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