kernel/sysctl.c

   1 /*
   2  * sysctl.c: General linux system control interface
   3  *
   4  * Begun 24 March 1995, Stephen Tweedie
   5  * Added /proc support, Dec 1995
   6  * Added bdflush entry and intvec min/max checking, 2/23/96, Tom Dyas.
   7  * Added hooks for /proc/sys/net (minor, minor patch), 96/4/1, Mike Shaver.
   8  * Added kernel/java-{interpreter,appletviewer}, 96/5/10, Mike Shaver.
   9  * Dynamic registration fixes, Stephen Tweedie.
  10  * Added kswapd-interval, ctrl-alt-del, printk stuff, 1/8/97, Chris Horn.
  11  * Made sysctl support optional via CONFIG_SYSCTL, 1/10/97, Chris
  12  *  Horn.
  13  * Added proc_doulongvec_ms_jiffies_minmax, 09/08/99, Carlos H. Bauer.
  14  * Added proc_doulongvec_minmax, 09/08/99, Carlos H. Bauer.
  15  * Changed linked lists to use list.h instead of lists.h, 02/24/00, Bill
  16  *  Wendling.
  17  * The list_for_each() macro wasn't appropriate for the sysctl loop.
  18  *  Removed it and replaced it with older style, 03/23/00, Bill Wendling
  19  */
  20
  21 #include <linux/module.h>
  22 #include <linux/mm.h>
  23 #include <linux/swap.h>
  24 #include <linux/slab.h>
  25 #include <linux/sysctl.h>
  26 #include <linux/proc_fs.h>
  27 #include <linux/capability.h>
  28 #include <linux/ctype.h>
  29 #include <linux/utsname.h>
  30 #include <linux/capability.h>
  31 #include <linux/smp_lock.h>
  32 #include <linux/fs.h>
  33 #include <linux/init.h>
  34 #include <linux/kernel.h>
  35 #include <linux/kobject.h>
  36 #include <linux/net.h>
  37 #include <linux/sysrq.h>
  38 #include <linux/highuid.h>
  39 #include <linux/writeback.h>
  40 #include <linux/hugetlb.h>
  41 #include <linux/security.h>
  42 #include <linux/initrd.h>
  43 #include <linux/times.h>
  44 #include <linux/limits.h>
  45 #include <linux/dcache.h>
  46 #include <linux/syscalls.h>
  47 #include <linux/nfs_fs.h>
  48 #include <linux/acpi.h>
  49 #include <linux/reboot.h>
  50
  51 #include <asm/uaccess.h>
  52 #include <asm/processor.h>
  53
  54 #ifdef CONFIG_X86
  55 #include <asm/nmi.h>
  56 #include <asm/stacktrace.h>
  57 #endif
  58
  59 #if defined(CONFIG_SYSCTL)
  60
  61 /* External variables not in a header file. */
  62 extern int C_A_D;
  63 extern int print_fatal_signals;
  64 extern int sysctl_overcommit_memory;
  65 extern int sysctl_overcommit_ratio;
  66 extern int sysctl_panic_on_oom;
  67 extern int max_threads;
  68 extern int core_uses_pid;
  69 extern int suid_dumpable;
  70 extern char core_pattern[];
  71 extern int pid_max;
  72 extern int min_free_kbytes;
  73 extern int printk_ratelimit_jiffies;
  74 extern int printk_ratelimit_burst;
  75 extern int pid_max_min, pid_max_max;
  76 extern int sysctl_drop_caches;
  77 extern int percpu_pagelist_fraction;
  78 extern int compat_log;
  79 extern int maps_protect;
  80 extern int sysctl_stat_interval;
  81 extern int audit_argv_kb;
  82
  83 /* this is needed for the proc_dointvec_minmax for [fs_]overflow UID and GID */
  84 static int maxolduid = 65535;
  85 static int minolduid;
  86 static int min_percpu_pagelist_fract = 8;
  87
  88 static int ngroups_max = NGROUPS_MAX;
  89
  90 #ifdef CONFIG_KMOD
  91 extern char modprobe_path[];
  92 #endif
  93 #ifdef CONFIG_CHR_DEV_SG
  94 extern int sg_big_buff;
  95 #endif
  96
  97 #ifdef __sparc__
  98 extern char reboot_command [];
  99 extern int stop_a_enabled;
 100 extern int scons_pwroff;
 101 #endif
 102
 103 #ifdef __hppa__
 104 extern int pwrsw_enabled;
 105 extern int unaligned_enabled;
 106 #endif
 107
 108 #ifdef CONFIG_S390
 109 #ifdef CONFIG_MATHEMU
 110 extern int sysctl_ieee_emulation_warnings;
 111 #endif
 112 extern int sysctl_userprocess_debug;
 113 extern int spin_retry;
 114 #endif
 115
 116 extern int sysctl_hz_timer;
 117
 118 #ifdef CONFIG_BSD_PROCESS_ACCT
 119 extern int acct_parm[];
 120 #endif
 121
 122 #ifdef CONFIG_IA64
 123 extern int no_unaligned_warning;
 124 #endif
 125
 126 #ifdef CONFIG_RT_MUTEXES
 127 extern int max_lock_depth;
 128 #endif
 129
 130 #ifdef CONFIG_SYSCTL_SYSCALL
 131 static int parse_table(int __user *, int, void __user *, size_t __user *,
 132                 void __user *, size_t, ctl_table *);
 133 #endif
 134
 135
 136 #ifdef CONFIG_PROC_SYSCTL
 137 static int proc_do_cad_pid(ctl_table *table, int write, struct file *filp,
 138                   void __user *buffer, size_t *lenp, loff_t *ppos);
 139 static int proc_dointvec_taint(ctl_table *table, int write, struct file *filp,
 140                                void __user *buffer, size_t *lenp, loff_t *ppos);
 141 #endif
 142
 143 static ctl_table root_table[];
 144 static struct ctl_table_header root_table_header =
 145         { root_table, LIST_HEAD_INIT(root_table_header.ctl_entry) };
 146
 147 static ctl_table kern_table[];
 148 static ctl_table vm_table[];
 149 static ctl_table fs_table[];
 150 static ctl_table debug_table[];
 151 static ctl_table dev_table[];
 152 extern ctl_table random_table[];
 153 #ifdef CONFIG_UNIX98_PTYS
 154 extern ctl_table pty_table[];
 155 #endif
 156 #ifdef CONFIG_INOTIFY_USER
 157 extern ctl_table inotify_table[];
 158 #endif
 159
 160 #ifdef HAVE_ARCH_PICK_MMAP_LAYOUT
 161 int sysctl_legacy_va_layout;
 162 #endif
 163
 164 extern int prove_locking;
 165 extern int lock_stat;
 166
 167 /* The default sysctl tables: */
 168
 169 static ctl_table root_table[] = {
 170         {
 171                 .ctl_name       = CTL_KERN,
 172                 .procname       = "kernel",
 173                 .mode           = 0555,
 174                 .child          = kern_table,
 175         },
 176         {
 177                 .ctl_name       = CTL_VM,
 178                 .procname       = "vm",
 179                 .mode           = 0555,
 180                 .child          = vm_table,
 181         },
 182 #ifdef CONFIG_NET
 183         {
 184                 .ctl_name       = CTL_NET,
 185                 .procname       = "net",
 186                 .mode           = 0555,
 187                 .child          = net_table,
 188         },
 189 #endif
 190         {
 191                 .ctl_name       = CTL_FS,
 192                 .procname       = "fs",
 193                 .mode           = 0555,
 194                 .child          = fs_table,
 195         },
 196         {
 197                 .ctl_name       = CTL_DEBUG,
 198                 .procname       = "debug",
 199                 .mode           = 0555,
 200                 .child          = debug_table,
 201         },
 202         {
 203                 .ctl_name       = CTL_DEV,
 204                 .procname       = "dev",
 205                 .mode           = 0555,
 206                 .child          = dev_table,
 207         },
 208 /*
 209  * NOTE: do not add new entries to this table unless you have read
 210  * Documentation/sysctl/ctl_unnumbered.txt
 211  */
 212         { .ctl_name = 0 }
 213 };
 214
 215 #ifdef CONFIG_SCHED_DEBUG
 216 static unsigned long min_sched_granularity_ns = 100000;         /* 100 usecs */
 217 static unsigned long max_sched_granularity_ns = 1000000000;     /* 1 second */
 218 static unsigned long min_wakeup_granularity_ns;                 /* 0 usecs */
 219 static unsigned long max_wakeup_granularity_ns = 1000000000;    /* 1 second */
 220 #endif
 221
 222 static ctl_table kern_table[] = {
 223 #ifdef CONFIG_SCHED_DEBUG
 224         {
 225                 .ctl_name       = CTL_UNNUMBERED,
 226                 .procname       = "sched_granularity_ns",
 227                 .data           = &sysctl_sched_granularity,
 228                 .maxlen         = sizeof(unsigned int),
 229                 .mode           = 0644,
 230                 .proc_handler   = &proc_dointvec_minmax,
 231                 .strategy       = &sysctl_intvec,
 232                 .extra1         = &min_sched_granularity_ns,
 233                 .extra2         = &max_sched_granularity_ns,
 234         },
 235         {
 236                 .ctl_name       = CTL_UNNUMBERED,
 237                 .procname       = "sched_wakeup_granularity_ns",
 238                 .data           = &sysctl_sched_wakeup_granularity,
 239                 .maxlen         = sizeof(unsigned int),
 240                 .mode           = 0644,
 241                 .proc_handler   = &proc_dointvec_minmax,
 242                 .strategy       = &sysctl_intvec,
 243                 .extra1         = &min_wakeup_granularity_ns,
 244                 .extra2         = &max_wakeup_granularity_ns,
 245         },
 246         {
 247                 .ctl_name       = CTL_UNNUMBERED,
 248                 .procname       = "sched_batch_wakeup_granularity_ns",
 249                 .data           = &sysctl_sched_batch_wakeup_granularity,
 250                 .maxlen         = sizeof(unsigned int),
 251                 .mode           = 0644,
 252                 .proc_handler   = &proc_dointvec_minmax,
 253                 .strategy       = &sysctl_intvec,
 254                 .extra1         = &min_wakeup_granularity_ns,
 255                 .extra2         = &max_wakeup_granularity_ns,
 256         },
 257         {
 258                 .ctl_name       = CTL_UNNUMBERED,
 259                 .procname       = "sched_stat_granularity_ns",
 260                 .data           = &sysctl_sched_stat_granularity,
 261                 .maxlen         = sizeof(unsigned int),
 262                 .mode           = 0644,
 263                 .proc_handler   = &proc_dointvec_minmax,
 264                 .strategy       = &sysctl_intvec,
 265                 .extra1         = &min_wakeup_granularity_ns,
 266                 .extra2         = &max_wakeup_granularity_ns,
 267         },
 268         {
 269                 .ctl_name       = CTL_UNNUMBERED,
 270                 .procname       = "sched_runtime_limit_ns",
 271                 .data           = &sysctl_sched_runtime_limit,
 272                 .maxlen         = sizeof(unsigned int),
 273                 .mode           = 0644,
 274                 .proc_handler   = &proc_dointvec_minmax,
 275                 .strategy       = &sysctl_intvec,
 276                 .extra1         = &min_sched_granularity_ns,
 277                 .extra2         = &max_sched_granularity_ns,
 278         },
 279         {
 280                 .ctl_name       = CTL_UNNUMBERED,
 281                 .procname       = "sched_child_runs_first",
 282                 .data           = &sysctl_sched_child_runs_first,
 283                 .maxlen         = sizeof(unsigned int),
 284                 .mode           = 0644,
 285                 .proc_handler   = &proc_dointvec,
 286         },
 287 #ifdef CONFIG_PROVE_LOCKING
 288         {
 289                 .ctl_name       = CTL_UNNUMBERED,
 290                 .procname       = "prove_locking",
 291                 .data           = &prove_locking,
 292                 .maxlen         = sizeof(int),
 293                 .mode           = 0644,
 294                 .proc_handler   = &proc_dointvec,
 295         },
 296 #endif
 297 #ifdef CONFIG_LOCK_STAT
 298         {
 299                 .ctl_name       = CTL_UNNUMBERED,
 300                 .procname       = "lock_stat",
 301                 .data           = &lock_stat,
 302                 .maxlen         = sizeof(int),
 303                 .mode           = 0644,
 304                 .proc_handler   = &proc_dointvec,
 305         },
 306 #endif
 307         {
 308                 .ctl_name       = CTL_UNNUMBERED,
 309                 .procname       = "sched_features",
 310                 .data           = &sysctl_sched_features,
 311                 .maxlen         = sizeof(unsigned int),
 312                 .mode           = 0644,
 313                 .proc_handler   = &proc_dointvec,
 314         },
 315 #endif
 316         {
 317                 .ctl_name       = KERN_PANIC,
 318                 .procname       = "panic",
 319                 .data           = &panic_timeout,
 320                 .maxlen         = sizeof(int),
 321                 .mode           = 0644,
 322                 .proc_handler   = &proc_dointvec,
 323         },
 324         {
 325                 .ctl_name       = KERN_CORE_USES_PID,
 326                 .procname       = "core_uses_pid",
 327                 .data           = &core_uses_pid,
 328                 .maxlen         = sizeof(int),
 329                 .mode           = 0644,
 330                 .proc_handler   = &proc_dointvec,
 331         },
 332 #ifdef CONFIG_AUDITSYSCALL
 333         {
 334                 .ctl_name       = CTL_UNNUMBERED,
 335                 .procname       = "audit_argv_kb",
 336                 .data           = &audit_argv_kb,
 337                 .maxlen         = sizeof(int),
 338                 .mode           = 0644,
 339                 .proc_handler   = &proc_dointvec,
 340         },
 341 #endif
 342         {
 343                 .ctl_name       = KERN_CORE_PATTERN,
 344                 .procname       = "core_pattern",
 345                 .data           = core_pattern,
 346                 .maxlen         = CORENAME_MAX_SIZE,
 347                 .mode           = 0644,
 348                 .proc_handler   = &proc_dostring,
 349                 .strategy       = &sysctl_string,
 350         },
 351 #ifdef CONFIG_PROC_SYSCTL
 352         {
 353                 .ctl_name       = KERN_TAINTED,
 354                 .procname       = "tainted",
 355                 .data           = &tainted,
 356                 .maxlen         = sizeof(int),
 357                 .mode           = 0644,
 358                 .proc_handler   = &proc_dointvec_taint,
 359         },
 360 #endif
 361         {
 362                 .ctl_name       = KERN_CAP_BSET,
 363                 .procname       = "cap-bound",
 364                 .data           = &cap_bset,
 365                 .maxlen         = sizeof(kernel_cap_t),
 366                 .mode           = 0600,
 367                 .proc_handler   = &proc_dointvec_bset,
 368         },
 369 #ifdef CONFIG_BLK_DEV_INITRD
 370         {
 371                 .ctl_name       = KERN_REALROOTDEV,
 372                 .procname       = "real-root-dev",
 373                 .data           = &real_root_dev,
 374                 .maxlen         = sizeof(int),
 375                 .mode           = 0644,
 376                 .proc_handler   = &proc_dointvec,
 377         },
 378 #endif
 379         {
 380                 .ctl_name       = CTL_UNNUMBERED,
 381                 .procname       = "print-fatal-signals",
 382                 .data           = &print_fatal_signals,
 383                 .maxlen         = sizeof(int),
 384                 .mode           = 0644,
 385                 .proc_handler   = &proc_dointvec,
 386         },
 387 #ifdef __sparc__
 388         {
 389                 .ctl_name       = KERN_SPARC_REBOOT,
 390                 .procname       = "reboot-cmd",
 391                 .data           = reboot_command,
 392                 .maxlen         = 256,
 393                 .mode           = 0644,
 394                 .proc_handler   = &proc_dostring,
 395                 .strategy       = &sysctl_string,
 396         },
 397         {
 398                 .ctl_name       = KERN_SPARC_STOP_A,
 399                 .procname       = "stop-a",
 400                 .data           = &stop_a_enabled,
 401                 .maxlen         = sizeof (int),
 402                 .mode           = 0644,
 403                 .proc_handler   = &proc_dointvec,
 404         },
 405         {
 406                 .ctl_name       = KERN_SPARC_SCONS_PWROFF,
 407                 .procname       = "scons-poweroff",
 408                 .data           = &scons_pwroff,
 409                 .maxlen         = sizeof (int),
 410                 .mode           = 0644,
 411                 .proc_handler   = &proc_dointvec,
 412         },
 413 #endif
 414 #ifdef __hppa__
 415         {
 416                 .ctl_name       = KERN_HPPA_PWRSW,
 417                 .procname       = "soft-power",
 418                 .data           = &pwrsw_enabled,
 419                 .maxlen         = sizeof (int),
 420                 .mode           = 0644,
 421                 .proc_handler   = &proc_dointvec,
 422         },
 423         {
 424                 .ctl_name       = KERN_HPPA_UNALIGNED,
 425                 .procname       = "unaligned-trap",
 426                 .data           = &unaligned_enabled,
 427                 .maxlen         = sizeof (int),
 428                 .mode           = 0644,
 429                 .proc_handler   = &proc_dointvec,
 430         },
 431 #endif
 432         {
 433                 .ctl_name       = KERN_CTLALTDEL,
 434                 .procname       = "ctrl-alt-del",
 435                 .data           = &C_A_D,
 436                 .maxlen         = sizeof(int),
 437                 .mode           = 0644,
 438                 .proc_handler   = &proc_dointvec,
 439         },
 440         {
 441                 .ctl_name       = KERN_PRINTK,
 442                 .procname       = "printk",
 443                 .data           = &console_loglevel,
 444                 .maxlen         = 4*sizeof(int),
 445                 .mode           = 0644,
 446                 .proc_handler   = &proc_dointvec,
 447         },
 448 #ifdef CONFIG_KMOD
 449         {
 450                 .ctl_name       = KERN_MODPROBE,
 451                 .procname       = "modprobe",
 452                 .data           = &modprobe_path,
 453                 .maxlen         = KMOD_PATH_LEN,
 454                 .mode           = 0644,
 455                 .proc_handler   = &proc_dostring,
 456                 .strategy       = &sysctl_string,
 457         },
 458 #endif
 459 #if defined(CONFIG_HOTPLUG) && defined(CONFIG_NET)
 460         {
 461                 .ctl_name       = KERN_HOTPLUG,
 462                 .procname       = "hotplug",
 463                 .data           = &uevent_helper,
 464                 .maxlen         = UEVENT_HELPER_PATH_LEN,
 465                 .mode           = 0644,
 466                 .proc_handler   = &proc_dostring,
 467                 .strategy       = &sysctl_string,
 468         },
 469 #endif
 470 #ifdef CONFIG_CHR_DEV_SG
 471         {
 472                 .ctl_name       = KERN_SG_BIG_BUFF,
 473                 .procname       = "sg-big-buff",
 474                 .data           = &sg_big_buff,
 475                 .maxlen         = sizeof (int),
 476                 .mode           = 0444,
 477                 .proc_handler   = &proc_dointvec,
 478         },
 479 #endif
 480 #ifdef CONFIG_BSD_PROCESS_ACCT
 481         {
 482                 .ctl_name       = KERN_ACCT,
 483                 .procname       = "acct",
 484                 .data           = &acct_parm,
 485                 .maxlen         = 3*sizeof(int),
 486                 .mode           = 0644,
 487                 .proc_handler   = &proc_dointvec,
 488         },
 489 #endif
 490 #ifdef CONFIG_MAGIC_SYSRQ
 491         {
 492                 .ctl_name       = KERN_SYSRQ,
 493                 .procname       = "sysrq",
 494                 .data           = &__sysrq_enabled,
 495                 .maxlen         = sizeof (int),
 496                 .mode           = 0644,
 497                 .proc_handler   = &proc_dointvec,
 498         },
 499 #endif
 500 #ifdef CONFIG_PROC_SYSCTL
 501         {
 502                 .ctl_name       = KERN_CADPID,
 503                 .procname       = "cad_pid",
 504                 .data           = NULL,
 505                 .maxlen         = sizeof (int),
 506                 .mode           = 0600,
 507                 .proc_handler   = &proc_do_cad_pid,
 508         },
 509 #endif
 510         {
 511                 .ctl_name       = KERN_MAX_THREADS,
 512                 .procname       = "threads-max",
 513                 .data           = &max_threads,
 514                 .maxlen         = sizeof(int),
 515                 .mode           = 0644,
 516                 .proc_handler   = &proc_dointvec,
 517         },
 518         {
 519                 .ctl_name       = KERN_RANDOM,
 520                 .procname       = "random",
 521                 .mode           = 0555,
 522                 .child          = random_table,
 523         },
 524 #ifdef CONFIG_UNIX98_PTYS
 525         {
 526                 .ctl_name       = KERN_PTY,
 527                 .procname       = "pty",
 528                 .mode           = 0555,
 529                 .child          = pty_table,
 530         },
 531 #endif
 532         {
 533                 .ctl_name       = KERN_OVERFLOWUID,
 534                 .procname       = "overflowuid",
 535                 .data           = &overflowuid,
 536                 .maxlen         = sizeof(int),
 537                 .mode           = 0644,
 538                 .proc_handler   = &proc_dointvec_minmax,
 539                 .strategy       = &sysctl_intvec,
 540                 .extra1         = &minolduid,
 541                 .extra2         = &maxolduid,
 542         },
 543         {
 544                 .ctl_name       = KERN_OVERFLOWGID,
 545                 .procname       = "overflowgid",
 546                 .data           = &overflowgid,
 547                 .maxlen         = sizeof(int),
 548                 .mode           = 0644,
 549                 .proc_handler   = &proc_dointvec_minmax,
 550                 .strategy       = &sysctl_intvec,
 551                 .extra1         = &minolduid,
 552                 .extra2         = &maxolduid,
 553         },
 554 #ifdef CONFIG_S390
 555 #ifdef CONFIG_MATHEMU
 556         {
 557                 .ctl_name       = KERN_IEEE_EMULATION_WARNINGS,
 558                 .procname       = "ieee_emulation_warnings",
 559                 .data           = &sysctl_ieee_emulation_warnings,
 560                 .maxlen         = sizeof(int),
 561                 .mode           = 0644,
 562                 .proc_handler   = &proc_dointvec,
 563         },
 564 #endif
 565 #ifdef CONFIG_NO_IDLE_HZ
 566         {
 567                 .ctl_name       = KERN_HZ_TIMER,
 568                 .procname       = "hz_timer",
 569                 .data           = &sysctl_hz_timer,
 570                 .maxlen         = sizeof(int),
 571                 .mode           = 0644,
 572                 .proc_handler   = &proc_dointvec,
 573         },
 574 #endif
 575         {
 576                 .ctl_name       = KERN_S390_USER_DEBUG_LOGGING,
 577                 .procname       = "userprocess_debug",
 578                 .data           = &sysctl_userprocess_debug,
 579                 .maxlen         = sizeof(int),
 580                 .mode           = 0644,
 581                 .proc_handler   = &proc_dointvec,
 582         },
 583 #endif
 584         {
 585                 .ctl_name       = KERN_PIDMAX,
 586                 .procname       = "pid_max",
 587                 .data           = &pid_max,
 588                 .maxlen         = sizeof (int),
 589                 .mode           = 0644,
 590                 .proc_handler   = &proc_dointvec_minmax,
 591                 .strategy       = sysctl_intvec,
 592                 .extra1         = &pid_max_min,
 593                 .extra2         = &pid_max_max,
 594         },
 595         {
 596                 .ctl_name       = KERN_PANIC_ON_OOPS,
 597                 .procname       = "panic_on_oops",
 598                 .data           = &panic_on_oops,
 599                 .maxlen         = sizeof(int),
 600                 .mode           = 0644,
 601                 .proc_handler   = &proc_dointvec,
 602         },
 603         {
 604                 .ctl_name       = KERN_PRINTK_RATELIMIT,
 605                 .procname       = "printk_ratelimit",
 606                 .data           = &printk_ratelimit_jiffies,
 607                 .maxlen         = sizeof(int),
 608                 .mode           = 0644,
 609                 .proc_handler   = &proc_dointvec_jiffies,
 610                 .strategy       = &sysctl_jiffies,
 611         },
 612         {
 613                 .ctl_name       = KERN_PRINTK_RATELIMIT_BURST,
 614                 .procname       = "printk_ratelimit_burst",
 615                 .data           = &printk_ratelimit_burst,
 616                 .maxlen         = sizeof(int),
 617                 .mode           = 0644,
 618                 .proc_handler   = &proc_dointvec,
 619         },
 620         {
 621                 .ctl_name       = KERN_NGROUPS_MAX,
 622                 .procname       = "ngroups_max",
 623                 .data           = &ngroups_max,
 624                 .maxlen         = sizeof (int),
 625                 .mode           = 0444,
 626                 .proc_handler   = &proc_dointvec,
 627         },
 628 #if defined(CONFIG_X86_LOCAL_APIC) && defined(CONFIG_X86)
 629         {
 630                 .ctl_name       = KERN_UNKNOWN_NMI_PANIC,
 631                 .procname       = "unknown_nmi_panic",
 632                 .data           = &unknown_nmi_panic,
 633                 .maxlen         = sizeof (int),
 634                 .mode           = 0644,
 635                 .proc_handler   = &proc_dointvec,
 636         },
 637         {
 638                 .ctl_name       = KERN_NMI_WATCHDOG,
 639                 .procname       = "nmi_watchdog",
 640                 .data           = &nmi_watchdog_enabled,
 641                 .maxlen         = sizeof (int),
 642                 .mode           = 0644,
 643                 .proc_handler   = &proc_nmi_enabled,
 644         },
 645 #endif
 646 #if defined(CONFIG_X86)
 647         {
 648                 .ctl_name       = KERN_PANIC_ON_NMI,
 649                 .procname       = "panic_on_unrecovered_nmi",
 650                 .data           = &panic_on_unrecovered_nmi,
 651                 .maxlen         = sizeof(int),
 652                 .mode           = 0644,
 653                 .proc_handler   = &proc_dointvec,
 654         },
 655         {
 656                 .ctl_name       = KERN_BOOTLOADER_TYPE,
 657                 .procname       = "bootloader_type",
 658                 .data           = &bootloader_type,
 659                 .maxlen         = sizeof (int),
 660                 .mode           = 0444,
 661                 .proc_handler   = &proc_dointvec,
 662         },
 663         {
 664                 .ctl_name       = CTL_UNNUMBERED,
 665                 .procname       = "kstack_depth_to_print",
 666                 .data           = &kstack_depth_to_print,
 667                 .maxlen         = sizeof(int),
 668                 .mode           = 0644,
 669                 .proc_handler   = &proc_dointvec,
 670         },
 671 #endif
 672 #if defined(CONFIG_MMU)
 673         {
 674                 .ctl_name       = KERN_RANDOMIZE,
 675                 .procname       = "randomize_va_space",
 676                 .data           = &randomize_va_space,
 677                 .maxlen         = sizeof(int),
 678                 .mode           = 0644,
 679                 .proc_handler   = &proc_dointvec,
 680         },
 681 #endif
 682 #if defined(CONFIG_S390) && defined(CONFIG_SMP)
 683         {
 684                 .ctl_name       = KERN_SPIN_RETRY,
 685                 .procname       = "spin_retry",
 686                 .data           = &spin_retry,
 687                 .maxlen         = sizeof (int),
 688                 .mode           = 0644,
 689                 .proc_handler   = &proc_dointvec,
 690         },
 691 #endif
 692 #ifdef CONFIG_ACPI_SLEEP
 693         {
 694                 .ctl_name       = KERN_ACPI_VIDEO_FLAGS,
 695                 .procname       = "acpi_video_flags",
 696                 .data           = &acpi_realmode_flags,
 697                 .maxlen         = sizeof (unsigned long),
 698                 .mode           = 0644,
 699                 .proc_handler   = &proc_doulongvec_minmax,
 700         },
 701 #endif
 702 #ifdef CONFIG_IA64
 703         {
 704                 .ctl_name       = KERN_IA64_UNALIGNED,
 705                 .procname       = "ignore-unaligned-usertrap",
 706                 .data           = &no_unaligned_warning,
 707                 .maxlen         = sizeof (int),
 708                 .mode           = 0644,
 709                 .proc_handler   = &proc_dointvec,
 710         },
 711 #endif
 712 #ifdef CONFIG_COMPAT
 713         {
 714                 .ctl_name       = KERN_COMPAT_LOG,
 715                 .procname       = "compat-log",
 716                 .data           = &compat_log,
 717                 .maxlen         = sizeof (int),
 718                 .mode           = 0644,
 719                 .proc_handler   = &proc_dointvec,
 720         },
 721 #endif
 722 #ifdef CONFIG_RT_MUTEXES
 723         {
 724                 .ctl_name       = KERN_MAX_LOCK_DEPTH,
 725                 .procname       = "max_lock_depth",
 726                 .data           = &max_lock_depth,
 727                 .maxlen         = sizeof(int),
 728                 .mode           = 0644,
 729                 .proc_handler   = &proc_dointvec,
 730         },
 731 #endif
 732 #ifdef CONFIG_PROC_FS
 733         {
 734                 .ctl_name       = CTL_UNNUMBERED,
 735                 .procname       = "maps_protect",
 736                 .data           = &maps_protect,
 737                 .maxlen         = sizeof(int),
 738                 .mode           = 0644,
 739                 .proc_handler   = &proc_dointvec,
 740         },
 741 #endif
 742         {
 743                 .ctl_name       = CTL_UNNUMBERED,
 744                 .procname       = "poweroff_cmd",
 745                 .data           = &poweroff_cmd,
 746                 .maxlen         = POWEROFF_CMD_PATH_LEN,
 747                 .mode           = 0644,
 748                 .proc_handler   = &proc_dostring,
 749                 .strategy       = &sysctl_string,
 750         },
 751 /*
 752  * NOTE: do not add new entries to this table unless you have read
 753  * Documentation/sysctl/ctl_unnumbered.txt
 754  */
 755         { .ctl_name = 0 }
 756 };
 757
 758 /* Constants for minimum and maximum testing in vm_table.
 759    We use these as one-element integer vectors. */
 760 static int zero;
 761 static int two = 2;
 762 static int one_hundred = 100;
 763
 764
 765 static ctl_table vm_table[] = {
 766         {
 767                 .ctl_name       = VM_OVERCOMMIT_MEMORY,
 768                 .procname       = "overcommit_memory",
 769                 .data           = &sysctl_overcommit_memory,
 770                 .maxlen         = sizeof(sysctl_overcommit_memory),
 771                 .mode           = 0644,
 772                 .proc_handler   = &proc_dointvec,
 773         },
 774         {
 775                 .ctl_name       = VM_PANIC_ON_OOM,
 776                 .procname       = "panic_on_oom",
 777                 .data           = &sysctl_panic_on_oom,
 778                 .maxlen         = sizeof(sysctl_panic_on_oom),
 779                 .mode           = 0644,
 780                 .proc_handler   = &proc_dointvec,
 781         },
 782         {
 783                 .ctl_name       = VM_OVERCOMMIT_RATIO,
 784                 .procname       = "overcommit_ratio",
 785                 .data           = &sysctl_overcommit_ratio,
 786                 .maxlen         = sizeof(sysctl_overcommit_ratio),
 787                 .mode           = 0644,
 788                 .proc_handler   = &proc_dointvec,
 789         },
 790         {
 791                 .ctl_name       = VM_PAGE_CLUSTER,
 792                 .procname       = "page-cluster",
 793                 .data           = &page_cluster,
 794                 .maxlen         = sizeof(int),
 795                 .mode           = 0644,
 796                 .proc_handler   = &proc_dointvec,
 797         },
 798         {
 799                 .ctl_name       = VM_DIRTY_BACKGROUND,
 800                 .procname       = "dirty_background_ratio",
 801                 .data           = &dirty_background_ratio,
 802                 .maxlen         = sizeof(dirty_background_ratio),
 803                 .mode           = 0644,
 804                 .proc_handler   = &proc_dointvec_minmax,
 805                 .strategy       = &sysctl_intvec,
 806                 .extra1         = &zero,
 807                 .extra2         = &one_hundred,
 808         },
 809         {
 810                 .ctl_name       = VM_DIRTY_RATIO,
 811                 .procname       = "dirty_ratio",
 812                 .data           = &vm_dirty_ratio,
 813                 .maxlen         = sizeof(vm_dirty_ratio),
 814                 .mode           = 0644,
 815                 .proc_handler   = &proc_dointvec_minmax,
 816                 .strategy       = &sysctl_intvec,
 817                 .extra1         = &zero,
 818                 .extra2         = &one_hundred,
 819         },
 820         {
 821                 .ctl_name       = VM_DIRTY_WB_CS,
 822                 .procname       = "dirty_writeback_centisecs",
 823                 .data           = &dirty_writeback_interval,
 824                 .maxlen         = sizeof(dirty_writeback_interval),
 825                 .mode           = 0644,
 826                 .proc_handler   = &dirty_writeback_centisecs_handler,
 827         },
 828         {
 829                 .ctl_name       = VM_DIRTY_EXPIRE_CS,
 830                 .procname       = "dirty_expire_centisecs",
 831                 .data           = &dirty_expire_interval,
 832                 .maxlen         = sizeof(dirty_expire_interval),
 833                 .mode           = 0644,
 834                 .proc_handler   = &proc_dointvec_userhz_jiffies,
 835         },
 836         {
 837                 .ctl_name       = VM_NR_PDFLUSH_THREADS,
 838                 .procname       = "nr_pdflush_threads",
 839                 .data           = &nr_pdflush_threads,
 840                 .maxlen         = sizeof nr_pdflush_threads,
 841                 .mode           = 0444 /* read-only*/,
 842                 .proc_handler   = &proc_dointvec,
 843         },
 844         {
 845                 .ctl_name       = VM_SWAPPINESS,
 846                 .procname       = "swappiness",
 847                 .data           = &vm_swappiness,
 848                 .maxlen         = sizeof(vm_swappiness),
 849                 .mode           = 0644,
 850                 .proc_handler   = &proc_dointvec_minmax,
 851                 .strategy       = &sysctl_intvec,
 852                 .extra1         = &zero,
 853                 .extra2         = &one_hundred,
 854         },
 855 #ifdef CONFIG_HUGETLB_PAGE
 856          {
 857                 .ctl_name       = VM_HUGETLB_PAGES,
 858                 .procname       = "nr_hugepages",
 859                 .data           = &max_huge_pages,
 860                 .maxlen         = sizeof(unsigned long),
 861                 .mode           = 0644,
 862                 .proc_handler   = &hugetlb_sysctl_handler,
 863                 .extra1         = (void *)&hugetlb_zero,
 864                 .extra2         = (void *)&hugetlb_infinity,
 865          },
 866          {
 867                 .ctl_name       = VM_HUGETLB_GROUP,
 868                 .procname       = "hugetlb_shm_group",
 869                 .data           = &sysctl_hugetlb_shm_group,
 870                 .maxlen         = sizeof(gid_t),
 871                 .mode           = 0644,
 872                 .proc_handler   = &proc_dointvec,
 873          },
 874          {
 875                 .ctl_name       = CTL_UNNUMBERED,
 876                 .procname       = "hugepages_treat_as_movable",
 877                 .data           = &hugepages_treat_as_movable,
 878                 .maxlen         = sizeof(int),
 879                 .mode           = 0644,
 880                 .proc_handler   = &hugetlb_treat_movable_handler,
 881         },
 882 #endif
 883         {
 884                 .ctl_name       = VM_LOWMEM_RESERVE_RATIO,
 885                 .procname       = "lowmem_reserve_ratio",
 886                 .data           = &sysctl_lowmem_reserve_ratio,
 887                 .maxlen         = sizeof(sysctl_lowmem_reserve_ratio),
 888                 .mode           = 0644,
 889                 .proc_handler   = &lowmem_reserve_ratio_sysctl_handler,
 890                 .strategy       = &sysctl_intvec,
 891         },
 892         {
 893                 .ctl_name       = VM_DROP_PAGECACHE,
 894                 .procname       = "drop_caches",
 895                 .data           = &sysctl_drop_caches,
 896                 .maxlen         = sizeof(int),
 897                 .mode           = 0644,
 898                 .proc_handler   = drop_caches_sysctl_handler,
 899                 .strategy       = &sysctl_intvec,
 900         },
 901         {
 902                 .ctl_name       = VM_MIN_FREE_KBYTES,
 903                 .procname       = "min_free_kbytes",
 904                 .data           = &min_free_kbytes,
 905                 .maxlen         = sizeof(min_free_kbytes),
 906                 .mode           = 0644,
 907                 .proc_handler   = &min_free_kbytes_sysctl_handler,
 908                 .strategy       = &sysctl_intvec,
 909                 .extra1         = &zero,
 910         },
 911         {
 912                 .ctl_name       = VM_PERCPU_PAGELIST_FRACTION,
 913                 .procname       = "percpu_pagelist_fraction",
 914                 .data           = &percpu_pagelist_fraction,
 915                 .maxlen         = sizeof(percpu_pagelist_fraction),
 916                 .mode           = 0644,
 917                 .proc_handler   = &percpu_pagelist_fraction_sysctl_handler,
 918                 .strategy       = &sysctl_intvec,
 919                 .extra1         = &min_percpu_pagelist_fract,
 920         },
 921 #ifdef CONFIG_MMU
 922         {
 923                 .ctl_name       = VM_MAX_MAP_COUNT,
 924                 .procname       = "max_map_count",
 925                 .data           = &sysctl_max_map_count,
 926                 .maxlen         = sizeof(sysctl_max_map_count),
 927                 .mode           = 0644,
 928                 .proc_handler   = &proc_dointvec
 929         },
 930 #endif
 931         {
 932                 .ctl_name       = VM_LAPTOP_MODE,
 933                 .procname       = "laptop_mode",
 934                 .data           = &laptop_mode,
 935                 .maxlen         = sizeof(laptop_mode),
 936                 .mode           = 0644,
 937                 .proc_handler   = &proc_dointvec_jiffies,
 938                 .strategy       = &sysctl_jiffies,
 939         },
 940         {
 941                 .ctl_name       = VM_BLOCK_DUMP,
 942                 .procname       = "block_dump",
 943                 .data           = &block_dump,
 944                 .maxlen         = sizeof(block_dump),
 945                 .mode           = 0644,
 946                 .proc_handler   = &proc_dointvec,
 947                 .strategy       = &sysctl_intvec,
 948                 .extra1         = &zero,
 949         },
 950         {
 951                 .ctl_name       = VM_VFS_CACHE_PRESSURE,
 952                 .procname       = "vfs_cache_pressure",
 953                 .data           = &sysctl_vfs_cache_pressure,
 954                 .maxlen         = sizeof(sysctl_vfs_cache_pressure),
 955                 .mode           = 0644,
 956                 .proc_handler   = &proc_dointvec,
 957                 .strategy       = &sysctl_intvec,
 958                 .extra1         = &zero,
 959         },
 960 #ifdef HAVE_ARCH_PICK_MMAP_LAYOUT
 961         {
 962                 .ctl_name       = VM_LEGACY_VA_LAYOUT,
 963                 .procname       = "legacy_va_layout",
 964                 .data           = &sysctl_legacy_va_layout,
 965                 .maxlen         = sizeof(sysctl_legacy_va_layout),
 966                 .mode           = 0644,
 967                 .proc_handler   = &proc_dointvec,
 968                 .strategy       = &sysctl_intvec,
 969                 .extra1         = &zero,
 970         },
 971 #endif
 972 #ifdef CONFIG_NUMA
 973         {
 974                 .ctl_name       = VM_ZONE_RECLAIM_MODE,
 975                 .procname       = "zone_reclaim_mode",
 976                 .data           = &zone_reclaim_mode,
 977                 .maxlen         = sizeof(zone_reclaim_mode),
 978                 .mode           = 0644,
 979                 .proc_handler   = &proc_dointvec,
 980                 .strategy       = &sysctl_intvec,
 981                 .extra1         = &zero,
 982         },
 983         {
 984                 .ctl_name       = VM_MIN_UNMAPPED,
 985                 .procname       = "min_unmapped_ratio",
 986                 .data           = &sysctl_min_unmapped_ratio,
 987                 .maxlen         = sizeof(sysctl_min_unmapped_ratio),
 988                 .mode           = 0644,
 989                 .proc_handler   = &sysctl_min_unmapped_ratio_sysctl_handler,
 990                 .strategy       = &sysctl_intvec,
 991                 .extra1         = &zero,
 992                 .extra2         = &one_hundred,
 993         },
 994         {
 995                 .ctl_name       = VM_MIN_SLAB,
 996                 .procname       = "min_slab_ratio",
 997                 .data           = &sysctl_min_slab_ratio,
 998                 .maxlen         = sizeof(sysctl_min_slab_ratio),
 999                 .mode           = 0644,
1000                 .proc_handler   = &sysctl_min_slab_ratio_sysctl_handler,
1001                 .strategy       = &sysctl_intvec,
1002                 .extra1         = &zero,
1003                 .extra2         = &one_hundred,
1004         },
1005 #endif
1006 #ifdef CONFIG_SMP
1007         {
1008                 .ctl_name       = CTL_UNNUMBERED,
1009                 .procname       = "stat_interval",
1010                 .data           = &sysctl_stat_interval,
1011                 .maxlen         = sizeof(sysctl_stat_interval),
1012                 .mode           = 0644,
1013                 .proc_handler   = &proc_dointvec_jiffies,
1014                 .strategy       = &sysctl_jiffies,
1015         },
1016 #endif
1017 #ifdef CONFIG_SECURITY
1018         {
1019                 .ctl_name       = CTL_UNNUMBERED,
1020                 .procname       = "mmap_min_addr",
1021                 .data           = &mmap_min_addr,
1022                 .maxlen         = sizeof(unsigned long),
1023                 .mode           = 0644,
1024                 .proc_handler   = &proc_doulongvec_minmax,
1025         },
1026 #ifdef CONFIG_NUMA
1027         {
1028                 .ctl_name       = CTL_UNNUMBERED,
1029                 .procname       = "numa_zonelist_order",
1030                 .data           = &numa_zonelist_order,
1031                 .maxlen         = NUMA_ZONELIST_ORDER_LEN,
1032                 .mode           = 0644,
1033                 .proc_handler   = &numa_zonelist_order_handler,
1034                 .strategy       = &sysctl_string,
1035         },
1036 #endif
1037 #endif
1038 #if defined(CONFIG_X86_32) || \
1039    (defined(CONFIG_SUPERH) && defined(CONFIG_VSYSCALL))
1040         {
1041                 .ctl_name       = VM_VDSO_ENABLED,
1042                 .procname       = "vdso_enabled",
1043                 .data           = &vdso_enabled,
1044                 .maxlen         = sizeof(vdso_enabled),
1045                 .mode           = 0644,
1046                 .proc_handler   = &proc_dointvec,
1047                 .strategy       = &sysctl_intvec,
1048                 .extra1         = &zero,
1049         },
1050 #endif
1051 /*
1052  * NOTE: do not add new entries to this table unless you have read
1053  * Documentation/sysctl/ctl_unnumbered.txt
1054  */
1055         { .ctl_name = 0 }
1056 };
1057
1058 #if defined(CONFIG_BINFMT_MISC) || defined(CONFIG_BINFMT_MISC_MODULE)
1059 static ctl_table binfmt_misc_table[] = {
1060         { .ctl_name = 0 }
1061 };
1062 #endif
1063
1064 static ctl_table fs_table[] = {
1065         {
1066                 .ctl_name       = FS_NRINODE,
1067                 .procname       = "inode-nr",
1068                 .data           = &inodes_stat,
1069                 .maxlen         = 2*sizeof(int),
1070                 .mode           = 0444,
1071                 .proc_handler   = &proc_dointvec,
1072         },
1073         {
1074                 .ctl_name       = FS_STATINODE,
1075                 .procname       = "inode-state",
1076                 .data           = &inodes_stat,
1077                 .maxlen         = 7*sizeof(int),
1078                 .mode           = 0444,
1079                 .proc_handler   = &proc_dointvec,
1080         },
1081         {
1082                 .ctl_name       = FS_NRFILE,
1083                 .procname       = "file-nr",
1084                 .data           = &files_stat,
1085                 .maxlen         = 3*sizeof(int),
1086                 .mode           = 0444,
1087                 .proc_handler   = &proc_nr_files,
1088         },
1089         {
1090                 .ctl_name       = FS_MAXFILE,
1091                 .procname       = "file-max",
1092                 .data           = &files_stat.max_files,
1093                 .maxlen         = sizeof(int),
1094                 .mode           = 0644,
1095                 .proc_handler   = &proc_dointvec,
1096         },
1097         {
1098                 .ctl_name       = FS_DENTRY,
1099                 .procname       = "dentry-state",
1100                 .data           = &dentry_stat,
1101                 .maxlen         = 6*sizeof(int),
1102                 .mode           = 0444,
1103                 .proc_handler   = &proc_dointvec,
1104         },
1105         {
1106                 .ctl_name       = FS_OVERFLOWUID,
1107                 .procname       = "overflowuid",
1108                 .data           = &fs_overflowuid,
1109                 .maxlen         = sizeof(int),
1110                 .mode           = 0644,
1111                 .proc_handler   = &proc_dointvec_minmax,
1112                 .strategy       = &sysctl_intvec,
1113                 .extra1         = &minolduid,
1114                 .extra2         = &maxolduid,
1115         },
1116         {
1117                 .ctl_name       = FS_OVERFLOWGID,
1118                 .procname       = "overflowgid",
1119                 .data           = &fs_overflowgid,
1120                 .maxlen         = sizeof(int),
1121                 .mode           = 0644,
1122                 .proc_handler   = &proc_dointvec_minmax,
1123                 .strategy       = &sysctl_intvec,
1124                 .extra1         = &minolduid,
1125                 .extra2         = &maxolduid,
1126         },
1127         {
1128                 .ctl_name       = FS_LEASES,
1129                 .procname       = "leases-enable",
1130                 .data           = &leases_enable,
1131                 .maxlen         = sizeof(int),
1132                 .mode           = 0644,
1133                 .proc_handler   = &proc_dointvec,
1134         },
1135 #ifdef CONFIG_DNOTIFY
1136         {
1137                 .ctl_name       = FS_DIR_NOTIFY,
1138                 .procname       = "dir-notify-enable",
1139                 .data           = &dir_notify_enable,
1140                 .maxlen         = sizeof(int),
1141                 .mode           = 0644,
1142                 .proc_handler   = &proc_dointvec,
1143         },
1144 #endif
1145 #ifdef CONFIG_MMU
1146         {
1147                 .ctl_name       = FS_LEASE_TIME,
1148                 .procname       = "lease-break-time",
1149                 .data           = &lease_break_time,
1150                 .maxlen         = sizeof(int),
1151                 .mode           = 0644,
1152                 .proc_handler   = &proc_dointvec_minmax,
1153                 .strategy       = &sysctl_intvec,
1154                 .extra1         = &zero,
1155                 .extra2         = &two,
1156         },
1157         {
1158                 .ctl_name       = FS_AIO_NR,
1159                 .procname       = "aio-nr",
1160                 .data           = &aio_nr,
1161                 .maxlen         = sizeof(aio_nr),
1162                 .mode           = 0444,
1163                 .proc_handler   = &proc_doulongvec_minmax,
1164         },
1165         {
1166                 .ctl_name       = FS_AIO_MAX_NR,
1167                 .procname       = "aio-max-nr",
1168                 .data           = &aio_max_nr,
1169                 .maxlen         = sizeof(aio_max_nr),
1170                 .mode           = 0644,
1171                 .proc_handler   = &proc_doulongvec_minmax,
1172         },
1173 #ifdef CONFIG_INOTIFY_USER
1174         {
1175                 .ctl_name       = FS_INOTIFY,
1176                 .procname       = "inotify",
1177                 .mode           = 0555,
1178                 .child          = inotify_table,
1179         },
1180 #endif
1181 #endif
1182         {
1183                 .ctl_name       = KERN_SETUID_DUMPABLE,
1184                 .procname       = "suid_dumpable",
1185                 .data           = &suid_dumpable,
1186                 .maxlen         = sizeof(int),
1187                 .mode           = 0644,
1188                 .proc_handler   = &proc_dointvec,
1189         },
1190 #if defined(CONFIG_BINFMT_MISC) || defined(CONFIG_BINFMT_MISC_MODULE)
1191         {
1192                 .ctl_name       = CTL_UNNUMBERED,
1193                 .procname       = "binfmt_misc",
1194                 .mode           = 0555,
1195                 .child          = binfmt_misc_table,
1196         },
1197 #endif
1198 /*
1199  * NOTE: do not add new entries to this table unless you have read
1200  * Documentation/sysctl/ctl_unnumbered.txt
1201  */
1202         { .ctl_name = 0 }
1203 };
1204
1205 static ctl_table debug_table[] = {
1206         { .ctl_name = 0 }
1207 };
1208
1209 static ctl_table dev_table[] = {
1210         { .ctl_name = 0 }
1211 };
1212
1213 static DEFINE_SPINLOCK(sysctl_lock);
1214
1215 /* called under sysctl_lock */
1216 static int use_table(struct ctl_table_header *p)
1217 {
1218         if (unlikely(p->unregistering))
1219                 return 0;
1220         p->used++;
1221         return 1;
1222 }
1223
1224 /* called under sysctl_lock */
1225 static void unuse_table(struct ctl_table_header *p)
1226 {
1227         if (!--p->used)
1228                 if (unlikely(p->unregistering))
1229                         complete(p->unregistering);
1230 }
1231
1232 /* called under sysctl_lock, will reacquire if has to wait */
1233 static void start_unregistering(struct ctl_table_header *p)
1234 {
1235         /*
1236          * if p->used is 0, nobody will ever touch that entry again;
1237          * we'll eliminate all paths to it before dropping sysctl_lock
1238          */
1239         if (unlikely(p->used)) {
1240                 struct completion wait;
1241                 init_completion(&wait);
1242                 p->unregistering = &wait;
1243                 spin_unlock(&sysctl_lock);
1244                 wait_for_completion(&wait);
1245                 spin_lock(&sysctl_lock);
1246         }
1247         /*
1248          * do not remove from the list until nobody holds it; walking the
1249          * list in do_sysctl() relies on that.
1250          */
1251         list_del_init(&p->ctl_entry);
1252 }
1253
1254 void sysctl_head_finish(struct ctl_table_header *head)
1255 {
1256         if (!head)
1257                 return;
1258         spin_lock(&sysctl_lock);
1259         unuse_table(head);
1260         spin_unlock(&sysctl_lock);
1261 }
1262
1263 struct ctl_table_header *sysctl_head_next(struct ctl_table_header *prev)
1264 {
1265         struct ctl_table_header *head;
1266         struct list_head *tmp;
1267         spin_lock(&sysctl_lock);
1268         if (prev) {
1269                 tmp = &prev->ctl_entry;
1270                 unuse_table(prev);
1271                 goto next;
1272         }
1273         tmp = &root_table_header.ctl_entry;
1274         for (;;) {
1275                 head = list_entry(tmp, struct ctl_table_header, ctl_entry);
1276
1277                 if (!use_table(head))
1278                         goto next;
1279                 spin_unlock(&sysctl_lock);
1280                 return head;
1281         next:
1282                 tmp = tmp->next;
1283                 if (tmp == &root_table_header.ctl_entry)
1284                         break;
1285         }
1286         spin_unlock(&sysctl_lock);
1287         return NULL;
1288 }
1289
1290 #ifdef CONFIG_SYSCTL_SYSCALL
1291 int do_sysctl(int __user *name, int nlen, void __user *oldval, size_t __user *oldlenp,
1292                void __user *newval, size_t newlen)
1293 {
1294         struct ctl_table_header *head;
1295         int error = -ENOTDIR;
1296
1297         if (nlen <= 0 || nlen >= CTL_MAXNAME)
1298                 return -ENOTDIR;
1299         if (oldval) {
1300                 int old_len;
1301                 if (!oldlenp || get_user(old_len, oldlenp))
1302                         return -EFAULT;
1303         }
1304
1305         for (head = sysctl_head_next(NULL); head;
1306                         head = sysctl_head_next(head)) {
1307                 error = parse_table(name, nlen, oldval, oldlenp,
1308                                         newval, newlen, head->ctl_table);
1309                 if (error != -ENOTDIR) {
1310                         sysctl_head_finish(head);
1311                         break;
1312                 }
1313         }
1314         return error;
1315 }
1316
1317 asmlinkage long sys_sysctl(struct __sysctl_args __user *args)
1318 {
1319         struct __sysctl_args tmp;
1320         int error;
1321
1322         if (copy_from_user(&tmp, args, sizeof(tmp)))
1323                 return -EFAULT;
1324
1325         lock_kernel();
1326         error = do_sysctl(tmp.name, tmp.nlen, tmp.oldval, tmp.oldlenp,
1327                           tmp.newval, tmp.newlen);
1328         unlock_kernel();
1329         return error;
1330 }
1331 #endif /* CONFIG_SYSCTL_SYSCALL */
1332
1333 /*
1334  * sysctl_perm does NOT grant the superuser all rights automatically, because
1335  * some sysctl variables are readonly even to root.
1336  */
1337
1338 static int test_perm(int mode, int op)
1339 {
1340         if (!current->euid)
1341                 mode >>= 6;
1342         else if (in_egroup_p(0))
1343                 mode >>= 3;
1344         if ((mode & op & 0007) == op)
1345                 return 0;
1346         return -EACCES;
1347 }
1348
1349 int sysctl_perm(ctl_table *table, int op)
1350 {
1351         int error;
1352         error = security_sysctl(table, op);
1353         if (error)
1354                 return error;
1355         return test_perm(table->mode, op);
1356 }
1357
1358 #ifdef CONFIG_SYSCTL_SYSCALL
1359 static int parse_table(int __user *name, int nlen,
1360                        void __user *oldval, size_t __user *oldlenp,
1361                        void __user *newval, size_t newlen,
1362                        ctl_table *table)
1363 {
1364         int n;
1365 repeat:
1366         if (!nlen)
1367                 return -ENOTDIR;
1368         if (get_user(n, name))
1369                 return -EFAULT;
1370         for ( ; table->ctl_name || table->procname; table++) {
1371                 if (!table->ctl_name)
1372                         continue;
1373                 if (n == table->ctl_name) {
1374                         int error;
1375                         if (table->child) {
1376                                 if (sysctl_perm(table, 001))
1377                                         return -EPERM;
1378                                 name++;
1379                                 nlen--;
1380                                 table = table->child;
1381                                 goto repeat;
1382                         }
1383                         error = do_sysctl_strategy(table, name, nlen,
1384                                                    oldval, oldlenp,
1385                                                    newval, newlen);
1386                         return error;
1387                 }
1388         }
1389         return -ENOTDIR;
1390 }
1391
1392 /* Perform the actual read/write of a sysctl table entry. */
1393 int do_sysctl_strategy (ctl_table *table,
1394                         int __user *name, int nlen,
1395                         void __user *oldval, size_t __user *oldlenp,
1396                         void __user *newval, size_t newlen)
1397 {
1398         int op = 0, rc;
1399         size_t len;
1400
1401         if (oldval)
1402                 op |= 004;
1403         if (newval)
1404                 op |= 002;
1405         if (sysctl_perm(table, op))
1406                 return -EPERM;
1407
1408         if (table->strategy) {
1409                 rc = table->strategy(table, name, nlen, oldval, oldlenp,
1410                                      newval, newlen);
1411                 if (rc < 0)
1412                         return rc;
1413                 if (rc > 0)
1414                         return 0;
1415         }
1416
1417         /* If there is no strategy routine, or if the strategy returns
1418          * zero, proceed with automatic r/w */
1419         if (table->data && table->maxlen) {
1420                 if (oldval && oldlenp) {
1421                         if (get_user(len, oldlenp))
1422                                 return -EFAULT;
1423                         if (len) {
1424                                 if (len > table->maxlen)
1425                                         len = table->maxlen;
1426                                 if(copy_to_user(oldval, table->data, len))
1427                                         return -EFAULT;
1428                                 if(put_user(len, oldlenp))
1429                                         return -EFAULT;
1430                         }
1431                 }
1432                 if (newval && newlen) {
1433                         len = newlen;
1434                         if (len > table->maxlen)
1435                                 len = table->maxlen;
1436                         if(copy_from_user(table->data, newval, len))
1437                                 return -EFAULT;
1438                 }
1439         }
1440         return 0;
1441 }
1442 #endif /* CONFIG_SYSCTL_SYSCALL */
1443
1444 static void sysctl_set_parent(struct ctl_table *parent, struct ctl_table *table)
1445 {
1446         for (; table->ctl_name || table->procname; table++) {
1447                 table->parent = parent;
1448                 if (table->child)
1449                         sysctl_set_parent(table, table->child);
1450         }
1451 }
1452
1453 static __init int sysctl_init(void)
1454 {
1455         sysctl_set_parent(NULL, root_table);
1456         return 0;
1457 }
1458
1459 core_initcall(sysctl_init);
1460
1461 /**
1462  * register_sysctl_table - register a sysctl hierarchy
1463  * @table: the top-level table structure
1464  *
1465  * Register a sysctl table hierarchy. @table should be a filled in ctl_table
1466  * array. An entry with a ctl_name of 0 terminates the table.
1467  *
1468  * The members of the &ctl_table structure are used as follows:
1469  *
1470  * ctl_name - This is the numeric sysctl value used by sysctl(2). The number
1471  *            must be unique within that level of sysctl
1472  *
1473  * procname - the name of the sysctl file under /proc/sys. Set to %NULL to not
1474  *            enter a sysctl file
1475  *
1476  * data - a pointer to data for use by proc_handler
1477  *
1478  * maxlen - the maximum size in bytes of the data
1479  *
1480  * mode - the file permissions for the /proc/sys file, and for sysctl(2)
1481  *
1482  * child - a pointer to the child sysctl table if this entry is a directory, or
1483  *         %NULL.
1484  *
1485  * proc_handler - the text handler routine (described below)
1486  *
1487  * strategy - the strategy routine (described below)
1488  *
1489  * de - for internal use by the sysctl routines
1490  *
1491  * extra1, extra2 - extra pointers usable by the proc handler routines
1492  *
1493  * Leaf nodes in the sysctl tree will be represented by a single file
1494  * under /proc; non-leaf nodes will be represented by directories.
1495  *
1496  * sysctl(2) can automatically manage read and write requests through
1497  * the sysctl table.  The data and maxlen fields of the ctl_table
1498  * struct enable minimal validation of the values being written to be
1499  * performed, and the mode field allows minimal authentication.
1500  *
1501  * More sophisticated management can be enabled by the provision of a
1502  * strategy routine with the table entry.  This will be called before
1503  * any automatic read or write of the data is performed.
1504  *
1505  * The strategy routine may return
1506  *
1507  * < 0 - Error occurred (error is passed to user process)
1508  *
1509  * 0   - OK - proceed with automatic read or write.
1510  *
1511  * > 0 - OK - read or write has been done by the strategy routine, so
1512  *       return immediately.
1513  *
1514  * There must be a proc_handler routine for any terminal nodes
1515  * mirrored under /proc/sys (non-terminals are handled by a built-in
1516  * directory handler).  Several default handlers are available to
1517  * cover common cases -
1518  *
1519  * proc_dostring(), proc_dointvec(), proc_dointvec_jiffies(),
1520  * proc_dointvec_userhz_jiffies(), proc_dointvec_minmax(),
1521  * proc_doulongvec_ms_jiffies_minmax(), proc_doulongvec_minmax()
1522  *
1523  * It is the handler's job to read the input buffer from user memory
1524  * and process it. The handler should return 0 on success.
1525  *
1526  * This routine returns %NULL on a failure to register, and a pointer
1527  * to the table header on success.
1528  */
1529 struct ctl_table_header *register_sysctl_table(ctl_table * table)
1530 {
1531         struct ctl_table_header *tmp;
1532         tmp = kmalloc(sizeof(struct ctl_table_header), GFP_KERNEL);
1533         if (!tmp)
1534                 return NULL;
1535         tmp->ctl_table = table;
1536         INIT_LIST_HEAD(&tmp->ctl_entry);
1537         tmp->used = 0;
1538         tmp->unregistering = NULL;
1539         sysctl_set_parent(NULL, table);
1540         spin_lock(&sysctl_lock);
1541         list_add_tail(&tmp->ctl_entry, &root_table_header.ctl_entry);
1542         spin_unlock(&sysctl_lock);
1543         return tmp;
1544 }
1545
1546 /**
1547  * unregister_sysctl_table - unregister a sysctl table hierarchy
1548  * @header: the header returned from register_sysctl_table
1549  *
1550  * Unregisters the sysctl table and all children. proc entries may not
1551  * actually be removed until they are no longer used by anyone.
1552  */
1553 void unregister_sysctl_table(struct ctl_table_header * header)
1554 {
1555         might_sleep();
1556         spin_lock(&sysctl_lock);
1557         start_unregistering(header);
1558         spin_unlock(&sysctl_lock);
1559         kfree(header);
1560 }
1561
1562 #else /* !CONFIG_SYSCTL */
1563 struct ctl_table_header *register_sysctl_table(ctl_table * table)
1564 {
1565         return NULL;
1566 }
1567
1568 void unregister_sysctl_table(struct ctl_table_header * table)
1569 {
1570 }
1571
1572 #endif /* CONFIG_SYSCTL */
1573
1574 /*
1575  * /proc/sys support
1576  */
1577
1578 #ifdef CONFIG_PROC_SYSCTL
1579
1580 static int _proc_do_string(void* data, int maxlen, int write,
1581                            struct file *filp, void __user *buffer,
1582                            size_t *lenp, loff_t *ppos)
1583 {
1584         size_t len;
1585         char __user *p;
1586         char c;
1587
1588         if (!data || !maxlen || !*lenp) {
1589                 *lenp = 0;
1590                 return 0;
1591         }
1592
1593         if (write) {
1594                 len = 0;
1595                 p = buffer;
1596                 while (len < *lenp) {
1597                         if (get_user(c, p++))
1598                                 return -EFAULT;
1599                         if (c == 0 || c == '\n')
1600                                 break;
1601                         len++;
1602                 }
1603                 if (len >= maxlen)
1604                         len = maxlen-1;
1605                 if(copy_from_user(data, buffer, len))
1606                         return -EFAULT;
1607                 ((char *) data)[len] = 0;
1608                 *ppos += *lenp;
1609         } else {
1610                 len = strlen(data);
1611                 if (len > maxlen)
1612                         len = maxlen;
1613
1614                 if (*ppos > len) {
1615                         *lenp = 0;
1616                         return 0;
1617                 }
1618
1619                 data += *ppos;
1620                 len  -= *ppos;
1621
1622                 if (len > *lenp)
1623                         len = *lenp;
1624                 if (len)
1625                         if(copy_to_user(buffer, data, len))
1626                                 return -EFAULT;
1627                 if (len < *lenp) {
1628                         if(put_user('\n', ((char __user *) buffer) + len))
1629                                 return -EFAULT;
1630                         len++;
1631                 }
1632                 *lenp = len;
1633                 *ppos += len;
1634         }
1635         return 0;
1636 }
1637
1638 /**
1639  * proc_dostring - read a string sysctl
1640  * @table: the sysctl table
1641  * @write: %TRUE if this is a write to the sysctl file
1642  * @filp: the file structure
1643  * @buffer: the user buffer
1644  * @lenp: the size of the user buffer
1645  * @ppos: file position
1646  *
1647  * Reads/writes a string from/to the user buffer. If the kernel
1648  * buffer provided is not large enough to hold the string, the
1649  * string is truncated. The copied string is %NULL-terminated.
1650  * If the string is being read by the user process, it is copied
1651  * and a newline '\n' is added. It is truncated if the buffer is
1652  * not large enough.
1653  *
1654  * Returns 0 on success.
1655  */
1656 int proc_dostring(ctl_table *table, int write, struct file *filp,
1657                   void __user *buffer, size_t *lenp, loff_t *ppos)
1658 {
1659         return _proc_do_string(table->data, table->maxlen, write, filp,
1660                                buffer, lenp, ppos);
1661 }
1662
1663
1664 static int do_proc_dointvec_conv(int *negp, unsigned long *lvalp,
1665                                  int *valp,
1666                                  int write, void *data)
1667 {
1668         if (write) {
1669                 *valp = *negp ? -*lvalp : *lvalp;
1670         } else {
1671                 int val = *valp;
1672                 if (val < 0) {
1673                         *negp = -1;
1674                         *lvalp = (unsigned long)-val;
1675                 } else {
1676                         *negp = 0;
1677                         *lvalp = (unsigned long)val;
1678                 }
1679         }
1680         return 0;
1681 }
1682
1683 static int __do_proc_dointvec(void *tbl_data, ctl_table *table,
1684                   int write, struct file *filp, void __user *buffer,
1685                   size_t *lenp, loff_t *ppos,
1686                   int (*conv)(int *negp, unsigned long *lvalp, int *valp,
1687                               int write, void *data),
1688                   void *data)
1689 {
1690 #define TMPBUFLEN 21
1691         int *i, vleft, first=1, neg, val;
1692         unsigned long lval;
1693         size_t left, len;
1694
1695         char buf[TMPBUFLEN], *p;
1696         char __user *s = buffer;
1697
1698         if (!tbl_data || !table->maxlen || !*lenp ||
1699             (*ppos && !write)) {
1700                 *lenp = 0;
1701                 return 0;
1702         }
1703
1704         i = (int *) tbl_data;
1705         vleft = table->maxlen / sizeof(*i);
1706         left = *lenp;
1707
1708         if (!conv)
1709                 conv = do_proc_dointvec_conv;
1710
1711         for (; left && vleft--; i++, first=0) {
1712                 if (write) {
1713                         while (left) {
1714                                 char c;
1715                                 if (get_user(c, s))
1716                                         return -EFAULT;
1717                                 if (!isspace(c))
1718                                         break;
1719                                 left--;
1720                                 s++;
1721                         }
1722                         if (!left)
1723                                 break;
1724                         neg = 0;
1725                         len = left;
1726                         if (len > sizeof(buf) - 1)
1727                                 len = sizeof(buf) - 1;
1728                         if (copy_from_user(buf, s, len))
1729                                 return -EFAULT;
1730                         buf[len] = 0;
1731                         p = buf;
1732                         if (*p == '-' && left > 1) {
1733                                 neg = 1;
1734                                 p++;
1735                         }
1736                         if (*p < '0' || *p > '9')
1737                                 break;
1738
1739                         lval = simple_strtoul(p, &p, 0);
1740
1741                         len = p-buf;
1742                         if ((len < left) && *p && !isspace(*p))
1743                                 break;
1744                         if (neg)
1745                                 val = -val;
1746                         s += len;
1747                         left -= len;
1748
1749                         if (conv(&neg, &lval, i, 1, data))
1750                                 break;
1751                 } else {
1752                         p = buf;
1753                         if (!first)
1754                                 *p++ = '\t';
1755
1756                         if (conv(&neg, &lval, i, 0, data))
1757                                 break;
1758
1759                         sprintf(p, "%s%lu", neg ? "-" : "", lval);
1760                         len = strlen(buf);
1761                         if (len > left)
1762                                 len = left;
1763                         if(copy_to_user(s, buf, len))
1764                                 return -EFAULT;
1765                         left -= len;
1766                         s += len;
1767                 }
1768         }
1769
1770         if (!write && !first && left) {
1771                 if(put_user('\n', s))
1772                         return -EFAULT;
1773                 left--, s++;
1774         }
1775         if (write) {
1776                 while (left) {
1777                         char c;
1778                         if (get_user(c, s++))
1779                                 return -EFAULT;
1780                         if (!isspace(c))
1781                                 break;
1782                         left--;
1783                 }
1784         }
1785         if (write && first)
1786                 return -EINVAL;
1787         *lenp -= left;
1788         *ppos += *lenp;
1789         return 0;
1790 #undef TMPBUFLEN
1791 }
1792
1793 static int do_proc_dointvec(ctl_table *table, int write, struct file *filp,
1794                   void __user *buffer, size_t *lenp, loff_t *ppos,
1795                   int (*conv)(int *negp, unsigned long *lvalp, int *valp,
1796                               int write, void *data),
1797                   void *data)
1798 {
1799         return __do_proc_dointvec(table->data, table, write, filp,
1800                         buffer, lenp, ppos, conv, data);
1801 }
1802
1803 /**
1804  * proc_dointvec - read a vector of integers
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  * Returns 0 on success.
1816  */
1817 int proc_dointvec(ctl_table *table, int write, struct file *filp,
1818                      void __user *buffer, size_t *lenp, loff_t *ppos)
1819 {
1820     return do_proc_dointvec(table,write,filp,buffer,lenp,ppos,
1821                             NULL,NULL);
1822 }
1823
1824 #define OP_SET  0
1825 #define OP_AND  1
1826 #define OP_OR   2
1827
1828 static int do_proc_dointvec_bset_conv(int *negp, unsigned long *lvalp,
1829                                       int *valp,
1830                                       int write, void *data)
1831 {
1832         int op = *(int *)data;
1833         if (write) {
1834                 int val = *negp ? -*lvalp : *lvalp;
1835                 switch(op) {
1836                 case OP_SET:    *valp = val; break;
1837                 case OP_AND:    *valp &= val; break;
1838                 case OP_OR:     *valp |= val; break;
1839                 }
1840         } else {
1841                 int val = *valp;
1842                 if (val < 0) {
1843                         *negp = -1;
1844                         *lvalp = (unsigned long)-val;
1845                 } else {
1846                         *negp = 0;
1847                         *lvalp = (unsigned long)val;
1848                 }
1849         }
1850         return 0;
1851 }
1852
1853 /*
1854  *      init may raise the set.
1855  */
1856
1857 int proc_dointvec_bset(ctl_table *table, int write, struct file *filp,
1858                         void __user *buffer, size_t *lenp, loff_t *ppos)
1859 {
1860         int op;
1861
1862         if (write && !capable(CAP_SYS_MODULE)) {
1863                 return -EPERM;
1864         }
1865
1866         op = is_init(current) ? OP_SET : OP_AND;
1867         return do_proc_dointvec(table,write,filp,buffer,lenp,ppos,
1868                                 do_proc_dointvec_bset_conv,&op);
1869 }
1870
1871 /*
1872  *      Taint values can only be increased
1873  */
1874 static int proc_dointvec_taint(ctl_table *table, int write, struct file *filp,
1875                                void __user *buffer, size_t *lenp, loff_t *ppos)
1876 {
1877         int op;
1878
1879         if (write && !capable(CAP_SYS_ADMIN))
1880                 return -EPERM;
1881
1882         op = OP_OR;
1883         return do_proc_dointvec(table,write,filp,buffer,lenp,ppos,
1884                                 do_proc_dointvec_bset_conv,&op);
1885 }
1886
1887 struct do_proc_dointvec_minmax_conv_param {
1888         int *min;
1889         int *max;
1890 };
1891
1892 static int do_proc_dointvec_minmax_conv(int *negp, unsigned long *lvalp,
1893                                         int *valp,
1894                                         int write, void *data)
1895 {
1896         struct do_proc_dointvec_minmax_conv_param *param = data;
1897         if (write) {
1898                 int val = *negp ? -*lvalp : *lvalp;
1899                 if ((param->min && *param->min > val) ||
1900                     (param->max && *param->max < val))
1901                         return -EINVAL;
1902                 *valp = val;
1903         } else {
1904                 int val = *valp;
1905                 if (val < 0) {
1906                         *negp = -1;
1907                         *lvalp = (unsigned long)-val;
1908                 } else {
1909                         *negp = 0;
1910                         *lvalp = (unsigned long)val;
1911                 }
1912         }
1913         return 0;
1914 }
1915
1916 /**
1917  * proc_dointvec_minmax - read a vector of integers with min/max values
1918  * @table: the sysctl table
1919  * @write: %TRUE if this is a write to the sysctl file
1920  * @filp: the file structure
1921  * @buffer: the user buffer
1922  * @lenp: the size of the user buffer
1923  * @ppos: file position
1924  *
1925  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
1926  * values from/to the user buffer, treated as an ASCII string.
1927  *
1928  * This routine will ensure the values are within the range specified by
1929  * table->extra1 (min) and table->extra2 (max).
1930  *
1931  * Returns 0 on success.
1932  */
1933 int proc_dointvec_minmax(ctl_table *table, int write, struct file *filp,
1934                   void __user *buffer, size_t *lenp, loff_t *ppos)
1935 {
1936         struct do_proc_dointvec_minmax_conv_param param = {
1937                 .min = (int *) table->extra1,
1938                 .max = (int *) table->extra2,
1939         };
1940         return do_proc_dointvec(table, write, filp, buffer, lenp, ppos,
1941                                 do_proc_dointvec_minmax_conv, &param);
1942 }
1943
1944 static int __do_proc_doulongvec_minmax(void *data, ctl_table *table, int write,
1945                                      struct file *filp,
1946                                      void __user *buffer,
1947                                      size_t *lenp, loff_t *ppos,
1948                                      unsigned long convmul,
1949                                      unsigned long convdiv)
1950 {
1951 #define TMPBUFLEN 21
1952         unsigned long *i, *min, *max, val;
1953         int vleft, first=1, neg;
1954         size_t len, left;
1955         char buf[TMPBUFLEN], *p;
1956         char __user *s = buffer;
1957
1958         if (!data || !table->maxlen || !*lenp ||
1959             (*ppos && !write)) {
1960                 *lenp = 0;
1961                 return 0;
1962         }
1963
1964         i = (unsigned long *) data;
1965         min = (unsigned long *) table->extra1;
1966         max = (unsigned long *) table->extra2;
1967         vleft = table->maxlen / sizeof(unsigned long);
1968         left = *lenp;
1969
1970         for (; left && vleft--; i++, min++, max++, first=0) {
1971                 if (write) {
1972                         while (left) {
1973                                 char c;
1974                                 if (get_user(c, s))
1975                                         return -EFAULT;
1976                                 if (!isspace(c))
1977                                         break;
1978                                 left--;
1979                                 s++;
1980                         }
1981                         if (!left)
1982                                 break;
1983                         neg = 0;
1984                         len = left;
1985                         if (len > TMPBUFLEN-1)
1986                                 len = TMPBUFLEN-1;
1987                         if (copy_from_user(buf, s, len))
1988                                 return -EFAULT;
1989                         buf[len] = 0;
1990                         p = buf;
1991                         if (*p == '-' && left > 1) {
1992                                 neg = 1;
1993                                 p++;
1994                         }
1995                         if (*p < '0' || *p > '9')
1996                                 break;
1997                         val = simple_strtoul(p, &p, 0) * convmul / convdiv ;
1998                         len = p-buf;
1999                         if ((len < left) && *p && !isspace(*p))
2000                                 break;
2001                         if (neg)
2002                                 val = -val;
2003                         s += len;
2004                         left -= len;
2005
2006                         if(neg)
2007                                 continue;
2008                         if ((min && val < *min) || (max && val > *max))
2009                                 continue;
2010                         *i = val;
2011                 } else {
2012                         p = buf;
2013                         if (!first)
2014                                 *p++ = '\t';
2015                         sprintf(p, "%lu", convdiv * (*i) / convmul);
2016                         len = strlen(buf);
2017                         if (len > left)
2018                                 len = left;
2019                         if(copy_to_user(s, buf, len))
2020                                 return -EFAULT;
2021                         left -= len;
2022                         s += len;
2023                 }
2024         }
2025
2026         if (!write && !first && left) {
2027                 if(put_user('\n', s))
2028                         return -EFAULT;
2029                 left--, s++;
2030         }
2031         if (write) {
2032                 while (left) {
2033                         char c;
2034                         if (get_user(c, s++))
2035                                 return -EFAULT;
2036                         if (!isspace(c))
2037                                 break;
2038                         left--;
2039                 }
2040         }
2041         if (write && first)
2042                 return -EINVAL;
2043         *lenp -= left;
2044         *ppos += *lenp;
2045         return 0;
2046 #undef TMPBUFLEN
2047 }
2048
2049 static int do_proc_doulongvec_minmax(ctl_table *table, int write,
2050                                      struct file *filp,
2051                                      void __user *buffer,
2052                                      size_t *lenp, loff_t *ppos,
2053                                      unsigned long convmul,
2054                                      unsigned long convdiv)
2055 {
2056         return __do_proc_doulongvec_minmax(table->data, table, write,
2057                         filp, buffer, lenp, ppos, convmul, convdiv);
2058 }
2059
2060 /**
2061  * proc_doulongvec_minmax - read a vector of long integers with min/max values
2062  * @table: the sysctl table
2063  * @write: %TRUE if this is a write to the sysctl file
2064  * @filp: the file structure
2065  * @buffer: the user buffer
2066  * @lenp: the size of the user buffer
2067  * @ppos: file position
2068  *
2069  * Reads/writes up to table->maxlen/sizeof(unsigned long) unsigned long
2070  * values from/to the user buffer, treated as an ASCII string.
2071  *
2072  * This routine will ensure the values are within the range specified by
2073  * table->extra1 (min) and table->extra2 (max).
2074  *
2075  * Returns 0 on success.
2076  */
2077 int proc_doulongvec_minmax(ctl_table *table, int write, struct file *filp,
2078                            void __user *buffer, size_t *lenp, loff_t *ppos)
2079 {
2080     return do_proc_doulongvec_minmax(table, write, filp, buffer, lenp, ppos, 1l, 1l);
2081 }
2082
2083 /**
2084  * proc_doulongvec_ms_jiffies_minmax - read a vector of millisecond values with min/max values
2085  * @table: the sysctl table
2086  * @write: %TRUE if this is a write to the sysctl file
2087  * @filp: the file structure
2088  * @buffer: the user buffer
2089  * @lenp: the size of the user buffer
2090  * @ppos: file position
2091  *
2092  * Reads/writes up to table->maxlen/sizeof(unsigned long) unsigned long
2093  * values from/to the user buffer, treated as an ASCII string. The values
2094  * are treated as milliseconds, and converted to jiffies when they are stored.
2095  *
2096  * This routine will ensure the values are within the range specified by
2097  * table->extra1 (min) and table->extra2 (max).
2098  *
2099  * Returns 0 on success.
2100  */
2101 int proc_doulongvec_ms_jiffies_minmax(ctl_table *table, int write,
2102                                       struct file *filp,
2103                                       void __user *buffer,
2104                                       size_t *lenp, loff_t *ppos)
2105 {
2106     return do_proc_doulongvec_minmax(table, write, filp, buffer,
2107                                      lenp, ppos, HZ, 1000l);
2108 }
2109
2110
2111 static int do_proc_dointvec_jiffies_conv(int *negp, unsigned long *lvalp,
2112                                          int *valp,
2113                                          int write, void *data)
2114 {
2115         if (write) {
2116                 if (*lvalp > LONG_MAX / HZ)
2117                         return 1;
2118                 *valp = *negp ? -(*lvalp*HZ) : (*lvalp*HZ);
2119         } else {
2120                 int val = *valp;
2121                 unsigned long lval;
2122                 if (val < 0) {
2123                         *negp = -1;
2124                         lval = (unsigned long)-val;
2125                 } else {
2126                         *negp = 0;
2127                         lval = (unsigned long)val;
2128                 }
2129                 *lvalp = lval / HZ;
2130         }
2131         return 0;
2132 }
2133
2134 static int do_proc_dointvec_userhz_jiffies_conv(int *negp, unsigned long *lvalp,
2135                                                 int *valp,
2136                                                 int write, void *data)
2137 {
2138         if (write) {
2139                 if (USER_HZ < HZ && *lvalp > (LONG_MAX / HZ) * USER_HZ)
2140                         return 1;
2141                 *valp = clock_t_to_jiffies(*negp ? -*lvalp : *lvalp);
2142         } else {
2143                 int val = *valp;
2144                 unsigned long lval;
2145                 if (val < 0) {
2146                         *negp = -1;
2147                         lval = (unsigned long)-val;
2148                 } else {
2149                         *negp = 0;
2150                         lval = (unsigned long)val;
2151                 }
2152                 *lvalp = jiffies_to_clock_t(lval);
2153         }
2154         return 0;
2155 }
2156
2157 static int do_proc_dointvec_ms_jiffies_conv(int *negp, unsigned long *lvalp,
2158                                             int *valp,
2159                                             int write, void *data)
2160 {
2161         if (write) {
2162                 *valp = msecs_to_jiffies(*negp ? -*lvalp : *lvalp);
2163         } else {
2164                 int val = *valp;
2165                 unsigned long lval;
2166                 if (val < 0) {
2167                         *negp = -1;
2168                         lval = (unsigned long)-val;
2169                 } else {
2170                         *negp = 0;
2171                         lval = (unsigned long)val;
2172                 }
2173                 *lvalp = jiffies_to_msecs(lval);
2174         }
2175         return 0;
2176 }
2177
2178 /**
2179  * proc_dointvec_jiffies - read a vector of integers as seconds
2180  * @table: the sysctl table
2181  * @write: %TRUE if this is a write to the sysctl file
2182  * @filp: the file structure
2183  * @buffer: the user buffer
2184  * @lenp: the size of the user buffer
2185  * @ppos: file position
2186  *
2187  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2188  * values from/to the user buffer, treated as an ASCII string.
2189  * The values read are assumed to be in seconds, and are converted into
2190  * jiffies.
2191  *
2192  * Returns 0 on success.
2193  */
2194 int proc_dointvec_jiffies(ctl_table *table, int write, struct file *filp,
2195                           void __user *buffer, size_t *lenp, loff_t *ppos)
2196 {
2197     return do_proc_dointvec(table,write,filp,buffer,lenp,ppos,
2198                             do_proc_dointvec_jiffies_conv,NULL);
2199 }
2200
2201 /**
2202  * proc_dointvec_userhz_jiffies - read a vector of integers as 1/USER_HZ seconds
2203  * @table: the sysctl table
2204  * @write: %TRUE if this is a write to the sysctl file
2205  * @filp: the file structure
2206  * @buffer: the user buffer
2207  * @lenp: the size of the user buffer
2208  * @ppos: pointer to the file position
2209  *
2210  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2211  * values from/to the user buffer, treated as an ASCII string.
2212  * The values read are assumed to be in 1/USER_HZ seconds, and
2213  * are converted into jiffies.
2214  *
2215  * Returns 0 on success.
2216  */
2217 int proc_dointvec_userhz_jiffies(ctl_table *table, int write, struct file *filp,
2218                                  void __user *buffer, size_t *lenp, loff_t *ppos)
2219 {
2220     return do_proc_dointvec(table,write,filp,buffer,lenp,ppos,
2221                             do_proc_dointvec_userhz_jiffies_conv,NULL);
2222 }
2223
2224 /**
2225  * proc_dointvec_ms_jiffies - read a vector of integers as 1 milliseconds
2226  * @table: the sysctl table
2227  * @write: %TRUE if this is a write to the sysctl file
2228  * @filp: the file structure
2229  * @buffer: the user buffer
2230  * @lenp: the size of the user buffer
2231  * @ppos: file position
2232  * @ppos: the current position in the file
2233  *
2234  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2235  * values from/to the user buffer, treated as an ASCII string.
2236  * The values read are assumed to be in 1/1000 seconds, and
2237  * are converted into jiffies.
2238  *
2239  * Returns 0 on success.
2240  */
2241 int proc_dointvec_ms_jiffies(ctl_table *table, int write, struct file *filp,
2242                              void __user *buffer, size_t *lenp, loff_t *ppos)
2243 {
2244         return do_proc_dointvec(table, write, filp, buffer, lenp, ppos,
2245                                 do_proc_dointvec_ms_jiffies_conv, NULL);
2246 }
2247
2248 static int proc_do_cad_pid(ctl_table *table, int write, struct file *filp,
2249                            void __user *buffer, size_t *lenp, loff_t *ppos)
2250 {
2251         struct pid *new_pid;
2252         pid_t tmp;
2253         int r;
2254
2255         tmp = pid_nr(cad_pid);
2256
2257         r = __do_proc_dointvec(&tmp, table, write, filp, buffer,
2258                                lenp, ppos, NULL, NULL);
2259         if (r || !write)
2260                 return r;
2261
2262         new_pid = find_get_pid(tmp);
2263         if (!new_pid)
2264                 return -ESRCH;
2265
2266         put_pid(xchg(&cad_pid, new_pid));
2267         return 0;
2268 }
2269
2270 #else /* CONFIG_PROC_FS */
2271
2272 int proc_dostring(ctl_table *table, int write, struct file *filp,
2273                   void __user *buffer, size_t *lenp, loff_t *ppos)
2274 {
2275         return -ENOSYS;
2276 }
2277
2278 int proc_dointvec(ctl_table *table, int write, struct file *filp,
2279                   void __user *buffer, size_t *lenp, loff_t *ppos)
2280 {
2281         return -ENOSYS;
2282 }
2283
2284 int proc_dointvec_bset(ctl_table *table, int write, struct file *filp,
2285                         void __user *buffer, size_t *lenp, loff_t *ppos)
2286 {
2287         return -ENOSYS;
2288 }
2289
2290 int proc_dointvec_minmax(ctl_table *table, int write, struct file *filp,
2291                     void __user *buffer, size_t *lenp, loff_t *ppos)
2292 {
2293         return -ENOSYS;
2294 }
2295
2296 int proc_dointvec_jiffies(ctl_table *table, int write, struct file *filp,
2297                     void __user *buffer, size_t *lenp, loff_t *ppos)
2298 {
2299         return -ENOSYS;
2300 }
2301
2302 int proc_dointvec_userhz_jiffies(ctl_table *table, int write, struct file *filp,
2303                     void __user *buffer, size_t *lenp, loff_t *ppos)
2304 {
2305         return -ENOSYS;
2306 }
2307
2308 int proc_dointvec_ms_jiffies(ctl_table *table, int write, struct file *filp,
2309                              void __user *buffer, size_t *lenp, loff_t *ppos)
2310 {
2311         return -ENOSYS;
2312 }
2313
2314 int proc_doulongvec_minmax(ctl_table *table, int write, struct file *filp,
2315                     void __user *buffer, size_t *lenp, loff_t *ppos)
2316 {
2317         return -ENOSYS;
2318 }
2319
2320 int proc_doulongvec_ms_jiffies_minmax(ctl_table *table, int write,
2321                                       struct file *filp,
2322                                       void __user *buffer,
2323                                       size_t *lenp, loff_t *ppos)
2324 {
2325     return -ENOSYS;
2326 }
2327
2328
2329 #endif /* CONFIG_PROC_FS */
2330
2331
2332 #ifdef CONFIG_SYSCTL_SYSCALL
2333 /*
2334  * General sysctl support routines
2335  */
2336
2337 /* The generic string strategy routine: */
2338 int sysctl_string(ctl_table *table, int __user *name, int nlen,
2339                   void __user *oldval, size_t __user *oldlenp,
2340                   void __user *newval, size_t newlen)
2341 {
2342         if (!table->data || !table->maxlen)
2343                 return -ENOTDIR;
2344
2345         if (oldval && oldlenp) {
2346                 size_t bufsize;
2347                 if (get_user(bufsize, oldlenp))
2348                         return -EFAULT;
2349                 if (bufsize) {
2350                         size_t len = strlen(table->data), copied;
2351
2352                         /* This shouldn't trigger for a well-formed sysctl */
2353                         if (len > table->maxlen)
2354                                 len = table->maxlen;
2355
2356                         /* Copy up to a max of bufsize-1 bytes of the string */
2357                         copied = (len >= bufsize) ? bufsize - 1 : len;
2358
2359                         if (copy_to_user(oldval, table->data, copied) ||
2360                             put_user(0, (char __user *)(oldval + copied)))
2361                                 return -EFAULT;
2362                         if (put_user(len, oldlenp))
2363                                 return -EFAULT;
2364                 }
2365         }
2366         if (newval && newlen) {
2367                 size_t len = newlen;
2368                 if (len > table->maxlen)
2369                         len = table->maxlen;
2370                 if(copy_from_user(table->data, newval, len))
2371                         return -EFAULT;
2372                 if (len == table->maxlen)
2373                         len--;
2374                 ((char *) table->data)[len] = 0;
2375         }
2376         return 1;
2377 }
2378
2379 /*
2380  * This function makes sure that all of the integers in the vector
2381  * are between the minimum and maximum values given in the arrays
2382  * table->extra1 and table->extra2, respectively.
2383  */
2384 int sysctl_intvec(ctl_table *table, int __user *name, int nlen,
2385                 void __user *oldval, size_t __user *oldlenp,
2386                 void __user *newval, size_t newlen)
2387 {
2388
2389         if (newval && newlen) {
2390                 int __user *vec = (int __user *) newval;
2391                 int *min = (int *) table->extra1;
2392                 int *max = (int *) table->extra2;
2393                 size_t length;
2394                 int i;
2395
2396                 if (newlen % sizeof(int) != 0)
2397                         return -EINVAL;
2398
2399                 if (!table->extra1 && !table->extra2)
2400                         return 0;
2401
2402                 if (newlen > table->maxlen)
2403                         newlen = table->maxlen;
2404                 length = newlen / sizeof(int);
2405
2406                 for (i = 0; i < length; i++) {
2407                         int value;
2408                         if (get_user(value, vec + i))
2409                                 return -EFAULT;
2410                         if (min && value < min[i])
2411                                 return -EINVAL;
2412                         if (max && value > max[i])
2413                                 return -EINVAL;
2414                 }
2415         }
2416         return 0;
2417 }
2418
2419 /* Strategy function to convert jiffies to seconds */
2420 int sysctl_jiffies(ctl_table *table, int __user *name, int nlen,
2421                 void __user *oldval, size_t __user *oldlenp,
2422                 void __user *newval, size_t newlen)
2423 {
2424         if (oldval && oldlenp) {
2425                 size_t olen;
2426
2427                 if (get_user(olen, oldlenp))
2428                         return -EFAULT;
2429                 if (olen) {
2430                         int val;
2431
2432                         if (olen < sizeof(int))
2433                                 return -EINVAL;
2434
2435                         val = *(int *)(table->data) / HZ;
2436                         if (put_user(val, (int __user *)oldval))
2437                                 return -EFAULT;
2438                         if (put_user(sizeof(int), oldlenp))
2439                                 return -EFAULT;
2440                 }
2441         }
2442         if (newval && newlen) {
2443                 int new;
2444                 if (newlen != sizeof(int))
2445                         return -EINVAL;
2446                 if (get_user(new, (int __user *)newval))
2447                         return -EFAULT;
2448                 *(int *)(table->data) = new*HZ;
2449         }
2450         return 1;
2451 }
2452
2453 /* Strategy function to convert jiffies to seconds */
2454 int sysctl_ms_jiffies(ctl_table *table, int __user *name, int nlen,
2455                 void __user *oldval, size_t __user *oldlenp,
2456                 void __user *newval, size_t newlen)
2457 {
2458         if (oldval && oldlenp) {
2459                 size_t olen;
2460
2461                 if (get_user(olen, oldlenp))
2462                         return -EFAULT;
2463                 if (olen) {
2464                         int val;
2465
2466                         if (olen < sizeof(int))
2467                                 return -EINVAL;
2468
2469                         val = jiffies_to_msecs(*(int *)(table->data));
2470                         if (put_user(val, (int __user *)oldval))
2471                                 return -EFAULT;
2472                         if (put_user(sizeof(int), oldlenp))
2473                                 return -EFAULT;
2474                 }
2475         }
2476         if (newval && newlen) {
2477                 int new;
2478                 if (newlen != sizeof(int))
2479                         return -EINVAL;
2480                 if (get_user(new, (int __user *)newval))
2481                         return -EFAULT;
2482                 *(int *)(table->data) = msecs_to_jiffies(new);
2483         }
2484         return 1;
2485 }
2486
2487
2488
2489 #else /* CONFIG_SYSCTL_SYSCALL */
2490
2491
2492 asmlinkage long sys_sysctl(struct __sysctl_args __user *args)
2493 {
2494         static int msg_count;
2495         struct __sysctl_args tmp;
2496         int name[CTL_MAXNAME];
2497         int i;
2498
2499         /* Read in the sysctl name for better debug message logging */
2500         if (copy_from_user(&tmp, args, sizeof(tmp)))
2501                 return -EFAULT;
2502         if (tmp.nlen <= 0 || tmp.nlen >= CTL_MAXNAME)
2503                 return -ENOTDIR;
2504         for (i = 0; i < tmp.nlen; i++)
2505                 if (get_user(name[i], tmp.name + i))
2506                         return -EFAULT;
2507
2508         /* Ignore accesses to kernel.version */
2509         if ((tmp.nlen == 2) && (name[0] == CTL_KERN) && (name[1] == KERN_VERSION))
2510                 goto out;
2511
2512         if (msg_count < 5) {
2513                 msg_count++;
2514                 printk(KERN_INFO
2515                         "warning: process `%s' used the removed sysctl "
2516                         "system call with ", current->comm);
2517                 for (i = 0; i < tmp.nlen; i++)
2518                         printk("%d.", name[i]);
2519                 printk("\n");
2520         }
2521 out:
2522         return -ENOSYS;
2523 }
2524
2525 int sysctl_string(ctl_table *table, int __user *name, int nlen,
2526                   void __user *oldval, size_t __user *oldlenp,
2527                   void __user *newval, size_t newlen)
2528 {
2529         return -ENOSYS;
2530 }
2531
2532 int sysctl_intvec(ctl_table *table, int __user *name, int nlen,
2533                 void __user *oldval, size_t __user *oldlenp,
2534                 void __user *newval, size_t newlen)
2535 {
2536         return -ENOSYS;
2537 }
2538
2539 int sysctl_jiffies(ctl_table *table, int __user *name, int nlen,
2540                 void __user *oldval, size_t __user *oldlenp,
2541                 void __user *newval, size_t newlen)
2542 {
2543         return -ENOSYS;
2544 }
2545
2546 int sysctl_ms_jiffies(ctl_table *table, int __user *name, int nlen,
2547                 void __user *oldval, size_t __user *oldlenp,
2548                 void __user *newval, size_t newlen)
2549 {
2550         return -ENOSYS;
2551 }
2552
2553 #endif /* CONFIG_SYSCTL_SYSCALL */
2554
2555 /*
2556  * No sense putting this after each symbol definition, twice,
2557  * exception granted :-)
2558  */
2559 EXPORT_SYMBOL(proc_dointvec);
2560 EXPORT_SYMBOL(proc_dointvec_jiffies);
2561 EXPORT_SYMBOL(proc_dointvec_minmax);
2562 EXPORT_SYMBOL(proc_dointvec_userhz_jiffies);
2563 EXPORT_SYMBOL(proc_dointvec_ms_jiffies);
2564 EXPORT_SYMBOL(proc_dostring);
2565 EXPORT_SYMBOL(proc_doulongvec_minmax);
2566 EXPORT_SYMBOL(proc_doulongvec_ms_jiffies_minmax);
2567 EXPORT_SYMBOL(register_sysctl_table);
2568 EXPORT_SYMBOL(sysctl_intvec);
2569 EXPORT_SYMBOL(sysctl_jiffies);
2570 EXPORT_SYMBOL(sysctl_ms_jiffies);
2571 EXPORT_SYMBOL(sysctl_string);
2572 EXPORT_SYMBOL(unregister_sysctl_table);