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/capability.h>
  29 #include <linux/ctype.h>
  30 #include <linux/utsname.h>
  31 #include <linux/capability.h>
  32 #include <linux/smp_lock.h>
  33 #include <linux/init.h>
  34 #include <linux/kernel.h>
  35 #include <linux/kobject.h>
  36 #include <linux/net.h>
  37 #include <linux/sysrq.h>
  38 #include <linux/highuid.h>
  39 #include <linux/writeback.h>
  40 #include <linux/hugetlb.h>
  41 #include <linux/security.h>
  42 #include <linux/initrd.h>
  43 #include <linux/times.h>
  44 #include <linux/limits.h>
  45 #include <linux/dcache.h>
  46 #include <linux/syscalls.h>
  47 #include <linux/nfs_fs.h>
  48 #include <linux/acpi.h>
  49 #include <linux/vserver/cvirt.h>
  50
  51 #include <asm/uaccess.h>
  52 #include <asm/processor.h>
  53
  54 extern int proc_nr_files(ctl_table *table, int write, struct file *filp,
  55                      void __user *buffer, size_t *lenp, loff_t *ppos);
  56
  57 #if defined(CONFIG_SYSCTL)
  58
  59 /* External variables not in a header file. */
  60 extern int C_A_D;
  61 extern int sysctl_overcommit_memory;
  62 extern int sysctl_overcommit_ratio;
  63 extern int max_threads;
  64 extern int sysrq_enabled;
  65 extern int core_uses_pid;
  66 extern int suid_dumpable;
  67 extern char core_pattern[];
  68 extern int cad_pid;
  69 extern int pid_max;
  70 extern int min_free_kbytes;
  71 extern int printk_ratelimit_jiffies;
  72 extern int printk_ratelimit_burst;
  73 extern int pid_max_min, pid_max_max;
  74 extern int sysctl_drop_caches;
  75 extern int percpu_pagelist_fraction;
  76
  77 #if defined(CONFIG_X86_LOCAL_APIC) && defined(CONFIG_X86)
  78 int unknown_nmi_panic;
  79 extern int proc_unknown_nmi_panic(ctl_table *, int, struct file *,
  80                                   void __user *, size_t *, loff_t *);
  81 #endif
  82
  83 /* this is needed for the proc_dointvec_minmax for [fs_]overflow UID and GID */
  84 static int maxolduid = 65535;
  85 static int minolduid;
  86 static int min_percpu_pagelist_fract = 8;
  87
  88 static int ngroups_max = NGROUPS_MAX;
  89
  90 #ifdef CONFIG_KMOD
  91 extern char modprobe_path[];
  92 #endif
  93 extern char vshelper_path[];
  94 #ifdef CONFIG_CHR_DEV_SG
  95 extern int sg_big_buff;
  96 #endif
  97 #ifdef CONFIG_SYSVIPC
  98 extern size_t shm_ctlmax;
  99 extern size_t shm_ctlall;
 100 extern int shm_ctlmni;
 101 extern int msg_ctlmax;
 102 extern int msg_ctlmnb;
 103 extern int msg_ctlmni;
 104 extern int sem_ctls[];
 105 #endif
 106
 107 #ifdef __sparc__
 108 extern char reboot_command [];
 109 extern int stop_a_enabled;
 110 extern int scons_pwroff;
 111 #endif
 112
 113 #ifdef __hppa__
 114 extern int pwrsw_enabled;
 115 extern int unaligned_enabled;
 116 #endif
 117
 118 #ifdef CONFIG_S390
 119 #ifdef CONFIG_MATHEMU
 120 extern int sysctl_ieee_emulation_warnings;
 121 #endif
 122 extern int sysctl_userprocess_debug;
 123 extern int spin_retry;
 124 #endif
 125
 126 extern int sysctl_hz_timer;
 127
 128 #ifdef CONFIG_BSD_PROCESS_ACCT
 129 extern int acct_parm[];
 130 #endif
 131
 132 #ifdef CONFIG_IA64
 133 extern int no_unaligned_warning;
 134 #endif
 135
 136 static int parse_table(int __user *, int, void __user *, size_t __user *, void __user *, size_t,
 137                        ctl_table *, void **);
 138 static int proc_doutsstring(ctl_table *table, int write, struct file *filp,
 139                   void __user *buffer, size_t *lenp, loff_t *ppos);
 140
 141 static ctl_table root_table[];
 142 static struct ctl_table_header root_table_header =
 143         { root_table, LIST_HEAD_INIT(root_table_header.ctl_entry) };
 144
 145 static ctl_table kern_table[];
 146 static ctl_table vm_table[];
 147 static ctl_table proc_table[];
 148 static ctl_table fs_table[];
 149 static ctl_table debug_table[];
 150 static ctl_table dev_table[];
 151 extern ctl_table random_table[];
 152 #ifdef CONFIG_UNIX98_PTYS
 153 extern ctl_table pty_table[];
 154 #endif
 155 #ifdef CONFIG_INOTIFY
 156 extern ctl_table inotify_table[];
 157 #endif
 158
 159 #ifdef HAVE_ARCH_PICK_MMAP_LAYOUT
 160 int sysctl_legacy_va_layout;
 161 #endif
 162
 163 /* /proc declarations: */
 164
 165 #ifdef CONFIG_PROC_FS
 166
 167 static ssize_t proc_readsys(struct file *, char __user *, size_t, loff_t *);
 168 static ssize_t proc_writesys(struct file *, const char __user *, size_t, loff_t *);
 169 static int proc_opensys(struct inode *, struct file *);
 170
 171 struct file_operations proc_sys_file_operations = {
 172         .open           = proc_opensys,
 173         .read           = proc_readsys,
 174         .write          = proc_writesys,
 175 };
 176
 177 extern struct proc_dir_entry *proc_sys_root;
 178
 179 static void register_proc_table(ctl_table *, struct proc_dir_entry *, void *);
 180 static void unregister_proc_table(ctl_table *, struct proc_dir_entry *);
 181 #endif
 182
 183 /* The default sysctl tables: */
 184
 185 static ctl_table root_table[] = {
 186         {
 187                 .ctl_name       = CTL_KERN,
 188                 .procname       = "kernel",
 189                 .mode           = 0555,
 190                 .child          = kern_table,
 191         },
 192         {
 193                 .ctl_name       = CTL_VM,
 194                 .procname       = "vm",
 195                 .mode           = 0555,
 196                 .child          = vm_table,
 197         },
 198 #ifdef CONFIG_NET
 199         {
 200                 .ctl_name       = CTL_NET,
 201                 .procname       = "net",
 202                 .mode           = 0555,
 203                 .child          = net_table,
 204         },
 205 #endif
 206         {
 207                 .ctl_name       = CTL_PROC,
 208                 .procname       = "proc",
 209                 .mode           = 0555,
 210                 .child          = proc_table,
 211         },
 212         {
 213                 .ctl_name       = CTL_FS,
 214                 .procname       = "fs",
 215                 .mode           = 0555,
 216                 .child          = fs_table,
 217         },
 218         {
 219                 .ctl_name       = CTL_DEBUG,
 220                 .procname       = "debug",
 221                 .mode           = 0555,
 222                 .child          = debug_table,
 223         },
 224         {
 225                 .ctl_name       = CTL_DEV,
 226                 .procname       = "dev",
 227                 .mode           = 0555,
 228                 .child          = dev_table,
 229         },
 230
 231         { .ctl_name = 0 }
 232 };
 233
 234 static ctl_table kern_table[] = {
 235         {
 236                 .ctl_name       = KERN_OSTYPE,
 237                 .procname       = "ostype",
 238                 .data           = system_utsname.sysname,
 239                 .maxlen         = sizeof(system_utsname.sysname),
 240                 .mode           = 0444,
 241                 .proc_handler   = &proc_doutsstring,
 242                 .virt_handler   = &vx_uts_virt_handler,
 243                 .strategy       = &sysctl_string,
 244         },
 245         {
 246                 .ctl_name       = KERN_OSRELEASE,
 247                 .procname       = "osrelease",
 248                 .data           = system_utsname.release,
 249                 .maxlen         = sizeof(system_utsname.release),
 250                 .mode           = 0444,
 251                 .proc_handler   = &proc_doutsstring,
 252                 .virt_handler   = &vx_uts_virt_handler,
 253                 .strategy       = &sysctl_string,
 254         },
 255         {
 256                 .ctl_name       = KERN_VERSION,
 257                 .procname       = "version",
 258                 .data           = system_utsname.version,
 259                 .maxlen         = sizeof(system_utsname.version),
 260                 .mode           = 0444,
 261                 .proc_handler   = &proc_doutsstring,
 262                 .virt_handler   = &vx_uts_virt_handler,
 263                 .strategy       = &sysctl_string,
 264         },
 265         {
 266                 .ctl_name       = KERN_NODENAME,
 267                 .procname       = "hostname",
 268                 .data           = system_utsname.nodename,
 269                 .maxlen         = sizeof(system_utsname.nodename),
 270                 .mode           = 0644,
 271                 .proc_handler   = &proc_doutsstring,
 272                 .virt_handler   = &vx_uts_virt_handler,
 273                 .strategy       = &sysctl_string,
 274         },
 275         {
 276                 .ctl_name       = KERN_DOMAINNAME,
 277                 .procname       = "domainname",
 278                 .data           = system_utsname.domainname,
 279                 .maxlen         = sizeof(system_utsname.domainname),
 280                 .mode           = 0644,
 281                 .proc_handler   = &proc_doutsstring,
 282                 .virt_handler   = &vx_uts_virt_handler,
 283                 .strategy       = &sysctl_string,
 284         },
 285         {
 286                 .ctl_name       = KERN_PANIC,
 287                 .procname       = "panic",
 288                 .data           = &panic_timeout,
 289                 .maxlen         = sizeof(int),
 290                 .mode           = 0644,
 291                 .proc_handler   = &proc_dointvec,
 292         },
 293         {
 294                 .ctl_name       = KERN_CORE_USES_PID,
 295                 .procname       = "core_uses_pid",
 296                 .data           = &core_uses_pid,
 297                 .maxlen         = sizeof(int),
 298                 .mode           = 0644,
 299                 .proc_handler   = &proc_dointvec,
 300         },
 301         {
 302                 .ctl_name       = KERN_CORE_PATTERN,
 303                 .procname       = "core_pattern",
 304                 .data           = core_pattern,
 305                 .maxlen         = 64,
 306                 .mode           = 0644,
 307                 .proc_handler   = &proc_dostring,
 308                 .strategy       = &sysctl_string,
 309         },
 310         {
 311                 .ctl_name       = KERN_TAINTED,
 312                 .procname       = "tainted",
 313                 .data           = &tainted,
 314                 .maxlen         = sizeof(int),
 315                 .mode           = 0444,
 316                 .proc_handler   = &proc_dointvec,
 317         },
 318         {
 319                 .ctl_name       = KERN_CAP_BSET,
 320                 .procname       = "cap-bound",
 321                 .data           = &cap_bset,
 322                 .maxlen         = sizeof(kernel_cap_t),
 323                 .mode           = 0600,
 324                 .proc_handler   = &proc_dointvec_bset,
 325         },
 326 #ifdef CONFIG_BLK_DEV_INITRD
 327         {
 328                 .ctl_name       = KERN_REALROOTDEV,
 329                 .procname       = "real-root-dev",
 330                 .data           = &real_root_dev,
 331                 .maxlen         = sizeof(int),
 332                 .mode           = 0644,
 333                 .proc_handler   = &proc_dointvec,
 334         },
 335 #endif
 336 #ifdef __sparc__
 337         {
 338                 .ctl_name       = KERN_SPARC_REBOOT,
 339                 .procname       = "reboot-cmd",
 340                 .data           = reboot_command,
 341                 .maxlen         = 256,
 342                 .mode           = 0644,
 343                 .proc_handler   = &proc_dostring,
 344                 .strategy       = &sysctl_string,
 345         },
 346         {
 347                 .ctl_name       = KERN_SPARC_STOP_A,
 348                 .procname       = "stop-a",
 349                 .data           = &stop_a_enabled,
 350                 .maxlen         = sizeof (int),
 351                 .mode           = 0644,
 352                 .proc_handler   = &proc_dointvec,
 353         },
 354         {
 355                 .ctl_name       = KERN_SPARC_SCONS_PWROFF,
 356                 .procname       = "scons-poweroff",
 357                 .data           = &scons_pwroff,
 358                 .maxlen         = sizeof (int),
 359                 .mode           = 0644,
 360                 .proc_handler   = &proc_dointvec,
 361         },
 362 #endif
 363 #ifdef __hppa__
 364         {
 365                 .ctl_name       = KERN_HPPA_PWRSW,
 366                 .procname       = "soft-power",
 367                 .data           = &pwrsw_enabled,
 368                 .maxlen         = sizeof (int),
 369                 .mode           = 0644,
 370                 .proc_handler   = &proc_dointvec,
 371         },
 372         {
 373                 .ctl_name       = KERN_HPPA_UNALIGNED,
 374                 .procname       = "unaligned-trap",
 375                 .data           = &unaligned_enabled,
 376                 .maxlen         = sizeof (int),
 377                 .mode           = 0644,
 378                 .proc_handler   = &proc_dointvec,
 379         },
 380 #endif
 381         {
 382                 .ctl_name       = KERN_CTLALTDEL,
 383                 .procname       = "ctrl-alt-del",
 384                 .data           = &C_A_D,
 385                 .maxlen         = sizeof(int),
 386                 .mode           = 0644,
 387                 .proc_handler   = &proc_dointvec,
 388         },
 389         {
 390                 .ctl_name       = KERN_PRINTK,
 391                 .procname       = "printk",
 392                 .data           = &console_loglevel,
 393                 .maxlen         = 4*sizeof(int),
 394                 .mode           = 0644,
 395                 .proc_handler   = &proc_dointvec,
 396         },
 397 #ifdef CONFIG_KMOD
 398         {
 399                 .ctl_name       = KERN_MODPROBE,
 400                 .procname       = "modprobe",
 401                 .data           = &modprobe_path,
 402                 .maxlen         = KMOD_PATH_LEN,
 403                 .mode           = 0644,
 404                 .proc_handler   = &proc_dostring,
 405                 .strategy       = &sysctl_string,
 406         },
 407 #endif
 408 #ifdef CONFIG_HOTPLUG
 409         {
 410                 .ctl_name       = KERN_HOTPLUG,
 411                 .procname       = "hotplug",
 412                 .data           = &uevent_helper,
 413                 .maxlen         = UEVENT_HELPER_PATH_LEN,
 414                 .mode           = 0644,
 415                 .proc_handler   = &proc_dostring,
 416                 .strategy       = &sysctl_string,
 417         },
 418 #endif
 419         {
 420                 .ctl_name       = KERN_VSHELPER,
 421                 .procname       = "vshelper",
 422                 .data           = &vshelper_path,
 423                 .maxlen         = 256,
 424                 .mode           = 0644,
 425                 .proc_handler   = &proc_dostring,
 426                 .strategy       = &sysctl_string,
 427         },
 428 #ifdef CONFIG_CHR_DEV_SG
 429         {
 430                 .ctl_name       = KERN_SG_BIG_BUFF,
 431                 .procname       = "sg-big-buff",
 432                 .data           = &sg_big_buff,
 433                 .maxlen         = sizeof (int),
 434                 .mode           = 0444,
 435                 .proc_handler   = &proc_dointvec,
 436         },
 437 #endif
 438 #ifdef CONFIG_BSD_PROCESS_ACCT
 439         {
 440                 .ctl_name       = KERN_ACCT,
 441                 .procname       = "acct",
 442                 .data           = &acct_parm,
 443                 .maxlen         = 3*sizeof(int),
 444                 .mode           = 0644,
 445                 .proc_handler   = &proc_dointvec,
 446         },
 447 #endif
 448 #ifdef CONFIG_SYSVIPC
 449         {
 450                 .ctl_name       = KERN_SHMMAX,
 451                 .procname       = "shmmax",
 452                 .data           = &shm_ctlmax,
 453                 .maxlen         = sizeof (size_t),
 454                 .mode           = 0644,
 455                 .proc_handler   = &proc_doulongvec_minmax,
 456         },
 457         {
 458                 .ctl_name       = KERN_SHMALL,
 459                 .procname       = "shmall",
 460                 .data           = &shm_ctlall,
 461                 .maxlen         = sizeof (size_t),
 462                 .mode           = 0644,
 463                 .proc_handler   = &proc_doulongvec_minmax,
 464         },
 465         {
 466                 .ctl_name       = KERN_SHMMNI,
 467                 .procname       = "shmmni",
 468                 .data           = &shm_ctlmni,
 469                 .maxlen         = sizeof (int),
 470                 .mode           = 0644,
 471                 .proc_handler   = &proc_dointvec,
 472         },
 473         {
 474                 .ctl_name       = KERN_MSGMAX,
 475                 .procname       = "msgmax",
 476                 .data           = &msg_ctlmax,
 477                 .maxlen         = sizeof (int),
 478                 .mode           = 0644,
 479                 .proc_handler   = &proc_dointvec,
 480         },
 481         {
 482                 .ctl_name       = KERN_MSGMNI,
 483                 .procname       = "msgmni",
 484                 .data           = &msg_ctlmni,
 485                 .maxlen         = sizeof (int),
 486                 .mode           = 0644,
 487                 .proc_handler   = &proc_dointvec,
 488         },
 489         {
 490                 .ctl_name       = KERN_MSGMNB,
 491                 .procname       =  "msgmnb",
 492                 .data           = &msg_ctlmnb,
 493                 .maxlen         = sizeof (int),
 494                 .mode           = 0644,
 495                 .proc_handler   = &proc_dointvec,
 496         },
 497         {
 498                 .ctl_name       = KERN_SEM,
 499                 .procname       = "sem",
 500                 .data           = &sem_ctls,
 501                 .maxlen         = 4*sizeof (int),
 502                 .mode           = 0644,
 503                 .proc_handler   = &proc_dointvec,
 504         },
 505 #endif
 506 #ifdef CONFIG_MAGIC_SYSRQ
 507         {
 508                 .ctl_name       = KERN_SYSRQ,
 509                 .procname       = "sysrq",
 510                 .data           = &sysrq_enabled,
 511                 .maxlen         = sizeof (int),
 512                 .mode           = 0644,
 513                 .proc_handler   = &proc_dointvec,
 514         },
 515 #endif
 516         {
 517                 .ctl_name       = KERN_CADPID,
 518                 .procname       = "cad_pid",
 519                 .data           = &cad_pid,
 520                 .maxlen         = sizeof (int),
 521                 .mode           = 0600,
 522                 .proc_handler   = &proc_dointvec,
 523         },
 524         {
 525                 .ctl_name       = KERN_MAX_THREADS,
 526                 .procname       = "threads-max",
 527                 .data           = &max_threads,
 528                 .maxlen         = sizeof(int),
 529                 .mode           = 0644,
 530                 .proc_handler   = &proc_dointvec,
 531         },
 532         {
 533                 .ctl_name       = KERN_RANDOM,
 534                 .procname       = "random",
 535                 .mode           = 0555,
 536                 .child          = random_table,
 537         },
 538 #ifdef CONFIG_UNIX98_PTYS
 539         {
 540                 .ctl_name       = KERN_PTY,
 541                 .procname       = "pty",
 542                 .mode           = 0555,
 543                 .child          = pty_table,
 544         },
 545 #endif
 546         {
 547                 .ctl_name       = KERN_OVERFLOWUID,
 548                 .procname       = "overflowuid",
 549                 .data           = &overflowuid,
 550                 .maxlen         = sizeof(int),
 551                 .mode           = 0644,
 552                 .proc_handler   = &proc_dointvec_minmax,
 553                 .strategy       = &sysctl_intvec,
 554                 .extra1         = &minolduid,
 555                 .extra2         = &maxolduid,
 556         },
 557         {
 558                 .ctl_name       = KERN_OVERFLOWGID,
 559                 .procname       = "overflowgid",
 560                 .data           = &overflowgid,
 561                 .maxlen         = sizeof(int),
 562                 .mode           = 0644,
 563                 .proc_handler   = &proc_dointvec_minmax,
 564                 .strategy       = &sysctl_intvec,
 565                 .extra1         = &minolduid,
 566                 .extra2         = &maxolduid,
 567         },
 568 #ifdef CONFIG_S390
 569 #ifdef CONFIG_MATHEMU
 570         {
 571                 .ctl_name       = KERN_IEEE_EMULATION_WARNINGS,
 572                 .procname       = "ieee_emulation_warnings",
 573                 .data           = &sysctl_ieee_emulation_warnings,
 574                 .maxlen         = sizeof(int),
 575                 .mode           = 0644,
 576                 .proc_handler   = &proc_dointvec,
 577         },
 578 #endif
 579 #ifdef CONFIG_NO_IDLE_HZ
 580         {
 581                 .ctl_name       = KERN_HZ_TIMER,
 582                 .procname       = "hz_timer",
 583                 .data           = &sysctl_hz_timer,
 584                 .maxlen         = sizeof(int),
 585                 .mode           = 0644,
 586                 .proc_handler   = &proc_dointvec,
 587         },
 588 #endif
 589         {
 590                 .ctl_name       = KERN_S390_USER_DEBUG_LOGGING,
 591                 .procname       = "userprocess_debug",
 592                 .data           = &sysctl_userprocess_debug,
 593                 .maxlen         = sizeof(int),
 594                 .mode           = 0644,
 595                 .proc_handler   = &proc_dointvec,
 596         },
 597 #endif
 598         {
 599                 .ctl_name       = KERN_PIDMAX,
 600                 .procname       = "pid_max",
 601                 .data           = &pid_max,
 602                 .maxlen         = sizeof (int),
 603                 .mode           = 0644,
 604                 .proc_handler   = &proc_dointvec_minmax,
 605                 .strategy       = sysctl_intvec,
 606                 .extra1         = &pid_max_min,
 607                 .extra2         = &pid_max_max,
 608         },
 609         {
 610                 .ctl_name       = KERN_PANIC_ON_OOPS,
 611                 .procname       = "panic_on_oops",
 612                 .data           = &panic_on_oops,
 613                 .maxlen         = sizeof(int),
 614                 .mode           = 0644,
 615                 .proc_handler   = &proc_dointvec,
 616         },
 617         {
 618                 .ctl_name       = KERN_PRINTK_RATELIMIT,
 619                 .procname       = "printk_ratelimit",
 620                 .data           = &printk_ratelimit_jiffies,
 621                 .maxlen         = sizeof(int),
 622                 .mode           = 0644,
 623                 .proc_handler   = &proc_dointvec_jiffies,
 624                 .strategy       = &sysctl_jiffies,
 625         },
 626         {
 627                 .ctl_name       = KERN_PRINTK_RATELIMIT_BURST,
 628                 .procname       = "printk_ratelimit_burst",
 629                 .data           = &printk_ratelimit_burst,
 630                 .maxlen         = sizeof(int),
 631                 .mode           = 0644,
 632                 .proc_handler   = &proc_dointvec,
 633         },
 634         {
 635                 .ctl_name       = KERN_NGROUPS_MAX,
 636                 .procname       = "ngroups_max",
 637                 .data           = &ngroups_max,
 638                 .maxlen         = sizeof (int),
 639                 .mode           = 0444,
 640                 .proc_handler   = &proc_dointvec,
 641         },
 642 #if defined(CONFIG_X86_LOCAL_APIC) && defined(CONFIG_X86)
 643         {
 644                 .ctl_name       = KERN_UNKNOWN_NMI_PANIC,
 645                 .procname       = "unknown_nmi_panic",
 646                 .data           = &unknown_nmi_panic,
 647                 .maxlen         = sizeof (int),
 648                 .mode           = 0644,
 649                 .proc_handler   = &proc_unknown_nmi_panic,
 650         },
 651 #endif
 652 #if defined(CONFIG_X86)
 653         {
 654                 .ctl_name       = KERN_BOOTLOADER_TYPE,
 655                 .procname       = "bootloader_type",
 656                 .data           = &bootloader_type,
 657                 .maxlen         = sizeof (int),
 658                 .mode           = 0444,
 659                 .proc_handler   = &proc_dointvec,
 660         },
 661 #endif
 662 #if defined(CONFIG_MMU)
 663         {
 664                 .ctl_name       = KERN_RANDOMIZE,
 665                 .procname       = "randomize_va_space",
 666                 .data           = &randomize_va_space,
 667                 .maxlen         = sizeof(int),
 668                 .mode           = 0644,
 669                 .proc_handler   = &proc_dointvec,
 670         },
 671 #endif
 672 #if defined(CONFIG_S390) && defined(CONFIG_SMP)
 673         {
 674                 .ctl_name       = KERN_SPIN_RETRY,
 675                 .procname       = "spin_retry",
 676                 .data           = &spin_retry,
 677                 .maxlen         = sizeof (int),
 678                 .mode           = 0644,
 679                 .proc_handler   = &proc_dointvec,
 680         },
 681 #endif
 682 #ifdef CONFIG_ACPI_SLEEP
 683         {
 684                 .ctl_name       = KERN_ACPI_VIDEO_FLAGS,
 685                 .procname       = "acpi_video_flags",
 686                 .data           = &acpi_video_flags,
 687                 .maxlen         = sizeof (unsigned long),
 688                 .mode           = 0644,
 689                 .proc_handler   = &proc_doulongvec_minmax,
 690         },
 691 #endif
 692 #ifdef CONFIG_IA64
 693         {
 694                 .ctl_name       = KERN_IA64_UNALIGNED,
 695                 .procname       = "ignore-unaligned-usertrap",
 696                 .data           = &no_unaligned_warning,
 697                 .maxlen         = sizeof (int),
 698                 .mode           = 0644,
 699                 .proc_handler   = &proc_dointvec,
 700         },
 701 #endif
 702         { .ctl_name = 0 }
 703 };
 704
 705 /* Constants for minimum and maximum testing in vm_table.
 706    We use these as one-element integer vectors. */
 707 static int zero;
 708 static int one_hundred = 100;
 709
 710
 711 static ctl_table vm_table[] = {
 712         {
 713                 .ctl_name       = VM_OVERCOMMIT_MEMORY,
 714                 .procname       = "overcommit_memory",
 715                 .data           = &sysctl_overcommit_memory,
 716                 .maxlen         = sizeof(sysctl_overcommit_memory),
 717                 .mode           = 0644,
 718                 .proc_handler   = &proc_dointvec,
 719         },
 720         {
 721                 .ctl_name       = VM_OVERCOMMIT_RATIO,
 722                 .procname       = "overcommit_ratio",
 723                 .data           = &sysctl_overcommit_ratio,
 724                 .maxlen         = sizeof(sysctl_overcommit_ratio),
 725                 .mode           = 0644,
 726                 .proc_handler   = &proc_dointvec,
 727         },
 728         {
 729                 .ctl_name       = VM_PAGE_CLUSTER,
 730                 .procname       = "page-cluster",
 731                 .data           = &page_cluster,
 732                 .maxlen         = sizeof(int),
 733                 .mode           = 0644,
 734                 .proc_handler   = &proc_dointvec,
 735         },
 736         {
 737                 .ctl_name       = VM_DIRTY_BACKGROUND,
 738                 .procname       = "dirty_background_ratio",
 739                 .data           = &dirty_background_ratio,
 740                 .maxlen         = sizeof(dirty_background_ratio),
 741                 .mode           = 0644,
 742                 .proc_handler   = &proc_dointvec_minmax,
 743                 .strategy       = &sysctl_intvec,
 744                 .extra1         = &zero,
 745                 .extra2         = &one_hundred,
 746         },
 747         {
 748                 .ctl_name       = VM_DIRTY_RATIO,
 749                 .procname       = "dirty_ratio",
 750                 .data           = &vm_dirty_ratio,
 751                 .maxlen         = sizeof(vm_dirty_ratio),
 752                 .mode           = 0644,
 753                 .proc_handler   = &proc_dointvec_minmax,
 754                 .strategy       = &sysctl_intvec,
 755                 .extra1         = &zero,
 756                 .extra2         = &one_hundred,
 757         },
 758         {
 759                 .ctl_name       = VM_DIRTY_WB_CS,
 760                 .procname       = "dirty_writeback_centisecs",
 761                 .data           = &dirty_writeback_centisecs,
 762                 .maxlen         = sizeof(dirty_writeback_centisecs),
 763                 .mode           = 0644,
 764                 .proc_handler   = &dirty_writeback_centisecs_handler,
 765         },
 766         {
 767                 .ctl_name       = VM_DIRTY_EXPIRE_CS,
 768                 .procname       = "dirty_expire_centisecs",
 769                 .data           = &dirty_expire_centisecs,
 770                 .maxlen         = sizeof(dirty_expire_centisecs),
 771                 .mode           = 0644,
 772                 .proc_handler   = &proc_dointvec,
 773         },
 774         {
 775                 .ctl_name       = VM_NR_PDFLUSH_THREADS,
 776                 .procname       = "nr_pdflush_threads",
 777                 .data           = &nr_pdflush_threads,
 778                 .maxlen         = sizeof nr_pdflush_threads,
 779                 .mode           = 0444 /* read-only*/,
 780                 .proc_handler   = &proc_dointvec,
 781         },
 782         {
 783                 .ctl_name       = VM_SWAPPINESS,
 784                 .procname       = "swappiness",
 785                 .data           = &vm_swappiness,
 786                 .maxlen         = sizeof(vm_swappiness),
 787                 .mode           = 0644,
 788                 .proc_handler   = &proc_dointvec_minmax,
 789                 .strategy       = &sysctl_intvec,
 790                 .extra1         = &zero,
 791                 .extra2         = &one_hundred,
 792         },
 793 #ifdef CONFIG_HUGETLB_PAGE
 794          {
 795                 .ctl_name       = VM_HUGETLB_PAGES,
 796                 .procname       = "nr_hugepages",
 797                 .data           = &max_huge_pages,
 798                 .maxlen         = sizeof(unsigned long),
 799                 .mode           = 0644,
 800                 .proc_handler   = &hugetlb_sysctl_handler,
 801                 .extra1         = (void *)&hugetlb_zero,
 802                 .extra2         = (void *)&hugetlb_infinity,
 803          },
 804          {
 805                 .ctl_name       = VM_HUGETLB_GROUP,
 806                 .procname       = "hugetlb_shm_group",
 807                 .data           = &sysctl_hugetlb_shm_group,
 808                 .maxlen         = sizeof(gid_t),
 809                 .mode           = 0644,
 810                 .proc_handler   = &proc_dointvec,
 811          },
 812 #endif
 813         {
 814                 .ctl_name       = VM_LOWMEM_RESERVE_RATIO,
 815                 .procname       = "lowmem_reserve_ratio",
 816                 .data           = &sysctl_lowmem_reserve_ratio,
 817                 .maxlen         = sizeof(sysctl_lowmem_reserve_ratio),
 818                 .mode           = 0644,
 819                 .proc_handler   = &lowmem_reserve_ratio_sysctl_handler,
 820                 .strategy       = &sysctl_intvec,
 821         },
 822         {
 823                 .ctl_name       = VM_DROP_PAGECACHE,
 824                 .procname       = "drop_caches",
 825                 .data           = &sysctl_drop_caches,
 826                 .maxlen         = sizeof(int),
 827                 .mode           = 0644,
 828                 .proc_handler   = drop_caches_sysctl_handler,
 829                 .strategy       = &sysctl_intvec,
 830         },
 831         {
 832                 .ctl_name       = VM_MIN_FREE_KBYTES,
 833                 .procname       = "min_free_kbytes",
 834                 .data           = &min_free_kbytes,
 835                 .maxlen         = sizeof(min_free_kbytes),
 836                 .mode           = 0644,
 837                 .proc_handler   = &min_free_kbytes_sysctl_handler,
 838                 .strategy       = &sysctl_intvec,
 839                 .extra1         = &zero,
 840         },
 841         {
 842                 .ctl_name       = VM_PERCPU_PAGELIST_FRACTION,
 843                 .procname       = "percpu_pagelist_fraction",
 844                 .data           = &percpu_pagelist_fraction,
 845                 .maxlen         = sizeof(percpu_pagelist_fraction),
 846                 .mode           = 0644,
 847                 .proc_handler   = &percpu_pagelist_fraction_sysctl_handler,
 848                 .strategy       = &sysctl_intvec,
 849                 .extra1         = &min_percpu_pagelist_fract,
 850         },
 851 #ifdef CONFIG_MMU
 852         {
 853                 .ctl_name       = VM_MAX_MAP_COUNT,
 854                 .procname       = "max_map_count",
 855                 .data           = &sysctl_max_map_count,
 856                 .maxlen         = sizeof(sysctl_max_map_count),
 857                 .mode           = 0644,
 858                 .proc_handler   = &proc_dointvec
 859         },
 860 #endif
 861         {
 862                 .ctl_name       = VM_LAPTOP_MODE,
 863                 .procname       = "laptop_mode",
 864                 .data           = &laptop_mode,
 865                 .maxlen         = sizeof(laptop_mode),
 866                 .mode           = 0644,
 867                 .proc_handler   = &proc_dointvec,
 868                 .strategy       = &sysctl_intvec,
 869                 .extra1         = &zero,
 870         },
 871         {
 872                 .ctl_name       = VM_BLOCK_DUMP,
 873                 .procname       = "block_dump",
 874                 .data           = &block_dump,
 875                 .maxlen         = sizeof(block_dump),
 876                 .mode           = 0644,
 877                 .proc_handler   = &proc_dointvec,
 878                 .strategy       = &sysctl_intvec,
 879                 .extra1         = &zero,
 880         },
 881         {
 882                 .ctl_name       = VM_VFS_CACHE_PRESSURE,
 883                 .procname       = "vfs_cache_pressure",
 884                 .data           = &sysctl_vfs_cache_pressure,
 885                 .maxlen         = sizeof(sysctl_vfs_cache_pressure),
 886                 .mode           = 0644,
 887                 .proc_handler   = &proc_dointvec,
 888                 .strategy       = &sysctl_intvec,
 889                 .extra1         = &zero,
 890         },
 891 #ifdef HAVE_ARCH_PICK_MMAP_LAYOUT
 892         {
 893                 .ctl_name       = VM_LEGACY_VA_LAYOUT,
 894                 .procname       = "legacy_va_layout",
 895                 .data           = &sysctl_legacy_va_layout,
 896                 .maxlen         = sizeof(sysctl_legacy_va_layout),
 897                 .mode           = 0644,
 898                 .proc_handler   = &proc_dointvec,
 899                 .strategy       = &sysctl_intvec,
 900                 .extra1         = &zero,
 901         },
 902 #endif
 903 #ifdef CONFIG_SWAP
 904         {
 905                 .ctl_name       = VM_SWAP_TOKEN_TIMEOUT,
 906                 .procname       = "swap_token_timeout",
 907                 .data           = &swap_token_default_timeout,
 908                 .maxlen         = sizeof(swap_token_default_timeout),
 909                 .mode           = 0644,
 910                 .proc_handler   = &proc_dointvec_jiffies,
 911                 .strategy       = &sysctl_jiffies,
 912         },
 913 #endif
 914 #ifdef CONFIG_NUMA
 915         {
 916                 .ctl_name       = VM_ZONE_RECLAIM_MODE,
 917                 .procname       = "zone_reclaim_mode",
 918                 .data           = &zone_reclaim_mode,
 919                 .maxlen         = sizeof(zone_reclaim_mode),
 920                 .mode           = 0644,
 921                 .proc_handler   = &proc_dointvec,
 922                 .strategy       = &sysctl_intvec,
 923                 .extra1         = &zero,
 924         },
 925         {
 926                 .ctl_name       = VM_ZONE_RECLAIM_INTERVAL,
 927                 .procname       = "zone_reclaim_interval",
 928                 .data           = &zone_reclaim_interval,
 929                 .maxlen         = sizeof(zone_reclaim_interval),
 930                 .mode           = 0644,
 931                 .proc_handler   = &proc_dointvec_jiffies,
 932                 .strategy       = &sysctl_jiffies,
 933         },
 934 #endif
 935         { .ctl_name = 0 }
 936 };
 937
 938 static ctl_table proc_table[] = {
 939         { .ctl_name = 0 }
 940 };
 941
 942 static ctl_table fs_table[] = {
 943         {
 944                 .ctl_name       = FS_NRINODE,
 945                 .procname       = "inode-nr",
 946                 .data           = &inodes_stat,
 947                 .maxlen         = 2*sizeof(int),
 948                 .mode           = 0444,
 949                 .proc_handler   = &proc_dointvec,
 950         },
 951         {
 952                 .ctl_name       = FS_STATINODE,
 953                 .procname       = "inode-state",
 954                 .data           = &inodes_stat,
 955                 .maxlen         = 7*sizeof(int),
 956                 .mode           = 0444,
 957                 .proc_handler   = &proc_dointvec,
 958         },
 959         {
 960                 .ctl_name       = FS_NRFILE,
 961                 .procname       = "file-nr",
 962                 .data           = &files_stat,
 963                 .maxlen         = 3*sizeof(int),
 964                 .mode           = 0444,
 965                 .proc_handler   = &proc_nr_files,
 966         },
 967         {
 968                 .ctl_name       = FS_MAXFILE,
 969                 .procname       = "file-max",
 970                 .data           = &files_stat.max_files,
 971                 .maxlen         = sizeof(int),
 972                 .mode           = 0644,
 973                 .proc_handler   = &proc_dointvec,
 974         },
 975         {
 976                 .ctl_name       = FS_DENTRY,
 977                 .procname       = "dentry-state",
 978                 .data           = &dentry_stat,
 979                 .maxlen         = 6*sizeof(int),
 980                 .mode           = 0444,
 981                 .proc_handler   = &proc_dointvec,
 982         },
 983         {
 984                 .ctl_name       = FS_OVERFLOWUID,
 985                 .procname       = "overflowuid",
 986                 .data           = &fs_overflowuid,
 987                 .maxlen         = sizeof(int),
 988                 .mode           = 0644,
 989                 .proc_handler   = &proc_dointvec_minmax,
 990                 .strategy       = &sysctl_intvec,
 991                 .extra1         = &minolduid,
 992                 .extra2         = &maxolduid,
 993         },
 994         {
 995                 .ctl_name       = FS_OVERFLOWGID,
 996                 .procname       = "overflowgid",
 997                 .data           = &fs_overflowgid,
 998                 .maxlen         = sizeof(int),
 999                 .mode           = 0644,
1000                 .proc_handler   = &proc_dointvec_minmax,
1001                 .strategy       = &sysctl_intvec,
1002                 .extra1         = &minolduid,
1003                 .extra2         = &maxolduid,
1004         },
1005         {
1006                 .ctl_name       = FS_LEASES,
1007                 .procname       = "leases-enable",
1008                 .data           = &leases_enable,
1009                 .maxlen         = sizeof(int),
1010                 .mode           = 0644,
1011                 .proc_handler   = &proc_dointvec,
1012         },
1013 #ifdef CONFIG_DNOTIFY
1014         {
1015                 .ctl_name       = FS_DIR_NOTIFY,
1016                 .procname       = "dir-notify-enable",
1017                 .data           = &dir_notify_enable,
1018                 .maxlen         = sizeof(int),
1019                 .mode           = 0644,
1020                 .proc_handler   = &proc_dointvec,
1021         },
1022 #endif
1023 #ifdef CONFIG_MMU
1024         {
1025                 .ctl_name       = FS_LEASE_TIME,
1026                 .procname       = "lease-break-time",
1027                 .data           = &lease_break_time,
1028                 .maxlen         = sizeof(int),
1029                 .mode           = 0644,
1030                 .proc_handler   = &proc_dointvec,
1031         },
1032         {
1033                 .ctl_name       = FS_AIO_NR,
1034                 .procname       = "aio-nr",
1035                 .data           = &aio_nr,
1036                 .maxlen         = sizeof(aio_nr),
1037                 .mode           = 0444,
1038                 .proc_handler   = &proc_doulongvec_minmax,
1039         },
1040         {
1041                 .ctl_name       = FS_AIO_MAX_NR,
1042                 .procname       = "aio-max-nr",
1043                 .data           = &aio_max_nr,
1044                 .maxlen         = sizeof(aio_max_nr),
1045                 .mode           = 0644,
1046                 .proc_handler   = &proc_doulongvec_minmax,
1047         },
1048 #ifdef CONFIG_INOTIFY
1049         {
1050                 .ctl_name       = FS_INOTIFY,
1051                 .procname       = "inotify",
1052                 .mode           = 0555,
1053                 .child          = inotify_table,
1054         },
1055 #endif
1056 #endif
1057         {
1058                 .ctl_name       = KERN_SETUID_DUMPABLE,
1059                 .procname       = "suid_dumpable",
1060                 .data           = &suid_dumpable,
1061                 .maxlen         = sizeof(int),
1062                 .mode           = 0644,
1063                 .proc_handler   = &proc_dointvec,
1064         },
1065         { .ctl_name = 0 }
1066 };
1067
1068 static ctl_table debug_table[] = {
1069         { .ctl_name = 0 }
1070 };
1071
1072 static ctl_table dev_table[] = {
1073         { .ctl_name = 0 }
1074 };
1075
1076 extern void init_irq_proc (void);
1077
1078 static DEFINE_SPINLOCK(sysctl_lock);
1079
1080 /* called under sysctl_lock */
1081 static int use_table(struct ctl_table_header *p)
1082 {
1083         if (unlikely(p->unregistering))
1084                 return 0;
1085         p->used++;
1086         return 1;
1087 }
1088
1089 /* called under sysctl_lock */
1090 static void unuse_table(struct ctl_table_header *p)
1091 {
1092         if (!--p->used)
1093                 if (unlikely(p->unregistering))
1094                         complete(p->unregistering);
1095 }
1096
1097 /* called under sysctl_lock, will reacquire if has to wait */
1098 static void start_unregistering(struct ctl_table_header *p)
1099 {
1100         /*
1101          * if p->used is 0, nobody will ever touch that entry again;
1102          * we'll eliminate all paths to it before dropping sysctl_lock
1103          */
1104         if (unlikely(p->used)) {
1105                 struct completion wait;
1106                 init_completion(&wait);
1107                 p->unregistering = &wait;
1108                 spin_unlock(&sysctl_lock);
1109                 wait_for_completion(&wait);
1110                 spin_lock(&sysctl_lock);
1111         }
1112         /*
1113          * do not remove from the list until nobody holds it; walking the
1114          * list in do_sysctl() relies on that.
1115          */
1116         list_del_init(&p->ctl_entry);
1117 }
1118
1119 void __init sysctl_init(void)
1120 {
1121 #ifdef CONFIG_PROC_FS
1122         register_proc_table(root_table, proc_sys_root, &root_table_header);
1123         init_irq_proc();
1124 #endif
1125 }
1126
1127 int do_sysctl(int __user *name, int nlen, void __user *oldval, size_t __user *oldlenp,
1128                void __user *newval, size_t newlen)
1129 {
1130         struct list_head *tmp;
1131         int error = -ENOTDIR;
1132
1133         if (nlen <= 0 || nlen >= CTL_MAXNAME)
1134                 return -ENOTDIR;
1135         if (oldval) {
1136                 int old_len;
1137                 if (!oldlenp || get_user(old_len, oldlenp))
1138                         return -EFAULT;
1139         }
1140         spin_lock(&sysctl_lock);
1141         tmp = &root_table_header.ctl_entry;
1142         do {
1143                 struct ctl_table_header *head =
1144                         list_entry(tmp, struct ctl_table_header, ctl_entry);
1145                 void *context = NULL;
1146
1147                 if (!use_table(head))
1148                         continue;
1149
1150                 spin_unlock(&sysctl_lock);
1151
1152                 error = parse_table(name, nlen, oldval, oldlenp,
1153                                         newval, newlen, head->ctl_table,
1154                                         &context);
1155                 kfree(context);
1156
1157                 spin_lock(&sysctl_lock);
1158                 unuse_table(head);
1159                 if (error != -ENOTDIR)
1160                         break;
1161         } while ((tmp = tmp->next) != &root_table_header.ctl_entry);
1162         spin_unlock(&sysctl_lock);
1163         return error;
1164 }
1165
1166 asmlinkage long sys_sysctl(struct __sysctl_args __user *args)
1167 {
1168         struct __sysctl_args tmp;
1169         int error;
1170
1171         if (copy_from_user(&tmp, args, sizeof(tmp)))
1172                 return -EFAULT;
1173
1174         lock_kernel();
1175         error = do_sysctl(tmp.name, tmp.nlen, tmp.oldval, tmp.oldlenp,
1176                           tmp.newval, tmp.newlen);
1177         unlock_kernel();
1178         return error;
1179 }
1180
1181 /*
1182  * ctl_perm does NOT grant the superuser all rights automatically, because
1183  * some sysctl variables are readonly even to root.
1184  */
1185
1186 static int test_perm(int mode, int op)
1187 {
1188         if (!current->euid)
1189                 mode >>= 6;
1190         else if (in_egroup_p(0))
1191                 mode >>= 3;
1192         if ((mode & op & 0007) == op)
1193                 return 0;
1194         return -EACCES;
1195 }
1196
1197 static inline int ctl_perm(ctl_table *table, int op)
1198 {
1199         int error;
1200         error = security_sysctl(table, op);
1201         if (error)
1202                 return error;
1203         return test_perm(table->mode, op);
1204 }
1205
1206 static int parse_table(int __user *name, int nlen,
1207                        void __user *oldval, size_t __user *oldlenp,
1208                        void __user *newval, size_t newlen,
1209                        ctl_table *table, void **context)
1210 {
1211         int n;
1212 repeat:
1213         if (!nlen)
1214                 return -ENOTDIR;
1215         if (get_user(n, name))
1216                 return -EFAULT;
1217         for ( ; table->ctl_name; table++) {
1218                 if (n == table->ctl_name || table->ctl_name == CTL_ANY) {
1219                         int error;
1220                         if (table->child) {
1221                                 if (ctl_perm(table, 001))
1222                                         return -EPERM;
1223                                 if (table->strategy) {
1224                                         error = table->strategy(
1225                                                 table, name, nlen,
1226                                                 oldval, oldlenp,
1227                                                 newval, newlen, context);
1228                                         if (error)
1229                                                 return error;
1230                                 }
1231                                 name++;
1232                                 nlen--;
1233                                 table = table->child;
1234                                 goto repeat;
1235                         }
1236                         error = do_sysctl_strategy(table, name, nlen,
1237                                                    oldval, oldlenp,
1238                                                    newval, newlen, context);
1239                         return error;
1240                 }
1241         }
1242         return -ENOTDIR;
1243 }
1244
1245 /* Perform the actual read/write of a sysctl table entry. */
1246 int do_sysctl_strategy (ctl_table *table,
1247                         int __user *name, int nlen,
1248                         void __user *oldval, size_t __user *oldlenp,
1249                         void __user *newval, size_t newlen, void **context)
1250 {
1251         int op = 0, rc;
1252         size_t len;
1253
1254         if (oldval)
1255                 op |= 004;
1256         if (newval)
1257                 op |= 002;
1258         if (ctl_perm(table, op))
1259                 return -EPERM;
1260
1261         if (table->strategy) {
1262                 rc = table->strategy(table, name, nlen, oldval, oldlenp,
1263                                      newval, newlen, context);
1264                 if (rc < 0)
1265                         return rc;
1266                 if (rc > 0)
1267                         return 0;
1268         }
1269
1270         /* If there is no strategy routine, or if the strategy returns
1271          * zero, proceed with automatic r/w */
1272         if (table->data && table->maxlen) {
1273                 if (oldval && oldlenp) {
1274                         if (get_user(len, oldlenp))
1275                                 return -EFAULT;
1276                         if (len) {
1277                                 if (len > table->maxlen)
1278                                         len = table->maxlen;
1279                                 if(copy_to_user(oldval, table->data, len))
1280                                         return -EFAULT;
1281                                 if(put_user(len, oldlenp))
1282                                         return -EFAULT;
1283                         }
1284                 }
1285                 if (newval && newlen) {
1286                         len = newlen;
1287                         if (len > table->maxlen)
1288                                 len = table->maxlen;
1289                         if(copy_from_user(table->data, newval, len))
1290                                 return -EFAULT;
1291                 }
1292         }
1293         return 0;
1294 }
1295
1296 /**
1297  * register_sysctl_table - register a sysctl hierarchy
1298  * @table: the top-level table structure
1299  * @insert_at_head: whether the entry should be inserted in front or at the end
1300  *
1301  * Register a sysctl table hierarchy. @table should be a filled in ctl_table
1302  * array. An entry with a ctl_name of 0 terminates the table.
1303  *
1304  * The members of the &ctl_table structure are used as follows:
1305  *
1306  * ctl_name - This is the numeric sysctl value used by sysctl(2). The number
1307  *            must be unique within that level of sysctl
1308  *
1309  * procname - the name of the sysctl file under /proc/sys. Set to %NULL to not
1310  *            enter a sysctl file
1311  *
1312  * data - a pointer to data for use by proc_handler
1313  *
1314  * maxlen - the maximum size in bytes of the data
1315  *
1316  * mode - the file permissions for the /proc/sys file, and for sysctl(2)
1317  *
1318  * child - a pointer to the child sysctl table if this entry is a directory, or
1319  *         %NULL.
1320  *
1321  * proc_handler - the text handler routine (described below)
1322  *
1323  * strategy - the strategy routine (described below)
1324  *
1325  * de - for internal use by the sysctl routines
1326  *
1327  * extra1, extra2 - extra pointers usable by the proc handler routines
1328  *
1329  * Leaf nodes in the sysctl tree will be represented by a single file
1330  * under /proc; non-leaf nodes will be represented by directories.
1331  *
1332  * sysctl(2) can automatically manage read and write requests through
1333  * the sysctl table.  The data and maxlen fields of the ctl_table
1334  * struct enable minimal validation of the values being written to be
1335  * performed, and the mode field allows minimal authentication.
1336  *
1337  * More sophisticated management can be enabled by the provision of a
1338  * strategy routine with the table entry.  This will be called before
1339  * any automatic read or write of the data is performed.
1340  *
1341  * The strategy routine may return
1342  *
1343  * < 0 - Error occurred (error is passed to user process)
1344  *
1345  * 0   - OK - proceed with automatic read or write.
1346  *
1347  * > 0 - OK - read or write has been done by the strategy routine, so
1348  *       return immediately.
1349  *
1350  * There must be a proc_handler routine for any terminal nodes
1351  * mirrored under /proc/sys (non-terminals are handled by a built-in
1352  * directory handler).  Several default handlers are available to
1353  * cover common cases -
1354  *
1355  * proc_dostring(), proc_dointvec(), proc_dointvec_jiffies(),
1356  * proc_dointvec_userhz_jiffies(), proc_dointvec_minmax(),
1357  * proc_doulongvec_ms_jiffies_minmax(), proc_doulongvec_minmax()
1358  *
1359  * It is the handler's job to read the input buffer from user memory
1360  * and process it. The handler should return 0 on success.
1361  *
1362  * This routine returns %NULL on a failure to register, and a pointer
1363  * to the table header on success.
1364  */
1365 struct ctl_table_header *register_sysctl_table(ctl_table * table,
1366                                                int insert_at_head)
1367 {
1368         struct ctl_table_header *tmp;
1369         tmp = kmalloc(sizeof(struct ctl_table_header), GFP_KERNEL);
1370         if (!tmp)
1371                 return NULL;
1372         tmp->ctl_table = table;
1373         INIT_LIST_HEAD(&tmp->ctl_entry);
1374         tmp->used = 0;
1375         tmp->unregistering = NULL;
1376         spin_lock(&sysctl_lock);
1377         if (insert_at_head)
1378                 list_add(&tmp->ctl_entry, &root_table_header.ctl_entry);
1379         else
1380                 list_add_tail(&tmp->ctl_entry, &root_table_header.ctl_entry);
1381         spin_unlock(&sysctl_lock);
1382 #ifdef CONFIG_PROC_FS
1383         register_proc_table(table, proc_sys_root, tmp);
1384 #endif
1385         return tmp;
1386 }
1387
1388 /**
1389  * unregister_sysctl_table - unregister a sysctl table hierarchy
1390  * @header: the header returned from register_sysctl_table
1391  *
1392  * Unregisters the sysctl table and all children. proc entries may not
1393  * actually be removed until they are no longer used by anyone.
1394  */
1395 void unregister_sysctl_table(struct ctl_table_header * header)
1396 {
1397         might_sleep();
1398         spin_lock(&sysctl_lock);
1399         start_unregistering(header);
1400 #ifdef CONFIG_PROC_FS
1401         unregister_proc_table(header->ctl_table, proc_sys_root);
1402 #endif
1403         spin_unlock(&sysctl_lock);
1404         kfree(header);
1405 }
1406
1407 /*
1408  * /proc/sys support
1409  */
1410
1411 #ifdef CONFIG_PROC_FS
1412
1413 /* Scan the sysctl entries in table and add them all into /proc */
1414 static void register_proc_table(ctl_table * table, struct proc_dir_entry *root, void *set)
1415 {
1416         struct proc_dir_entry *de;
1417         int len;
1418         mode_t mode;
1419
1420         for (; table->ctl_name; table++) {
1421                 /* Can't do anything without a proc name. */
1422                 if (!table->procname)
1423                         continue;
1424                 /* Maybe we can't do anything with it... */
1425                 if (!table->proc_handler && !table->child) {
1426                         printk(KERN_WARNING "SYSCTL: Can't register %s\n",
1427                                 table->procname);
1428                         continue;
1429                 }
1430
1431                 len = strlen(table->procname);
1432                 mode = table->mode;
1433
1434                 de = NULL;
1435                 if (table->proc_handler)
1436                         mode |= S_IFREG;
1437                 else {
1438                         mode |= S_IFDIR;
1439                         for (de = root->subdir; de; de = de->next) {
1440                                 if (proc_match(len, table->procname, de))
1441                                         break;
1442                         }
1443                         /* If the subdir exists already, de is non-NULL */
1444                 }
1445
1446                 if (!de) {
1447                         de = create_proc_entry(table->procname, mode, root);
1448                         if (!de)
1449                                 continue;
1450                         de->set = set;
1451                         de->data = (void *) table;
1452                         if (table->proc_handler)
1453                                 de->proc_fops = &proc_sys_file_operations;
1454                 }
1455                 table->de = de;
1456                 if (de->mode & S_IFDIR)
1457                         register_proc_table(table->child, de, set);
1458         }
1459 }
1460
1461 /*
1462  * Unregister a /proc sysctl table and any subdirectories.
1463  */
1464 static void unregister_proc_table(ctl_table * table, struct proc_dir_entry *root)
1465 {
1466         struct proc_dir_entry *de;
1467         for (; table->ctl_name; table++) {
1468                 if (!(de = table->de))
1469                         continue;
1470                 if (de->mode & S_IFDIR) {
1471                         if (!table->child) {
1472                                 printk (KERN_ALERT "Help - malformed sysctl tree on free\n");
1473                                 continue;
1474                         }
1475                         unregister_proc_table(table->child, de);
1476
1477                         /* Don't unregister directories which still have entries.. */
1478                         if (de->subdir)
1479                                 continue;
1480                 }
1481
1482                 /*
1483                  * In any case, mark the entry as goner; we'll keep it
1484                  * around if it's busy, but we'll know to do nothing with
1485                  * its fields.  We are under sysctl_lock here.
1486                  */
1487                 de->data = NULL;
1488
1489                 /* Don't unregister proc entries that are still being used.. */
1490                 if (atomic_read(&de->count))
1491                         continue;
1492
1493                 table->de = NULL;
1494                 remove_proc_entry(table->procname, root);
1495         }
1496 }
1497
1498 static ssize_t do_rw_proc(int write, struct file * file, char __user * buf,
1499                           size_t count, loff_t *ppos)
1500 {
1501         int op;
1502         struct proc_dir_entry *de = PDE(file->f_dentry->d_inode);
1503         struct ctl_table *table;
1504         size_t res;
1505         ssize_t error = -ENOTDIR;
1506
1507         spin_lock(&sysctl_lock);
1508         if (de && de->data && use_table(de->set)) {
1509                 /*
1510                  * at that point we know that sysctl was not unregistered
1511                  * and won't be until we finish
1512                  */
1513                 spin_unlock(&sysctl_lock);
1514                 table = (struct ctl_table *) de->data;
1515                 if (!table || !table->proc_handler)
1516                         goto out;
1517                 error = -EPERM;
1518                 op = (write ? 002 : 004);
1519                 if (ctl_perm(table, op))
1520                         goto out;
1521
1522                 /* careful: calling conventions are nasty here */
1523                 res = count;
1524                 error = (*table->proc_handler)(table, write, file,
1525                                                 buf, &res, ppos);
1526                 if (!error)
1527                         error = res;
1528         out:
1529                 spin_lock(&sysctl_lock);
1530                 unuse_table(de->set);
1531         }
1532         spin_unlock(&sysctl_lock);
1533         return error;
1534 }
1535
1536 static int proc_opensys(struct inode *inode, struct file *file)
1537 {
1538         if (file->f_mode & FMODE_WRITE) {
1539                 /*
1540                  * sysctl entries that are not writable,
1541                  * are _NOT_ writable, capabilities or not.
1542                  */
1543                 if (!(inode->i_mode & S_IWUSR))
1544                         return -EPERM;
1545         }
1546
1547         return 0;
1548 }
1549
1550 static ssize_t proc_readsys(struct file * file, char __user * buf,
1551                             size_t count, loff_t *ppos)
1552 {
1553         return do_rw_proc(0, file, buf, count, ppos);
1554 }
1555
1556 static ssize_t proc_writesys(struct file * file, const char __user * buf,
1557                              size_t count, loff_t *ppos)
1558 {
1559         return do_rw_proc(1, file, (char __user *) buf, count, ppos);
1560 }
1561
1562 /**
1563  * proc_dostring - read a string sysctl
1564  * @table: the sysctl table
1565  * @write: %TRUE if this is a write to the sysctl file
1566  * @filp: the file structure
1567  * @buffer: the user buffer
1568  * @lenp: the size of the user buffer
1569  * @ppos: file position
1570  *
1571  * Reads/writes a string from/to the user buffer. If the kernel
1572  * buffer provided is not large enough to hold the string, the
1573  * string is truncated. The copied string is %NULL-terminated.
1574  * If the string is being read by the user process, it is copied
1575  * and a newline '\n' is added. It is truncated if the buffer is
1576  * not large enough.
1577  *
1578  * Returns 0 on success.
1579  */
1580 int proc_dostring(ctl_table *table, int write, struct file *filp,
1581                   void __user *buffer, size_t *lenp, loff_t *ppos)
1582 {
1583         size_t len, maxlen;
1584         char __user *p;
1585         char c;
1586         void *data;
1587
1588         data = table->data;
1589         maxlen = table->maxlen;
1590
1591         if (!data || !maxlen || !*lenp || (*ppos && !write))
1592                 return (*lenp = 0);
1593
1594         if (table->virt_handler)
1595                 table->virt_handler(table, write, filp->f_xid, &data, &maxlen);
1596
1597         if (write) {
1598                 len = 0;
1599                 p = buffer;
1600                 while (len < *lenp) {
1601                         if (get_user(c, p++))
1602                                 return -EFAULT;
1603                         if (c == 0 || c == '\n')
1604                                 break;
1605                         len++;
1606                 }
1607                 if (len >= maxlen)
1608                         len = maxlen-1;
1609                 if(copy_from_user(data, buffer, len))
1610                         return -EFAULT;
1611                 ((char *) data)[len] = 0;
1612                 *ppos += *lenp;
1613         } else {
1614                 len = strlen(data);
1615                 if (len > maxlen)
1616                         len = maxlen;
1617                 if (len > *lenp)
1618                         len = *lenp;
1619                 if (len)
1620                         if(copy_to_user(buffer, data, len))
1621                                 return -EFAULT;
1622                 if (len < *lenp) {
1623                         if(put_user('\n', ((char __user *) buffer) + len))
1624                                 return -EFAULT;
1625                         len++;
1626                 }
1627                 *lenp = len;
1628                 *ppos += len;
1629         }
1630         return 0;
1631 }
1632
1633 /*
1634  *      Special case of dostring for the UTS structure. This has locks
1635  *      to observe. Should this be in kernel/sys.c ????
1636  */
1637
1638 static int proc_doutsstring(ctl_table *table, int write, struct file *filp,
1639                   void __user *buffer, size_t *lenp, loff_t *ppos)
1640 {
1641         int r;
1642
1643         if (!write) {
1644                 down_read(&uts_sem);
1645                 r=proc_dostring(table,0,filp,buffer,lenp, ppos);
1646                 up_read(&uts_sem);
1647         } else {
1648                 down_write(&uts_sem);
1649                 r=proc_dostring(table,1,filp,buffer,lenp, ppos);
1650                 up_write(&uts_sem);
1651         }
1652         return r;
1653 }
1654
1655 static int do_proc_dointvec_conv(int *negp, unsigned long *lvalp,
1656                                  int *valp,
1657                                  int write, void *data)
1658 {
1659         if (write) {
1660                 *valp = *negp ? -*lvalp : *lvalp;
1661         } else {
1662                 int val = *valp;
1663                 if (val < 0) {
1664                         *negp = -1;
1665                         *lvalp = (unsigned long)-val;
1666                 } else {
1667                         *negp = 0;
1668                         *lvalp = (unsigned long)val;
1669                 }
1670         }
1671         return 0;
1672 }
1673
1674 static int do_proc_dointvec(ctl_table *table, int write, struct file *filp,
1675                   void __user *buffer, size_t *lenp, loff_t *ppos,
1676                   int (*conv)(int *negp, unsigned long *lvalp, int *valp,
1677                               int write, void *data),
1678                   void *data)
1679 {
1680 #define TMPBUFLEN 21
1681         int *i, vleft, first=1, neg, val;
1682         unsigned long lval;
1683         size_t left, len;
1684
1685         char buf[TMPBUFLEN], *p;
1686         char __user *s = buffer;
1687
1688         if (!table->data || !table->maxlen || !*lenp ||
1689             (*ppos && !write)) {
1690                 *lenp = 0;
1691                 return 0;
1692         }
1693
1694         i = (int *) table->data;
1695         vleft = table->maxlen / sizeof(*i);
1696         left = *lenp;
1697
1698         if (!conv)
1699                 conv = do_proc_dointvec_conv;
1700
1701         for (; left && vleft--; i++, first=0) {
1702                 if (write) {
1703                         while (left) {
1704                                 char c;
1705                                 if (get_user(c, s))
1706                                         return -EFAULT;
1707                                 if (!isspace(c))
1708                                         break;
1709                                 left--;
1710                                 s++;
1711                         }
1712                         if (!left)
1713                                 break;
1714                         neg = 0;
1715                         len = left;
1716                         if (len > sizeof(buf) - 1)
1717                                 len = sizeof(buf) - 1;
1718                         if (copy_from_user(buf, s, len))
1719                                 return -EFAULT;
1720                         buf[len] = 0;
1721                         p = buf;
1722                         if (*p == '-' && left > 1) {
1723                                 neg = 1;
1724                                 left--, p++;
1725                         }
1726                         if (*p < '0' || *p > '9')
1727                                 break;
1728
1729                         lval = simple_strtoul(p, &p, 0);
1730
1731                         len = p-buf;
1732                         if ((len < left) && *p && !isspace(*p))
1733                                 break;
1734                         if (neg)
1735                                 val = -val;
1736                         s += len;
1737                         left -= len;
1738
1739                         if (conv(&neg, &lval, i, 1, data))
1740                                 break;
1741                 } else {
1742                         p = buf;
1743                         if (!first)
1744                                 *p++ = '\t';
1745
1746                         if (conv(&neg, &lval, i, 0, data))
1747                                 break;
1748
1749                         sprintf(p, "%s%lu", neg ? "-" : "", lval);
1750                         len = strlen(buf);
1751                         if (len > left)
1752                                 len = left;
1753                         if(copy_to_user(s, buf, len))
1754                                 return -EFAULT;
1755                         left -= len;
1756                         s += len;
1757                 }
1758         }
1759
1760         if (!write && !first && left) {
1761                 if(put_user('\n', s))
1762                         return -EFAULT;
1763                 left--, s++;
1764         }
1765         if (write) {
1766                 while (left) {
1767                         char c;
1768                         if (get_user(c, s++))
1769                                 return -EFAULT;
1770                         if (!isspace(c))
1771                                 break;
1772                         left--;
1773                 }
1774         }
1775         if (write && first)
1776                 return -EINVAL;
1777         *lenp -= left;
1778         *ppos += *lenp;
1779         return 0;
1780 #undef TMPBUFLEN
1781 }
1782
1783 /**
1784  * proc_dointvec - read a vector of integers
1785  * @table: the sysctl table
1786  * @write: %TRUE if this is a write to the sysctl file
1787  * @filp: the file structure
1788  * @buffer: the user buffer
1789  * @lenp: the size of the user buffer
1790  * @ppos: file position
1791  *
1792  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
1793  * values from/to the user buffer, treated as an ASCII string.
1794  *
1795  * Returns 0 on success.
1796  */
1797 int proc_dointvec(ctl_table *table, int write, struct file *filp,
1798                      void __user *buffer, size_t *lenp, loff_t *ppos)
1799 {
1800     return do_proc_dointvec(table,write,filp,buffer,lenp,ppos,
1801                             NULL,NULL);
1802 }
1803
1804 #define OP_SET  0
1805 #define OP_AND  1
1806 #define OP_OR   2
1807 #define OP_MAX  3
1808 #define OP_MIN  4
1809
1810 static int do_proc_dointvec_bset_conv(int *negp, unsigned long *lvalp,
1811                                       int *valp,
1812                                       int write, void *data)
1813 {
1814         int op = *(int *)data;
1815         if (write) {
1816                 int val = *negp ? -*lvalp : *lvalp;
1817                 switch(op) {
1818                 case OP_SET:    *valp = val; break;
1819                 case OP_AND:    *valp &= val; break;
1820                 case OP_OR:     *valp |= val; break;
1821                 case OP_MAX:    if(*valp < val)
1822                                         *valp = val;
1823                                 break;
1824                 case OP_MIN:    if(*valp > val)
1825                                 *valp = val;
1826                                 break;
1827                 }
1828         } else {
1829                 int val = *valp;
1830                 if (val < 0) {
1831                         *negp = -1;
1832                         *lvalp = (unsigned long)-val;
1833                 } else {
1834                         *negp = 0;
1835                         *lvalp = (unsigned long)val;
1836                 }
1837         }
1838         return 0;
1839 }
1840
1841 /*
1842  *      init may raise the set.
1843  */
1844
1845 int proc_dointvec_bset(ctl_table *table, int write, struct file *filp,
1846                         void __user *buffer, size_t *lenp, loff_t *ppos)
1847 {
1848         int op;
1849
1850         if (!capable(CAP_SYS_MODULE)) {
1851                 return -EPERM;
1852         }
1853
1854         op = (current->pid == 1) ? OP_SET : OP_AND;
1855         return do_proc_dointvec(table,write,filp,buffer,lenp,ppos,
1856                                 do_proc_dointvec_bset_conv,&op);
1857 }
1858
1859 struct do_proc_dointvec_minmax_conv_param {
1860         int *min;
1861         int *max;
1862 };
1863
1864 static int do_proc_dointvec_minmax_conv(int *negp, unsigned long *lvalp,
1865                                         int *valp,
1866                                         int write, void *data)
1867 {
1868         struct do_proc_dointvec_minmax_conv_param *param = data;
1869         if (write) {
1870                 int val = *negp ? -*lvalp : *lvalp;
1871                 if ((param->min && *param->min > val) ||
1872                     (param->max && *param->max < val))
1873                         return -EINVAL;
1874                 *valp = val;
1875         } else {
1876                 int val = *valp;
1877                 if (val < 0) {
1878                         *negp = -1;
1879                         *lvalp = (unsigned long)-val;
1880                 } else {
1881                         *negp = 0;
1882                         *lvalp = (unsigned long)val;
1883                 }
1884         }
1885         return 0;
1886 }
1887
1888 /**
1889  * proc_dointvec_minmax - read a vector of integers with min/max values
1890  * @table: the sysctl table
1891  * @write: %TRUE if this is a write to the sysctl file
1892  * @filp: the file structure
1893  * @buffer: the user buffer
1894  * @lenp: the size of the user buffer
1895  * @ppos: file position
1896  *
1897  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
1898  * values from/to the user buffer, treated as an ASCII string.
1899  *
1900  * This routine will ensure the values are within the range specified by
1901  * table->extra1 (min) and table->extra2 (max).
1902  *
1903  * Returns 0 on success.
1904  */
1905 int proc_dointvec_minmax(ctl_table *table, int write, struct file *filp,
1906                   void __user *buffer, size_t *lenp, loff_t *ppos)
1907 {
1908         struct do_proc_dointvec_minmax_conv_param param = {
1909                 .min = (int *) table->extra1,
1910                 .max = (int *) table->extra2,
1911         };
1912         return do_proc_dointvec(table, write, filp, buffer, lenp, ppos,
1913                                 do_proc_dointvec_minmax_conv, &param);
1914 }
1915
1916 static int do_proc_doulongvec_minmax(ctl_table *table, int write,
1917                                      struct file *filp,
1918                                      void __user *buffer,
1919                                      size_t *lenp, loff_t *ppos,
1920                                      unsigned long convmul,
1921                                      unsigned long convdiv)
1922 {
1923 #define TMPBUFLEN 21
1924         unsigned long *i, *min, *max, val;
1925         int vleft, first=1, neg;
1926         size_t len, left;
1927         char buf[TMPBUFLEN], *p;
1928         char __user *s = buffer;
1929
1930         if (!table->data || !table->maxlen || !*lenp ||
1931             (*ppos && !write)) {
1932                 *lenp = 0;
1933                 return 0;
1934         }
1935
1936         i = (unsigned long *) table->data;
1937         min = (unsigned long *) table->extra1;
1938         max = (unsigned long *) table->extra2;
1939         vleft = table->maxlen / sizeof(unsigned long);
1940         left = *lenp;
1941
1942         for (; left && vleft--; i++, min++, max++, first=0) {
1943                 if (write) {
1944                         while (left) {
1945                                 char c;
1946                                 if (get_user(c, s))
1947                                         return -EFAULT;
1948                                 if (!isspace(c))
1949                                         break;
1950                                 left--;
1951                                 s++;
1952                         }
1953                         if (!left)
1954                                 break;
1955                         neg = 0;
1956                         len = left;
1957                         if (len > TMPBUFLEN-1)
1958                                 len = TMPBUFLEN-1;
1959                         if (copy_from_user(buf, s, len))
1960                                 return -EFAULT;
1961                         buf[len] = 0;
1962                         p = buf;
1963                         if (*p == '-' && left > 1) {
1964                                 neg = 1;
1965                                 left--, p++;
1966                         }
1967                         if (*p < '0' || *p > '9')
1968                                 break;
1969                         val = simple_strtoul(p, &p, 0) * convmul / convdiv ;
1970                         len = p-buf;
1971                         if ((len < left) && *p && !isspace(*p))
1972                                 break;
1973                         if (neg)
1974                                 val = -val;
1975                         s += len;
1976                         left -= len;
1977
1978                         if(neg)
1979                                 continue;
1980                         if ((min && val < *min) || (max && val > *max))
1981                                 continue;
1982                         *i = val;
1983                 } else {
1984                         p = buf;
1985                         if (!first)
1986                                 *p++ = '\t';
1987                         sprintf(p, "%lu", convdiv * (*i) / convmul);
1988                         len = strlen(buf);
1989                         if (len > left)
1990                                 len = left;
1991                         if(copy_to_user(s, buf, len))
1992                                 return -EFAULT;
1993                         left -= len;
1994                         s += len;
1995                 }
1996         }
1997
1998         if (!write && !first && left) {
1999                 if(put_user('\n', s))
2000                         return -EFAULT;
2001                 left--, s++;
2002         }
2003         if (write) {
2004                 while (left) {
2005                         char c;
2006                         if (get_user(c, s++))
2007                                 return -EFAULT;
2008                         if (!isspace(c))
2009                                 break;
2010                         left--;
2011                 }
2012         }
2013         if (write && first)
2014                 return -EINVAL;
2015         *lenp -= left;
2016         *ppos += *lenp;
2017         return 0;
2018 #undef TMPBUFLEN
2019 }
2020
2021 /**
2022  * proc_doulongvec_minmax - read a vector of long integers with min/max values
2023  * @table: the sysctl table
2024  * @write: %TRUE if this is a write to the sysctl file
2025  * @filp: the file structure
2026  * @buffer: the user buffer
2027  * @lenp: the size of the user buffer
2028  * @ppos: file position
2029  *
2030  * Reads/writes up to table->maxlen/sizeof(unsigned long) unsigned long
2031  * values from/to the user buffer, treated as an ASCII string.
2032  *
2033  * This routine will ensure the values are within the range specified by
2034  * table->extra1 (min) and table->extra2 (max).
2035  *
2036  * Returns 0 on success.
2037  */
2038 int proc_doulongvec_minmax(ctl_table *table, int write, struct file *filp,
2039                            void __user *buffer, size_t *lenp, loff_t *ppos)
2040 {
2041     return do_proc_doulongvec_minmax(table, write, filp, buffer, lenp, ppos, 1l, 1l);
2042 }
2043
2044 /**
2045  * proc_doulongvec_ms_jiffies_minmax - read a vector of millisecond values with min/max values
2046  * @table: the sysctl table
2047  * @write: %TRUE if this is a write to the sysctl file
2048  * @filp: the file structure
2049  * @buffer: the user buffer
2050  * @lenp: the size of the user buffer
2051  * @ppos: file position
2052  *
2053  * Reads/writes up to table->maxlen/sizeof(unsigned long) unsigned long
2054  * values from/to the user buffer, treated as an ASCII string. The values
2055  * are treated as milliseconds, and converted to jiffies when they are stored.
2056  *
2057  * This routine will ensure the values are within the range specified by
2058  * table->extra1 (min) and table->extra2 (max).
2059  *
2060  * Returns 0 on success.
2061  */
2062 int proc_doulongvec_ms_jiffies_minmax(ctl_table *table, int write,
2063                                       struct file *filp,
2064                                       void __user *buffer,
2065                                       size_t *lenp, loff_t *ppos)
2066 {
2067     return do_proc_doulongvec_minmax(table, write, filp, buffer,
2068                                      lenp, ppos, HZ, 1000l);
2069 }
2070
2071
2072 static int do_proc_dointvec_jiffies_conv(int *negp, unsigned long *lvalp,
2073                                          int *valp,
2074                                          int write, void *data)
2075 {
2076         if (write) {
2077                 *valp = *negp ? -(*lvalp*HZ) : (*lvalp*HZ);
2078         } else {
2079                 int val = *valp;
2080                 unsigned long lval;
2081                 if (val < 0) {
2082                         *negp = -1;
2083                         lval = (unsigned long)-val;
2084                 } else {
2085                         *negp = 0;
2086                         lval = (unsigned long)val;
2087                 }
2088                 *lvalp = lval / HZ;
2089         }
2090         return 0;
2091 }
2092
2093 static int do_proc_dointvec_userhz_jiffies_conv(int *negp, unsigned long *lvalp,
2094                                                 int *valp,
2095                                                 int write, void *data)
2096 {
2097         if (write) {
2098                 *valp = clock_t_to_jiffies(*negp ? -*lvalp : *lvalp);
2099         } else {
2100                 int val = *valp;
2101                 unsigned long lval;
2102                 if (val < 0) {
2103                         *negp = -1;
2104                         lval = (unsigned long)-val;
2105                 } else {
2106                         *negp = 0;
2107                         lval = (unsigned long)val;
2108                 }
2109                 *lvalp = jiffies_to_clock_t(lval);
2110         }
2111         return 0;
2112 }
2113
2114 static int do_proc_dointvec_ms_jiffies_conv(int *negp, unsigned long *lvalp,
2115                                             int *valp,
2116                                             int write, void *data)
2117 {
2118         if (write) {
2119                 *valp = msecs_to_jiffies(*negp ? -*lvalp : *lvalp);
2120         } else {
2121                 int val = *valp;
2122                 unsigned long lval;
2123                 if (val < 0) {
2124                         *negp = -1;
2125                         lval = (unsigned long)-val;
2126                 } else {
2127                         *negp = 0;
2128                         lval = (unsigned long)val;
2129                 }
2130                 *lvalp = jiffies_to_msecs(lval);
2131         }
2132         return 0;
2133 }
2134
2135 /**
2136  * proc_dointvec_jiffies - read a vector of integers as seconds
2137  * @table: the sysctl table
2138  * @write: %TRUE if this is a write to the sysctl file
2139  * @filp: the file structure
2140  * @buffer: the user buffer
2141  * @lenp: the size of the user buffer
2142  * @ppos: file position
2143  *
2144  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2145  * values from/to the user buffer, treated as an ASCII string.
2146  * The values read are assumed to be in seconds, and are converted into
2147  * jiffies.
2148  *
2149  * Returns 0 on success.
2150  */
2151 int proc_dointvec_jiffies(ctl_table *table, int write, struct file *filp,
2152                           void __user *buffer, size_t *lenp, loff_t *ppos)
2153 {
2154     return do_proc_dointvec(table,write,filp,buffer,lenp,ppos,
2155                             do_proc_dointvec_jiffies_conv,NULL);
2156 }
2157
2158 /**
2159  * proc_dointvec_userhz_jiffies - read a vector of integers as 1/USER_HZ seconds
2160  * @table: the sysctl table
2161  * @write: %TRUE if this is a write to the sysctl file
2162  * @filp: the file structure
2163  * @buffer: the user buffer
2164  * @lenp: the size of the user buffer
2165  * @ppos: pointer to the file position
2166  *
2167  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2168  * values from/to the user buffer, treated as an ASCII string.
2169  * The values read are assumed to be in 1/USER_HZ seconds, and
2170  * are converted into jiffies.
2171  *
2172  * Returns 0 on success.
2173  */
2174 int proc_dointvec_userhz_jiffies(ctl_table *table, int write, struct file *filp,
2175                                  void __user *buffer, size_t *lenp, loff_t *ppos)
2176 {
2177     return do_proc_dointvec(table,write,filp,buffer,lenp,ppos,
2178                             do_proc_dointvec_userhz_jiffies_conv,NULL);
2179 }
2180
2181 /**
2182  * proc_dointvec_ms_jiffies - read a vector of integers as 1 milliseconds
2183  * @table: the sysctl table
2184  * @write: %TRUE if this is a write to the sysctl file
2185  * @filp: the file structure
2186  * @buffer: the user buffer
2187  * @lenp: the size of the user buffer
2188  * @ppos: file position
2189  * @ppos: the current position in the file
2190  *
2191  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2192  * values from/to the user buffer, treated as an ASCII string.
2193  * The values read are assumed to be in 1/1000 seconds, and
2194  * are converted into jiffies.
2195  *
2196  * Returns 0 on success.
2197  */
2198 int proc_dointvec_ms_jiffies(ctl_table *table, int write, struct file *filp,
2199                              void __user *buffer, size_t *lenp, loff_t *ppos)
2200 {
2201         return do_proc_dointvec(table, write, filp, buffer, lenp, ppos,
2202                                 do_proc_dointvec_ms_jiffies_conv, NULL);
2203 }
2204
2205 #else /* CONFIG_PROC_FS */
2206
2207 int proc_dostring(ctl_table *table, int write, struct file *filp,
2208                   void __user *buffer, size_t *lenp, loff_t *ppos)
2209 {
2210         return -ENOSYS;
2211 }
2212
2213 static int proc_doutsstring(ctl_table *table, int write, struct file *filp,
2214                             void __user *buffer, size_t *lenp, loff_t *ppos)
2215 {
2216         return -ENOSYS;
2217 }
2218
2219 int proc_dointvec(ctl_table *table, int write, struct file *filp,
2220                   void __user *buffer, size_t *lenp, loff_t *ppos)
2221 {
2222         return -ENOSYS;
2223 }
2224
2225 int proc_dointvec_bset(ctl_table *table, int write, struct file *filp,
2226                         void __user *buffer, size_t *lenp, loff_t *ppos)
2227 {
2228         return -ENOSYS;
2229 }
2230
2231 int proc_dointvec_minmax(ctl_table *table, int write, struct file *filp,
2232                     void __user *buffer, size_t *lenp, loff_t *ppos)
2233 {
2234         return -ENOSYS;
2235 }
2236
2237 int proc_dointvec_jiffies(ctl_table *table, int write, struct file *filp,
2238                     void __user *buffer, size_t *lenp, loff_t *ppos)
2239 {
2240         return -ENOSYS;
2241 }
2242
2243 int proc_dointvec_userhz_jiffies(ctl_table *table, int write, struct file *filp,
2244                     void __user *buffer, size_t *lenp, loff_t *ppos)
2245 {
2246         return -ENOSYS;
2247 }
2248
2249 int proc_dointvec_ms_jiffies(ctl_table *table, int write, struct file *filp,
2250                              void __user *buffer, size_t *lenp, loff_t *ppos)
2251 {
2252         return -ENOSYS;
2253 }
2254
2255 int proc_doulongvec_minmax(ctl_table *table, int write, struct file *filp,
2256                     void __user *buffer, size_t *lenp, loff_t *ppos)
2257 {
2258         return -ENOSYS;
2259 }
2260
2261 int proc_doulongvec_ms_jiffies_minmax(ctl_table *table, int write,
2262                                       struct file *filp,
2263                                       void __user *buffer,
2264                                       size_t *lenp, loff_t *ppos)
2265 {
2266     return -ENOSYS;
2267 }
2268
2269
2270 #endif /* CONFIG_PROC_FS */
2271
2272
2273 /*
2274  * General sysctl support routines
2275  */
2276
2277 /* The generic string strategy routine: */
2278 int sysctl_string(ctl_table *table, int __user *name, int nlen,
2279                   void __user *oldval, size_t __user *oldlenp,
2280                   void __user *newval, size_t newlen, void **context)
2281 {
2282         if (!table->data || !table->maxlen)
2283                 return -ENOTDIR;
2284
2285         if (oldval && oldlenp) {
2286                 size_t bufsize;
2287                 if (get_user(bufsize, oldlenp))
2288                         return -EFAULT;
2289                 if (bufsize) {
2290                         size_t len = strlen(table->data), copied;
2291
2292                         /* This shouldn't trigger for a well-formed sysctl */
2293                         if (len > table->maxlen)
2294                                 len = table->maxlen;
2295
2296                         /* Copy up to a max of bufsize-1 bytes of the string */
2297                         copied = (len >= bufsize) ? bufsize - 1 : len;
2298
2299                         if (copy_to_user(oldval, table->data, copied) ||
2300                             put_user(0, (char __user *)(oldval + copied)))
2301                                 return -EFAULT;
2302                         if (put_user(len, oldlenp))
2303                                 return -EFAULT;
2304                 }
2305         }
2306         if (newval && newlen) {
2307                 size_t len = newlen;
2308                 if (len > table->maxlen)
2309                         len = table->maxlen;
2310                 if(copy_from_user(table->data, newval, len))
2311                         return -EFAULT;
2312                 if (len == table->maxlen)
2313                         len--;
2314                 ((char *) table->data)[len] = 0;
2315         }
2316         return 1;
2317 }
2318
2319 /*
2320  * This function makes sure that all of the integers in the vector
2321  * are between the minimum and maximum values given in the arrays
2322  * table->extra1 and table->extra2, respectively.
2323  */
2324 int sysctl_intvec(ctl_table *table, int __user *name, int nlen,
2325                 void __user *oldval, size_t __user *oldlenp,
2326                 void __user *newval, size_t newlen, void **context)
2327 {
2328
2329         if (newval && newlen) {
2330                 int __user *vec = (int __user *) newval;
2331                 int *min = (int *) table->extra1;
2332                 int *max = (int *) table->extra2;
2333                 size_t length;
2334                 int i;
2335
2336                 if (newlen % sizeof(int) != 0)
2337                         return -EINVAL;
2338
2339                 if (!table->extra1 && !table->extra2)
2340                         return 0;
2341
2342                 if (newlen > table->maxlen)
2343                         newlen = table->maxlen;
2344                 length = newlen / sizeof(int);
2345
2346                 for (i = 0; i < length; i++) {
2347                         int value;
2348                         if (get_user(value, vec + i))
2349                                 return -EFAULT;
2350                         if (min && value < min[i])
2351                                 return -EINVAL;
2352                         if (max && value > max[i])
2353                                 return -EINVAL;
2354                 }
2355         }
2356         return 0;
2357 }
2358
2359 /* Strategy function to convert jiffies to seconds */
2360 int sysctl_jiffies(ctl_table *table, int __user *name, int nlen,
2361                 void __user *oldval, size_t __user *oldlenp,
2362                 void __user *newval, size_t newlen, void **context)
2363 {
2364         if (oldval) {
2365                 size_t olen;
2366                 if (oldlenp) {
2367                         if (get_user(olen, oldlenp))
2368                                 return -EFAULT;
2369                         if (olen!=sizeof(int))
2370                                 return -EINVAL;
2371                 }
2372                 if (put_user(*(int *)(table->data)/HZ, (int __user *)oldval) ||
2373                     (oldlenp && put_user(sizeof(int),oldlenp)))
2374                         return -EFAULT;
2375         }
2376         if (newval && newlen) {
2377                 int new;
2378                 if (newlen != sizeof(int))
2379                         return -EINVAL;
2380                 if (get_user(new, (int __user *)newval))
2381                         return -EFAULT;
2382                 *(int *)(table->data) = new*HZ;
2383         }
2384         return 1;
2385 }
2386
2387 /* Strategy function to convert jiffies to seconds */
2388 int sysctl_ms_jiffies(ctl_table *table, int __user *name, int nlen,
2389                 void __user *oldval, size_t __user *oldlenp,
2390                 void __user *newval, size_t newlen, void **context)
2391 {
2392         if (oldval) {
2393                 size_t olen;
2394                 if (oldlenp) {
2395                         if (get_user(olen, oldlenp))
2396                                 return -EFAULT;
2397                         if (olen!=sizeof(int))
2398                                 return -EINVAL;
2399                 }
2400                 if (put_user(jiffies_to_msecs(*(int *)(table->data)), (int __user *)oldval) ||
2401                     (oldlenp && put_user(sizeof(int),oldlenp)))
2402                         return -EFAULT;
2403         }
2404         if (newval && newlen) {
2405                 int new;
2406                 if (newlen != sizeof(int))
2407                         return -EINVAL;
2408                 if (get_user(new, (int __user *)newval))
2409                         return -EFAULT;
2410                 *(int *)(table->data) = msecs_to_jiffies(new);
2411         }
2412         return 1;
2413 }
2414
2415 #else /* CONFIG_SYSCTL */
2416
2417
2418 asmlinkage long sys_sysctl(struct __sysctl_args __user *args)
2419 {
2420         return -ENOSYS;
2421 }
2422
2423 int sysctl_string(ctl_table *table, int __user *name, int nlen,
2424                   void __user *oldval, size_t __user *oldlenp,
2425                   void __user *newval, size_t newlen, void **context)
2426 {
2427         return -ENOSYS;
2428 }
2429
2430 int sysctl_intvec(ctl_table *table, int __user *name, int nlen,
2431                 void __user *oldval, size_t __user *oldlenp,
2432                 void __user *newval, size_t newlen, void **context)
2433 {
2434         return -ENOSYS;
2435 }
2436
2437 int sysctl_jiffies(ctl_table *table, int __user *name, int nlen,
2438                 void __user *oldval, size_t __user *oldlenp,
2439                 void __user *newval, size_t newlen, void **context)
2440 {
2441         return -ENOSYS;
2442 }
2443
2444 int sysctl_ms_jiffies(ctl_table *table, int __user *name, int nlen,
2445                 void __user *oldval, size_t __user *oldlenp,
2446                 void __user *newval, size_t newlen, void **context)
2447 {
2448         return -ENOSYS;
2449 }
2450
2451 int proc_dostring(ctl_table *table, int write, struct file *filp,
2452                   void __user *buffer, size_t *lenp, loff_t *ppos)
2453 {
2454         return -ENOSYS;
2455 }
2456
2457 int proc_dointvec(ctl_table *table, int write, struct file *filp,
2458                   void __user *buffer, size_t *lenp, loff_t *ppos)
2459 {
2460         return -ENOSYS;
2461 }
2462
2463 int proc_dointvec_bset(ctl_table *table, int write, struct file *filp,
2464                         void __user *buffer, size_t *lenp, loff_t *ppos)
2465 {
2466         return -ENOSYS;
2467 }
2468
2469 int proc_dointvec_minmax(ctl_table *table, int write, struct file *filp,
2470                     void __user *buffer, size_t *lenp, loff_t *ppos)
2471 {
2472         return -ENOSYS;
2473 }
2474
2475 int proc_dointvec_jiffies(ctl_table *table, int write, struct file *filp,
2476                           void __user *buffer, size_t *lenp, loff_t *ppos)
2477 {
2478         return -ENOSYS;
2479 }
2480
2481 int proc_dointvec_userhz_jiffies(ctl_table *table, int write, struct file *filp,
2482                           void __user *buffer, size_t *lenp, loff_t *ppos)
2483 {
2484         return -ENOSYS;
2485 }
2486
2487 int proc_dointvec_ms_jiffies(ctl_table *table, int write, struct file *filp,
2488                              void __user *buffer, size_t *lenp, loff_t *ppos)
2489 {
2490         return -ENOSYS;
2491 }
2492
2493 int proc_doulongvec_minmax(ctl_table *table, int write, struct file *filp,
2494                     void __user *buffer, size_t *lenp, loff_t *ppos)
2495 {
2496         return -ENOSYS;
2497 }
2498
2499 int proc_doulongvec_ms_jiffies_minmax(ctl_table *table, int write,
2500                                       struct file *filp,
2501                                       void __user *buffer,
2502                                       size_t *lenp, loff_t *ppos)
2503 {
2504     return -ENOSYS;
2505 }
2506
2507 struct ctl_table_header * register_sysctl_table(ctl_table * table,
2508                                                 int insert_at_head)
2509 {
2510         return NULL;
2511 }
2512
2513 void unregister_sysctl_table(struct ctl_table_header * table)
2514 {
2515 }
2516
2517 #endif /* CONFIG_SYSCTL */
2518
2519 /*
2520  * No sense putting this after each symbol definition, twice,
2521  * exception granted :-)
2522  */
2523 EXPORT_SYMBOL(proc_dointvec);
2524 EXPORT_SYMBOL(proc_dointvec_jiffies);
2525 EXPORT_SYMBOL(proc_dointvec_minmax);
2526 EXPORT_SYMBOL(proc_dointvec_userhz_jiffies);
2527 EXPORT_SYMBOL(proc_dointvec_ms_jiffies);
2528 EXPORT_SYMBOL(proc_dostring);
2529 EXPORT_SYMBOL(proc_doulongvec_minmax);
2530 EXPORT_SYMBOL(proc_doulongvec_ms_jiffies_minmax);
2531 EXPORT_SYMBOL(register_sysctl_table);
2532 EXPORT_SYMBOL(sysctl_intvec);
2533 EXPORT_SYMBOL(sysctl_jiffies);
2534 EXPORT_SYMBOL(sysctl_ms_jiffies);
2535 EXPORT_SYMBOL(sysctl_string);
2536 EXPORT_SYMBOL(unregister_sysctl_table);