kernel/sysctl.c

   1 /*
   2  * sysctl.c: General linux system control interface
   3  *
   4  * Begun 24 March 1995, Stephen Tweedie
   5  * Added /proc support, Dec 1995
   6  * Added bdflush entry and intvec min/max checking, 2/23/96, Tom Dyas.
   7  * Added hooks for /proc/sys/net (minor, minor patch), 96/4/1, Mike Shaver.
   8  * Added kernel/java-{interpreter,appletviewer}, 96/5/10, Mike Shaver.
   9  * Dynamic registration fixes, Stephen Tweedie.
  10  * Added kswapd-interval, ctrl-alt-del, printk stuff, 1/8/97, Chris Horn.
  11  * Made sysctl support optional via CONFIG_SYSCTL, 1/10/97, Chris
  12  *  Horn.
  13  * Added proc_doulongvec_ms_jiffies_minmax, 09/08/99, Carlos H. Bauer.
  14  * Added proc_doulongvec_minmax, 09/08/99, Carlos H. Bauer.
  15  * Changed linked lists to use list.h instead of lists.h, 02/24/00, Bill
  16  *  Wendling.
  17  * The list_for_each() macro wasn't appropriate for the sysctl loop.
  18  *  Removed it and replaced it with older style, 03/23/00, Bill Wendling
  19  */
  20
  21 #include <linux/module.h>
  22 #include <linux/mm.h>
  23 #include <linux/swap.h>
  24 #include <linux/slab.h>
  25 #include <linux/sysctl.h>
  26 #include <linux/proc_fs.h>
  27 #include <linux/capability.h>
  28 #include <linux/ctype.h>
  29 #include <linux/utsname.h>
  30 #include <linux/capability.h>
  31 #include <linux/smp_lock.h>
  32 #include <linux/init.h>
  33 #include <linux/kernel.h>
  34 #include <linux/kobject.h>
  35 #include <linux/net.h>
  36 #include <linux/sysrq.h>
  37 #include <linux/highuid.h>
  38 #include <linux/writeback.h>
  39 #include <linux/hugetlb.h>
  40 #include <linux/security.h>
  41 #include <linux/initrd.h>
  42 #include <linux/times.h>
  43 #include <linux/limits.h>
  44 #include <linux/dcache.h>
  45 #include <linux/syscalls.h>
  46 #include <linux/nfs_fs.h>
  47 #include <linux/acpi.h>
  48
  49 #include <asm/uaccess.h>
  50 #include <asm/processor.h>
  51
  52 extern int proc_nr_files(ctl_table *table, int write, struct file *filp,
  53                      void __user *buffer, size_t *lenp, loff_t *ppos);
  54
  55 #if defined(CONFIG_SYSCTL)
  56
  57 /* External variables not in a header file. */
  58 extern int C_A_D;
  59 extern int sysctl_overcommit_memory;
  60 extern int sysctl_overcommit_ratio;
  61 extern int sysctl_panic_on_oom;
  62 extern int max_threads;
  63 extern int sysrq_enabled;
  64 extern int core_uses_pid;
  65 extern int suid_dumpable;
  66 extern char core_pattern[];
  67 extern int cad_pid;
  68 extern int pid_max;
  69 extern int min_free_kbytes;
  70 extern int printk_ratelimit_jiffies;
  71 extern int printk_ratelimit_burst;
  72 extern int pid_max_min, pid_max_max;
  73 extern int sysctl_drop_caches;
  74 extern int percpu_pagelist_fraction;
  75 extern int compat_log;
  76
  77 #if defined(CONFIG_X86_LOCAL_APIC) && defined(CONFIG_X86)
  78 int unknown_nmi_panic;
  79 int nmi_watchdog_enabled;
  80 extern int proc_nmi_enabled(struct ctl_table *, int , struct file *,
  81                         void __user *, size_t *, loff_t *);
  82 #endif
  83
  84 /* this is needed for the proc_dointvec_minmax for [fs_]overflow UID and GID */
  85 static int maxolduid = 65535;
  86 static int minolduid;
  87 static int min_percpu_pagelist_fract = 8;
  88
  89 static int ngroups_max = NGROUPS_MAX;
  90
  91 #ifdef CONFIG_KMOD
  92 extern char modprobe_path[];
  93 #endif
  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 #ifdef CONFIG_RT_MUTEXES
 137 extern int max_lock_depth;
 138 #endif
 139
 140 #ifdef CONFIG_SYSCTL_SYSCALL
 141 static int parse_table(int __user *, int, void __user *, size_t __user *,
 142                 void __user *, size_t, ctl_table *, void **);
 143 #endif
 144
 145 static int proc_doutsstring(ctl_table *table, int write, struct file *filp,
 146                   void __user *buffer, size_t *lenp, loff_t *ppos);
 147
 148 static ctl_table root_table[];
 149 static struct ctl_table_header root_table_header =
 150         { root_table, LIST_HEAD_INIT(root_table_header.ctl_entry) };
 151
 152 static ctl_table kern_table[];
 153 static ctl_table vm_table[];
 154 static ctl_table fs_table[];
 155 static ctl_table debug_table[];
 156 static ctl_table dev_table[];
 157 extern ctl_table random_table[];
 158 #ifdef CONFIG_UNIX98_PTYS
 159 extern ctl_table pty_table[];
 160 #endif
 161 #ifdef CONFIG_INOTIFY_USER
 162 extern ctl_table inotify_table[];
 163 #endif
 164
 165 #ifdef HAVE_ARCH_PICK_MMAP_LAYOUT
 166 int sysctl_legacy_va_layout;
 167 #endif
 168
 169 /* /proc declarations: */
 170
 171 #ifdef CONFIG_PROC_SYSCTL
 172
 173 static ssize_t proc_readsys(struct file *, char __user *, size_t, loff_t *);
 174 static ssize_t proc_writesys(struct file *, const char __user *, size_t, loff_t *);
 175 static int proc_opensys(struct inode *, struct file *);
 176
 177 struct file_operations proc_sys_file_operations = {
 178         .open           = proc_opensys,
 179         .read           = proc_readsys,
 180         .write          = proc_writesys,
 181 };
 182
 183 extern struct proc_dir_entry *proc_sys_root;
 184
 185 static void register_proc_table(ctl_table *, struct proc_dir_entry *, void *);
 186 static void unregister_proc_table(ctl_table *, struct proc_dir_entry *);
 187 #endif
 188
 189 /* The default sysctl tables: */
 190
 191 static ctl_table root_table[] = {
 192         {
 193                 .ctl_name       = CTL_KERN,
 194                 .procname       = "kernel",
 195                 .mode           = 0555,
 196                 .child          = kern_table,
 197         },
 198         {
 199                 .ctl_name       = CTL_VM,
 200                 .procname       = "vm",
 201                 .mode           = 0555,
 202                 .child          = vm_table,
 203         },
 204 #ifdef CONFIG_NET
 205         {
 206                 .ctl_name       = CTL_NET,
 207                 .procname       = "net",
 208                 .mode           = 0555,
 209                 .child          = net_table,
 210         },
 211 #endif
 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                 .strategy       = &sysctl_string,
 243         },
 244         {
 245                 .ctl_name       = KERN_OSRELEASE,
 246                 .procname       = "osrelease",
 247                 .data           = system_utsname.release,
 248                 .maxlen         = sizeof(system_utsname.release),
 249                 .mode           = 0444,
 250                 .proc_handler   = &proc_doutsstring,
 251                 .strategy       = &sysctl_string,
 252         },
 253         {
 254                 .ctl_name       = KERN_VERSION,
 255                 .procname       = "version",
 256                 .data           = system_utsname.version,
 257                 .maxlen         = sizeof(system_utsname.version),
 258                 .mode           = 0444,
 259                 .proc_handler   = &proc_doutsstring,
 260                 .strategy       = &sysctl_string,
 261         },
 262         {
 263                 .ctl_name       = KERN_NODENAME,
 264                 .procname       = "hostname",
 265                 .data           = system_utsname.nodename,
 266                 .maxlen         = sizeof(system_utsname.nodename),
 267                 .mode           = 0644,
 268                 .proc_handler   = &proc_doutsstring,
 269                 .strategy       = &sysctl_string,
 270         },
 271         {
 272                 .ctl_name       = KERN_DOMAINNAME,
 273                 .procname       = "domainname",
 274                 .data           = system_utsname.domainname,
 275                 .maxlen         = sizeof(system_utsname.domainname),
 276                 .mode           = 0644,
 277                 .proc_handler   = &proc_doutsstring,
 278                 .strategy       = &sysctl_string,
 279         },
 280         {
 281                 .ctl_name       = KERN_PANIC,
 282                 .procname       = "panic",
 283                 .data           = &panic_timeout,
 284                 .maxlen         = sizeof(int),
 285                 .mode           = 0644,
 286                 .proc_handler   = &proc_dointvec,
 287         },
 288         {
 289                 .ctl_name       = KERN_CORE_USES_PID,
 290                 .procname       = "core_uses_pid",
 291                 .data           = &core_uses_pid,
 292                 .maxlen         = sizeof(int),
 293                 .mode           = 0644,
 294                 .proc_handler   = &proc_dointvec,
 295         },
 296         {
 297                 .ctl_name       = KERN_CORE_PATTERN,
 298                 .procname       = "core_pattern",
 299                 .data           = core_pattern,
 300                 .maxlen         = 64,
 301                 .mode           = 0644,
 302                 .proc_handler   = &proc_dostring,
 303                 .strategy       = &sysctl_string,
 304         },
 305         {
 306                 .ctl_name       = KERN_TAINTED,
 307                 .procname       = "tainted",
 308                 .data           = &tainted,
 309                 .maxlen         = sizeof(int),
 310                 .mode           = 0444,
 311                 .proc_handler   = &proc_dointvec,
 312         },
 313         {
 314                 .ctl_name       = KERN_CAP_BSET,
 315                 .procname       = "cap-bound",
 316                 .data           = &cap_bset,
 317                 .maxlen         = sizeof(kernel_cap_t),
 318                 .mode           = 0600,
 319                 .proc_handler   = &proc_dointvec_bset,
 320         },
 321 #ifdef CONFIG_BLK_DEV_INITRD
 322         {
 323                 .ctl_name       = KERN_REALROOTDEV,
 324                 .procname       = "real-root-dev",
 325                 .data           = &real_root_dev,
 326                 .maxlen         = sizeof(int),
 327                 .mode           = 0644,
 328                 .proc_handler   = &proc_dointvec,
 329         },
 330 #endif
 331 #ifdef __sparc__
 332         {
 333                 .ctl_name       = KERN_SPARC_REBOOT,
 334                 .procname       = "reboot-cmd",
 335                 .data           = reboot_command,
 336                 .maxlen         = 256,
 337                 .mode           = 0644,
 338                 .proc_handler   = &proc_dostring,
 339                 .strategy       = &sysctl_string,
 340         },
 341         {
 342                 .ctl_name       = KERN_SPARC_STOP_A,
 343                 .procname       = "stop-a",
 344                 .data           = &stop_a_enabled,
 345                 .maxlen         = sizeof (int),
 346                 .mode           = 0644,
 347                 .proc_handler   = &proc_dointvec,
 348         },
 349         {
 350                 .ctl_name       = KERN_SPARC_SCONS_PWROFF,
 351                 .procname       = "scons-poweroff",
 352                 .data           = &scons_pwroff,
 353                 .maxlen         = sizeof (int),
 354                 .mode           = 0644,
 355                 .proc_handler   = &proc_dointvec,
 356         },
 357 #endif
 358 #ifdef __hppa__
 359         {
 360                 .ctl_name       = KERN_HPPA_PWRSW,
 361                 .procname       = "soft-power",
 362                 .data           = &pwrsw_enabled,
 363                 .maxlen         = sizeof (int),
 364                 .mode           = 0644,
 365                 .proc_handler   = &proc_dointvec,
 366         },
 367         {
 368                 .ctl_name       = KERN_HPPA_UNALIGNED,
 369                 .procname       = "unaligned-trap",
 370                 .data           = &unaligned_enabled,
 371                 .maxlen         = sizeof (int),
 372                 .mode           = 0644,
 373                 .proc_handler   = &proc_dointvec,
 374         },
 375 #endif
 376         {
 377                 .ctl_name       = KERN_CTLALTDEL,
 378                 .procname       = "ctrl-alt-del",
 379                 .data           = &C_A_D,
 380                 .maxlen         = sizeof(int),
 381                 .mode           = 0644,
 382                 .proc_handler   = &proc_dointvec,
 383         },
 384         {
 385                 .ctl_name       = KERN_PRINTK,
 386                 .procname       = "printk",
 387                 .data           = &console_loglevel,
 388                 .maxlen         = 4*sizeof(int),
 389                 .mode           = 0644,
 390                 .proc_handler   = &proc_dointvec,
 391         },
 392 #ifdef CONFIG_KMOD
 393         {
 394                 .ctl_name       = KERN_MODPROBE,
 395                 .procname       = "modprobe",
 396                 .data           = &modprobe_path,
 397                 .maxlen         = KMOD_PATH_LEN,
 398                 .mode           = 0644,
 399                 .proc_handler   = &proc_dostring,
 400                 .strategy       = &sysctl_string,
 401         },
 402 #endif
 403 #if defined(CONFIG_HOTPLUG) && defined(CONFIG_NET)
 404         {
 405                 .ctl_name       = KERN_HOTPLUG,
 406                 .procname       = "hotplug",
 407                 .data           = &uevent_helper,
 408                 .maxlen         = UEVENT_HELPER_PATH_LEN,
 409                 .mode           = 0644,
 410                 .proc_handler   = &proc_dostring,
 411                 .strategy       = &sysctl_string,
 412         },
 413 #endif
 414 #ifdef CONFIG_CHR_DEV_SG
 415         {
 416                 .ctl_name       = KERN_SG_BIG_BUFF,
 417                 .procname       = "sg-big-buff",
 418                 .data           = &sg_big_buff,
 419                 .maxlen         = sizeof (int),
 420                 .mode           = 0444,
 421                 .proc_handler   = &proc_dointvec,
 422         },
 423 #endif
 424 #ifdef CONFIG_BSD_PROCESS_ACCT
 425         {
 426                 .ctl_name       = KERN_ACCT,
 427                 .procname       = "acct",
 428                 .data           = &acct_parm,
 429                 .maxlen         = 3*sizeof(int),
 430                 .mode           = 0644,
 431                 .proc_handler   = &proc_dointvec,
 432         },
 433 #endif
 434 #ifdef CONFIG_SYSVIPC
 435         {
 436                 .ctl_name       = KERN_SHMMAX,
 437                 .procname       = "shmmax",
 438                 .data           = &shm_ctlmax,
 439                 .maxlen         = sizeof (size_t),
 440                 .mode           = 0644,
 441                 .proc_handler   = &proc_doulongvec_minmax,
 442         },
 443         {
 444                 .ctl_name       = KERN_SHMALL,
 445                 .procname       = "shmall",
 446                 .data           = &shm_ctlall,
 447                 .maxlen         = sizeof (size_t),
 448                 .mode           = 0644,
 449                 .proc_handler   = &proc_doulongvec_minmax,
 450         },
 451         {
 452                 .ctl_name       = KERN_SHMMNI,
 453                 .procname       = "shmmni",
 454                 .data           = &shm_ctlmni,
 455                 .maxlen         = sizeof (int),
 456                 .mode           = 0644,
 457                 .proc_handler   = &proc_dointvec,
 458         },
 459         {
 460                 .ctl_name       = KERN_MSGMAX,
 461                 .procname       = "msgmax",
 462                 .data           = &msg_ctlmax,
 463                 .maxlen         = sizeof (int),
 464                 .mode           = 0644,
 465                 .proc_handler   = &proc_dointvec,
 466         },
 467         {
 468                 .ctl_name       = KERN_MSGMNI,
 469                 .procname       = "msgmni",
 470                 .data           = &msg_ctlmni,
 471                 .maxlen         = sizeof (int),
 472                 .mode           = 0644,
 473                 .proc_handler   = &proc_dointvec,
 474         },
 475         {
 476                 .ctl_name       = KERN_MSGMNB,
 477                 .procname       =  "msgmnb",
 478                 .data           = &msg_ctlmnb,
 479                 .maxlen         = sizeof (int),
 480                 .mode           = 0644,
 481                 .proc_handler   = &proc_dointvec,
 482         },
 483         {
 484                 .ctl_name       = KERN_SEM,
 485                 .procname       = "sem",
 486                 .data           = &sem_ctls,
 487                 .maxlen         = 4*sizeof (int),
 488                 .mode           = 0644,
 489                 .proc_handler   = &proc_dointvec,
 490         },
 491 #endif
 492 #ifdef CONFIG_MAGIC_SYSRQ
 493         {
 494                 .ctl_name       = KERN_SYSRQ,
 495                 .procname       = "sysrq",
 496                 .data           = &sysrq_enabled,
 497                 .maxlen         = sizeof (int),
 498                 .mode           = 0644,
 499                 .proc_handler   = &proc_dointvec,
 500         },
 501 #endif
 502         {
 503                 .ctl_name       = KERN_CADPID,
 504                 .procname       = "cad_pid",
 505                 .data           = &cad_pid,
 506                 .maxlen         = sizeof (int),
 507                 .mode           = 0600,
 508                 .proc_handler   = &proc_dointvec,
 509         },
 510         {
 511                 .ctl_name       = KERN_MAX_THREADS,
 512                 .procname       = "threads-max",
 513                 .data           = &max_threads,
 514                 .maxlen         = sizeof(int),
 515                 .mode           = 0644,
 516                 .proc_handler   = &proc_dointvec,
 517         },
 518         {
 519                 .ctl_name       = KERN_RANDOM,
 520                 .procname       = "random",
 521                 .mode           = 0555,
 522                 .child          = random_table,
 523         },
 524 #ifdef CONFIG_UNIX98_PTYS
 525         {
 526                 .ctl_name       = KERN_PTY,
 527                 .procname       = "pty",
 528                 .mode           = 0555,
 529                 .child          = pty_table,
 530         },
 531 #endif
 532         {
 533                 .ctl_name       = KERN_OVERFLOWUID,
 534                 .procname       = "overflowuid",
 535                 .data           = &overflowuid,
 536                 .maxlen         = sizeof(int),
 537                 .mode           = 0644,
 538                 .proc_handler   = &proc_dointvec_minmax,
 539                 .strategy       = &sysctl_intvec,
 540                 .extra1         = &minolduid,
 541                 .extra2         = &maxolduid,
 542         },
 543         {
 544                 .ctl_name       = KERN_OVERFLOWGID,
 545                 .procname       = "overflowgid",
 546                 .data           = &overflowgid,
 547                 .maxlen         = sizeof(int),
 548                 .mode           = 0644,
 549                 .proc_handler   = &proc_dointvec_minmax,
 550                 .strategy       = &sysctl_intvec,
 551                 .extra1         = &minolduid,
 552                 .extra2         = &maxolduid,
 553         },
 554 #ifdef CONFIG_S390
 555 #ifdef CONFIG_MATHEMU
 556         {
 557                 .ctl_name       = KERN_IEEE_EMULATION_WARNINGS,
 558                 .procname       = "ieee_emulation_warnings",
 559                 .data           = &sysctl_ieee_emulation_warnings,
 560                 .maxlen         = sizeof(int),
 561                 .mode           = 0644,
 562                 .proc_handler   = &proc_dointvec,
 563         },
 564 #endif
 565 #ifdef CONFIG_NO_IDLE_HZ
 566         {
 567                 .ctl_name       = KERN_HZ_TIMER,
 568                 .procname       = "hz_timer",
 569                 .data           = &sysctl_hz_timer,
 570                 .maxlen         = sizeof(int),
 571                 .mode           = 0644,
 572                 .proc_handler   = &proc_dointvec,
 573         },
 574 #endif
 575         {
 576                 .ctl_name       = KERN_S390_USER_DEBUG_LOGGING,
 577                 .procname       = "userprocess_debug",
 578                 .data           = &sysctl_userprocess_debug,
 579                 .maxlen         = sizeof(int),
 580                 .mode           = 0644,
 581                 .proc_handler   = &proc_dointvec,
 582         },
 583 #endif
 584         {
 585                 .ctl_name       = KERN_PIDMAX,
 586                 .procname       = "pid_max",
 587                 .data           = &pid_max,
 588                 .maxlen         = sizeof (int),
 589                 .mode           = 0644,
 590                 .proc_handler   = &proc_dointvec_minmax,
 591                 .strategy       = sysctl_intvec,
 592                 .extra1         = &pid_max_min,
 593                 .extra2         = &pid_max_max,
 594         },
 595         {
 596                 .ctl_name       = KERN_PANIC_ON_OOPS,
 597                 .procname       = "panic_on_oops",
 598                 .data           = &panic_on_oops,
 599                 .maxlen         = sizeof(int),
 600                 .mode           = 0644,
 601                 .proc_handler   = &proc_dointvec,
 602         },
 603         {
 604                 .ctl_name       = KERN_PRINTK_RATELIMIT,
 605                 .procname       = "printk_ratelimit",
 606                 .data           = &printk_ratelimit_jiffies,
 607                 .maxlen         = sizeof(int),
 608                 .mode           = 0644,
 609                 .proc_handler   = &proc_dointvec_jiffies,
 610                 .strategy       = &sysctl_jiffies,
 611         },
 612         {
 613                 .ctl_name       = KERN_PRINTK_RATELIMIT_BURST,
 614                 .procname       = "printk_ratelimit_burst",
 615                 .data           = &printk_ratelimit_burst,
 616                 .maxlen         = sizeof(int),
 617                 .mode           = 0644,
 618                 .proc_handler   = &proc_dointvec,
 619         },
 620         {
 621                 .ctl_name       = KERN_NGROUPS_MAX,
 622                 .procname       = "ngroups_max",
 623                 .data           = &ngroups_max,
 624                 .maxlen         = sizeof (int),
 625                 .mode           = 0444,
 626                 .proc_handler   = &proc_dointvec,
 627         },
 628 #if defined(CONFIG_X86_LOCAL_APIC) && defined(CONFIG_X86)
 629         {
 630                 .ctl_name       = KERN_UNKNOWN_NMI_PANIC,
 631                 .procname       = "unknown_nmi_panic",
 632                 .data           = &unknown_nmi_panic,
 633                 .maxlen         = sizeof (int),
 634                 .mode           = 0644,
 635                 .proc_handler   = &proc_dointvec,
 636         },
 637         {
 638                 .ctl_name       = KERN_NMI_WATCHDOG,
 639                 .procname       = "nmi_watchdog",
 640                 .data           = &nmi_watchdog_enabled,
 641                 .maxlen         = sizeof (int),
 642                 .mode           = 0644,
 643                 .proc_handler   = &proc_nmi_enabled,
 644         },
 645 #endif
 646 #if defined(CONFIG_X86)
 647         {
 648                 .ctl_name       = KERN_PANIC_ON_NMI,
 649                 .procname       = "panic_on_unrecovered_nmi",
 650                 .data           = &panic_on_unrecovered_nmi,
 651                 .maxlen         = sizeof(int),
 652                 .mode           = 0644,
 653                 .proc_handler   = &proc_dointvec,
 654         },
 655         {
 656                 .ctl_name       = KERN_BOOTLOADER_TYPE,
 657                 .procname       = "bootloader_type",
 658                 .data           = &bootloader_type,
 659                 .maxlen         = sizeof (int),
 660                 .mode           = 0444,
 661                 .proc_handler   = &proc_dointvec,
 662         },
 663 #endif
 664 #if defined(CONFIG_MMU)
 665         {
 666                 .ctl_name       = KERN_RANDOMIZE,
 667                 .procname       = "randomize_va_space",
 668                 .data           = &randomize_va_space,
 669                 .maxlen         = sizeof(int),
 670                 .mode           = 0644,
 671                 .proc_handler   = &proc_dointvec,
 672         },
 673 #endif
 674 #if defined(CONFIG_S390) && defined(CONFIG_SMP)
 675         {
 676                 .ctl_name       = KERN_SPIN_RETRY,
 677                 .procname       = "spin_retry",
 678                 .data           = &spin_retry,
 679                 .maxlen         = sizeof (int),
 680                 .mode           = 0644,
 681                 .proc_handler   = &proc_dointvec,
 682         },
 683 #endif
 684 #ifdef CONFIG_ACPI_SLEEP
 685         {
 686                 .ctl_name       = KERN_ACPI_VIDEO_FLAGS,
 687                 .procname       = "acpi_video_flags",
 688                 .data           = &acpi_video_flags,
 689                 .maxlen         = sizeof (unsigned long),
 690                 .mode           = 0644,
 691                 .proc_handler   = &proc_doulongvec_minmax,
 692         },
 693 #endif
 694 #ifdef CONFIG_IA64
 695         {
 696                 .ctl_name       = KERN_IA64_UNALIGNED,
 697                 .procname       = "ignore-unaligned-usertrap",
 698                 .data           = &no_unaligned_warning,
 699                 .maxlen         = sizeof (int),
 700                 .mode           = 0644,
 701                 .proc_handler   = &proc_dointvec,
 702         },
 703 #endif
 704 #ifdef CONFIG_COMPAT
 705         {
 706                 .ctl_name       = KERN_COMPAT_LOG,
 707                 .procname       = "compat-log",
 708                 .data           = &compat_log,
 709                 .maxlen         = sizeof (int),
 710                 .mode           = 0644,
 711                 .proc_handler   = &proc_dointvec,
 712         },
 713 #endif
 714 #ifdef CONFIG_RT_MUTEXES
 715         {
 716                 .ctl_name       = KERN_MAX_LOCK_DEPTH,
 717                 .procname       = "max_lock_depth",
 718                 .data           = &max_lock_depth,
 719                 .maxlen         = sizeof(int),
 720                 .mode           = 0644,
 721                 .proc_handler   = &proc_dointvec,
 722         },
 723 #endif
 724
 725         { .ctl_name = 0 }
 726 };
 727
 728 /* Constants for minimum and maximum testing in vm_table.
 729    We use these as one-element integer vectors. */
 730 static int zero;
 731 static int one_hundred = 100;
 732
 733
 734 static ctl_table vm_table[] = {
 735         {
 736                 .ctl_name       = VM_OVERCOMMIT_MEMORY,
 737                 .procname       = "overcommit_memory",
 738                 .data           = &sysctl_overcommit_memory,
 739                 .maxlen         = sizeof(sysctl_overcommit_memory),
 740                 .mode           = 0644,
 741                 .proc_handler   = &proc_dointvec,
 742         },
 743         {
 744                 .ctl_name       = VM_PANIC_ON_OOM,
 745                 .procname       = "panic_on_oom",
 746                 .data           = &sysctl_panic_on_oom,
 747                 .maxlen         = sizeof(sysctl_panic_on_oom),
 748                 .mode           = 0644,
 749                 .proc_handler   = &proc_dointvec,
 750         },
 751         {
 752                 .ctl_name       = VM_OVERCOMMIT_RATIO,
 753                 .procname       = "overcommit_ratio",
 754                 .data           = &sysctl_overcommit_ratio,
 755                 .maxlen         = sizeof(sysctl_overcommit_ratio),
 756                 .mode           = 0644,
 757                 .proc_handler   = &proc_dointvec,
 758         },
 759         {
 760                 .ctl_name       = VM_PAGE_CLUSTER,
 761                 .procname       = "page-cluster",
 762                 .data           = &page_cluster,
 763                 .maxlen         = sizeof(int),
 764                 .mode           = 0644,
 765                 .proc_handler   = &proc_dointvec,
 766         },
 767         {
 768                 .ctl_name       = VM_DIRTY_BACKGROUND,
 769                 .procname       = "dirty_background_ratio",
 770                 .data           = &dirty_background_ratio,
 771                 .maxlen         = sizeof(dirty_background_ratio),
 772                 .mode           = 0644,
 773                 .proc_handler   = &proc_dointvec_minmax,
 774                 .strategy       = &sysctl_intvec,
 775                 .extra1         = &zero,
 776                 .extra2         = &one_hundred,
 777         },
 778         {
 779                 .ctl_name       = VM_DIRTY_RATIO,
 780                 .procname       = "dirty_ratio",
 781                 .data           = &vm_dirty_ratio,
 782                 .maxlen         = sizeof(vm_dirty_ratio),
 783                 .mode           = 0644,
 784                 .proc_handler   = &proc_dointvec_minmax,
 785                 .strategy       = &sysctl_intvec,
 786                 .extra1         = &zero,
 787                 .extra2         = &one_hundred,
 788         },
 789         {
 790                 .ctl_name       = VM_DIRTY_WB_CS,
 791                 .procname       = "dirty_writeback_centisecs",
 792                 .data           = &dirty_writeback_interval,
 793                 .maxlen         = sizeof(dirty_writeback_interval),
 794                 .mode           = 0644,
 795                 .proc_handler   = &dirty_writeback_centisecs_handler,
 796         },
 797         {
 798                 .ctl_name       = VM_DIRTY_EXPIRE_CS,
 799                 .procname       = "dirty_expire_centisecs",
 800                 .data           = &dirty_expire_interval,
 801                 .maxlen         = sizeof(dirty_expire_interval),
 802                 .mode           = 0644,
 803                 .proc_handler   = &proc_dointvec_userhz_jiffies,
 804         },
 805         {
 806                 .ctl_name       = VM_NR_PDFLUSH_THREADS,
 807                 .procname       = "nr_pdflush_threads",
 808                 .data           = &nr_pdflush_threads,
 809                 .maxlen         = sizeof nr_pdflush_threads,
 810                 .mode           = 0444 /* read-only*/,
 811                 .proc_handler   = &proc_dointvec,
 812         },
 813         {
 814                 .ctl_name       = VM_SWAPPINESS,
 815                 .procname       = "swappiness",
 816                 .data           = &vm_swappiness,
 817                 .maxlen         = sizeof(vm_swappiness),
 818                 .mode           = 0644,
 819                 .proc_handler   = &proc_dointvec_minmax,
 820                 .strategy       = &sysctl_intvec,
 821                 .extra1         = &zero,
 822                 .extra2         = &one_hundred,
 823         },
 824 #ifdef CONFIG_HUGETLB_PAGE
 825          {
 826                 .ctl_name       = VM_HUGETLB_PAGES,
 827                 .procname       = "nr_hugepages",
 828                 .data           = &max_huge_pages,
 829                 .maxlen         = sizeof(unsigned long),
 830                 .mode           = 0644,
 831                 .proc_handler   = &hugetlb_sysctl_handler,
 832                 .extra1         = (void *)&hugetlb_zero,
 833                 .extra2         = (void *)&hugetlb_infinity,
 834          },
 835          {
 836                 .ctl_name       = VM_HUGETLB_GROUP,
 837                 .procname       = "hugetlb_shm_group",
 838                 .data           = &sysctl_hugetlb_shm_group,
 839                 .maxlen         = sizeof(gid_t),
 840                 .mode           = 0644,
 841                 .proc_handler   = &proc_dointvec,
 842          },
 843 #endif
 844         {
 845                 .ctl_name       = VM_LOWMEM_RESERVE_RATIO,
 846                 .procname       = "lowmem_reserve_ratio",
 847                 .data           = &sysctl_lowmem_reserve_ratio,
 848                 .maxlen         = sizeof(sysctl_lowmem_reserve_ratio),
 849                 .mode           = 0644,
 850                 .proc_handler   = &lowmem_reserve_ratio_sysctl_handler,
 851                 .strategy       = &sysctl_intvec,
 852         },
 853         {
 854                 .ctl_name       = VM_DROP_PAGECACHE,
 855                 .procname       = "drop_caches",
 856                 .data           = &sysctl_drop_caches,
 857                 .maxlen         = sizeof(int),
 858                 .mode           = 0644,
 859                 .proc_handler   = drop_caches_sysctl_handler,
 860                 .strategy       = &sysctl_intvec,
 861         },
 862         {
 863                 .ctl_name       = VM_MIN_FREE_KBYTES,
 864                 .procname       = "min_free_kbytes",
 865                 .data           = &min_free_kbytes,
 866                 .maxlen         = sizeof(min_free_kbytes),
 867                 .mode           = 0644,
 868                 .proc_handler   = &min_free_kbytes_sysctl_handler,
 869                 .strategy       = &sysctl_intvec,
 870                 .extra1         = &zero,
 871         },
 872         {
 873                 .ctl_name       = VM_PERCPU_PAGELIST_FRACTION,
 874                 .procname       = "percpu_pagelist_fraction",
 875                 .data           = &percpu_pagelist_fraction,
 876                 .maxlen         = sizeof(percpu_pagelist_fraction),
 877                 .mode           = 0644,
 878                 .proc_handler   = &percpu_pagelist_fraction_sysctl_handler,
 879                 .strategy       = &sysctl_intvec,
 880                 .extra1         = &min_percpu_pagelist_fract,
 881         },
 882 #ifdef CONFIG_MMU
 883         {
 884                 .ctl_name       = VM_MAX_MAP_COUNT,
 885                 .procname       = "max_map_count",
 886                 .data           = &sysctl_max_map_count,
 887                 .maxlen         = sizeof(sysctl_max_map_count),
 888                 .mode           = 0644,
 889                 .proc_handler   = &proc_dointvec
 890         },
 891 #endif
 892         {
 893                 .ctl_name       = VM_LAPTOP_MODE,
 894                 .procname       = "laptop_mode",
 895                 .data           = &laptop_mode,
 896                 .maxlen         = sizeof(laptop_mode),
 897                 .mode           = 0644,
 898                 .proc_handler   = &proc_dointvec_jiffies,
 899                 .strategy       = &sysctl_jiffies,
 900         },
 901         {
 902                 .ctl_name       = VM_BLOCK_DUMP,
 903                 .procname       = "block_dump",
 904                 .data           = &block_dump,
 905                 .maxlen         = sizeof(block_dump),
 906                 .mode           = 0644,
 907                 .proc_handler   = &proc_dointvec,
 908                 .strategy       = &sysctl_intvec,
 909                 .extra1         = &zero,
 910         },
 911         {
 912                 .ctl_name       = VM_VFS_CACHE_PRESSURE,
 913                 .procname       = "vfs_cache_pressure",
 914                 .data           = &sysctl_vfs_cache_pressure,
 915                 .maxlen         = sizeof(sysctl_vfs_cache_pressure),
 916                 .mode           = 0644,
 917                 .proc_handler   = &proc_dointvec,
 918                 .strategy       = &sysctl_intvec,
 919                 .extra1         = &zero,
 920         },
 921 #ifdef HAVE_ARCH_PICK_MMAP_LAYOUT
 922         {
 923                 .ctl_name       = VM_LEGACY_VA_LAYOUT,
 924                 .procname       = "legacy_va_layout",
 925                 .data           = &sysctl_legacy_va_layout,
 926                 .maxlen         = sizeof(sysctl_legacy_va_layout),
 927                 .mode           = 0644,
 928                 .proc_handler   = &proc_dointvec,
 929                 .strategy       = &sysctl_intvec,
 930                 .extra1         = &zero,
 931         },
 932 #endif
 933 #ifdef CONFIG_SWAP
 934         {
 935                 .ctl_name       = VM_SWAP_TOKEN_TIMEOUT,
 936                 .procname       = "swap_token_timeout",
 937                 .data           = &swap_token_default_timeout,
 938                 .maxlen         = sizeof(swap_token_default_timeout),
 939                 .mode           = 0644,
 940                 .proc_handler   = &proc_dointvec_jiffies,
 941                 .strategy       = &sysctl_jiffies,
 942         },
 943 #endif
 944 #ifdef CONFIG_NUMA
 945         {
 946                 .ctl_name       = VM_ZONE_RECLAIM_MODE,
 947                 .procname       = "zone_reclaim_mode",
 948                 .data           = &zone_reclaim_mode,
 949                 .maxlen         = sizeof(zone_reclaim_mode),
 950                 .mode           = 0644,
 951                 .proc_handler   = &proc_dointvec,
 952                 .strategy       = &sysctl_intvec,
 953                 .extra1         = &zero,
 954         },
 955         {
 956                 .ctl_name       = VM_MIN_UNMAPPED,
 957                 .procname       = "min_unmapped_ratio",
 958                 .data           = &sysctl_min_unmapped_ratio,
 959                 .maxlen         = sizeof(sysctl_min_unmapped_ratio),
 960                 .mode           = 0644,
 961                 .proc_handler   = &sysctl_min_unmapped_ratio_sysctl_handler,
 962                 .strategy       = &sysctl_intvec,
 963                 .extra1         = &zero,
 964                 .extra2         = &one_hundred,
 965         },
 966         {
 967                 .ctl_name       = VM_MIN_SLAB,
 968                 .procname       = "min_slab_ratio",
 969                 .data           = &sysctl_min_slab_ratio,
 970                 .maxlen         = sizeof(sysctl_min_slab_ratio),
 971                 .mode           = 0644,
 972                 .proc_handler   = &sysctl_min_slab_ratio_sysctl_handler,
 973                 .strategy       = &sysctl_intvec,
 974                 .extra1         = &zero,
 975                 .extra2         = &one_hundred,
 976         },
 977 #endif
 978 #ifdef CONFIG_X86_32
 979         {
 980                 .ctl_name       = VM_VDSO_ENABLED,
 981                 .procname       = "vdso_enabled",
 982                 .data           = &vdso_enabled,
 983                 .maxlen         = sizeof(vdso_enabled),
 984                 .mode           = 0644,
 985                 .proc_handler   = &proc_dointvec,
 986                 .strategy       = &sysctl_intvec,
 987                 .extra1         = &zero,
 988         },
 989 #endif
 990         { .ctl_name = 0 }
 991 };
 992
 993 static ctl_table fs_table[] = {
 994         {
 995                 .ctl_name       = FS_NRINODE,
 996                 .procname       = "inode-nr",
 997                 .data           = &inodes_stat,
 998                 .maxlen         = 2*sizeof(int),
 999                 .mode           = 0444,
1000                 .proc_handler   = &proc_dointvec,
1001         },
1002         {
1003                 .ctl_name       = FS_STATINODE,
1004                 .procname       = "inode-state",
1005                 .data           = &inodes_stat,
1006                 .maxlen         = 7*sizeof(int),
1007                 .mode           = 0444,
1008                 .proc_handler   = &proc_dointvec,
1009         },
1010         {
1011                 .ctl_name       = FS_NRFILE,
1012                 .procname       = "file-nr",
1013                 .data           = &files_stat,
1014                 .maxlen         = 3*sizeof(int),
1015                 .mode           = 0444,
1016                 .proc_handler   = &proc_nr_files,
1017         },
1018         {
1019                 .ctl_name       = FS_MAXFILE,
1020                 .procname       = "file-max",
1021                 .data           = &files_stat.max_files,
1022                 .maxlen         = sizeof(int),
1023                 .mode           = 0644,
1024                 .proc_handler   = &proc_dointvec,
1025         },
1026         {
1027                 .ctl_name       = FS_DENTRY,
1028                 .procname       = "dentry-state",
1029                 .data           = &dentry_stat,
1030                 .maxlen         = 6*sizeof(int),
1031                 .mode           = 0444,
1032                 .proc_handler   = &proc_dointvec,
1033         },
1034         {
1035                 .ctl_name       = FS_OVERFLOWUID,
1036                 .procname       = "overflowuid",
1037                 .data           = &fs_overflowuid,
1038                 .maxlen         = sizeof(int),
1039                 .mode           = 0644,
1040                 .proc_handler   = &proc_dointvec_minmax,
1041                 .strategy       = &sysctl_intvec,
1042                 .extra1         = &minolduid,
1043                 .extra2         = &maxolduid,
1044         },
1045         {
1046                 .ctl_name       = FS_OVERFLOWGID,
1047                 .procname       = "overflowgid",
1048                 .data           = &fs_overflowgid,
1049                 .maxlen         = sizeof(int),
1050                 .mode           = 0644,
1051                 .proc_handler   = &proc_dointvec_minmax,
1052                 .strategy       = &sysctl_intvec,
1053                 .extra1         = &minolduid,
1054                 .extra2         = &maxolduid,
1055         },
1056         {
1057                 .ctl_name       = FS_LEASES,
1058                 .procname       = "leases-enable",
1059                 .data           = &leases_enable,
1060                 .maxlen         = sizeof(int),
1061                 .mode           = 0644,
1062                 .proc_handler   = &proc_dointvec,
1063         },
1064 #ifdef CONFIG_DNOTIFY
1065         {
1066                 .ctl_name       = FS_DIR_NOTIFY,
1067                 .procname       = "dir-notify-enable",
1068                 .data           = &dir_notify_enable,
1069                 .maxlen         = sizeof(int),
1070                 .mode           = 0644,
1071                 .proc_handler   = &proc_dointvec,
1072         },
1073 #endif
1074 #ifdef CONFIG_MMU
1075         {
1076                 .ctl_name       = FS_LEASE_TIME,
1077                 .procname       = "lease-break-time",
1078                 .data           = &lease_break_time,
1079                 .maxlen         = sizeof(int),
1080                 .mode           = 0644,
1081                 .proc_handler   = &proc_dointvec,
1082         },
1083         {
1084                 .ctl_name       = FS_AIO_NR,
1085                 .procname       = "aio-nr",
1086                 .data           = &aio_nr,
1087                 .maxlen         = sizeof(aio_nr),
1088                 .mode           = 0444,
1089                 .proc_handler   = &proc_doulongvec_minmax,
1090         },
1091         {
1092                 .ctl_name       = FS_AIO_MAX_NR,
1093                 .procname       = "aio-max-nr",
1094                 .data           = &aio_max_nr,
1095                 .maxlen         = sizeof(aio_max_nr),
1096                 .mode           = 0644,
1097                 .proc_handler   = &proc_doulongvec_minmax,
1098         },
1099 #ifdef CONFIG_INOTIFY_USER
1100         {
1101                 .ctl_name       = FS_INOTIFY,
1102                 .procname       = "inotify",
1103                 .mode           = 0555,
1104                 .child          = inotify_table,
1105         },
1106 #endif
1107 #endif
1108         {
1109                 .ctl_name       = KERN_SETUID_DUMPABLE,
1110                 .procname       = "suid_dumpable",
1111                 .data           = &suid_dumpable,
1112                 .maxlen         = sizeof(int),
1113                 .mode           = 0644,
1114                 .proc_handler   = &proc_dointvec,
1115         },
1116         { .ctl_name = 0 }
1117 };
1118
1119 static ctl_table debug_table[] = {
1120         { .ctl_name = 0 }
1121 };
1122
1123 static ctl_table dev_table[] = {
1124         { .ctl_name = 0 }
1125 };
1126
1127 extern void init_irq_proc (void);
1128
1129 static DEFINE_SPINLOCK(sysctl_lock);
1130
1131 /* called under sysctl_lock */
1132 static int use_table(struct ctl_table_header *p)
1133 {
1134         if (unlikely(p->unregistering))
1135                 return 0;
1136         p->used++;
1137         return 1;
1138 }
1139
1140 /* called under sysctl_lock */
1141 static void unuse_table(struct ctl_table_header *p)
1142 {
1143         if (!--p->used)
1144                 if (unlikely(p->unregistering))
1145                         complete(p->unregistering);
1146 }
1147
1148 /* called under sysctl_lock, will reacquire if has to wait */
1149 static void start_unregistering(struct ctl_table_header *p)
1150 {
1151         /*
1152          * if p->used is 0, nobody will ever touch that entry again;
1153          * we'll eliminate all paths to it before dropping sysctl_lock
1154          */
1155         if (unlikely(p->used)) {
1156                 struct completion wait;
1157                 init_completion(&wait);
1158                 p->unregistering = &wait;
1159                 spin_unlock(&sysctl_lock);
1160                 wait_for_completion(&wait);
1161                 spin_lock(&sysctl_lock);
1162         }
1163         /*
1164          * do not remove from the list until nobody holds it; walking the
1165          * list in do_sysctl() relies on that.
1166          */
1167         list_del_init(&p->ctl_entry);
1168 }
1169
1170 void __init sysctl_init(void)
1171 {
1172 #ifdef CONFIG_PROC_SYSCTL
1173         register_proc_table(root_table, proc_sys_root, &root_table_header);
1174         init_irq_proc();
1175 #endif
1176 }
1177
1178 #ifdef CONFIG_SYSCTL_SYSCALL
1179 int do_sysctl(int __user *name, int nlen, void __user *oldval, size_t __user *oldlenp,
1180                void __user *newval, size_t newlen)
1181 {
1182         struct list_head *tmp;
1183         int error = -ENOTDIR;
1184
1185         if (nlen <= 0 || nlen >= CTL_MAXNAME)
1186                 return -ENOTDIR;
1187         if (oldval) {
1188                 int old_len;
1189                 if (!oldlenp || get_user(old_len, oldlenp))
1190                         return -EFAULT;
1191         }
1192         spin_lock(&sysctl_lock);
1193         tmp = &root_table_header.ctl_entry;
1194         do {
1195                 struct ctl_table_header *head =
1196                         list_entry(tmp, struct ctl_table_header, ctl_entry);
1197                 void *context = NULL;
1198
1199                 if (!use_table(head))
1200                         continue;
1201
1202                 spin_unlock(&sysctl_lock);
1203
1204                 error = parse_table(name, nlen, oldval, oldlenp,
1205                                         newval, newlen, head->ctl_table,
1206                                         &context);
1207                 kfree(context);
1208
1209                 spin_lock(&sysctl_lock);
1210                 unuse_table(head);
1211                 if (error != -ENOTDIR)
1212                         break;
1213         } while ((tmp = tmp->next) != &root_table_header.ctl_entry);
1214         spin_unlock(&sysctl_lock);
1215         return error;
1216 }
1217
1218 asmlinkage long sys_sysctl(struct __sysctl_args __user *args)
1219 {
1220         struct __sysctl_args tmp;
1221         int error;
1222
1223         if (copy_from_user(&tmp, args, sizeof(tmp)))
1224                 return -EFAULT;
1225
1226         lock_kernel();
1227         error = do_sysctl(tmp.name, tmp.nlen, tmp.oldval, tmp.oldlenp,
1228                           tmp.newval, tmp.newlen);
1229         unlock_kernel();
1230         return error;
1231 }
1232 #endif /* CONFIG_SYSCTL_SYSCALL */
1233
1234 /*
1235  * ctl_perm does NOT grant the superuser all rights automatically, because
1236  * some sysctl variables are readonly even to root.
1237  */
1238
1239 static int test_perm(int mode, int op)
1240 {
1241         if (!current->euid)
1242                 mode >>= 6;
1243         else if (in_egroup_p(0))
1244                 mode >>= 3;
1245         if ((mode & op & 0007) == op)
1246                 return 0;
1247         return -EACCES;
1248 }
1249
1250 static inline int ctl_perm(ctl_table *table, int op)
1251 {
1252         int error;
1253         error = security_sysctl(table, op);
1254         if (error)
1255                 return error;
1256         return test_perm(table->mode, op);
1257 }
1258
1259 #ifdef CONFIG_SYSCTL_SYSCALL
1260 static int parse_table(int __user *name, int nlen,
1261                        void __user *oldval, size_t __user *oldlenp,
1262                        void __user *newval, size_t newlen,
1263                        ctl_table *table, void **context)
1264 {
1265         int n;
1266 repeat:
1267         if (!nlen)
1268                 return -ENOTDIR;
1269         if (get_user(n, name))
1270                 return -EFAULT;
1271         for ( ; table->ctl_name; table++) {
1272                 if (n == table->ctl_name || table->ctl_name == CTL_ANY) {
1273                         int error;
1274                         if (table->child) {
1275                                 if (ctl_perm(table, 001))
1276                                         return -EPERM;
1277                                 if (table->strategy) {
1278                                         error = table->strategy(
1279                                                 table, name, nlen,
1280                                                 oldval, oldlenp,
1281                                                 newval, newlen, context);
1282                                         if (error)
1283                                                 return error;
1284                                 }
1285                                 name++;
1286                                 nlen--;
1287                                 table = table->child;
1288                                 goto repeat;
1289                         }
1290                         error = do_sysctl_strategy(table, name, nlen,
1291                                                    oldval, oldlenp,
1292                                                    newval, newlen, context);
1293                         return error;
1294                 }
1295         }
1296         return -ENOTDIR;
1297 }
1298
1299 /* Perform the actual read/write of a sysctl table entry. */
1300 int do_sysctl_strategy (ctl_table *table,
1301                         int __user *name, int nlen,
1302                         void __user *oldval, size_t __user *oldlenp,
1303                         void __user *newval, size_t newlen, void **context)
1304 {
1305         int op = 0, rc;
1306         size_t len;
1307
1308         if (oldval)
1309                 op |= 004;
1310         if (newval)
1311                 op |= 002;
1312         if (ctl_perm(table, op))
1313                 return -EPERM;
1314
1315         if (table->strategy) {
1316                 rc = table->strategy(table, name, nlen, oldval, oldlenp,
1317                                      newval, newlen, context);
1318                 if (rc < 0)
1319                         return rc;
1320                 if (rc > 0)
1321                         return 0;
1322         }
1323
1324         /* If there is no strategy routine, or if the strategy returns
1325          * zero, proceed with automatic r/w */
1326         if (table->data && table->maxlen) {
1327                 if (oldval && oldlenp) {
1328                         if (get_user(len, oldlenp))
1329                                 return -EFAULT;
1330                         if (len) {
1331                                 if (len > table->maxlen)
1332                                         len = table->maxlen;
1333                                 if(copy_to_user(oldval, table->data, len))
1334                                         return -EFAULT;
1335                                 if(put_user(len, oldlenp))
1336                                         return -EFAULT;
1337                         }
1338                 }
1339                 if (newval && newlen) {
1340                         len = newlen;
1341                         if (len > table->maxlen)
1342                                 len = table->maxlen;
1343                         if(copy_from_user(table->data, newval, len))
1344                                 return -EFAULT;
1345                 }
1346         }
1347         return 0;
1348 }
1349 #endif /* CONFIG_SYSCTL_SYSCALL */
1350
1351 /**
1352  * register_sysctl_table - register a sysctl hierarchy
1353  * @table: the top-level table structure
1354  * @insert_at_head: whether the entry should be inserted in front or at the end
1355  *
1356  * Register a sysctl table hierarchy. @table should be a filled in ctl_table
1357  * array. An entry with a ctl_name of 0 terminates the table.
1358  *
1359  * The members of the &ctl_table structure are used as follows:
1360  *
1361  * ctl_name - This is the numeric sysctl value used by sysctl(2). The number
1362  *            must be unique within that level of sysctl
1363  *
1364  * procname - the name of the sysctl file under /proc/sys. Set to %NULL to not
1365  *            enter a sysctl file
1366  *
1367  * data - a pointer to data for use by proc_handler
1368  *
1369  * maxlen - the maximum size in bytes of the data
1370  *
1371  * mode - the file permissions for the /proc/sys file, and for sysctl(2)
1372  *
1373  * child - a pointer to the child sysctl table if this entry is a directory, or
1374  *         %NULL.
1375  *
1376  * proc_handler - the text handler routine (described below)
1377  *
1378  * strategy - the strategy routine (described below)
1379  *
1380  * de - for internal use by the sysctl routines
1381  *
1382  * extra1, extra2 - extra pointers usable by the proc handler routines
1383  *
1384  * Leaf nodes in the sysctl tree will be represented by a single file
1385  * under /proc; non-leaf nodes will be represented by directories.
1386  *
1387  * sysctl(2) can automatically manage read and write requests through
1388  * the sysctl table.  The data and maxlen fields of the ctl_table
1389  * struct enable minimal validation of the values being written to be
1390  * performed, and the mode field allows minimal authentication.
1391  *
1392  * More sophisticated management can be enabled by the provision of a
1393  * strategy routine with the table entry.  This will be called before
1394  * any automatic read or write of the data is performed.
1395  *
1396  * The strategy routine may return
1397  *
1398  * < 0 - Error occurred (error is passed to user process)
1399  *
1400  * 0   - OK - proceed with automatic read or write.
1401  *
1402  * > 0 - OK - read or write has been done by the strategy routine, so
1403  *       return immediately.
1404  *
1405  * There must be a proc_handler routine for any terminal nodes
1406  * mirrored under /proc/sys (non-terminals are handled by a built-in
1407  * directory handler).  Several default handlers are available to
1408  * cover common cases -
1409  *
1410  * proc_dostring(), proc_dointvec(), proc_dointvec_jiffies(),
1411  * proc_dointvec_userhz_jiffies(), proc_dointvec_minmax(),
1412  * proc_doulongvec_ms_jiffies_minmax(), proc_doulongvec_minmax()
1413  *
1414  * It is the handler's job to read the input buffer from user memory
1415  * and process it. The handler should return 0 on success.
1416  *
1417  * This routine returns %NULL on a failure to register, and a pointer
1418  * to the table header on success.
1419  */
1420 struct ctl_table_header *register_sysctl_table(ctl_table * table,
1421                                                int insert_at_head)
1422 {
1423         struct ctl_table_header *tmp;
1424         tmp = kmalloc(sizeof(struct ctl_table_header), GFP_KERNEL);
1425         if (!tmp)
1426                 return NULL;
1427         tmp->ctl_table = table;
1428         INIT_LIST_HEAD(&tmp->ctl_entry);
1429         tmp->used = 0;
1430         tmp->unregistering = NULL;
1431         spin_lock(&sysctl_lock);
1432         if (insert_at_head)
1433                 list_add(&tmp->ctl_entry, &root_table_header.ctl_entry);
1434         else
1435                 list_add_tail(&tmp->ctl_entry, &root_table_header.ctl_entry);
1436         spin_unlock(&sysctl_lock);
1437 #ifdef CONFIG_PROC_SYSCTL
1438         register_proc_table(table, proc_sys_root, tmp);
1439 #endif
1440         return tmp;
1441 }
1442
1443 /**
1444  * unregister_sysctl_table - unregister a sysctl table hierarchy
1445  * @header: the header returned from register_sysctl_table
1446  *
1447  * Unregisters the sysctl table and all children. proc entries may not
1448  * actually be removed until they are no longer used by anyone.
1449  */
1450 void unregister_sysctl_table(struct ctl_table_header * header)
1451 {
1452         might_sleep();
1453         spin_lock(&sysctl_lock);
1454         start_unregistering(header);
1455 #ifdef CONFIG_PROC_SYSCTL
1456         unregister_proc_table(header->ctl_table, proc_sys_root);
1457 #endif
1458         spin_unlock(&sysctl_lock);
1459         kfree(header);
1460 }
1461
1462 #else /* !CONFIG_SYSCTL */
1463 struct ctl_table_header * register_sysctl_table(ctl_table * table,
1464                                                 int insert_at_head)
1465 {
1466         return NULL;
1467 }
1468
1469 void unregister_sysctl_table(struct ctl_table_header * table)
1470 {
1471 }
1472
1473 #endif /* CONFIG_SYSCTL */
1474
1475 /*
1476  * /proc/sys support
1477  */
1478
1479 #ifdef CONFIG_PROC_SYSCTL
1480
1481 /* Scan the sysctl entries in table and add them all into /proc */
1482 static void register_proc_table(ctl_table * table, struct proc_dir_entry *root, void *set)
1483 {
1484         struct proc_dir_entry *de;
1485         int len;
1486         mode_t mode;
1487
1488         for (; table->ctl_name; table++) {
1489                 /* Can't do anything without a proc name. */
1490                 if (!table->procname)
1491                         continue;
1492                 /* Maybe we can't do anything with it... */
1493                 if (!table->proc_handler && !table->child) {
1494                         printk(KERN_WARNING "SYSCTL: Can't register %s\n",
1495                                 table->procname);
1496                         continue;
1497                 }
1498
1499                 len = strlen(table->procname);
1500                 mode = table->mode;
1501
1502                 de = NULL;
1503                 if (table->proc_handler)
1504                         mode |= S_IFREG;
1505                 else {
1506                         mode |= S_IFDIR;
1507                         for (de = root->subdir; de; de = de->next) {
1508                                 if (proc_match(len, table->procname, de))
1509                                         break;
1510                         }
1511                         /* If the subdir exists already, de is non-NULL */
1512                 }
1513
1514                 if (!de) {
1515                         de = create_proc_entry(table->procname, mode, root);
1516                         if (!de)
1517                                 continue;
1518                         de->set = set;
1519                         de->data = (void *) table;
1520                         if (table->proc_handler)
1521                                 de->proc_fops = &proc_sys_file_operations;
1522                 }
1523                 table->de = de;
1524                 if (de->mode & S_IFDIR)
1525                         register_proc_table(table->child, de, set);
1526         }
1527 }
1528
1529 /*
1530  * Unregister a /proc sysctl table and any subdirectories.
1531  */
1532 static void unregister_proc_table(ctl_table * table, struct proc_dir_entry *root)
1533 {
1534         struct proc_dir_entry *de;
1535         for (; table->ctl_name; table++) {
1536                 if (!(de = table->de))
1537                         continue;
1538                 if (de->mode & S_IFDIR) {
1539                         if (!table->child) {
1540                                 printk (KERN_ALERT "Help - malformed sysctl tree on free\n");
1541                                 continue;
1542                         }
1543                         unregister_proc_table(table->child, de);
1544
1545                         /* Don't unregister directories which still have entries.. */
1546                         if (de->subdir)
1547                                 continue;
1548                 }
1549
1550                 /*
1551                  * In any case, mark the entry as goner; we'll keep it
1552                  * around if it's busy, but we'll know to do nothing with
1553                  * its fields.  We are under sysctl_lock here.
1554                  */
1555                 de->data = NULL;
1556
1557                 /* Don't unregister proc entries that are still being used.. */
1558                 if (atomic_read(&de->count))
1559                         continue;
1560
1561                 table->de = NULL;
1562                 remove_proc_entry(table->procname, root);
1563         }
1564 }
1565
1566 static ssize_t do_rw_proc(int write, struct file * file, char __user * buf,
1567                           size_t count, loff_t *ppos)
1568 {
1569         int op;
1570         struct proc_dir_entry *de = PDE(file->f_dentry->d_inode);
1571         struct ctl_table *table;
1572         size_t res;
1573         ssize_t error = -ENOTDIR;
1574
1575         spin_lock(&sysctl_lock);
1576         if (de && de->data && use_table(de->set)) {
1577                 /*
1578                  * at that point we know that sysctl was not unregistered
1579                  * and won't be until we finish
1580                  */
1581                 spin_unlock(&sysctl_lock);
1582                 table = (struct ctl_table *) de->data;
1583                 if (!table || !table->proc_handler)
1584                         goto out;
1585                 error = -EPERM;
1586                 op = (write ? 002 : 004);
1587                 if (ctl_perm(table, op))
1588                         goto out;
1589
1590                 /* careful: calling conventions are nasty here */
1591                 res = count;
1592                 error = (*table->proc_handler)(table, write, file,
1593                                                 buf, &res, ppos);
1594                 if (!error)
1595                         error = res;
1596         out:
1597                 spin_lock(&sysctl_lock);
1598                 unuse_table(de->set);
1599         }
1600         spin_unlock(&sysctl_lock);
1601         return error;
1602 }
1603
1604 static int proc_opensys(struct inode *inode, struct file *file)
1605 {
1606         if (file->f_mode & FMODE_WRITE) {
1607                 /*
1608                  * sysctl entries that are not writable,
1609                  * are _NOT_ writable, capabilities or not.
1610                  */
1611                 if (!(inode->i_mode & S_IWUSR))
1612                         return -EPERM;
1613         }
1614
1615         return 0;
1616 }
1617
1618 static ssize_t proc_readsys(struct file * file, char __user * buf,
1619                             size_t count, loff_t *ppos)
1620 {
1621         return do_rw_proc(0, file, buf, count, ppos);
1622 }
1623
1624 static ssize_t proc_writesys(struct file * file, const char __user * buf,
1625                              size_t count, loff_t *ppos)
1626 {
1627         return do_rw_proc(1, file, (char __user *) buf, count, ppos);
1628 }
1629
1630 /**
1631  * proc_dostring - read a string sysctl
1632  * @table: the sysctl table
1633  * @write: %TRUE if this is a write to the sysctl file
1634  * @filp: the file structure
1635  * @buffer: the user buffer
1636  * @lenp: the size of the user buffer
1637  * @ppos: file position
1638  *
1639  * Reads/writes a string from/to the user buffer. If the kernel
1640  * buffer provided is not large enough to hold the string, the
1641  * string is truncated. The copied string is %NULL-terminated.
1642  * If the string is being read by the user process, it is copied
1643  * and a newline '\n' is added. It is truncated if the buffer is
1644  * not large enough.
1645  *
1646  * Returns 0 on success.
1647  */
1648 int proc_dostring(ctl_table *table, int write, struct file *filp,
1649                   void __user *buffer, size_t *lenp, loff_t *ppos)
1650 {
1651         size_t len;
1652         char __user *p;
1653         char c;
1654
1655         if (!table->data || !table->maxlen || !*lenp ||
1656             (*ppos && !write)) {
1657                 *lenp = 0;
1658                 return 0;
1659         }
1660
1661         if (write) {
1662                 len = 0;
1663                 p = buffer;
1664                 while (len < *lenp) {
1665                         if (get_user(c, p++))
1666                                 return -EFAULT;
1667                         if (c == 0 || c == '\n')
1668                                 break;
1669                         len++;
1670                 }
1671                 if (len >= table->maxlen)
1672                         len = table->maxlen-1;
1673                 if(copy_from_user(table->data, buffer, len))
1674                         return -EFAULT;
1675                 ((char *) table->data)[len] = 0;
1676                 *ppos += *lenp;
1677         } else {
1678                 len = strlen(table->data);
1679                 if (len > table->maxlen)
1680                         len = table->maxlen;
1681                 if (len > *lenp)
1682                         len = *lenp;
1683                 if (len)
1684                         if(copy_to_user(buffer, table->data, len))
1685                                 return -EFAULT;
1686                 if (len < *lenp) {
1687                         if(put_user('\n', ((char __user *) buffer) + len))
1688                                 return -EFAULT;
1689                         len++;
1690                 }
1691                 *lenp = len;
1692                 *ppos += len;
1693         }
1694         return 0;
1695 }
1696
1697 /*
1698  *      Special case of dostring for the UTS structure. This has locks
1699  *      to observe. Should this be in kernel/sys.c ????
1700  */
1701
1702 static int proc_doutsstring(ctl_table *table, int write, struct file *filp,
1703                   void __user *buffer, size_t *lenp, loff_t *ppos)
1704 {
1705         int r;
1706
1707         if (!write) {
1708                 down_read(&uts_sem);
1709                 r=proc_dostring(table,0,filp,buffer,lenp, ppos);
1710                 up_read(&uts_sem);
1711         } else {
1712                 down_write(&uts_sem);
1713                 r=proc_dostring(table,1,filp,buffer,lenp, ppos);
1714                 up_write(&uts_sem);
1715         }
1716         return r;
1717 }
1718
1719 static int do_proc_dointvec_conv(int *negp, unsigned long *lvalp,
1720                                  int *valp,
1721                                  int write, void *data)
1722 {
1723         if (write) {
1724                 *valp = *negp ? -*lvalp : *lvalp;
1725         } else {
1726                 int val = *valp;
1727                 if (val < 0) {
1728                         *negp = -1;
1729                         *lvalp = (unsigned long)-val;
1730                 } else {
1731                         *negp = 0;
1732                         *lvalp = (unsigned long)val;
1733                 }
1734         }
1735         return 0;
1736 }
1737
1738 static int do_proc_dointvec(ctl_table *table, int write, struct file *filp,
1739                   void __user *buffer, size_t *lenp, loff_t *ppos,
1740                   int (*conv)(int *negp, unsigned long *lvalp, int *valp,
1741                               int write, void *data),
1742                   void *data)
1743 {
1744 #define TMPBUFLEN 21
1745         int *i, vleft, first=1, neg, val;
1746         unsigned long lval;
1747         size_t left, len;
1748
1749         char buf[TMPBUFLEN], *p;
1750         char __user *s = buffer;
1751
1752         if (!table->data || !table->maxlen || !*lenp ||
1753             (*ppos && !write)) {
1754                 *lenp = 0;
1755                 return 0;
1756         }
1757
1758         i = (int *) table->data;
1759         vleft = table->maxlen / sizeof(*i);
1760         left = *lenp;
1761
1762         if (!conv)
1763                 conv = do_proc_dointvec_conv;
1764
1765         for (; left && vleft--; i++, first=0) {
1766                 if (write) {
1767                         while (left) {
1768                                 char c;
1769                                 if (get_user(c, s))
1770                                         return -EFAULT;
1771                                 if (!isspace(c))
1772                                         break;
1773                                 left--;
1774                                 s++;
1775                         }
1776                         if (!left)
1777                                 break;
1778                         neg = 0;
1779                         len = left;
1780                         if (len > sizeof(buf) - 1)
1781                                 len = sizeof(buf) - 1;
1782                         if (copy_from_user(buf, s, len))
1783                                 return -EFAULT;
1784                         buf[len] = 0;
1785                         p = buf;
1786                         if (*p == '-' && left > 1) {
1787                                 neg = 1;
1788                                 left--, p++;
1789                         }
1790                         if (*p < '0' || *p > '9')
1791                                 break;
1792
1793                         lval = simple_strtoul(p, &p, 0);
1794
1795                         len = p-buf;
1796                         if ((len < left) && *p && !isspace(*p))
1797                                 break;
1798                         if (neg)
1799                                 val = -val;
1800                         s += len;
1801                         left -= len;
1802
1803                         if (conv(&neg, &lval, i, 1, data))
1804                                 break;
1805                 } else {
1806                         p = buf;
1807                         if (!first)
1808                                 *p++ = '\t';
1809
1810                         if (conv(&neg, &lval, i, 0, data))
1811                                 break;
1812
1813                         sprintf(p, "%s%lu", neg ? "-" : "", lval);
1814                         len = strlen(buf);
1815                         if (len > left)
1816                                 len = left;
1817                         if(copy_to_user(s, buf, len))
1818                                 return -EFAULT;
1819                         left -= len;
1820                         s += len;
1821                 }
1822         }
1823
1824         if (!write && !first && left) {
1825                 if(put_user('\n', s))
1826                         return -EFAULT;
1827                 left--, s++;
1828         }
1829         if (write) {
1830                 while (left) {
1831                         char c;
1832                         if (get_user(c, s++))
1833                                 return -EFAULT;
1834                         if (!isspace(c))
1835                                 break;
1836                         left--;
1837                 }
1838         }
1839         if (write && first)
1840                 return -EINVAL;
1841         *lenp -= left;
1842         *ppos += *lenp;
1843         return 0;
1844 #undef TMPBUFLEN
1845 }
1846
1847 /**
1848  * proc_dointvec - read a vector of integers
1849  * @table: the sysctl table
1850  * @write: %TRUE if this is a write to the sysctl file
1851  * @filp: the file structure
1852  * @buffer: the user buffer
1853  * @lenp: the size of the user buffer
1854  * @ppos: file position
1855  *
1856  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
1857  * values from/to the user buffer, treated as an ASCII string.
1858  *
1859  * Returns 0 on success.
1860  */
1861 int proc_dointvec(ctl_table *table, int write, struct file *filp,
1862                      void __user *buffer, size_t *lenp, loff_t *ppos)
1863 {
1864     return do_proc_dointvec(table,write,filp,buffer,lenp,ppos,
1865                             NULL,NULL);
1866 }
1867
1868 #define OP_SET  0
1869 #define OP_AND  1
1870 #define OP_OR   2
1871 #define OP_MAX  3
1872 #define OP_MIN  4
1873
1874 static int do_proc_dointvec_bset_conv(int *negp, unsigned long *lvalp,
1875                                       int *valp,
1876                                       int write, void *data)
1877 {
1878         int op = *(int *)data;
1879         if (write) {
1880                 int val = *negp ? -*lvalp : *lvalp;
1881                 switch(op) {
1882                 case OP_SET:    *valp = val; break;
1883                 case OP_AND:    *valp &= val; break;
1884                 case OP_OR:     *valp |= val; break;
1885                 case OP_MAX:    if(*valp < val)
1886                                         *valp = val;
1887                                 break;
1888                 case OP_MIN:    if(*valp > val)
1889                                 *valp = val;
1890                                 break;
1891                 }
1892         } else {
1893                 int val = *valp;
1894                 if (val < 0) {
1895                         *negp = -1;
1896                         *lvalp = (unsigned long)-val;
1897                 } else {
1898                         *negp = 0;
1899                         *lvalp = (unsigned long)val;
1900                 }
1901         }
1902         return 0;
1903 }
1904
1905 /*
1906  *      init may raise the set.
1907  */
1908
1909 int proc_dointvec_bset(ctl_table *table, int write, struct file *filp,
1910                         void __user *buffer, size_t *lenp, loff_t *ppos)
1911 {
1912         int op;
1913
1914         if (!capable(CAP_SYS_MODULE)) {
1915                 return -EPERM;
1916         }
1917
1918         op = (current->pid == 1) ? OP_SET : OP_AND;
1919         return do_proc_dointvec(table,write,filp,buffer,lenp,ppos,
1920                                 do_proc_dointvec_bset_conv,&op);
1921 }
1922
1923 struct do_proc_dointvec_minmax_conv_param {
1924         int *min;
1925         int *max;
1926 };
1927
1928 static int do_proc_dointvec_minmax_conv(int *negp, unsigned long *lvalp,
1929                                         int *valp,
1930                                         int write, void *data)
1931 {
1932         struct do_proc_dointvec_minmax_conv_param *param = data;
1933         if (write) {
1934                 int val = *negp ? -*lvalp : *lvalp;
1935                 if ((param->min && *param->min > val) ||
1936                     (param->max && *param->max < val))
1937                         return -EINVAL;
1938                 *valp = val;
1939         } else {
1940                 int val = *valp;
1941                 if (val < 0) {
1942                         *negp = -1;
1943                         *lvalp = (unsigned long)-val;
1944                 } else {
1945                         *negp = 0;
1946                         *lvalp = (unsigned long)val;
1947                 }
1948         }
1949         return 0;
1950 }
1951
1952 /**
1953  * proc_dointvec_minmax - read a vector of integers with min/max values
1954  * @table: the sysctl table
1955  * @write: %TRUE if this is a write to the sysctl file
1956  * @filp: the file structure
1957  * @buffer: the user buffer
1958  * @lenp: the size of the user buffer
1959  * @ppos: file position
1960  *
1961  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
1962  * values from/to the user buffer, treated as an ASCII string.
1963  *
1964  * This routine will ensure the values are within the range specified by
1965  * table->extra1 (min) and table->extra2 (max).
1966  *
1967  * Returns 0 on success.
1968  */
1969 int proc_dointvec_minmax(ctl_table *table, int write, struct file *filp,
1970                   void __user *buffer, size_t *lenp, loff_t *ppos)
1971 {
1972         struct do_proc_dointvec_minmax_conv_param param = {
1973                 .min = (int *) table->extra1,
1974                 .max = (int *) table->extra2,
1975         };
1976         return do_proc_dointvec(table, write, filp, buffer, lenp, ppos,
1977                                 do_proc_dointvec_minmax_conv, &param);
1978 }
1979
1980 static int do_proc_doulongvec_minmax(ctl_table *table, int write,
1981                                      struct file *filp,
1982                                      void __user *buffer,
1983                                      size_t *lenp, loff_t *ppos,
1984                                      unsigned long convmul,
1985                                      unsigned long convdiv)
1986 {
1987 #define TMPBUFLEN 21
1988         unsigned long *i, *min, *max, val;
1989         int vleft, first=1, neg;
1990         size_t len, left;
1991         char buf[TMPBUFLEN], *p;
1992         char __user *s = buffer;
1993
1994         if (!table->data || !table->maxlen || !*lenp ||
1995             (*ppos && !write)) {
1996                 *lenp = 0;
1997                 return 0;
1998         }
1999
2000         i = (unsigned long *) table->data;
2001         min = (unsigned long *) table->extra1;
2002         max = (unsigned long *) table->extra2;
2003         vleft = table->maxlen / sizeof(unsigned long);
2004         left = *lenp;
2005
2006         for (; left && vleft--; i++, min++, max++, first=0) {
2007                 if (write) {
2008                         while (left) {
2009                                 char c;
2010                                 if (get_user(c, s))
2011                                         return -EFAULT;
2012                                 if (!isspace(c))
2013                                         break;
2014                                 left--;
2015                                 s++;
2016                         }
2017                         if (!left)
2018                                 break;
2019                         neg = 0;
2020                         len = left;
2021                         if (len > TMPBUFLEN-1)
2022                                 len = TMPBUFLEN-1;
2023                         if (copy_from_user(buf, s, len))
2024                                 return -EFAULT;
2025                         buf[len] = 0;
2026                         p = buf;
2027                         if (*p == '-' && left > 1) {
2028                                 neg = 1;
2029                                 left--, p++;
2030                         }
2031                         if (*p < '0' || *p > '9')
2032                                 break;
2033                         val = simple_strtoul(p, &p, 0) * convmul / convdiv ;
2034                         len = p-buf;
2035                         if ((len < left) && *p && !isspace(*p))
2036                                 break;
2037                         if (neg)
2038                                 val = -val;
2039                         s += len;
2040                         left -= len;
2041
2042                         if(neg)
2043                                 continue;
2044                         if ((min && val < *min) || (max && val > *max))
2045                                 continue;
2046                         *i = val;
2047                 } else {
2048                         p = buf;
2049                         if (!first)
2050                                 *p++ = '\t';
2051                         sprintf(p, "%lu", convdiv * (*i) / convmul);
2052                         len = strlen(buf);
2053                         if (len > left)
2054                                 len = left;
2055                         if(copy_to_user(s, buf, len))
2056                                 return -EFAULT;
2057                         left -= len;
2058                         s += len;
2059                 }
2060         }
2061
2062         if (!write && !first && left) {
2063                 if(put_user('\n', s))
2064                         return -EFAULT;
2065                 left--, s++;
2066         }
2067         if (write) {
2068                 while (left) {
2069                         char c;
2070                         if (get_user(c, s++))
2071                                 return -EFAULT;
2072                         if (!isspace(c))
2073                                 break;
2074                         left--;
2075                 }
2076         }
2077         if (write && first)
2078                 return -EINVAL;
2079         *lenp -= left;
2080         *ppos += *lenp;
2081         return 0;
2082 #undef TMPBUFLEN
2083 }
2084
2085 /**
2086  * proc_doulongvec_minmax - read a vector of long integers with min/max values
2087  * @table: the sysctl table
2088  * @write: %TRUE if this is a write to the sysctl file
2089  * @filp: the file structure
2090  * @buffer: the user buffer
2091  * @lenp: the size of the user buffer
2092  * @ppos: file position
2093  *
2094  * Reads/writes up to table->maxlen/sizeof(unsigned long) unsigned long
2095  * values from/to the user buffer, treated as an ASCII string.
2096  *
2097  * This routine will ensure the values are within the range specified by
2098  * table->extra1 (min) and table->extra2 (max).
2099  *
2100  * Returns 0 on success.
2101  */
2102 int proc_doulongvec_minmax(ctl_table *table, int write, struct file *filp,
2103                            void __user *buffer, size_t *lenp, loff_t *ppos)
2104 {
2105     return do_proc_doulongvec_minmax(table, write, filp, buffer, lenp, ppos, 1l, 1l);
2106 }
2107
2108 /**
2109  * proc_doulongvec_ms_jiffies_minmax - read a vector of millisecond values with min/max values
2110  * @table: the sysctl table
2111  * @write: %TRUE if this is a write to the sysctl file
2112  * @filp: the file structure
2113  * @buffer: the user buffer
2114  * @lenp: the size of the user buffer
2115  * @ppos: file position
2116  *
2117  * Reads/writes up to table->maxlen/sizeof(unsigned long) unsigned long
2118  * values from/to the user buffer, treated as an ASCII string. The values
2119  * are treated as milliseconds, and converted to jiffies when they are stored.
2120  *
2121  * This routine will ensure the values are within the range specified by
2122  * table->extra1 (min) and table->extra2 (max).
2123  *
2124  * Returns 0 on success.
2125  */
2126 int proc_doulongvec_ms_jiffies_minmax(ctl_table *table, int write,
2127                                       struct file *filp,
2128                                       void __user *buffer,
2129                                       size_t *lenp, loff_t *ppos)
2130 {
2131     return do_proc_doulongvec_minmax(table, write, filp, buffer,
2132                                      lenp, ppos, HZ, 1000l);
2133 }
2134
2135
2136 static int do_proc_dointvec_jiffies_conv(int *negp, unsigned long *lvalp,
2137                                          int *valp,
2138                                          int write, void *data)
2139 {
2140         if (write) {
2141                 if (*lvalp > LONG_MAX / HZ)
2142                         return 1;
2143                 *valp = *negp ? -(*lvalp*HZ) : (*lvalp*HZ);
2144         } else {
2145                 int val = *valp;
2146                 unsigned long lval;
2147                 if (val < 0) {
2148                         *negp = -1;
2149                         lval = (unsigned long)-val;
2150                 } else {
2151                         *negp = 0;
2152                         lval = (unsigned long)val;
2153                 }
2154                 *lvalp = lval / HZ;
2155         }
2156         return 0;
2157 }
2158
2159 static int do_proc_dointvec_userhz_jiffies_conv(int *negp, unsigned long *lvalp,
2160                                                 int *valp,
2161                                                 int write, void *data)
2162 {
2163         if (write) {
2164                 if (USER_HZ < HZ && *lvalp > (LONG_MAX / HZ) * USER_HZ)
2165                         return 1;
2166                 *valp = clock_t_to_jiffies(*negp ? -*lvalp : *lvalp);
2167         } else {
2168                 int val = *valp;
2169                 unsigned long lval;
2170                 if (val < 0) {
2171                         *negp = -1;
2172                         lval = (unsigned long)-val;
2173                 } else {
2174                         *negp = 0;
2175                         lval = (unsigned long)val;
2176                 }
2177                 *lvalp = jiffies_to_clock_t(lval);
2178         }
2179         return 0;
2180 }
2181
2182 static int do_proc_dointvec_ms_jiffies_conv(int *negp, unsigned long *lvalp,
2183                                             int *valp,
2184                                             int write, void *data)
2185 {
2186         if (write) {
2187                 *valp = msecs_to_jiffies(*negp ? -*lvalp : *lvalp);
2188         } else {
2189                 int val = *valp;
2190                 unsigned long lval;
2191                 if (val < 0) {
2192                         *negp = -1;
2193                         lval = (unsigned long)-val;
2194                 } else {
2195                         *negp = 0;
2196                         lval = (unsigned long)val;
2197                 }
2198                 *lvalp = jiffies_to_msecs(lval);
2199         }
2200         return 0;
2201 }
2202
2203 /**
2204  * proc_dointvec_jiffies - read a vector of integers as seconds
2205  * @table: the sysctl table
2206  * @write: %TRUE if this is a write to the sysctl file
2207  * @filp: the file structure
2208  * @buffer: the user buffer
2209  * @lenp: the size of the user buffer
2210  * @ppos: file position
2211  *
2212  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2213  * values from/to the user buffer, treated as an ASCII string.
2214  * The values read are assumed to be in seconds, and are converted into
2215  * jiffies.
2216  *
2217  * Returns 0 on success.
2218  */
2219 int proc_dointvec_jiffies(ctl_table *table, int write, struct file *filp,
2220                           void __user *buffer, size_t *lenp, loff_t *ppos)
2221 {
2222     return do_proc_dointvec(table,write,filp,buffer,lenp,ppos,
2223                             do_proc_dointvec_jiffies_conv,NULL);
2224 }
2225
2226 /**
2227  * proc_dointvec_userhz_jiffies - read a vector of integers as 1/USER_HZ seconds
2228  * @table: the sysctl table
2229  * @write: %TRUE if this is a write to the sysctl file
2230  * @filp: the file structure
2231  * @buffer: the user buffer
2232  * @lenp: the size of the user buffer
2233  * @ppos: pointer to the file position
2234  *
2235  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2236  * values from/to the user buffer, treated as an ASCII string.
2237  * The values read are assumed to be in 1/USER_HZ seconds, and
2238  * are converted into jiffies.
2239  *
2240  * Returns 0 on success.
2241  */
2242 int proc_dointvec_userhz_jiffies(ctl_table *table, int write, struct file *filp,
2243                                  void __user *buffer, size_t *lenp, loff_t *ppos)
2244 {
2245     return do_proc_dointvec(table,write,filp,buffer,lenp,ppos,
2246                             do_proc_dointvec_userhz_jiffies_conv,NULL);
2247 }
2248
2249 /**
2250  * proc_dointvec_ms_jiffies - read a vector of integers as 1 milliseconds
2251  * @table: the sysctl table
2252  * @write: %TRUE if this is a write to the sysctl file
2253  * @filp: the file structure
2254  * @buffer: the user buffer
2255  * @lenp: the size of the user buffer
2256  * @ppos: file position
2257  * @ppos: the current position in the file
2258  *
2259  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2260  * values from/to the user buffer, treated as an ASCII string.
2261  * The values read are assumed to be in 1/1000 seconds, and
2262  * are converted into jiffies.
2263  *
2264  * Returns 0 on success.
2265  */
2266 int proc_dointvec_ms_jiffies(ctl_table *table, int write, struct file *filp,
2267                              void __user *buffer, size_t *lenp, loff_t *ppos)
2268 {
2269         return do_proc_dointvec(table, write, filp, buffer, lenp, ppos,
2270                                 do_proc_dointvec_ms_jiffies_conv, NULL);
2271 }
2272
2273 #else /* CONFIG_PROC_FS */
2274
2275 int proc_dostring(ctl_table *table, int write, struct file *filp,
2276                   void __user *buffer, size_t *lenp, loff_t *ppos)
2277 {
2278         return -ENOSYS;
2279 }
2280
2281 static int proc_doutsstring(ctl_table *table, int write, struct file *filp,
2282                             void __user *buffer, size_t *lenp, loff_t *ppos)
2283 {
2284         return -ENOSYS;
2285 }
2286
2287 int proc_dointvec(ctl_table *table, int write, struct file *filp,
2288                   void __user *buffer, size_t *lenp, loff_t *ppos)
2289 {
2290         return -ENOSYS;
2291 }
2292
2293 int proc_dointvec_bset(ctl_table *table, int write, struct file *filp,
2294                         void __user *buffer, size_t *lenp, loff_t *ppos)
2295 {
2296         return -ENOSYS;
2297 }
2298
2299 int proc_dointvec_minmax(ctl_table *table, int write, struct file *filp,
2300                     void __user *buffer, size_t *lenp, loff_t *ppos)
2301 {
2302         return -ENOSYS;
2303 }
2304
2305 int proc_dointvec_jiffies(ctl_table *table, int write, struct file *filp,
2306                     void __user *buffer, size_t *lenp, loff_t *ppos)
2307 {
2308         return -ENOSYS;
2309 }
2310
2311 int proc_dointvec_userhz_jiffies(ctl_table *table, int write, struct file *filp,
2312                     void __user *buffer, size_t *lenp, loff_t *ppos)
2313 {
2314         return -ENOSYS;
2315 }
2316
2317 int proc_dointvec_ms_jiffies(ctl_table *table, int write, struct file *filp,
2318                              void __user *buffer, size_t *lenp, loff_t *ppos)
2319 {
2320         return -ENOSYS;
2321 }
2322
2323 int proc_doulongvec_minmax(ctl_table *table, int write, struct file *filp,
2324                     void __user *buffer, size_t *lenp, loff_t *ppos)
2325 {
2326         return -ENOSYS;
2327 }
2328
2329 int proc_doulongvec_ms_jiffies_minmax(ctl_table *table, int write,
2330                                       struct file *filp,
2331                                       void __user *buffer,
2332                                       size_t *lenp, loff_t *ppos)
2333 {
2334     return -ENOSYS;
2335 }
2336
2337
2338 #endif /* CONFIG_PROC_FS */
2339
2340
2341 #ifdef CONFIG_SYSCTL_SYSCALL
2342 /*
2343  * General sysctl support routines
2344  */
2345
2346 /* The generic string strategy routine: */
2347 int sysctl_string(ctl_table *table, int __user *name, int nlen,
2348                   void __user *oldval, size_t __user *oldlenp,
2349                   void __user *newval, size_t newlen, void **context)
2350 {
2351         if (!table->data || !table->maxlen)
2352                 return -ENOTDIR;
2353
2354         if (oldval && oldlenp) {
2355                 size_t bufsize;
2356                 if (get_user(bufsize, oldlenp))
2357                         return -EFAULT;
2358                 if (bufsize) {
2359                         size_t len = strlen(table->data), copied;
2360
2361                         /* This shouldn't trigger for a well-formed sysctl */
2362                         if (len > table->maxlen)
2363                                 len = table->maxlen;
2364
2365                         /* Copy up to a max of bufsize-1 bytes of the string */
2366                         copied = (len >= bufsize) ? bufsize - 1 : len;
2367
2368                         if (copy_to_user(oldval, table->data, copied) ||
2369                             put_user(0, (char __user *)(oldval + copied)))
2370                                 return -EFAULT;
2371                         if (put_user(len, oldlenp))
2372                                 return -EFAULT;
2373                 }
2374         }
2375         if (newval && newlen) {
2376                 size_t len = newlen;
2377                 if (len > table->maxlen)
2378                         len = table->maxlen;
2379                 if(copy_from_user(table->data, newval, len))
2380                         return -EFAULT;
2381                 if (len == table->maxlen)
2382                         len--;
2383                 ((char *) table->data)[len] = 0;
2384         }
2385         return 1;
2386 }
2387
2388 /*
2389  * This function makes sure that all of the integers in the vector
2390  * are between the minimum and maximum values given in the arrays
2391  * table->extra1 and table->extra2, respectively.
2392  */
2393 int sysctl_intvec(ctl_table *table, int __user *name, int nlen,
2394                 void __user *oldval, size_t __user *oldlenp,
2395                 void __user *newval, size_t newlen, void **context)
2396 {
2397
2398         if (newval && newlen) {
2399                 int __user *vec = (int __user *) newval;
2400                 int *min = (int *) table->extra1;
2401                 int *max = (int *) table->extra2;
2402                 size_t length;
2403                 int i;
2404
2405                 if (newlen % sizeof(int) != 0)
2406                         return -EINVAL;
2407
2408                 if (!table->extra1 && !table->extra2)
2409                         return 0;
2410
2411                 if (newlen > table->maxlen)
2412                         newlen = table->maxlen;
2413                 length = newlen / sizeof(int);
2414
2415                 for (i = 0; i < length; i++) {
2416                         int value;
2417                         if (get_user(value, vec + i))
2418                                 return -EFAULT;
2419                         if (min && value < min[i])
2420                                 return -EINVAL;
2421                         if (max && value > max[i])
2422                                 return -EINVAL;
2423                 }
2424         }
2425         return 0;
2426 }
2427
2428 /* Strategy function to convert jiffies to seconds */
2429 int sysctl_jiffies(ctl_table *table, int __user *name, int nlen,
2430                 void __user *oldval, size_t __user *oldlenp,
2431                 void __user *newval, size_t newlen, void **context)
2432 {
2433         if (oldval) {
2434                 size_t olen;
2435                 if (oldlenp) {
2436                         if (get_user(olen, oldlenp))
2437                                 return -EFAULT;
2438                         if (olen!=sizeof(int))
2439                                 return -EINVAL;
2440                 }
2441                 if (put_user(*(int *)(table->data)/HZ, (int __user *)oldval) ||
2442                     (oldlenp && put_user(sizeof(int),oldlenp)))
2443                         return -EFAULT;
2444         }
2445         if (newval && newlen) {
2446                 int new;
2447                 if (newlen != sizeof(int))
2448                         return -EINVAL;
2449                 if (get_user(new, (int __user *)newval))
2450                         return -EFAULT;
2451                 *(int *)(table->data) = new*HZ;
2452         }
2453         return 1;
2454 }
2455
2456 /* Strategy function to convert jiffies to seconds */
2457 int sysctl_ms_jiffies(ctl_table *table, int __user *name, int nlen,
2458                 void __user *oldval, size_t __user *oldlenp,
2459                 void __user *newval, size_t newlen, void **context)
2460 {
2461         if (oldval) {
2462                 size_t olen;
2463                 if (oldlenp) {
2464                         if (get_user(olen, oldlenp))
2465                                 return -EFAULT;
2466                         if (olen!=sizeof(int))
2467                                 return -EINVAL;
2468                 }
2469                 if (put_user(jiffies_to_msecs(*(int *)(table->data)), (int __user *)oldval) ||
2470                     (oldlenp && put_user(sizeof(int),oldlenp)))
2471                         return -EFAULT;
2472         }
2473         if (newval && newlen) {
2474                 int new;
2475                 if (newlen != sizeof(int))
2476                         return -EINVAL;
2477                 if (get_user(new, (int __user *)newval))
2478                         return -EFAULT;
2479                 *(int *)(table->data) = msecs_to_jiffies(new);
2480         }
2481         return 1;
2482 }
2483
2484 #else /* CONFIG_SYSCTL_SYSCALL */
2485
2486
2487 asmlinkage long sys_sysctl(struct __sysctl_args __user *args)
2488 {
2489         static int msg_count;
2490
2491         if (msg_count < 5) {
2492                 msg_count++;
2493                 printk(KERN_INFO
2494                         "warning: process `%s' used the removed sysctl "
2495                         "system call\n", current->comm);
2496         }
2497         return -ENOSYS;
2498 }
2499
2500 int sysctl_string(ctl_table *table, int __user *name, int nlen,
2501                   void __user *oldval, size_t __user *oldlenp,
2502                   void __user *newval, size_t newlen, void **context)
2503 {
2504         return -ENOSYS;
2505 }
2506
2507 int sysctl_intvec(ctl_table *table, int __user *name, int nlen,
2508                 void __user *oldval, size_t __user *oldlenp,
2509                 void __user *newval, size_t newlen, void **context)
2510 {
2511         return -ENOSYS;
2512 }
2513
2514 int sysctl_jiffies(ctl_table *table, int __user *name, int nlen,
2515                 void __user *oldval, size_t __user *oldlenp,
2516                 void __user *newval, size_t newlen, void **context)
2517 {
2518         return -ENOSYS;
2519 }
2520
2521 int sysctl_ms_jiffies(ctl_table *table, int __user *name, int nlen,
2522                 void __user *oldval, size_t __user *oldlenp,
2523                 void __user *newval, size_t newlen, void **context)
2524 {
2525         return -ENOSYS;
2526 }
2527
2528 #endif /* CONFIG_SYSCTL_SYSCALL */
2529
2530 /*
2531  * No sense putting this after each symbol definition, twice,
2532  * exception granted :-)
2533  */
2534 EXPORT_SYMBOL(proc_dointvec);
2535 EXPORT_SYMBOL(proc_dointvec_jiffies);
2536 EXPORT_SYMBOL(proc_dointvec_minmax);
2537 EXPORT_SYMBOL(proc_dointvec_userhz_jiffies);
2538 EXPORT_SYMBOL(proc_dointvec_ms_jiffies);
2539 EXPORT_SYMBOL(proc_dostring);
2540 EXPORT_SYMBOL(proc_doulongvec_minmax);
2541 EXPORT_SYMBOL(proc_doulongvec_ms_jiffies_minmax);
2542 EXPORT_SYMBOL(register_sysctl_table);
2543 EXPORT_SYMBOL(sysctl_intvec);
2544 EXPORT_SYMBOL(sysctl_jiffies);
2545 EXPORT_SYMBOL(sysctl_ms_jiffies);
2546 EXPORT_SYMBOL(sysctl_string);
2547 EXPORT_SYMBOL(unregister_sysctl_table);