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