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 extern int proc_unknown_nmi_panic(ctl_table *, int, struct file *,
  80                                   void __user *, size_t *, loff_t *);
  81 #endif
  82
  83 /* this is needed for the proc_dointvec_minmax for [fs_]overflow UID and GID */
  84 static int maxolduid = 65535;
  85 static int minolduid;
  86 static int min_percpu_pagelist_fract = 8;
  87
  88 static int ngroups_max = NGROUPS_MAX;
  89
  90 #ifdef CONFIG_KMOD
  91 extern char modprobe_path[];
  92 #endif
  93 #ifdef CONFIG_CHR_DEV_SG
  94 extern int sg_big_buff;
  95 #endif
  96 #ifdef CONFIG_SYSVIPC
  97 extern size_t shm_ctlmax;
  98 extern size_t shm_ctlall;
  99 extern int shm_ctlmni;
 100 extern int msg_ctlmax;
 101 extern int msg_ctlmnb;
 102 extern int msg_ctlmni;
 103 extern int sem_ctls[];
 104 #endif
 105
 106 #ifdef __sparc__
 107 extern char reboot_command [];
 108 extern int stop_a_enabled;
 109 extern int scons_pwroff;
 110 #endif
 111
 112 #ifdef __hppa__
 113 extern int pwrsw_enabled;
 114 extern int unaligned_enabled;
 115 #endif
 116
 117 #ifdef CONFIG_S390
 118 #ifdef CONFIG_MATHEMU
 119 extern int sysctl_ieee_emulation_warnings;
 120 #endif
 121 extern int sysctl_userprocess_debug;
 122 extern int spin_retry;
 123 #endif
 124
 125 extern int sysctl_hz_timer;
 126
 127 #ifdef CONFIG_BSD_PROCESS_ACCT
 128 extern int acct_parm[];
 129 #endif
 130
 131 #ifdef CONFIG_IA64
 132 extern int no_unaligned_warning;
 133 #endif
 134
 135 #ifdef CONFIG_RT_MUTEXES
 136 extern int max_lock_depth;
 137 #endif
 138
 139 static int parse_table(int __user *, int, void __user *, size_t __user *, void __user *, size_t,
 140                        ctl_table *, void **);
 141 static int proc_doutsstring(ctl_table *table, int write, struct file *filp,
 142                   void __user *buffer, size_t *lenp, loff_t *ppos);
 143
 144 static ctl_table root_table[];
 145 static struct ctl_table_header root_table_header =
 146         { root_table, LIST_HEAD_INIT(root_table_header.ctl_entry) };
 147
 148 static ctl_table kern_table[];
 149 static ctl_table vm_table[];
 150 static ctl_table fs_table[];
 151 static ctl_table debug_table[];
 152 static ctl_table dev_table[];
 153 extern ctl_table random_table[];
 154 #ifdef CONFIG_UNIX98_PTYS
 155 extern ctl_table pty_table[];
 156 #endif
 157 #ifdef CONFIG_INOTIFY_USER
 158 extern ctl_table inotify_table[];
 159 #endif
 160
 161 #ifdef HAVE_ARCH_PICK_MMAP_LAYOUT
 162 int sysctl_legacy_va_layout;
 163 #endif
 164
 165 /* /proc declarations: */
 166
 167 #ifdef CONFIG_PROC_FS
 168
 169 static ssize_t proc_readsys(struct file *, char __user *, size_t, loff_t *);
 170 static ssize_t proc_writesys(struct file *, const char __user *, size_t, loff_t *);
 171 static int proc_opensys(struct inode *, struct file *);
 172
 173 struct file_operations proc_sys_file_operations = {
 174         .open           = proc_opensys,
 175         .read           = proc_readsys,
 176         .write          = proc_writesys,
 177 };
 178
 179 extern struct proc_dir_entry *proc_sys_root;
 180
 181 static void register_proc_table(ctl_table *, struct proc_dir_entry *, void *);
 182 static void unregister_proc_table(ctl_table *, struct proc_dir_entry *);
 183 #endif
 184
 185 /* The default sysctl tables: */
 186
 187 static ctl_table root_table[] = {
 188         {
 189                 .ctl_name       = CTL_KERN,
 190                 .procname       = "kernel",
 191                 .mode           = 0555,
 192                 .child          = kern_table,
 193         },
 194         {
 195                 .ctl_name       = CTL_VM,
 196                 .procname       = "vm",
 197                 .mode           = 0555,
 198                 .child          = vm_table,
 199         },
 200 #ifdef CONFIG_NET
 201         {
 202                 .ctl_name       = CTL_NET,
 203                 .procname       = "net",
 204                 .mode           = 0555,
 205                 .child          = net_table,
 206         },
 207 #endif
 208         {
 209                 .ctl_name       = CTL_FS,
 210                 .procname       = "fs",
 211                 .mode           = 0555,
 212                 .child          = fs_table,
 213         },
 214         {
 215                 .ctl_name       = CTL_DEBUG,
 216                 .procname       = "debug",
 217                 .mode           = 0555,
 218                 .child          = debug_table,
 219         },
 220         {
 221                 .ctl_name       = CTL_DEV,
 222                 .procname       = "dev",
 223                 .mode           = 0555,
 224                 .child          = dev_table,
 225         },
 226
 227         { .ctl_name = 0 }
 228 };
 229
 230 static ctl_table kern_table[] = {
 231         {
 232                 .ctl_name       = KERN_OSTYPE,
 233                 .procname       = "ostype",
 234                 .data           = system_utsname.sysname,
 235                 .maxlen         = sizeof(system_utsname.sysname),
 236                 .mode           = 0444,
 237                 .proc_handler   = &proc_doutsstring,
 238                 .strategy       = &sysctl_string,
 239         },
 240         {
 241                 .ctl_name       = KERN_OSRELEASE,
 242                 .procname       = "osrelease",
 243                 .data           = system_utsname.release,
 244                 .maxlen         = sizeof(system_utsname.release),
 245                 .mode           = 0444,
 246                 .proc_handler   = &proc_doutsstring,
 247                 .strategy       = &sysctl_string,
 248         },
 249         {
 250                 .ctl_name       = KERN_VERSION,
 251                 .procname       = "version",
 252                 .data           = system_utsname.version,
 253                 .maxlen         = sizeof(system_utsname.version),
 254                 .mode           = 0444,
 255                 .proc_handler   = &proc_doutsstring,
 256                 .strategy       = &sysctl_string,
 257         },
 258         {
 259                 .ctl_name       = KERN_NODENAME,
 260                 .procname       = "hostname",
 261                 .data           = system_utsname.nodename,
 262                 .maxlen         = sizeof(system_utsname.nodename),
 263                 .mode           = 0644,
 264                 .proc_handler   = &proc_doutsstring,
 265                 .strategy       = &sysctl_string,
 266         },
 267         {
 268                 .ctl_name       = KERN_DOMAINNAME,
 269                 .procname       = "domainname",
 270                 .data           = system_utsname.domainname,
 271                 .maxlen         = sizeof(system_utsname.domainname),
 272                 .mode           = 0644,
 273                 .proc_handler   = &proc_doutsstring,
 274                 .strategy       = &sysctl_string,
 275         },
 276         {
 277                 .ctl_name       = KERN_PANIC,
 278                 .procname       = "panic",
 279                 .data           = &panic_timeout,
 280                 .maxlen         = sizeof(int),
 281                 .mode           = 0644,
 282                 .proc_handler   = &proc_dointvec,
 283         },
 284         {
 285                 .ctl_name       = KERN_CORE_USES_PID,
 286                 .procname       = "core_uses_pid",
 287                 .data           = &core_uses_pid,
 288                 .maxlen         = sizeof(int),
 289                 .mode           = 0644,
 290                 .proc_handler   = &proc_dointvec,
 291         },
 292         {
 293                 .ctl_name       = KERN_CORE_PATTERN,
 294                 .procname       = "core_pattern",
 295                 .data           = core_pattern,
 296                 .maxlen         = 64,
 297                 .mode           = 0644,
 298                 .proc_handler   = &proc_dostring,
 299                 .strategy       = &sysctl_string,
 300         },
 301         {
 302                 .ctl_name       = KERN_TAINTED,
 303                 .procname       = "tainted",
 304                 .data           = &tainted,
 305                 .maxlen         = sizeof(int),
 306                 .mode           = 0444,
 307                 .proc_handler   = &proc_dointvec,
 308         },
 309         {
 310                 .ctl_name       = KERN_CAP_BSET,
 311                 .procname       = "cap-bound",
 312                 .data           = &cap_bset,
 313                 .maxlen         = sizeof(kernel_cap_t),
 314                 .mode           = 0600,
 315                 .proc_handler   = &proc_dointvec_bset,
 316         },
 317 #ifdef CONFIG_BLK_DEV_INITRD
 318         {
 319                 .ctl_name       = KERN_REALROOTDEV,
 320                 .procname       = "real-root-dev",
 321                 .data           = &real_root_dev,
 322                 .maxlen         = sizeof(int),
 323                 .mode           = 0644,
 324                 .proc_handler   = &proc_dointvec,
 325         },
 326 #endif
 327 #ifdef __sparc__
 328         {
 329                 .ctl_name       = KERN_SPARC_REBOOT,
 330                 .procname       = "reboot-cmd",
 331                 .data           = reboot_command,
 332                 .maxlen         = 256,
 333                 .mode           = 0644,
 334                 .proc_handler   = &proc_dostring,
 335                 .strategy       = &sysctl_string,
 336         },
 337         {
 338                 .ctl_name       = KERN_SPARC_STOP_A,
 339                 .procname       = "stop-a",
 340                 .data           = &stop_a_enabled,
 341                 .maxlen         = sizeof (int),
 342                 .mode           = 0644,
 343                 .proc_handler   = &proc_dointvec,
 344         },
 345         {
 346                 .ctl_name       = KERN_SPARC_SCONS_PWROFF,
 347                 .procname       = "scons-poweroff",
 348                 .data           = &scons_pwroff,
 349                 .maxlen         = sizeof (int),
 350                 .mode           = 0644,
 351                 .proc_handler   = &proc_dointvec,
 352         },
 353 #endif
 354 #ifdef __hppa__
 355         {
 356                 .ctl_name       = KERN_HPPA_PWRSW,
 357                 .procname       = "soft-power",
 358                 .data           = &pwrsw_enabled,
 359                 .maxlen         = sizeof (int),
 360                 .mode           = 0644,
 361                 .proc_handler   = &proc_dointvec,
 362         },
 363         {
 364                 .ctl_name       = KERN_HPPA_UNALIGNED,
 365                 .procname       = "unaligned-trap",
 366                 .data           = &unaligned_enabled,
 367                 .maxlen         = sizeof (int),
 368                 .mode           = 0644,
 369                 .proc_handler   = &proc_dointvec,
 370         },
 371 #endif
 372         {
 373                 .ctl_name       = KERN_CTLALTDEL,
 374                 .procname       = "ctrl-alt-del",
 375                 .data           = &C_A_D,
 376                 .maxlen         = sizeof(int),
 377                 .mode           = 0644,
 378                 .proc_handler   = &proc_dointvec,
 379         },
 380         {
 381                 .ctl_name       = KERN_PRINTK,
 382                 .procname       = "printk",
 383                 .data           = &console_loglevel,
 384                 .maxlen         = 4*sizeof(int),
 385                 .mode           = 0644,
 386                 .proc_handler   = &proc_dointvec,
 387         },
 388 #ifdef CONFIG_KMOD
 389         {
 390                 .ctl_name       = KERN_MODPROBE,
 391                 .procname       = "modprobe",
 392                 .data           = &modprobe_path,
 393                 .maxlen         = KMOD_PATH_LEN,
 394                 .mode           = 0644,
 395                 .proc_handler   = &proc_dostring,
 396                 .strategy       = &sysctl_string,
 397         },
 398 #endif
 399 #if defined(CONFIG_HOTPLUG) && defined(CONFIG_NET)
 400         {
 401                 .ctl_name       = KERN_HOTPLUG,
 402                 .procname       = "hotplug",
 403                 .data           = &uevent_helper,
 404                 .maxlen         = UEVENT_HELPER_PATH_LEN,
 405                 .mode           = 0644,
 406                 .proc_handler   = &proc_dostring,
 407                 .strategy       = &sysctl_string,
 408         },
 409 #endif
 410 #ifdef CONFIG_CHR_DEV_SG
 411         {
 412                 .ctl_name       = KERN_SG_BIG_BUFF,
 413                 .procname       = "sg-big-buff",
 414                 .data           = &sg_big_buff,
 415                 .maxlen         = sizeof (int),
 416                 .mode           = 0444,
 417                 .proc_handler   = &proc_dointvec,
 418         },
 419 #endif
 420 #ifdef CONFIG_BSD_PROCESS_ACCT
 421         {
 422                 .ctl_name       = KERN_ACCT,
 423                 .procname       = "acct",
 424                 .data           = &acct_parm,
 425                 .maxlen         = 3*sizeof(int),
 426                 .mode           = 0644,
 427                 .proc_handler   = &proc_dointvec,
 428         },
 429 #endif
 430 #ifdef CONFIG_SYSVIPC
 431         {
 432                 .ctl_name       = KERN_SHMMAX,
 433                 .procname       = "shmmax",
 434                 .data           = &shm_ctlmax,
 435                 .maxlen         = sizeof (size_t),
 436                 .mode           = 0644,
 437                 .proc_handler   = &proc_doulongvec_minmax,
 438         },
 439         {
 440                 .ctl_name       = KERN_SHMALL,
 441                 .procname       = "shmall",
 442                 .data           = &shm_ctlall,
 443                 .maxlen         = sizeof (size_t),
 444                 .mode           = 0644,
 445                 .proc_handler   = &proc_doulongvec_minmax,
 446         },
 447         {
 448                 .ctl_name       = KERN_SHMMNI,
 449                 .procname       = "shmmni",
 450                 .data           = &shm_ctlmni,
 451                 .maxlen         = sizeof (int),
 452                 .mode           = 0644,
 453                 .proc_handler   = &proc_dointvec,
 454         },
 455         {
 456                 .ctl_name       = KERN_MSGMAX,
 457                 .procname       = "msgmax",
 458                 .data           = &msg_ctlmax,
 459                 .maxlen         = sizeof (int),
 460                 .mode           = 0644,
 461                 .proc_handler   = &proc_dointvec,
 462         },
 463         {
 464                 .ctl_name       = KERN_MSGMNI,
 465                 .procname       = "msgmni",
 466                 .data           = &msg_ctlmni,
 467                 .maxlen         = sizeof (int),
 468                 .mode           = 0644,
 469                 .proc_handler   = &proc_dointvec,
 470         },
 471         {
 472                 .ctl_name       = KERN_MSGMNB,
 473                 .procname       =  "msgmnb",
 474                 .data           = &msg_ctlmnb,
 475                 .maxlen         = sizeof (int),
 476                 .mode           = 0644,
 477                 .proc_handler   = &proc_dointvec,
 478         },
 479         {
 480                 .ctl_name       = KERN_SEM,
 481                 .procname       = "sem",
 482                 .data           = &sem_ctls,
 483                 .maxlen         = 4*sizeof (int),
 484                 .mode           = 0644,
 485                 .proc_handler   = &proc_dointvec,
 486         },
 487 #endif
 488 #ifdef CONFIG_MAGIC_SYSRQ
 489         {
 490                 .ctl_name       = KERN_SYSRQ,
 491                 .procname       = "sysrq",
 492                 .data           = &sysrq_enabled,
 493                 .maxlen         = sizeof (int),
 494                 .mode           = 0644,
 495                 .proc_handler   = &proc_dointvec,
 496         },
 497 #endif
 498         {
 499                 .ctl_name       = KERN_CADPID,
 500                 .procname       = "cad_pid",
 501                 .data           = &cad_pid,
 502                 .maxlen         = sizeof (int),
 503                 .mode           = 0600,
 504                 .proc_handler   = &proc_dointvec,
 505         },
 506         {
 507                 .ctl_name       = KERN_MAX_THREADS,
 508                 .procname       = "threads-max",
 509                 .data           = &max_threads,
 510                 .maxlen         = sizeof(int),
 511                 .mode           = 0644,
 512                 .proc_handler   = &proc_dointvec,
 513         },
 514         {
 515                 .ctl_name       = KERN_RANDOM,
 516                 .procname       = "random",
 517                 .mode           = 0555,
 518                 .child          = random_table,
 519         },
 520 #ifdef CONFIG_UNIX98_PTYS
 521         {
 522                 .ctl_name       = KERN_PTY,
 523                 .procname       = "pty",
 524                 .mode           = 0555,
 525                 .child          = pty_table,
 526         },
 527 #endif
 528         {
 529                 .ctl_name       = KERN_OVERFLOWUID,
 530                 .procname       = "overflowuid",
 531                 .data           = &overflowuid,
 532                 .maxlen         = sizeof(int),
 533                 .mode           = 0644,
 534                 .proc_handler   = &proc_dointvec_minmax,
 535                 .strategy       = &sysctl_intvec,
 536                 .extra1         = &minolduid,
 537                 .extra2         = &maxolduid,
 538         },
 539         {
 540                 .ctl_name       = KERN_OVERFLOWGID,
 541                 .procname       = "overflowgid",
 542                 .data           = &overflowgid,
 543                 .maxlen         = sizeof(int),
 544                 .mode           = 0644,
 545                 .proc_handler   = &proc_dointvec_minmax,
 546                 .strategy       = &sysctl_intvec,
 547                 .extra1         = &minolduid,
 548                 .extra2         = &maxolduid,
 549         },
 550 #ifdef CONFIG_S390
 551 #ifdef CONFIG_MATHEMU
 552         {
 553                 .ctl_name       = KERN_IEEE_EMULATION_WARNINGS,
 554                 .procname       = "ieee_emulation_warnings",
 555                 .data           = &sysctl_ieee_emulation_warnings,
 556                 .maxlen         = sizeof(int),
 557                 .mode           = 0644,
 558                 .proc_handler   = &proc_dointvec,
 559         },
 560 #endif
 561 #ifdef CONFIG_NO_IDLE_HZ
 562         {
 563                 .ctl_name       = KERN_HZ_TIMER,
 564                 .procname       = "hz_timer",
 565                 .data           = &sysctl_hz_timer,
 566                 .maxlen         = sizeof(int),
 567                 .mode           = 0644,
 568                 .proc_handler   = &proc_dointvec,
 569         },
 570 #endif
 571         {
 572                 .ctl_name       = KERN_S390_USER_DEBUG_LOGGING,
 573                 .procname       = "userprocess_debug",
 574                 .data           = &sysctl_userprocess_debug,
 575                 .maxlen         = sizeof(int),
 576                 .mode           = 0644,
 577                 .proc_handler   = &proc_dointvec,
 578         },
 579 #endif
 580         {
 581                 .ctl_name       = KERN_PIDMAX,
 582                 .procname       = "pid_max",
 583                 .data           = &pid_max,
 584                 .maxlen         = sizeof (int),
 585                 .mode           = 0644,
 586                 .proc_handler   = &proc_dointvec_minmax,
 587                 .strategy       = sysctl_intvec,
 588                 .extra1         = &pid_max_min,
 589                 .extra2         = &pid_max_max,
 590         },
 591         {
 592                 .ctl_name       = KERN_PANIC_ON_OOPS,
 593                 .procname       = "panic_on_oops",
 594                 .data           = &panic_on_oops,
 595                 .maxlen         = sizeof(int),
 596                 .mode           = 0644,
 597                 .proc_handler   = &proc_dointvec,
 598         },
 599         {
 600                 .ctl_name       = KERN_PRINTK_RATELIMIT,
 601                 .procname       = "printk_ratelimit",
 602                 .data           = &printk_ratelimit_jiffies,
 603                 .maxlen         = sizeof(int),
 604                 .mode           = 0644,
 605                 .proc_handler   = &proc_dointvec_jiffies,
 606                 .strategy       = &sysctl_jiffies,
 607         },
 608         {
 609                 .ctl_name       = KERN_PRINTK_RATELIMIT_BURST,
 610                 .procname       = "printk_ratelimit_burst",
 611                 .data           = &printk_ratelimit_burst,
 612                 .maxlen         = sizeof(int),
 613                 .mode           = 0644,
 614                 .proc_handler   = &proc_dointvec,
 615         },
 616         {
 617                 .ctl_name       = KERN_NGROUPS_MAX,
 618                 .procname       = "ngroups_max",
 619                 .data           = &ngroups_max,
 620                 .maxlen         = sizeof (int),
 621                 .mode           = 0444,
 622                 .proc_handler   = &proc_dointvec,
 623         },
 624 #if defined(CONFIG_X86_LOCAL_APIC) && defined(CONFIG_X86)
 625         {
 626                 .ctl_name       = KERN_UNKNOWN_NMI_PANIC,
 627                 .procname       = "unknown_nmi_panic",
 628                 .data           = &unknown_nmi_panic,
 629                 .maxlen         = sizeof (int),
 630                 .mode           = 0644,
 631                 .proc_handler   = &proc_unknown_nmi_panic,
 632         },
 633 #endif
 634 #if defined(CONFIG_X86)
 635         {
 636                 .ctl_name       = KERN_BOOTLOADER_TYPE,
 637                 .procname       = "bootloader_type",
 638                 .data           = &bootloader_type,
 639                 .maxlen         = sizeof (int),
 640                 .mode           = 0444,
 641                 .proc_handler   = &proc_dointvec,
 642         },
 643 #endif
 644 #if defined(CONFIG_MMU)
 645         {
 646                 .ctl_name       = KERN_RANDOMIZE,
 647                 .procname       = "randomize_va_space",
 648                 .data           = &randomize_va_space,
 649                 .maxlen         = sizeof(int),
 650                 .mode           = 0644,
 651                 .proc_handler   = &proc_dointvec,
 652         },
 653 #endif
 654 #if defined(CONFIG_S390) && defined(CONFIG_SMP)
 655         {
 656                 .ctl_name       = KERN_SPIN_RETRY,
 657                 .procname       = "spin_retry",
 658                 .data           = &spin_retry,
 659                 .maxlen         = sizeof (int),
 660                 .mode           = 0644,
 661                 .proc_handler   = &proc_dointvec,
 662         },
 663 #endif
 664 #ifdef CONFIG_ACPI_SLEEP
 665         {
 666                 .ctl_name       = KERN_ACPI_VIDEO_FLAGS,
 667                 .procname       = "acpi_video_flags",
 668                 .data           = &acpi_video_flags,
 669                 .maxlen         = sizeof (unsigned long),
 670                 .mode           = 0644,
 671                 .proc_handler   = &proc_doulongvec_minmax,
 672         },
 673 #endif
 674 #ifdef CONFIG_IA64
 675         {
 676                 .ctl_name       = KERN_IA64_UNALIGNED,
 677                 .procname       = "ignore-unaligned-usertrap",
 678                 .data           = &no_unaligned_warning,
 679                 .maxlen         = sizeof (int),
 680                 .mode           = 0644,
 681                 .proc_handler   = &proc_dointvec,
 682         },
 683 #endif
 684 #ifdef CONFIG_COMPAT
 685         {
 686                 .ctl_name       = KERN_COMPAT_LOG,
 687                 .procname       = "compat-log",
 688                 .data           = &compat_log,
 689                 .maxlen         = sizeof (int),
 690                 .mode           = 0644,
 691                 .proc_handler   = &proc_dointvec,
 692         },
 693 #endif
 694 #ifdef CONFIG_RT_MUTEXES
 695         {
 696                 .ctl_name       = KERN_MAX_LOCK_DEPTH,
 697                 .procname       = "max_lock_depth",
 698                 .data           = &max_lock_depth,
 699                 .maxlen         = sizeof(int),
 700                 .mode           = 0644,
 701                 .proc_handler   = &proc_dointvec,
 702         },
 703 #endif
 704
 705         { .ctl_name = 0 }
 706 };
 707
 708 /* Constants for minimum and maximum testing in vm_table.
 709    We use these as one-element integer vectors. */
 710 static int zero;
 711 static int one_hundred = 100;
 712
 713
 714 static ctl_table vm_table[] = {
 715         {
 716                 .ctl_name       = VM_OVERCOMMIT_MEMORY,
 717                 .procname       = "overcommit_memory",
 718                 .data           = &sysctl_overcommit_memory,
 719                 .maxlen         = sizeof(sysctl_overcommit_memory),
 720                 .mode           = 0644,
 721                 .proc_handler   = &proc_dointvec,
 722         },
 723         {
 724                 .ctl_name       = VM_PANIC_ON_OOM,
 725                 .procname       = "panic_on_oom",
 726                 .data           = &sysctl_panic_on_oom,
 727                 .maxlen         = sizeof(sysctl_panic_on_oom),
 728                 .mode           = 0644,
 729                 .proc_handler   = &proc_dointvec,
 730         },
 731         {
 732                 .ctl_name       = VM_OVERCOMMIT_RATIO,
 733                 .procname       = "overcommit_ratio",
 734                 .data           = &sysctl_overcommit_ratio,
 735                 .maxlen         = sizeof(sysctl_overcommit_ratio),
 736                 .mode           = 0644,
 737                 .proc_handler   = &proc_dointvec,
 738         },
 739         {
 740                 .ctl_name       = VM_PAGE_CLUSTER,
 741                 .procname       = "page-cluster",
 742                 .data           = &page_cluster,
 743                 .maxlen         = sizeof(int),
 744                 .mode           = 0644,
 745                 .proc_handler   = &proc_dointvec,
 746         },
 747         {
 748                 .ctl_name       = VM_DIRTY_BACKGROUND,
 749                 .procname       = "dirty_background_ratio",
 750                 .data           = &dirty_background_ratio,
 751                 .maxlen         = sizeof(dirty_background_ratio),
 752                 .mode           = 0644,
 753                 .proc_handler   = &proc_dointvec_minmax,
 754                 .strategy       = &sysctl_intvec,
 755                 .extra1         = &zero,
 756                 .extra2         = &one_hundred,
 757         },
 758         {
 759                 .ctl_name       = VM_DIRTY_RATIO,
 760                 .procname       = "dirty_ratio",
 761                 .data           = &vm_dirty_ratio,
 762                 .maxlen         = sizeof(vm_dirty_ratio),
 763                 .mode           = 0644,
 764                 .proc_handler   = &proc_dointvec_minmax,
 765                 .strategy       = &sysctl_intvec,
 766                 .extra1         = &zero,
 767                 .extra2         = &one_hundred,
 768         },
 769         {
 770                 .ctl_name       = VM_DIRTY_WB_CS,
 771                 .procname       = "dirty_writeback_centisecs",
 772                 .data           = &dirty_writeback_interval,
 773                 .maxlen         = sizeof(dirty_writeback_interval),
 774                 .mode           = 0644,
 775                 .proc_handler   = &dirty_writeback_centisecs_handler,
 776         },
 777         {
 778                 .ctl_name       = VM_DIRTY_EXPIRE_CS,
 779                 .procname       = "dirty_expire_centisecs",
 780                 .data           = &dirty_expire_interval,
 781                 .maxlen         = sizeof(dirty_expire_interval),
 782                 .mode           = 0644,
 783                 .proc_handler   = &proc_dointvec_userhz_jiffies,
 784         },
 785         {
 786                 .ctl_name       = VM_NR_PDFLUSH_THREADS,
 787                 .procname       = "nr_pdflush_threads",
 788                 .data           = &nr_pdflush_threads,
 789                 .maxlen         = sizeof nr_pdflush_threads,
 790                 .mode           = 0444 /* read-only*/,
 791                 .proc_handler   = &proc_dointvec,
 792         },
 793         {
 794                 .ctl_name       = VM_SWAPPINESS,
 795                 .procname       = "swappiness",
 796                 .data           = &vm_swappiness,
 797                 .maxlen         = sizeof(vm_swappiness),
 798                 .mode           = 0644,
 799                 .proc_handler   = &proc_dointvec_minmax,
 800                 .strategy       = &sysctl_intvec,
 801                 .extra1         = &zero,
 802                 .extra2         = &one_hundred,
 803         },
 804 #ifdef CONFIG_HUGETLB_PAGE
 805          {
 806                 .ctl_name       = VM_HUGETLB_PAGES,
 807                 .procname       = "nr_hugepages",
 808                 .data           = &max_huge_pages,
 809                 .maxlen         = sizeof(unsigned long),
 810                 .mode           = 0644,
 811                 .proc_handler   = &hugetlb_sysctl_handler,
 812                 .extra1         = (void *)&hugetlb_zero,
 813                 .extra2         = (void *)&hugetlb_infinity,
 814          },
 815          {
 816                 .ctl_name       = VM_HUGETLB_GROUP,
 817                 .procname       = "hugetlb_shm_group",
 818                 .data           = &sysctl_hugetlb_shm_group,
 819                 .maxlen         = sizeof(gid_t),
 820                 .mode           = 0644,
 821                 .proc_handler   = &proc_dointvec,
 822          },
 823 #endif
 824         {
 825                 .ctl_name       = VM_LOWMEM_RESERVE_RATIO,
 826                 .procname       = "lowmem_reserve_ratio",
 827                 .data           = &sysctl_lowmem_reserve_ratio,
 828                 .maxlen         = sizeof(sysctl_lowmem_reserve_ratio),
 829                 .mode           = 0644,
 830                 .proc_handler   = &lowmem_reserve_ratio_sysctl_handler,
 831                 .strategy       = &sysctl_intvec,
 832         },
 833         {
 834                 .ctl_name       = VM_DROP_PAGECACHE,
 835                 .procname       = "drop_caches",
 836                 .data           = &sysctl_drop_caches,
 837                 .maxlen         = sizeof(int),
 838                 .mode           = 0644,
 839                 .proc_handler   = drop_caches_sysctl_handler,
 840                 .strategy       = &sysctl_intvec,
 841         },
 842         {
 843                 .ctl_name       = VM_MIN_FREE_KBYTES,
 844                 .procname       = "min_free_kbytes",
 845                 .data           = &min_free_kbytes,
 846                 .maxlen         = sizeof(min_free_kbytes),
 847                 .mode           = 0644,
 848                 .proc_handler   = &min_free_kbytes_sysctl_handler,
 849                 .strategy       = &sysctl_intvec,
 850                 .extra1         = &zero,
 851         },
 852         {
 853                 .ctl_name       = VM_PERCPU_PAGELIST_FRACTION,
 854                 .procname       = "percpu_pagelist_fraction",
 855                 .data           = &percpu_pagelist_fraction,
 856                 .maxlen         = sizeof(percpu_pagelist_fraction),
 857                 .mode           = 0644,
 858                 .proc_handler   = &percpu_pagelist_fraction_sysctl_handler,
 859                 .strategy       = &sysctl_intvec,
 860                 .extra1         = &min_percpu_pagelist_fract,
 861         },
 862 #ifdef CONFIG_MMU
 863         {
 864                 .ctl_name       = VM_MAX_MAP_COUNT,
 865                 .procname       = "max_map_count",
 866                 .data           = &sysctl_max_map_count,
 867                 .maxlen         = sizeof(sysctl_max_map_count),
 868                 .mode           = 0644,
 869                 .proc_handler   = &proc_dointvec
 870         },
 871 #endif
 872         {
 873                 .ctl_name       = VM_LAPTOP_MODE,
 874                 .procname       = "laptop_mode",
 875                 .data           = &laptop_mode,
 876                 .maxlen         = sizeof(laptop_mode),
 877                 .mode           = 0644,
 878                 .proc_handler   = &proc_dointvec_jiffies,
 879                 .strategy       = &sysctl_jiffies,
 880         },
 881         {
 882                 .ctl_name       = VM_BLOCK_DUMP,
 883                 .procname       = "block_dump",
 884                 .data           = &block_dump,
 885                 .maxlen         = sizeof(block_dump),
 886                 .mode           = 0644,
 887                 .proc_handler   = &proc_dointvec,
 888                 .strategy       = &sysctl_intvec,
 889                 .extra1         = &zero,
 890         },
 891         {
 892                 .ctl_name       = VM_VFS_CACHE_PRESSURE,
 893                 .procname       = "vfs_cache_pressure",
 894                 .data           = &sysctl_vfs_cache_pressure,
 895                 .maxlen         = sizeof(sysctl_vfs_cache_pressure),
 896                 .mode           = 0644,
 897                 .proc_handler   = &proc_dointvec,
 898                 .strategy       = &sysctl_intvec,
 899                 .extra1         = &zero,
 900         },
 901 #ifdef HAVE_ARCH_PICK_MMAP_LAYOUT
 902         {
 903                 .ctl_name       = VM_LEGACY_VA_LAYOUT,
 904                 .procname       = "legacy_va_layout",
 905                 .data           = &sysctl_legacy_va_layout,
 906                 .maxlen         = sizeof(sysctl_legacy_va_layout),
 907                 .mode           = 0644,
 908                 .proc_handler   = &proc_dointvec,
 909                 .strategy       = &sysctl_intvec,
 910                 .extra1         = &zero,
 911         },
 912 #endif
 913 #ifdef CONFIG_SWAP
 914         {
 915                 .ctl_name       = VM_SWAP_TOKEN_TIMEOUT,
 916                 .procname       = "swap_token_timeout",
 917                 .data           = &swap_token_default_timeout,
 918                 .maxlen         = sizeof(swap_token_default_timeout),
 919                 .mode           = 0644,
 920                 .proc_handler   = &proc_dointvec_jiffies,
 921                 .strategy       = &sysctl_jiffies,
 922         },
 923 #endif
 924 #ifdef CONFIG_NUMA
 925         {
 926                 .ctl_name       = VM_ZONE_RECLAIM_MODE,
 927                 .procname       = "zone_reclaim_mode",
 928                 .data           = &zone_reclaim_mode,
 929                 .maxlen         = sizeof(zone_reclaim_mode),
 930                 .mode           = 0644,
 931                 .proc_handler   = &proc_dointvec,
 932                 .strategy       = &sysctl_intvec,
 933                 .extra1         = &zero,
 934         },
 935         {
 936                 .ctl_name       = VM_MIN_UNMAPPED,
 937                 .procname       = "min_unmapped_ratio",
 938                 .data           = &sysctl_min_unmapped_ratio,
 939                 .maxlen         = sizeof(sysctl_min_unmapped_ratio),
 940                 .mode           = 0644,
 941                 .proc_handler   = &sysctl_min_unmapped_ratio_sysctl_handler,
 942                 .strategy       = &sysctl_intvec,
 943                 .extra1         = &zero,
 944                 .extra2         = &one_hundred,
 945         },
 946         {
 947                 .ctl_name       = VM_MIN_SLAB,
 948                 .procname       = "min_slab_ratio",
 949                 .data           = &sysctl_min_slab_ratio,
 950                 .maxlen         = sizeof(sysctl_min_slab_ratio),
 951                 .mode           = 0644,
 952                 .proc_handler   = &sysctl_min_slab_ratio_sysctl_handler,
 953                 .strategy       = &sysctl_intvec,
 954                 .extra1         = &zero,
 955                 .extra2         = &one_hundred,
 956         },
 957 #endif
 958 #ifdef CONFIG_X86_32
 959         {
 960                 .ctl_name       = VM_VDSO_ENABLED,
 961                 .procname       = "vdso_enabled",
 962                 .data           = &vdso_enabled,
 963                 .maxlen         = sizeof(vdso_enabled),
 964                 .mode           = 0644,
 965                 .proc_handler   = &proc_dointvec,
 966                 .strategy       = &sysctl_intvec,
 967                 .extra1         = &zero,
 968         },
 969 #endif
 970         { .ctl_name = 0 }
 971 };
 972
 973 static ctl_table fs_table[] = {
 974         {
 975                 .ctl_name       = FS_NRINODE,
 976                 .procname       = "inode-nr",
 977                 .data           = &inodes_stat,
 978                 .maxlen         = 2*sizeof(int),
 979                 .mode           = 0444,
 980                 .proc_handler   = &proc_dointvec,
 981         },
 982         {
 983                 .ctl_name       = FS_STATINODE,
 984                 .procname       = "inode-state",
 985                 .data           = &inodes_stat,
 986                 .maxlen         = 7*sizeof(int),
 987                 .mode           = 0444,
 988                 .proc_handler   = &proc_dointvec,
 989         },
 990         {
 991                 .ctl_name       = FS_NRFILE,
 992                 .procname       = "file-nr",
 993                 .data           = &files_stat,
 994                 .maxlen         = 3*sizeof(int),
 995                 .mode           = 0444,
 996                 .proc_handler   = &proc_nr_files,
 997         },
 998         {
 999                 .ctl_name       = FS_MAXFILE,
1000                 .procname       = "file-max",
1001                 .data           = &files_stat.max_files,
1002                 .maxlen         = sizeof(int),
1003                 .mode           = 0644,
1004                 .proc_handler   = &proc_dointvec,
1005         },
1006         {
1007                 .ctl_name       = FS_DENTRY,
1008                 .procname       = "dentry-state",
1009                 .data           = &dentry_stat,
1010                 .maxlen         = 6*sizeof(int),
1011                 .mode           = 0444,
1012                 .proc_handler   = &proc_dointvec,
1013         },
1014         {
1015                 .ctl_name       = FS_OVERFLOWUID,
1016                 .procname       = "overflowuid",
1017                 .data           = &fs_overflowuid,
1018                 .maxlen         = sizeof(int),
1019                 .mode           = 0644,
1020                 .proc_handler   = &proc_dointvec_minmax,
1021                 .strategy       = &sysctl_intvec,
1022                 .extra1         = &minolduid,
1023                 .extra2         = &maxolduid,
1024         },
1025         {
1026                 .ctl_name       = FS_OVERFLOWGID,
1027                 .procname       = "overflowgid",
1028                 .data           = &fs_overflowgid,
1029                 .maxlen         = sizeof(int),
1030                 .mode           = 0644,
1031                 .proc_handler   = &proc_dointvec_minmax,
1032                 .strategy       = &sysctl_intvec,
1033                 .extra1         = &minolduid,
1034                 .extra2         = &maxolduid,
1035         },
1036         {
1037                 .ctl_name       = FS_LEASES,
1038                 .procname       = "leases-enable",
1039                 .data           = &leases_enable,
1040                 .maxlen         = sizeof(int),
1041                 .mode           = 0644,
1042                 .proc_handler   = &proc_dointvec,
1043         },
1044 #ifdef CONFIG_DNOTIFY
1045         {
1046                 .ctl_name       = FS_DIR_NOTIFY,
1047                 .procname       = "dir-notify-enable",
1048                 .data           = &dir_notify_enable,
1049                 .maxlen         = sizeof(int),
1050                 .mode           = 0644,
1051                 .proc_handler   = &proc_dointvec,
1052         },
1053 #endif
1054 #ifdef CONFIG_MMU
1055         {
1056                 .ctl_name       = FS_LEASE_TIME,
1057                 .procname       = "lease-break-time",
1058                 .data           = &lease_break_time,
1059                 .maxlen         = sizeof(int),
1060                 .mode           = 0644,
1061                 .proc_handler   = &proc_dointvec,
1062         },
1063         {
1064                 .ctl_name       = FS_AIO_NR,
1065                 .procname       = "aio-nr",
1066                 .data           = &aio_nr,
1067                 .maxlen         = sizeof(aio_nr),
1068                 .mode           = 0444,
1069                 .proc_handler   = &proc_doulongvec_minmax,
1070         },
1071         {
1072                 .ctl_name       = FS_AIO_MAX_NR,
1073                 .procname       = "aio-max-nr",
1074                 .data           = &aio_max_nr,
1075                 .maxlen         = sizeof(aio_max_nr),
1076                 .mode           = 0644,
1077                 .proc_handler   = &proc_doulongvec_minmax,
1078         },
1079 #ifdef CONFIG_INOTIFY_USER
1080         {
1081                 .ctl_name       = FS_INOTIFY,
1082                 .procname       = "inotify",
1083                 .mode           = 0555,
1084                 .child          = inotify_table,
1085         },
1086 #endif
1087 #endif
1088         {
1089                 .ctl_name       = KERN_SETUID_DUMPABLE,
1090                 .procname       = "suid_dumpable",
1091                 .data           = &suid_dumpable,
1092                 .maxlen         = sizeof(int),
1093                 .mode           = 0644,
1094                 .proc_handler   = &proc_dointvec,
1095         },
1096         { .ctl_name = 0 }
1097 };
1098
1099 static ctl_table debug_table[] = {
1100         { .ctl_name = 0 }
1101 };
1102
1103 static ctl_table dev_table[] = {
1104         { .ctl_name = 0 }
1105 };
1106
1107 extern void init_irq_proc (void);
1108
1109 static DEFINE_SPINLOCK(sysctl_lock);
1110
1111 /* called under sysctl_lock */
1112 static int use_table(struct ctl_table_header *p)
1113 {
1114         if (unlikely(p->unregistering))
1115                 return 0;
1116         p->used++;
1117         return 1;
1118 }
1119
1120 /* called under sysctl_lock */
1121 static void unuse_table(struct ctl_table_header *p)
1122 {
1123         if (!--p->used)
1124                 if (unlikely(p->unregistering))
1125                         complete(p->unregistering);
1126 }
1127
1128 /* called under sysctl_lock, will reacquire if has to wait */
1129 static void start_unregistering(struct ctl_table_header *p)
1130 {
1131         /*
1132          * if p->used is 0, nobody will ever touch that entry again;
1133          * we'll eliminate all paths to it before dropping sysctl_lock
1134          */
1135         if (unlikely(p->used)) {
1136                 struct completion wait;
1137                 init_completion(&wait);
1138                 p->unregistering = &wait;
1139                 spin_unlock(&sysctl_lock);
1140                 wait_for_completion(&wait);
1141                 spin_lock(&sysctl_lock);
1142         }
1143         /*
1144          * do not remove from the list until nobody holds it; walking the
1145          * list in do_sysctl() relies on that.
1146          */
1147         list_del_init(&p->ctl_entry);
1148 }
1149
1150 void __init sysctl_init(void)
1151 {
1152 #ifdef CONFIG_PROC_FS
1153         register_proc_table(root_table, proc_sys_root, &root_table_header);
1154         init_irq_proc();
1155 #endif
1156 }
1157
1158 int do_sysctl(int __user *name, int nlen, void __user *oldval, size_t __user *oldlenp,
1159                void __user *newval, size_t newlen)
1160 {
1161         struct list_head *tmp;
1162         int error = -ENOTDIR;
1163
1164         if (nlen <= 0 || nlen >= CTL_MAXNAME)
1165                 return -ENOTDIR;
1166         if (oldval) {
1167                 int old_len;
1168                 if (!oldlenp || get_user(old_len, oldlenp))
1169                         return -EFAULT;
1170         }
1171         spin_lock(&sysctl_lock);
1172         tmp = &root_table_header.ctl_entry;
1173         do {
1174                 struct ctl_table_header *head =
1175                         list_entry(tmp, struct ctl_table_header, ctl_entry);
1176                 void *context = NULL;
1177
1178                 if (!use_table(head))
1179                         continue;
1180
1181                 spin_unlock(&sysctl_lock);
1182
1183                 error = parse_table(name, nlen, oldval, oldlenp,
1184                                         newval, newlen, head->ctl_table,
1185                                         &context);
1186                 kfree(context);
1187
1188                 spin_lock(&sysctl_lock);
1189                 unuse_table(head);
1190                 if (error != -ENOTDIR)
1191                         break;
1192         } while ((tmp = tmp->next) != &root_table_header.ctl_entry);
1193         spin_unlock(&sysctl_lock);
1194         return error;
1195 }
1196
1197 asmlinkage long sys_sysctl(struct __sysctl_args __user *args)
1198 {
1199         struct __sysctl_args tmp;
1200         int error;
1201
1202         if (copy_from_user(&tmp, args, sizeof(tmp)))
1203                 return -EFAULT;
1204
1205         lock_kernel();
1206         error = do_sysctl(tmp.name, tmp.nlen, tmp.oldval, tmp.oldlenp,
1207                           tmp.newval, tmp.newlen);
1208         unlock_kernel();
1209         return error;
1210 }
1211
1212 /*
1213  * ctl_perm does NOT grant the superuser all rights automatically, because
1214  * some sysctl variables are readonly even to root.
1215  */
1216
1217 static int test_perm(int mode, int op)
1218 {
1219         if (!current->euid)
1220                 mode >>= 6;
1221         else if (in_egroup_p(0))
1222                 mode >>= 3;
1223         if ((mode & op & 0007) == op)
1224                 return 0;
1225         return -EACCES;
1226 }
1227
1228 static inline int ctl_perm(ctl_table *table, int op)
1229 {
1230         int error;
1231         error = security_sysctl(table, op);
1232         if (error)
1233                 return error;
1234         return test_perm(table->mode, op);
1235 }
1236
1237 static int parse_table(int __user *name, int nlen,
1238                        void __user *oldval, size_t __user *oldlenp,
1239                        void __user *newval, size_t newlen,
1240                        ctl_table *table, void **context)
1241 {
1242         int n;
1243 repeat:
1244         if (!nlen)
1245                 return -ENOTDIR;
1246         if (get_user(n, name))
1247                 return -EFAULT;
1248         for ( ; table->ctl_name; table++) {
1249                 if (n == table->ctl_name || table->ctl_name == CTL_ANY) {
1250                         int error;
1251                         if (table->child) {
1252                                 if (ctl_perm(table, 001))
1253                                         return -EPERM;
1254                                 if (table->strategy) {
1255                                         error = table->strategy(
1256                                                 table, name, nlen,
1257                                                 oldval, oldlenp,
1258                                                 newval, newlen, context);
1259                                         if (error)
1260                                                 return error;
1261                                 }
1262                                 name++;
1263                                 nlen--;
1264                                 table = table->child;
1265                                 goto repeat;
1266                         }
1267                         error = do_sysctl_strategy(table, name, nlen,
1268                                                    oldval, oldlenp,
1269                                                    newval, newlen, context);
1270                         return error;
1271                 }
1272         }
1273         return -ENOTDIR;
1274 }
1275
1276 /* Perform the actual read/write of a sysctl table entry. */
1277 int do_sysctl_strategy (ctl_table *table,
1278                         int __user *name, int nlen,
1279                         void __user *oldval, size_t __user *oldlenp,
1280                         void __user *newval, size_t newlen, void **context)
1281 {
1282         int op = 0, rc;
1283         size_t len;
1284
1285         if (oldval)
1286                 op |= 004;
1287         if (newval)
1288                 op |= 002;
1289         if (ctl_perm(table, op))
1290                 return -EPERM;
1291
1292         if (table->strategy) {
1293                 rc = table->strategy(table, name, nlen, oldval, oldlenp,
1294                                      newval, newlen, context);
1295                 if (rc < 0)
1296                         return rc;
1297                 if (rc > 0)
1298                         return 0;
1299         }
1300
1301         /* If there is no strategy routine, or if the strategy returns
1302          * zero, proceed with automatic r/w */
1303         if (table->data && table->maxlen) {
1304                 if (oldval && oldlenp) {
1305                         if (get_user(len, oldlenp))
1306                                 return -EFAULT;
1307                         if (len) {
1308                                 if (len > table->maxlen)
1309                                         len = table->maxlen;
1310                                 if(copy_to_user(oldval, table->data, len))
1311                                         return -EFAULT;
1312                                 if(put_user(len, oldlenp))
1313                                         return -EFAULT;
1314                         }
1315                 }
1316                 if (newval && newlen) {
1317                         len = newlen;
1318                         if (len > table->maxlen)
1319                                 len = table->maxlen;
1320                         if(copy_from_user(table->data, newval, len))
1321                                 return -EFAULT;
1322                 }
1323         }
1324         return 0;
1325 }
1326
1327 /**
1328  * register_sysctl_table - register a sysctl hierarchy
1329  * @table: the top-level table structure
1330  * @insert_at_head: whether the entry should be inserted in front or at the end
1331  *
1332  * Register a sysctl table hierarchy. @table should be a filled in ctl_table
1333  * array. An entry with a ctl_name of 0 terminates the table.
1334  *
1335  * The members of the &ctl_table structure are used as follows:
1336  *
1337  * ctl_name - This is the numeric sysctl value used by sysctl(2). The number
1338  *            must be unique within that level of sysctl
1339  *
1340  * procname - the name of the sysctl file under /proc/sys. Set to %NULL to not
1341  *            enter a sysctl file
1342  *
1343  * data - a pointer to data for use by proc_handler
1344  *
1345  * maxlen - the maximum size in bytes of the data
1346  *
1347  * mode - the file permissions for the /proc/sys file, and for sysctl(2)
1348  *
1349  * child - a pointer to the child sysctl table if this entry is a directory, or
1350  *         %NULL.
1351  *
1352  * proc_handler - the text handler routine (described below)
1353  *
1354  * strategy - the strategy routine (described below)
1355  *
1356  * de - for internal use by the sysctl routines
1357  *
1358  * extra1, extra2 - extra pointers usable by the proc handler routines
1359  *
1360  * Leaf nodes in the sysctl tree will be represented by a single file
1361  * under /proc; non-leaf nodes will be represented by directories.
1362  *
1363  * sysctl(2) can automatically manage read and write requests through
1364  * the sysctl table.  The data and maxlen fields of the ctl_table
1365  * struct enable minimal validation of the values being written to be
1366  * performed, and the mode field allows minimal authentication.
1367  *
1368  * More sophisticated management can be enabled by the provision of a
1369  * strategy routine with the table entry.  This will be called before
1370  * any automatic read or write of the data is performed.
1371  *
1372  * The strategy routine may return
1373  *
1374  * < 0 - Error occurred (error is passed to user process)
1375  *
1376  * 0   - OK - proceed with automatic read or write.
1377  *
1378  * > 0 - OK - read or write has been done by the strategy routine, so
1379  *       return immediately.
1380  *
1381  * There must be a proc_handler routine for any terminal nodes
1382  * mirrored under /proc/sys (non-terminals are handled by a built-in
1383  * directory handler).  Several default handlers are available to
1384  * cover common cases -
1385  *
1386  * proc_dostring(), proc_dointvec(), proc_dointvec_jiffies(),
1387  * proc_dointvec_userhz_jiffies(), proc_dointvec_minmax(),
1388  * proc_doulongvec_ms_jiffies_minmax(), proc_doulongvec_minmax()
1389  *
1390  * It is the handler's job to read the input buffer from user memory
1391  * and process it. The handler should return 0 on success.
1392  *
1393  * This routine returns %NULL on a failure to register, and a pointer
1394  * to the table header on success.
1395  */
1396 struct ctl_table_header *register_sysctl_table(ctl_table * table,
1397                                                int insert_at_head)
1398 {
1399         struct ctl_table_header *tmp;
1400         tmp = kmalloc(sizeof(struct ctl_table_header), GFP_KERNEL);
1401         if (!tmp)
1402                 return NULL;
1403         tmp->ctl_table = table;
1404         INIT_LIST_HEAD(&tmp->ctl_entry);
1405         tmp->used = 0;
1406         tmp->unregistering = NULL;
1407         spin_lock(&sysctl_lock);
1408         if (insert_at_head)
1409                 list_add(&tmp->ctl_entry, &root_table_header.ctl_entry);
1410         else
1411                 list_add_tail(&tmp->ctl_entry, &root_table_header.ctl_entry);
1412         spin_unlock(&sysctl_lock);
1413 #ifdef CONFIG_PROC_FS
1414         register_proc_table(table, proc_sys_root, tmp);
1415 #endif
1416         return tmp;
1417 }
1418
1419 /**
1420  * unregister_sysctl_table - unregister a sysctl table hierarchy
1421  * @header: the header returned from register_sysctl_table
1422  *
1423  * Unregisters the sysctl table and all children. proc entries may not
1424  * actually be removed until they are no longer used by anyone.
1425  */
1426 void unregister_sysctl_table(struct ctl_table_header * header)
1427 {
1428         might_sleep();
1429         spin_lock(&sysctl_lock);
1430         start_unregistering(header);
1431 #ifdef CONFIG_PROC_FS
1432         unregister_proc_table(header->ctl_table, proc_sys_root);
1433 #endif
1434         spin_unlock(&sysctl_lock);
1435         kfree(header);
1436 }
1437
1438 /*
1439  * /proc/sys support
1440  */
1441
1442 #ifdef CONFIG_PROC_FS
1443
1444 /* Scan the sysctl entries in table and add them all into /proc */
1445 static void register_proc_table(ctl_table * table, struct proc_dir_entry *root, void *set)
1446 {
1447         struct proc_dir_entry *de;
1448         int len;
1449         mode_t mode;
1450
1451         for (; table->ctl_name; table++) {
1452                 /* Can't do anything without a proc name. */
1453                 if (!table->procname)
1454                         continue;
1455                 /* Maybe we can't do anything with it... */
1456                 if (!table->proc_handler && !table->child) {
1457                         printk(KERN_WARNING "SYSCTL: Can't register %s\n",
1458                                 table->procname);
1459                         continue;
1460                 }
1461
1462                 len = strlen(table->procname);
1463                 mode = table->mode;
1464
1465                 de = NULL;
1466                 if (table->proc_handler)
1467                         mode |= S_IFREG;
1468                 else {
1469                         mode |= S_IFDIR;
1470                         for (de = root->subdir; de; de = de->next) {
1471                                 if (proc_match(len, table->procname, de))
1472                                         break;
1473                         }
1474                         /* If the subdir exists already, de is non-NULL */
1475                 }
1476
1477                 if (!de) {
1478                         de = create_proc_entry(table->procname, mode, root);
1479                         if (!de)
1480                                 continue;
1481                         de->set = set;
1482                         de->data = (void *) table;
1483                         if (table->proc_handler)
1484                                 de->proc_fops = &proc_sys_file_operations;
1485                 }
1486                 table->de = de;
1487                 if (de->mode & S_IFDIR)
1488                         register_proc_table(table->child, de, set);
1489         }
1490 }
1491
1492 /*
1493  * Unregister a /proc sysctl table and any subdirectories.
1494  */
1495 static void unregister_proc_table(ctl_table * table, struct proc_dir_entry *root)
1496 {
1497         struct proc_dir_entry *de;
1498         for (; table->ctl_name; table++) {
1499                 if (!(de = table->de))
1500                         continue;
1501                 if (de->mode & S_IFDIR) {
1502                         if (!table->child) {
1503                                 printk (KERN_ALERT "Help - malformed sysctl tree on free\n");
1504                                 continue;
1505                         }
1506                         unregister_proc_table(table->child, de);
1507
1508                         /* Don't unregister directories which still have entries.. */
1509                         if (de->subdir)
1510                                 continue;
1511                 }
1512
1513                 /*
1514                  * In any case, mark the entry as goner; we'll keep it
1515                  * around if it's busy, but we'll know to do nothing with
1516                  * its fields.  We are under sysctl_lock here.
1517                  */
1518                 de->data = NULL;
1519
1520                 /* Don't unregister proc entries that are still being used.. */
1521                 if (atomic_read(&de->count))
1522                         continue;
1523
1524                 table->de = NULL;
1525                 remove_proc_entry(table->procname, root);
1526         }
1527 }
1528
1529 static ssize_t do_rw_proc(int write, struct file * file, char __user * buf,
1530                           size_t count, loff_t *ppos)
1531 {
1532         int op;
1533         struct proc_dir_entry *de = PDE(file->f_dentry->d_inode);
1534         struct ctl_table *table;
1535         size_t res;
1536         ssize_t error = -ENOTDIR;
1537
1538         spin_lock(&sysctl_lock);
1539         if (de && de->data && use_table(de->set)) {
1540                 /*
1541                  * at that point we know that sysctl was not unregistered
1542                  * and won't be until we finish
1543                  */
1544                 spin_unlock(&sysctl_lock);
1545                 table = (struct ctl_table *) de->data;
1546                 if (!table || !table->proc_handler)
1547                         goto out;
1548                 error = -EPERM;
1549                 op = (write ? 002 : 004);
1550                 if (ctl_perm(table, op))
1551                         goto out;
1552
1553                 /* careful: calling conventions are nasty here */
1554                 res = count;
1555                 error = (*table->proc_handler)(table, write, file,
1556                                                 buf, &res, ppos);
1557                 if (!error)
1558                         error = res;
1559         out:
1560                 spin_lock(&sysctl_lock);
1561                 unuse_table(de->set);
1562         }
1563         spin_unlock(&sysctl_lock);
1564         return error;
1565 }
1566
1567 static int proc_opensys(struct inode *inode, struct file *file)
1568 {
1569         if (file->f_mode & FMODE_WRITE) {
1570                 /*
1571                  * sysctl entries that are not writable,
1572                  * are _NOT_ writable, capabilities or not.
1573                  */
1574                 if (!(inode->i_mode & S_IWUSR))
1575                         return -EPERM;
1576         }
1577
1578         return 0;
1579 }
1580
1581 static ssize_t proc_readsys(struct file * file, char __user * buf,
1582                             size_t count, loff_t *ppos)
1583 {
1584         return do_rw_proc(0, file, buf, count, ppos);
1585 }
1586
1587 static ssize_t proc_writesys(struct file * file, const char __user * buf,
1588                              size_t count, loff_t *ppos)
1589 {
1590         return do_rw_proc(1, file, (char __user *) buf, count, ppos);
1591 }
1592
1593 /**
1594  * proc_dostring - read a string sysctl
1595  * @table: the sysctl table
1596  * @write: %TRUE if this is a write to the sysctl file
1597  * @filp: the file structure
1598  * @buffer: the user buffer
1599  * @lenp: the size of the user buffer
1600  * @ppos: file position
1601  *
1602  * Reads/writes a string from/to the user buffer. If the kernel
1603  * buffer provided is not large enough to hold the string, the
1604  * string is truncated. The copied string is %NULL-terminated.
1605  * If the string is being read by the user process, it is copied
1606  * and a newline '\n' is added. It is truncated if the buffer is
1607  * not large enough.
1608  *
1609  * Returns 0 on success.
1610  */
1611 int proc_dostring(ctl_table *table, int write, struct file *filp,
1612                   void __user *buffer, size_t *lenp, loff_t *ppos)
1613 {
1614         size_t len;
1615         char __user *p;
1616         char c;
1617
1618         if (!table->data || !table->maxlen || !*lenp ||
1619             (*ppos && !write)) {
1620                 *lenp = 0;
1621                 return 0;
1622         }
1623
1624         if (write) {
1625                 len = 0;
1626                 p = buffer;
1627                 while (len < *lenp) {
1628                         if (get_user(c, p++))
1629                                 return -EFAULT;
1630                         if (c == 0 || c == '\n')
1631                                 break;
1632                         len++;
1633                 }
1634                 if (len >= table->maxlen)
1635                         len = table->maxlen-1;
1636                 if(copy_from_user(table->data, buffer, len))
1637                         return -EFAULT;
1638                 ((char *) table->data)[len] = 0;
1639                 *ppos += *lenp;
1640         } else {
1641                 len = strlen(table->data);
1642                 if (len > table->maxlen)
1643                         len = table->maxlen;
1644                 if (len > *lenp)
1645                         len = *lenp;
1646                 if (len)
1647                         if(copy_to_user(buffer, table->data, len))
1648                                 return -EFAULT;
1649                 if (len < *lenp) {
1650                         if(put_user('\n', ((char __user *) buffer) + len))
1651                                 return -EFAULT;
1652                         len++;
1653                 }
1654                 *lenp = len;
1655                 *ppos += len;
1656         }
1657         return 0;
1658 }
1659
1660 /*
1661  *      Special case of dostring for the UTS structure. This has locks
1662  *      to observe. Should this be in kernel/sys.c ????
1663  */
1664
1665 static int proc_doutsstring(ctl_table *table, int write, struct file *filp,
1666                   void __user *buffer, size_t *lenp, loff_t *ppos)
1667 {
1668         int r;
1669
1670         if (!write) {
1671                 down_read(&uts_sem);
1672                 r=proc_dostring(table,0,filp,buffer,lenp, ppos);
1673                 up_read(&uts_sem);
1674         } else {
1675                 down_write(&uts_sem);
1676                 r=proc_dostring(table,1,filp,buffer,lenp, ppos);
1677                 up_write(&uts_sem);
1678         }
1679         return r;
1680 }
1681
1682 static int do_proc_dointvec_conv(int *negp, unsigned long *lvalp,
1683                                  int *valp,
1684                                  int write, void *data)
1685 {
1686         if (write) {
1687                 *valp = *negp ? -*lvalp : *lvalp;
1688         } else {
1689                 int val = *valp;
1690                 if (val < 0) {
1691                         *negp = -1;
1692                         *lvalp = (unsigned long)-val;
1693                 } else {
1694                         *negp = 0;
1695                         *lvalp = (unsigned long)val;
1696                 }
1697         }
1698         return 0;
1699 }
1700
1701 static int do_proc_dointvec(ctl_table *table, int write, struct file *filp,
1702                   void __user *buffer, size_t *lenp, loff_t *ppos,
1703                   int (*conv)(int *negp, unsigned long *lvalp, int *valp,
1704                               int write, void *data),
1705                   void *data)
1706 {
1707 #define TMPBUFLEN 21
1708         int *i, vleft, first=1, neg, val;
1709         unsigned long lval;
1710         size_t left, len;
1711
1712         char buf[TMPBUFLEN], *p;
1713         char __user *s = buffer;
1714
1715         if (!table->data || !table->maxlen || !*lenp ||
1716             (*ppos && !write)) {
1717                 *lenp = 0;
1718                 return 0;
1719         }
1720
1721         i = (int *) table->data;
1722         vleft = table->maxlen / sizeof(*i);
1723         left = *lenp;
1724
1725         if (!conv)
1726                 conv = do_proc_dointvec_conv;
1727
1728         for (; left && vleft--; i++, first=0) {
1729                 if (write) {
1730                         while (left) {
1731                                 char c;
1732                                 if (get_user(c, s))
1733                                         return -EFAULT;
1734                                 if (!isspace(c))
1735                                         break;
1736                                 left--;
1737                                 s++;
1738                         }
1739                         if (!left)
1740                                 break;
1741                         neg = 0;
1742                         len = left;
1743                         if (len > sizeof(buf) - 1)
1744                                 len = sizeof(buf) - 1;
1745                         if (copy_from_user(buf, s, len))
1746                                 return -EFAULT;
1747                         buf[len] = 0;
1748                         p = buf;
1749                         if (*p == '-' && left > 1) {
1750                                 neg = 1;
1751                                 left--, p++;
1752                         }
1753                         if (*p < '0' || *p > '9')
1754                                 break;
1755
1756                         lval = simple_strtoul(p, &p, 0);
1757
1758                         len = p-buf;
1759                         if ((len < left) && *p && !isspace(*p))
1760                                 break;
1761                         if (neg)
1762                                 val = -val;
1763                         s += len;
1764                         left -= len;
1765
1766                         if (conv(&neg, &lval, i, 1, data))
1767                                 break;
1768                 } else {
1769                         p = buf;
1770                         if (!first)
1771                                 *p++ = '\t';
1772
1773                         if (conv(&neg, &lval, i, 0, data))
1774                                 break;
1775
1776                         sprintf(p, "%s%lu", neg ? "-" : "", lval);
1777                         len = strlen(buf);
1778                         if (len > left)
1779                                 len = left;
1780                         if(copy_to_user(s, buf, len))
1781                                 return -EFAULT;
1782                         left -= len;
1783                         s += len;
1784                 }
1785         }
1786
1787         if (!write && !first && left) {
1788                 if(put_user('\n', s))
1789                         return -EFAULT;
1790                 left--, s++;
1791         }
1792         if (write) {
1793                 while (left) {
1794                         char c;
1795                         if (get_user(c, s++))
1796                                 return -EFAULT;
1797                         if (!isspace(c))
1798                                 break;
1799                         left--;
1800                 }
1801         }
1802         if (write && first)
1803                 return -EINVAL;
1804         *lenp -= left;
1805         *ppos += *lenp;
1806         return 0;
1807 #undef TMPBUFLEN
1808 }
1809
1810 /**
1811  * proc_dointvec - read a vector of integers
1812  * @table: the sysctl table
1813  * @write: %TRUE if this is a write to the sysctl file
1814  * @filp: the file structure
1815  * @buffer: the user buffer
1816  * @lenp: the size of the user buffer
1817  * @ppos: file position
1818  *
1819  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
1820  * values from/to the user buffer, treated as an ASCII string.
1821  *
1822  * Returns 0 on success.
1823  */
1824 int proc_dointvec(ctl_table *table, int write, struct file *filp,
1825                      void __user *buffer, size_t *lenp, loff_t *ppos)
1826 {
1827     return do_proc_dointvec(table,write,filp,buffer,lenp,ppos,
1828                             NULL,NULL);
1829 }
1830
1831 #define OP_SET  0
1832 #define OP_AND  1
1833 #define OP_OR   2
1834 #define OP_MAX  3
1835 #define OP_MIN  4
1836
1837 static int do_proc_dointvec_bset_conv(int *negp, unsigned long *lvalp,
1838                                       int *valp,
1839                                       int write, void *data)
1840 {
1841         int op = *(int *)data;
1842         if (write) {
1843                 int val = *negp ? -*lvalp : *lvalp;
1844                 switch(op) {
1845                 case OP_SET:    *valp = val; break;
1846                 case OP_AND:    *valp &= val; break;
1847                 case OP_OR:     *valp |= val; break;
1848                 case OP_MAX:    if(*valp < val)
1849                                         *valp = val;
1850                                 break;
1851                 case OP_MIN:    if(*valp > val)
1852                                 *valp = val;
1853                                 break;
1854                 }
1855         } else {
1856                 int val = *valp;
1857                 if (val < 0) {
1858                         *negp = -1;
1859                         *lvalp = (unsigned long)-val;
1860                 } else {
1861                         *negp = 0;
1862                         *lvalp = (unsigned long)val;
1863                 }
1864         }
1865         return 0;
1866 }
1867
1868 /*
1869  *      init may raise the set.
1870  */
1871
1872 int proc_dointvec_bset(ctl_table *table, int write, struct file *filp,
1873                         void __user *buffer, size_t *lenp, loff_t *ppos)
1874 {
1875         int op;
1876
1877         if (!capable(CAP_SYS_MODULE)) {
1878                 return -EPERM;
1879         }
1880
1881         op = (current->pid == 1) ? OP_SET : OP_AND;
1882         return do_proc_dointvec(table,write,filp,buffer,lenp,ppos,
1883                                 do_proc_dointvec_bset_conv,&op);
1884 }
1885
1886 struct do_proc_dointvec_minmax_conv_param {
1887         int *min;
1888         int *max;
1889 };
1890
1891 static int do_proc_dointvec_minmax_conv(int *negp, unsigned long *lvalp,
1892                                         int *valp,
1893                                         int write, void *data)
1894 {
1895         struct do_proc_dointvec_minmax_conv_param *param = data;
1896         if (write) {
1897                 int val = *negp ? -*lvalp : *lvalp;
1898                 if ((param->min && *param->min > val) ||
1899                     (param->max && *param->max < val))
1900                         return -EINVAL;
1901                 *valp = val;
1902         } else {
1903                 int val = *valp;
1904                 if (val < 0) {
1905                         *negp = -1;
1906                         *lvalp = (unsigned long)-val;
1907                 } else {
1908                         *negp = 0;
1909                         *lvalp = (unsigned long)val;
1910                 }
1911         }
1912         return 0;
1913 }
1914
1915 /**
1916  * proc_dointvec_minmax - read a vector of integers with min/max values
1917  * @table: the sysctl table
1918  * @write: %TRUE if this is a write to the sysctl file
1919  * @filp: the file structure
1920  * @buffer: the user buffer
1921  * @lenp: the size of the user buffer
1922  * @ppos: file position
1923  *
1924  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
1925  * values from/to the user buffer, treated as an ASCII string.
1926  *
1927  * This routine will ensure the values are within the range specified by
1928  * table->extra1 (min) and table->extra2 (max).
1929  *
1930  * Returns 0 on success.
1931  */
1932 int proc_dointvec_minmax(ctl_table *table, int write, struct file *filp,
1933                   void __user *buffer, size_t *lenp, loff_t *ppos)
1934 {
1935         struct do_proc_dointvec_minmax_conv_param param = {
1936                 .min = (int *) table->extra1,
1937                 .max = (int *) table->extra2,
1938         };
1939         return do_proc_dointvec(table, write, filp, buffer, lenp, ppos,
1940                                 do_proc_dointvec_minmax_conv, &param);
1941 }
1942
1943 static int do_proc_doulongvec_minmax(ctl_table *table, int write,
1944                                      struct file *filp,
1945                                      void __user *buffer,
1946                                      size_t *lenp, loff_t *ppos,
1947                                      unsigned long convmul,
1948                                      unsigned long convdiv)
1949 {
1950 #define TMPBUFLEN 21
1951         unsigned long *i, *min, *max, val;
1952         int vleft, first=1, neg;
1953         size_t len, left;
1954         char buf[TMPBUFLEN], *p;
1955         char __user *s = buffer;
1956
1957         if (!table->data || !table->maxlen || !*lenp ||
1958             (*ppos && !write)) {
1959                 *lenp = 0;
1960                 return 0;
1961         }
1962
1963         i = (unsigned long *) table->data;
1964         min = (unsigned long *) table->extra1;
1965         max = (unsigned long *) table->extra2;
1966         vleft = table->maxlen / sizeof(unsigned long);
1967         left = *lenp;
1968
1969         for (; left && vleft--; i++, min++, max++, first=0) {
1970                 if (write) {
1971                         while (left) {
1972                                 char c;
1973                                 if (get_user(c, s))
1974                                         return -EFAULT;
1975                                 if (!isspace(c))
1976                                         break;
1977                                 left--;
1978                                 s++;
1979                         }
1980                         if (!left)
1981                                 break;
1982                         neg = 0;
1983                         len = left;
1984                         if (len > TMPBUFLEN-1)
1985                                 len = TMPBUFLEN-1;
1986                         if (copy_from_user(buf, s, len))
1987                                 return -EFAULT;
1988                         buf[len] = 0;
1989                         p = buf;
1990                         if (*p == '-' && left > 1) {
1991                                 neg = 1;
1992                                 left--, p++;
1993                         }
1994                         if (*p < '0' || *p > '9')
1995                                 break;
1996                         val = simple_strtoul(p, &p, 0) * convmul / convdiv ;
1997                         len = p-buf;
1998                         if ((len < left) && *p && !isspace(*p))
1999                                 break;
2000                         if (neg)
2001                                 val = -val;
2002                         s += len;
2003                         left -= len;
2004
2005                         if(neg)
2006                                 continue;
2007                         if ((min && val < *min) || (max && val > *max))
2008                                 continue;
2009                         *i = val;
2010                 } else {
2011                         p = buf;
2012                         if (!first)
2013                                 *p++ = '\t';
2014                         sprintf(p, "%lu", convdiv * (*i) / convmul);
2015                         len = strlen(buf);
2016                         if (len > left)
2017                                 len = left;
2018                         if(copy_to_user(s, buf, len))
2019                                 return -EFAULT;
2020                         left -= len;
2021                         s += len;
2022                 }
2023         }
2024
2025         if (!write && !first && left) {
2026                 if(put_user('\n', s))
2027                         return -EFAULT;
2028                 left--, s++;
2029         }
2030         if (write) {
2031                 while (left) {
2032                         char c;
2033                         if (get_user(c, s++))
2034                                 return -EFAULT;
2035                         if (!isspace(c))
2036                                 break;
2037                         left--;
2038                 }
2039         }
2040         if (write && first)
2041                 return -EINVAL;
2042         *lenp -= left;
2043         *ppos += *lenp;
2044         return 0;
2045 #undef TMPBUFLEN
2046 }
2047
2048 /**
2049  * proc_doulongvec_minmax - read a vector of long integers with min/max values
2050  * @table: the sysctl table
2051  * @write: %TRUE if this is a write to the sysctl file
2052  * @filp: the file structure
2053  * @buffer: the user buffer
2054  * @lenp: the size of the user buffer
2055  * @ppos: file position
2056  *
2057  * Reads/writes up to table->maxlen/sizeof(unsigned long) unsigned long
2058  * values from/to the user buffer, treated as an ASCII string.
2059  *
2060  * This routine will ensure the values are within the range specified by
2061  * table->extra1 (min) and table->extra2 (max).
2062  *
2063  * Returns 0 on success.
2064  */
2065 int proc_doulongvec_minmax(ctl_table *table, int write, struct file *filp,
2066                            void __user *buffer, size_t *lenp, loff_t *ppos)
2067 {
2068     return do_proc_doulongvec_minmax(table, write, filp, buffer, lenp, ppos, 1l, 1l);
2069 }
2070
2071 /**
2072  * proc_doulongvec_ms_jiffies_minmax - read a vector of millisecond values with min/max values
2073  * @table: the sysctl table
2074  * @write: %TRUE if this is a write to the sysctl file
2075  * @filp: the file structure
2076  * @buffer: the user buffer
2077  * @lenp: the size of the user buffer
2078  * @ppos: file position
2079  *
2080  * Reads/writes up to table->maxlen/sizeof(unsigned long) unsigned long
2081  * values from/to the user buffer, treated as an ASCII string. The values
2082  * are treated as milliseconds, and converted to jiffies when they are stored.
2083  *
2084  * This routine will ensure the values are within the range specified by
2085  * table->extra1 (min) and table->extra2 (max).
2086  *
2087  * Returns 0 on success.
2088  */
2089 int proc_doulongvec_ms_jiffies_minmax(ctl_table *table, int write,
2090                                       struct file *filp,
2091                                       void __user *buffer,
2092                                       size_t *lenp, loff_t *ppos)
2093 {
2094     return do_proc_doulongvec_minmax(table, write, filp, buffer,
2095                                      lenp, ppos, HZ, 1000l);
2096 }
2097
2098
2099 static int do_proc_dointvec_jiffies_conv(int *negp, unsigned long *lvalp,
2100                                          int *valp,
2101                                          int write, void *data)
2102 {
2103         if (write) {
2104                 if (*lvalp > LONG_MAX / HZ)
2105                         return 1;
2106                 *valp = *negp ? -(*lvalp*HZ) : (*lvalp*HZ);
2107         } else {
2108                 int val = *valp;
2109                 unsigned long lval;
2110                 if (val < 0) {
2111                         *negp = -1;
2112                         lval = (unsigned long)-val;
2113                 } else {
2114                         *negp = 0;
2115                         lval = (unsigned long)val;
2116                 }
2117                 *lvalp = lval / HZ;
2118         }
2119         return 0;
2120 }
2121
2122 static int do_proc_dointvec_userhz_jiffies_conv(int *negp, unsigned long *lvalp,
2123                                                 int *valp,
2124                                                 int write, void *data)
2125 {
2126         if (write) {
2127                 if (USER_HZ < HZ && *lvalp > (LONG_MAX / HZ) * USER_HZ)
2128                         return 1;
2129                 *valp = clock_t_to_jiffies(*negp ? -*lvalp : *lvalp);
2130         } else {
2131                 int val = *valp;
2132                 unsigned long lval;
2133                 if (val < 0) {
2134                         *negp = -1;
2135                         lval = (unsigned long)-val;
2136                 } else {
2137                         *negp = 0;
2138                         lval = (unsigned long)val;
2139                 }
2140                 *lvalp = jiffies_to_clock_t(lval);
2141         }
2142         return 0;
2143 }
2144
2145 static int do_proc_dointvec_ms_jiffies_conv(int *negp, unsigned long *lvalp,
2146                                             int *valp,
2147                                             int write, void *data)
2148 {
2149         if (write) {
2150                 *valp = msecs_to_jiffies(*negp ? -*lvalp : *lvalp);
2151         } else {
2152                 int val = *valp;
2153                 unsigned long lval;
2154                 if (val < 0) {
2155                         *negp = -1;
2156                         lval = (unsigned long)-val;
2157                 } else {
2158                         *negp = 0;
2159                         lval = (unsigned long)val;
2160                 }
2161                 *lvalp = jiffies_to_msecs(lval);
2162         }
2163         return 0;
2164 }
2165
2166 /**
2167  * proc_dointvec_jiffies - read a vector of integers as seconds
2168  * @table: the sysctl table
2169  * @write: %TRUE if this is a write to the sysctl file
2170  * @filp: the file structure
2171  * @buffer: the user buffer
2172  * @lenp: the size of the user buffer
2173  * @ppos: file position
2174  *
2175  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2176  * values from/to the user buffer, treated as an ASCII string.
2177  * The values read are assumed to be in seconds, and are converted into
2178  * jiffies.
2179  *
2180  * Returns 0 on success.
2181  */
2182 int proc_dointvec_jiffies(ctl_table *table, int write, struct file *filp,
2183                           void __user *buffer, size_t *lenp, loff_t *ppos)
2184 {
2185     return do_proc_dointvec(table,write,filp,buffer,lenp,ppos,
2186                             do_proc_dointvec_jiffies_conv,NULL);
2187 }
2188
2189 /**
2190  * proc_dointvec_userhz_jiffies - read a vector of integers as 1/USER_HZ seconds
2191  * @table: the sysctl table
2192  * @write: %TRUE if this is a write to the sysctl file
2193  * @filp: the file structure
2194  * @buffer: the user buffer
2195  * @lenp: the size of the user buffer
2196  * @ppos: pointer to the file position
2197  *
2198  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2199  * values from/to the user buffer, treated as an ASCII string.
2200  * The values read are assumed to be in 1/USER_HZ seconds, and
2201  * are converted into jiffies.
2202  *
2203  * Returns 0 on success.
2204  */
2205 int proc_dointvec_userhz_jiffies(ctl_table *table, int write, struct file *filp,
2206                                  void __user *buffer, size_t *lenp, loff_t *ppos)
2207 {
2208     return do_proc_dointvec(table,write,filp,buffer,lenp,ppos,
2209                             do_proc_dointvec_userhz_jiffies_conv,NULL);
2210 }
2211
2212 /**
2213  * proc_dointvec_ms_jiffies - read a vector of integers as 1 milliseconds
2214  * @table: the sysctl table
2215  * @write: %TRUE if this is a write to the sysctl file
2216  * @filp: the file structure
2217  * @buffer: the user buffer
2218  * @lenp: the size of the user buffer
2219  * @ppos: file position
2220  * @ppos: the current position in the file
2221  *
2222  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2223  * values from/to the user buffer, treated as an ASCII string.
2224  * The values read are assumed to be in 1/1000 seconds, and
2225  * are converted into jiffies.
2226  *
2227  * Returns 0 on success.
2228  */
2229 int proc_dointvec_ms_jiffies(ctl_table *table, int write, struct file *filp,
2230                              void __user *buffer, size_t *lenp, loff_t *ppos)
2231 {
2232         return do_proc_dointvec(table, write, filp, buffer, lenp, ppos,
2233                                 do_proc_dointvec_ms_jiffies_conv, NULL);
2234 }
2235
2236 #else /* CONFIG_PROC_FS */
2237
2238 int proc_dostring(ctl_table *table, int write, struct file *filp,
2239                   void __user *buffer, size_t *lenp, loff_t *ppos)
2240 {
2241         return -ENOSYS;
2242 }
2243
2244 static int proc_doutsstring(ctl_table *table, int write, struct file *filp,
2245                             void __user *buffer, size_t *lenp, loff_t *ppos)
2246 {
2247         return -ENOSYS;
2248 }
2249
2250 int proc_dointvec(ctl_table *table, int write, struct file *filp,
2251                   void __user *buffer, size_t *lenp, loff_t *ppos)
2252 {
2253         return -ENOSYS;
2254 }
2255
2256 int proc_dointvec_bset(ctl_table *table, int write, struct file *filp,
2257                         void __user *buffer, size_t *lenp, loff_t *ppos)
2258 {
2259         return -ENOSYS;
2260 }
2261
2262 int proc_dointvec_minmax(ctl_table *table, int write, struct file *filp,
2263                     void __user *buffer, size_t *lenp, loff_t *ppos)
2264 {
2265         return -ENOSYS;
2266 }
2267
2268 int proc_dointvec_jiffies(ctl_table *table, int write, struct file *filp,
2269                     void __user *buffer, size_t *lenp, loff_t *ppos)
2270 {
2271         return -ENOSYS;
2272 }
2273
2274 int proc_dointvec_userhz_jiffies(ctl_table *table, int write, struct file *filp,
2275                     void __user *buffer, size_t *lenp, loff_t *ppos)
2276 {
2277         return -ENOSYS;
2278 }
2279
2280 int proc_dointvec_ms_jiffies(ctl_table *table, int write, struct file *filp,
2281                              void __user *buffer, size_t *lenp, loff_t *ppos)
2282 {
2283         return -ENOSYS;
2284 }
2285
2286 int proc_doulongvec_minmax(ctl_table *table, int write, struct file *filp,
2287                     void __user *buffer, size_t *lenp, loff_t *ppos)
2288 {
2289         return -ENOSYS;
2290 }
2291
2292 int proc_doulongvec_ms_jiffies_minmax(ctl_table *table, int write,
2293                                       struct file *filp,
2294                                       void __user *buffer,
2295                                       size_t *lenp, loff_t *ppos)
2296 {
2297     return -ENOSYS;
2298 }
2299
2300
2301 #endif /* CONFIG_PROC_FS */
2302
2303
2304 /*
2305  * General sysctl support routines
2306  */
2307
2308 /* The generic string strategy routine: */
2309 int sysctl_string(ctl_table *table, int __user *name, int nlen,
2310                   void __user *oldval, size_t __user *oldlenp,
2311                   void __user *newval, size_t newlen, void **context)
2312 {
2313         if (!table->data || !table->maxlen)
2314                 return -ENOTDIR;
2315
2316         if (oldval && oldlenp) {
2317                 size_t bufsize;
2318                 if (get_user(bufsize, oldlenp))
2319                         return -EFAULT;
2320                 if (bufsize) {
2321                         size_t len = strlen(table->data), copied;
2322
2323                         /* This shouldn't trigger for a well-formed sysctl */
2324                         if (len > table->maxlen)
2325                                 len = table->maxlen;
2326
2327                         /* Copy up to a max of bufsize-1 bytes of the string */
2328                         copied = (len >= bufsize) ? bufsize - 1 : len;
2329
2330                         if (copy_to_user(oldval, table->data, copied) ||
2331                             put_user(0, (char __user *)(oldval + copied)))
2332                                 return -EFAULT;
2333                         if (put_user(len, oldlenp))
2334                                 return -EFAULT;
2335                 }
2336         }
2337         if (newval && newlen) {
2338                 size_t len = newlen;
2339                 if (len > table->maxlen)
2340                         len = table->maxlen;
2341                 if(copy_from_user(table->data, newval, len))
2342                         return -EFAULT;
2343                 if (len == table->maxlen)
2344                         len--;
2345                 ((char *) table->data)[len] = 0;
2346         }
2347         return 1;
2348 }
2349
2350 /*
2351  * This function makes sure that all of the integers in the vector
2352  * are between the minimum and maximum values given in the arrays
2353  * table->extra1 and table->extra2, respectively.
2354  */
2355 int sysctl_intvec(ctl_table *table, int __user *name, int nlen,
2356                 void __user *oldval, size_t __user *oldlenp,
2357                 void __user *newval, size_t newlen, void **context)
2358 {
2359
2360         if (newval && newlen) {
2361                 int __user *vec = (int __user *) newval;
2362                 int *min = (int *) table->extra1;
2363                 int *max = (int *) table->extra2;
2364                 size_t length;
2365                 int i;
2366
2367                 if (newlen % sizeof(int) != 0)
2368                         return -EINVAL;
2369
2370                 if (!table->extra1 && !table->extra2)
2371                         return 0;
2372
2373                 if (newlen > table->maxlen)
2374                         newlen = table->maxlen;
2375                 length = newlen / sizeof(int);
2376
2377                 for (i = 0; i < length; i++) {
2378                         int value;
2379                         if (get_user(value, vec + i))
2380                                 return -EFAULT;
2381                         if (min && value < min[i])
2382                                 return -EINVAL;
2383                         if (max && value > max[i])
2384                                 return -EINVAL;
2385                 }
2386         }
2387         return 0;
2388 }
2389
2390 /* Strategy function to convert jiffies to seconds */
2391 int sysctl_jiffies(ctl_table *table, int __user *name, int nlen,
2392                 void __user *oldval, size_t __user *oldlenp,
2393                 void __user *newval, size_t newlen, void **context)
2394 {
2395         if (oldval) {
2396                 size_t olen;
2397                 if (oldlenp) {
2398                         if (get_user(olen, oldlenp))
2399                                 return -EFAULT;
2400                         if (olen!=sizeof(int))
2401                                 return -EINVAL;
2402                 }
2403                 if (put_user(*(int *)(table->data)/HZ, (int __user *)oldval) ||
2404                     (oldlenp && put_user(sizeof(int),oldlenp)))
2405                         return -EFAULT;
2406         }
2407         if (newval && newlen) {
2408                 int new;
2409                 if (newlen != sizeof(int))
2410                         return -EINVAL;
2411                 if (get_user(new, (int __user *)newval))
2412                         return -EFAULT;
2413                 *(int *)(table->data) = new*HZ;
2414         }
2415         return 1;
2416 }
2417
2418 /* Strategy function to convert jiffies to seconds */
2419 int sysctl_ms_jiffies(ctl_table *table, int __user *name, int nlen,
2420                 void __user *oldval, size_t __user *oldlenp,
2421                 void __user *newval, size_t newlen, void **context)
2422 {
2423         if (oldval) {
2424                 size_t olen;
2425                 if (oldlenp) {
2426                         if (get_user(olen, oldlenp))
2427                                 return -EFAULT;
2428                         if (olen!=sizeof(int))
2429                                 return -EINVAL;
2430                 }
2431                 if (put_user(jiffies_to_msecs(*(int *)(table->data)), (int __user *)oldval) ||
2432                     (oldlenp && put_user(sizeof(int),oldlenp)))
2433                         return -EFAULT;
2434         }
2435         if (newval && newlen) {
2436                 int new;
2437                 if (newlen != sizeof(int))
2438                         return -EINVAL;
2439                 if (get_user(new, (int __user *)newval))
2440                         return -EFAULT;
2441                 *(int *)(table->data) = msecs_to_jiffies(new);
2442         }
2443         return 1;
2444 }
2445
2446 #else /* CONFIG_SYSCTL */
2447
2448
2449 asmlinkage long sys_sysctl(struct __sysctl_args __user *args)
2450 {
2451         return -ENOSYS;
2452 }
2453
2454 int sysctl_string(ctl_table *table, int __user *name, int nlen,
2455                   void __user *oldval, size_t __user *oldlenp,
2456                   void __user *newval, size_t newlen, void **context)
2457 {
2458         return -ENOSYS;
2459 }
2460
2461 int sysctl_intvec(ctl_table *table, int __user *name, int nlen,
2462                 void __user *oldval, size_t __user *oldlenp,
2463                 void __user *newval, size_t newlen, void **context)
2464 {
2465         return -ENOSYS;
2466 }
2467
2468 int sysctl_jiffies(ctl_table *table, int __user *name, int nlen,
2469                 void __user *oldval, size_t __user *oldlenp,
2470                 void __user *newval, size_t newlen, void **context)
2471 {
2472         return -ENOSYS;
2473 }
2474
2475 int sysctl_ms_jiffies(ctl_table *table, int __user *name, int nlen,
2476                 void __user *oldval, size_t __user *oldlenp,
2477                 void __user *newval, size_t newlen, void **context)
2478 {
2479         return -ENOSYS;
2480 }
2481
2482 int proc_dostring(ctl_table *table, int write, struct file *filp,
2483                   void __user *buffer, size_t *lenp, loff_t *ppos)
2484 {
2485         return -ENOSYS;
2486 }
2487
2488 int proc_dointvec(ctl_table *table, int write, struct file *filp,
2489                   void __user *buffer, size_t *lenp, loff_t *ppos)
2490 {
2491         return -ENOSYS;
2492 }
2493
2494 int proc_dointvec_bset(ctl_table *table, int write, struct file *filp,
2495                         void __user *buffer, size_t *lenp, loff_t *ppos)
2496 {
2497         return -ENOSYS;
2498 }
2499
2500 int proc_dointvec_minmax(ctl_table *table, int write, struct file *filp,
2501                     void __user *buffer, size_t *lenp, loff_t *ppos)
2502 {
2503         return -ENOSYS;
2504 }
2505
2506 int proc_dointvec_jiffies(ctl_table *table, int write, struct file *filp,
2507                           void __user *buffer, size_t *lenp, loff_t *ppos)
2508 {
2509         return -ENOSYS;
2510 }
2511
2512 int proc_dointvec_userhz_jiffies(ctl_table *table, int write, struct file *filp,
2513                           void __user *buffer, size_t *lenp, loff_t *ppos)
2514 {
2515         return -ENOSYS;
2516 }
2517
2518 int proc_dointvec_ms_jiffies(ctl_table *table, int write, struct file *filp,
2519                              void __user *buffer, size_t *lenp, loff_t *ppos)
2520 {
2521         return -ENOSYS;
2522 }
2523
2524 int proc_doulongvec_minmax(ctl_table *table, int write, struct file *filp,
2525                     void __user *buffer, size_t *lenp, loff_t *ppos)
2526 {
2527         return -ENOSYS;
2528 }
2529
2530 int proc_doulongvec_ms_jiffies_minmax(ctl_table *table, int write,
2531                                       struct file *filp,
2532                                       void __user *buffer,
2533                                       size_t *lenp, loff_t *ppos)
2534 {
2535     return -ENOSYS;
2536 }
2537
2538 struct ctl_table_header * register_sysctl_table(ctl_table * table,
2539                                                 int insert_at_head)
2540 {
2541         return NULL;
2542 }
2543
2544 void unregister_sysctl_table(struct ctl_table_header * table)
2545 {
2546 }
2547
2548 #endif /* CONFIG_SYSCTL */
2549
2550 /*
2551  * No sense putting this after each symbol definition, twice,
2552  * exception granted :-)
2553  */
2554 EXPORT_SYMBOL(proc_dointvec);
2555 EXPORT_SYMBOL(proc_dointvec_jiffies);
2556 EXPORT_SYMBOL(proc_dointvec_minmax);
2557 EXPORT_SYMBOL(proc_dointvec_userhz_jiffies);
2558 EXPORT_SYMBOL(proc_dointvec_ms_jiffies);
2559 EXPORT_SYMBOL(proc_dostring);
2560 EXPORT_SYMBOL(proc_doulongvec_minmax);
2561 EXPORT_SYMBOL(proc_doulongvec_ms_jiffies_minmax);
2562 EXPORT_SYMBOL(register_sysctl_table);
2563 EXPORT_SYMBOL(sysctl_intvec);
2564 EXPORT_SYMBOL(sysctl_jiffies);
2565 EXPORT_SYMBOL(sysctl_ms_jiffies);
2566 EXPORT_SYMBOL(sysctl_string);
2567 EXPORT_SYMBOL(unregister_sysctl_table);