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         { .ctl_name = 0 }
 753 };
 754
 755 /* Constants for minimum and maximum testing in vm_table.
 756    We use these as one-element integer vectors. */
 757 static int zero;
 758 static int two = 2;
 759 static int one_hundred = 100;
 760
 761
 762 static ctl_table vm_table[] = {
 763         {
 764                 .ctl_name       = VM_OVERCOMMIT_MEMORY,
 765                 .procname       = "overcommit_memory",
 766                 .data           = &sysctl_overcommit_memory,
 767                 .maxlen         = sizeof(sysctl_overcommit_memory),
 768                 .mode           = 0644,
 769                 .proc_handler   = &proc_dointvec,
 770         },
 771         {
 772                 .ctl_name       = VM_PANIC_ON_OOM,
 773                 .procname       = "panic_on_oom",
 774                 .data           = &sysctl_panic_on_oom,
 775                 .maxlen         = sizeof(sysctl_panic_on_oom),
 776                 .mode           = 0644,
 777                 .proc_handler   = &proc_dointvec,
 778         },
 779         {
 780                 .ctl_name       = VM_OVERCOMMIT_RATIO,
 781                 .procname       = "overcommit_ratio",
 782                 .data           = &sysctl_overcommit_ratio,
 783                 .maxlen         = sizeof(sysctl_overcommit_ratio),
 784                 .mode           = 0644,
 785                 .proc_handler   = &proc_dointvec,
 786         },
 787         {
 788                 .ctl_name       = VM_PAGE_CLUSTER,
 789                 .procname       = "page-cluster",
 790                 .data           = &page_cluster,
 791                 .maxlen         = sizeof(int),
 792                 .mode           = 0644,
 793                 .proc_handler   = &proc_dointvec,
 794         },
 795         {
 796                 .ctl_name       = VM_DIRTY_BACKGROUND,
 797                 .procname       = "dirty_background_ratio",
 798                 .data           = &dirty_background_ratio,
 799                 .maxlen         = sizeof(dirty_background_ratio),
 800                 .mode           = 0644,
 801                 .proc_handler   = &proc_dointvec_minmax,
 802                 .strategy       = &sysctl_intvec,
 803                 .extra1         = &zero,
 804                 .extra2         = &one_hundred,
 805         },
 806         {
 807                 .ctl_name       = VM_DIRTY_RATIO,
 808                 .procname       = "dirty_ratio",
 809                 .data           = &vm_dirty_ratio,
 810                 .maxlen         = sizeof(vm_dirty_ratio),
 811                 .mode           = 0644,
 812                 .proc_handler   = &proc_dointvec_minmax,
 813                 .strategy       = &sysctl_intvec,
 814                 .extra1         = &zero,
 815                 .extra2         = &one_hundred,
 816         },
 817         {
 818                 .ctl_name       = VM_DIRTY_WB_CS,
 819                 .procname       = "dirty_writeback_centisecs",
 820                 .data           = &dirty_writeback_interval,
 821                 .maxlen         = sizeof(dirty_writeback_interval),
 822                 .mode           = 0644,
 823                 .proc_handler   = &dirty_writeback_centisecs_handler,
 824         },
 825         {
 826                 .ctl_name       = VM_DIRTY_EXPIRE_CS,
 827                 .procname       = "dirty_expire_centisecs",
 828                 .data           = &dirty_expire_interval,
 829                 .maxlen         = sizeof(dirty_expire_interval),
 830                 .mode           = 0644,
 831                 .proc_handler   = &proc_dointvec_userhz_jiffies,
 832         },
 833         {
 834                 .ctl_name       = VM_NR_PDFLUSH_THREADS,
 835                 .procname       = "nr_pdflush_threads",
 836                 .data           = &nr_pdflush_threads,
 837                 .maxlen         = sizeof nr_pdflush_threads,
 838                 .mode           = 0444 /* read-only*/,
 839                 .proc_handler   = &proc_dointvec,
 840         },
 841         {
 842                 .ctl_name       = VM_SWAPPINESS,
 843                 .procname       = "swappiness",
 844                 .data           = &vm_swappiness,
 845                 .maxlen         = sizeof(vm_swappiness),
 846                 .mode           = 0644,
 847                 .proc_handler   = &proc_dointvec_minmax,
 848                 .strategy       = &sysctl_intvec,
 849                 .extra1         = &zero,
 850                 .extra2         = &one_hundred,
 851         },
 852 #ifdef CONFIG_HUGETLB_PAGE
 853          {
 854                 .ctl_name       = VM_HUGETLB_PAGES,
 855                 .procname       = "nr_hugepages",
 856                 .data           = &max_huge_pages,
 857                 .maxlen         = sizeof(unsigned long),
 858                 .mode           = 0644,
 859                 .proc_handler   = &hugetlb_sysctl_handler,
 860                 .extra1         = (void *)&hugetlb_zero,
 861                 .extra2         = (void *)&hugetlb_infinity,
 862          },
 863          {
 864                 .ctl_name       = VM_HUGETLB_GROUP,
 865                 .procname       = "hugetlb_shm_group",
 866                 .data           = &sysctl_hugetlb_shm_group,
 867                 .maxlen         = sizeof(gid_t),
 868                 .mode           = 0644,
 869                 .proc_handler   = &proc_dointvec,
 870          },
 871          {
 872                 .ctl_name       = CTL_UNNUMBERED,
 873                 .procname       = "hugepages_treat_as_movable",
 874                 .data           = &hugepages_treat_as_movable,
 875                 .maxlen         = sizeof(int),
 876                 .mode           = 0644,
 877                 .proc_handler   = &hugetlb_treat_movable_handler,
 878         },
 879 #endif
 880         {
 881                 .ctl_name       = VM_LOWMEM_RESERVE_RATIO,
 882                 .procname       = "lowmem_reserve_ratio",
 883                 .data           = &sysctl_lowmem_reserve_ratio,
 884                 .maxlen         = sizeof(sysctl_lowmem_reserve_ratio),
 885                 .mode           = 0644,
 886                 .proc_handler   = &lowmem_reserve_ratio_sysctl_handler,
 887                 .strategy       = &sysctl_intvec,
 888         },
 889         {
 890                 .ctl_name       = VM_DROP_PAGECACHE,
 891                 .procname       = "drop_caches",
 892                 .data           = &sysctl_drop_caches,
 893                 .maxlen         = sizeof(int),
 894                 .mode           = 0644,
 895                 .proc_handler   = drop_caches_sysctl_handler,
 896                 .strategy       = &sysctl_intvec,
 897         },
 898         {
 899                 .ctl_name       = VM_MIN_FREE_KBYTES,
 900                 .procname       = "min_free_kbytes",
 901                 .data           = &min_free_kbytes,
 902                 .maxlen         = sizeof(min_free_kbytes),
 903                 .mode           = 0644,
 904                 .proc_handler   = &min_free_kbytes_sysctl_handler,
 905                 .strategy       = &sysctl_intvec,
 906                 .extra1         = &zero,
 907         },
 908         {
 909                 .ctl_name       = VM_PERCPU_PAGELIST_FRACTION,
 910                 .procname       = "percpu_pagelist_fraction",
 911                 .data           = &percpu_pagelist_fraction,
 912                 .maxlen         = sizeof(percpu_pagelist_fraction),
 913                 .mode           = 0644,
 914                 .proc_handler   = &percpu_pagelist_fraction_sysctl_handler,
 915                 .strategy       = &sysctl_intvec,
 916                 .extra1         = &min_percpu_pagelist_fract,
 917         },
 918 #ifdef CONFIG_MMU
 919         {
 920                 .ctl_name       = VM_MAX_MAP_COUNT,
 921                 .procname       = "max_map_count",
 922                 .data           = &sysctl_max_map_count,
 923                 .maxlen         = sizeof(sysctl_max_map_count),
 924                 .mode           = 0644,
 925                 .proc_handler   = &proc_dointvec
 926         },
 927 #endif
 928         {
 929                 .ctl_name       = VM_LAPTOP_MODE,
 930                 .procname       = "laptop_mode",
 931                 .data           = &laptop_mode,
 932                 .maxlen         = sizeof(laptop_mode),
 933                 .mode           = 0644,
 934                 .proc_handler   = &proc_dointvec_jiffies,
 935                 .strategy       = &sysctl_jiffies,
 936         },
 937         {
 938                 .ctl_name       = VM_BLOCK_DUMP,
 939                 .procname       = "block_dump",
 940                 .data           = &block_dump,
 941                 .maxlen         = sizeof(block_dump),
 942                 .mode           = 0644,
 943                 .proc_handler   = &proc_dointvec,
 944                 .strategy       = &sysctl_intvec,
 945                 .extra1         = &zero,
 946         },
 947         {
 948                 .ctl_name       = VM_VFS_CACHE_PRESSURE,
 949                 .procname       = "vfs_cache_pressure",
 950                 .data           = &sysctl_vfs_cache_pressure,
 951                 .maxlen         = sizeof(sysctl_vfs_cache_pressure),
 952                 .mode           = 0644,
 953                 .proc_handler   = &proc_dointvec,
 954                 .strategy       = &sysctl_intvec,
 955                 .extra1         = &zero,
 956         },
 957 #ifdef HAVE_ARCH_PICK_MMAP_LAYOUT
 958         {
 959                 .ctl_name       = VM_LEGACY_VA_LAYOUT,
 960                 .procname       = "legacy_va_layout",
 961                 .data           = &sysctl_legacy_va_layout,
 962                 .maxlen         = sizeof(sysctl_legacy_va_layout),
 963                 .mode           = 0644,
 964                 .proc_handler   = &proc_dointvec,
 965                 .strategy       = &sysctl_intvec,
 966                 .extra1         = &zero,
 967         },
 968 #endif
 969 #ifdef CONFIG_NUMA
 970         {
 971                 .ctl_name       = VM_ZONE_RECLAIM_MODE,
 972                 .procname       = "zone_reclaim_mode",
 973                 .data           = &zone_reclaim_mode,
 974                 .maxlen         = sizeof(zone_reclaim_mode),
 975                 .mode           = 0644,
 976                 .proc_handler   = &proc_dointvec,
 977                 .strategy       = &sysctl_intvec,
 978                 .extra1         = &zero,
 979         },
 980         {
 981                 .ctl_name       = VM_MIN_UNMAPPED,
 982                 .procname       = "min_unmapped_ratio",
 983                 .data           = &sysctl_min_unmapped_ratio,
 984                 .maxlen         = sizeof(sysctl_min_unmapped_ratio),
 985                 .mode           = 0644,
 986                 .proc_handler   = &sysctl_min_unmapped_ratio_sysctl_handler,
 987                 .strategy       = &sysctl_intvec,
 988                 .extra1         = &zero,
 989                 .extra2         = &one_hundred,
 990         },
 991         {
 992                 .ctl_name       = VM_MIN_SLAB,
 993                 .procname       = "min_slab_ratio",
 994                 .data           = &sysctl_min_slab_ratio,
 995                 .maxlen         = sizeof(sysctl_min_slab_ratio),
 996                 .mode           = 0644,
 997                 .proc_handler   = &sysctl_min_slab_ratio_sysctl_handler,
 998                 .strategy       = &sysctl_intvec,
 999                 .extra1         = &zero,
1000                 .extra2         = &one_hundred,
1001         },
1002 #endif
1003 #ifdef CONFIG_SMP
1004         {
1005                 .ctl_name       = CTL_UNNUMBERED,
1006                 .procname       = "stat_interval",
1007                 .data           = &sysctl_stat_interval,
1008                 .maxlen         = sizeof(sysctl_stat_interval),
1009                 .mode           = 0644,
1010                 .proc_handler   = &proc_dointvec_jiffies,
1011                 .strategy       = &sysctl_jiffies,
1012         },
1013 #endif
1014 #ifdef CONFIG_SECURITY
1015         {
1016                 .ctl_name       = CTL_UNNUMBERED,
1017                 .procname       = "mmap_min_addr",
1018                 .data           = &mmap_min_addr,
1019                 .maxlen         = sizeof(unsigned long),
1020                 .mode           = 0644,
1021                 .proc_handler   = &proc_doulongvec_minmax,
1022         },
1023 #ifdef CONFIG_NUMA
1024         {
1025                 .ctl_name       = CTL_UNNUMBERED,
1026                 .procname       = "numa_zonelist_order",
1027                 .data           = &numa_zonelist_order,
1028                 .maxlen         = NUMA_ZONELIST_ORDER_LEN,
1029                 .mode           = 0644,
1030                 .proc_handler   = &numa_zonelist_order_handler,
1031                 .strategy       = &sysctl_string,
1032         },
1033 #endif
1034 #endif
1035 #if defined(CONFIG_X86_32) || \
1036    (defined(CONFIG_SUPERH) && defined(CONFIG_VSYSCALL))
1037         {
1038                 .ctl_name       = VM_VDSO_ENABLED,
1039                 .procname       = "vdso_enabled",
1040                 .data           = &vdso_enabled,
1041                 .maxlen         = sizeof(vdso_enabled),
1042                 .mode           = 0644,
1043                 .proc_handler   = &proc_dointvec,
1044                 .strategy       = &sysctl_intvec,
1045                 .extra1         = &zero,
1046         },
1047 #endif
1048 /*
1049  * NOTE: do not add new entries to this table unless you have read
1050  * Documentation/sysctl/ctl_unnumbered.txt
1051  */
1052         { .ctl_name = 0 }
1053 };
1054
1055 #if defined(CONFIG_BINFMT_MISC) || defined(CONFIG_BINFMT_MISC_MODULE)
1056 static ctl_table binfmt_misc_table[] = {
1057         { .ctl_name = 0 }
1058 };
1059 #endif
1060
1061 static ctl_table fs_table[] = {
1062         {
1063                 .ctl_name       = FS_NRINODE,
1064                 .procname       = "inode-nr",
1065                 .data           = &inodes_stat,
1066                 .maxlen         = 2*sizeof(int),
1067                 .mode           = 0444,
1068                 .proc_handler   = &proc_dointvec,
1069         },
1070         {
1071                 .ctl_name       = FS_STATINODE,
1072                 .procname       = "inode-state",
1073                 .data           = &inodes_stat,
1074                 .maxlen         = 7*sizeof(int),
1075                 .mode           = 0444,
1076                 .proc_handler   = &proc_dointvec,
1077         },
1078         {
1079                 .ctl_name       = FS_NRFILE,
1080                 .procname       = "file-nr",
1081                 .data           = &files_stat,
1082                 .maxlen         = 3*sizeof(int),
1083                 .mode           = 0444,
1084                 .proc_handler   = &proc_nr_files,
1085         },
1086         {
1087                 .ctl_name       = FS_MAXFILE,
1088                 .procname       = "file-max",
1089                 .data           = &files_stat.max_files,
1090                 .maxlen         = sizeof(int),
1091                 .mode           = 0644,
1092                 .proc_handler   = &proc_dointvec,
1093         },
1094         {
1095                 .ctl_name       = FS_DENTRY,
1096                 .procname       = "dentry-state",
1097                 .data           = &dentry_stat,
1098                 .maxlen         = 6*sizeof(int),
1099                 .mode           = 0444,
1100                 .proc_handler   = &proc_dointvec,
1101         },
1102         {
1103                 .ctl_name       = FS_OVERFLOWUID,
1104                 .procname       = "overflowuid",
1105                 .data           = &fs_overflowuid,
1106                 .maxlen         = sizeof(int),
1107                 .mode           = 0644,
1108                 .proc_handler   = &proc_dointvec_minmax,
1109                 .strategy       = &sysctl_intvec,
1110                 .extra1         = &minolduid,
1111                 .extra2         = &maxolduid,
1112         },
1113         {
1114                 .ctl_name       = FS_OVERFLOWGID,
1115                 .procname       = "overflowgid",
1116                 .data           = &fs_overflowgid,
1117                 .maxlen         = sizeof(int),
1118                 .mode           = 0644,
1119                 .proc_handler   = &proc_dointvec_minmax,
1120                 .strategy       = &sysctl_intvec,
1121                 .extra1         = &minolduid,
1122                 .extra2         = &maxolduid,
1123         },
1124         {
1125                 .ctl_name       = FS_LEASES,
1126                 .procname       = "leases-enable",
1127                 .data           = &leases_enable,
1128                 .maxlen         = sizeof(int),
1129                 .mode           = 0644,
1130                 .proc_handler   = &proc_dointvec,
1131         },
1132 #ifdef CONFIG_DNOTIFY
1133         {
1134                 .ctl_name       = FS_DIR_NOTIFY,
1135                 .procname       = "dir-notify-enable",
1136                 .data           = &dir_notify_enable,
1137                 .maxlen         = sizeof(int),
1138                 .mode           = 0644,
1139                 .proc_handler   = &proc_dointvec,
1140         },
1141 #endif
1142 #ifdef CONFIG_MMU
1143         {
1144                 .ctl_name       = FS_LEASE_TIME,
1145                 .procname       = "lease-break-time",
1146                 .data           = &lease_break_time,
1147                 .maxlen         = sizeof(int),
1148                 .mode           = 0644,
1149                 .proc_handler   = &proc_dointvec_minmax,
1150                 .strategy       = &sysctl_intvec,
1151                 .extra1         = &zero,
1152                 .extra2         = &two,
1153         },
1154         {
1155                 .ctl_name       = FS_AIO_NR,
1156                 .procname       = "aio-nr",
1157                 .data           = &aio_nr,
1158                 .maxlen         = sizeof(aio_nr),
1159                 .mode           = 0444,
1160                 .proc_handler   = &proc_doulongvec_minmax,
1161         },
1162         {
1163                 .ctl_name       = FS_AIO_MAX_NR,
1164                 .procname       = "aio-max-nr",
1165                 .data           = &aio_max_nr,
1166                 .maxlen         = sizeof(aio_max_nr),
1167                 .mode           = 0644,
1168                 .proc_handler   = &proc_doulongvec_minmax,
1169         },
1170 #ifdef CONFIG_INOTIFY_USER
1171         {
1172                 .ctl_name       = FS_INOTIFY,
1173                 .procname       = "inotify",
1174                 .mode           = 0555,
1175                 .child          = inotify_table,
1176         },
1177 #endif
1178 #endif
1179         {
1180                 .ctl_name       = KERN_SETUID_DUMPABLE,
1181                 .procname       = "suid_dumpable",
1182                 .data           = &suid_dumpable,
1183                 .maxlen         = sizeof(int),
1184                 .mode           = 0644,
1185                 .proc_handler   = &proc_dointvec,
1186         },
1187 #if defined(CONFIG_BINFMT_MISC) || defined(CONFIG_BINFMT_MISC_MODULE)
1188         {
1189                 .ctl_name       = CTL_UNNUMBERED,
1190                 .procname       = "binfmt_misc",
1191                 .mode           = 0555,
1192                 .child          = binfmt_misc_table,
1193         },
1194 #endif
1195 /*
1196  * NOTE: do not add new entries to this table unless you have read
1197  * Documentation/sysctl/ctl_unnumbered.txt
1198  */
1199         { .ctl_name = 0 }
1200 };
1201
1202 static ctl_table debug_table[] = {
1203         { .ctl_name = 0 }
1204 };
1205
1206 static ctl_table dev_table[] = {
1207         { .ctl_name = 0 }
1208 };
1209
1210 static DEFINE_SPINLOCK(sysctl_lock);
1211
1212 /* called under sysctl_lock */
1213 static int use_table(struct ctl_table_header *p)
1214 {
1215         if (unlikely(p->unregistering))
1216                 return 0;
1217         p->used++;
1218         return 1;
1219 }
1220
1221 /* called under sysctl_lock */
1222 static void unuse_table(struct ctl_table_header *p)
1223 {
1224         if (!--p->used)
1225                 if (unlikely(p->unregistering))
1226                         complete(p->unregistering);
1227 }
1228
1229 /* called under sysctl_lock, will reacquire if has to wait */
1230 static void start_unregistering(struct ctl_table_header *p)
1231 {
1232         /*
1233          * if p->used is 0, nobody will ever touch that entry again;
1234          * we'll eliminate all paths to it before dropping sysctl_lock
1235          */
1236         if (unlikely(p->used)) {
1237                 struct completion wait;
1238                 init_completion(&wait);
1239                 p->unregistering = &wait;
1240                 spin_unlock(&sysctl_lock);
1241                 wait_for_completion(&wait);
1242                 spin_lock(&sysctl_lock);
1243         }
1244         /*
1245          * do not remove from the list until nobody holds it; walking the
1246          * list in do_sysctl() relies on that.
1247          */
1248         list_del_init(&p->ctl_entry);
1249 }
1250
1251 void sysctl_head_finish(struct ctl_table_header *head)
1252 {
1253         if (!head)
1254                 return;
1255         spin_lock(&sysctl_lock);
1256         unuse_table(head);
1257         spin_unlock(&sysctl_lock);
1258 }
1259
1260 struct ctl_table_header *sysctl_head_next(struct ctl_table_header *prev)
1261 {
1262         struct ctl_table_header *head;
1263         struct list_head *tmp;
1264         spin_lock(&sysctl_lock);
1265         if (prev) {
1266                 tmp = &prev->ctl_entry;
1267                 unuse_table(prev);
1268                 goto next;
1269         }
1270         tmp = &root_table_header.ctl_entry;
1271         for (;;) {
1272                 head = list_entry(tmp, struct ctl_table_header, ctl_entry);
1273
1274                 if (!use_table(head))
1275                         goto next;
1276                 spin_unlock(&sysctl_lock);
1277                 return head;
1278         next:
1279                 tmp = tmp->next;
1280                 if (tmp == &root_table_header.ctl_entry)
1281                         break;
1282         }
1283         spin_unlock(&sysctl_lock);
1284         return NULL;
1285 }
1286
1287 #ifdef CONFIG_SYSCTL_SYSCALL
1288 int do_sysctl(int __user *name, int nlen, void __user *oldval, size_t __user *oldlenp,
1289                void __user *newval, size_t newlen)
1290 {
1291         struct ctl_table_header *head;
1292         int error = -ENOTDIR;
1293
1294         if (nlen <= 0 || nlen >= CTL_MAXNAME)
1295                 return -ENOTDIR;
1296         if (oldval) {
1297                 int old_len;
1298                 if (!oldlenp || get_user(old_len, oldlenp))
1299                         return -EFAULT;
1300         }
1301
1302         for (head = sysctl_head_next(NULL); head;
1303                         head = sysctl_head_next(head)) {
1304                 error = parse_table(name, nlen, oldval, oldlenp,
1305                                         newval, newlen, head->ctl_table);
1306                 if (error != -ENOTDIR) {
1307                         sysctl_head_finish(head);
1308                         break;
1309                 }
1310         }
1311         return error;
1312 }
1313
1314 asmlinkage long sys_sysctl(struct __sysctl_args __user *args)
1315 {
1316         struct __sysctl_args tmp;
1317         int error;
1318
1319         if (copy_from_user(&tmp, args, sizeof(tmp)))
1320                 return -EFAULT;
1321
1322         lock_kernel();
1323         error = do_sysctl(tmp.name, tmp.nlen, tmp.oldval, tmp.oldlenp,
1324                           tmp.newval, tmp.newlen);
1325         unlock_kernel();
1326         return error;
1327 }
1328 #endif /* CONFIG_SYSCTL_SYSCALL */
1329
1330 /*
1331  * sysctl_perm does NOT grant the superuser all rights automatically, because
1332  * some sysctl variables are readonly even to root.
1333  */
1334
1335 static int test_perm(int mode, int op)
1336 {
1337         if (!current->euid)
1338                 mode >>= 6;
1339         else if (in_egroup_p(0))
1340                 mode >>= 3;
1341         if ((mode & op & 0007) == op)
1342                 return 0;
1343         return -EACCES;
1344 }
1345
1346 int sysctl_perm(ctl_table *table, int op)
1347 {
1348         int error;
1349         error = security_sysctl(table, op);
1350         if (error)
1351                 return error;
1352         return test_perm(table->mode, op);
1353 }
1354
1355 #ifdef CONFIG_SYSCTL_SYSCALL
1356 static int parse_table(int __user *name, int nlen,
1357                        void __user *oldval, size_t __user *oldlenp,
1358                        void __user *newval, size_t newlen,
1359                        ctl_table *table)
1360 {
1361         int n;
1362 repeat:
1363         if (!nlen)
1364                 return -ENOTDIR;
1365         if (get_user(n, name))
1366                 return -EFAULT;
1367         for ( ; table->ctl_name || table->procname; table++) {
1368                 if (!table->ctl_name)
1369                         continue;
1370                 if (n == table->ctl_name) {
1371                         int error;
1372                         if (table->child) {
1373                                 if (sysctl_perm(table, 001))
1374                                         return -EPERM;
1375                                 name++;
1376                                 nlen--;
1377                                 table = table->child;
1378                                 goto repeat;
1379                         }
1380                         error = do_sysctl_strategy(table, name, nlen,
1381                                                    oldval, oldlenp,
1382                                                    newval, newlen);
1383                         return error;
1384                 }
1385         }
1386         return -ENOTDIR;
1387 }
1388
1389 /* Perform the actual read/write of a sysctl table entry. */
1390 int do_sysctl_strategy (ctl_table *table,
1391                         int __user *name, int nlen,
1392                         void __user *oldval, size_t __user *oldlenp,
1393                         void __user *newval, size_t newlen)
1394 {
1395         int op = 0, rc;
1396         size_t len;
1397
1398         if (oldval)
1399                 op |= 004;
1400         if (newval)
1401                 op |= 002;
1402         if (sysctl_perm(table, op))
1403                 return -EPERM;
1404
1405         if (table->strategy) {
1406                 rc = table->strategy(table, name, nlen, oldval, oldlenp,
1407                                      newval, newlen);
1408                 if (rc < 0)
1409                         return rc;
1410                 if (rc > 0)
1411                         return 0;
1412         }
1413
1414         /* If there is no strategy routine, or if the strategy returns
1415          * zero, proceed with automatic r/w */
1416         if (table->data && table->maxlen) {
1417                 if (oldval && oldlenp) {
1418                         if (get_user(len, oldlenp))
1419                                 return -EFAULT;
1420                         if (len) {
1421                                 if (len > table->maxlen)
1422                                         len = table->maxlen;
1423                                 if(copy_to_user(oldval, table->data, len))
1424                                         return -EFAULT;
1425                                 if(put_user(len, oldlenp))
1426                                         return -EFAULT;
1427                         }
1428                 }
1429                 if (newval && newlen) {
1430                         len = newlen;
1431                         if (len > table->maxlen)
1432                                 len = table->maxlen;
1433                         if(copy_from_user(table->data, newval, len))
1434                                 return -EFAULT;
1435                 }
1436         }
1437         return 0;
1438 }
1439 #endif /* CONFIG_SYSCTL_SYSCALL */
1440
1441 static void sysctl_set_parent(struct ctl_table *parent, struct ctl_table *table)
1442 {
1443         for (; table->ctl_name || table->procname; table++) {
1444                 table->parent = parent;
1445                 if (table->child)
1446                         sysctl_set_parent(table, table->child);
1447         }
1448 }
1449
1450 static __init int sysctl_init(void)
1451 {
1452         sysctl_set_parent(NULL, root_table);
1453         return 0;
1454 }
1455
1456 core_initcall(sysctl_init);
1457
1458 /**
1459  * register_sysctl_table - register a sysctl hierarchy
1460  * @table: the top-level table structure
1461  *
1462  * Register a sysctl table hierarchy. @table should be a filled in ctl_table
1463  * array. An entry with a ctl_name of 0 terminates the table.
1464  *
1465  * The members of the &ctl_table structure are used as follows:
1466  *
1467  * ctl_name - This is the numeric sysctl value used by sysctl(2). The number
1468  *            must be unique within that level of sysctl
1469  *
1470  * procname - the name of the sysctl file under /proc/sys. Set to %NULL to not
1471  *            enter a sysctl file
1472  *
1473  * data - a pointer to data for use by proc_handler
1474  *
1475  * maxlen - the maximum size in bytes of the data
1476  *
1477  * mode - the file permissions for the /proc/sys file, and for sysctl(2)
1478  *
1479  * child - a pointer to the child sysctl table if this entry is a directory, or
1480  *         %NULL.
1481  *
1482  * proc_handler - the text handler routine (described below)
1483  *
1484  * strategy - the strategy routine (described below)
1485  *
1486  * de - for internal use by the sysctl routines
1487  *
1488  * extra1, extra2 - extra pointers usable by the proc handler routines
1489  *
1490  * Leaf nodes in the sysctl tree will be represented by a single file
1491  * under /proc; non-leaf nodes will be represented by directories.
1492  *
1493  * sysctl(2) can automatically manage read and write requests through
1494  * the sysctl table.  The data and maxlen fields of the ctl_table
1495  * struct enable minimal validation of the values being written to be
1496  * performed, and the mode field allows minimal authentication.
1497  *
1498  * More sophisticated management can be enabled by the provision of a
1499  * strategy routine with the table entry.  This will be called before
1500  * any automatic read or write of the data is performed.
1501  *
1502  * The strategy routine may return
1503  *
1504  * < 0 - Error occurred (error is passed to user process)
1505  *
1506  * 0   - OK - proceed with automatic read or write.
1507  *
1508  * > 0 - OK - read or write has been done by the strategy routine, so
1509  *       return immediately.
1510  *
1511  * There must be a proc_handler routine for any terminal nodes
1512  * mirrored under /proc/sys (non-terminals are handled by a built-in
1513  * directory handler).  Several default handlers are available to
1514  * cover common cases -
1515  *
1516  * proc_dostring(), proc_dointvec(), proc_dointvec_jiffies(),
1517  * proc_dointvec_userhz_jiffies(), proc_dointvec_minmax(),
1518  * proc_doulongvec_ms_jiffies_minmax(), proc_doulongvec_minmax()
1519  *
1520  * It is the handler's job to read the input buffer from user memory
1521  * and process it. The handler should return 0 on success.
1522  *
1523  * This routine returns %NULL on a failure to register, and a pointer
1524  * to the table header on success.
1525  */
1526 struct ctl_table_header *register_sysctl_table(ctl_table * table)
1527 {
1528         struct ctl_table_header *tmp;
1529         tmp = kmalloc(sizeof(struct ctl_table_header), GFP_KERNEL);
1530         if (!tmp)
1531                 return NULL;
1532         tmp->ctl_table = table;
1533         INIT_LIST_HEAD(&tmp->ctl_entry);
1534         tmp->used = 0;
1535         tmp->unregistering = NULL;
1536         sysctl_set_parent(NULL, table);
1537         spin_lock(&sysctl_lock);
1538         list_add_tail(&tmp->ctl_entry, &root_table_header.ctl_entry);
1539         spin_unlock(&sysctl_lock);
1540         return tmp;
1541 }
1542
1543 /**
1544  * unregister_sysctl_table - unregister a sysctl table hierarchy
1545  * @header: the header returned from register_sysctl_table
1546  *
1547  * Unregisters the sysctl table and all children. proc entries may not
1548  * actually be removed until they are no longer used by anyone.
1549  */
1550 void unregister_sysctl_table(struct ctl_table_header * header)
1551 {
1552         might_sleep();
1553         spin_lock(&sysctl_lock);
1554         start_unregistering(header);
1555         spin_unlock(&sysctl_lock);
1556         kfree(header);
1557 }
1558
1559 #else /* !CONFIG_SYSCTL */
1560 struct ctl_table_header *register_sysctl_table(ctl_table * table)
1561 {
1562         return NULL;
1563 }
1564
1565 void unregister_sysctl_table(struct ctl_table_header * table)
1566 {
1567 }
1568
1569 #endif /* CONFIG_SYSCTL */
1570
1571 /*
1572  * /proc/sys support
1573  */
1574
1575 #ifdef CONFIG_PROC_SYSCTL
1576
1577 static int _proc_do_string(void* data, int maxlen, int write,
1578                            struct file *filp, void __user *buffer,
1579                            size_t *lenp, loff_t *ppos)
1580 {
1581         size_t len;
1582         char __user *p;
1583         char c;
1584
1585         if (!data || !maxlen || !*lenp) {
1586                 *lenp = 0;
1587                 return 0;
1588         }
1589
1590         if (write) {
1591                 len = 0;
1592                 p = buffer;
1593                 while (len < *lenp) {
1594                         if (get_user(c, p++))
1595                                 return -EFAULT;
1596                         if (c == 0 || c == '\n')
1597                                 break;
1598                         len++;
1599                 }
1600                 if (len >= maxlen)
1601                         len = maxlen-1;
1602                 if(copy_from_user(data, buffer, len))
1603                         return -EFAULT;
1604                 ((char *) data)[len] = 0;
1605                 *ppos += *lenp;
1606         } else {
1607                 len = strlen(data);
1608                 if (len > maxlen)
1609                         len = maxlen;
1610
1611                 if (*ppos > len) {
1612                         *lenp = 0;
1613                         return 0;
1614                 }
1615
1616                 data += *ppos;
1617                 len  -= *ppos;
1618
1619                 if (len > *lenp)
1620                         len = *lenp;
1621                 if (len)
1622                         if(copy_to_user(buffer, data, len))
1623                                 return -EFAULT;
1624                 if (len < *lenp) {
1625                         if(put_user('\n', ((char __user *) buffer) + len))
1626                                 return -EFAULT;
1627                         len++;
1628                 }
1629                 *lenp = len;
1630                 *ppos += len;
1631         }
1632         return 0;
1633 }
1634
1635 /**
1636  * proc_dostring - read a string sysctl
1637  * @table: the sysctl table
1638  * @write: %TRUE if this is a write to the sysctl file
1639  * @filp: the file structure
1640  * @buffer: the user buffer
1641  * @lenp: the size of the user buffer
1642  * @ppos: file position
1643  *
1644  * Reads/writes a string from/to the user buffer. If the kernel
1645  * buffer provided is not large enough to hold the string, the
1646  * string is truncated. The copied string is %NULL-terminated.
1647  * If the string is being read by the user process, it is copied
1648  * and a newline '\n' is added. It is truncated if the buffer is
1649  * not large enough.
1650  *
1651  * Returns 0 on success.
1652  */
1653 int proc_dostring(ctl_table *table, int write, struct file *filp,
1654                   void __user *buffer, size_t *lenp, loff_t *ppos)
1655 {
1656         return _proc_do_string(table->data, table->maxlen, write, filp,
1657                                buffer, lenp, ppos);
1658 }
1659
1660
1661 static int do_proc_dointvec_conv(int *negp, unsigned long *lvalp,
1662                                  int *valp,
1663                                  int write, void *data)
1664 {
1665         if (write) {
1666                 *valp = *negp ? -*lvalp : *lvalp;
1667         } else {
1668                 int val = *valp;
1669                 if (val < 0) {
1670                         *negp = -1;
1671                         *lvalp = (unsigned long)-val;
1672                 } else {
1673                         *negp = 0;
1674                         *lvalp = (unsigned long)val;
1675                 }
1676         }
1677         return 0;
1678 }
1679
1680 static int __do_proc_dointvec(void *tbl_data, ctl_table *table,
1681                   int write, struct file *filp, void __user *buffer,
1682                   size_t *lenp, loff_t *ppos,
1683                   int (*conv)(int *negp, unsigned long *lvalp, int *valp,
1684                               int write, void *data),
1685                   void *data)
1686 {
1687 #define TMPBUFLEN 21
1688         int *i, vleft, first=1, neg, val;
1689         unsigned long lval;
1690         size_t left, len;
1691
1692         char buf[TMPBUFLEN], *p;
1693         char __user *s = buffer;
1694
1695         if (!tbl_data || !table->maxlen || !*lenp ||
1696             (*ppos && !write)) {
1697                 *lenp = 0;
1698                 return 0;
1699         }
1700
1701         i = (int *) tbl_data;
1702         vleft = table->maxlen / sizeof(*i);
1703         left = *lenp;
1704
1705         if (!conv)
1706                 conv = do_proc_dointvec_conv;
1707
1708         for (; left && vleft--; i++, first=0) {
1709                 if (write) {
1710                         while (left) {
1711                                 char c;
1712                                 if (get_user(c, s))
1713                                         return -EFAULT;
1714                                 if (!isspace(c))
1715                                         break;
1716                                 left--;
1717                                 s++;
1718                         }
1719                         if (!left)
1720                                 break;
1721                         neg = 0;
1722                         len = left;
1723                         if (len > sizeof(buf) - 1)
1724                                 len = sizeof(buf) - 1;
1725                         if (copy_from_user(buf, s, len))
1726                                 return -EFAULT;
1727                         buf[len] = 0;
1728                         p = buf;
1729                         if (*p == '-' && left > 1) {
1730                                 neg = 1;
1731                                 p++;
1732                         }
1733                         if (*p < '0' || *p > '9')
1734                                 break;
1735
1736                         lval = simple_strtoul(p, &p, 0);
1737
1738                         len = p-buf;
1739                         if ((len < left) && *p && !isspace(*p))
1740                                 break;
1741                         if (neg)
1742                                 val = -val;
1743                         s += len;
1744                         left -= len;
1745
1746                         if (conv(&neg, &lval, i, 1, data))
1747                                 break;
1748                 } else {
1749                         p = buf;
1750                         if (!first)
1751                                 *p++ = '\t';
1752
1753                         if (conv(&neg, &lval, i, 0, data))
1754                                 break;
1755
1756                         sprintf(p, "%s%lu", neg ? "-" : "", lval);
1757                         len = strlen(buf);
1758                         if (len > left)
1759                                 len = left;
1760                         if(copy_to_user(s, buf, len))
1761                                 return -EFAULT;
1762                         left -= len;
1763                         s += len;
1764                 }
1765         }
1766
1767         if (!write && !first && left) {
1768                 if(put_user('\n', s))
1769                         return -EFAULT;
1770                 left--, s++;
1771         }
1772         if (write) {
1773                 while (left) {
1774                         char c;
1775                         if (get_user(c, s++))
1776                                 return -EFAULT;
1777                         if (!isspace(c))
1778                                 break;
1779                         left--;
1780                 }
1781         }
1782         if (write && first)
1783                 return -EINVAL;
1784         *lenp -= left;
1785         *ppos += *lenp;
1786         return 0;
1787 #undef TMPBUFLEN
1788 }
1789
1790 static int do_proc_dointvec(ctl_table *table, int write, struct file *filp,
1791                   void __user *buffer, size_t *lenp, loff_t *ppos,
1792                   int (*conv)(int *negp, unsigned long *lvalp, int *valp,
1793                               int write, void *data),
1794                   void *data)
1795 {
1796         return __do_proc_dointvec(table->data, table, write, filp,
1797                         buffer, lenp, ppos, conv, data);
1798 }
1799
1800 /**
1801  * proc_dointvec - read a vector of integers
1802  * @table: the sysctl table
1803  * @write: %TRUE if this is a write to the sysctl file
1804  * @filp: the file structure
1805  * @buffer: the user buffer
1806  * @lenp: the size of the user buffer
1807  * @ppos: file position
1808  *
1809  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
1810  * values from/to the user buffer, treated as an ASCII string.
1811  *
1812  * Returns 0 on success.
1813  */
1814 int proc_dointvec(ctl_table *table, int write, struct file *filp,
1815                      void __user *buffer, size_t *lenp, loff_t *ppos)
1816 {
1817     return do_proc_dointvec(table,write,filp,buffer,lenp,ppos,
1818                             NULL,NULL);
1819 }
1820
1821 #define OP_SET  0
1822 #define OP_AND  1
1823 #define OP_OR   2
1824
1825 static int do_proc_dointvec_bset_conv(int *negp, unsigned long *lvalp,
1826                                       int *valp,
1827                                       int write, void *data)
1828 {
1829         int op = *(int *)data;
1830         if (write) {
1831                 int val = *negp ? -*lvalp : *lvalp;
1832                 switch(op) {
1833                 case OP_SET:    *valp = val; break;
1834                 case OP_AND:    *valp &= val; break;
1835                 case OP_OR:     *valp |= val; break;
1836                 }
1837         } else {
1838                 int val = *valp;
1839                 if (val < 0) {
1840                         *negp = -1;
1841                         *lvalp = (unsigned long)-val;
1842                 } else {
1843                         *negp = 0;
1844                         *lvalp = (unsigned long)val;
1845                 }
1846         }
1847         return 0;
1848 }
1849
1850 /*
1851  *      init may raise the set.
1852  */
1853
1854 int proc_dointvec_bset(ctl_table *table, int write, struct file *filp,
1855                         void __user *buffer, size_t *lenp, loff_t *ppos)
1856 {
1857         int op;
1858
1859         if (write && !capable(CAP_SYS_MODULE)) {
1860                 return -EPERM;
1861         }
1862
1863         op = is_init(current) ? OP_SET : OP_AND;
1864         return do_proc_dointvec(table,write,filp,buffer,lenp,ppos,
1865                                 do_proc_dointvec_bset_conv,&op);
1866 }
1867
1868 /*
1869  *      Taint values can only be increased
1870  */
1871 static int proc_dointvec_taint(ctl_table *table, int write, struct file *filp,
1872                                void __user *buffer, size_t *lenp, loff_t *ppos)
1873 {
1874         int op;
1875
1876         if (write && !capable(CAP_SYS_ADMIN))
1877                 return -EPERM;
1878
1879         op = OP_OR;
1880         return do_proc_dointvec(table,write,filp,buffer,lenp,ppos,
1881                                 do_proc_dointvec_bset_conv,&op);
1882 }
1883
1884 struct do_proc_dointvec_minmax_conv_param {
1885         int *min;
1886         int *max;
1887 };
1888
1889 static int do_proc_dointvec_minmax_conv(int *negp, unsigned long *lvalp,
1890                                         int *valp,
1891                                         int write, void *data)
1892 {
1893         struct do_proc_dointvec_minmax_conv_param *param = data;
1894         if (write) {
1895                 int val = *negp ? -*lvalp : *lvalp;
1896                 if ((param->min && *param->min > val) ||
1897                     (param->max && *param->max < val))
1898                         return -EINVAL;
1899                 *valp = val;
1900         } else {
1901                 int val = *valp;
1902                 if (val < 0) {
1903                         *negp = -1;
1904                         *lvalp = (unsigned long)-val;
1905                 } else {
1906                         *negp = 0;
1907                         *lvalp = (unsigned long)val;
1908                 }
1909         }
1910         return 0;
1911 }
1912
1913 /**
1914  * proc_dointvec_minmax - read a vector of integers with min/max values
1915  * @table: the sysctl table
1916  * @write: %TRUE if this is a write to the sysctl file
1917  * @filp: the file structure
1918  * @buffer: the user buffer
1919  * @lenp: the size of the user buffer
1920  * @ppos: file position
1921  *
1922  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
1923  * values from/to the user buffer, treated as an ASCII string.
1924  *
1925  * This routine will ensure the values are within the range specified by
1926  * table->extra1 (min) and table->extra2 (max).
1927  *
1928  * Returns 0 on success.
1929  */
1930 int proc_dointvec_minmax(ctl_table *table, int write, struct file *filp,
1931                   void __user *buffer, size_t *lenp, loff_t *ppos)
1932 {
1933         struct do_proc_dointvec_minmax_conv_param param = {
1934                 .min = (int *) table->extra1,
1935                 .max = (int *) table->extra2,
1936         };
1937         return do_proc_dointvec(table, write, filp, buffer, lenp, ppos,
1938                                 do_proc_dointvec_minmax_conv, &param);
1939 }
1940
1941 static int __do_proc_doulongvec_minmax(void *data, ctl_table *table, int write,
1942                                      struct file *filp,
1943                                      void __user *buffer,
1944                                      size_t *lenp, loff_t *ppos,
1945                                      unsigned long convmul,
1946                                      unsigned long convdiv)
1947 {
1948 #define TMPBUFLEN 21
1949         unsigned long *i, *min, *max, val;
1950         int vleft, first=1, neg;
1951         size_t len, left;
1952         char buf[TMPBUFLEN], *p;
1953         char __user *s = buffer;
1954
1955         if (!data || !table->maxlen || !*lenp ||
1956             (*ppos && !write)) {
1957                 *lenp = 0;
1958                 return 0;
1959         }
1960
1961         i = (unsigned long *) data;
1962         min = (unsigned long *) table->extra1;
1963         max = (unsigned long *) table->extra2;
1964         vleft = table->maxlen / sizeof(unsigned long);
1965         left = *lenp;
1966
1967         for (; left && vleft--; i++, min++, max++, first=0) {
1968                 if (write) {
1969                         while (left) {
1970                                 char c;
1971                                 if (get_user(c, s))
1972                                         return -EFAULT;
1973                                 if (!isspace(c))
1974                                         break;
1975                                 left--;
1976                                 s++;
1977                         }
1978                         if (!left)
1979                                 break;
1980                         neg = 0;
1981                         len = left;
1982                         if (len > TMPBUFLEN-1)
1983                                 len = TMPBUFLEN-1;
1984                         if (copy_from_user(buf, s, len))
1985                                 return -EFAULT;
1986                         buf[len] = 0;
1987                         p = buf;
1988                         if (*p == '-' && left > 1) {
1989                                 neg = 1;
1990                                 p++;
1991                         }
1992                         if (*p < '0' || *p > '9')
1993                                 break;
1994                         val = simple_strtoul(p, &p, 0) * convmul / convdiv ;
1995                         len = p-buf;
1996                         if ((len < left) && *p && !isspace(*p))
1997                                 break;
1998                         if (neg)
1999                                 val = -val;
2000                         s += len;
2001                         left -= len;
2002
2003                         if(neg)
2004                                 continue;
2005                         if ((min && val < *min) || (max && val > *max))
2006                                 continue;
2007                         *i = val;
2008                 } else {
2009                         p = buf;
2010                         if (!first)
2011                                 *p++ = '\t';
2012                         sprintf(p, "%lu", convdiv * (*i) / convmul);
2013                         len = strlen(buf);
2014                         if (len > left)
2015                                 len = left;
2016                         if(copy_to_user(s, buf, len))
2017                                 return -EFAULT;
2018                         left -= len;
2019                         s += len;
2020                 }
2021         }
2022
2023         if (!write && !first && left) {
2024                 if(put_user('\n', s))
2025                         return -EFAULT;
2026                 left--, s++;
2027         }
2028         if (write) {
2029                 while (left) {
2030                         char c;
2031                         if (get_user(c, s++))
2032                                 return -EFAULT;
2033                         if (!isspace(c))
2034                                 break;
2035                         left--;
2036                 }
2037         }
2038         if (write && first)
2039                 return -EINVAL;
2040         *lenp -= left;
2041         *ppos += *lenp;
2042         return 0;
2043 #undef TMPBUFLEN
2044 }
2045
2046 static int do_proc_doulongvec_minmax(ctl_table *table, int write,
2047                                      struct file *filp,
2048                                      void __user *buffer,
2049                                      size_t *lenp, loff_t *ppos,
2050                                      unsigned long convmul,
2051                                      unsigned long convdiv)
2052 {
2053         return __do_proc_doulongvec_minmax(table->data, table, write,
2054                         filp, buffer, lenp, ppos, convmul, convdiv);
2055 }
2056
2057 /**
2058  * proc_doulongvec_minmax - read a vector of long integers with min/max values
2059  * @table: the sysctl table
2060  * @write: %TRUE if this is a write to the sysctl file
2061  * @filp: the file structure
2062  * @buffer: the user buffer
2063  * @lenp: the size of the user buffer
2064  * @ppos: file position
2065  *
2066  * Reads/writes up to table->maxlen/sizeof(unsigned long) unsigned long
2067  * values from/to the user buffer, treated as an ASCII string.
2068  *
2069  * This routine will ensure the values are within the range specified by
2070  * table->extra1 (min) and table->extra2 (max).
2071  *
2072  * Returns 0 on success.
2073  */
2074 int proc_doulongvec_minmax(ctl_table *table, int write, struct file *filp,
2075                            void __user *buffer, size_t *lenp, loff_t *ppos)
2076 {
2077     return do_proc_doulongvec_minmax(table, write, filp, buffer, lenp, ppos, 1l, 1l);
2078 }
2079
2080 /**
2081  * proc_doulongvec_ms_jiffies_minmax - read a vector of millisecond values with min/max values
2082  * @table: the sysctl table
2083  * @write: %TRUE if this is a write to the sysctl file
2084  * @filp: the file structure
2085  * @buffer: the user buffer
2086  * @lenp: the size of the user buffer
2087  * @ppos: file position
2088  *
2089  * Reads/writes up to table->maxlen/sizeof(unsigned long) unsigned long
2090  * values from/to the user buffer, treated as an ASCII string. The values
2091  * are treated as milliseconds, and converted to jiffies when they are stored.
2092  *
2093  * This routine will ensure the values are within the range specified by
2094  * table->extra1 (min) and table->extra2 (max).
2095  *
2096  * Returns 0 on success.
2097  */
2098 int proc_doulongvec_ms_jiffies_minmax(ctl_table *table, int write,
2099                                       struct file *filp,
2100                                       void __user *buffer,
2101                                       size_t *lenp, loff_t *ppos)
2102 {
2103     return do_proc_doulongvec_minmax(table, write, filp, buffer,
2104                                      lenp, ppos, HZ, 1000l);
2105 }
2106
2107
2108 static int do_proc_dointvec_jiffies_conv(int *negp, unsigned long *lvalp,
2109                                          int *valp,
2110                                          int write, void *data)
2111 {
2112         if (write) {
2113                 if (*lvalp > LONG_MAX / HZ)
2114                         return 1;
2115                 *valp = *negp ? -(*lvalp*HZ) : (*lvalp*HZ);
2116         } else {
2117                 int val = *valp;
2118                 unsigned long lval;
2119                 if (val < 0) {
2120                         *negp = -1;
2121                         lval = (unsigned long)-val;
2122                 } else {
2123                         *negp = 0;
2124                         lval = (unsigned long)val;
2125                 }
2126                 *lvalp = lval / HZ;
2127         }
2128         return 0;
2129 }
2130
2131 static int do_proc_dointvec_userhz_jiffies_conv(int *negp, unsigned long *lvalp,
2132                                                 int *valp,
2133                                                 int write, void *data)
2134 {
2135         if (write) {
2136                 if (USER_HZ < HZ && *lvalp > (LONG_MAX / HZ) * USER_HZ)
2137                         return 1;
2138                 *valp = clock_t_to_jiffies(*negp ? -*lvalp : *lvalp);
2139         } else {
2140                 int val = *valp;
2141                 unsigned long lval;
2142                 if (val < 0) {
2143                         *negp = -1;
2144                         lval = (unsigned long)-val;
2145                 } else {
2146                         *negp = 0;
2147                         lval = (unsigned long)val;
2148                 }
2149                 *lvalp = jiffies_to_clock_t(lval);
2150         }
2151         return 0;
2152 }
2153
2154 static int do_proc_dointvec_ms_jiffies_conv(int *negp, unsigned long *lvalp,
2155                                             int *valp,
2156                                             int write, void *data)
2157 {
2158         if (write) {
2159                 *valp = msecs_to_jiffies(*negp ? -*lvalp : *lvalp);
2160         } else {
2161                 int val = *valp;
2162                 unsigned long lval;
2163                 if (val < 0) {
2164                         *negp = -1;
2165                         lval = (unsigned long)-val;
2166                 } else {
2167                         *negp = 0;
2168                         lval = (unsigned long)val;
2169                 }
2170                 *lvalp = jiffies_to_msecs(lval);
2171         }
2172         return 0;
2173 }
2174
2175 /**
2176  * proc_dointvec_jiffies - read a vector of integers as seconds
2177  * @table: the sysctl table
2178  * @write: %TRUE if this is a write to the sysctl file
2179  * @filp: the file structure
2180  * @buffer: the user buffer
2181  * @lenp: the size of the user buffer
2182  * @ppos: file position
2183  *
2184  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2185  * values from/to the user buffer, treated as an ASCII string.
2186  * The values read are assumed to be in seconds, and are converted into
2187  * jiffies.
2188  *
2189  * Returns 0 on success.
2190  */
2191 int proc_dointvec_jiffies(ctl_table *table, int write, struct file *filp,
2192                           void __user *buffer, size_t *lenp, loff_t *ppos)
2193 {
2194     return do_proc_dointvec(table,write,filp,buffer,lenp,ppos,
2195                             do_proc_dointvec_jiffies_conv,NULL);
2196 }
2197
2198 /**
2199  * proc_dointvec_userhz_jiffies - read a vector of integers as 1/USER_HZ seconds
2200  * @table: the sysctl table
2201  * @write: %TRUE if this is a write to the sysctl file
2202  * @filp: the file structure
2203  * @buffer: the user buffer
2204  * @lenp: the size of the user buffer
2205  * @ppos: pointer to the file position
2206  *
2207  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2208  * values from/to the user buffer, treated as an ASCII string.
2209  * The values read are assumed to be in 1/USER_HZ seconds, and
2210  * are converted into jiffies.
2211  *
2212  * Returns 0 on success.
2213  */
2214 int proc_dointvec_userhz_jiffies(ctl_table *table, int write, struct file *filp,
2215                                  void __user *buffer, size_t *lenp, loff_t *ppos)
2216 {
2217     return do_proc_dointvec(table,write,filp,buffer,lenp,ppos,
2218                             do_proc_dointvec_userhz_jiffies_conv,NULL);
2219 }
2220
2221 /**
2222  * proc_dointvec_ms_jiffies - read a vector of integers as 1 milliseconds
2223  * @table: the sysctl table
2224  * @write: %TRUE if this is a write to the sysctl file
2225  * @filp: the file structure
2226  * @buffer: the user buffer
2227  * @lenp: the size of the user buffer
2228  * @ppos: file position
2229  * @ppos: the current position in the file
2230  *
2231  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2232  * values from/to the user buffer, treated as an ASCII string.
2233  * The values read are assumed to be in 1/1000 seconds, and
2234  * are converted into jiffies.
2235  *
2236  * Returns 0 on success.
2237  */
2238 int proc_dointvec_ms_jiffies(ctl_table *table, int write, struct file *filp,
2239                              void __user *buffer, size_t *lenp, loff_t *ppos)
2240 {
2241         return do_proc_dointvec(table, write, filp, buffer, lenp, ppos,
2242                                 do_proc_dointvec_ms_jiffies_conv, NULL);
2243 }
2244
2245 static int proc_do_cad_pid(ctl_table *table, int write, struct file *filp,
2246                            void __user *buffer, size_t *lenp, loff_t *ppos)
2247 {
2248         struct pid *new_pid;
2249         pid_t tmp;
2250         int r;
2251
2252         tmp = pid_nr(cad_pid);
2253
2254         r = __do_proc_dointvec(&tmp, table, write, filp, buffer,
2255                                lenp, ppos, NULL, NULL);
2256         if (r || !write)
2257                 return r;
2258
2259         new_pid = find_get_pid(tmp);
2260         if (!new_pid)
2261                 return -ESRCH;
2262
2263         put_pid(xchg(&cad_pid, new_pid));
2264         return 0;
2265 }
2266
2267 #else /* CONFIG_PROC_FS */
2268
2269 int proc_dostring(ctl_table *table, int write, struct file *filp,
2270                   void __user *buffer, size_t *lenp, loff_t *ppos)
2271 {
2272         return -ENOSYS;
2273 }
2274
2275 int proc_dointvec(ctl_table *table, int write, struct file *filp,
2276                   void __user *buffer, size_t *lenp, loff_t *ppos)
2277 {
2278         return -ENOSYS;
2279 }
2280
2281 int proc_dointvec_bset(ctl_table *table, int write, struct file *filp,
2282                         void __user *buffer, size_t *lenp, loff_t *ppos)
2283 {
2284         return -ENOSYS;
2285 }
2286
2287 int proc_dointvec_minmax(ctl_table *table, int write, struct file *filp,
2288                     void __user *buffer, size_t *lenp, loff_t *ppos)
2289 {
2290         return -ENOSYS;
2291 }
2292
2293 int proc_dointvec_jiffies(ctl_table *table, int write, struct file *filp,
2294                     void __user *buffer, size_t *lenp, loff_t *ppos)
2295 {
2296         return -ENOSYS;
2297 }
2298
2299 int proc_dointvec_userhz_jiffies(ctl_table *table, int write, struct file *filp,
2300                     void __user *buffer, size_t *lenp, loff_t *ppos)
2301 {
2302         return -ENOSYS;
2303 }
2304
2305 int proc_dointvec_ms_jiffies(ctl_table *table, int write, struct file *filp,
2306                              void __user *buffer, size_t *lenp, loff_t *ppos)
2307 {
2308         return -ENOSYS;
2309 }
2310
2311 int proc_doulongvec_minmax(ctl_table *table, int write, struct file *filp,
2312                     void __user *buffer, size_t *lenp, loff_t *ppos)
2313 {
2314         return -ENOSYS;
2315 }
2316
2317 int proc_doulongvec_ms_jiffies_minmax(ctl_table *table, int write,
2318                                       struct file *filp,
2319                                       void __user *buffer,
2320                                       size_t *lenp, loff_t *ppos)
2321 {
2322     return -ENOSYS;
2323 }
2324
2325
2326 #endif /* CONFIG_PROC_FS */
2327
2328
2329 #ifdef CONFIG_SYSCTL_SYSCALL
2330 /*
2331  * General sysctl support routines
2332  */
2333
2334 /* The generic string strategy routine: */
2335 int sysctl_string(ctl_table *table, int __user *name, int nlen,
2336                   void __user *oldval, size_t __user *oldlenp,
2337                   void __user *newval, size_t newlen)
2338 {
2339         if (!table->data || !table->maxlen)
2340                 return -ENOTDIR;
2341
2342         if (oldval && oldlenp) {
2343                 size_t bufsize;
2344                 if (get_user(bufsize, oldlenp))
2345                         return -EFAULT;
2346                 if (bufsize) {
2347                         size_t len = strlen(table->data), copied;
2348
2349                         /* This shouldn't trigger for a well-formed sysctl */
2350                         if (len > table->maxlen)
2351                                 len = table->maxlen;
2352
2353                         /* Copy up to a max of bufsize-1 bytes of the string */
2354                         copied = (len >= bufsize) ? bufsize - 1 : len;
2355
2356                         if (copy_to_user(oldval, table->data, copied) ||
2357                             put_user(0, (char __user *)(oldval + copied)))
2358                                 return -EFAULT;
2359                         if (put_user(len, oldlenp))
2360                                 return -EFAULT;
2361                 }
2362         }
2363         if (newval && newlen) {
2364                 size_t len = newlen;
2365                 if (len > table->maxlen)
2366                         len = table->maxlen;
2367                 if(copy_from_user(table->data, newval, len))
2368                         return -EFAULT;
2369                 if (len == table->maxlen)
2370                         len--;
2371                 ((char *) table->data)[len] = 0;
2372         }
2373         return 1;
2374 }
2375
2376 /*
2377  * This function makes sure that all of the integers in the vector
2378  * are between the minimum and maximum values given in the arrays
2379  * table->extra1 and table->extra2, respectively.
2380  */
2381 int sysctl_intvec(ctl_table *table, int __user *name, int nlen,
2382                 void __user *oldval, size_t __user *oldlenp,
2383                 void __user *newval, size_t newlen)
2384 {
2385
2386         if (newval && newlen) {
2387                 int __user *vec = (int __user *) newval;
2388                 int *min = (int *) table->extra1;
2389                 int *max = (int *) table->extra2;
2390                 size_t length;
2391                 int i;
2392
2393                 if (newlen % sizeof(int) != 0)
2394                         return -EINVAL;
2395
2396                 if (!table->extra1 && !table->extra2)
2397                         return 0;
2398
2399                 if (newlen > table->maxlen)
2400                         newlen = table->maxlen;
2401                 length = newlen / sizeof(int);
2402
2403                 for (i = 0; i < length; i++) {
2404                         int value;
2405                         if (get_user(value, vec + i))
2406                                 return -EFAULT;
2407                         if (min && value < min[i])
2408                                 return -EINVAL;
2409                         if (max && value > max[i])
2410                                 return -EINVAL;
2411                 }
2412         }
2413         return 0;
2414 }
2415
2416 /* Strategy function to convert jiffies to seconds */
2417 int sysctl_jiffies(ctl_table *table, int __user *name, int nlen,
2418                 void __user *oldval, size_t __user *oldlenp,
2419                 void __user *newval, size_t newlen)
2420 {
2421         if (oldval && oldlenp) {
2422                 size_t olen;
2423
2424                 if (get_user(olen, oldlenp))
2425                         return -EFAULT;
2426                 if (olen) {
2427                         int val;
2428
2429                         if (olen < sizeof(int))
2430                                 return -EINVAL;
2431
2432                         val = *(int *)(table->data) / HZ;
2433                         if (put_user(val, (int __user *)oldval))
2434                                 return -EFAULT;
2435                         if (put_user(sizeof(int), oldlenp))
2436                                 return -EFAULT;
2437                 }
2438         }
2439         if (newval && newlen) {
2440                 int new;
2441                 if (newlen != sizeof(int))
2442                         return -EINVAL;
2443                 if (get_user(new, (int __user *)newval))
2444                         return -EFAULT;
2445                 *(int *)(table->data) = new*HZ;
2446         }
2447         return 1;
2448 }
2449
2450 /* Strategy function to convert jiffies to seconds */
2451 int sysctl_ms_jiffies(ctl_table *table, int __user *name, int nlen,
2452                 void __user *oldval, size_t __user *oldlenp,
2453                 void __user *newval, size_t newlen)
2454 {
2455         if (oldval && oldlenp) {
2456                 size_t olen;
2457
2458                 if (get_user(olen, oldlenp))
2459                         return -EFAULT;
2460                 if (olen) {
2461                         int val;
2462
2463                         if (olen < sizeof(int))
2464                                 return -EINVAL;
2465
2466                         val = jiffies_to_msecs(*(int *)(table->data));
2467                         if (put_user(val, (int __user *)oldval))
2468                                 return -EFAULT;
2469                         if (put_user(sizeof(int), oldlenp))
2470                                 return -EFAULT;
2471                 }
2472         }
2473         if (newval && newlen) {
2474                 int new;
2475                 if (newlen != sizeof(int))
2476                         return -EINVAL;
2477                 if (get_user(new, (int __user *)newval))
2478                         return -EFAULT;
2479                 *(int *)(table->data) = msecs_to_jiffies(new);
2480         }
2481         return 1;
2482 }
2483
2484
2485
2486 #else /* CONFIG_SYSCTL_SYSCALL */
2487
2488
2489 asmlinkage long sys_sysctl(struct __sysctl_args __user *args)
2490 {
2491         static int msg_count;
2492         struct __sysctl_args tmp;
2493         int name[CTL_MAXNAME];
2494         int i;
2495
2496         /* Read in the sysctl name for better debug message logging */
2497         if (copy_from_user(&tmp, args, sizeof(tmp)))
2498                 return -EFAULT;
2499         if (tmp.nlen <= 0 || tmp.nlen >= CTL_MAXNAME)
2500                 return -ENOTDIR;
2501         for (i = 0; i < tmp.nlen; i++)
2502                 if (get_user(name[i], tmp.name + i))
2503                         return -EFAULT;
2504
2505         /* Ignore accesses to kernel.version */
2506         if ((tmp.nlen == 2) && (name[0] == CTL_KERN) && (name[1] == KERN_VERSION))
2507                 goto out;
2508
2509         if (msg_count < 5) {
2510                 msg_count++;
2511                 printk(KERN_INFO
2512                         "warning: process `%s' used the removed sysctl "
2513                         "system call with ", current->comm);
2514                 for (i = 0; i < tmp.nlen; i++)
2515                         printk("%d.", name[i]);
2516                 printk("\n");
2517         }
2518 out:
2519         return -ENOSYS;
2520 }
2521
2522 int sysctl_string(ctl_table *table, int __user *name, int nlen,
2523                   void __user *oldval, size_t __user *oldlenp,
2524                   void __user *newval, size_t newlen)
2525 {
2526         return -ENOSYS;
2527 }
2528
2529 int sysctl_intvec(ctl_table *table, int __user *name, int nlen,
2530                 void __user *oldval, size_t __user *oldlenp,
2531                 void __user *newval, size_t newlen)
2532 {
2533         return -ENOSYS;
2534 }
2535
2536 int sysctl_jiffies(ctl_table *table, int __user *name, int nlen,
2537                 void __user *oldval, size_t __user *oldlenp,
2538                 void __user *newval, size_t newlen)
2539 {
2540         return -ENOSYS;
2541 }
2542
2543 int sysctl_ms_jiffies(ctl_table *table, int __user *name, int nlen,
2544                 void __user *oldval, size_t __user *oldlenp,
2545                 void __user *newval, size_t newlen)
2546 {
2547         return -ENOSYS;
2548 }
2549
2550 #endif /* CONFIG_SYSCTL_SYSCALL */
2551
2552 /*
2553  * No sense putting this after each symbol definition, twice,
2554  * exception granted :-)
2555  */
2556 EXPORT_SYMBOL(proc_dointvec);
2557 EXPORT_SYMBOL(proc_dointvec_jiffies);
2558 EXPORT_SYMBOL(proc_dointvec_minmax);
2559 EXPORT_SYMBOL(proc_dointvec_userhz_jiffies);
2560 EXPORT_SYMBOL(proc_dointvec_ms_jiffies);
2561 EXPORT_SYMBOL(proc_dostring);
2562 EXPORT_SYMBOL(proc_doulongvec_minmax);
2563 EXPORT_SYMBOL(proc_doulongvec_ms_jiffies_minmax);
2564 EXPORT_SYMBOL(register_sysctl_table);
2565 EXPORT_SYMBOL(sysctl_intvec);
2566 EXPORT_SYMBOL(sysctl_jiffies);
2567 EXPORT_SYMBOL(sysctl_ms_jiffies);
2568 EXPORT_SYMBOL(sysctl_string);
2569 EXPORT_SYMBOL(unregister_sysctl_table);