kernel/sysctl.c

   1 /*
   2  * sysctl.c: General linux system control interface
   3  *
   4  * Begun 24 March 1995, Stephen Tweedie
   5  * Added /proc support, Dec 1995
   6  * Added bdflush entry and intvec min/max checking, 2/23/96, Tom Dyas.
   7  * Added hooks for /proc/sys/net (minor, minor patch), 96/4/1, Mike Shaver.
   8  * Added kernel/java-{interpreter,appletviewer}, 96/5/10, Mike Shaver.
   9  * Dynamic registration fixes, Stephen Tweedie.
  10  * Added kswapd-interval, ctrl-alt-del, printk stuff, 1/8/97, Chris Horn.
  11  * Made sysctl support optional via CONFIG_SYSCTL, 1/10/97, Chris
  12  *  Horn.
  13  * Added proc_doulongvec_ms_jiffies_minmax, 09/08/99, Carlos H. Bauer.
  14  * Added proc_doulongvec_minmax, 09/08/99, Carlos H. Bauer.
  15  * Changed linked lists to use list.h instead of lists.h, 02/24/00, Bill
  16  *  Wendling.
  17  * The list_for_each() macro wasn't appropriate for the sysctl loop.
  18  *  Removed it and replaced it with older style, 03/23/00, Bill Wendling
  19  */
  20
  21 #include <linux/config.h>
  22 #include <linux/module.h>
  23 #include <linux/mm.h>
  24 #include <linux/swap.h>
  25 #include <linux/slab.h>
  26 #include <linux/sysctl.h>
  27 #include <linux/proc_fs.h>
  28 #include <linux/ctype.h>
  29 #include <linux/utsname.h>
  30 #include <linux/capability.h>
  31 #include <linux/smp_lock.h>
  32 #include <linux/init.h>
  33 #include <linux/kernel.h>
  34 #include <linux/sysrq.h>
  35 #include <linux/highuid.h>
  36 #include <linux/writeback.h>
  37 #include <linux/hugetlb.h>
  38 #include <linux/security.h>
  39 #include <linux/initrd.h>
  40 #include <linux/times.h>
  41 #include <linux/limits.h>
  42 #include <linux/dcache.h>
  43 #include <linux/syscalls.h>
  44
  45 #include <asm/uaccess.h>
  46 #include <asm/processor.h>
  47
  48 #ifdef CONFIG_ROOT_NFS
  49 #include <linux/nfs_fs.h>
  50 #endif
  51
  52 #if defined(CONFIG_SYSCTL)
  53
  54 /* External variables not in a header file. */
  55 extern int C_A_D;
  56 extern int sysctl_overcommit_memory;
  57 extern int sysctl_overcommit_ratio;
  58 extern int max_threads;
  59 extern int sysrq_enabled;
  60 extern int core_uses_pid;
  61 extern int suid_dumpable;
  62 extern char core_pattern[];
  63 extern int cad_pid;
  64 extern int pid_max;
  65 extern int min_free_kbytes;
  66 extern int printk_ratelimit_jiffies;
  67 extern int printk_ratelimit_burst;
  68 extern int pid_max_min, pid_max_max;
  69
  70 #if defined(CONFIG_X86_LOCAL_APIC) && defined(CONFIG_X86)
  71 int unknown_nmi_panic;
  72 extern int proc_unknown_nmi_panic(ctl_table *, int, struct file *,
  73                                   void __user *, size_t *, loff_t *);
  74 #endif
  75
  76 /* this is needed for the proc_dointvec_minmax for [fs_]overflow UID and GID */
  77 static int maxolduid = 65535;
  78 static int minolduid;
  79
  80 static int ngroups_max = NGROUPS_MAX;
  81
  82 #ifdef CONFIG_KMOD
  83 extern char modprobe_path[];
  84 #endif
  85 #ifdef CONFIG_HOTPLUG
  86 extern char hotplug_path[];
  87 #endif
  88 #ifdef CONFIG_CHR_DEV_SG
  89 extern int sg_big_buff;
  90 #endif
  91 #ifdef CONFIG_SYSVIPC
  92 extern size_t shm_ctlmax;
  93 extern size_t shm_ctlall;
  94 extern int shm_ctlmni;
  95 extern int msg_ctlmax;
  96 extern int msg_ctlmnb;
  97 extern int msg_ctlmni;
  98 extern int sem_ctls[];
  99 #endif
 100
 101 #ifdef __sparc__
 102 extern char reboot_command [];
 103 extern int stop_a_enabled;
 104 extern int scons_pwroff;
 105 #endif
 106
 107 #ifdef __hppa__
 108 extern int pwrsw_enabled;
 109 extern int unaligned_enabled;
 110 #endif
 111
 112 #ifdef CONFIG_ARCH_S390
 113 #ifdef CONFIG_MATHEMU
 114 extern int sysctl_ieee_emulation_warnings;
 115 #endif
 116 extern int sysctl_userprocess_debug;
 117 extern int spin_retry;
 118 #endif
 119
 120 extern int sysctl_hz_timer;
 121
 122 #ifdef CONFIG_BSD_PROCESS_ACCT
 123 extern int acct_parm[];
 124 #endif
 125
 126 int randomize_va_space = 1;
 127
 128 static int parse_table(int __user *, int, void __user *, size_t __user *, void __user *, size_t,
 129                        ctl_table *, void **);
 130 static int proc_doutsstring(ctl_table *table, int write, struct file *filp,
 131                   void __user *buffer, size_t *lenp, loff_t *ppos);
 132
 133 static ctl_table root_table[];
 134 static struct ctl_table_header root_table_header =
 135         { root_table, LIST_HEAD_INIT(root_table_header.ctl_entry) };
 136
 137 static ctl_table kern_table[];
 138 static ctl_table vm_table[];
 139 #ifdef CONFIG_NET
 140 extern ctl_table net_table[];
 141 #endif
 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           = &hotplug_path,
 403                 .maxlen         = HOTPLUG_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_ARCH_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_ARCH_S390)
 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_MIN_FREE_KBYTES,
 783                 .procname       = "min_free_kbytes",
 784                 .data           = &min_free_kbytes,
 785                 .maxlen         = sizeof(min_free_kbytes),
 786                 .mode           = 0644,
 787                 .proc_handler   = &min_free_kbytes_sysctl_handler,
 788                 .strategy       = &sysctl_intvec,
 789                 .extra1         = &zero,
 790         },
 791 #ifdef CONFIG_MMU
 792         {
 793                 .ctl_name       = VM_MAX_MAP_COUNT,
 794                 .procname       = "max_map_count",
 795                 .data           = &sysctl_max_map_count,
 796                 .maxlen         = sizeof(sysctl_max_map_count),
 797                 .mode           = 0644,
 798                 .proc_handler   = &proc_dointvec
 799         },
 800 #endif
 801         {
 802                 .ctl_name       = VM_LAPTOP_MODE,
 803                 .procname       = "laptop_mode",
 804                 .data           = &laptop_mode,
 805                 .maxlen         = sizeof(laptop_mode),
 806                 .mode           = 0644,
 807                 .proc_handler   = &proc_dointvec,
 808                 .strategy       = &sysctl_intvec,
 809                 .extra1         = &zero,
 810         },
 811         {
 812                 .ctl_name       = VM_BLOCK_DUMP,
 813                 .procname       = "block_dump",
 814                 .data           = &block_dump,
 815                 .maxlen         = sizeof(block_dump),
 816                 .mode           = 0644,
 817                 .proc_handler   = &proc_dointvec,
 818                 .strategy       = &sysctl_intvec,
 819                 .extra1         = &zero,
 820         },
 821         {
 822                 .ctl_name       = VM_VFS_CACHE_PRESSURE,
 823                 .procname       = "vfs_cache_pressure",
 824                 .data           = &sysctl_vfs_cache_pressure,
 825                 .maxlen         = sizeof(sysctl_vfs_cache_pressure),
 826                 .mode           = 0644,
 827                 .proc_handler   = &proc_dointvec,
 828                 .strategy       = &sysctl_intvec,
 829                 .extra1         = &zero,
 830         },
 831 #ifdef HAVE_ARCH_PICK_MMAP_LAYOUT
 832         {
 833                 .ctl_name       = VM_LEGACY_VA_LAYOUT,
 834                 .procname       = "legacy_va_layout",
 835                 .data           = &sysctl_legacy_va_layout,
 836                 .maxlen         = sizeof(sysctl_legacy_va_layout),
 837                 .mode           = 0644,
 838                 .proc_handler   = &proc_dointvec,
 839                 .strategy       = &sysctl_intvec,
 840                 .extra1         = &zero,
 841         },
 842 #endif
 843 #ifdef CONFIG_SWAP
 844         {
 845                 .ctl_name       = VM_SWAP_TOKEN_TIMEOUT,
 846                 .procname       = "swap_token_timeout",
 847                 .data           = &swap_token_default_timeout,
 848                 .maxlen         = sizeof(swap_token_default_timeout),
 849                 .mode           = 0644,
 850                 .proc_handler   = &proc_dointvec_jiffies,
 851                 .strategy       = &sysctl_jiffies,
 852         },
 853 #endif
 854         { .ctl_name = 0 }
 855 };
 856
 857 static ctl_table proc_table[] = {
 858         { .ctl_name = 0 }
 859 };
 860
 861 static ctl_table fs_table[] = {
 862         {
 863                 .ctl_name       = FS_NRINODE,
 864                 .procname       = "inode-nr",
 865                 .data           = &inodes_stat,
 866                 .maxlen         = 2*sizeof(int),
 867                 .mode           = 0444,
 868                 .proc_handler   = &proc_dointvec,
 869         },
 870         {
 871                 .ctl_name       = FS_STATINODE,
 872                 .procname       = "inode-state",
 873                 .data           = &inodes_stat,
 874                 .maxlen         = 7*sizeof(int),
 875                 .mode           = 0444,
 876                 .proc_handler   = &proc_dointvec,
 877         },
 878         {
 879                 .ctl_name       = FS_NRFILE,
 880                 .procname       = "file-nr",
 881                 .data           = &files_stat,
 882                 .maxlen         = 3*sizeof(int),
 883                 .mode           = 0444,
 884                 .proc_handler   = &proc_dointvec,
 885         },
 886         {
 887                 .ctl_name       = FS_MAXFILE,
 888                 .procname       = "file-max",
 889                 .data           = &files_stat.max_files,
 890                 .maxlen         = sizeof(int),
 891                 .mode           = 0644,
 892                 .proc_handler   = &proc_dointvec,
 893         },
 894         {
 895                 .ctl_name       = FS_DENTRY,
 896                 .procname       = "dentry-state",
 897                 .data           = &dentry_stat,
 898                 .maxlen         = 6*sizeof(int),
 899                 .mode           = 0444,
 900                 .proc_handler   = &proc_dointvec,
 901         },
 902         {
 903                 .ctl_name       = FS_OVERFLOWUID,
 904                 .procname       = "overflowuid",
 905                 .data           = &fs_overflowuid,
 906                 .maxlen         = sizeof(int),
 907                 .mode           = 0644,
 908                 .proc_handler   = &proc_dointvec_minmax,
 909                 .strategy       = &sysctl_intvec,
 910                 .extra1         = &minolduid,
 911                 .extra2         = &maxolduid,
 912         },
 913         {
 914                 .ctl_name       = FS_OVERFLOWGID,
 915                 .procname       = "overflowgid",
 916                 .data           = &fs_overflowgid,
 917                 .maxlen         = sizeof(int),
 918                 .mode           = 0644,
 919                 .proc_handler   = &proc_dointvec_minmax,
 920                 .strategy       = &sysctl_intvec,
 921                 .extra1         = &minolduid,
 922                 .extra2         = &maxolduid,
 923         },
 924         {
 925                 .ctl_name       = FS_LEASES,
 926                 .procname       = "leases-enable",
 927                 .data           = &leases_enable,
 928                 .maxlen         = sizeof(int),
 929                 .mode           = 0644,
 930                 .proc_handler   = &proc_dointvec,
 931         },
 932 #ifdef CONFIG_DNOTIFY
 933         {
 934                 .ctl_name       = FS_DIR_NOTIFY,
 935                 .procname       = "dir-notify-enable",
 936                 .data           = &dir_notify_enable,
 937                 .maxlen         = sizeof(int),
 938                 .mode           = 0644,
 939                 .proc_handler   = &proc_dointvec,
 940         },
 941 #endif
 942 #ifdef CONFIG_MMU
 943         {
 944                 .ctl_name       = FS_LEASE_TIME,
 945                 .procname       = "lease-break-time",
 946                 .data           = &lease_break_time,
 947                 .maxlen         = sizeof(int),
 948                 .mode           = 0644,
 949                 .proc_handler   = &proc_dointvec,
 950         },
 951         {
 952                 .ctl_name       = FS_AIO_NR,
 953                 .procname       = "aio-nr",
 954                 .data           = &aio_nr,
 955                 .maxlen         = sizeof(aio_nr),
 956                 .mode           = 0444,
 957                 .proc_handler   = &proc_dointvec,
 958         },
 959         {
 960                 .ctl_name       = FS_AIO_MAX_NR,
 961                 .procname       = "aio-max-nr",
 962                 .data           = &aio_max_nr,
 963                 .maxlen         = sizeof(aio_max_nr),
 964                 .mode           = 0644,
 965                 .proc_handler   = &proc_dointvec,
 966         },
 967 #ifdef CONFIG_INOTIFY
 968         {
 969                 .ctl_name       = FS_INOTIFY,
 970                 .procname       = "inotify",
 971                 .mode           = 0555,
 972                 .child          = inotify_table,
 973         },
 974 #endif
 975 #endif
 976         {
 977                 .ctl_name       = KERN_SETUID_DUMPABLE,
 978                 .procname       = "suid_dumpable",
 979                 .data           = &suid_dumpable,
 980                 .maxlen         = sizeof(int),
 981                 .mode           = 0644,
 982                 .proc_handler   = &proc_dointvec,
 983         },
 984         { .ctl_name = 0 }
 985 };
 986
 987 static ctl_table debug_table[] = {
 988         { .ctl_name = 0 }
 989 };
 990
 991 static ctl_table dev_table[] = {
 992         { .ctl_name = 0 }
 993 };
 994
 995 extern void init_irq_proc (void);
 996
 997 static DEFINE_SPINLOCK(sysctl_lock);
 998
 999 /* called under sysctl_lock */
1000 static int use_table(struct ctl_table_header *p)
1001 {
1002         if (unlikely(p->unregistering))
1003                 return 0;
1004         p->used++;
1005         return 1;
1006 }
1007
1008 /* called under sysctl_lock */
1009 static void unuse_table(struct ctl_table_header *p)
1010 {
1011         if (!--p->used)
1012                 if (unlikely(p->unregistering))
1013                         complete(p->unregistering);
1014 }
1015
1016 /* called under sysctl_lock, will reacquire if has to wait */
1017 static void start_unregistering(struct ctl_table_header *p)
1018 {
1019         /*
1020          * if p->used is 0, nobody will ever touch that entry again;
1021          * we'll eliminate all paths to it before dropping sysctl_lock
1022          */
1023         if (unlikely(p->used)) {
1024                 struct completion wait;
1025                 init_completion(&wait);
1026                 p->unregistering = &wait;
1027                 spin_unlock(&sysctl_lock);
1028                 wait_for_completion(&wait);
1029                 spin_lock(&sysctl_lock);
1030         }
1031         /*
1032          * do not remove from the list until nobody holds it; walking the
1033          * list in do_sysctl() relies on that.
1034          */
1035         list_del_init(&p->ctl_entry);
1036 }
1037
1038 void __init sysctl_init(void)
1039 {
1040 #ifdef CONFIG_PROC_FS
1041         register_proc_table(root_table, proc_sys_root, &root_table_header);
1042         init_irq_proc();
1043 #endif
1044 }
1045
1046 int do_sysctl(int __user *name, int nlen, void __user *oldval, size_t __user *oldlenp,
1047                void __user *newval, size_t newlen)
1048 {
1049         struct list_head *tmp;
1050         int error = -ENOTDIR;
1051
1052         if (nlen <= 0 || nlen >= CTL_MAXNAME)
1053                 return -ENOTDIR;
1054         if (oldval) {
1055                 int old_len;
1056                 if (!oldlenp || get_user(old_len, oldlenp))
1057                         return -EFAULT;
1058         }
1059         spin_lock(&sysctl_lock);
1060         tmp = &root_table_header.ctl_entry;
1061         do {
1062                 struct ctl_table_header *head =
1063                         list_entry(tmp, struct ctl_table_header, ctl_entry);
1064                 void *context = NULL;
1065
1066                 if (!use_table(head))
1067                         continue;
1068
1069                 spin_unlock(&sysctl_lock);
1070
1071                 error = parse_table(name, nlen, oldval, oldlenp,
1072                                         newval, newlen, head->ctl_table,
1073                                         &context);
1074                 kfree(context);
1075
1076                 spin_lock(&sysctl_lock);
1077                 unuse_table(head);
1078                 if (error != -ENOTDIR)
1079                         break;
1080         } while ((tmp = tmp->next) != &root_table_header.ctl_entry);
1081         spin_unlock(&sysctl_lock);
1082         return error;
1083 }
1084
1085 asmlinkage long sys_sysctl(struct __sysctl_args __user *args)
1086 {
1087         struct __sysctl_args tmp;
1088         int error;
1089
1090         if (copy_from_user(&tmp, args, sizeof(tmp)))
1091                 return -EFAULT;
1092
1093         lock_kernel();
1094         error = do_sysctl(tmp.name, tmp.nlen, tmp.oldval, tmp.oldlenp,
1095                           tmp.newval, tmp.newlen);
1096         unlock_kernel();
1097         return error;
1098 }
1099
1100 /*
1101  * ctl_perm does NOT grant the superuser all rights automatically, because
1102  * some sysctl variables are readonly even to root.
1103  */
1104
1105 static int test_perm(int mode, int op)
1106 {
1107         if (!current->euid)
1108                 mode >>= 6;
1109         else if (in_egroup_p(0))
1110                 mode >>= 3;
1111         if ((mode & op & 0007) == op)
1112                 return 0;
1113         return -EACCES;
1114 }
1115
1116 static inline int ctl_perm(ctl_table *table, int op)
1117 {
1118         int error;
1119         error = security_sysctl(table, op);
1120         if (error)
1121                 return error;
1122         return test_perm(table->mode, op);
1123 }
1124
1125 static int parse_table(int __user *name, int nlen,
1126                        void __user *oldval, size_t __user *oldlenp,
1127                        void __user *newval, size_t newlen,
1128                        ctl_table *table, void **context)
1129 {
1130         int n;
1131 repeat:
1132         if (!nlen)
1133                 return -ENOTDIR;
1134         if (get_user(n, name))
1135                 return -EFAULT;
1136         for ( ; table->ctl_name; table++) {
1137                 if (n == table->ctl_name || table->ctl_name == CTL_ANY) {
1138                         int error;
1139                         if (table->child) {
1140                                 if (ctl_perm(table, 001))
1141                                         return -EPERM;
1142                                 if (table->strategy) {
1143                                         error = table->strategy(
1144                                                 table, name, nlen,
1145                                                 oldval, oldlenp,
1146                                                 newval, newlen, context);
1147                                         if (error)
1148                                                 return error;
1149                                 }
1150                                 name++;
1151                                 nlen--;
1152                                 table = table->child;
1153                                 goto repeat;
1154                         }
1155                         error = do_sysctl_strategy(table, name, nlen,
1156                                                    oldval, oldlenp,
1157                                                    newval, newlen, context);
1158                         return error;
1159                 }
1160         }
1161         return -ENOTDIR;
1162 }
1163
1164 /* Perform the actual read/write of a sysctl table entry. */
1165 int do_sysctl_strategy (ctl_table *table,
1166                         int __user *name, int nlen,
1167                         void __user *oldval, size_t __user *oldlenp,
1168                         void __user *newval, size_t newlen, void **context)
1169 {
1170         int op = 0, rc;
1171         size_t len;
1172
1173         if (oldval)
1174                 op |= 004;
1175         if (newval)
1176                 op |= 002;
1177         if (ctl_perm(table, op))
1178                 return -EPERM;
1179
1180         if (table->strategy) {
1181                 rc = table->strategy(table, name, nlen, oldval, oldlenp,
1182                                      newval, newlen, context);
1183                 if (rc < 0)
1184                         return rc;
1185                 if (rc > 0)
1186                         return 0;
1187         }
1188
1189         /* If there is no strategy routine, or if the strategy returns
1190          * zero, proceed with automatic r/w */
1191         if (table->data && table->maxlen) {
1192                 if (oldval && oldlenp) {
1193                         if (get_user(len, oldlenp))
1194                                 return -EFAULT;
1195                         if (len) {
1196                                 if (len > table->maxlen)
1197                                         len = table->maxlen;
1198                                 if(copy_to_user(oldval, table->data, len))
1199                                         return -EFAULT;
1200                                 if(put_user(len, oldlenp))
1201                                         return -EFAULT;
1202                         }
1203                 }
1204                 if (newval && newlen) {
1205                         len = newlen;
1206                         if (len > table->maxlen)
1207                                 len = table->maxlen;
1208                         if(copy_from_user(table->data, newval, len))
1209                                 return -EFAULT;
1210                 }
1211         }
1212         return 0;
1213 }
1214
1215 /**
1216  * register_sysctl_table - register a sysctl hierarchy
1217  * @table: the top-level table structure
1218  * @insert_at_head: whether the entry should be inserted in front or at the end
1219  *
1220  * Register a sysctl table hierarchy. @table should be a filled in ctl_table
1221  * array. An entry with a ctl_name of 0 terminates the table.
1222  *
1223  * The members of the &ctl_table structure are used as follows:
1224  *
1225  * ctl_name - This is the numeric sysctl value used by sysctl(2). The number
1226  *            must be unique within that level of sysctl
1227  *
1228  * procname - the name of the sysctl file under /proc/sys. Set to %NULL to not
1229  *            enter a sysctl file
1230  *
1231  * data - a pointer to data for use by proc_handler
1232  *
1233  * maxlen - the maximum size in bytes of the data
1234  *
1235  * mode - the file permissions for the /proc/sys file, and for sysctl(2)
1236  *
1237  * child - a pointer to the child sysctl table if this entry is a directory, or
1238  *         %NULL.
1239  *
1240  * proc_handler - the text handler routine (described below)
1241  *
1242  * strategy - the strategy routine (described below)
1243  *
1244  * de - for internal use by the sysctl routines
1245  *
1246  * extra1, extra2 - extra pointers usable by the proc handler routines
1247  *
1248  * Leaf nodes in the sysctl tree will be represented by a single file
1249  * under /proc; non-leaf nodes will be represented by directories.
1250  *
1251  * sysctl(2) can automatically manage read and write requests through
1252  * the sysctl table.  The data and maxlen fields of the ctl_table
1253  * struct enable minimal validation of the values being written to be
1254  * performed, and the mode field allows minimal authentication.
1255  *
1256  * More sophisticated management can be enabled by the provision of a
1257  * strategy routine with the table entry.  This will be called before
1258  * any automatic read or write of the data is performed.
1259  *
1260  * The strategy routine may return
1261  *
1262  * < 0 - Error occurred (error is passed to user process)
1263  *
1264  * 0   - OK - proceed with automatic read or write.
1265  *
1266  * > 0 - OK - read or write has been done by the strategy routine, so
1267  *       return immediately.
1268  *
1269  * There must be a proc_handler routine for any terminal nodes
1270  * mirrored under /proc/sys (non-terminals are handled by a built-in
1271  * directory handler).  Several default handlers are available to
1272  * cover common cases -
1273  *
1274  * proc_dostring(), proc_dointvec(), proc_dointvec_jiffies(),
1275  * proc_dointvec_userhz_jiffies(), proc_dointvec_minmax(),
1276  * proc_doulongvec_ms_jiffies_minmax(), proc_doulongvec_minmax()
1277  *
1278  * It is the handler's job to read the input buffer from user memory
1279  * and process it. The handler should return 0 on success.
1280  *
1281  * This routine returns %NULL on a failure to register, and a pointer
1282  * to the table header on success.
1283  */
1284 struct ctl_table_header *register_sysctl_table(ctl_table * table,
1285                                                int insert_at_head)
1286 {
1287         struct ctl_table_header *tmp;
1288         tmp = kmalloc(sizeof(struct ctl_table_header), GFP_KERNEL);
1289         if (!tmp)
1290                 return NULL;
1291         tmp->ctl_table = table;
1292         INIT_LIST_HEAD(&tmp->ctl_entry);
1293         tmp->used = 0;
1294         tmp->unregistering = NULL;
1295         spin_lock(&sysctl_lock);
1296         if (insert_at_head)
1297                 list_add(&tmp->ctl_entry, &root_table_header.ctl_entry);
1298         else
1299                 list_add_tail(&tmp->ctl_entry, &root_table_header.ctl_entry);
1300         spin_unlock(&sysctl_lock);
1301 #ifdef CONFIG_PROC_FS
1302         register_proc_table(table, proc_sys_root, tmp);
1303 #endif
1304         return tmp;
1305 }
1306
1307 /**
1308  * unregister_sysctl_table - unregister a sysctl table hierarchy
1309  * @header: the header returned from register_sysctl_table
1310  *
1311  * Unregisters the sysctl table and all children. proc entries may not
1312  * actually be removed until they are no longer used by anyone.
1313  */
1314 void unregister_sysctl_table(struct ctl_table_header * header)
1315 {
1316         might_sleep();
1317         spin_lock(&sysctl_lock);
1318         start_unregistering(header);
1319 #ifdef CONFIG_PROC_FS
1320         unregister_proc_table(header->ctl_table, proc_sys_root);
1321 #endif
1322         spin_unlock(&sysctl_lock);
1323         kfree(header);
1324 }
1325
1326 /*
1327  * /proc/sys support
1328  */
1329
1330 #ifdef CONFIG_PROC_FS
1331
1332 /* Scan the sysctl entries in table and add them all into /proc */
1333 static void register_proc_table(ctl_table * table, struct proc_dir_entry *root, void *set)
1334 {
1335         struct proc_dir_entry *de;
1336         int len;
1337         mode_t mode;
1338
1339         for (; table->ctl_name; table++) {
1340                 /* Can't do anything without a proc name. */
1341                 if (!table->procname)
1342                         continue;
1343                 /* Maybe we can't do anything with it... */
1344                 if (!table->proc_handler && !table->child) {
1345                         printk(KERN_WARNING "SYSCTL: Can't register %s\n",
1346                                 table->procname);
1347                         continue;
1348                 }
1349
1350                 len = strlen(table->procname);
1351                 mode = table->mode;
1352
1353                 de = NULL;
1354                 if (table->proc_handler)
1355                         mode |= S_IFREG;
1356                 else {
1357                         mode |= S_IFDIR;
1358                         for (de = root->subdir; de; de = de->next) {
1359                                 if (proc_match(len, table->procname, de))
1360                                         break;
1361                         }
1362                         /* If the subdir exists already, de is non-NULL */
1363                 }
1364
1365                 if (!de) {
1366                         de = create_proc_entry(table->procname, mode, root);
1367                         if (!de)
1368                                 continue;
1369                         de->set = set;
1370                         de->data = (void *) table;
1371                         if (table->proc_handler)
1372                                 de->proc_fops = &proc_sys_file_operations;
1373                 }
1374                 table->de = de;
1375                 if (de->mode & S_IFDIR)
1376                         register_proc_table(table->child, de, set);
1377         }
1378 }
1379
1380 /*
1381  * Unregister a /proc sysctl table and any subdirectories.
1382  */
1383 static void unregister_proc_table(ctl_table * table, struct proc_dir_entry *root)
1384 {
1385         struct proc_dir_entry *de;
1386         for (; table->ctl_name; table++) {
1387                 if (!(de = table->de))
1388                         continue;
1389                 if (de->mode & S_IFDIR) {
1390                         if (!table->child) {
1391                                 printk (KERN_ALERT "Help - malformed sysctl tree on free\n");
1392                                 continue;
1393                         }
1394                         unregister_proc_table(table->child, de);
1395
1396                         /* Don't unregister directories which still have entries.. */
1397                         if (de->subdir)
1398                                 continue;
1399                 }
1400
1401                 /*
1402                  * In any case, mark the entry as goner; we'll keep it
1403                  * around if it's busy, but we'll know to do nothing with
1404                  * its fields.  We are under sysctl_lock here.
1405                  */
1406                 de->data = NULL;
1407
1408                 /* Don't unregister proc entries that are still being used.. */
1409                 if (atomic_read(&de->count))
1410                         continue;
1411
1412                 table->de = NULL;
1413                 remove_proc_entry(table->procname, root);
1414         }
1415 }
1416
1417 static ssize_t do_rw_proc(int write, struct file * file, char __user * buf,
1418                           size_t count, loff_t *ppos)
1419 {
1420         int op;
1421         struct proc_dir_entry *de = PDE(file->f_dentry->d_inode);
1422         struct ctl_table *table;
1423         size_t res;
1424         ssize_t error = -ENOTDIR;
1425
1426         spin_lock(&sysctl_lock);
1427         if (de && de->data && use_table(de->set)) {
1428                 /*
1429                  * at that point we know that sysctl was not unregistered
1430                  * and won't be until we finish
1431                  */
1432                 spin_unlock(&sysctl_lock);
1433                 table = (struct ctl_table *) de->data;
1434                 if (!table || !table->proc_handler)
1435                         goto out;
1436                 error = -EPERM;
1437                 op = (write ? 002 : 004);
1438                 if (ctl_perm(table, op))
1439                         goto out;
1440
1441                 /* careful: calling conventions are nasty here */
1442                 res = count;
1443                 error = (*table->proc_handler)(table, write, file,
1444                                                 buf, &res, ppos);
1445                 if (!error)
1446                         error = res;
1447         out:
1448                 spin_lock(&sysctl_lock);
1449                 unuse_table(de->set);
1450         }
1451         spin_unlock(&sysctl_lock);
1452         return error;
1453 }
1454
1455 static int proc_opensys(struct inode *inode, struct file *file)
1456 {
1457         if (file->f_mode & FMODE_WRITE) {
1458                 /*
1459                  * sysctl entries that are not writable,
1460                  * are _NOT_ writable, capabilities or not.
1461                  */
1462                 if (!(inode->i_mode & S_IWUSR))
1463                         return -EPERM;
1464         }
1465
1466         return 0;
1467 }
1468
1469 static ssize_t proc_readsys(struct file * file, char __user * buf,
1470                             size_t count, loff_t *ppos)
1471 {
1472         return do_rw_proc(0, file, buf, count, ppos);
1473 }
1474
1475 static ssize_t proc_writesys(struct file * file, const char __user * buf,
1476                              size_t count, loff_t *ppos)
1477 {
1478         return do_rw_proc(1, file, (char __user *) buf, count, ppos);
1479 }
1480
1481 /**
1482  * proc_dostring - read a string sysctl
1483  * @table: the sysctl table
1484  * @write: %TRUE if this is a write to the sysctl file
1485  * @filp: the file structure
1486  * @buffer: the user buffer
1487  * @lenp: the size of the user buffer
1488  * @ppos: file position
1489  *
1490  * Reads/writes a string from/to the user buffer. If the kernel
1491  * buffer provided is not large enough to hold the string, the
1492  * string is truncated. The copied string is %NULL-terminated.
1493  * If the string is being read by the user process, it is copied
1494  * and a newline '\n' is added. It is truncated if the buffer is
1495  * not large enough.
1496  *
1497  * Returns 0 on success.
1498  */
1499 int proc_dostring(ctl_table *table, int write, struct file *filp,
1500                   void __user *buffer, size_t *lenp, loff_t *ppos)
1501 {
1502         size_t len;
1503         char __user *p;
1504         char c;
1505
1506         if (!table->data || !table->maxlen || !*lenp ||
1507             (*ppos && !write)) {
1508                 *lenp = 0;
1509                 return 0;
1510         }
1511
1512         if (write) {
1513                 len = 0;
1514                 p = buffer;
1515                 while (len < *lenp) {
1516                         if (get_user(c, p++))
1517                                 return -EFAULT;
1518                         if (c == 0 || c == '\n')
1519                                 break;
1520                         len++;
1521                 }
1522                 if (len >= table->maxlen)
1523                         len = table->maxlen-1;
1524                 if(copy_from_user(table->data, buffer, len))
1525                         return -EFAULT;
1526                 ((char *) table->data)[len] = 0;
1527                 *ppos += *lenp;
1528         } else {
1529                 len = strlen(table->data);
1530                 if (len > table->maxlen)
1531                         len = table->maxlen;
1532                 if (len > *lenp)
1533                         len = *lenp;
1534                 if (len)
1535                         if(copy_to_user(buffer, table->data, len))
1536                                 return -EFAULT;
1537                 if (len < *lenp) {
1538                         if(put_user('\n', ((char __user *) buffer) + len))
1539                                 return -EFAULT;
1540                         len++;
1541                 }
1542                 *lenp = len;
1543                 *ppos += len;
1544         }
1545         return 0;
1546 }
1547
1548 /*
1549  *      Special case of dostring for the UTS structure. This has locks
1550  *      to observe. Should this be in kernel/sys.c ????
1551  */
1552
1553 static int proc_doutsstring(ctl_table *table, int write, struct file *filp,
1554                   void __user *buffer, size_t *lenp, loff_t *ppos)
1555 {
1556         int r;
1557
1558         if (!write) {
1559                 down_read(&uts_sem);
1560                 r=proc_dostring(table,0,filp,buffer,lenp, ppos);
1561                 up_read(&uts_sem);
1562         } else {
1563                 down_write(&uts_sem);
1564                 r=proc_dostring(table,1,filp,buffer,lenp, ppos);
1565                 up_write(&uts_sem);
1566         }
1567         return r;
1568 }
1569
1570 static int do_proc_dointvec_conv(int *negp, unsigned long *lvalp,
1571                                  int *valp,
1572                                  int write, void *data)
1573 {
1574         if (write) {
1575                 *valp = *negp ? -*lvalp : *lvalp;
1576         } else {
1577                 int val = *valp;
1578                 if (val < 0) {
1579                         *negp = -1;
1580                         *lvalp = (unsigned long)-val;
1581                 } else {
1582                         *negp = 0;
1583                         *lvalp = (unsigned long)val;
1584                 }
1585         }
1586         return 0;
1587 }
1588
1589 static int do_proc_dointvec(ctl_table *table, int write, struct file *filp,
1590                   void __user *buffer, size_t *lenp, loff_t *ppos,
1591                   int (*conv)(int *negp, unsigned long *lvalp, int *valp,
1592                               int write, void *data),
1593                   void *data)
1594 {
1595 #define TMPBUFLEN 21
1596         int *i, vleft, first=1, neg, val;
1597         unsigned long lval;
1598         size_t left, len;
1599
1600         char buf[TMPBUFLEN], *p;
1601         char __user *s = buffer;
1602
1603         if (!table->data || !table->maxlen || !*lenp ||
1604             (*ppos && !write)) {
1605                 *lenp = 0;
1606                 return 0;
1607         }
1608
1609         i = (int *) table->data;
1610         vleft = table->maxlen / sizeof(*i);
1611         left = *lenp;
1612
1613         if (!conv)
1614                 conv = do_proc_dointvec_conv;
1615
1616         for (; left && vleft--; i++, first=0) {
1617                 if (write) {
1618                         while (left) {
1619                                 char c;
1620                                 if (get_user(c, s))
1621                                         return -EFAULT;
1622                                 if (!isspace(c))
1623                                         break;
1624                                 left--;
1625                                 s++;
1626                         }
1627                         if (!left)
1628                                 break;
1629                         neg = 0;
1630                         len = left;
1631                         if (len > sizeof(buf) - 1)
1632                                 len = sizeof(buf) - 1;
1633                         if (copy_from_user(buf, s, len))
1634                                 return -EFAULT;
1635                         buf[len] = 0;
1636                         p = buf;
1637                         if (*p == '-' && left > 1) {
1638                                 neg = 1;
1639                                 left--, p++;
1640                         }
1641                         if (*p < '0' || *p > '9')
1642                                 break;
1643
1644                         lval = simple_strtoul(p, &p, 0);
1645
1646                         len = p-buf;
1647                         if ((len < left) && *p && !isspace(*p))
1648                                 break;
1649                         if (neg)
1650                                 val = -val;
1651                         s += len;
1652                         left -= len;
1653
1654                         if (conv(&neg, &lval, i, 1, data))
1655                                 break;
1656                 } else {
1657                         p = buf;
1658                         if (!first)
1659                                 *p++ = '\t';
1660
1661                         if (conv(&neg, &lval, i, 0, data))
1662                                 break;
1663
1664                         sprintf(p, "%s%lu", neg ? "-" : "", lval);
1665                         len = strlen(buf);
1666                         if (len > left)
1667                                 len = left;
1668                         if(copy_to_user(s, buf, len))
1669                                 return -EFAULT;
1670                         left -= len;
1671                         s += len;
1672                 }
1673         }
1674
1675         if (!write && !first && left) {
1676                 if(put_user('\n', s))
1677                         return -EFAULT;
1678                 left--, s++;
1679         }
1680         if (write) {
1681                 while (left) {
1682                         char c;
1683                         if (get_user(c, s++))
1684                                 return -EFAULT;
1685                         if (!isspace(c))
1686                                 break;
1687                         left--;
1688                 }
1689         }
1690         if (write && first)
1691                 return -EINVAL;
1692         *lenp -= left;
1693         *ppos += *lenp;
1694         return 0;
1695 #undef TMPBUFLEN
1696 }
1697
1698 /**
1699  * proc_dointvec - read a vector of integers
1700  * @table: the sysctl table
1701  * @write: %TRUE if this is a write to the sysctl file
1702  * @filp: the file structure
1703  * @buffer: the user buffer
1704  * @lenp: the size of the user buffer
1705  * @ppos: file position
1706  *
1707  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
1708  * values from/to the user buffer, treated as an ASCII string.
1709  *
1710  * Returns 0 on success.
1711  */
1712 int proc_dointvec(ctl_table *table, int write, struct file *filp,
1713                      void __user *buffer, size_t *lenp, loff_t *ppos)
1714 {
1715     return do_proc_dointvec(table,write,filp,buffer,lenp,ppos,
1716                             NULL,NULL);
1717 }
1718
1719 #define OP_SET  0
1720 #define OP_AND  1
1721 #define OP_OR   2
1722 #define OP_MAX  3
1723 #define OP_MIN  4
1724
1725 static int do_proc_dointvec_bset_conv(int *negp, unsigned long *lvalp,
1726                                       int *valp,
1727                                       int write, void *data)
1728 {
1729         int op = *(int *)data;
1730         if (write) {
1731                 int val = *negp ? -*lvalp : *lvalp;
1732                 switch(op) {
1733                 case OP_SET:    *valp = val; break;
1734                 case OP_AND:    *valp &= val; break;
1735                 case OP_OR:     *valp |= val; break;
1736                 case OP_MAX:    if(*valp < val)
1737                                         *valp = val;
1738                                 break;
1739                 case OP_MIN:    if(*valp > val)
1740                                 *valp = val;
1741                                 break;
1742                 }
1743         } else {
1744                 int val = *valp;
1745                 if (val < 0) {
1746                         *negp = -1;
1747                         *lvalp = (unsigned long)-val;
1748                 } else {
1749                         *negp = 0;
1750                         *lvalp = (unsigned long)val;
1751                 }
1752         }
1753         return 0;
1754 }
1755
1756 /*
1757  *      init may raise the set.
1758  */
1759
1760 int proc_dointvec_bset(ctl_table *table, int write, struct file *filp,
1761                         void __user *buffer, size_t *lenp, loff_t *ppos)
1762 {
1763         int op;
1764
1765         if (!capable(CAP_SYS_MODULE)) {
1766                 return -EPERM;
1767         }
1768
1769         op = (current->pid == 1) ? OP_SET : OP_AND;
1770         return do_proc_dointvec(table,write,filp,buffer,lenp,ppos,
1771                                 do_proc_dointvec_bset_conv,&op);
1772 }
1773
1774 struct do_proc_dointvec_minmax_conv_param {
1775         int *min;
1776         int *max;
1777 };
1778
1779 static int do_proc_dointvec_minmax_conv(int *negp, unsigned long *lvalp,
1780                                         int *valp,
1781                                         int write, void *data)
1782 {
1783         struct do_proc_dointvec_minmax_conv_param *param = data;
1784         if (write) {
1785                 int val = *negp ? -*lvalp : *lvalp;
1786                 if ((param->min && *param->min > val) ||
1787                     (param->max && *param->max < val))
1788                         return -EINVAL;
1789                 *valp = val;
1790         } else {
1791                 int val = *valp;
1792                 if (val < 0) {
1793                         *negp = -1;
1794                         *lvalp = (unsigned long)-val;
1795                 } else {
1796                         *negp = 0;
1797                         *lvalp = (unsigned long)val;
1798                 }
1799         }
1800         return 0;
1801 }
1802
1803 /**
1804  * proc_dointvec_minmax - read a vector of integers with min/max values
1805  * @table: the sysctl table
1806  * @write: %TRUE if this is a write to the sysctl file
1807  * @filp: the file structure
1808  * @buffer: the user buffer
1809  * @lenp: the size of the user buffer
1810  * @ppos: file position
1811  *
1812  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
1813  * values from/to the user buffer, treated as an ASCII string.
1814  *
1815  * This routine will ensure the values are within the range specified by
1816  * table->extra1 (min) and table->extra2 (max).
1817  *
1818  * Returns 0 on success.
1819  */
1820 int proc_dointvec_minmax(ctl_table *table, int write, struct file *filp,
1821                   void __user *buffer, size_t *lenp, loff_t *ppos)
1822 {
1823         struct do_proc_dointvec_minmax_conv_param param = {
1824                 .min = (int *) table->extra1,
1825                 .max = (int *) table->extra2,
1826         };
1827         return do_proc_dointvec(table, write, filp, buffer, lenp, ppos,
1828                                 do_proc_dointvec_minmax_conv, &param);
1829 }
1830
1831 static int do_proc_doulongvec_minmax(ctl_table *table, int write,
1832                                      struct file *filp,
1833                                      void __user *buffer,
1834                                      size_t *lenp, loff_t *ppos,
1835                                      unsigned long convmul,
1836                                      unsigned long convdiv)
1837 {
1838 #define TMPBUFLEN 21
1839         unsigned long *i, *min, *max, val;
1840         int vleft, first=1, neg;
1841         size_t len, left;
1842         char buf[TMPBUFLEN], *p;
1843         char __user *s = buffer;
1844
1845         if (!table->data || !table->maxlen || !*lenp ||
1846             (*ppos && !write)) {
1847                 *lenp = 0;
1848                 return 0;
1849         }
1850
1851         i = (unsigned long *) table->data;
1852         min = (unsigned long *) table->extra1;
1853         max = (unsigned long *) table->extra2;
1854         vleft = table->maxlen / sizeof(unsigned long);
1855         left = *lenp;
1856
1857         for (; left && vleft--; i++, min++, max++, first=0) {
1858                 if (write) {
1859                         while (left) {
1860                                 char c;
1861                                 if (get_user(c, s))
1862                                         return -EFAULT;
1863                                 if (!isspace(c))
1864                                         break;
1865                                 left--;
1866                                 s++;
1867                         }
1868                         if (!left)
1869                                 break;
1870                         neg = 0;
1871                         len = left;
1872                         if (len > TMPBUFLEN-1)
1873                                 len = TMPBUFLEN-1;
1874                         if (copy_from_user(buf, s, len))
1875                                 return -EFAULT;
1876                         buf[len] = 0;
1877                         p = buf;
1878                         if (*p == '-' && left > 1) {
1879                                 neg = 1;
1880                                 left--, p++;
1881                         }
1882                         if (*p < '0' || *p > '9')
1883                                 break;
1884                         val = simple_strtoul(p, &p, 0) * convmul / convdiv ;
1885                         len = p-buf;
1886                         if ((len < left) && *p && !isspace(*p))
1887                                 break;
1888                         if (neg)
1889                                 val = -val;
1890                         s += len;
1891                         left -= len;
1892
1893                         if(neg)
1894                                 continue;
1895                         if ((min && val < *min) || (max && val > *max))
1896                                 continue;
1897                         *i = val;
1898                 } else {
1899                         p = buf;
1900                         if (!first)
1901                                 *p++ = '\t';
1902                         sprintf(p, "%lu", convdiv * (*i) / convmul);
1903                         len = strlen(buf);
1904                         if (len > left)
1905                                 len = left;
1906                         if(copy_to_user(s, buf, len))
1907                                 return -EFAULT;
1908                         left -= len;
1909                         s += len;
1910                 }
1911         }
1912
1913         if (!write && !first && left) {
1914                 if(put_user('\n', s))
1915                         return -EFAULT;
1916                 left--, s++;
1917         }
1918         if (write) {
1919                 while (left) {
1920                         char c;
1921                         if (get_user(c, s++))
1922                                 return -EFAULT;
1923                         if (!isspace(c))
1924                                 break;
1925                         left--;
1926                 }
1927         }
1928         if (write && first)
1929                 return -EINVAL;
1930         *lenp -= left;
1931         *ppos += *lenp;
1932         return 0;
1933 #undef TMPBUFLEN
1934 }
1935
1936 /**
1937  * proc_doulongvec_minmax - read a vector of long integers with min/max values
1938  * @table: the sysctl table
1939  * @write: %TRUE if this is a write to the sysctl file
1940  * @filp: the file structure
1941  * @buffer: the user buffer
1942  * @lenp: the size of the user buffer
1943  * @ppos: file position
1944  *
1945  * Reads/writes up to table->maxlen/sizeof(unsigned long) unsigned long
1946  * values from/to the user buffer, treated as an ASCII string.
1947  *
1948  * This routine will ensure the values are within the range specified by
1949  * table->extra1 (min) and table->extra2 (max).
1950  *
1951  * Returns 0 on success.
1952  */
1953 int proc_doulongvec_minmax(ctl_table *table, int write, struct file *filp,
1954                            void __user *buffer, size_t *lenp, loff_t *ppos)
1955 {
1956     return do_proc_doulongvec_minmax(table, write, filp, buffer, lenp, ppos, 1l, 1l);
1957 }
1958
1959 /**
1960  * proc_doulongvec_ms_jiffies_minmax - read a vector of millisecond values with min/max values
1961  * @table: the sysctl table
1962  * @write: %TRUE if this is a write to the sysctl file
1963  * @filp: the file structure
1964  * @buffer: the user buffer
1965  * @lenp: the size of the user buffer
1966  * @ppos: file position
1967  *
1968  * Reads/writes up to table->maxlen/sizeof(unsigned long) unsigned long
1969  * values from/to the user buffer, treated as an ASCII string. The values
1970  * are treated as milliseconds, and converted to jiffies when they are stored.
1971  *
1972  * This routine will ensure the values are within the range specified by
1973  * table->extra1 (min) and table->extra2 (max).
1974  *
1975  * Returns 0 on success.
1976  */
1977 int proc_doulongvec_ms_jiffies_minmax(ctl_table *table, int write,
1978                                       struct file *filp,
1979                                       void __user *buffer,
1980                                       size_t *lenp, loff_t *ppos)
1981 {
1982     return do_proc_doulongvec_minmax(table, write, filp, buffer,
1983                                      lenp, ppos, HZ, 1000l);
1984 }
1985
1986
1987 static int do_proc_dointvec_jiffies_conv(int *negp, unsigned long *lvalp,
1988                                          int *valp,
1989                                          int write, void *data)
1990 {
1991         if (write) {
1992                 *valp = *negp ? -(*lvalp*HZ) : (*lvalp*HZ);
1993         } else {
1994                 int val = *valp;
1995                 unsigned long lval;
1996                 if (val < 0) {
1997                         *negp = -1;
1998                         lval = (unsigned long)-val;
1999                 } else {
2000                         *negp = 0;
2001                         lval = (unsigned long)val;
2002                 }
2003                 *lvalp = lval / HZ;
2004         }
2005         return 0;
2006 }
2007
2008 static int do_proc_dointvec_userhz_jiffies_conv(int *negp, unsigned long *lvalp,
2009                                                 int *valp,
2010                                                 int write, void *data)
2011 {
2012         if (write) {
2013                 *valp = clock_t_to_jiffies(*negp ? -*lvalp : *lvalp);
2014         } else {
2015                 int val = *valp;
2016                 unsigned long lval;
2017                 if (val < 0) {
2018                         *negp = -1;
2019                         lval = (unsigned long)-val;
2020                 } else {
2021                         *negp = 0;
2022                         lval = (unsigned long)val;
2023                 }
2024                 *lvalp = jiffies_to_clock_t(lval);
2025         }
2026         return 0;
2027 }
2028
2029 static int do_proc_dointvec_ms_jiffies_conv(int *negp, unsigned long *lvalp,
2030                                             int *valp,
2031                                             int write, void *data)
2032 {
2033         if (write) {
2034                 *valp = msecs_to_jiffies(*negp ? -*lvalp : *lvalp);
2035         } else {
2036                 int val = *valp;
2037                 unsigned long lval;
2038                 if (val < 0) {
2039                         *negp = -1;
2040                         lval = (unsigned long)-val;
2041                 } else {
2042                         *negp = 0;
2043                         lval = (unsigned long)val;
2044                 }
2045                 *lvalp = jiffies_to_msecs(lval);
2046         }
2047         return 0;
2048 }
2049
2050 /**
2051  * proc_dointvec_jiffies - read a vector of integers as seconds
2052  * @table: the sysctl table
2053  * @write: %TRUE if this is a write to the sysctl file
2054  * @filp: the file structure
2055  * @buffer: the user buffer
2056  * @lenp: the size of the user buffer
2057  * @ppos: file position
2058  *
2059  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2060  * values from/to the user buffer, treated as an ASCII string.
2061  * The values read are assumed to be in seconds, and are converted into
2062  * jiffies.
2063  *
2064  * Returns 0 on success.
2065  */
2066 int proc_dointvec_jiffies(ctl_table *table, int write, struct file *filp,
2067                           void __user *buffer, size_t *lenp, loff_t *ppos)
2068 {
2069     return do_proc_dointvec(table,write,filp,buffer,lenp,ppos,
2070                             do_proc_dointvec_jiffies_conv,NULL);
2071 }
2072
2073 /**
2074  * proc_dointvec_userhz_jiffies - read a vector of integers as 1/USER_HZ seconds
2075  * @table: the sysctl table
2076  * @write: %TRUE if this is a write to the sysctl file
2077  * @filp: the file structure
2078  * @buffer: the user buffer
2079  * @lenp: the size of the user buffer
2080  *
2081  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2082  * values from/to the user buffer, treated as an ASCII string.
2083  * The values read are assumed to be in 1/USER_HZ seconds, and
2084  * are converted into jiffies.
2085  *
2086  * Returns 0 on success.
2087  */
2088 int proc_dointvec_userhz_jiffies(ctl_table *table, int write, struct file *filp,
2089                                  void __user *buffer, size_t *lenp, loff_t *ppos)
2090 {
2091     return do_proc_dointvec(table,write,filp,buffer,lenp,ppos,
2092                             do_proc_dointvec_userhz_jiffies_conv,NULL);
2093 }
2094
2095 /**
2096  * proc_dointvec_ms_jiffies - read a vector of integers as 1 milliseconds
2097  * @table: the sysctl table
2098  * @write: %TRUE if this is a write to the sysctl file
2099  * @filp: the file structure
2100  * @buffer: the user buffer
2101  * @lenp: the size of the user buffer
2102  * @ppos: file position
2103  * @ppos: the current position in the file
2104  *
2105  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2106  * values from/to the user buffer, treated as an ASCII string.
2107  * The values read are assumed to be in 1/1000 seconds, and
2108  * are converted into jiffies.
2109  *
2110  * Returns 0 on success.
2111  */
2112 int proc_dointvec_ms_jiffies(ctl_table *table, int write, struct file *filp,
2113                              void __user *buffer, size_t *lenp, loff_t *ppos)
2114 {
2115         return do_proc_dointvec(table, write, filp, buffer, lenp, ppos,
2116                                 do_proc_dointvec_ms_jiffies_conv, NULL);
2117 }
2118
2119 #else /* CONFIG_PROC_FS */
2120
2121 int proc_dostring(ctl_table *table, int write, struct file *filp,
2122                   void __user *buffer, size_t *lenp, loff_t *ppos)
2123 {
2124         return -ENOSYS;
2125 }
2126
2127 static int proc_doutsstring(ctl_table *table, int write, struct file *filp,
2128                             void __user *buffer, size_t *lenp, loff_t *ppos)
2129 {
2130         return -ENOSYS;
2131 }
2132
2133 int proc_dointvec(ctl_table *table, int write, struct file *filp,
2134                   void __user *buffer, size_t *lenp, loff_t *ppos)
2135 {
2136         return -ENOSYS;
2137 }
2138
2139 int proc_dointvec_bset(ctl_table *table, int write, struct file *filp,
2140                         void __user *buffer, size_t *lenp, loff_t *ppos)
2141 {
2142         return -ENOSYS;
2143 }
2144
2145 int proc_dointvec_minmax(ctl_table *table, int write, struct file *filp,
2146                     void __user *buffer, size_t *lenp, loff_t *ppos)
2147 {
2148         return -ENOSYS;
2149 }
2150
2151 int proc_dointvec_jiffies(ctl_table *table, int write, struct file *filp,
2152                     void __user *buffer, size_t *lenp, loff_t *ppos)
2153 {
2154         return -ENOSYS;
2155 }
2156
2157 int proc_dointvec_userhz_jiffies(ctl_table *table, int write, struct file *filp,
2158                     void __user *buffer, size_t *lenp, loff_t *ppos)
2159 {
2160         return -ENOSYS;
2161 }
2162
2163 int proc_dointvec_ms_jiffies(ctl_table *table, int write, struct file *filp,
2164                              void __user *buffer, size_t *lenp, loff_t *ppos)
2165 {
2166         return -ENOSYS;
2167 }
2168
2169 int proc_doulongvec_minmax(ctl_table *table, int write, struct file *filp,
2170                     void __user *buffer, size_t *lenp, loff_t *ppos)
2171 {
2172         return -ENOSYS;
2173 }
2174
2175 int proc_doulongvec_ms_jiffies_minmax(ctl_table *table, int write,
2176                                       struct file *filp,
2177                                       void __user *buffer,
2178                                       size_t *lenp, loff_t *ppos)
2179 {
2180     return -ENOSYS;
2181 }
2182
2183
2184 #endif /* CONFIG_PROC_FS */
2185
2186
2187 /*
2188  * General sysctl support routines
2189  */
2190
2191 /* The generic string strategy routine: */
2192 int sysctl_string(ctl_table *table, int __user *name, int nlen,
2193                   void __user *oldval, size_t __user *oldlenp,
2194                   void __user *newval, size_t newlen, void **context)
2195 {
2196         size_t l, len;
2197
2198         if (!table->data || !table->maxlen)
2199                 return -ENOTDIR;
2200
2201         if (oldval && oldlenp) {
2202                 if (get_user(len, oldlenp))
2203                         return -EFAULT;
2204                 if (len) {
2205                         l = strlen(table->data);
2206                         if (len > l) len = l;
2207                         if (len >= table->maxlen)
2208                                 len = table->maxlen;
2209                         if(copy_to_user(oldval, table->data, len))
2210                                 return -EFAULT;
2211                         if(put_user(0, ((char __user *) oldval) + len))
2212                                 return -EFAULT;
2213                         if(put_user(len, oldlenp))
2214                                 return -EFAULT;
2215                 }
2216         }
2217         if (newval && newlen) {
2218                 len = newlen;
2219                 if (len > table->maxlen)
2220                         len = table->maxlen;
2221                 if(copy_from_user(table->data, newval, len))
2222                         return -EFAULT;
2223                 if (len == table->maxlen)
2224                         len--;
2225                 ((char *) table->data)[len] = 0;
2226         }
2227         return 0;
2228 }
2229
2230 /*
2231  * This function makes sure that all of the integers in the vector
2232  * are between the minimum and maximum values given in the arrays
2233  * table->extra1 and table->extra2, respectively.
2234  */
2235 int sysctl_intvec(ctl_table *table, int __user *name, int nlen,
2236                 void __user *oldval, size_t __user *oldlenp,
2237                 void __user *newval, size_t newlen, void **context)
2238 {
2239
2240         if (newval && newlen) {
2241                 int __user *vec = (int __user *) newval;
2242                 int *min = (int *) table->extra1;
2243                 int *max = (int *) table->extra2;
2244                 size_t length;
2245                 int i;
2246
2247                 if (newlen % sizeof(int) != 0)
2248                         return -EINVAL;
2249
2250                 if (!table->extra1 && !table->extra2)
2251                         return 0;
2252
2253                 if (newlen > table->maxlen)
2254                         newlen = table->maxlen;
2255                 length = newlen / sizeof(int);
2256
2257                 for (i = 0; i < length; i++) {
2258                         int value;
2259                         if (get_user(value, vec + i))
2260                                 return -EFAULT;
2261                         if (min && value < min[i])
2262                                 return -EINVAL;
2263                         if (max && value > max[i])
2264                                 return -EINVAL;
2265                 }
2266         }
2267         return 0;
2268 }
2269
2270 /* Strategy function to convert jiffies to seconds */
2271 int sysctl_jiffies(ctl_table *table, int __user *name, int nlen,
2272                 void __user *oldval, size_t __user *oldlenp,
2273                 void __user *newval, size_t newlen, void **context)
2274 {
2275         if (oldval) {
2276                 size_t olen;
2277                 if (oldlenp) {
2278                         if (get_user(olen, oldlenp))
2279                                 return -EFAULT;
2280                         if (olen!=sizeof(int))
2281                                 return -EINVAL;
2282                 }
2283                 if (put_user(*(int *)(table->data)/HZ, (int __user *)oldval) ||
2284                     (oldlenp && put_user(sizeof(int),oldlenp)))
2285                         return -EFAULT;
2286         }
2287         if (newval && newlen) {
2288                 int new;
2289                 if (newlen != sizeof(int))
2290                         return -EINVAL;
2291                 if (get_user(new, (int __user *)newval))
2292                         return -EFAULT;
2293                 *(int *)(table->data) = new*HZ;
2294         }
2295         return 1;
2296 }
2297
2298 /* Strategy function to convert jiffies to seconds */
2299 int sysctl_ms_jiffies(ctl_table *table, int __user *name, int nlen,
2300                 void __user *oldval, size_t __user *oldlenp,
2301                 void __user *newval, size_t newlen, void **context)
2302 {
2303         if (oldval) {
2304                 size_t olen;
2305                 if (oldlenp) {
2306                         if (get_user(olen, oldlenp))
2307                                 return -EFAULT;
2308                         if (olen!=sizeof(int))
2309                                 return -EINVAL;
2310                 }
2311                 if (put_user(jiffies_to_msecs(*(int *)(table->data)), (int __user *)oldval) ||
2312                     (oldlenp && put_user(sizeof(int),oldlenp)))
2313                         return -EFAULT;
2314         }
2315         if (newval && newlen) {
2316                 int new;
2317                 if (newlen != sizeof(int))
2318                         return -EINVAL;
2319                 if (get_user(new, (int __user *)newval))
2320                         return -EFAULT;
2321                 *(int *)(table->data) = msecs_to_jiffies(new);
2322         }
2323         return 1;
2324 }
2325
2326 #else /* CONFIG_SYSCTL */
2327
2328
2329 asmlinkage long sys_sysctl(struct __sysctl_args __user *args)
2330 {
2331         return -ENOSYS;
2332 }
2333
2334 int sysctl_string(ctl_table *table, int __user *name, int nlen,
2335                   void __user *oldval, size_t __user *oldlenp,
2336                   void __user *newval, size_t newlen, void **context)
2337 {
2338         return -ENOSYS;
2339 }
2340
2341 int sysctl_intvec(ctl_table *table, int __user *name, int nlen,
2342                 void __user *oldval, size_t __user *oldlenp,
2343                 void __user *newval, size_t newlen, void **context)
2344 {
2345         return -ENOSYS;
2346 }
2347
2348 int sysctl_jiffies(ctl_table *table, int __user *name, int nlen,
2349                 void __user *oldval, size_t __user *oldlenp,
2350                 void __user *newval, size_t newlen, void **context)
2351 {
2352         return -ENOSYS;
2353 }
2354
2355 int sysctl_ms_jiffies(ctl_table *table, int __user *name, int nlen,
2356                 void __user *oldval, size_t __user *oldlenp,
2357                 void __user *newval, size_t newlen, void **context)
2358 {
2359         return -ENOSYS;
2360 }
2361
2362 int proc_dostring(ctl_table *table, int write, struct file *filp,
2363                   void __user *buffer, size_t *lenp, loff_t *ppos)
2364 {
2365         return -ENOSYS;
2366 }
2367
2368 int proc_dointvec(ctl_table *table, int write, struct file *filp,
2369                   void __user *buffer, size_t *lenp, loff_t *ppos)
2370 {
2371         return -ENOSYS;
2372 }
2373
2374 int proc_dointvec_bset(ctl_table *table, int write, struct file *filp,
2375                         void __user *buffer, size_t *lenp, loff_t *ppos)
2376 {
2377         return -ENOSYS;
2378 }
2379
2380 int proc_dointvec_minmax(ctl_table *table, int write, struct file *filp,
2381                     void __user *buffer, size_t *lenp, loff_t *ppos)
2382 {
2383         return -ENOSYS;
2384 }
2385
2386 int proc_dointvec_jiffies(ctl_table *table, int write, struct file *filp,
2387                           void __user *buffer, size_t *lenp, loff_t *ppos)
2388 {
2389         return -ENOSYS;
2390 }
2391
2392 int proc_dointvec_userhz_jiffies(ctl_table *table, int write, struct file *filp,
2393                           void __user *buffer, size_t *lenp, loff_t *ppos)
2394 {
2395         return -ENOSYS;
2396 }
2397
2398 int proc_dointvec_ms_jiffies(ctl_table *table, int write, struct file *filp,
2399                              void __user *buffer, size_t *lenp, loff_t *ppos)
2400 {
2401         return -ENOSYS;
2402 }
2403
2404 int proc_doulongvec_minmax(ctl_table *table, int write, struct file *filp,
2405                     void __user *buffer, size_t *lenp, loff_t *ppos)
2406 {
2407         return -ENOSYS;
2408 }
2409
2410 int proc_doulongvec_ms_jiffies_minmax(ctl_table *table, int write,
2411                                       struct file *filp,
2412                                       void __user *buffer,
2413                                       size_t *lenp, loff_t *ppos)
2414 {
2415     return -ENOSYS;
2416 }
2417
2418 struct ctl_table_header * register_sysctl_table(ctl_table * table,
2419                                                 int insert_at_head)
2420 {
2421         return NULL;
2422 }
2423
2424 void unregister_sysctl_table(struct ctl_table_header * table)
2425 {
2426 }
2427
2428 #endif /* CONFIG_SYSCTL */
2429
2430 /*
2431  * No sense putting this after each symbol definition, twice,
2432  * exception granted :-)
2433  */
2434 EXPORT_SYMBOL(proc_dointvec);
2435 EXPORT_SYMBOL(proc_dointvec_jiffies);
2436 EXPORT_SYMBOL(proc_dointvec_minmax);
2437 EXPORT_SYMBOL(proc_dointvec_userhz_jiffies);
2438 EXPORT_SYMBOL(proc_dointvec_ms_jiffies);
2439 EXPORT_SYMBOL(proc_dostring);
2440 EXPORT_SYMBOL(proc_doulongvec_minmax);
2441 EXPORT_SYMBOL(proc_doulongvec_ms_jiffies_minmax);
2442 EXPORT_SYMBOL(register_sysctl_table);
2443 EXPORT_SYMBOL(sysctl_intvec);
2444 EXPORT_SYMBOL(sysctl_jiffies);
2445 EXPORT_SYMBOL(sysctl_ms_jiffies);
2446 EXPORT_SYMBOL(sysctl_string);
2447 EXPORT_SYMBOL(unregister_sysctl_table);