_jiffies);
2445 EXPORT_SYMBOL(sysctl_ms_jiffies);
2446 EXPORT_SYMBOL(sysctl_string);
2447 EXPORT_SYMBOL(unregister_sysctl_table);