git.oblomov.eu Git - linux-2.6/blob - 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/security.h>
  28 #include <linux/ctype.h>
  29 #include <linux/utsname.h>
  30 #include <linux/smp_lock.h>
  31 #include <linux/fs.h>
  32 #include <linux/init.h>
  33 #include <linux/kernel.h>
  34 #include <linux/kobject.h>
  35 #include <linux/net.h>
  36 #include <linux/sysrq.h>
  37 #include <linux/highuid.h>
  38 #include <linux/writeback.h>
  39 #include <linux/hugetlb.h>
  40 #include <linux/initrd.h>
  41 #include <linux/key.h>
  42 #include <linux/times.h>
  43 #include <linux/limits.h>
  44 #include <linux/dcache.h>
  45 #include <linux/syscalls.h>
  46 #include <linux/vmstat.h>
  47 #include <linux/nfs_fs.h>
  48 #include <linux/acpi.h>
  49 #include <linux/reboot.h>
  50 #include <linux/ftrace.h>
  51 #include <linux/slow-work.h>
  52
  53 #include <asm/uaccess.h>
  54 #include <asm/processor.h>
  55
  56 #ifdef CONFIG_X86
  57 #include <asm/nmi.h>
  58 #include <asm/stacktrace.h>
  59 #include <asm/io.h>
  60 #endif
  61
  62 static int deprecated_sysctl_warning(struct __sysctl_args *args);
  63
  64 #if defined(CONFIG_SYSCTL)
  65
  66 /* External variables not in a header file. */
  67 extern int C_A_D;
  68 extern int print_fatal_signals;
  69 extern int sysctl_overcommit_memory;
  70 extern int sysctl_overcommit_ratio;
  71 extern int sysctl_panic_on_oom;
  72 extern int sysctl_oom_kill_allocating_task;
  73 extern int sysctl_oom_dump_tasks;
  74 extern int max_threads;
  75 extern int core_uses_pid;
  76 extern int suid_dumpable;
  77 extern char core_pattern[];
  78 extern int pid_max;
  79 extern int min_free_kbytes;
  80 extern int pid_max_min, pid_max_max;
  81 extern int sysctl_drop_caches;
  82 extern int percpu_pagelist_fraction;
  83 extern int compat_log;
  84 extern int latencytop_enabled;
  85 extern int sysctl_nr_open_min, sysctl_nr_open_max;
  86 #ifndef CONFIG_MMU
  87 extern int sysctl_nr_trim_pages;
  88 #endif
  89 #ifdef CONFIG_RCU_TORTURE_TEST
  90 extern int rcutorture_runnable;
  91 #endif /* #ifdef CONFIG_RCU_TORTURE_TEST */
  92
  93 /* Constants used for minimum and  maximum */
  94 #ifdef CONFIG_DETECT_SOFTLOCKUP
  95 static int sixty = 60;
  96 static int neg_one = -1;
  97 #endif
  98
  99 static int zero;
 100 static int one = 1;
 101 static int two = 2;
 102 static unsigned long one_ul = 1;
 103 static int one_hundred = 100;
 104
 105 /* this is needed for the proc_dointvec_minmax for [fs_]overflow UID and GID */
 106 static int maxolduid = 65535;
 107 static int minolduid;
 108 static int min_percpu_pagelist_fract = 8;
 109
 110 static int ngroups_max = NGROUPS_MAX;
 111
 112 #ifdef CONFIG_MODULES
 113 extern char modprobe_path[];
 114 #endif
 115 #ifdef CONFIG_CHR_DEV_SG
 116 extern int sg_big_buff;
 117 #endif
 118
 119 #ifdef CONFIG_SPARC
 120 #include <asm/system.h>
 121 #endif
 122
 123 #ifdef CONFIG_SPARC64
 124 extern int sysctl_tsb_ratio;
 125 #endif
 126
 127 #ifdef __hppa__
 128 extern int pwrsw_enabled;
 129 extern int unaligned_enabled;
 130 #endif
 131
 132 #ifdef CONFIG_S390
 133 #ifdef CONFIG_MATHEMU
 134 extern int sysctl_ieee_emulation_warnings;
 135 #endif
 136 extern int sysctl_userprocess_debug;
 137 extern int spin_retry;
 138 #endif
 139
 140 #ifdef CONFIG_BSD_PROCESS_ACCT
 141 extern int acct_parm[];
 142 #endif
 143
 144 #ifdef CONFIG_IA64
 145 extern int no_unaligned_warning;
 146 extern int unaligned_dump_stack;
 147 #endif
 148
 149 #ifdef CONFIG_RT_MUTEXES
 150 extern int max_lock_depth;
 151 #endif
 152
 153 #ifdef CONFIG_PROC_SYSCTL
 154 static int proc_do_cad_pid(struct ctl_table *table, int write, struct file *filp,
 155                   void __user *buffer, size_t *lenp, loff_t *ppos);
 156 static int proc_taint(struct ctl_table *table, int write, struct file *filp,
 157                                void __user *buffer, size_t *lenp, loff_t *ppos);
 158 #endif
 159
 160 static struct ctl_table root_table[];
 161 static struct ctl_table_root sysctl_table_root;
 162 static struct ctl_table_header root_table_header = {
 163         .count = 1,
 164         .ctl_table = root_table,
 165         .ctl_entry = LIST_HEAD_INIT(sysctl_table_root.default_set.list),
 166         .root = &sysctl_table_root,
 167         .set = &sysctl_table_root.default_set,
 168 };
 169 static struct ctl_table_root sysctl_table_root = {
 170         .root_list = LIST_HEAD_INIT(sysctl_table_root.root_list),
 171         .default_set.list = LIST_HEAD_INIT(root_table_header.ctl_entry),
 172 };
 173
 174 static struct ctl_table kern_table[];
 175 static struct ctl_table vm_table[];
 176 static struct ctl_table fs_table[];
 177 static struct ctl_table debug_table[];
 178 static struct ctl_table dev_table[];
 179 extern struct ctl_table random_table[];
 180 #ifdef CONFIG_INOTIFY_USER
 181 extern struct ctl_table inotify_table[];
 182 #endif
 183 #ifdef CONFIG_EPOLL
 184 extern struct ctl_table epoll_table[];
 185 #endif
 186
 187 #ifdef HAVE_ARCH_PICK_MMAP_LAYOUT
 188 int sysctl_legacy_va_layout;
 189 #endif
 190
 191 extern int prove_locking;
 192 extern int lock_stat;
 193
 194 /* The default sysctl tables: */
 195
 196 static struct ctl_table root_table[] = {
 197         {
 198                 .ctl_name       = CTL_KERN,
 199                 .procname       = "kernel",
 200                 .mode           = 0555,
 201                 .child          = kern_table,
 202         },
 203         {
 204                 .ctl_name       = CTL_VM,
 205                 .procname       = "vm",
 206                 .mode           = 0555,
 207                 .child          = vm_table,
 208         },
 209         {
 210                 .ctl_name       = CTL_FS,
 211                 .procname       = "fs",
 212                 .mode           = 0555,
 213                 .child          = fs_table,
 214         },
 215         {
 216                 .ctl_name       = CTL_DEBUG,
 217                 .procname       = "debug",
 218                 .mode           = 0555,
 219                 .child          = debug_table,
 220         },
 221         {
 222                 .ctl_name       = CTL_DEV,
 223                 .procname       = "dev",
 224                 .mode           = 0555,
 225                 .child          = dev_table,
 226         },
 227 /*
 228  * NOTE: do not add new entries to this table unless you have read
 229  * Documentation/sysctl/ctl_unnumbered.txt
 230  */
 231         { .ctl_name = 0 }
 232 };
 233
 234 #ifdef CONFIG_SCHED_DEBUG
 235 static int min_sched_granularity_ns = 100000;           /* 100 usecs */
 236 static int max_sched_granularity_ns = NSEC_PER_SEC;     /* 1 second */
 237 static int min_wakeup_granularity_ns;                   /* 0 usecs */
 238 static int max_wakeup_granularity_ns = NSEC_PER_SEC;    /* 1 second */
 239 #endif
 240
 241 static struct ctl_table kern_table[] = {
 242 #ifdef CONFIG_SCHED_DEBUG
 243         {
 244                 .ctl_name       = CTL_UNNUMBERED,
 245                 .procname       = "sched_min_granularity_ns",
 246                 .data           = &sysctl_sched_min_granularity,
 247                 .maxlen         = sizeof(unsigned int),
 248                 .mode           = 0644,
 249                 .proc_handler   = &sched_nr_latency_handler,
 250                 .strategy       = &sysctl_intvec,
 251                 .extra1         = &min_sched_granularity_ns,
 252                 .extra2         = &max_sched_granularity_ns,
 253         },
 254         {
 255                 .ctl_name       = CTL_UNNUMBERED,
 256                 .procname       = "sched_latency_ns",
 257                 .data           = &sysctl_sched_latency,
 258                 .maxlen         = sizeof(unsigned int),
 259                 .mode           = 0644,
 260                 .proc_handler   = &sched_nr_latency_handler,
 261                 .strategy       = &sysctl_intvec,
 262                 .extra1         = &min_sched_granularity_ns,
 263                 .extra2         = &max_sched_granularity_ns,
 264         },
 265         {
 266                 .ctl_name       = CTL_UNNUMBERED,
 267                 .procname       = "sched_wakeup_granularity_ns",
 268                 .data           = &sysctl_sched_wakeup_granularity,
 269                 .maxlen         = sizeof(unsigned int),
 270                 .mode           = 0644,
 271                 .proc_handler   = &proc_dointvec_minmax,
 272                 .strategy       = &sysctl_intvec,
 273                 .extra1         = &min_wakeup_granularity_ns,
 274                 .extra2         = &max_wakeup_granularity_ns,
 275         },
 276         {
 277                 .ctl_name       = CTL_UNNUMBERED,
 278                 .procname       = "sched_shares_ratelimit",
 279                 .data           = &sysctl_sched_shares_ratelimit,
 280                 .maxlen         = sizeof(unsigned int),
 281                 .mode           = 0644,
 282                 .proc_handler   = &proc_dointvec,
 283         },
 284         {
 285                 .ctl_name       = CTL_UNNUMBERED,
 286                 .procname       = "sched_shares_thresh",
 287                 .data           = &sysctl_sched_shares_thresh,
 288                 .maxlen         = sizeof(unsigned int),
 289                 .mode           = 0644,
 290                 .proc_handler   = &proc_dointvec_minmax,
 291                 .strategy       = &sysctl_intvec,
 292                 .extra1         = &zero,
 293         },
 294         {
 295                 .ctl_name       = CTL_UNNUMBERED,
 296                 .procname       = "sched_child_runs_first",
 297                 .data           = &sysctl_sched_child_runs_first,
 298                 .maxlen         = sizeof(unsigned int),
 299                 .mode           = 0644,
 300                 .proc_handler   = &proc_dointvec,
 301         },
 302         {
 303                 .ctl_name       = CTL_UNNUMBERED,
 304                 .procname       = "sched_features",
 305                 .data           = &sysctl_sched_features,
 306                 .maxlen         = sizeof(unsigned int),
 307                 .mode           = 0644,
 308                 .proc_handler   = &proc_dointvec,
 309         },
 310         {
 311                 .ctl_name       = CTL_UNNUMBERED,
 312                 .procname       = "sched_migration_cost",
 313                 .data           = &sysctl_sched_migration_cost,
 314                 .maxlen         = sizeof(unsigned int),
 315                 .mode           = 0644,
 316                 .proc_handler   = &proc_dointvec,
 317         },
 318         {
 319                 .ctl_name       = CTL_UNNUMBERED,
 320                 .procname       = "sched_nr_migrate",
 321                 .data           = &sysctl_sched_nr_migrate,
 322                 .maxlen         = sizeof(unsigned int),
 323                 .mode           = 0644,
 324                 .proc_handler   = &proc_dointvec,
 325         },
 326 #endif
 327         {
 328                 .ctl_name       = CTL_UNNUMBERED,
 329                 .procname       = "sched_rt_period_us",
 330                 .data           = &sysctl_sched_rt_period,
 331                 .maxlen         = sizeof(unsigned int),
 332                 .mode           = 0644,
 333                 .proc_handler   = &sched_rt_handler,
 334         },
 335         {
 336                 .ctl_name       = CTL_UNNUMBERED,
 337                 .procname       = "sched_rt_runtime_us",
 338                 .data           = &sysctl_sched_rt_runtime,
 339                 .maxlen         = sizeof(int),
 340                 .mode           = 0644,
 341                 .proc_handler   = &sched_rt_handler,
 342         },
 343         {
 344                 .ctl_name       = CTL_UNNUMBERED,
 345                 .procname       = "sched_compat_yield",
 346                 .data           = &sysctl_sched_compat_yield,
 347                 .maxlen         = sizeof(unsigned int),
 348                 .mode           = 0644,
 349                 .proc_handler   = &proc_dointvec,
 350         },
 351 #ifdef CONFIG_PROVE_LOCKING
 352         {
 353                 .ctl_name       = CTL_UNNUMBERED,
 354                 .procname       = "prove_locking",
 355                 .data           = &prove_locking,
 356                 .maxlen         = sizeof(int),
 357                 .mode           = 0644,
 358                 .proc_handler   = &proc_dointvec,
 359         },
 360 #endif
 361 #ifdef CONFIG_LOCK_STAT
 362         {
 363                 .ctl_name       = CTL_UNNUMBERED,
 364                 .procname       = "lock_stat",
 365                 .data           = &lock_stat,
 366                 .maxlen         = sizeof(int),
 367                 .mode           = 0644,
 368                 .proc_handler   = &proc_dointvec,
 369         },
 370 #endif
 371         {
 372                 .ctl_name       = KERN_PANIC,
 373                 .procname       = "panic",
 374                 .data           = &panic_timeout,
 375                 .maxlen         = sizeof(int),
 376                 .mode           = 0644,
 377                 .proc_handler   = &proc_dointvec,
 378         },
 379         {
 380                 .ctl_name       = KERN_CORE_USES_PID,
 381                 .procname       = "core_uses_pid",
 382                 .data           = &core_uses_pid,
 383                 .maxlen         = sizeof(int),
 384                 .mode           = 0644,
 385                 .proc_handler   = &proc_dointvec,
 386         },
 387         {
 388                 .ctl_name       = KERN_CORE_PATTERN,
 389                 .procname       = "core_pattern",
 390                 .data           = core_pattern,
 391                 .maxlen         = CORENAME_MAX_SIZE,
 392                 .mode           = 0644,
 393                 .proc_handler   = &proc_dostring,
 394                 .strategy       = &sysctl_string,
 395         },
 396 #ifdef CONFIG_PROC_SYSCTL
 397         {
 398                 .procname       = "tainted",
 399                 .maxlen         = sizeof(long),
 400                 .mode           = 0644,
 401                 .proc_handler   = &proc_taint,
 402         },
 403 #endif
 404 #ifdef CONFIG_LATENCYTOP
 405         {
 406                 .procname       = "latencytop",
 407                 .data           = &latencytop_enabled,
 408                 .maxlen         = sizeof(int),
 409                 .mode           = 0644,
 410                 .proc_handler   = &proc_dointvec,
 411         },
 412 #endif
 413 #ifdef CONFIG_BLK_DEV_INITRD
 414         {
 415                 .ctl_name       = KERN_REALROOTDEV,
 416                 .procname       = "real-root-dev",
 417                 .data           = &real_root_dev,
 418                 .maxlen         = sizeof(int),
 419                 .mode           = 0644,
 420                 .proc_handler   = &proc_dointvec,
 421         },
 422 #endif
 423         {
 424                 .ctl_name       = CTL_UNNUMBERED,
 425                 .procname       = "print-fatal-signals",
 426                 .data           = &print_fatal_signals,
 427                 .maxlen         = sizeof(int),
 428                 .mode           = 0644,
 429                 .proc_handler   = &proc_dointvec,
 430         },
 431 #ifdef CONFIG_SPARC
 432         {
 433                 .ctl_name       = KERN_SPARC_REBOOT,
 434                 .procname       = "reboot-cmd",
 435                 .data           = reboot_command,
 436                 .maxlen         = 256,
 437                 .mode           = 0644,
 438                 .proc_handler   = &proc_dostring,
 439                 .strategy       = &sysctl_string,
 440         },
 441         {
 442                 .ctl_name       = KERN_SPARC_STOP_A,
 443                 .procname       = "stop-a",
 444                 .data           = &stop_a_enabled,
 445                 .maxlen         = sizeof (int),
 446                 .mode           = 0644,
 447                 .proc_handler   = &proc_dointvec,
 448         },
 449         {
 450                 .ctl_name       = KERN_SPARC_SCONS_PWROFF,
 451                 .procname       = "scons-poweroff",
 452                 .data           = &scons_pwroff,
 453                 .maxlen         = sizeof (int),
 454                 .mode           = 0644,
 455                 .proc_handler   = &proc_dointvec,
 456         },
 457 #endif
 458 #ifdef CONFIG_SPARC64
 459         {
 460                 .ctl_name       = CTL_UNNUMBERED,
 461                 .procname       = "tsb-ratio",
 462                 .data           = &sysctl_tsb_ratio,
 463                 .maxlen         = sizeof (int),
 464                 .mode           = 0644,
 465                 .proc_handler   = &proc_dointvec,
 466         },
 467 #endif
 468 #ifdef __hppa__
 469         {
 470                 .ctl_name       = KERN_HPPA_PWRSW,
 471                 .procname       = "soft-power",
 472                 .data           = &pwrsw_enabled,
 473                 .maxlen         = sizeof (int),
 474                 .mode           = 0644,
 475                 .proc_handler   = &proc_dointvec,
 476         },
 477         {
 478                 .ctl_name       = KERN_HPPA_UNALIGNED,
 479                 .procname       = "unaligned-trap",
 480                 .data           = &unaligned_enabled,
 481                 .maxlen         = sizeof (int),
 482                 .mode           = 0644,
 483                 .proc_handler   = &proc_dointvec,
 484         },
 485 #endif
 486         {
 487                 .ctl_name       = KERN_CTLALTDEL,
 488                 .procname       = "ctrl-alt-del",
 489                 .data           = &C_A_D,
 490                 .maxlen         = sizeof(int),
 491                 .mode           = 0644,
 492                 .proc_handler   = &proc_dointvec,
 493         },
 494 #ifdef CONFIG_FUNCTION_TRACER
 495         {
 496                 .ctl_name       = CTL_UNNUMBERED,
 497                 .procname       = "ftrace_enabled",
 498                 .data           = &ftrace_enabled,
 499                 .maxlen         = sizeof(int),
 500                 .mode           = 0644,
 501                 .proc_handler   = &ftrace_enable_sysctl,
 502         },
 503 #endif
 504 #ifdef CONFIG_STACK_TRACER
 505         {
 506                 .ctl_name       = CTL_UNNUMBERED,
 507                 .procname       = "stack_tracer_enabled",
 508                 .data           = &stack_tracer_enabled,
 509                 .maxlen         = sizeof(int),
 510                 .mode           = 0644,
 511                 .proc_handler   = &stack_trace_sysctl,
 512         },
 513 #endif
 514 #ifdef CONFIG_TRACING
 515         {
 516                 .ctl_name       = CTL_UNNUMBERED,
 517                 .procname       = "ftrace_dump_on_oops",
 518                 .data           = &ftrace_dump_on_oops,
 519                 .maxlen         = sizeof(int),
 520                 .mode           = 0644,
 521                 .proc_handler   = &proc_dointvec,
 522         },
 523 #endif
 524 #ifdef CONFIG_MODULES
 525         {
 526                 .ctl_name       = KERN_MODPROBE,
 527                 .procname       = "modprobe",
 528                 .data           = &modprobe_path,
 529                 .maxlen         = KMOD_PATH_LEN,
 530                 .mode           = 0644,
 531                 .proc_handler   = &proc_dostring,
 532                 .strategy       = &sysctl_string,
 533         },
 534 #endif
 535 #if defined(CONFIG_HOTPLUG) && defined(CONFIG_NET)
 536         {
 537                 .ctl_name       = KERN_HOTPLUG,
 538                 .procname       = "hotplug",
 539                 .data           = &uevent_helper,
 540                 .maxlen         = UEVENT_HELPER_PATH_LEN,
 541                 .mode           = 0644,
 542                 .proc_handler   = &proc_dostring,
 543                 .strategy       = &sysctl_string,
 544         },
 545 #endif
 546 #ifdef CONFIG_CHR_DEV_SG
 547         {
 548                 .ctl_name       = KERN_SG_BIG_BUFF,
 549                 .procname       = "sg-big-buff",
 550                 .data           = &sg_big_buff,
 551                 .maxlen         = sizeof (int),
 552                 .mode           = 0444,
 553                 .proc_handler   = &proc_dointvec,
 554         },
 555 #endif
 556 #ifdef CONFIG_BSD_PROCESS_ACCT
 557         {
 558                 .ctl_name       = KERN_ACCT,
 559                 .procname       = "acct",
 560                 .data           = &acct_parm,
 561                 .maxlen         = 3*sizeof(int),
 562                 .mode           = 0644,
 563                 .proc_handler   = &proc_dointvec,
 564         },
 565 #endif
 566 #ifdef CONFIG_MAGIC_SYSRQ
 567         {
 568                 .ctl_name       = KERN_SYSRQ,
 569                 .procname       = "sysrq",
 570                 .data           = &__sysrq_enabled,
 571                 .maxlen         = sizeof (int),
 572                 .mode           = 0644,
 573                 .proc_handler   = &proc_dointvec,
 574         },
 575 #endif
 576 #ifdef CONFIG_PROC_SYSCTL
 577         {
 578                 .procname       = "cad_pid",
 579                 .data           = NULL,
 580                 .maxlen         = sizeof (int),
 581                 .mode           = 0600,
 582                 .proc_handler   = &proc_do_cad_pid,
 583         },
 584 #endif
 585         {
 586                 .ctl_name       = KERN_MAX_THREADS,
 587                 .procname       = "threads-max",
 588                 .data           = &max_threads,
 589                 .maxlen         = sizeof(int),
 590                 .mode           = 0644,
 591                 .proc_handler   = &proc_dointvec,
 592         },
 593         {
 594                 .ctl_name       = KERN_RANDOM,
 595                 .procname       = "random",
 596                 .mode           = 0555,
 597                 .child          = random_table,
 598         },
 599         {
 600                 .ctl_name       = KERN_OVERFLOWUID,
 601                 .procname       = "overflowuid",
 602                 .data           = &overflowuid,
 603                 .maxlen         = sizeof(int),
 604                 .mode           = 0644,
 605                 .proc_handler   = &proc_dointvec_minmax,
 606                 .strategy       = &sysctl_intvec,
 607                 .extra1         = &minolduid,
 608                 .extra2         = &maxolduid,
 609         },
 610         {
 611                 .ctl_name       = KERN_OVERFLOWGID,
 612                 .procname       = "overflowgid",
 613                 .data           = &overflowgid,
 614                 .maxlen         = sizeof(int),
 615                 .mode           = 0644,
 616                 .proc_handler   = &proc_dointvec_minmax,
 617                 .strategy       = &sysctl_intvec,
 618                 .extra1         = &minolduid,
 619                 .extra2         = &maxolduid,
 620         },
 621 #ifdef CONFIG_S390
 622 #ifdef CONFIG_MATHEMU
 623         {
 624                 .ctl_name       = KERN_IEEE_EMULATION_WARNINGS,
 625                 .procname       = "ieee_emulation_warnings",
 626                 .data           = &sysctl_ieee_emulation_warnings,
 627                 .maxlen         = sizeof(int),
 628                 .mode           = 0644,
 629                 .proc_handler   = &proc_dointvec,
 630         },
 631 #endif
 632         {
 633                 .ctl_name       = KERN_S390_USER_DEBUG_LOGGING,
 634                 .procname       = "userprocess_debug",
 635                 .data           = &sysctl_userprocess_debug,
 636                 .maxlen         = sizeof(int),
 637                 .mode           = 0644,
 638                 .proc_handler   = &proc_dointvec,
 639         },
 640 #endif
 641         {
 642                 .ctl_name       = KERN_PIDMAX,
 643                 .procname       = "pid_max",
 644                 .data           = &pid_max,
 645                 .maxlen         = sizeof (int),
 646                 .mode           = 0644,
 647                 .proc_handler   = &proc_dointvec_minmax,
 648                 .strategy       = sysctl_intvec,
 649                 .extra1         = &pid_max_min,
 650                 .extra2         = &pid_max_max,
 651         },
 652         {
 653                 .ctl_name       = KERN_PANIC_ON_OOPS,
 654                 .procname       = "panic_on_oops",
 655                 .data           = &panic_on_oops,
 656                 .maxlen         = sizeof(int),
 657                 .mode           = 0644,
 658                 .proc_handler   = &proc_dointvec,
 659         },
 660 #if defined CONFIG_PRINTK
 661         {
 662                 .ctl_name       = KERN_PRINTK,
 663                 .procname       = "printk",
 664                 .data           = &console_loglevel,
 665                 .maxlen         = 4*sizeof(int),
 666                 .mode           = 0644,
 667                 .proc_handler   = &proc_dointvec,
 668         },
 669         {
 670                 .ctl_name       = KERN_PRINTK_RATELIMIT,
 671                 .procname       = "printk_ratelimit",
 672                 .data           = &printk_ratelimit_state.interval,
 673                 .maxlen         = sizeof(int),
 674                 .mode           = 0644,
 675                 .proc_handler   = &proc_dointvec_jiffies,
 676                 .strategy       = &sysctl_jiffies,
 677         },
 678         {
 679                 .ctl_name       = KERN_PRINTK_RATELIMIT_BURST,
 680                 .procname       = "printk_ratelimit_burst",
 681                 .data           = &printk_ratelimit_state.burst,
 682                 .maxlen         = sizeof(int),
 683                 .mode           = 0644,
 684                 .proc_handler   = &proc_dointvec,
 685         },
 686 #endif
 687         {
 688                 .ctl_name       = KERN_NGROUPS_MAX,
 689                 .procname       = "ngroups_max",
 690                 .data           = &ngroups_max,
 691                 .maxlen         = sizeof (int),
 692                 .mode           = 0444,
 693                 .proc_handler   = &proc_dointvec,
 694         },
 695 #if defined(CONFIG_X86_LOCAL_APIC) && defined(CONFIG_X86)
 696         {
 697                 .ctl_name       = KERN_UNKNOWN_NMI_PANIC,
 698                 .procname       = "unknown_nmi_panic",
 699                 .data           = &unknown_nmi_panic,
 700                 .maxlen         = sizeof (int),
 701                 .mode           = 0644,
 702                 .proc_handler   = &proc_dointvec,
 703         },
 704         {
 705                 .procname       = "nmi_watchdog",
 706                 .data           = &nmi_watchdog_enabled,
 707                 .maxlen         = sizeof (int),
 708                 .mode           = 0644,
 709                 .proc_handler   = &proc_nmi_enabled,
 710         },
 711 #endif
 712 #if defined(CONFIG_X86)
 713         {
 714                 .ctl_name       = KERN_PANIC_ON_NMI,
 715                 .procname       = "panic_on_unrecovered_nmi",
 716                 .data           = &panic_on_unrecovered_nmi,
 717                 .maxlen         = sizeof(int),
 718                 .mode           = 0644,
 719                 .proc_handler   = &proc_dointvec,
 720         },
 721         {
 722                 .ctl_name       = KERN_BOOTLOADER_TYPE,
 723                 .procname       = "bootloader_type",
 724                 .data           = &bootloader_type,
 725                 .maxlen         = sizeof (int),
 726                 .mode           = 0444,
 727                 .proc_handler   = &proc_dointvec,
 728         },
 729         {
 730                 .ctl_name       = CTL_UNNUMBERED,
 731                 .procname       = "kstack_depth_to_print",
 732                 .data           = &kstack_depth_to_print,
 733                 .maxlen         = sizeof(int),
 734                 .mode           = 0644,
 735                 .proc_handler   = &proc_dointvec,
 736         },
 737         {
 738                 .ctl_name       = CTL_UNNUMBERED,
 739                 .procname       = "io_delay_type",
 740                 .data           = &io_delay_type,
 741                 .maxlen         = sizeof(int),
 742                 .mode           = 0644,
 743                 .proc_handler   = &proc_dointvec,
 744         },
 745 #endif
 746 #if defined(CONFIG_MMU)
 747         {
 748                 .ctl_name       = KERN_RANDOMIZE,
 749                 .procname       = "randomize_va_space",
 750                 .data           = &randomize_va_space,
 751                 .maxlen         = sizeof(int),
 752                 .mode           = 0644,
 753                 .proc_handler   = &proc_dointvec,
 754         },
 755 #endif
 756 #if defined(CONFIG_S390) && defined(CONFIG_SMP)
 757         {
 758                 .ctl_name       = KERN_SPIN_RETRY,
 759                 .procname       = "spin_retry",
 760                 .data           = &spin_retry,
 761                 .maxlen         = sizeof (int),
 762                 .mode           = 0644,
 763                 .proc_handler   = &proc_dointvec,
 764         },
 765 #endif
 766 #if     defined(CONFIG_ACPI_SLEEP) && defined(CONFIG_X86)
 767         {
 768                 .procname       = "acpi_video_flags",
 769                 .data           = &acpi_realmode_flags,
 770                 .maxlen         = sizeof (unsigned long),
 771                 .mode           = 0644,
 772                 .proc_handler   = &proc_doulongvec_minmax,
 773         },
 774 #endif
 775 #ifdef CONFIG_IA64
 776         {
 777                 .ctl_name       = KERN_IA64_UNALIGNED,
 778                 .procname       = "ignore-unaligned-usertrap",
 779                 .data           = &no_unaligned_warning,
 780                 .maxlen         = sizeof (int),
 781                 .mode           = 0644,
 782                 .proc_handler   = &proc_dointvec,
 783         },
 784         {
 785                 .ctl_name       = CTL_UNNUMBERED,
 786                 .procname       = "unaligned-dump-stack",
 787                 .data           = &unaligned_dump_stack,
 788                 .maxlen         = sizeof (int),
 789                 .mode           = 0644,
 790                 .proc_handler   = &proc_dointvec,
 791         },
 792 #endif
 793 #ifdef CONFIG_DETECT_SOFTLOCKUP
 794         {
 795                 .ctl_name       = CTL_UNNUMBERED,
 796                 .procname       = "softlockup_panic",
 797                 .data           = &softlockup_panic,
 798                 .maxlen         = sizeof(int),
 799                 .mode           = 0644,
 800                 .proc_handler   = &proc_dointvec_minmax,
 801                 .strategy       = &sysctl_intvec,
 802                 .extra1         = &zero,
 803                 .extra2         = &one,
 804         },
 805         {
 806                 .ctl_name       = CTL_UNNUMBERED,
 807                 .procname       = "softlockup_thresh",
 808                 .data           = &softlockup_thresh,
 809                 .maxlen         = sizeof(int),
 810                 .mode           = 0644,
 811                 .proc_handler   = &proc_dosoftlockup_thresh,
 812                 .strategy       = &sysctl_intvec,
 813                 .extra1         = &neg_one,
 814                 .extra2         = &sixty,
 815         },
 816         {
 817                 .ctl_name       = CTL_UNNUMBERED,
 818                 .procname       = "hung_task_check_count",
 819                 .data           = &sysctl_hung_task_check_count,
 820                 .maxlen         = sizeof(unsigned long),
 821                 .mode           = 0644,
 822                 .proc_handler   = &proc_doulongvec_minmax,
 823                 .strategy       = &sysctl_intvec,
 824         },
 825         {
 826                 .ctl_name       = CTL_UNNUMBERED,
 827                 .procname       = "hung_task_timeout_secs",
 828                 .data           = &sysctl_hung_task_timeout_secs,
 829                 .maxlen         = sizeof(unsigned long),
 830                 .mode           = 0644,
 831                 .proc_handler   = &proc_doulongvec_minmax,
 832                 .strategy       = &sysctl_intvec,
 833         },
 834         {
 835                 .ctl_name       = CTL_UNNUMBERED,
 836                 .procname       = "hung_task_warnings",
 837                 .data           = &sysctl_hung_task_warnings,
 838                 .maxlen         = sizeof(unsigned long),
 839                 .mode           = 0644,
 840                 .proc_handler   = &proc_doulongvec_minmax,
 841                 .strategy       = &sysctl_intvec,
 842         },
 843 #endif
 844 #ifdef CONFIG_COMPAT
 845         {
 846                 .ctl_name       = KERN_COMPAT_LOG,
 847                 .procname       = "compat-log",
 848                 .data           = &compat_log,
 849                 .maxlen         = sizeof (int),
 850                 .mode           = 0644,
 851                 .proc_handler   = &proc_dointvec,
 852         },
 853 #endif
 854 #ifdef CONFIG_RT_MUTEXES
 855         {
 856                 .ctl_name       = KERN_MAX_LOCK_DEPTH,
 857                 .procname       = "max_lock_depth",
 858                 .data           = &max_lock_depth,
 859                 .maxlen         = sizeof(int),
 860                 .mode           = 0644,
 861                 .proc_handler   = &proc_dointvec,
 862         },
 863 #endif
 864         {
 865                 .ctl_name       = CTL_UNNUMBERED,
 866                 .procname       = "poweroff_cmd",
 867                 .data           = &poweroff_cmd,
 868                 .maxlen         = POWEROFF_CMD_PATH_LEN,
 869                 .mode           = 0644,
 870                 .proc_handler   = &proc_dostring,
 871                 .strategy       = &sysctl_string,
 872         },
 873 #ifdef CONFIG_KEYS
 874         {
 875                 .ctl_name       = CTL_UNNUMBERED,
 876                 .procname       = "keys",
 877                 .mode           = 0555,
 878                 .child          = key_sysctls,
 879         },
 880 #endif
 881 #ifdef CONFIG_RCU_TORTURE_TEST
 882         {
 883                 .ctl_name       = CTL_UNNUMBERED,
 884                 .procname       = "rcutorture_runnable",
 885                 .data           = &rcutorture_runnable,
 886                 .maxlen         = sizeof(int),
 887                 .mode           = 0644,
 888                 .proc_handler   = &proc_dointvec,
 889         },
 890 #endif
 891 #ifdef CONFIG_UNEVICTABLE_LRU
 892         {
 893                 .ctl_name       = CTL_UNNUMBERED,
 894                 .procname       = "scan_unevictable_pages",
 895                 .data           = &scan_unevictable_pages,
 896                 .maxlen         = sizeof(scan_unevictable_pages),
 897                 .mode           = 0644,
 898                 .proc_handler   = &scan_unevictable_handler,
 899         },
 900 #endif
 901 #ifdef CONFIG_SLOW_WORK
 902         {
 903                 .ctl_name       = CTL_UNNUMBERED,
 904                 .procname       = "slow-work",
 905                 .mode           = 0555,
 906                 .child          = slow_work_sysctls,
 907         },
 908 #endif
 909 /*
 910  * NOTE: do not add new entries to this table unless you have read
 911  * Documentation/sysctl/ctl_unnumbered.txt
 912  */
 913         { .ctl_name = 0 }
 914 };
 915
 916 static struct ctl_table vm_table[] = {
 917         {
 918                 .ctl_name       = VM_OVERCOMMIT_MEMORY,
 919                 .procname       = "overcommit_memory",
 920                 .data           = &sysctl_overcommit_memory,
 921                 .maxlen         = sizeof(sysctl_overcommit_memory),
 922                 .mode           = 0644,
 923                 .proc_handler   = &proc_dointvec,
 924         },
 925         {
 926                 .ctl_name       = VM_PANIC_ON_OOM,
 927                 .procname       = "panic_on_oom",
 928                 .data           = &sysctl_panic_on_oom,
 929                 .maxlen         = sizeof(sysctl_panic_on_oom),
 930                 .mode           = 0644,
 931                 .proc_handler   = &proc_dointvec,
 932         },
 933         {
 934                 .ctl_name       = CTL_UNNUMBERED,
 935                 .procname       = "oom_kill_allocating_task",
 936                 .data           = &sysctl_oom_kill_allocating_task,
 937                 .maxlen         = sizeof(sysctl_oom_kill_allocating_task),
 938                 .mode           = 0644,
 939                 .proc_handler   = &proc_dointvec,
 940         },
 941         {
 942                 .ctl_name       = CTL_UNNUMBERED,
 943                 .procname       = "oom_dump_tasks",
 944                 .data           = &sysctl_oom_dump_tasks,
 945                 .maxlen         = sizeof(sysctl_oom_dump_tasks),
 946                 .mode           = 0644,
 947                 .proc_handler   = &proc_dointvec,
 948         },
 949         {
 950                 .ctl_name       = VM_OVERCOMMIT_RATIO,
 951                 .procname       = "overcommit_ratio",
 952                 .data           = &sysctl_overcommit_ratio,
 953                 .maxlen         = sizeof(sysctl_overcommit_ratio),
 954                 .mode           = 0644,
 955                 .proc_handler   = &proc_dointvec,
 956         },
 957         {
 958                 .ctl_name       = VM_PAGE_CLUSTER,
 959                 .procname       = "page-cluster",
 960                 .data           = &page_cluster,
 961                 .maxlen         = sizeof(int),
 962                 .mode           = 0644,
 963                 .proc_handler   = &proc_dointvec,
 964         },
 965         {
 966                 .ctl_name       = VM_DIRTY_BACKGROUND,
 967                 .procname       = "dirty_background_ratio",
 968                 .data           = &dirty_background_ratio,
 969                 .maxlen         = sizeof(dirty_background_ratio),
 970                 .mode           = 0644,
 971                 .proc_handler   = &dirty_background_ratio_handler,
 972                 .strategy       = &sysctl_intvec,
 973                 .extra1         = &zero,
 974                 .extra2         = &one_hundred,
 975         },
 976         {
 977                 .ctl_name       = CTL_UNNUMBERED,
 978                 .procname       = "dirty_background_bytes",
 979                 .data           = &dirty_background_bytes,
 980                 .maxlen         = sizeof(dirty_background_bytes),
 981                 .mode           = 0644,
 982                 .proc_handler   = &dirty_background_bytes_handler,
 983                 .strategy       = &sysctl_intvec,
 984                 .extra1         = &one_ul,
 985         },
 986         {
 987                 .ctl_name       = VM_DIRTY_RATIO,
 988                 .procname       = "dirty_ratio",
 989                 .data           = &vm_dirty_ratio,
 990                 .maxlen         = sizeof(vm_dirty_ratio),
 991                 .mode           = 0644,
 992                 .proc_handler   = &dirty_ratio_handler,
 993                 .strategy       = &sysctl_intvec,
 994                 .extra1         = &zero,
 995                 .extra2         = &one_hundred,
 996         },
 997         {
 998                 .ctl_name       = CTL_UNNUMBERED,
 999                 .procname       = "dirty_bytes",
1000                 .data           = &vm_dirty_bytes,
1001                 .maxlen         = sizeof(vm_dirty_bytes),
1002                 .mode           = 0644,
1003                 .proc_handler   = &dirty_bytes_handler,
1004                 .strategy       = &sysctl_intvec,
1005                 .extra1         = &one_ul,
1006         },
1007         {
1008                 .procname       = "dirty_writeback_centisecs",
1009                 .data           = &dirty_writeback_interval,
1010                 .maxlen         = sizeof(dirty_writeback_interval),
1011                 .mode           = 0644,
1012                 .proc_handler   = &dirty_writeback_centisecs_handler,
1013         },
1014         {
1015                 .procname       = "dirty_expire_centisecs",
1016                 .data           = &dirty_expire_interval,
1017                 .maxlen         = sizeof(dirty_expire_interval),
1018                 .mode           = 0644,
1019                 .proc_handler   = &proc_dointvec,
1020         },
1021         {
1022                 .ctl_name       = VM_NR_PDFLUSH_THREADS,
1023                 .procname       = "nr_pdflush_threads",
1024                 .data           = &nr_pdflush_threads,
1025                 .maxlen         = sizeof nr_pdflush_threads,
1026                 .mode           = 0444 /* read-only*/,
1027                 .proc_handler   = &proc_dointvec,
1028         },
1029         {
1030                 .ctl_name       = VM_SWAPPINESS,
1031                 .procname       = "swappiness",
1032                 .data           = &vm_swappiness,
1033                 .maxlen         = sizeof(vm_swappiness),
1034                 .mode           = 0644,
1035                 .proc_handler   = &proc_dointvec_minmax,
1036                 .strategy       = &sysctl_intvec,
1037                 .extra1         = &zero,
1038                 .extra2         = &one_hundred,
1039         },
1040 #ifdef CONFIG_HUGETLB_PAGE
1041          {
1042                 .procname       = "nr_hugepages",
1043                 .data           = NULL,
1044                 .maxlen         = sizeof(unsigned long),
1045                 .mode           = 0644,
1046                 .proc_handler   = &hugetlb_sysctl_handler,
1047                 .extra1         = (void *)&hugetlb_zero,
1048                 .extra2         = (void *)&hugetlb_infinity,
1049          },
1050          {
1051                 .ctl_name       = VM_HUGETLB_GROUP,
1052                 .procname       = "hugetlb_shm_group",
1053                 .data           = &sysctl_hugetlb_shm_group,
1054                 .maxlen         = sizeof(gid_t),
1055                 .mode           = 0644,
1056                 .proc_handler   = &proc_dointvec,
1057          },
1058          {
1059                 .ctl_name       = CTL_UNNUMBERED,
1060                 .procname       = "hugepages_treat_as_movable",
1061                 .data           = &hugepages_treat_as_movable,
1062                 .maxlen         = sizeof(int),
1063                 .mode           = 0644,
1064                 .proc_handler   = &hugetlb_treat_movable_handler,
1065         },
1066         {
1067                 .ctl_name       = CTL_UNNUMBERED,
1068                 .procname       = "nr_overcommit_hugepages",
1069                 .data           = NULL,
1070                 .maxlen         = sizeof(unsigned long),
1071                 .mode           = 0644,
1072                 .proc_handler   = &hugetlb_overcommit_handler,
1073                 .extra1         = (void *)&hugetlb_zero,
1074                 .extra2         = (void *)&hugetlb_infinity,
1075         },
1076 #endif
1077         {
1078                 .ctl_name       = VM_LOWMEM_RESERVE_RATIO,
1079                 .procname       = "lowmem_reserve_ratio",
1080                 .data           = &sysctl_lowmem_reserve_ratio,
1081                 .maxlen         = sizeof(sysctl_lowmem_reserve_ratio),
1082                 .mode           = 0644,
1083                 .proc_handler   = &lowmem_reserve_ratio_sysctl_handler,
1084                 .strategy       = &sysctl_intvec,
1085         },
1086         {
1087                 .ctl_name       = VM_DROP_PAGECACHE,
1088                 .procname       = "drop_caches",
1089                 .data           = &sysctl_drop_caches,
1090                 .maxlen         = sizeof(int),
1091                 .mode           = 0644,
1092                 .proc_handler   = drop_caches_sysctl_handler,
1093                 .strategy       = &sysctl_intvec,
1094         },
1095         {
1096                 .ctl_name       = VM_MIN_FREE_KBYTES,
1097                 .procname       = "min_free_kbytes",
1098                 .data           = &min_free_kbytes,
1099                 .maxlen         = sizeof(min_free_kbytes),
1100                 .mode           = 0644,
1101                 .proc_handler   = &min_free_kbytes_sysctl_handler,
1102                 .strategy       = &sysctl_intvec,
1103                 .extra1         = &zero,
1104         },
1105         {
1106                 .ctl_name       = VM_PERCPU_PAGELIST_FRACTION,
1107                 .procname       = "percpu_pagelist_fraction",
1108                 .data           = &percpu_pagelist_fraction,
1109                 .maxlen         = sizeof(percpu_pagelist_fraction),
1110                 .mode           = 0644,
1111                 .proc_handler   = &percpu_pagelist_fraction_sysctl_handler,
1112                 .strategy       = &sysctl_intvec,
1113                 .extra1         = &min_percpu_pagelist_fract,
1114         },
1115 #ifdef CONFIG_MMU
1116         {
1117                 .ctl_name       = VM_MAX_MAP_COUNT,
1118                 .procname       = "max_map_count",
1119                 .data           = &sysctl_max_map_count,
1120                 .maxlen         = sizeof(sysctl_max_map_count),
1121                 .mode           = 0644,
1122                 .proc_handler   = &proc_dointvec
1123         },
1124 #else
1125         {
1126                 .ctl_name       = CTL_UNNUMBERED,
1127                 .procname       = "nr_trim_pages",
1128                 .data           = &sysctl_nr_trim_pages,
1129                 .maxlen         = sizeof(sysctl_nr_trim_pages),
1130                 .mode           = 0644,
1131                 .proc_handler   = &proc_dointvec_minmax,
1132                 .strategy       = &sysctl_intvec,
1133                 .extra1         = &zero,
1134         },
1135 #endif
1136         {
1137                 .ctl_name       = VM_LAPTOP_MODE,
1138                 .procname       = "laptop_mode",
1139                 .data           = &laptop_mode,
1140                 .maxlen         = sizeof(laptop_mode),
1141                 .mode           = 0644,
1142                 .proc_handler   = &proc_dointvec_jiffies,
1143                 .strategy       = &sysctl_jiffies,
1144         },
1145         {
1146                 .ctl_name       = VM_BLOCK_DUMP,
1147                 .procname       = "block_dump",
1148                 .data           = &block_dump,
1149                 .maxlen         = sizeof(block_dump),
1150                 .mode           = 0644,
1151                 .proc_handler   = &proc_dointvec,
1152                 .strategy       = &sysctl_intvec,
1153                 .extra1         = &zero,
1154         },
1155         {
1156                 .ctl_name       = VM_VFS_CACHE_PRESSURE,
1157                 .procname       = "vfs_cache_pressure",
1158                 .data           = &sysctl_vfs_cache_pressure,
1159                 .maxlen         = sizeof(sysctl_vfs_cache_pressure),
1160                 .mode           = 0644,
1161                 .proc_handler   = &proc_dointvec,
1162                 .strategy       = &sysctl_intvec,
1163                 .extra1         = &zero,
1164         },
1165 #ifdef HAVE_ARCH_PICK_MMAP_LAYOUT
1166         {
1167                 .ctl_name       = VM_LEGACY_VA_LAYOUT,
1168                 .procname       = "legacy_va_layout",
1169                 .data           = &sysctl_legacy_va_layout,
1170                 .maxlen         = sizeof(sysctl_legacy_va_layout),
1171                 .mode           = 0644,
1172                 .proc_handler   = &proc_dointvec,
1173                 .strategy       = &sysctl_intvec,
1174                 .extra1         = &zero,
1175         },
1176 #endif
1177 #ifdef CONFIG_NUMA
1178         {
1179                 .ctl_name       = VM_ZONE_RECLAIM_MODE,
1180                 .procname       = "zone_reclaim_mode",
1181                 .data           = &zone_reclaim_mode,
1182                 .maxlen         = sizeof(zone_reclaim_mode),
1183                 .mode           = 0644,
1184                 .proc_handler   = &proc_dointvec,
1185                 .strategy       = &sysctl_intvec,
1186                 .extra1         = &zero,
1187         },
1188         {
1189                 .ctl_name       = VM_MIN_UNMAPPED,
1190                 .procname       = "min_unmapped_ratio",
1191                 .data           = &sysctl_min_unmapped_ratio,
1192                 .maxlen         = sizeof(sysctl_min_unmapped_ratio),
1193                 .mode           = 0644,
1194                 .proc_handler   = &sysctl_min_unmapped_ratio_sysctl_handler,
1195                 .strategy       = &sysctl_intvec,
1196                 .extra1         = &zero,
1197                 .extra2         = &one_hundred,
1198         },
1199         {
1200                 .ctl_name       = VM_MIN_SLAB,
1201                 .procname       = "min_slab_ratio",
1202                 .data           = &sysctl_min_slab_ratio,
1203                 .maxlen         = sizeof(sysctl_min_slab_ratio),
1204                 .mode           = 0644,
1205                 .proc_handler   = &sysctl_min_slab_ratio_sysctl_handler,
1206                 .strategy       = &sysctl_intvec,
1207                 .extra1         = &zero,
1208                 .extra2         = &one_hundred,
1209         },
1210 #endif
1211 #ifdef CONFIG_SMP
1212         {
1213                 .ctl_name       = CTL_UNNUMBERED,
1214                 .procname       = "stat_interval",
1215                 .data           = &sysctl_stat_interval,
1216                 .maxlen         = sizeof(sysctl_stat_interval),
1217                 .mode           = 0644,
1218                 .proc_handler   = &proc_dointvec_jiffies,
1219                 .strategy       = &sysctl_jiffies,
1220         },
1221 #endif
1222 #ifdef CONFIG_SECURITY
1223         {
1224                 .ctl_name       = CTL_UNNUMBERED,
1225                 .procname       = "mmap_min_addr",
1226                 .data           = &mmap_min_addr,
1227                 .maxlen         = sizeof(unsigned long),
1228                 .mode           = 0644,
1229                 .proc_handler   = &proc_doulongvec_minmax,
1230         },
1231 #endif
1232 #ifdef CONFIG_NUMA
1233         {
1234                 .ctl_name       = CTL_UNNUMBERED,
1235                 .procname       = "numa_zonelist_order",
1236                 .data           = &numa_zonelist_order,
1237                 .maxlen         = NUMA_ZONELIST_ORDER_LEN,
1238                 .mode           = 0644,
1239                 .proc_handler   = &numa_zonelist_order_handler,
1240                 .strategy       = &sysctl_string,
1241         },
1242 #endif
1243 #if (defined(CONFIG_X86_32) && !defined(CONFIG_UML))|| \
1244    (defined(CONFIG_SUPERH) && defined(CONFIG_VSYSCALL))
1245         {
1246                 .ctl_name       = VM_VDSO_ENABLED,
1247                 .procname       = "vdso_enabled",
1248                 .data           = &vdso_enabled,
1249                 .maxlen         = sizeof(vdso_enabled),
1250                 .mode           = 0644,
1251                 .proc_handler   = &proc_dointvec,
1252                 .strategy       = &sysctl_intvec,
1253                 .extra1         = &zero,
1254         },
1255 #endif
1256 #ifdef CONFIG_HIGHMEM
1257         {
1258                 .ctl_name       = CTL_UNNUMBERED,
1259                 .procname       = "highmem_is_dirtyable",
1260                 .data           = &vm_highmem_is_dirtyable,
1261                 .maxlen         = sizeof(vm_highmem_is_dirtyable),
1262                 .mode           = 0644,
1263                 .proc_handler   = &proc_dointvec_minmax,
1264                 .strategy       = &sysctl_intvec,
1265                 .extra1         = &zero,
1266                 .extra2         = &one,
1267         },
1268 #endif
1269 /*
1270  * NOTE: do not add new entries to this table unless you have read
1271  * Documentation/sysctl/ctl_unnumbered.txt
1272  */
1273         { .ctl_name = 0 }
1274 };
1275
1276 #if defined(CONFIG_BINFMT_MISC) || defined(CONFIG_BINFMT_MISC_MODULE)
1277 static struct ctl_table binfmt_misc_table[] = {
1278         { .ctl_name = 0 }
1279 };
1280 #endif
1281
1282 static struct ctl_table fs_table[] = {
1283         {
1284                 .ctl_name       = FS_NRINODE,
1285                 .procname       = "inode-nr",
1286                 .data           = &inodes_stat,
1287                 .maxlen         = 2*sizeof(int),
1288                 .mode           = 0444,
1289                 .proc_handler   = &proc_dointvec,
1290         },
1291         {
1292                 .ctl_name       = FS_STATINODE,
1293                 .procname       = "inode-state",
1294                 .data           = &inodes_stat,
1295                 .maxlen         = 7*sizeof(int),
1296                 .mode           = 0444,
1297                 .proc_handler   = &proc_dointvec,
1298         },
1299         {
1300                 .procname       = "file-nr",
1301                 .data           = &files_stat,
1302                 .maxlen         = 3*sizeof(int),
1303                 .mode           = 0444,
1304                 .proc_handler   = &proc_nr_files,
1305         },
1306         {
1307                 .ctl_name       = FS_MAXFILE,
1308                 .procname       = "file-max",
1309                 .data           = &files_stat.max_files,
1310                 .maxlen         = sizeof(int),
1311                 .mode           = 0644,
1312                 .proc_handler   = &proc_dointvec,
1313         },
1314         {
1315                 .ctl_name       = CTL_UNNUMBERED,
1316                 .procname       = "nr_open",
1317                 .data           = &sysctl_nr_open,
1318                 .maxlen         = sizeof(int),
1319                 .mode           = 0644,
1320                 .proc_handler   = &proc_dointvec_minmax,
1321                 .extra1         = &sysctl_nr_open_min,
1322                 .extra2         = &sysctl_nr_open_max,
1323         },
1324         {
1325                 .ctl_name       = FS_DENTRY,
1326                 .procname       = "dentry-state",
1327                 .data           = &dentry_stat,
1328                 .maxlen         = 6*sizeof(int),
1329                 .mode           = 0444,
1330                 .proc_handler   = &proc_dointvec,
1331         },
1332         {
1333                 .ctl_name       = FS_OVERFLOWUID,
1334                 .procname       = "overflowuid",
1335                 .data           = &fs_overflowuid,
1336                 .maxlen         = sizeof(int),
1337                 .mode           = 0644,
1338                 .proc_handler   = &proc_dointvec_minmax,
1339                 .strategy       = &sysctl_intvec,
1340                 .extra1         = &minolduid,
1341                 .extra2         = &maxolduid,
1342         },
1343         {
1344                 .ctl_name       = FS_OVERFLOWGID,
1345                 .procname       = "overflowgid",
1346                 .data           = &fs_overflowgid,
1347                 .maxlen         = sizeof(int),
1348                 .mode           = 0644,
1349                 .proc_handler   = &proc_dointvec_minmax,
1350                 .strategy       = &sysctl_intvec,
1351                 .extra1         = &minolduid,
1352                 .extra2         = &maxolduid,
1353         },
1354 #ifdef CONFIG_FILE_LOCKING
1355         {
1356                 .ctl_name       = FS_LEASES,
1357                 .procname       = "leases-enable",
1358                 .data           = &leases_enable,
1359                 .maxlen         = sizeof(int),
1360                 .mode           = 0644,
1361                 .proc_handler   = &proc_dointvec,
1362         },
1363 #endif
1364 #ifdef CONFIG_DNOTIFY
1365         {
1366                 .ctl_name       = FS_DIR_NOTIFY,
1367                 .procname       = "dir-notify-enable",
1368                 .data           = &dir_notify_enable,
1369                 .maxlen         = sizeof(int),
1370                 .mode           = 0644,
1371                 .proc_handler   = &proc_dointvec,
1372         },
1373 #endif
1374 #ifdef CONFIG_MMU
1375 #ifdef CONFIG_FILE_LOCKING
1376         {
1377                 .ctl_name       = FS_LEASE_TIME,
1378                 .procname       = "lease-break-time",
1379                 .data           = &lease_break_time,
1380                 .maxlen         = sizeof(int),
1381                 .mode           = 0644,
1382                 .proc_handler   = &proc_dointvec,
1383         },
1384 #endif
1385 #ifdef CONFIG_AIO
1386         {
1387                 .procname       = "aio-nr",
1388                 .data           = &aio_nr,
1389                 .maxlen         = sizeof(aio_nr),
1390                 .mode           = 0444,
1391                 .proc_handler   = &proc_doulongvec_minmax,
1392         },
1393         {
1394                 .procname       = "aio-max-nr",
1395                 .data           = &aio_max_nr,
1396                 .maxlen         = sizeof(aio_max_nr),
1397                 .mode           = 0644,
1398                 .proc_handler   = &proc_doulongvec_minmax,
1399         },
1400 #endif /* CONFIG_AIO */
1401 #ifdef CONFIG_INOTIFY_USER
1402         {
1403                 .ctl_name       = FS_INOTIFY,
1404                 .procname       = "inotify",
1405                 .mode           = 0555,
1406                 .child          = inotify_table,
1407         },
1408 #endif
1409 #ifdef CONFIG_EPOLL
1410         {
1411                 .procname       = "epoll",
1412                 .mode           = 0555,
1413                 .child          = epoll_table,
1414         },
1415 #endif
1416 #endif
1417         {
1418                 .ctl_name       = KERN_SETUID_DUMPABLE,
1419                 .procname       = "suid_dumpable",
1420                 .data           = &suid_dumpable,
1421                 .maxlen         = sizeof(int),
1422                 .mode           = 0644,
1423                 .proc_handler   = &proc_dointvec_minmax,
1424                 .strategy       = &sysctl_intvec,
1425                 .extra1         = &zero,
1426                 .extra2         = &two,
1427         },
1428 #if defined(CONFIG_BINFMT_MISC) || defined(CONFIG_BINFMT_MISC_MODULE)
1429         {
1430                 .ctl_name       = CTL_UNNUMBERED,
1431                 .procname       = "binfmt_misc",
1432                 .mode           = 0555,
1433                 .child          = binfmt_misc_table,
1434         },
1435 #endif
1436 /*
1437  * NOTE: do not add new entries to this table unless you have read
1438  * Documentation/sysctl/ctl_unnumbered.txt
1439  */
1440         { .ctl_name = 0 }
1441 };
1442
1443 static struct ctl_table debug_table[] = {
1444 #if defined(CONFIG_X86) || defined(CONFIG_PPC)
1445         {
1446                 .ctl_name       = CTL_UNNUMBERED,
1447                 .procname       = "exception-trace",
1448                 .data           = &show_unhandled_signals,
1449                 .maxlen         = sizeof(int),
1450                 .mode           = 0644,
1451                 .proc_handler   = proc_dointvec
1452         },
1453 #endif
1454         { .ctl_name = 0 }
1455 };
1456
1457 static struct ctl_table dev_table[] = {
1458         { .ctl_name = 0 }
1459 };
1460
1461 static DEFINE_SPINLOCK(sysctl_lock);
1462
1463 /* called under sysctl_lock */
1464 static int use_table(struct ctl_table_header *p)
1465 {
1466         if (unlikely(p->unregistering))
1467                 return 0;
1468         p->used++;
1469         return 1;
1470 }
1471
1472 /* called under sysctl_lock */
1473 static void unuse_table(struct ctl_table_header *p)
1474 {
1475         if (!--p->used)
1476                 if (unlikely(p->unregistering))
1477                         complete(p->unregistering);
1478 }
1479
1480 /* called under sysctl_lock, will reacquire if has to wait */
1481 static void start_unregistering(struct ctl_table_header *p)
1482 {
1483         /*
1484          * if p->used is 0, nobody will ever touch that entry again;
1485          * we'll eliminate all paths to it before dropping sysctl_lock
1486          */
1487         if (unlikely(p->used)) {
1488                 struct completion wait;
1489                 init_completion(&wait);
1490                 p->unregistering = &wait;
1491                 spin_unlock(&sysctl_lock);
1492                 wait_for_completion(&wait);
1493                 spin_lock(&sysctl_lock);
1494         } else {
1495                 /* anything non-NULL; we'll never dereference it */
1496                 p->unregistering = ERR_PTR(-EINVAL);
1497         }
1498         /*
1499          * do not remove from the list until nobody holds it; walking the
1500          * list in do_sysctl() relies on that.
1501          */
1502         list_del_init(&p->ctl_entry);
1503 }
1504
1505 void sysctl_head_get(struct ctl_table_header *head)
1506 {
1507         spin_lock(&sysctl_lock);
1508         head->count++;
1509         spin_unlock(&sysctl_lock);
1510 }
1511
1512 void sysctl_head_put(struct ctl_table_header *head)
1513 {
1514         spin_lock(&sysctl_lock);
1515         if (!--head->count)
1516                 kfree(head);
1517         spin_unlock(&sysctl_lock);
1518 }
1519
1520 struct ctl_table_header *sysctl_head_grab(struct ctl_table_header *head)
1521 {
1522         if (!head)
1523                 BUG();
1524         spin_lock(&sysctl_lock);
1525         if (!use_table(head))
1526                 head = ERR_PTR(-ENOENT);
1527         spin_unlock(&sysctl_lock);
1528         return head;
1529 }
1530
1531 void sysctl_head_finish(struct ctl_table_header *head)
1532 {
1533         if (!head)
1534                 return;
1535         spin_lock(&sysctl_lock);
1536         unuse_table(head);
1537         spin_unlock(&sysctl_lock);
1538 }
1539
1540 static struct ctl_table_set *
1541 lookup_header_set(struct ctl_table_root *root, struct nsproxy *namespaces)
1542 {
1543         struct ctl_table_set *set = &root->default_set;
1544         if (root->lookup)
1545                 set = root->lookup(root, namespaces);
1546         return set;
1547 }
1548
1549 static struct list_head *
1550 lookup_header_list(struct ctl_table_root *root, struct nsproxy *namespaces)
1551 {
1552         struct ctl_table_set *set = lookup_header_set(root, namespaces);
1553         return &set->list;
1554 }
1555
1556 struct ctl_table_header *__sysctl_head_next(struct nsproxy *namespaces,
1557                                             struct ctl_table_header *prev)
1558 {
1559         struct ctl_table_root *root;
1560         struct list_head *header_list;
1561         struct ctl_table_header *head;
1562         struct list_head *tmp;
1563
1564         spin_lock(&sysctl_lock);
1565         if (prev) {
1566                 head = prev;
1567                 tmp = &prev->ctl_entry;
1568                 unuse_table(prev);
1569                 goto next;
1570         }
1571         tmp = &root_table_header.ctl_entry;
1572         for (;;) {
1573                 head = list_entry(tmp, struct ctl_table_header, ctl_entry);
1574
1575                 if (!use_table(head))
1576                         goto next;
1577                 spin_unlock(&sysctl_lock);
1578                 return head;
1579         next:
1580                 root = head->root;
1581                 tmp = tmp->next;
1582                 header_list = lookup_header_list(root, namespaces);
1583                 if (tmp != header_list)
1584                         continue;
1585
1586                 do {
1587                         root = list_entry(root->root_list.next,
1588                                         struct ctl_table_root, root_list);
1589                         if (root == &sysctl_table_root)
1590                                 goto out;
1591                         header_list = lookup_header_list(root, namespaces);
1592                 } while (list_empty(header_list));
1593                 tmp = header_list->next;
1594         }
1595 out:
1596         spin_unlock(&sysctl_lock);
1597         return NULL;
1598 }
1599
1600 struct ctl_table_header *sysctl_head_next(struct ctl_table_header *prev)
1601 {
1602         return __sysctl_head_next(current->nsproxy, prev);
1603 }
1604
1605 void register_sysctl_root(struct ctl_table_root *root)
1606 {
1607         spin_lock(&sysctl_lock);
1608         list_add_tail(&root->root_list, &sysctl_table_root.root_list);
1609         spin_unlock(&sysctl_lock);
1610 }
1611
1612 #ifdef CONFIG_SYSCTL_SYSCALL
1613 /* Perform the actual read/write of a sysctl table entry. */
1614 static int do_sysctl_strategy(struct ctl_table_root *root,
1615                         struct ctl_table *table,
1616                         void __user *oldval, size_t __user *oldlenp,
1617                         void __user *newval, size_t newlen)
1618 {
1619         int op = 0, rc;
1620
1621         if (oldval)
1622                 op |= MAY_READ;
1623         if (newval)
1624                 op |= MAY_WRITE;
1625         if (sysctl_perm(root, table, op))
1626                 return -EPERM;
1627
1628         if (table->strategy) {
1629                 rc = table->strategy(table, oldval, oldlenp, newval, newlen);
1630                 if (rc < 0)
1631                         return rc;
1632                 if (rc > 0)
1633                         return 0;
1634         }
1635
1636         /* If there is no strategy routine, or if the strategy returns
1637          * zero, proceed with automatic r/w */
1638         if (table->data && table->maxlen) {
1639                 rc = sysctl_data(table, oldval, oldlenp, newval, newlen);
1640                 if (rc < 0)
1641                         return rc;
1642         }
1643         return 0;
1644 }
1645
1646 static int parse_table(int __user *name, int nlen,
1647                        void __user *oldval, size_t __user *oldlenp,
1648                        void __user *newval, size_t newlen,
1649                        struct ctl_table_root *root,
1650                        struct ctl_table *table)
1651 {
1652         int n;
1653 repeat:
1654         if (!nlen)
1655                 return -ENOTDIR;
1656         if (get_user(n, name))
1657                 return -EFAULT;
1658         for ( ; table->ctl_name || table->procname; table++) {
1659                 if (!table->ctl_name)
1660                         continue;
1661                 if (n == table->ctl_name) {
1662                         int error;
1663                         if (table->child) {
1664                                 if (sysctl_perm(root, table, MAY_EXEC))
1665                                         return -EPERM;
1666                                 name++;
1667                                 nlen--;
1668                                 table = table->child;
1669                                 goto repeat;
1670                         }
1671                         error = do_sysctl_strategy(root, table,
1672                                                    oldval, oldlenp,
1673                                                    newval, newlen);
1674                         return error;
1675                 }
1676         }
1677         return -ENOTDIR;
1678 }
1679
1680 int do_sysctl(int __user *name, int nlen, void __user *oldval, size_t __user *oldlenp,
1681                void __user *newval, size_t newlen)
1682 {
1683         struct ctl_table_header *head;
1684         int error = -ENOTDIR;
1685
1686         if (nlen <= 0 || nlen >= CTL_MAXNAME)
1687                 return -ENOTDIR;
1688         if (oldval) {
1689                 int old_len;
1690                 if (!oldlenp || get_user(old_len, oldlenp))
1691                         return -EFAULT;
1692         }
1693
1694         for (head = sysctl_head_next(NULL); head;
1695                         head = sysctl_head_next(head)) {
1696                 error = parse_table(name, nlen, oldval, oldlenp,
1697                                         newval, newlen,
1698                                         head->root, head->ctl_table);
1699                 if (error != -ENOTDIR) {
1700                         sysctl_head_finish(head);
1701                         break;
1702                 }
1703         }
1704         return error;
1705 }
1706
1707 SYSCALL_DEFINE1(sysctl, struct __sysctl_args __user *, args)
1708 {
1709         struct __sysctl_args tmp;
1710         int error;
1711
1712         if (copy_from_user(&tmp, args, sizeof(tmp)))
1713                 return -EFAULT;
1714
1715         error = deprecated_sysctl_warning(&tmp);
1716         if (error)
1717                 goto out;
1718
1719         lock_kernel();
1720         error = do_sysctl(tmp.name, tmp.nlen, tmp.oldval, tmp.oldlenp,
1721                           tmp.newval, tmp.newlen);
1722         unlock_kernel();
1723 out:
1724         return error;
1725 }
1726 #endif /* CONFIG_SYSCTL_SYSCALL */
1727
1728 /*
1729  * sysctl_perm does NOT grant the superuser all rights automatically, because
1730  * some sysctl variables are readonly even to root.
1731  */
1732
1733 static int test_perm(int mode, int op)
1734 {
1735         if (!current_euid())
1736                 mode >>= 6;
1737         else if (in_egroup_p(0))
1738                 mode >>= 3;
1739         if ((op & ~mode & (MAY_READ|MAY_WRITE|MAY_EXEC)) == 0)
1740                 return 0;
1741         return -EACCES;
1742 }
1743
1744 int sysctl_perm(struct ctl_table_root *root, struct ctl_table *table, int op)
1745 {
1746         int error;
1747         int mode;
1748
1749         error = security_sysctl(table, op & (MAY_READ | MAY_WRITE | MAY_EXEC));
1750         if (error)
1751                 return error;
1752
1753         if (root->permissions)
1754                 mode = root->permissions(root, current->nsproxy, table);
1755         else
1756                 mode = table->mode;
1757
1758         return test_perm(mode, op);
1759 }
1760
1761 static void sysctl_set_parent(struct ctl_table *parent, struct ctl_table *table)
1762 {
1763         for (; table->ctl_name || table->procname; table++) {
1764                 table->parent = parent;
1765                 if (table->child)
1766                         sysctl_set_parent(table, table->child);
1767         }
1768 }
1769
1770 static __init int sysctl_init(void)
1771 {
1772         sysctl_set_parent(NULL, root_table);
1773 #ifdef CONFIG_SYSCTL_SYSCALL_CHECK
1774         {
1775                 int err;
1776                 err = sysctl_check_table(current->nsproxy, root_table);
1777         }
1778 #endif
1779         return 0;
1780 }
1781
1782 core_initcall(sysctl_init);
1783
1784 static struct ctl_table *is_branch_in(struct ctl_table *branch,
1785                                       struct ctl_table *table)
1786 {
1787         struct ctl_table *p;
1788         const char *s = branch->procname;
1789
1790         /* branch should have named subdirectory as its first element */
1791         if (!s || !branch->child)
1792                 return NULL;
1793
1794         /* ... and nothing else */
1795         if (branch[1].procname || branch[1].ctl_name)
1796                 return NULL;
1797
1798         /* table should contain subdirectory with the same name */
1799         for (p = table; p->procname || p->ctl_name; p++) {
1800                 if (!p->child)
1801                         continue;
1802                 if (p->procname && strcmp(p->procname, s) == 0)
1803                         return p;
1804         }
1805         return NULL;
1806 }
1807
1808 /* see if attaching q to p would be an improvement */
1809 static void try_attach(struct ctl_table_header *p, struct ctl_table_header *q)
1810 {
1811         struct ctl_table *to = p->ctl_table, *by = q->ctl_table;
1812         struct ctl_table *next;
1813         int is_better = 0;
1814         int not_in_parent = !p->attached_by;
1815
1816         while ((next = is_branch_in(by, to)) != NULL) {
1817                 if (by == q->attached_by)
1818                         is_better = 1;
1819                 if (to == p->attached_by)
1820                         not_in_parent = 1;
1821                 by = by->child;
1822                 to = next->child;
1823         }
1824
1825         if (is_better && not_in_parent) {
1826                 q->attached_by = by;
1827                 q->attached_to = to;
1828                 q->parent = p;
1829         }
1830 }
1831
1832 /**
1833  * __register_sysctl_paths - register a sysctl hierarchy
1834  * @root: List of sysctl headers to register on
1835  * @namespaces: Data to compute which lists of sysctl entries are visible
1836  * @path: The path to the directory the sysctl table is in.
1837  * @table: the top-level table structure
1838  *
1839  * Register a sysctl table hierarchy. @table should be a filled in ctl_table
1840  * array. A completely 0 filled entry terminates the table.
1841  *
1842  * The members of the &struct ctl_table structure are used as follows:
1843  *
1844  * ctl_name - This is the numeric sysctl value used by sysctl(2). The number
1845  *            must be unique within that level of sysctl
1846  *
1847  * procname - the name of the sysctl file under /proc/sys. Set to %NULL to not
1848  *            enter a sysctl file
1849  *
1850  * data - a pointer to data for use by proc_handler
1851  *
1852  * maxlen - the maximum size in bytes of the data
1853  *
1854  * mode - the file permissions for the /proc/sys file, and for sysctl(2)
1855  *
1856  * child - a pointer to the child sysctl table if this entry is a directory, or
1857  *         %NULL.
1858  *
1859  * proc_handler - the text handler routine (described below)
1860  *
1861  * strategy - the strategy routine (described below)
1862  *
1863  * de - for internal use by the sysctl routines
1864  *
1865  * extra1, extra2 - extra pointers usable by the proc handler routines
1866  *
1867  * Leaf nodes in the sysctl tree will be represented by a single file
1868  * under /proc; non-leaf nodes will be represented by directories.
1869  *
1870  * sysctl(2) can automatically manage read and write requests through
1871  * the sysctl table.  The data and maxlen fields of the ctl_table
1872  * struct enable minimal validation of the values being written to be
1873  * performed, and the mode field allows minimal authentication.
1874  *
1875  * More sophisticated management can be enabled by the provision of a
1876  * strategy routine with the table entry.  This will be called before
1877  * any automatic read or write of the data is performed.
1878  *
1879  * The strategy routine may return
1880  *
1881  * < 0 - Error occurred (error is passed to user process)
1882  *
1883  * 0   - OK - proceed with automatic read or write.
1884  *
1885  * > 0 - OK - read or write has been done by the strategy routine, so
1886  *       return immediately.
1887  *
1888  * There must be a proc_handler routine for any terminal nodes
1889  * mirrored under /proc/sys (non-terminals are handled by a built-in
1890  * directory handler).  Several default handlers are available to
1891  * cover common cases -
1892  *
1893  * proc_dostring(), proc_dointvec(), proc_dointvec_jiffies(),
1894  * proc_dointvec_userhz_jiffies(), proc_dointvec_minmax(),
1895  * proc_doulongvec_ms_jiffies_minmax(), proc_doulongvec_minmax()
1896  *
1897  * It is the handler's job to read the input buffer from user memory
1898  * and process it. The handler should return 0 on success.
1899  *
1900  * This routine returns %NULL on a failure to register, and a pointer
1901  * to the table header on success.
1902  */
1903 struct ctl_table_header *__register_sysctl_paths(
1904         struct ctl_table_root *root,
1905         struct nsproxy *namespaces,
1906         const struct ctl_path *path, struct ctl_table *table)
1907 {
1908         struct ctl_table_header *header;
1909         struct ctl_table *new, **prevp;
1910         unsigned int n, npath;
1911         struct ctl_table_set *set;
1912
1913         /* Count the path components */
1914         for (npath = 0; path[npath].ctl_name || path[npath].procname; ++npath)
1915                 ;
1916
1917         /*
1918          * For each path component, allocate a 2-element ctl_table array.
1919          * The first array element will be filled with the sysctl entry
1920          * for this, the second will be the sentinel (ctl_name == 0).
1921          *
1922          * We allocate everything in one go so that we don't have to
1923          * worry about freeing additional memory in unregister_sysctl_table.
1924          */
1925         header = kzalloc(sizeof(struct ctl_table_header) +
1926                          (2 * npath * sizeof(struct ctl_table)), GFP_KERNEL);
1927         if (!header)
1928                 return NULL;
1929
1930         new = (struct ctl_table *) (header + 1);
1931
1932         /* Now connect the dots */
1933         prevp = &header->ctl_table;
1934         for (n = 0; n < npath; ++n, ++path) {
1935                 /* Copy the procname */
1936                 new->procname = path->procname;
1937                 new->ctl_name = path->ctl_name;
1938                 new->mode     = 0555;
1939
1940                 *prevp = new;
1941                 prevp = &new->child;
1942
1943                 new += 2;
1944         }
1945         *prevp = table;
1946         header->ctl_table_arg = table;
1947
1948         INIT_LIST_HEAD(&header->ctl_entry);
1949         header->used = 0;
1950         header->unregistering = NULL;
1951         header->root = root;
1952         sysctl_set_parent(NULL, header->ctl_table);
1953         header->count = 1;
1954 #ifdef CONFIG_SYSCTL_SYSCALL_CHECK
1955         if (sysctl_check_table(namespaces, header->ctl_table)) {
1956                 kfree(header);
1957                 return NULL;
1958         }
1959 #endif
1960         spin_lock(&sysctl_lock);
1961         header->set = lookup_header_set(root, namespaces);
1962         header->attached_by = header->ctl_table;
1963         header->attached_to = root_table;
1964         header->parent = &root_table_header;
1965         for (set = header->set; set; set = set->parent) {
1966                 struct ctl_table_header *p;
1967                 list_for_each_entry(p, &set->list, ctl_entry) {
1968                         if (p->unregistering)
1969                                 continue;
1970                         try_attach(p, header);
1971                 }
1972         }
1973         header->parent->count++;
1974         list_add_tail(&header->ctl_entry, &header->set->list);
1975         spin_unlock(&sysctl_lock);
1976
1977         return header;
1978 }
1979
1980 /**
1981  * register_sysctl_table_path - register a sysctl table hierarchy
1982  * @path: The path to the directory the sysctl table is in.
1983  * @table: the top-level table structure
1984  *
1985  * Register a sysctl table hierarchy. @table should be a filled in ctl_table
1986  * array. A completely 0 filled entry terminates the table.
1987  *
1988  * See __register_sysctl_paths for more details.
1989  */
1990 struct ctl_table_header *register_sysctl_paths(const struct ctl_path *path,
1991                                                 struct ctl_table *table)
1992 {
1993         return __register_sysctl_paths(&sysctl_table_root, current->nsproxy,
1994                                         path, table);
1995 }
1996
1997 /**
1998  * register_sysctl_table - register a sysctl table hierarchy
1999  * @table: the top-level table structure
2000  *
2001  * Register a sysctl table hierarchy. @table should be a filled in ctl_table
2002  * array. A completely 0 filled entry terminates the table.
2003  *
2004  * See register_sysctl_paths for more details.
2005  */
2006 struct ctl_table_header *register_sysctl_table(struct ctl_table *table)
2007 {
2008         static const struct ctl_path null_path[] = { {} };
2009
2010         return register_sysctl_paths(null_path, table);
2011 }
2012
2013 /**
2014  * unregister_sysctl_table - unregister a sysctl table hierarchy
2015  * @header: the header returned from register_sysctl_table
2016  *
2017  * Unregisters the sysctl table and all children. proc entries may not
2018  * actually be removed until they are no longer used by anyone.
2019  */
2020 void unregister_sysctl_table(struct ctl_table_header * header)
2021 {
2022         might_sleep();
2023
2024         if (header == NULL)
2025                 return;
2026
2027         spin_lock(&sysctl_lock);
2028         start_unregistering(header);
2029         if (!--header->parent->count) {
2030                 WARN_ON(1);
2031                 kfree(header->parent);
2032         }
2033         if (!--header->count)
2034                 kfree(header);
2035         spin_unlock(&sysctl_lock);
2036 }
2037
2038 int sysctl_is_seen(struct ctl_table_header *p)
2039 {
2040         struct ctl_table_set *set = p->set;
2041         int res;
2042         spin_lock(&sysctl_lock);
2043         if (p->unregistering)
2044                 res = 0;
2045         else if (!set->is_seen)
2046                 res = 1;
2047         else
2048                 res = set->is_seen(set);
2049         spin_unlock(&sysctl_lock);
2050         return res;
2051 }
2052
2053 void setup_sysctl_set(struct ctl_table_set *p,
2054         struct ctl_table_set *parent,
2055         int (*is_seen)(struct ctl_table_set *))
2056 {
2057         INIT_LIST_HEAD(&p->list);
2058         p->parent = parent ? parent : &sysctl_table_root.default_set;
2059         p->is_seen = is_seen;
2060 }
2061
2062 #else /* !CONFIG_SYSCTL */
2063 struct ctl_table_header *register_sysctl_table(struct ctl_table * table)
2064 {
2065         return NULL;
2066 }
2067
2068 struct ctl_table_header *register_sysctl_paths(const struct ctl_path *path,
2069                                                     struct ctl_table *table)
2070 {
2071         return NULL;
2072 }
2073
2074 void unregister_sysctl_table(struct ctl_table_header * table)
2075 {
2076 }
2077
2078 void setup_sysctl_set(struct ctl_table_set *p,
2079         struct ctl_table_set *parent,
2080         int (*is_seen)(struct ctl_table_set *))
2081 {
2082 }
2083
2084 void sysctl_head_put(struct ctl_table_header *head)
2085 {
2086 }
2087
2088 #endif /* CONFIG_SYSCTL */
2089
2090 /*
2091  * /proc/sys support
2092  */
2093
2094 #ifdef CONFIG_PROC_SYSCTL
2095
2096 static int _proc_do_string(void* data, int maxlen, int write,
2097                            struct file *filp, void __user *buffer,
2098                            size_t *lenp, loff_t *ppos)
2099 {
2100         size_t len;
2101         char __user *p;
2102         char c;
2103
2104         if (!data || !maxlen || !*lenp) {
2105                 *lenp = 0;
2106                 return 0;
2107         }
2108
2109         if (write) {
2110                 len = 0;
2111                 p = buffer;
2112                 while (len < *lenp) {
2113                         if (get_user(c, p++))
2114                                 return -EFAULT;
2115                         if (c == 0 || c == '\n')
2116                                 break;
2117                         len++;
2118                 }
2119                 if (len >= maxlen)
2120                         len = maxlen-1;
2121                 if(copy_from_user(data, buffer, len))
2122                         return -EFAULT;
2123                 ((char *) data)[len] = 0;
2124                 *ppos += *lenp;
2125         } else {
2126                 len = strlen(data);
2127                 if (len > maxlen)
2128                         len = maxlen;
2129
2130                 if (*ppos > len) {
2131                         *lenp = 0;
2132                         return 0;
2133                 }
2134
2135                 data += *ppos;
2136                 len  -= *ppos;
2137
2138                 if (len > *lenp)
2139                         len = *lenp;
2140                 if (len)
2141                         if(copy_to_user(buffer, data, len))
2142                                 return -EFAULT;
2143                 if (len < *lenp) {
2144                         if(put_user('\n', ((char __user *) buffer) + len))
2145                                 return -EFAULT;
2146                         len++;
2147                 }
2148                 *lenp = len;
2149                 *ppos += len;
2150         }
2151         return 0;
2152 }
2153
2154 /**
2155  * proc_dostring - read a string sysctl
2156  * @table: the sysctl table
2157  * @write: %TRUE if this is a write to the sysctl file
2158  * @filp: the file structure
2159  * @buffer: the user buffer
2160  * @lenp: the size of the user buffer
2161  * @ppos: file position
2162  *
2163  * Reads/writes a string from/to the user buffer. If the kernel
2164  * buffer provided is not large enough to hold the string, the
2165  * string is truncated. The copied string is %NULL-terminated.
2166  * If the string is being read by the user process, it is copied
2167  * and a newline '\n' is added. It is truncated if the buffer is
2168  * not large enough.
2169  *
2170  * Returns 0 on success.
2171  */
2172 int proc_dostring(struct ctl_table *table, int write, struct file *filp,
2173                   void __user *buffer, size_t *lenp, loff_t *ppos)
2174 {
2175         return _proc_do_string(table->data, table->maxlen, write, filp,
2176                                buffer, lenp, ppos);
2177 }
2178
2179
2180 static int do_proc_dointvec_conv(int *negp, unsigned long *lvalp,
2181                                  int *valp,
2182                                  int write, void *data)
2183 {
2184         if (write) {
2185                 *valp = *negp ? -*lvalp : *lvalp;
2186         } else {
2187                 int val = *valp;
2188                 if (val < 0) {
2189                         *negp = -1;
2190                         *lvalp = (unsigned long)-val;
2191                 } else {
2192                         *negp = 0;
2193                         *lvalp = (unsigned long)val;
2194                 }
2195         }
2196         return 0;
2197 }
2198
2199 static int __do_proc_dointvec(void *tbl_data, struct ctl_table *table,
2200                   int write, struct file *filp, void __user *buffer,
2201                   size_t *lenp, loff_t *ppos,
2202                   int (*conv)(int *negp, unsigned long *lvalp, int *valp,
2203                               int write, void *data),
2204                   void *data)
2205 {
2206 #define TMPBUFLEN 21
2207         int *i, vleft, first=1, neg, val;
2208         unsigned long lval;
2209         size_t left, len;
2210
2211         char buf[TMPBUFLEN], *p;
2212         char __user *s = buffer;
2213
2214         if (!tbl_data || !table->maxlen || !*lenp ||
2215             (*ppos && !write)) {
2216                 *lenp = 0;
2217                 return 0;
2218         }
2219
2220         i = (int *) tbl_data;
2221         vleft = table->maxlen / sizeof(*i);
2222         left = *lenp;
2223
2224         if (!conv)
2225                 conv = do_proc_dointvec_conv;
2226
2227         for (; left && vleft--; i++, first=0) {
2228                 if (write) {
2229                         while (left) {
2230                                 char c;
2231                                 if (get_user(c, s))
2232                                         return -EFAULT;
2233                                 if (!isspace(c))
2234                                         break;
2235                                 left--;
2236                                 s++;
2237                         }
2238                         if (!left)
2239                                 break;
2240                         neg = 0;
2241                         len = left;
2242                         if (len > sizeof(buf) - 1)
2243                                 len = sizeof(buf) - 1;
2244                         if (copy_from_user(buf, s, len))
2245                                 return -EFAULT;
2246                         buf[len] = 0;
2247                         p = buf;
2248                         if (*p == '-' && left > 1) {
2249                                 neg = 1;
2250                                 p++;
2251                         }
2252                         if (*p < '0' || *p > '9')
2253                                 break;
2254
2255                         lval = simple_strtoul(p, &p, 0);
2256
2257                         len = p-buf;
2258                         if ((len < left) && *p && !isspace(*p))
2259                                 break;
2260                         if (neg)
2261                                 val = -val;
2262                         s += len;
2263                         left -= len;
2264
2265                         if (conv(&neg, &lval, i, 1, data))
2266                                 break;
2267                 } else {
2268                         p = buf;
2269                         if (!first)
2270                                 *p++ = '\t';
2271
2272                         if (conv(&neg, &lval, i, 0, data))
2273                                 break;
2274
2275                         sprintf(p, "%s%lu", neg ? "-" : "", lval);
2276                         len = strlen(buf);
2277                         if (len > left)
2278                                 len = left;
2279                         if(copy_to_user(s, buf, len))
2280                                 return -EFAULT;
2281                         left -= len;
2282                         s += len;
2283                 }
2284         }
2285
2286         if (!write && !first && left) {
2287                 if(put_user('\n', s))
2288                         return -EFAULT;
2289                 left--, s++;
2290         }
2291         if (write) {
2292                 while (left) {
2293                         char c;
2294                         if (get_user(c, s++))
2295                                 return -EFAULT;
2296                         if (!isspace(c))
2297                                 break;
2298                         left--;
2299                 }
2300         }
2301         if (write && first)
2302                 return -EINVAL;
2303         *lenp -= left;
2304         *ppos += *lenp;
2305         return 0;
2306 #undef TMPBUFLEN
2307 }
2308
2309 static int do_proc_dointvec(struct ctl_table *table, int write, struct file *filp,
2310                   void __user *buffer, size_t *lenp, loff_t *ppos,
2311                   int (*conv)(int *negp, unsigned long *lvalp, int *valp,
2312                               int write, void *data),
2313                   void *data)
2314 {
2315         return __do_proc_dointvec(table->data, table, write, filp,
2316                         buffer, lenp, ppos, conv, data);
2317 }
2318
2319 /**
2320  * proc_dointvec - read a vector of integers
2321  * @table: the sysctl table
2322  * @write: %TRUE if this is a write to the sysctl file
2323  * @filp: the file structure
2324  * @buffer: the user buffer
2325  * @lenp: the size of the user buffer
2326  * @ppos: file position
2327  *
2328  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2329  * values from/to the user buffer, treated as an ASCII string.
2330  *
2331  * Returns 0 on success.
2332  */
2333 int proc_dointvec(struct ctl_table *table, int write, struct file *filp,
2334                      void __user *buffer, size_t *lenp, loff_t *ppos)
2335 {
2336     return do_proc_dointvec(table,write,filp,buffer,lenp,ppos,
2337                             NULL,NULL);
2338 }
2339
2340 /*
2341  * Taint values can only be increased
2342  * This means we can safely use a temporary.
2343  */
2344 static int proc_taint(struct ctl_table *table, int write, struct file *filp,
2345                                void __user *buffer, size_t *lenp, loff_t *ppos)
2346 {
2347         struct ctl_table t;
2348         unsigned long tmptaint = get_taint();
2349         int err;
2350
2351         if (write && !capable(CAP_SYS_ADMIN))
2352                 return -EPERM;
2353
2354         t = *table;
2355         t.data = &tmptaint;
2356         err = proc_doulongvec_minmax(&t, write, filp, buffer, lenp, ppos);
2357         if (err < 0)
2358                 return err;
2359
2360         if (write) {
2361                 /*
2362                  * Poor man's atomic or. Not worth adding a primitive
2363                  * to everyone's atomic.h for this
2364                  */
2365                 int i;
2366                 for (i = 0; i < BITS_PER_LONG && tmptaint >> i; i++) {
2367                         if ((tmptaint >> i) & 1)
2368                                 add_taint(i);
2369                 }
2370         }
2371
2372         return err;
2373 }
2374
2375 struct do_proc_dointvec_minmax_conv_param {
2376         int *min;
2377         int *max;
2378 };
2379
2380 static int do_proc_dointvec_minmax_conv(int *negp, unsigned long *lvalp,
2381                                         int *valp,
2382                                         int write, void *data)
2383 {
2384         struct do_proc_dointvec_minmax_conv_param *param = data;
2385         if (write) {
2386                 int val = *negp ? -*lvalp : *lvalp;
2387                 if ((param->min && *param->min > val) ||
2388                     (param->max && *param->max < val))
2389                         return -EINVAL;
2390                 *valp = val;
2391         } else {
2392                 int val = *valp;
2393                 if (val < 0) {
2394                         *negp = -1;
2395                         *lvalp = (unsigned long)-val;
2396                 } else {
2397                         *negp = 0;
2398                         *lvalp = (unsigned long)val;
2399                 }
2400         }
2401         return 0;
2402 }
2403
2404 /**
2405  * proc_dointvec_minmax - read a vector of integers with min/max values
2406  * @table: the sysctl table
2407  * @write: %TRUE if this is a write to the sysctl file
2408  * @filp: the file structure
2409  * @buffer: the user buffer
2410  * @lenp: the size of the user buffer
2411  * @ppos: file position
2412  *
2413  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2414  * values from/to the user buffer, treated as an ASCII string.
2415  *
2416  * This routine will ensure the values are within the range specified by
2417  * table->extra1 (min) and table->extra2 (max).
2418  *
2419  * Returns 0 on success.
2420  */
2421 int proc_dointvec_minmax(struct ctl_table *table, int write, struct file *filp,
2422                   void __user *buffer, size_t *lenp, loff_t *ppos)
2423 {
2424         struct do_proc_dointvec_minmax_conv_param param = {
2425                 .min = (int *) table->extra1,
2426                 .max = (int *) table->extra2,
2427         };
2428         return do_proc_dointvec(table, write, filp, buffer, lenp, ppos,
2429                                 do_proc_dointvec_minmax_conv, &param);
2430 }
2431
2432 static int __do_proc_doulongvec_minmax(void *data, struct ctl_table *table, int write,
2433                                      struct file *filp,
2434                                      void __user *buffer,
2435                                      size_t *lenp, loff_t *ppos,
2436                                      unsigned long convmul,
2437                                      unsigned long convdiv)
2438 {
2439 #define TMPBUFLEN 21
2440         unsigned long *i, *min, *max, val;
2441         int vleft, first=1, neg;
2442         size_t len, left;
2443         char buf[TMPBUFLEN], *p;
2444         char __user *s = buffer;
2445
2446         if (!data || !table->maxlen || !*lenp ||
2447             (*ppos && !write)) {
2448                 *lenp = 0;
2449                 return 0;
2450         }
2451
2452         i = (unsigned long *) data;
2453         min = (unsigned long *) table->extra1;
2454         max = (unsigned long *) table->extra2;
2455         vleft = table->maxlen / sizeof(unsigned long);
2456         left = *lenp;
2457
2458         for (; left && vleft--; i++, min++, max++, first=0) {
2459                 if (write) {
2460                         while (left) {
2461                                 char c;
2462                                 if (get_user(c, s))
2463                                         return -EFAULT;
2464                                 if (!isspace(c))
2465                                         break;
2466                                 left--;
2467                                 s++;
2468                         }
2469                         if (!left)
2470                                 break;
2471                         neg = 0;
2472                         len = left;
2473                         if (len > TMPBUFLEN-1)
2474                                 len = TMPBUFLEN-1;
2475                         if (copy_from_user(buf, s, len))
2476                                 return -EFAULT;
2477                         buf[len] = 0;
2478                         p = buf;
2479                         if (*p == '-' && left > 1) {
2480                                 neg = 1;
2481                                 p++;
2482                         }
2483                         if (*p < '0' || *p > '9')
2484                                 break;
2485                         val = simple_strtoul(p, &p, 0) * convmul / convdiv ;
2486                         len = p-buf;
2487                         if ((len < left) && *p && !isspace(*p))
2488                                 break;
2489                         if (neg)
2490                                 val = -val;
2491                         s += len;
2492                         left -= len;
2493
2494                         if(neg)
2495                                 continue;
2496                         if ((min && val < *min) || (max && val > *max))
2497                                 continue;
2498                         *i = val;
2499                 } else {
2500                         p = buf;
2501                         if (!first)
2502                                 *p++ = '\t';
2503                         sprintf(p, "%lu", convdiv * (*i) / convmul);
2504                         len = strlen(buf);
2505                         if (len > left)
2506                                 len = left;
2507                         if(copy_to_user(s, buf, len))
2508                                 return -EFAULT;
2509                         left -= len;
2510                         s += len;
2511                 }
2512         }
2513
2514         if (!write && !first && left) {
2515                 if(put_user('\n', s))
2516                         return -EFAULT;
2517                 left--, s++;
2518         }
2519         if (write) {
2520                 while (left) {
2521                         char c;
2522                         if (get_user(c, s++))
2523                                 return -EFAULT;
2524                         if (!isspace(c))
2525                                 break;
2526                         left--;
2527                 }
2528         }
2529         if (write && first)
2530                 return -EINVAL;
2531         *lenp -= left;
2532         *ppos += *lenp;
2533         return 0;
2534 #undef TMPBUFLEN
2535 }
2536
2537 static int do_proc_doulongvec_minmax(struct ctl_table *table, int write,
2538                                      struct file *filp,
2539                                      void __user *buffer,
2540                                      size_t *lenp, loff_t *ppos,
2541                                      unsigned long convmul,
2542                                      unsigned long convdiv)
2543 {
2544         return __do_proc_doulongvec_minmax(table->data, table, write,
2545                         filp, buffer, lenp, ppos, convmul, convdiv);
2546 }
2547
2548 /**
2549  * proc_doulongvec_minmax - read a vector of long integers with min/max values
2550  * @table: the sysctl table
2551  * @write: %TRUE if this is a write to the sysctl file
2552  * @filp: the file structure
2553  * @buffer: the user buffer
2554  * @lenp: the size of the user buffer
2555  * @ppos: file position
2556  *
2557  * Reads/writes up to table->maxlen/sizeof(unsigned long) unsigned long
2558  * values from/to the user buffer, treated as an ASCII string.
2559  *
2560  * This routine will ensure the values are within the range specified by
2561  * table->extra1 (min) and table->extra2 (max).
2562  *
2563  * Returns 0 on success.
2564  */
2565 int proc_doulongvec_minmax(struct ctl_table *table, int write, struct file *filp,
2566                            void __user *buffer, size_t *lenp, loff_t *ppos)
2567 {
2568     return do_proc_doulongvec_minmax(table, write, filp, buffer, lenp, ppos, 1l, 1l);
2569 }
2570
2571 /**
2572  * proc_doulongvec_ms_jiffies_minmax - read a vector of millisecond values with min/max values
2573  * @table: the sysctl table
2574  * @write: %TRUE if this is a write to the sysctl file
2575  * @filp: the file structure
2576  * @buffer: the user buffer
2577  * @lenp: the size of the user buffer
2578  * @ppos: file position
2579  *
2580  * Reads/writes up to table->maxlen/sizeof(unsigned long) unsigned long
2581  * values from/to the user buffer, treated as an ASCII string. The values
2582  * are treated as milliseconds, and converted to jiffies when they are stored.
2583  *
2584  * This routine will ensure the values are within the range specified by
2585  * table->extra1 (min) and table->extra2 (max).
2586  *
2587  * Returns 0 on success.
2588  */
2589 int proc_doulongvec_ms_jiffies_minmax(struct ctl_table *table, int write,
2590                                       struct file *filp,
2591                                       void __user *buffer,
2592                                       size_t *lenp, loff_t *ppos)
2593 {
2594     return do_proc_doulongvec_minmax(table, write, filp, buffer,
2595                                      lenp, ppos, HZ, 1000l);
2596 }
2597
2598
2599 static int do_proc_dointvec_jiffies_conv(int *negp, unsigned long *lvalp,
2600                                          int *valp,
2601                                          int write, void *data)
2602 {
2603         if (write) {
2604                 if (*lvalp > LONG_MAX / HZ)
2605                         return 1;
2606                 *valp = *negp ? -(*lvalp*HZ) : (*lvalp*HZ);
2607         } else {
2608                 int val = *valp;
2609                 unsigned long lval;
2610                 if (val < 0) {
2611                         *negp = -1;
2612                         lval = (unsigned long)-val;
2613                 } else {
2614                         *negp = 0;
2615                         lval = (unsigned long)val;
2616                 }
2617                 *lvalp = lval / HZ;
2618         }
2619         return 0;
2620 }
2621
2622 static int do_proc_dointvec_userhz_jiffies_conv(int *negp, unsigned long *lvalp,
2623                                                 int *valp,
2624                                                 int write, void *data)
2625 {
2626         if (write) {
2627                 if (USER_HZ < HZ && *lvalp > (LONG_MAX / HZ) * USER_HZ)
2628                         return 1;
2629                 *valp = clock_t_to_jiffies(*negp ? -*lvalp : *lvalp);
2630         } else {
2631                 int val = *valp;
2632                 unsigned long lval;
2633                 if (val < 0) {
2634                         *negp = -1;
2635                         lval = (unsigned long)-val;
2636                 } else {
2637                         *negp = 0;
2638                         lval = (unsigned long)val;
2639                 }
2640                 *lvalp = jiffies_to_clock_t(lval);
2641         }
2642         return 0;
2643 }
2644
2645 static int do_proc_dointvec_ms_jiffies_conv(int *negp, unsigned long *lvalp,
2646                                             int *valp,
2647                                             int write, void *data)
2648 {
2649         if (write) {
2650                 *valp = msecs_to_jiffies(*negp ? -*lvalp : *lvalp);
2651         } else {
2652                 int val = *valp;
2653                 unsigned long lval;
2654                 if (val < 0) {
2655                         *negp = -1;
2656                         lval = (unsigned long)-val;
2657                 } else {
2658                         *negp = 0;
2659                         lval = (unsigned long)val;
2660                 }
2661                 *lvalp = jiffies_to_msecs(lval);
2662         }
2663         return 0;
2664 }
2665
2666 /**
2667  * proc_dointvec_jiffies - read a vector of integers as seconds
2668  * @table: the sysctl table
2669  * @write: %TRUE if this is a write to the sysctl file
2670  * @filp: the file structure
2671  * @buffer: the user buffer
2672  * @lenp: the size of the user buffer
2673  * @ppos: file position
2674  *
2675  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2676  * values from/to the user buffer, treated as an ASCII string.
2677  * The values read are assumed to be in seconds, and are converted into
2678  * jiffies.
2679  *
2680  * Returns 0 on success.
2681  */
2682 int proc_dointvec_jiffies(struct ctl_table *table, int write, struct file *filp,
2683                           void __user *buffer, size_t *lenp, loff_t *ppos)
2684 {
2685     return do_proc_dointvec(table,write,filp,buffer,lenp,ppos,
2686                             do_proc_dointvec_jiffies_conv,NULL);
2687 }
2688
2689 /**
2690  * proc_dointvec_userhz_jiffies - read a vector of integers as 1/USER_HZ seconds
2691  * @table: the sysctl table
2692  * @write: %TRUE if this is a write to the sysctl file
2693  * @filp: the file structure
2694  * @buffer: the user buffer
2695  * @lenp: the size of the user buffer
2696  * @ppos: pointer to the file position
2697  *
2698  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2699  * values from/to the user buffer, treated as an ASCII string.
2700  * The values read are assumed to be in 1/USER_HZ seconds, and
2701  * are converted into jiffies.
2702  *
2703  * Returns 0 on success.
2704  */
2705 int proc_dointvec_userhz_jiffies(struct ctl_table *table, int write, struct file *filp,
2706                                  void __user *buffer, size_t *lenp, loff_t *ppos)
2707 {
2708     return do_proc_dointvec(table,write,filp,buffer,lenp,ppos,
2709                             do_proc_dointvec_userhz_jiffies_conv,NULL);
2710 }
2711
2712 /**
2713  * proc_dointvec_ms_jiffies - read a vector of integers as 1 milliseconds
2714  * @table: the sysctl table
2715  * @write: %TRUE if this is a write to the sysctl file
2716  * @filp: the file structure
2717  * @buffer: the user buffer
2718  * @lenp: the size of the user buffer
2719  * @ppos: file position
2720  * @ppos: the current position in the file
2721  *
2722  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2723  * values from/to the user buffer, treated as an ASCII string.
2724  * The values read are assumed to be in 1/1000 seconds, and
2725  * are converted into jiffies.
2726  *
2727  * Returns 0 on success.
2728  */
2729 int proc_dointvec_ms_jiffies(struct ctl_table *table, int write, struct file *filp,
2730                              void __user *buffer, size_t *lenp, loff_t *ppos)
2731 {
2732         return do_proc_dointvec(table, write, filp, buffer, lenp, ppos,
2733                                 do_proc_dointvec_ms_jiffies_conv, NULL);
2734 }
2735
2736 static int proc_do_cad_pid(struct ctl_table *table, int write, struct file *filp,
2737                            void __user *buffer, size_t *lenp, loff_t *ppos)
2738 {
2739         struct pid *new_pid;
2740         pid_t tmp;
2741         int r;
2742
2743         tmp = pid_vnr(cad_pid);
2744
2745         r = __do_proc_dointvec(&tmp, table, write, filp, buffer,
2746                                lenp, ppos, NULL, NULL);
2747         if (r || !write)
2748                 return r;
2749
2750         new_pid = find_get_pid(tmp);
2751         if (!new_pid)
2752                 return -ESRCH;
2753
2754         put_pid(xchg(&cad_pid, new_pid));
2755         return 0;
2756 }
2757
2758 #else /* CONFIG_PROC_FS */
2759
2760 int proc_dostring(struct ctl_table *table, int write, struct file *filp,
2761                   void __user *buffer, size_t *lenp, loff_t *ppos)
2762 {
2763         return -ENOSYS;
2764 }
2765
2766 int proc_dointvec(struct ctl_table *table, int write, struct file *filp,
2767                   void __user *buffer, size_t *lenp, loff_t *ppos)
2768 {
2769         return -ENOSYS;
2770 }
2771
2772 int proc_dointvec_minmax(struct ctl_table *table, int write, struct file *filp,
2773                     void __user *buffer, size_t *lenp, loff_t *ppos)
2774 {
2775         return -ENOSYS;
2776 }
2777
2778 int proc_dointvec_jiffies(struct ctl_table *table, int write, struct file *filp,
2779                     void __user *buffer, size_t *lenp, loff_t *ppos)
2780 {
2781         return -ENOSYS;
2782 }
2783
2784 int proc_dointvec_userhz_jiffies(struct ctl_table *table, int write, struct file *filp,
2785                     void __user *buffer, size_t *lenp, loff_t *ppos)
2786 {
2787         return -ENOSYS;
2788 }
2789
2790 int proc_dointvec_ms_jiffies(struct ctl_table *table, int write, struct file *filp,
2791                              void __user *buffer, size_t *lenp, loff_t *ppos)
2792 {
2793         return -ENOSYS;
2794 }
2795
2796 int proc_doulongvec_minmax(struct ctl_table *table, int write, struct file *filp,
2797                     void __user *buffer, size_t *lenp, loff_t *ppos)
2798 {
2799         return -ENOSYS;
2800 }
2801
2802 int proc_doulongvec_ms_jiffies_minmax(struct ctl_table *table, int write,
2803                                       struct file *filp,
2804                                       void __user *buffer,
2805                                       size_t *lenp, loff_t *ppos)
2806 {
2807     return -ENOSYS;
2808 }
2809
2810
2811 #endif /* CONFIG_PROC_FS */
2812
2813
2814 #ifdef CONFIG_SYSCTL_SYSCALL
2815 /*
2816  * General sysctl support routines
2817  */
2818
2819 /* The generic sysctl data routine (used if no strategy routine supplied) */
2820 int sysctl_data(struct ctl_table *table,
2821                 void __user *oldval, size_t __user *oldlenp,
2822                 void __user *newval, size_t newlen)
2823 {
2824         size_t len;
2825
2826         /* Get out of I don't have a variable */
2827         if (!table->data || !table->maxlen)
2828                 return -ENOTDIR;
2829
2830         if (oldval && oldlenp) {
2831                 if (get_user(len, oldlenp))
2832                         return -EFAULT;
2833                 if (len) {
2834                         if (len > table->maxlen)
2835                                 len = table->maxlen;
2836                         if (copy_to_user(oldval, table->data, len))
2837                                 return -EFAULT;
2838                         if (put_user(len, oldlenp))
2839                                 return -EFAULT;
2840                 }
2841         }
2842
2843         if (newval && newlen) {
2844                 if (newlen > table->maxlen)
2845                         newlen = table->maxlen;
2846
2847                 if (copy_from_user(table->data, newval, newlen))
2848                         return -EFAULT;
2849         }
2850         return 1;
2851 }
2852
2853 /* The generic string strategy routine: */
2854 int sysctl_string(struct ctl_table *table,
2855                   void __user *oldval, size_t __user *oldlenp,
2856                   void __user *newval, size_t newlen)
2857 {
2858         if (!table->data || !table->maxlen)
2859                 return -ENOTDIR;
2860
2861         if (oldval && oldlenp) {
2862                 size_t bufsize;
2863                 if (get_user(bufsize, oldlenp))
2864                         return -EFAULT;
2865                 if (bufsize) {
2866                         size_t len = strlen(table->data), copied;
2867
2868                         /* This shouldn't trigger for a well-formed sysctl */
2869                         if (len > table->maxlen)
2870                                 len = table->maxlen;
2871
2872                         /* Copy up to a max of bufsize-1 bytes of the string */
2873                         copied = (len >= bufsize) ? bufsize - 1 : len;
2874
2875                         if (copy_to_user(oldval, table->data, copied) ||
2876                             put_user(0, (char __user *)(oldval + copied)))
2877                                 return -EFAULT;
2878                         if (put_user(len, oldlenp))
2879                                 return -EFAULT;
2880                 }
2881         }
2882         if (newval && newlen) {
2883                 size_t len = newlen;
2884                 if (len > table->maxlen)
2885                         len = table->maxlen;
2886                 if(copy_from_user(table->data, newval, len))
2887                         return -EFAULT;
2888                 if (len == table->maxlen)
2889                         len--;
2890                 ((char *) table->data)[len] = 0;
2891         }
2892         return 1;
2893 }
2894
2895 /*
2896  * This function makes sure that all of the integers in the vector
2897  * are between the minimum and maximum values given in the arrays
2898  * table->extra1 and table->extra2, respectively.
2899  */
2900 int sysctl_intvec(struct ctl_table *table,
2901                 void __user *oldval, size_t __user *oldlenp,
2902                 void __user *newval, size_t newlen)
2903 {
2904
2905         if (newval && newlen) {
2906                 int __user *vec = (int __user *) newval;
2907                 int *min = (int *) table->extra1;
2908                 int *max = (int *) table->extra2;
2909                 size_t length;
2910                 int i;
2911
2912                 if (newlen % sizeof(int) != 0)
2913                         return -EINVAL;
2914
2915                 if (!table->extra1 && !table->extra2)
2916                         return 0;
2917
2918                 if (newlen > table->maxlen)
2919                         newlen = table->maxlen;
2920                 length = newlen / sizeof(int);
2921
2922                 for (i = 0; i < length; i++) {
2923                         int value;
2924                         if (get_user(value, vec + i))
2925                                 return -EFAULT;
2926                         if (min && value < min[i])
2927                                 return -EINVAL;
2928                         if (max && value > max[i])
2929                                 return -EINVAL;
2930                 }
2931         }
2932         return 0;
2933 }
2934
2935 /* Strategy function to convert jiffies to seconds */
2936 int sysctl_jiffies(struct ctl_table *table,
2937                 void __user *oldval, size_t __user *oldlenp,
2938                 void __user *newval, size_t newlen)
2939 {
2940         if (oldval && oldlenp) {
2941                 size_t olen;
2942
2943                 if (get_user(olen, oldlenp))
2944                         return -EFAULT;
2945                 if (olen) {
2946                         int val;
2947
2948                         if (olen < sizeof(int))
2949                                 return -EINVAL;
2950
2951                         val = *(int *)(table->data) / HZ;
2952                         if (put_user(val, (int __user *)oldval))
2953                                 return -EFAULT;
2954                         if (put_user(sizeof(int), oldlenp))
2955                                 return -EFAULT;
2956                 }
2957         }
2958         if (newval && newlen) {
2959                 int new;
2960                 if (newlen != sizeof(int))
2961                         return -EINVAL;
2962                 if (get_user(new, (int __user *)newval))
2963                         return -EFAULT;
2964                 *(int *)(table->data) = new*HZ;
2965         }
2966         return 1;
2967 }
2968
2969 /* Strategy function to convert jiffies to seconds */
2970 int sysctl_ms_jiffies(struct ctl_table *table,
2971                 void __user *oldval, size_t __user *oldlenp,
2972                 void __user *newval, size_t newlen)
2973 {
2974         if (oldval && oldlenp) {
2975                 size_t olen;
2976
2977                 if (get_user(olen, oldlenp))
2978                         return -EFAULT;
2979                 if (olen) {
2980                         int val;
2981
2982                         if (olen < sizeof(int))
2983                                 return -EINVAL;
2984
2985                         val = jiffies_to_msecs(*(int *)(table->data));
2986                         if (put_user(val, (int __user *)oldval))
2987                                 return -EFAULT;
2988                         if (put_user(sizeof(int), oldlenp))
2989                                 return -EFAULT;
2990                 }
2991         }
2992         if (newval && newlen) {
2993                 int new;
2994                 if (newlen != sizeof(int))
2995                         return -EINVAL;
2996                 if (get_user(new, (int __user *)newval))
2997                         return -EFAULT;
2998                 *(int *)(table->data) = msecs_to_jiffies(new);
2999         }
3000         return 1;
3001 }
3002
3003
3004
3005 #else /* CONFIG_SYSCTL_SYSCALL */
3006
3007
3008 SYSCALL_DEFINE1(sysctl, struct __sysctl_args __user *, args)
3009 {
3010         struct __sysctl_args tmp;
3011         int error;
3012
3013         if (copy_from_user(&tmp, args, sizeof(tmp)))
3014                 return -EFAULT;
3015
3016         error = deprecated_sysctl_warning(&tmp);
3017
3018         /* If no error reading the parameters then just -ENOSYS ... */
3019         if (!error)
3020                 error = -ENOSYS;
3021
3022         return error;
3023 }
3024
3025 int sysctl_data(struct ctl_table *table,
3026                   void __user *oldval, size_t __user *oldlenp,
3027                   void __user *newval, size_t newlen)
3028 {
3029         return -ENOSYS;
3030 }
3031
3032 int sysctl_string(struct ctl_table *table,
3033                   void __user *oldval, size_t __user *oldlenp,
3034                   void __user *newval, size_t newlen)
3035 {
3036         return -ENOSYS;
3037 }
3038
3039 int sysctl_intvec(struct ctl_table *table,
3040                 void __user *oldval, size_t __user *oldlenp,
3041                 void __user *newval, size_t newlen)
3042 {
3043         return -ENOSYS;
3044 }
3045
3046 int sysctl_jiffies(struct ctl_table *table,
3047                 void __user *oldval, size_t __user *oldlenp,
3048                 void __user *newval, size_t newlen)
3049 {
3050         return -ENOSYS;
3051 }
3052
3053 int sysctl_ms_jiffies(struct ctl_table *table,
3054                 void __user *oldval, size_t __user *oldlenp,
3055                 void __user *newval, size_t newlen)
3056 {
3057         return -ENOSYS;
3058 }
3059
3060 #endif /* CONFIG_SYSCTL_SYSCALL */
3061
3062 static int deprecated_sysctl_warning(struct __sysctl_args *args)
3063 {
3064         static int msg_count;
3065         int name[CTL_MAXNAME];
3066         int i;
3067
3068         /* Check args->nlen. */
3069         if (args->nlen < 0 || args->nlen > CTL_MAXNAME)
3070                 return -ENOTDIR;
3071
3072         /* Read in the sysctl name for better debug message logging */
3073         for (i = 0; i < args->nlen; i++)
3074                 if (get_user(name[i], args->name + i))
3075                         return -EFAULT;
3076
3077         /* Ignore accesses to kernel.version */
3078         if ((args->nlen == 2) && (name[0] == CTL_KERN) && (name[1] == KERN_VERSION))
3079                 return 0;
3080
3081         if (msg_count < 5) {
3082                 msg_count++;
3083                 printk(KERN_INFO
3084                         "warning: process `%s' used the deprecated sysctl "
3085                         "system call with ", current->comm);
3086                 for (i = 0; i < args->nlen; i++)
3087                         printk("%d.", name[i]);
3088                 printk("\n");
3089         }
3090         return 0;
3091 }
3092
3093 /*
3094  * No sense putting this after each symbol definition, twice,
3095  * exception granted :-)
3096  */
3097 EXPORT_SYMBOL(proc_dointvec);
3098 EXPORT_SYMBOL(proc_dointvec_jiffies);
3099 EXPORT_SYMBOL(proc_dointvec_minmax);
3100 EXPORT_SYMBOL(proc_dointvec_userhz_jiffies);
3101 EXPORT_SYMBOL(proc_dointvec_ms_jiffies);
3102 EXPORT_SYMBOL(proc_dostring);
3103 EXPORT_SYMBOL(proc_doulongvec_minmax);
3104 EXPORT_SYMBOL(proc_doulongvec_ms_jiffies_minmax);
3105 EXPORT_SYMBOL(register_sysctl_table);
3106 EXPORT_SYMBOL(register_sysctl_paths);
3107 EXPORT_SYMBOL(sysctl_intvec);
3108 EXPORT_SYMBOL(sysctl_jiffies);
3109 EXPORT_SYMBOL(sysctl_ms_jiffies);
3110 EXPORT_SYMBOL(sysctl_string);
3111 EXPORT_SYMBOL(sysctl_data);
3112 EXPORT_SYMBOL(unregister_sysctl_table);