kernel/sysctl.c

   1 /*
   2  * sysctl.c: General linux system control interface
   3  *
   4  * Begun 24 March 1995, Stephen Tweedie
   5  * Added /proc support, Dec 1995
   6  * Added bdflush entry and intvec min/max checking, 2/23/96, Tom Dyas.
   7  * Added hooks for /proc/sys/net (minor, minor patch), 96/4/1, Mike Shaver.
   8  * Added kernel/java-{interpreter,appletviewer}, 96/5/10, Mike Shaver.
   9  * Dynamic registration fixes, Stephen Tweedie.
  10  * Added kswapd-interval, ctrl-alt-del, printk stuff, 1/8/97, Chris Horn.
  11  * Made sysctl support optional via CONFIG_SYSCTL, 1/10/97, Chris
  12  *  Horn.
  13  * Added proc_doulongvec_ms_jiffies_minmax, 09/08/99, Carlos H. Bauer.
  14  * Added proc_doulongvec_minmax, 09/08/99, Carlos H. Bauer.
  15  * Changed linked lists to use list.h instead of lists.h, 02/24/00, Bill
  16  *  Wendling.
  17  * The list_for_each() macro wasn't appropriate for the sysctl loop.
  18  *  Removed it and replaced it with older style, 03/23/00, Bill Wendling
  19  */
  20
  21 #include <linux/config.h>
  22 #include <linux/module.h>
  23 #include <linux/mm.h>
  24 #include <linux/swap.h>
  25 #include <linux/slab.h>
  26 #include <linux/sysctl.h>
  27 #include <linux/proc_fs.h>
  28 #include <linux/capability.h>
  29 #include <linux/ctype.h>
  30 #include <linux/utsname.h>
  31 #include <linux/capability.h>
  32 #include <linux/smp_lock.h>
  33 #include <linux/init.h>
  34 #include <linux/kernel.h>
  35 #include <linux/kobject.h>
  36 #include <linux/net.h>
  37 #include <linux/sysrq.h>
  38 #include <linux/highuid.h>
  39 #include <linux/writeback.h>
  40 #include <linux/hugetlb.h>
  41 #include <linux/security.h>
  42 #include <linux/initrd.h>
  43 #include <linux/times.h>
  44 #include <linux/limits.h>
  45 #include <linux/dcache.h>
  46 #include <linux/syscalls.h>
  47
  48 #include <asm/uaccess.h>
  49 #include <asm/processor.h>
  50
  51 #ifdef CONFIG_ROOT_NFS
  52 #include <linux/nfs_fs.h>
  53 #endif
  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 max_threads;
  62 extern int sysrq_enabled;
  63 extern int core_uses_pid;
  64 extern int suid_dumpable;
  65 extern char core_pattern[];
  66 extern int cad_pid;
  67 extern int pid_max;
  68 extern int min_free_kbytes;
  69 extern int printk_ratelimit_jiffies;
  70 extern int printk_ratelimit_burst;
  71 extern int pid_max_min, pid_max_max;
  72 extern int sysctl_drop_caches;
  73 extern int percpu_pagelist_fraction;
  74
  75 #if defined(CONFIG_X86_LOCAL_APIC) && defined(CONFIG_X86)
  76 int unknown_nmi_panic;
  77 extern int proc_unknown_nmi_panic(ctl_table *, int, struct file *,
  78                                   void __user *, size_t *, loff_t *);
  79 #endif
  80
  81 /* this is needed for the proc_dointvec_minmax for [fs_]overflow UID and GID */
  82 static int maxolduid = 65535;
  83 static int minolduid;
  84 static int min_percpu_pagelist_fract = 8;
  85
  86 static int ngroups_max = NGROUPS_MAX;
  87
  88 #ifdef CONFIG_KMOD
  89 extern char modprobe_path[];
  90 #endif
  91 #ifdef CONFIG_CHR_DEV_SG
  92 extern int sg_big_buff;
  93 #endif
  94 #ifdef CONFIG_SYSVIPC
  95 extern size_t shm_ctlmax;
  96 extern size_t shm_ctlall;
  97 extern int shm_ctlmni;
  98 extern int msg_ctlmax;
  99 extern int msg_ctlmnb;
 100 extern int msg_ctlmni;
 101 extern int sem_ctls[];
 102 #endif
 103
 104 #ifdef __sparc__
 105 extern char reboot_command [];
 106 extern int stop_a_enabled;
 107 extern int scons_pwroff;
 108 #endif
 109
 110 #ifdef __hppa__
 111 extern int pwrsw_enabled;
 112 extern int unaligned_enabled;
 113 #endif
 114
 115 #ifdef CONFIG_S390
 116 #ifdef CONFIG_MATHEMU
 117 extern int sysctl_ieee_emulation_warnings;
 118 #endif
 119 extern int sysctl_userprocess_debug;
 120 extern int spin_retry;
 121 #endif
 122
 123 extern int sysctl_hz_timer;
 124
 125 #ifdef CONFIG_BSD_PROCESS_ACCT
 126 extern int acct_parm[];
 127 #endif
 128
 129 int randomize_va_space = 1;
 130
 131 static int parse_table(int __user *, int, void __user *, size_t __user *, void __user *, size_t,
 132                        ctl_table *, void **);
 133 static int proc_doutsstring(ctl_table *table, int write, struct file *filp,
 134                   void __user *buffer, size_t *lenp, loff_t *ppos);
 135
 136 static ctl_table root_table[];
 137 static struct ctl_table_header root_table_header =
 138         { root_table, LIST_HEAD_INIT(root_table_header.ctl_entry) };
 139
 140 static ctl_table kern_table[];
 141 static ctl_table vm_table[];
 142 static ctl_table proc_table[];
 143 static ctl_table fs_table[];
 144 static ctl_table debug_table[];
 145 static ctl_table dev_table[];
 146 extern ctl_table random_table[];
 147 #ifdef CONFIG_UNIX98_PTYS
 148 extern ctl_table pty_table[];
 149 #endif
 150 #ifdef CONFIG_INOTIFY
 151 extern ctl_table inotify_table[];
 152 #endif
 153
 154 #ifdef HAVE_ARCH_PICK_MMAP_LAYOUT
 155 int sysctl_legacy_va_layout;
 156 #endif
 157
 158 /* /proc declarations: */
 159
 160 #ifdef CONFIG_PROC_FS
 161
 162 static ssize_t proc_readsys(struct file *, char __user *, size_t, loff_t *);
 163 static ssize_t proc_writesys(struct file *, const char __user *, size_t, loff_t *);
 164 static int proc_opensys(struct inode *, struct file *);
 165
 166 struct file_operations proc_sys_file_operations = {
 167         .open           = proc_opensys,
 168         .read           = proc_readsys,
 169         .write          = proc_writesys,
 170 };
 171
 172 extern struct proc_dir_entry *proc_sys_root;
 173
 174 static void register_proc_table(ctl_table *, struct proc_dir_entry *, void *);
 175 static void unregister_proc_table(ctl_table *, struct proc_dir_entry *);
 176 #endif
 177
 178 /* The default sysctl tables: */
 179
 180 static ctl_table root_table[] = {
 181         {
 182                 .ctl_name       = CTL_KERN,
 183                 .procname       = "kernel",
 184                 .mode           = 0555,
 185                 .child          = kern_table,
 186         },
 187         {
 188                 .ctl_name       = CTL_VM,
 189                 .procname       = "vm",
 190                 .mode           = 0555,
 191                 .child          = vm_table,
 192         },
 193 #ifdef CONFIG_NET
 194         {
 195                 .ctl_name       = CTL_NET,
 196                 .procname       = "net",
 197                 .mode           = 0555,
 198                 .child          = net_table,
 199         },
 200 #endif
 201         {
 202                 .ctl_name       = CTL_PROC,
 203                 .procname       = "proc",
 204                 .mode           = 0555,
 205                 .child          = proc_table,
 206         },
 207         {
 208                 .ctl_name       = CTL_FS,
 209                 .procname       = "fs",
 210                 .mode           = 0555,
 211                 .child          = fs_table,
 212         },
 213         {
 214                 .ctl_name       = CTL_DEBUG,
 215                 .procname       = "debug",
 216                 .mode           = 0555,
 217                 .child          = debug_table,
 218         },
 219         {
 220                 .ctl_name       = CTL_DEV,
 221                 .procname       = "dev",
 222                 .mode           = 0555,
 223                 .child          = dev_table,
 224         },
 225
 226         { .ctl_name = 0 }
 227 };
 228
 229 static ctl_table kern_table[] = {
 230         {
 231                 .ctl_name       = KERN_OSTYPE,
 232                 .procname       = "ostype",
 233                 .data           = system_utsname.sysname,
 234                 .maxlen         = sizeof(system_utsname.sysname),
 235                 .mode           = 0444,
 236                 .proc_handler   = &proc_doutsstring,
 237                 .strategy       = &sysctl_string,
 238         },
 239         {
 240                 .ctl_name       = KERN_OSRELEASE,
 241                 .procname       = "osrelease",
 242                 .data           = system_utsname.release,
 243                 .maxlen         = sizeof(system_utsname.release),
 244                 .mode           = 0444,
 245                 .proc_handler   = &proc_doutsstring,
 246                 .strategy       = &sysctl_string,
 247         },
 248         {
 249                 .ctl_name       = KERN_VERSION,
 250                 .procname       = "version",
 251                 .data           = system_utsname.version,
 252                 .maxlen         = sizeof(system_utsname.version),
 253                 .mode           = 0444,
 254                 .proc_handler   = &proc_doutsstring,
 255                 .strategy       = &sysctl_string,
 256         },
 257         {
 258                 .ctl_name       = KERN_NODENAME,
 259                 .procname       = "hostname",
 260                 .data           = system_utsname.nodename,
 261                 .maxlen         = sizeof(system_utsname.nodename),
 262                 .mode           = 0644,
 263                 .proc_handler   = &proc_doutsstring,
 264                 .strategy       = &sysctl_string,
 265         },
 266         {
 267                 .ctl_name       = KERN_DOMAINNAME,
 268                 .procname       = "domainname",
 269                 .data           = system_utsname.domainname,
 270                 .maxlen         = sizeof(system_utsname.domainname),
 271                 .mode           = 0644,
 272                 .proc_handler   = &proc_doutsstring,
 273                 .strategy       = &sysctl_string,
 274         },
 275         {
 276                 .ctl_name       = KERN_PANIC,
 277                 .procname       = "panic",
 278                 .data           = &panic_timeout,
 279                 .maxlen         = sizeof(int),
 280                 .mode           = 0644,
 281                 .proc_handler   = &proc_dointvec,
 282         },
 283         {
 284                 .ctl_name       = KERN_CORE_USES_PID,
 285                 .procname       = "core_uses_pid",
 286                 .data           = &core_uses_pid,
 287                 .maxlen         = sizeof(int),
 288                 .mode           = 0644,
 289                 .proc_handler   = &proc_dointvec,
 290         },
 291         {
 292                 .ctl_name       = KERN_CORE_PATTERN,
 293                 .procname       = "core_pattern",
 294                 .data           = core_pattern,
 295                 .maxlen         = 64,
 296                 .mode           = 0644,
 297                 .proc_handler   = &proc_dostring,
 298                 .strategy       = &sysctl_string,
 299         },
 300         {
 301                 .ctl_name       = KERN_TAINTED,
 302                 .procname       = "tainted",
 303                 .data           = &tainted,
 304                 .maxlen         = sizeof(int),
 305                 .mode           = 0444,
 306                 .proc_handler   = &proc_dointvec,
 307         },
 308         {
 309                 .ctl_name       = KERN_CAP_BSET,
 310                 .procname       = "cap-bound",
 311                 .data           = &cap_bset,
 312                 .maxlen         = sizeof(kernel_cap_t),
 313                 .mode           = 0600,
 314                 .proc_handler   = &proc_dointvec_bset,
 315         },
 316 #ifdef CONFIG_BLK_DEV_INITRD
 317         {
 318                 .ctl_name       = KERN_REALROOTDEV,
 319                 .procname       = "real-root-dev",
 320                 .data           = &real_root_dev,
 321                 .maxlen         = sizeof(int),
 322                 .mode           = 0644,
 323                 .proc_handler   = &proc_dointvec,
 324         },
 325 #endif
 326 #ifdef __sparc__
 327         {
 328                 .ctl_name       = KERN_SPARC_REBOOT,
 329                 .procname       = "reboot-cmd",
 330                 .data           = reboot_command,
 331                 .maxlen         = 256,
 332                 .mode           = 0644,
 333                 .proc_handler   = &proc_dostring,
 334                 .strategy       = &sysctl_string,
 335         },
 336         {
 337                 .ctl_name       = KERN_SPARC_STOP_A,
 338                 .procname       = "stop-a",
 339                 .data           = &stop_a_enabled,
 340                 .maxlen         = sizeof (int),
 341                 .mode           = 0644,
 342                 .proc_handler   = &proc_dointvec,
 343         },
 344         {
 345                 .ctl_name       = KERN_SPARC_SCONS_PWROFF,
 346                 .procname       = "scons-poweroff",
 347                 .data           = &scons_pwroff,
 348                 .maxlen         = sizeof (int),
 349                 .mode           = 0644,
 350                 .proc_handler   = &proc_dointvec,
 351         },
 352 #endif
 353 #ifdef __hppa__
 354         {
 355                 .ctl_name       = KERN_HPPA_PWRSW,
 356                 .procname       = "soft-power",
 357                 .data           = &pwrsw_enabled,
 358                 .maxlen         = sizeof (int),
 359                 .mode           = 0644,
 360                 .proc_handler   = &proc_dointvec,
 361         },
 362         {
 363                 .ctl_name       = KERN_HPPA_UNALIGNED,
 364                 .procname       = "unaligned-trap",
 365                 .data           = &unaligned_enabled,
 366                 .maxlen         = sizeof (int),
 367                 .mode           = 0644,
 368                 .proc_handler   = &proc_dointvec,
 369         },
 370 #endif
 371         {
 372                 .ctl_name       = KERN_CTLALTDEL,
 373                 .procname       = "ctrl-alt-del",
 374                 .data           = &C_A_D,
 375                 .maxlen         = sizeof(int),
 376                 .mode           = 0644,
 377                 .proc_handler   = &proc_dointvec,
 378         },
 379         {
 380                 .ctl_name       = KERN_PRINTK,
 381                 .procname       = "printk",
 382                 .data           = &console_loglevel,
 383                 .maxlen         = 4*sizeof(int),
 384                 .mode           = 0644,
 385                 .proc_handler   = &proc_dointvec,
 386         },
 387 #ifdef CONFIG_KMOD
 388         {
 389                 .ctl_name       = KERN_MODPROBE,
 390                 .procname       = "modprobe",
 391                 .data           = &modprobe_path,
 392                 .maxlen         = KMOD_PATH_LEN,
 393                 .mode           = 0644,
 394                 .proc_handler   = &proc_dostring,
 395                 .strategy       = &sysctl_string,
 396         },
 397 #endif
 398 #ifdef CONFIG_HOTPLUG
 399         {
 400                 .ctl_name       = KERN_HOTPLUG,
 401                 .procname       = "hotplug",
 402                 .data           = &uevent_helper,
 403                 .maxlen         = UEVENT_HELPER_PATH_LEN,
 404                 .mode           = 0644,
 405                 .proc_handler   = &proc_dostring,
 406                 .strategy       = &sysctl_string,
 407         },
 408 #endif
 409 #ifdef CONFIG_CHR_DEV_SG
 410         {
 411                 .ctl_name       = KERN_SG_BIG_BUFF,
 412                 .procname       = "sg-big-buff",
 413                 .data           = &sg_big_buff,
 414                 .maxlen         = sizeof (int),
 415                 .mode           = 0444,
 416                 .proc_handler   = &proc_dointvec,
 417         },
 418 #endif
 419 #ifdef CONFIG_BSD_PROCESS_ACCT
 420         {
 421                 .ctl_name       = KERN_ACCT,
 422                 .procname       = "acct",
 423                 .data           = &acct_parm,
 424                 .maxlen         = 3*sizeof(int),
 425                 .mode           = 0644,
 426                 .proc_handler   = &proc_dointvec,
 427         },
 428 #endif
 429 #ifdef CONFIG_SYSVIPC
 430         {
 431                 .ctl_name       = KERN_SHMMAX,
 432                 .procname       = "shmmax",
 433                 .data           = &shm_ctlmax,
 434                 .maxlen         = sizeof (size_t),
 435                 .mode           = 0644,
 436                 .proc_handler   = &proc_doulongvec_minmax,
 437         },
 438         {
 439                 .ctl_name       = KERN_SHMALL,
 440                 .procname       = "shmall",
 441                 .data           = &shm_ctlall,
 442                 .maxlen         = sizeof (size_t),
 443                 .mode           = 0644,
 444                 .proc_handler   = &proc_doulongvec_minmax,
 445         },
 446         {
 447                 .ctl_name       = KERN_SHMMNI,
 448                 .procname       = "shmmni",
 449                 .data           = &shm_ctlmni,
 450                 .maxlen         = sizeof (int),
 451                 .mode           = 0644,
 452                 .proc_handler   = &proc_dointvec,
 453         },
 454         {
 455                 .ctl_name       = KERN_MSGMAX,
 456                 .procname       = "msgmax",
 457                 .data           = &msg_ctlmax,
 458                 .maxlen         = sizeof (int),
 459                 .mode           = 0644,
 460                 .proc_handler   = &proc_dointvec,
 461         },
 462         {
 463                 .ctl_name       = KERN_MSGMNI,
 464                 .procname       = "msgmni",
 465                 .data           = &msg_ctlmni,
 466                 .maxlen         = sizeof (int),
 467                 .mode           = 0644,
 468                 .proc_handler   = &proc_dointvec,
 469         },
 470         {
 471                 .ctl_name       = KERN_MSGMNB,
 472                 .procname       =  "msgmnb",
 473                 .data           = &msg_ctlmnb,
 474                 .maxlen         = sizeof (int),
 475                 .mode           = 0644,
 476                 .proc_handler   = &proc_dointvec,
 477         },
 478         {
 479                 .ctl_name       = KERN_SEM,
 480                 .procname       = "sem",
 481                 .data           = &sem_ctls,
 482                 .maxlen         = 4*sizeof (int),
 483                 .mode           = 0644,
 484                 .proc_handler   = &proc_dointvec,
 485         },
 486 #endif
 487 #ifdef CONFIG_MAGIC_SYSRQ
 488         {
 489                 .ctl_name       = KERN_SYSRQ,
 490                 .procname       = "sysrq",
 491                 .data           = &sysrq_enabled,
 492                 .maxlen         = sizeof (int),
 493                 .mode           = 0644,
 494                 .proc_handler   = &proc_dointvec,
 495         },
 496 #endif
 497         {
 498                 .ctl_name       = KERN_CADPID,
 499                 .procname       = "cad_pid",
 500                 .data           = &cad_pid,
 501                 .maxlen         = sizeof (int),
 502                 .mode           = 0600,
 503                 .proc_handler   = &proc_dointvec,
 504         },
 505         {
 506                 .ctl_name       = KERN_MAX_THREADS,
 507                 .procname       = "threads-max",
 508                 .data           = &max_threads,
 509                 .maxlen         = sizeof(int),
 510                 .mode           = 0644,
 511                 .proc_handler   = &proc_dointvec,
 512         },
 513         {
 514                 .ctl_name       = KERN_RANDOM,
 515                 .procname       = "random",
 516                 .mode           = 0555,
 517                 .child          = random_table,
 518         },
 519 #ifdef CONFIG_UNIX98_PTYS
 520         {
 521                 .ctl_name       = KERN_PTY,
 522                 .procname       = "pty",
 523                 .mode           = 0555,
 524                 .child          = pty_table,
 525         },
 526 #endif
 527         {
 528                 .ctl_name       = KERN_OVERFLOWUID,
 529                 .procname       = "overflowuid",
 530                 .data           = &overflowuid,
 531                 .maxlen         = sizeof(int),
 532                 .mode           = 0644,
 533                 .proc_handler   = &proc_dointvec_minmax,
 534                 .strategy       = &sysctl_intvec,
 535                 .extra1         = &minolduid,
 536                 .extra2         = &maxolduid,
 537         },
 538         {
 539                 .ctl_name       = KERN_OVERFLOWGID,
 540                 .procname       = "overflowgid",
 541                 .data           = &overflowgid,
 542                 .maxlen         = sizeof(int),
 543                 .mode           = 0644,
 544                 .proc_handler   = &proc_dointvec_minmax,
 545                 .strategy       = &sysctl_intvec,
 546                 .extra1         = &minolduid,
 547                 .extra2         = &maxolduid,
 548         },
 549 #ifdef CONFIG_S390
 550 #ifdef CONFIG_MATHEMU
 551         {
 552                 .ctl_name       = KERN_IEEE_EMULATION_WARNINGS,
 553                 .procname       = "ieee_emulation_warnings",
 554                 .data           = &sysctl_ieee_emulation_warnings,
 555                 .maxlen         = sizeof(int),
 556                 .mode           = 0644,
 557                 .proc_handler   = &proc_dointvec,
 558         },
 559 #endif
 560 #ifdef CONFIG_NO_IDLE_HZ
 561         {
 562                 .ctl_name       = KERN_HZ_TIMER,
 563                 .procname       = "hz_timer",
 564                 .data           = &sysctl_hz_timer,
 565                 .maxlen         = sizeof(int),
 566                 .mode           = 0644,
 567                 .proc_handler   = &proc_dointvec,
 568         },
 569 #endif
 570         {
 571                 .ctl_name       = KERN_S390_USER_DEBUG_LOGGING,
 572                 .procname       = "userprocess_debug",
 573                 .data           = &sysctl_userprocess_debug,
 574                 .maxlen         = sizeof(int),
 575                 .mode           = 0644,
 576                 .proc_handler   = &proc_dointvec,
 577         },
 578 #endif
 579         {
 580                 .ctl_name       = KERN_PIDMAX,
 581                 .procname       = "pid_max",
 582                 .data           = &pid_max,
 583                 .maxlen         = sizeof (int),
 584                 .mode           = 0644,
 585                 .proc_handler   = &proc_dointvec_minmax,
 586                 .strategy       = sysctl_intvec,
 587                 .extra1         = &pid_max_min,
 588                 .extra2         = &pid_max_max,
 589         },
 590         {
 591                 .ctl_name       = KERN_PANIC_ON_OOPS,
 592                 .procname       = "panic_on_oops",
 593                 .data           = &panic_on_oops,
 594                 .maxlen         = sizeof(int),
 595                 .mode           = 0644,
 596                 .proc_handler   = &proc_dointvec,
 597         },
 598         {
 599                 .ctl_name       = KERN_PRINTK_RATELIMIT,
 600                 .procname       = "printk_ratelimit",
 601                 .data           = &printk_ratelimit_jiffies,
 602                 .maxlen         = sizeof(int),
 603                 .mode           = 0644,
 604                 .proc_handler   = &proc_dointvec_jiffies,
 605                 .strategy       = &sysctl_jiffies,
 606         },
 607         {
 608                 .ctl_name       = KERN_PRINTK_RATELIMIT_BURST,
 609                 .procname       = "printk_ratelimit_burst",
 610                 .data           = &printk_ratelimit_burst,
 611                 .maxlen         = sizeof(int),
 612                 .mode           = 0644,
 613                 .proc_handler   = &proc_dointvec,
 614         },
 615         {
 616                 .ctl_name       = KERN_NGROUPS_MAX,
 617                 .procname       = "ngroups_max",
 618                 .data           = &ngroups_max,
 619                 .maxlen         = sizeof (int),
 620                 .mode           = 0444,
 621                 .proc_handler   = &proc_dointvec,
 622         },
 623 #if defined(CONFIG_X86_LOCAL_APIC) && defined(CONFIG_X86)
 624         {
 625                 .ctl_name       = KERN_UNKNOWN_NMI_PANIC,
 626                 .procname       = "unknown_nmi_panic",
 627                 .data           = &unknown_nmi_panic,
 628                 .maxlen         = sizeof (int),
 629                 .mode           = 0644,
 630                 .proc_handler   = &proc_unknown_nmi_panic,
 631         },
 632 #endif
 633 #if defined(CONFIG_X86)
 634         {
 635                 .ctl_name       = KERN_BOOTLOADER_TYPE,
 636                 .procname       = "bootloader_type",
 637                 .data           = &bootloader_type,
 638                 .maxlen         = sizeof (int),
 639                 .mode           = 0444,
 640                 .proc_handler   = &proc_dointvec,
 641         },
 642 #endif
 643         {
 644                 .ctl_name       = KERN_RANDOMIZE,
 645                 .procname       = "randomize_va_space",
 646                 .data           = &randomize_va_space,
 647                 .maxlen         = sizeof(int),
 648                 .mode           = 0644,
 649                 .proc_handler   = &proc_dointvec,
 650         },
 651 #if defined(CONFIG_S390) && defined(CONFIG_SMP)
 652         {
 653                 .ctl_name       = KERN_SPIN_RETRY,
 654                 .procname       = "spin_retry",
 655                 .data           = &spin_retry,
 656                 .maxlen         = sizeof (int),
 657                 .mode           = 0644,
 658                 .proc_handler   = &proc_dointvec,
 659         },
 660 #endif
 661         { .ctl_name = 0 }
 662 };
 663
 664 /* Constants for minimum and maximum testing in vm_table.
 665    We use these as one-element integer vectors. */
 666 static int zero;
 667 static int one_hundred = 100;
 668
 669
 670 static ctl_table vm_table[] = {
 671         {
 672                 .ctl_name       = VM_OVERCOMMIT_MEMORY,
 673                 .procname       = "overcommit_memory",
 674                 .data           = &sysctl_overcommit_memory,
 675                 .maxlen         = sizeof(sysctl_overcommit_memory),
 676                 .mode           = 0644,
 677                 .proc_handler   = &proc_dointvec,
 678         },
 679         {
 680                 .ctl_name       = VM_OVERCOMMIT_RATIO,
 681                 .procname       = "overcommit_ratio",
 682                 .data           = &sysctl_overcommit_ratio,
 683                 .maxlen         = sizeof(sysctl_overcommit_ratio),
 684                 .mode           = 0644,
 685                 .proc_handler   = &proc_dointvec,
 686         },
 687         {
 688                 .ctl_name       = VM_PAGE_CLUSTER,
 689                 .procname       = "page-cluster",
 690                 .data           = &page_cluster,
 691                 .maxlen         = sizeof(int),
 692                 .mode           = 0644,
 693                 .proc_handler   = &proc_dointvec,
 694         },
 695         {
 696                 .ctl_name       = VM_DIRTY_BACKGROUND,
 697                 .procname       = "dirty_background_ratio",
 698                 .data           = &dirty_background_ratio,
 699                 .maxlen         = sizeof(dirty_background_ratio),
 700                 .mode           = 0644,
 701                 .proc_handler   = &proc_dointvec_minmax,
 702                 .strategy       = &sysctl_intvec,
 703                 .extra1         = &zero,
 704                 .extra2         = &one_hundred,
 705         },
 706         {
 707                 .ctl_name       = VM_DIRTY_RATIO,
 708                 .procname       = "dirty_ratio",
 709                 .data           = &vm_dirty_ratio,
 710                 .maxlen         = sizeof(vm_dirty_ratio),
 711                 .mode           = 0644,
 712                 .proc_handler   = &proc_dointvec_minmax,
 713                 .strategy       = &sysctl_intvec,
 714                 .extra1         = &zero,
 715                 .extra2         = &one_hundred,
 716         },
 717         {
 718                 .ctl_name       = VM_DIRTY_WB_CS,
 719                 .procname       = "dirty_writeback_centisecs",
 720                 .data           = &dirty_writeback_centisecs,
 721                 .maxlen         = sizeof(dirty_writeback_centisecs),
 722                 .mode           = 0644,
 723                 .proc_handler   = &dirty_writeback_centisecs_handler,
 724         },
 725         {
 726                 .ctl_name       = VM_DIRTY_EXPIRE_CS,
 727                 .procname       = "dirty_expire_centisecs",
 728                 .data           = &dirty_expire_centisecs,
 729                 .maxlen         = sizeof(dirty_expire_centisecs),
 730                 .mode           = 0644,
 731                 .proc_handler   = &proc_dointvec,
 732         },
 733         {
 734                 .ctl_name       = VM_NR_PDFLUSH_THREADS,
 735                 .procname       = "nr_pdflush_threads",
 736                 .data           = &nr_pdflush_threads,
 737                 .maxlen         = sizeof nr_pdflush_threads,
 738                 .mode           = 0444 /* read-only*/,
 739                 .proc_handler   = &proc_dointvec,
 740         },
 741         {
 742                 .ctl_name       = VM_SWAPPINESS,
 743                 .procname       = "swappiness",
 744                 .data           = &vm_swappiness,
 745                 .maxlen         = sizeof(vm_swappiness),
 746                 .mode           = 0644,
 747                 .proc_handler   = &proc_dointvec_minmax,
 748                 .strategy       = &sysctl_intvec,
 749                 .extra1         = &zero,
 750                 .extra2         = &one_hundred,
 751         },
 752 #ifdef CONFIG_HUGETLB_PAGE
 753          {
 754                 .ctl_name       = VM_HUGETLB_PAGES,
 755                 .procname       = "nr_hugepages",
 756                 .data           = &max_huge_pages,
 757                 .maxlen         = sizeof(unsigned long),
 758                 .mode           = 0644,
 759                 .proc_handler   = &hugetlb_sysctl_handler,
 760                 .extra1         = (void *)&hugetlb_zero,
 761                 .extra2         = (void *)&hugetlb_infinity,
 762          },
 763          {
 764                 .ctl_name       = VM_HUGETLB_GROUP,
 765                 .procname       = "hugetlb_shm_group",
 766                 .data           = &sysctl_hugetlb_shm_group,
 767                 .maxlen         = sizeof(gid_t),
 768                 .mode           = 0644,
 769                 .proc_handler   = &proc_dointvec,
 770          },
 771 #endif
 772         {
 773                 .ctl_name       = VM_LOWMEM_RESERVE_RATIO,
 774                 .procname       = "lowmem_reserve_ratio",
 775                 .data           = &sysctl_lowmem_reserve_ratio,
 776                 .maxlen         = sizeof(sysctl_lowmem_reserve_ratio),
 777                 .mode           = 0644,
 778                 .proc_handler   = &lowmem_reserve_ratio_sysctl_handler,
 779                 .strategy       = &sysctl_intvec,
 780         },
 781         {
 782                 .ctl_name       = VM_DROP_PAGECACHE,
 783                 .procname       = "drop_caches",
 784                 .data           = &sysctl_drop_caches,
 785                 .maxlen         = sizeof(int),
 786                 .mode           = 0644,
 787                 .proc_handler   = drop_caches_sysctl_handler,
 788                 .strategy       = &sysctl_intvec,
 789         },
 790         {
 791                 .ctl_name       = VM_MIN_FREE_KBYTES,
 792                 .procname       = "min_free_kbytes",
 793                 .data           = &min_free_kbytes,
 794                 .maxlen         = sizeof(min_free_kbytes),
 795                 .mode           = 0644,
 796                 .proc_handler   = &min_free_kbytes_sysctl_handler,
 797                 .strategy       = &sysctl_intvec,
 798                 .extra1         = &zero,
 799         },
 800         {
 801                 .ctl_name       = VM_PERCPU_PAGELIST_FRACTION,
 802                 .procname       = "percpu_pagelist_fraction",
 803                 .data           = &percpu_pagelist_fraction,
 804                 .maxlen         = sizeof(percpu_pagelist_fraction),
 805                 .mode           = 0644,
 806                 .proc_handler   = &percpu_pagelist_fraction_sysctl_handler,
 807                 .strategy       = &sysctl_intvec,
 808                 .extra1         = &min_percpu_pagelist_fract,
 809         },
 810 #ifdef CONFIG_MMU
 811         {
 812                 .ctl_name       = VM_MAX_MAP_COUNT,
 813                 .procname       = "max_map_count",
 814                 .data           = &sysctl_max_map_count,
 815                 .maxlen         = sizeof(sysctl_max_map_count),
 816                 .mode           = 0644,
 817                 .proc_handler   = &proc_dointvec
 818         },
 819 #endif
 820         {
 821                 .ctl_name       = VM_LAPTOP_MODE,
 822                 .procname       = "laptop_mode",
 823                 .data           = &laptop_mode,
 824                 .maxlen         = sizeof(laptop_mode),
 825                 .mode           = 0644,
 826                 .proc_handler   = &proc_dointvec,
 827                 .strategy       = &sysctl_intvec,
 828                 .extra1         = &zero,
 829         },
 830         {
 831                 .ctl_name       = VM_BLOCK_DUMP,
 832                 .procname       = "block_dump",
 833                 .data           = &block_dump,
 834                 .maxlen         = sizeof(block_dump),
 835                 .mode           = 0644,
 836                 .proc_handler   = &proc_dointvec,
 837                 .strategy       = &sysctl_intvec,
 838                 .extra1         = &zero,
 839         },
 840         {
 841                 .ctl_name       = VM_VFS_CACHE_PRESSURE,
 842                 .procname       = "vfs_cache_pressure",
 843                 .data           = &sysctl_vfs_cache_pressure,
 844                 .maxlen         = sizeof(sysctl_vfs_cache_pressure),
 845                 .mode           = 0644,
 846                 .proc_handler   = &proc_dointvec,
 847                 .strategy       = &sysctl_intvec,
 848                 .extra1         = &zero,
 849         },
 850 #ifdef HAVE_ARCH_PICK_MMAP_LAYOUT
 851         {
 852                 .ctl_name       = VM_LEGACY_VA_LAYOUT,
 853                 .procname       = "legacy_va_layout",
 854                 .data           = &sysctl_legacy_va_layout,
 855                 .maxlen         = sizeof(sysctl_legacy_va_layout),
 856                 .mode           = 0644,
 857                 .proc_handler   = &proc_dointvec,
 858                 .strategy       = &sysctl_intvec,
 859                 .extra1         = &zero,
 860         },
 861 #endif
 862 #ifdef CONFIG_SWAP
 863         {
 864                 .ctl_name       = VM_SWAP_TOKEN_TIMEOUT,
 865                 .procname       = "swap_token_timeout",
 866                 .data           = &swap_token_default_timeout,
 867                 .maxlen         = sizeof(swap_token_default_timeout),
 868                 .mode           = 0644,
 869                 .proc_handler   = &proc_dointvec_jiffies,
 870                 .strategy       = &sysctl_jiffies,
 871         },
 872 #endif
 873 #ifdef CONFIG_NUMA
 874         {
 875                 .ctl_name       = VM_ZONE_RECLAIM_MODE,
 876                 .procname       = "zone_reclaim_mode",
 877                 .data           = &zone_reclaim_mode,
 878                 .maxlen         = sizeof(zone_reclaim_mode),
 879                 .mode           = 0644,
 880                 .proc_handler   = &proc_dointvec,
 881                 .strategy       = &sysctl_intvec,
 882                 .extra1         = &zero,
 883         },
 884 #endif
 885         { .ctl_name = 0 }
 886 };
 887
 888 static ctl_table proc_table[] = {
 889         { .ctl_name = 0 }
 890 };
 891
 892 static ctl_table fs_table[] = {
 893         {
 894                 .ctl_name       = FS_NRINODE,
 895                 .procname       = "inode-nr",
 896                 .data           = &inodes_stat,
 897                 .maxlen         = 2*sizeof(int),
 898                 .mode           = 0444,
 899                 .proc_handler   = &proc_dointvec,
 900         },
 901         {
 902                 .ctl_name       = FS_STATINODE,
 903                 .procname       = "inode-state",
 904                 .data           = &inodes_stat,
 905                 .maxlen         = 7*sizeof(int),
 906                 .mode           = 0444,
 907                 .proc_handler   = &proc_dointvec,
 908         },
 909         {
 910                 .ctl_name       = FS_NRFILE,
 911                 .procname       = "file-nr",
 912                 .data           = &files_stat,
 913                 .maxlen         = 3*sizeof(int),
 914                 .mode           = 0444,
 915                 .proc_handler   = &proc_dointvec,
 916         },
 917         {
 918                 .ctl_name       = FS_MAXFILE,
 919                 .procname       = "file-max",
 920                 .data           = &files_stat.max_files,
 921                 .maxlen         = sizeof(int),
 922                 .mode           = 0644,
 923                 .proc_handler   = &proc_dointvec,
 924         },
 925         {
 926                 .ctl_name       = FS_DENTRY,
 927                 .procname       = "dentry-state",
 928                 .data           = &dentry_stat,
 929                 .maxlen         = 6*sizeof(int),
 930                 .mode           = 0444,
 931                 .proc_handler   = &proc_dointvec,
 932         },
 933         {
 934                 .ctl_name       = FS_OVERFLOWUID,
 935                 .procname       = "overflowuid",
 936                 .data           = &fs_overflowuid,
 937                 .maxlen         = sizeof(int),
 938                 .mode           = 0644,
 939                 .proc_handler   = &proc_dointvec_minmax,
 940                 .strategy       = &sysctl_intvec,
 941                 .extra1         = &minolduid,
 942                 .extra2         = &maxolduid,
 943         },
 944         {
 945                 .ctl_name       = FS_OVERFLOWGID,
 946                 .procname       = "overflowgid",
 947                 .data           = &fs_overflowgid,
 948                 .maxlen         = sizeof(int),
 949                 .mode           = 0644,
 950                 .proc_handler   = &proc_dointvec_minmax,
 951                 .strategy       = &sysctl_intvec,
 952                 .extra1         = &minolduid,
 953                 .extra2         = &maxolduid,
 954         },
 955         {
 956                 .ctl_name       = FS_LEASES,
 957                 .procname       = "leases-enable",
 958                 .data           = &leases_enable,
 959                 .maxlen         = sizeof(int),
 960                 .mode           = 0644,
 961                 .proc_handler   = &proc_dointvec,
 962         },
 963 #ifdef CONFIG_DNOTIFY
 964         {
 965                 .ctl_name       = FS_DIR_NOTIFY,
 966                 .procname       = "dir-notify-enable",
 967                 .data           = &dir_notify_enable,
 968                 .maxlen         = sizeof(int),
 969                 .mode           = 0644,
 970                 .proc_handler   = &proc_dointvec,
 971         },
 972 #endif
 973 #ifdef CONFIG_MMU
 974         {
 975                 .ctl_name       = FS_LEASE_TIME,
 976                 .procname       = "lease-break-time",
 977                 .data           = &lease_break_time,
 978                 .maxlen         = sizeof(int),
 979                 .mode           = 0644,
 980                 .proc_handler   = &proc_dointvec,
 981         },
 982         {
 983                 .ctl_name       = FS_AIO_NR,
 984                 .procname       = "aio-nr",
 985                 .data           = &aio_nr,
 986                 .maxlen         = sizeof(aio_nr),
 987                 .mode           = 0444,
 988                 .proc_handler   = &proc_doulongvec_minmax,
 989         },
 990         {
 991                 .ctl_name       = FS_AIO_MAX_NR,
 992                 .procname       = "aio-max-nr",
 993                 .data           = &aio_max_nr,
 994                 .maxlen         = sizeof(aio_max_nr),
 995                 .mode           = 0644,
 996                 .proc_handler   = &proc_doulongvec_minmax,
 997         },
 998 #ifdef CONFIG_INOTIFY
 999         {
1000                 .ctl_name       = FS_INOTIFY,
1001                 .procname       = "inotify",
1002                 .mode           = 0555,
1003                 .child          = inotify_table,
1004         },
1005 #endif
1006 #endif
1007         {
1008                 .ctl_name       = KERN_SETUID_DUMPABLE,
1009                 .procname       = "suid_dumpable",
1010                 .data           = &suid_dumpable,
1011                 .maxlen         = sizeof(int),
1012                 .mode           = 0644,
1013                 .proc_handler   = &proc_dointvec,
1014         },
1015         { .ctl_name = 0 }
1016 };
1017
1018 static ctl_table debug_table[] = {
1019         { .ctl_name = 0 }
1020 };
1021
1022 static ctl_table dev_table[] = {
1023         { .ctl_name = 0 }
1024 };
1025
1026 extern void init_irq_proc (void);
1027
1028 static DEFINE_SPINLOCK(sysctl_lock);
1029
1030 /* called under sysctl_lock */
1031 static int use_table(struct ctl_table_header *p)
1032 {
1033         if (unlikely(p->unregistering))
1034                 return 0;
1035         p->used++;
1036         return 1;
1037 }
1038
1039 /* called under sysctl_lock */
1040 static void unuse_table(struct ctl_table_header *p)
1041 {
1042         if (!--p->used)
1043                 if (unlikely(p->unregistering))
1044                         complete(p->unregistering);
1045 }
1046
1047 /* called under sysctl_lock, will reacquire if has to wait */
1048 static void start_unregistering(struct ctl_table_header *p)
1049 {
1050         /*
1051          * if p->used is 0, nobody will ever touch that entry again;
1052          * we'll eliminate all paths to it before dropping sysctl_lock
1053          */
1054         if (unlikely(p->used)) {
1055                 struct completion wait;
1056                 init_completion(&wait);
1057                 p->unregistering = &wait;
1058                 spin_unlock(&sysctl_lock);
1059                 wait_for_completion(&wait);
1060                 spin_lock(&sysctl_lock);
1061         }
1062         /*
1063          * do not remove from the list until nobody holds it; walking the
1064          * list in do_sysctl() relies on that.
1065          */
1066         list_del_init(&p->ctl_entry);
1067 }
1068
1069 void __init sysctl_init(void)
1070 {
1071 #ifdef CONFIG_PROC_FS
1072         register_proc_table(root_table, proc_sys_root, &root_table_header);
1073         init_irq_proc();
1074 #endif
1075 }
1076
1077 int do_sysctl(int __user *name, int nlen, void __user *oldval, size_t __user *oldlenp,
1078                void __user *newval, size_t newlen)
1079 {
1080         struct list_head *tmp;
1081         int error = -ENOTDIR;
1082
1083         if (nlen <= 0 || nlen >= CTL_MAXNAME)
1084                 return -ENOTDIR;
1085         if (oldval) {
1086                 int old_len;
1087                 if (!oldlenp || get_user(old_len, oldlenp))
1088                         return -EFAULT;
1089         }
1090         spin_lock(&sysctl_lock);
1091         tmp = &root_table_header.ctl_entry;
1092         do {
1093                 struct ctl_table_header *head =
1094                         list_entry(tmp, struct ctl_table_header, ctl_entry);
1095                 void *context = NULL;
1096
1097                 if (!use_table(head))
1098                         continue;
1099
1100                 spin_unlock(&sysctl_lock);
1101
1102                 error = parse_table(name, nlen, oldval, oldlenp,
1103                                         newval, newlen, head->ctl_table,
1104                                         &context);
1105                 kfree(context);
1106
1107                 spin_lock(&sysctl_lock);
1108                 unuse_table(head);
1109                 if (error != -ENOTDIR)
1110                         break;
1111         } while ((tmp = tmp->next) != &root_table_header.ctl_entry);
1112         spin_unlock(&sysctl_lock);
1113         return error;
1114 }
1115
1116 asmlinkage long sys_sysctl(struct __sysctl_args __user *args)
1117 {
1118         struct __sysctl_args tmp;
1119         int error;
1120
1121         if (copy_from_user(&tmp, args, sizeof(tmp)))
1122                 return -EFAULT;
1123
1124         lock_kernel();
1125         error = do_sysctl(tmp.name, tmp.nlen, tmp.oldval, tmp.oldlenp,
1126                           tmp.newval, tmp.newlen);
1127         unlock_kernel();
1128         return error;
1129 }
1130
1131 /*
1132  * ctl_perm does NOT grant the superuser all rights automatically, because
1133  * some sysctl variables are readonly even to root.
1134  */
1135
1136 static int test_perm(int mode, int op)
1137 {
1138         if (!current->euid)
1139                 mode >>= 6;
1140         else if (in_egroup_p(0))
1141                 mode >>= 3;
1142         if ((mode & op & 0007) == op)
1143                 return 0;
1144         return -EACCES;
1145 }
1146
1147 static inline int ctl_perm(ctl_table *table, int op)
1148 {
1149         int error;
1150         error = security_sysctl(table, op);
1151         if (error)
1152                 return error;
1153         return test_perm(table->mode, op);
1154 }
1155
1156 static int parse_table(int __user *name, int nlen,
1157                        void __user *oldval, size_t __user *oldlenp,
1158                        void __user *newval, size_t newlen,
1159                        ctl_table *table, void **context)
1160 {
1161         int n;
1162 repeat:
1163         if (!nlen)
1164                 return -ENOTDIR;
1165         if (get_user(n, name))
1166                 return -EFAULT;
1167         for ( ; table->ctl_name; table++) {
1168                 if (n == table->ctl_name || table->ctl_name == CTL_ANY) {
1169                         int error;
1170                         if (table->child) {
1171                                 if (ctl_perm(table, 001))
1172                                         return -EPERM;
1173                                 if (table->strategy) {
1174                                         error = table->strategy(
1175                                                 table, name, nlen,
1176                                                 oldval, oldlenp,
1177                                                 newval, newlen, context);
1178                                         if (error)
1179                                                 return error;
1180                                 }
1181                                 name++;
1182                                 nlen--;
1183                                 table = table->child;
1184                                 goto repeat;
1185                         }
1186                         error = do_sysctl_strategy(table, name, nlen,
1187                                                    oldval, oldlenp,
1188                                                    newval, newlen, context);
1189                         return error;
1190                 }
1191         }
1192         return -ENOTDIR;
1193 }
1194
1195 /* Perform the actual read/write of a sysctl table entry. */
1196 int do_sysctl_strategy (ctl_table *table,
1197                         int __user *name, int nlen,
1198                         void __user *oldval, size_t __user *oldlenp,
1199                         void __user *newval, size_t newlen, void **context)
1200 {
1201         int op = 0, rc;
1202         size_t len;
1203
1204         if (oldval)
1205                 op |= 004;
1206         if (newval)
1207                 op |= 002;
1208         if (ctl_perm(table, op))
1209                 return -EPERM;
1210
1211         if (table->strategy) {
1212                 rc = table->strategy(table, name, nlen, oldval, oldlenp,
1213                                      newval, newlen, context);
1214                 if (rc < 0)
1215                         return rc;
1216                 if (rc > 0)
1217                         return 0;
1218         }
1219
1220         /* If there is no strategy routine, or if the strategy returns
1221          * zero, proceed with automatic r/w */
1222         if (table->data && table->maxlen) {
1223                 if (oldval && oldlenp) {
1224                         if (get_user(len, oldlenp))
1225                                 return -EFAULT;
1226                         if (len) {
1227                                 if (len > table->maxlen)
1228                                         len = table->maxlen;
1229                                 if(copy_to_user(oldval, table->data, len))
1230                                         return -EFAULT;
1231                                 if(put_user(len, oldlenp))
1232                                         return -EFAULT;
1233                         }
1234                 }
1235                 if (newval && newlen) {
1236                         len = newlen;
1237                         if (len > table->maxlen)
1238                                 len = table->maxlen;
1239                         if(copy_from_user(table->data, newval, len))
1240                                 return -EFAULT;
1241                 }
1242         }
1243         return 0;
1244 }
1245
1246 /**
1247  * register_sysctl_table - register a sysctl hierarchy
1248  * @table: the top-level table structure
1249  * @insert_at_head: whether the entry should be inserted in front or at the end
1250  *
1251  * Register a sysctl table hierarchy. @table should be a filled in ctl_table
1252  * array. An entry with a ctl_name of 0 terminates the table.
1253  *
1254  * The members of the &ctl_table structure are used as follows:
1255  *
1256  * ctl_name - This is the numeric sysctl value used by sysctl(2). The number
1257  *            must be unique within that level of sysctl
1258  *
1259  * procname - the name of the sysctl file under /proc/sys. Set to %NULL to not
1260  *            enter a sysctl file
1261  *
1262  * data - a pointer to data for use by proc_handler
1263  *
1264  * maxlen - the maximum size in bytes of the data
1265  *
1266  * mode - the file permissions for the /proc/sys file, and for sysctl(2)
1267  *
1268  * child - a pointer to the child sysctl table if this entry is a directory, or
1269  *         %NULL.
1270  *
1271  * proc_handler - the text handler routine (described below)
1272  *
1273  * strategy - the strategy routine (described below)
1274  *
1275  * de - for internal use by the sysctl routines
1276  *
1277  * extra1, extra2 - extra pointers usable by the proc handler routines
1278  *
1279  * Leaf nodes in the sysctl tree will be represented by a single file
1280  * under /proc; non-leaf nodes will be represented by directories.
1281  *
1282  * sysctl(2) can automatically manage read and write requests through
1283  * the sysctl table.  The data and maxlen fields of the ctl_table
1284  * struct enable minimal validation of the values being written to be
1285  * performed, and the mode field allows minimal authentication.
1286  *
1287  * More sophisticated management can be enabled by the provision of a
1288  * strategy routine with the table entry.  This will be called before
1289  * any automatic read or write of the data is performed.
1290  *
1291  * The strategy routine may return
1292  *
1293  * < 0 - Error occurred (error is passed to user process)
1294  *
1295  * 0   - OK - proceed with automatic read or write.
1296  *
1297  * > 0 - OK - read or write has been done by the strategy routine, so
1298  *       return immediately.
1299  *
1300  * There must be a proc_handler routine for any terminal nodes
1301  * mirrored under /proc/sys (non-terminals are handled by a built-in
1302  * directory handler).  Several default handlers are available to
1303  * cover common cases -
1304  *
1305  * proc_dostring(), proc_dointvec(), proc_dointvec_jiffies(),
1306  * proc_dointvec_userhz_jiffies(), proc_dointvec_minmax(),
1307  * proc_doulongvec_ms_jiffies_minmax(), proc_doulongvec_minmax()
1308  *
1309  * It is the handler's job to read the input buffer from user memory
1310  * and process it. The handler should return 0 on success.
1311  *
1312  * This routine returns %NULL on a failure to register, and a pointer
1313  * to the table header on success.
1314  */
1315 struct ctl_table_header *register_sysctl_table(ctl_table * table,
1316                                                int insert_at_head)
1317 {
1318         struct ctl_table_header *tmp;
1319         tmp = kmalloc(sizeof(struct ctl_table_header), GFP_KERNEL);
1320         if (!tmp)
1321                 return NULL;
1322         tmp->ctl_table = table;
1323         INIT_LIST_HEAD(&tmp->ctl_entry);
1324         tmp->used = 0;
1325         tmp->unregistering = NULL;
1326         spin_lock(&sysctl_lock);
1327         if (insert_at_head)
1328                 list_add(&tmp->ctl_entry, &root_table_header.ctl_entry);
1329         else
1330                 list_add_tail(&tmp->ctl_entry, &root_table_header.ctl_entry);
1331         spin_unlock(&sysctl_lock);
1332 #ifdef CONFIG_PROC_FS
1333         register_proc_table(table, proc_sys_root, tmp);
1334 #endif
1335         return tmp;
1336 }
1337
1338 /**
1339  * unregister_sysctl_table - unregister a sysctl table hierarchy
1340  * @header: the header returned from register_sysctl_table
1341  *
1342  * Unregisters the sysctl table and all children. proc entries may not
1343  * actually be removed until they are no longer used by anyone.
1344  */
1345 void unregister_sysctl_table(struct ctl_table_header * header)
1346 {
1347         might_sleep();
1348         spin_lock(&sysctl_lock);
1349         start_unregistering(header);
1350 #ifdef CONFIG_PROC_FS
1351         unregister_proc_table(header->ctl_table, proc_sys_root);
1352 #endif
1353         spin_unlock(&sysctl_lock);
1354         kfree(header);
1355 }
1356
1357 /*
1358  * /proc/sys support
1359  */
1360
1361 #ifdef CONFIG_PROC_FS
1362
1363 /* Scan the sysctl entries in table and add them all into /proc */
1364 static void register_proc_table(ctl_table * table, struct proc_dir_entry *root, void *set)
1365 {
1366         struct proc_dir_entry *de;
1367         int len;
1368         mode_t mode;
1369
1370         for (; table->ctl_name; table++) {
1371                 /* Can't do anything without a proc name. */
1372                 if (!table->procname)
1373                         continue;
1374                 /* Maybe we can't do anything with it... */
1375                 if (!table->proc_handler && !table->child) {
1376                         printk(KERN_WARNING "SYSCTL: Can't register %s\n",
1377                                 table->procname);
1378                         continue;
1379                 }
1380
1381                 len = strlen(table->procname);
1382                 mode = table->mode;
1383
1384                 de = NULL;
1385                 if (table->proc_handler)
1386                         mode |= S_IFREG;
1387                 else {
1388                         mode |= S_IFDIR;
1389                         for (de = root->subdir; de; de = de->next) {
1390                                 if (proc_match(len, table->procname, de))
1391                                         break;
1392                         }
1393                         /* If the subdir exists already, de is non-NULL */
1394                 }
1395
1396                 if (!de) {
1397                         de = create_proc_entry(table->procname, mode, root);
1398                         if (!de)
1399                                 continue;
1400                         de->set = set;
1401                         de->data = (void *) table;
1402                         if (table->proc_handler)
1403                                 de->proc_fops = &proc_sys_file_operations;
1404                 }
1405                 table->de = de;
1406                 if (de->mode & S_IFDIR)
1407                         register_proc_table(table->child, de, set);
1408         }
1409 }
1410
1411 /*
1412  * Unregister a /proc sysctl table and any subdirectories.
1413  */
1414 static void unregister_proc_table(ctl_table * table, struct proc_dir_entry *root)
1415 {
1416         struct proc_dir_entry *de;
1417         for (; table->ctl_name; table++) {
1418                 if (!(de = table->de))
1419                         continue;
1420                 if (de->mode & S_IFDIR) {
1421                         if (!table->child) {
1422                                 printk (KERN_ALERT "Help - malformed sysctl tree on free\n");
1423                                 continue;
1424                         }
1425                         unregister_proc_table(table->child, de);
1426
1427                         /* Don't unregister directories which still have entries.. */
1428                         if (de->subdir)
1429                                 continue;
1430                 }
1431
1432                 /*
1433                  * In any case, mark the entry as goner; we'll keep it
1434                  * around if it's busy, but we'll know to do nothing with
1435                  * its fields.  We are under sysctl_lock here.
1436                  */
1437                 de->data = NULL;
1438
1439                 /* Don't unregister proc entries that are still being used.. */
1440                 if (atomic_read(&de->count))
1441                         continue;
1442
1443                 table->de = NULL;
1444                 remove_proc_entry(table->procname, root);
1445         }
1446 }
1447
1448 static ssize_t do_rw_proc(int write, struct file * file, char __user * buf,
1449                           size_t count, loff_t *ppos)
1450 {
1451         int op;
1452         struct proc_dir_entry *de = PDE(file->f_dentry->d_inode);
1453         struct ctl_table *table;
1454         size_t res;
1455         ssize_t error = -ENOTDIR;
1456
1457         spin_lock(&sysctl_lock);
1458         if (de && de->data && use_table(de->set)) {
1459                 /*
1460                  * at that point we know that sysctl was not unregistered
1461                  * and won't be until we finish
1462                  */
1463                 spin_unlock(&sysctl_lock);
1464                 table = (struct ctl_table *) de->data;
1465                 if (!table || !table->proc_handler)
1466                         goto out;
1467                 error = -EPERM;
1468                 op = (write ? 002 : 004);
1469                 if (ctl_perm(table, op))
1470                         goto out;
1471
1472                 /* careful: calling conventions are nasty here */
1473                 res = count;
1474                 error = (*table->proc_handler)(table, write, file,
1475                                                 buf, &res, ppos);
1476                 if (!error)
1477                         error = res;
1478         out:
1479                 spin_lock(&sysctl_lock);
1480                 unuse_table(de->set);
1481         }
1482         spin_unlock(&sysctl_lock);
1483         return error;
1484 }
1485
1486 static int proc_opensys(struct inode *inode, struct file *file)
1487 {
1488         if (file->f_mode & FMODE_WRITE) {
1489                 /*
1490                  * sysctl entries that are not writable,
1491                  * are _NOT_ writable, capabilities or not.
1492                  */
1493                 if (!(inode->i_mode & S_IWUSR))
1494                         return -EPERM;
1495         }
1496
1497         return 0;
1498 }
1499
1500 static ssize_t proc_readsys(struct file * file, char __user * buf,
1501                             size_t count, loff_t *ppos)
1502 {
1503         return do_rw_proc(0, file, buf, count, ppos);
1504 }
1505
1506 static ssize_t proc_writesys(struct file * file, const char __user * buf,
1507                              size_t count, loff_t *ppos)
1508 {
1509         return do_rw_proc(1, file, (char __user *) buf, count, ppos);
1510 }
1511
1512 /**
1513  * proc_dostring - read a string sysctl
1514  * @table: the sysctl table
1515  * @write: %TRUE if this is a write to the sysctl file
1516  * @filp: the file structure
1517  * @buffer: the user buffer
1518  * @lenp: the size of the user buffer
1519  * @ppos: file position
1520  *
1521  * Reads/writes a string from/to the user buffer. If the kernel
1522  * buffer provided is not large enough to hold the string, the
1523  * string is truncated. The copied string is %NULL-terminated.
1524  * If the string is being read by the user process, it is copied
1525  * and a newline '\n' is added. It is truncated if the buffer is
1526  * not large enough.
1527  *
1528  * Returns 0 on success.
1529  */
1530 int proc_dostring(ctl_table *table, int write, struct file *filp,
1531                   void __user *buffer, size_t *lenp, loff_t *ppos)
1532 {
1533         size_t len;
1534         char __user *p;
1535         char c;
1536
1537         if (!table->data || !table->maxlen || !*lenp ||
1538             (*ppos && !write)) {
1539                 *lenp = 0;
1540                 return 0;
1541         }
1542
1543         if (write) {
1544                 len = 0;
1545                 p = buffer;
1546                 while (len < *lenp) {
1547                         if (get_user(c, p++))
1548                                 return -EFAULT;
1549                         if (c == 0 || c == '\n')
1550                                 break;
1551                         len++;
1552                 }
1553                 if (len >= table->maxlen)
1554                         len = table->maxlen-1;
1555                 if(copy_from_user(table->data, buffer, len))
1556                         return -EFAULT;
1557                 ((char *) table->data)[len] = 0;
1558                 *ppos += *lenp;
1559         } else {
1560                 len = strlen(table->data);
1561                 if (len > table->maxlen)
1562                         len = table->maxlen;
1563                 if (len > *lenp)
1564                         len = *lenp;
1565                 if (len)
1566                         if(copy_to_user(buffer, table->data, len))
1567                                 return -EFAULT;
1568                 if (len < *lenp) {
1569                         if(put_user('\n', ((char __user *) buffer) + len))
1570                                 return -EFAULT;
1571                         len++;
1572                 }
1573                 *lenp = len;
1574                 *ppos += len;
1575         }
1576         return 0;
1577 }
1578
1579 /*
1580  *      Special case of dostring for the UTS structure. This has locks
1581  *      to observe. Should this be in kernel/sys.c ????
1582  */
1583
1584 static int proc_doutsstring(ctl_table *table, int write, struct file *filp,
1585                   void __user *buffer, size_t *lenp, loff_t *ppos)
1586 {
1587         int r;
1588
1589         if (!write) {
1590                 down_read(&uts_sem);
1591                 r=proc_dostring(table,0,filp,buffer,lenp, ppos);
1592                 up_read(&uts_sem);
1593         } else {
1594                 down_write(&uts_sem);
1595                 r=proc_dostring(table,1,filp,buffer,lenp, ppos);
1596                 up_write(&uts_sem);
1597         }
1598         return r;
1599 }
1600
1601 static int do_proc_dointvec_conv(int *negp, unsigned long *lvalp,
1602                                  int *valp,
1603                                  int write, void *data)
1604 {
1605         if (write) {
1606                 *valp = *negp ? -*lvalp : *lvalp;
1607         } else {
1608                 int val = *valp;
1609                 if (val < 0) {
1610                         *negp = -1;
1611                         *lvalp = (unsigned long)-val;
1612                 } else {
1613                         *negp = 0;
1614                         *lvalp = (unsigned long)val;
1615                 }
1616         }
1617         return 0;
1618 }
1619
1620 static int do_proc_dointvec(ctl_table *table, int write, struct file *filp,
1621                   void __user *buffer, size_t *lenp, loff_t *ppos,
1622                   int (*conv)(int *negp, unsigned long *lvalp, int *valp,
1623                               int write, void *data),
1624                   void *data)
1625 {
1626 #define TMPBUFLEN 21
1627         int *i, vleft, first=1, neg, val;
1628         unsigned long lval;
1629         size_t left, len;
1630
1631         char buf[TMPBUFLEN], *p;
1632         char __user *s = buffer;
1633
1634         if (!table->data || !table->maxlen || !*lenp ||
1635             (*ppos && !write)) {
1636                 *lenp = 0;
1637                 return 0;
1638         }
1639
1640         i = (int *) table->data;
1641         vleft = table->maxlen / sizeof(*i);
1642         left = *lenp;
1643
1644         if (!conv)
1645                 conv = do_proc_dointvec_conv;
1646
1647         for (; left && vleft--; i++, first=0) {
1648                 if (write) {
1649                         while (left) {
1650                                 char c;
1651                                 if (get_user(c, s))
1652                                         return -EFAULT;
1653                                 if (!isspace(c))
1654                                         break;
1655                                 left--;
1656                                 s++;
1657                         }
1658                         if (!left)
1659                                 break;
1660                         neg = 0;
1661                         len = left;
1662                         if (len > sizeof(buf) - 1)
1663                                 len = sizeof(buf) - 1;
1664                         if (copy_from_user(buf, s, len))
1665                                 return -EFAULT;
1666                         buf[len] = 0;
1667                         p = buf;
1668                         if (*p == '-' && left > 1) {
1669                                 neg = 1;
1670                                 left--, p++;
1671                         }
1672                         if (*p < '0' || *p > '9')
1673                                 break;
1674
1675                         lval = simple_strtoul(p, &p, 0);
1676
1677                         len = p-buf;
1678                         if ((len < left) && *p && !isspace(*p))
1679                                 break;
1680                         if (neg)
1681                                 val = -val;
1682                         s += len;
1683                         left -= len;
1684
1685                         if (conv(&neg, &lval, i, 1, data))
1686                                 break;
1687                 } else {
1688                         p = buf;
1689                         if (!first)
1690                                 *p++ = '\t';
1691
1692                         if (conv(&neg, &lval, i, 0, data))
1693                                 break;
1694
1695                         sprintf(p, "%s%lu", neg ? "-" : "", lval);
1696                         len = strlen(buf);
1697                         if (len > left)
1698                                 len = left;
1699                         if(copy_to_user(s, buf, len))
1700                                 return -EFAULT;
1701                         left -= len;
1702                         s += len;
1703                 }
1704         }
1705
1706         if (!write && !first && left) {
1707                 if(put_user('\n', s))
1708                         return -EFAULT;
1709                 left--, s++;
1710         }
1711         if (write) {
1712                 while (left) {
1713                         char c;
1714                         if (get_user(c, s++))
1715                                 return -EFAULT;
1716                         if (!isspace(c))
1717                                 break;
1718                         left--;
1719                 }
1720         }
1721         if (write && first)
1722                 return -EINVAL;
1723         *lenp -= left;
1724         *ppos += *lenp;
1725         return 0;
1726 #undef TMPBUFLEN
1727 }
1728
1729 /**
1730  * proc_dointvec - read a vector of integers
1731  * @table: the sysctl table
1732  * @write: %TRUE if this is a write to the sysctl file
1733  * @filp: the file structure
1734  * @buffer: the user buffer
1735  * @lenp: the size of the user buffer
1736  * @ppos: file position
1737  *
1738  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
1739  * values from/to the user buffer, treated as an ASCII string.
1740  *
1741  * Returns 0 on success.
1742  */
1743 int proc_dointvec(ctl_table *table, int write, struct file *filp,
1744                      void __user *buffer, size_t *lenp, loff_t *ppos)
1745 {
1746     return do_proc_dointvec(table,write,filp,buffer,lenp,ppos,
1747                             NULL,NULL);
1748 }
1749
1750 #define OP_SET  0
1751 #define OP_AND  1
1752 #define OP_OR   2
1753 #define OP_MAX  3
1754 #define OP_MIN  4
1755
1756 static int do_proc_dointvec_bset_conv(int *negp, unsigned long *lvalp,
1757                                       int *valp,
1758                                       int write, void *data)
1759 {
1760         int op = *(int *)data;
1761         if (write) {
1762                 int val = *negp ? -*lvalp : *lvalp;
1763                 switch(op) {
1764                 case OP_SET:    *valp = val; break;
1765                 case OP_AND:    *valp &= val; break;
1766                 case OP_OR:     *valp |= val; break;
1767                 case OP_MAX:    if(*valp < val)
1768                                         *valp = val;
1769                                 break;
1770                 case OP_MIN:    if(*valp > val)
1771                                 *valp = val;
1772                                 break;
1773                 }
1774         } else {
1775                 int val = *valp;
1776                 if (val < 0) {
1777                         *negp = -1;
1778                         *lvalp = (unsigned long)-val;
1779                 } else {
1780                         *negp = 0;
1781                         *lvalp = (unsigned long)val;
1782                 }
1783         }
1784         return 0;
1785 }
1786
1787 /*
1788  *      init may raise the set.
1789  */
1790
1791 int proc_dointvec_bset(ctl_table *table, int write, struct file *filp,
1792                         void __user *buffer, size_t *lenp, loff_t *ppos)
1793 {
1794         int op;
1795
1796         if (!capable(CAP_SYS_MODULE)) {
1797                 return -EPERM;
1798         }
1799
1800         op = (current->pid == 1) ? OP_SET : OP_AND;
1801         return do_proc_dointvec(table,write,filp,buffer,lenp,ppos,
1802                                 do_proc_dointvec_bset_conv,&op);
1803 }
1804
1805 struct do_proc_dointvec_minmax_conv_param {
1806         int *min;
1807         int *max;
1808 };
1809
1810 static int do_proc_dointvec_minmax_conv(int *negp, unsigned long *lvalp,
1811                                         int *valp,
1812                                         int write, void *data)
1813 {
1814         struct do_proc_dointvec_minmax_conv_param *param = data;
1815         if (write) {
1816                 int val = *negp ? -*lvalp : *lvalp;
1817                 if ((param->min && *param->min > val) ||
1818                     (param->max && *param->max < val))
1819                         return -EINVAL;
1820                 *valp = val;
1821         } else {
1822                 int val = *valp;
1823                 if (val < 0) {
1824                         *negp = -1;
1825                         *lvalp = (unsigned long)-val;
1826                 } else {
1827                         *negp = 0;
1828                         *lvalp = (unsigned long)val;
1829                 }
1830         }
1831         return 0;
1832 }
1833
1834 /**
1835  * proc_dointvec_minmax - read a vector of integers with min/max values
1836  * @table: the sysctl table
1837  * @write: %TRUE if this is a write to the sysctl file
1838  * @filp: the file structure
1839  * @buffer: the user buffer
1840  * @lenp: the size of the user buffer
1841  * @ppos: file position
1842  *
1843  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
1844  * values from/to the user buffer, treated as an ASCII string.
1845  *
1846  * This routine will ensure the values are within the range specified by
1847  * table->extra1 (min) and table->extra2 (max).
1848  *
1849  * Returns 0 on success.
1850  */
1851 int proc_dointvec_minmax(ctl_table *table, int write, struct file *filp,
1852                   void __user *buffer, size_t *lenp, loff_t *ppos)
1853 {
1854         struct do_proc_dointvec_minmax_conv_param param = {
1855                 .min = (int *) table->extra1,
1856                 .max = (int *) table->extra2,
1857         };
1858         return do_proc_dointvec(table, write, filp, buffer, lenp, ppos,
1859                                 do_proc_dointvec_minmax_conv, &param);
1860 }
1861
1862 static int do_proc_doulongvec_minmax(ctl_table *table, int write,
1863                                      struct file *filp,
1864                                      void __user *buffer,
1865                                      size_t *lenp, loff_t *ppos,
1866                                      unsigned long convmul,
1867                                      unsigned long convdiv)
1868 {
1869 #define TMPBUFLEN 21
1870         unsigned long *i, *min, *max, val;
1871         int vleft, first=1, neg;
1872         size_t len, left;
1873         char buf[TMPBUFLEN], *p;
1874         char __user *s = buffer;
1875
1876         if (!table->data || !table->maxlen || !*lenp ||
1877             (*ppos && !write)) {
1878                 *lenp = 0;
1879                 return 0;
1880         }
1881
1882         i = (unsigned long *) table->data;
1883         min = (unsigned long *) table->extra1;
1884         max = (unsigned long *) table->extra2;
1885         vleft = table->maxlen / sizeof(unsigned long);
1886         left = *lenp;
1887
1888         for (; left && vleft--; i++, min++, max++, first=0) {
1889                 if (write) {
1890                         while (left) {
1891                                 char c;
1892                                 if (get_user(c, s))
1893                                         return -EFAULT;
1894                                 if (!isspace(c))
1895                                         break;
1896                                 left--;
1897                                 s++;
1898                         }
1899                         if (!left)
1900                                 break;
1901                         neg = 0;
1902                         len = left;
1903                         if (len > TMPBUFLEN-1)
1904                                 len = TMPBUFLEN-1;
1905                         if (copy_from_user(buf, s, len))
1906                                 return -EFAULT;
1907                         buf[len] = 0;
1908                         p = buf;
1909                         if (*p == '-' && left > 1) {
1910                                 neg = 1;
1911                                 left--, p++;
1912                         }
1913                         if (*p < '0' || *p > '9')
1914                                 break;
1915                         val = simple_strtoul(p, &p, 0) * convmul / convdiv ;
1916                         len = p-buf;
1917                         if ((len < left) && *p && !isspace(*p))
1918                                 break;
1919                         if (neg)
1920                                 val = -val;
1921                         s += len;
1922                         left -= len;
1923
1924                         if(neg)
1925                                 continue;
1926                         if ((min && val < *min) || (max && val > *max))
1927                                 continue;
1928                         *i = val;
1929                 } else {
1930                         p = buf;
1931                         if (!first)
1932                                 *p++ = '\t';
1933                         sprintf(p, "%lu", convdiv * (*i) / convmul);
1934                         len = strlen(buf);
1935                         if (len > left)
1936                                 len = left;
1937                         if(copy_to_user(s, buf, len))
1938                                 return -EFAULT;
1939                         left -= len;
1940                         s += len;
1941                 }
1942         }
1943
1944         if (!write && !first && left) {
1945                 if(put_user('\n', s))
1946                         return -EFAULT;
1947                 left--, s++;
1948         }
1949         if (write) {
1950                 while (left) {
1951                         char c;
1952                         if (get_user(c, s++))
1953                                 return -EFAULT;
1954                         if (!isspace(c))
1955                                 break;
1956                         left--;
1957                 }
1958         }
1959         if (write && first)
1960                 return -EINVAL;
1961         *lenp -= left;
1962         *ppos += *lenp;
1963         return 0;
1964 #undef TMPBUFLEN
1965 }
1966
1967 /**
1968  * proc_doulongvec_minmax - read a vector of long integers with min/max values
1969  * @table: the sysctl table
1970  * @write: %TRUE if this is a write to the sysctl file
1971  * @filp: the file structure
1972  * @buffer: the user buffer
1973  * @lenp: the size of the user buffer
1974  * @ppos: file position
1975  *
1976  * Reads/writes up to table->maxlen/sizeof(unsigned long) unsigned long
1977  * values from/to the user buffer, treated as an ASCII string.
1978  *
1979  * This routine will ensure the values are within the range specified by
1980  * table->extra1 (min) and table->extra2 (max).
1981  *
1982  * Returns 0 on success.
1983  */
1984 int proc_doulongvec_minmax(ctl_table *table, int write, struct file *filp,
1985                            void __user *buffer, size_t *lenp, loff_t *ppos)
1986 {
1987     return do_proc_doulongvec_minmax(table, write, filp, buffer, lenp, ppos, 1l, 1l);
1988 }
1989
1990 /**
1991  * proc_doulongvec_ms_jiffies_minmax - read a vector of millisecond values with min/max values
1992  * @table: the sysctl table
1993  * @write: %TRUE if this is a write to the sysctl file
1994  * @filp: the file structure
1995  * @buffer: the user buffer
1996  * @lenp: the size of the user buffer
1997  * @ppos: file position
1998  *
1999  * Reads/writes up to table->maxlen/sizeof(unsigned long) unsigned long
2000  * values from/to the user buffer, treated as an ASCII string. The values
2001  * are treated as milliseconds, and converted to jiffies when they are stored.
2002  *
2003  * This routine will ensure the values are within the range specified by
2004  * table->extra1 (min) and table->extra2 (max).
2005  *
2006  * Returns 0 on success.
2007  */
2008 int proc_doulongvec_ms_jiffies_minmax(ctl_table *table, int write,
2009                                       struct file *filp,
2010                                       void __user *buffer,
2011                                       size_t *lenp, loff_t *ppos)
2012 {
2013     return do_proc_doulongvec_minmax(table, write, filp, buffer,
2014                                      lenp, ppos, HZ, 1000l);
2015 }
2016
2017
2018 static int do_proc_dointvec_jiffies_conv(int *negp, unsigned long *lvalp,
2019                                          int *valp,
2020                                          int write, void *data)
2021 {
2022         if (write) {
2023                 *valp = *negp ? -(*lvalp*HZ) : (*lvalp*HZ);
2024         } else {
2025                 int val = *valp;
2026                 unsigned long lval;
2027                 if (val < 0) {
2028                         *negp = -1;
2029                         lval = (unsigned long)-val;
2030                 } else {
2031                         *negp = 0;
2032                         lval = (unsigned long)val;
2033                 }
2034                 *lvalp = lval / HZ;
2035         }
2036         return 0;
2037 }
2038
2039 static int do_proc_dointvec_userhz_jiffies_conv(int *negp, unsigned long *lvalp,
2040                                                 int *valp,
2041                                                 int write, void *data)
2042 {
2043         if (write) {
2044                 *valp = clock_t_to_jiffies(*negp ? -*lvalp : *lvalp);
2045         } else {
2046                 int val = *valp;
2047                 unsigned long lval;
2048                 if (val < 0) {
2049                         *negp = -1;
2050                         lval = (unsigned long)-val;
2051                 } else {
2052                         *negp = 0;
2053                         lval = (unsigned long)val;
2054                 }
2055                 *lvalp = jiffies_to_clock_t(lval);
2056         }
2057         return 0;
2058 }
2059
2060 static int do_proc_dointvec_ms_jiffies_conv(int *negp, unsigned long *lvalp,
2061                                             int *valp,
2062                                             int write, void *data)
2063 {
2064         if (write) {
2065                 *valp = msecs_to_jiffies(*negp ? -*lvalp : *lvalp);
2066         } else {
2067                 int val = *valp;
2068                 unsigned long lval;
2069                 if (val < 0) {
2070                         *negp = -1;
2071                         lval = (unsigned long)-val;
2072                 } else {
2073                         *negp = 0;
2074                         lval = (unsigned long)val;
2075                 }
2076                 *lvalp = jiffies_to_msecs(lval);
2077         }
2078         return 0;
2079 }
2080
2081 /**
2082  * proc_dointvec_jiffies - read a vector of integers as seconds
2083  * @table: the sysctl table
2084  * @write: %TRUE if this is a write to the sysctl file
2085  * @filp: the file structure
2086  * @buffer: the user buffer
2087  * @lenp: the size of the user buffer
2088  * @ppos: file position
2089  *
2090  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2091  * values from/to the user buffer, treated as an ASCII string.
2092  * The values read are assumed to be in seconds, and are converted into
2093  * jiffies.
2094  *
2095  * Returns 0 on success.
2096  */
2097 int proc_dointvec_jiffies(ctl_table *table, int write, struct file *filp,
2098                           void __user *buffer, size_t *lenp, loff_t *ppos)
2099 {
2100     return do_proc_dointvec(table,write,filp,buffer,lenp,ppos,
2101                             do_proc_dointvec_jiffies_conv,NULL);
2102 }
2103
2104 /**
2105  * proc_dointvec_userhz_jiffies - read a vector of integers as 1/USER_HZ seconds
2106  * @table: the sysctl table
2107  * @write: %TRUE if this is a write to the sysctl file
2108  * @filp: the file structure
2109  * @buffer: the user buffer
2110  * @lenp: the size of the user buffer
2111  * @ppos: pointer to the file position
2112  *
2113  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2114  * values from/to the user buffer, treated as an ASCII string.
2115  * The values read are assumed to be in 1/USER_HZ seconds, and
2116  * are converted into jiffies.
2117  *
2118  * Returns 0 on success.
2119  */
2120 int proc_dointvec_userhz_jiffies(ctl_table *table, int write, struct file *filp,
2121                                  void __user *buffer, size_t *lenp, loff_t *ppos)
2122 {
2123     return do_proc_dointvec(table,write,filp,buffer,lenp,ppos,
2124                             do_proc_dointvec_userhz_jiffies_conv,NULL);
2125 }
2126
2127 /**
2128  * proc_dointvec_ms_jiffies - read a vector of integers as 1 milliseconds
2129  * @table: the sysctl table
2130  * @write: %TRUE if this is a write to the sysctl file
2131  * @filp: the file structure
2132  * @buffer: the user buffer
2133  * @lenp: the size of the user buffer
2134  * @ppos: file position
2135  * @ppos: the current position in the file
2136  *
2137  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2138  * values from/to the user buffer, treated as an ASCII string.
2139  * The values read are assumed to be in 1/1000 seconds, and
2140  * are converted into jiffies.
2141  *
2142  * Returns 0 on success.
2143  */
2144 int proc_dointvec_ms_jiffies(ctl_table *table, int write, struct file *filp,
2145                              void __user *buffer, size_t *lenp, loff_t *ppos)
2146 {
2147         return do_proc_dointvec(table, write, filp, buffer, lenp, ppos,
2148                                 do_proc_dointvec_ms_jiffies_conv, NULL);
2149 }
2150
2151 #else /* CONFIG_PROC_FS */
2152
2153 int proc_dostring(ctl_table *table, int write, struct file *filp,
2154                   void __user *buffer, size_t *lenp, loff_t *ppos)
2155 {
2156         return -ENOSYS;
2157 }
2158
2159 static int proc_doutsstring(ctl_table *table, int write, struct file *filp,
2160                             void __user *buffer, size_t *lenp, loff_t *ppos)
2161 {
2162         return -ENOSYS;
2163 }
2164
2165 int proc_dointvec(ctl_table *table, int write, struct file *filp,
2166                   void __user *buffer, size_t *lenp, loff_t *ppos)
2167 {
2168         return -ENOSYS;
2169 }
2170
2171 int proc_dointvec_bset(ctl_table *table, int write, struct file *filp,
2172                         void __user *buffer, size_t *lenp, loff_t *ppos)
2173 {
2174         return -ENOSYS;
2175 }
2176
2177 int proc_dointvec_minmax(ctl_table *table, int write, struct file *filp,
2178                     void __user *buffer, size_t *lenp, loff_t *ppos)
2179 {
2180         return -ENOSYS;
2181 }
2182
2183 int proc_dointvec_jiffies(ctl_table *table, int write, struct file *filp,
2184                     void __user *buffer, size_t *lenp, loff_t *ppos)
2185 {
2186         return -ENOSYS;
2187 }
2188
2189 int proc_dointvec_userhz_jiffies(ctl_table *table, int write, struct file *filp,
2190                     void __user *buffer, size_t *lenp, loff_t *ppos)
2191 {
2192         return -ENOSYS;
2193 }
2194
2195 int proc_dointvec_ms_jiffies(ctl_table *table, int write, struct file *filp,
2196                              void __user *buffer, size_t *lenp, loff_t *ppos)
2197 {
2198         return -ENOSYS;
2199 }
2200
2201 int proc_doulongvec_minmax(ctl_table *table, int write, struct file *filp,
2202                     void __user *buffer, size_t *lenp, loff_t *ppos)
2203 {
2204         return -ENOSYS;
2205 }
2206
2207 int proc_doulongvec_ms_jiffies_minmax(ctl_table *table, int write,
2208                                       struct file *filp,
2209                                       void __user *buffer,
2210                                       size_t *lenp, loff_t *ppos)
2211 {
2212     return -ENOSYS;
2213 }
2214
2215
2216 #endif /* CONFIG_PROC_FS */
2217
2218
2219 /*
2220  * General sysctl support routines
2221  */
2222
2223 /* The generic string strategy routine: */
2224 int sysctl_string(ctl_table *table, int __user *name, int nlen,
2225                   void __user *oldval, size_t __user *oldlenp,
2226                   void __user *newval, size_t newlen, void **context)
2227 {
2228         if (!table->data || !table->maxlen)
2229                 return -ENOTDIR;
2230
2231         if (oldval && oldlenp) {
2232                 size_t bufsize;
2233                 if (get_user(bufsize, oldlenp))
2234                         return -EFAULT;
2235                 if (bufsize) {
2236                         size_t len = strlen(table->data), copied;
2237
2238                         /* This shouldn't trigger for a well-formed sysctl */
2239                         if (len > table->maxlen)
2240                                 len = table->maxlen;
2241
2242                         /* Copy up to a max of bufsize-1 bytes of the string */
2243                         copied = (len >= bufsize) ? bufsize - 1 : len;
2244
2245                         if (copy_to_user(oldval, table->data, copied) ||
2246                             put_user(0, (char __user *)(oldval + copied)))
2247                                 return -EFAULT;
2248                         if (put_user(len, oldlenp))
2249                                 return -EFAULT;
2250                 }
2251         }
2252         if (newval && newlen) {
2253                 size_t len = newlen;
2254                 if (len > table->maxlen)
2255                         len = table->maxlen;
2256                 if(copy_from_user(table->data, newval, len))
2257                         return -EFAULT;
2258                 if (len == table->maxlen)
2259                         len--;
2260                 ((char *) table->data)[len] = 0;
2261         }
2262         return 1;
2263 }
2264
2265 /*
2266  * This function makes sure that all of the integers in the vector
2267  * are between the minimum and maximum values given in the arrays
2268  * table->extra1 and table->extra2, respectively.
2269  */
2270 int sysctl_intvec(ctl_table *table, int __user *name, int nlen,
2271                 void __user *oldval, size_t __user *oldlenp,
2272                 void __user *newval, size_t newlen, void **context)
2273 {
2274
2275         if (newval && newlen) {
2276                 int __user *vec = (int __user *) newval;
2277                 int *min = (int *) table->extra1;
2278                 int *max = (int *) table->extra2;
2279                 size_t length;
2280                 int i;
2281
2282                 if (newlen % sizeof(int) != 0)
2283                         return -EINVAL;
2284
2285                 if (!table->extra1 && !table->extra2)
2286                         return 0;
2287
2288                 if (newlen > table->maxlen)
2289                         newlen = table->maxlen;
2290                 length = newlen / sizeof(int);
2291
2292                 for (i = 0; i < length; i++) {
2293                         int value;
2294                         if (get_user(value, vec + i))
2295                                 return -EFAULT;
2296                         if (min && value < min[i])
2297                                 return -EINVAL;
2298                         if (max && value > max[i])
2299                                 return -EINVAL;
2300                 }
2301         }
2302         return 0;
2303 }
2304
2305 /* Strategy function to convert jiffies to seconds */
2306 int sysctl_jiffies(ctl_table *table, int __user *name, int nlen,
2307                 void __user *oldval, size_t __user *oldlenp,
2308                 void __user *newval, size_t newlen, void **context)
2309 {
2310         if (oldval) {
2311                 size_t olen;
2312                 if (oldlenp) {
2313                         if (get_user(olen, oldlenp))
2314                                 return -EFAULT;
2315                         if (olen!=sizeof(int))
2316                                 return -EINVAL;
2317                 }
2318                 if (put_user(*(int *)(table->data)/HZ, (int __user *)oldval) ||
2319                     (oldlenp && put_user(sizeof(int),oldlenp)))
2320                         return -EFAULT;
2321         }
2322         if (newval && newlen) {
2323                 int new;
2324                 if (newlen != sizeof(int))
2325                         return -EINVAL;
2326                 if (get_user(new, (int __user *)newval))
2327                         return -EFAULT;
2328                 *(int *)(table->data) = new*HZ;
2329         }
2330         return 1;
2331 }
2332
2333 /* Strategy function to convert jiffies to seconds */
2334 int sysctl_ms_jiffies(ctl_table *table, int __user *name, int nlen,
2335                 void __user *oldval, size_t __user *oldlenp,
2336                 void __user *newval, size_t newlen, void **context)
2337 {
2338         if (oldval) {
2339                 size_t olen;
2340                 if (oldlenp) {
2341                         if (get_user(olen, oldlenp))
2342                                 return -EFAULT;
2343                         if (olen!=sizeof(int))
2344                                 return -EINVAL;
2345                 }
2346                 if (put_user(jiffies_to_msecs(*(int *)(table->data)), (int __user *)oldval) ||
2347                     (oldlenp && put_user(sizeof(int),oldlenp)))
2348                         return -EFAULT;
2349         }
2350         if (newval && newlen) {
2351                 int new;
2352                 if (newlen != sizeof(int))
2353                         return -EINVAL;
2354                 if (get_user(new, (int __user *)newval))
2355                         return -EFAULT;
2356                 *(int *)(table->data) = msecs_to_jiffies(new);
2357         }
2358         return 1;
2359 }
2360
2361 #else /* CONFIG_SYSCTL */
2362
2363
2364 asmlinkage long sys_sysctl(struct __sysctl_args __user *args)
2365 {
2366         return -ENOSYS;
2367 }
2368
2369 int sysctl_string(ctl_table *table, int __user *name, int nlen,
2370                   void __user *oldval, size_t __user *oldlenp,
2371                   void __user *newval, size_t newlen, void **context)
2372 {
2373         return -ENOSYS;
2374 }
2375
2376 int sysctl_intvec(ctl_table *table, int __user *name, int nlen,
2377                 void __user *oldval, size_t __user *oldlenp,
2378                 void __user *newval, size_t newlen, void **context)
2379 {
2380         return -ENOSYS;
2381 }
2382
2383 int sysctl_jiffies(ctl_table *table, int __user *name, int nlen,
2384                 void __user *oldval, size_t __user *oldlenp,
2385                 void __user *newval, size_t newlen, void **context)
2386 {
2387         return -ENOSYS;
2388 }
2389
2390 int sysctl_ms_jiffies(ctl_table *table, int __user *name, int nlen,
2391                 void __user *oldval, size_t __user *oldlenp,
2392                 void __user *newval, size_t newlen, void **context)
2393 {
2394         return -ENOSYS;
2395 }
2396
2397 int proc_dostring(ctl_table *table, int write, struct file *filp,
2398                   void __user *buffer, size_t *lenp, loff_t *ppos)
2399 {
2400         return -ENOSYS;
2401 }
2402
2403 int proc_dointvec(ctl_table *table, int write, struct file *filp,
2404                   void __user *buffer, size_t *lenp, loff_t *ppos)
2405 {
2406         return -ENOSYS;
2407 }
2408
2409 int proc_dointvec_bset(ctl_table *table, int write, struct file *filp,
2410                         void __user *buffer, size_t *lenp, loff_t *ppos)
2411 {
2412         return -ENOSYS;
2413 }
2414
2415 int proc_dointvec_minmax(ctl_table *table, int write, struct file *filp,
2416                     void __user *buffer, size_t *lenp, loff_t *ppos)
2417 {
2418         return -ENOSYS;
2419 }
2420
2421 int proc_dointvec_jiffies(ctl_table *table, int write, struct file *filp,
2422                           void __user *buffer, size_t *lenp, loff_t *ppos)
2423 {
2424         return -ENOSYS;
2425 }
2426
2427 int proc_dointvec_userhz_jiffies(ctl_table *table, int write, struct file *filp,
2428                           void __user *buffer, size_t *lenp, loff_t *ppos)
2429 {
2430         return -ENOSYS;
2431 }
2432
2433 int proc_dointvec_ms_jiffies(ctl_table *table, int write, struct file *filp,
2434                              void __user *buffer, size_t *lenp, loff_t *ppos)
2435 {
2436         return -ENOSYS;
2437 }
2438
2439 int proc_doulongvec_minmax(ctl_table *table, int write, struct file *filp,
2440                     void __user *buffer, size_t *lenp, loff_t *ppos)
2441 {
2442         return -ENOSYS;
2443 }
2444
2445 int proc_doulongvec_ms_jiffies_minmax(ctl_table *table, int write,
2446                                       struct file *filp,
2447                                       void __user *buffer,
2448                                       size_t *lenp, loff_t *ppos)
2449 {
2450     return -ENOSYS;
2451 }
2452
2453 struct ctl_table_header * register_sysctl_table(ctl_table * table,
2454                                                 int insert_at_head)
2455 {
2456         return NULL;
2457 }
2458
2459 void unregister_sysctl_table(struct ctl_table_header * table)
2460 {
2461 }
2462
2463 #endif /* CONFIG_SYSCTL */
2464
2465 /*
2466  * No sense putting this after each symbol definition, twice,
2467  * exception granted :-)
2468  */
2469 EXPORT_SYMBOL(proc_dointvec);
2470 EXPORT_SYMBOL(proc_dointvec_jiffies);
2471 EXPORT_SYMBOL(proc_dointvec_minmax);
2472 EXPORT_SYMBOL(proc_dointvec_userhz_jiffies);
2473 EXPORT_SYMBOL(proc_dointvec_ms_jiffies);
2474 EXPORT_SYMBOL(proc_dostring);
2475 EXPORT_SYMBOL(proc_doulongvec_minmax);
2476 EXPORT_SYMBOL(proc_doulongvec_ms_jiffies_minmax);
2477 EXPORT_SYMBOL(register_sysctl_table);
2478 EXPORT_SYMBOL(sysctl_intvec);
2479 EXPORT_SYMBOL(sysctl_jiffies);
2480 EXPORT_SYMBOL(sysctl_ms_jiffies);
2481 EXPORT_SYMBOL(sysctl_string);
2482 EXPORT_SYMBOL(unregister_sysctl_table);