SAFE public projects git trees. - safe/jmp/linux-2.6/blob - kernel/sysctl.c

   1 /*
   2  * sysctl.c: General linux system control interface
   3  *
   4  * Begun 24 March 1995, Stephen Tweedie
   5  * Added /proc support, Dec 1995
   6  * Added bdflush entry and intvec min/max checking, 2/23/96, Tom Dyas.
   7  * Added hooks for /proc/sys/net (minor, minor patch), 96/4/1, Mike Shaver.
   8  * Added kernel/java-{interpreter,appletviewer}, 96/5/10, Mike Shaver.
   9  * Dynamic registration fixes, Stephen Tweedie.
  10  * Added kswapd-interval, ctrl-alt-del, printk stuff, 1/8/97, Chris Horn.
  11  * Made sysctl support optional via CONFIG_SYSCTL, 1/10/97, Chris
  12  *  Horn.
  13  * Added proc_doulongvec_ms_jiffies_minmax, 09/08/99, Carlos H. Bauer.
  14  * Added proc_doulongvec_minmax, 09/08/99, Carlos H. Bauer.
  15  * Changed linked lists to use list.h instead of lists.h, 02/24/00, Bill
  16  *  Wendling.
  17  * The list_for_each() macro wasn't appropriate for the sysctl loop.
  18  *  Removed it and replaced it with older style, 03/23/00, Bill Wendling
  19  */
  20
  21 #include <linux/module.h>
  22 #include <linux/mm.h>
  23 #include <linux/swap.h>
  24 #include <linux/slab.h>
  25 #include <linux/sysctl.h>
  26 #include <linux/proc_fs.h>
  27 #include <linux/security.h>
  28 #include <linux/ctype.h>
  29 #include <linux/kmemcheck.h>
  30 #include <linux/smp_lock.h>
  31 #include <linux/fs.h>
  32 #include <linux/init.h>
  33 #include <linux/kernel.h>
  34 #include <linux/kobject.h>
  35 #include <linux/net.h>
  36 #include <linux/sysrq.h>
  37 #include <linux/highuid.h>
  38 #include <linux/writeback.h>
  39 #include <linux/hugetlb.h>
  40 #include <linux/initrd.h>
  41 #include <linux/key.h>
  42 #include <linux/times.h>
  43 #include <linux/limits.h>
  44 #include <linux/dcache.h>
  45 #include <linux/syscalls.h>
  46 #include <linux/vmstat.h>
  47 #include <linux/nfs_fs.h>
  48 #include <linux/acpi.h>
  49 #include <linux/reboot.h>
  50 #include <linux/ftrace.h>
  51 #include <linux/slow-work.h>
  52 #include <linux/perf_event.h>
  53
  54 #include <asm/uaccess.h>
  55 #include <asm/processor.h>
  56
  57 #ifdef CONFIG_X86
  58 #include <asm/nmi.h>
  59 #include <asm/stacktrace.h>
  60 #include <asm/io.h>
  61 #endif
  62
  63 static int deprecated_sysctl_warning(struct __sysctl_args *args);
  64
  65 #if defined(CONFIG_SYSCTL)
  66
  67 /* External variables not in a header file. */
  68 extern int C_A_D;
  69 extern int print_fatal_signals;
  70 extern int sysctl_overcommit_memory;
  71 extern int sysctl_overcommit_ratio;
  72 extern int sysctl_panic_on_oom;
  73 extern int sysctl_oom_kill_allocating_task;
  74 extern int sysctl_oom_dump_tasks;
  75 extern int max_threads;
  76 extern int core_uses_pid;
  77 extern int suid_dumpable;
  78 extern char core_pattern[];
  79 extern unsigned int core_pipe_limit;
  80 extern int pid_max;
  81 extern int min_free_kbytes;
  82 extern int pid_max_min, pid_max_max;
  83 extern int sysctl_drop_caches;
  84 extern int percpu_pagelist_fraction;
  85 extern int compat_log;
  86 extern int latencytop_enabled;
  87 extern int sysctl_nr_open_min, sysctl_nr_open_max;
  88 #ifndef CONFIG_MMU
  89 extern int sysctl_nr_trim_pages;
  90 #endif
  91 #ifdef CONFIG_RCU_TORTURE_TEST
  92 extern int rcutorture_runnable;
  93 #endif /* #ifdef CONFIG_RCU_TORTURE_TEST */
  94 #ifdef CONFIG_BLOCK
  95 extern int blk_iopoll_enabled;
  96 #endif
  97
  98 /* Constants used for minimum and  maximum */
  99 #ifdef CONFIG_DETECT_SOFTLOCKUP
 100 static int sixty = 60;
 101 static int neg_one = -1;
 102 #endif
 103
 104 static int zero;
 105 static int __maybe_unused one = 1;
 106 static int __maybe_unused two = 2;
 107 static unsigned long one_ul = 1;
 108 static int one_hundred = 100;
 109 #ifdef CONFIG_PRINTK
 110 static int ten_thousand = 10000;
 111 #endif
 112
 113 /* this is needed for the proc_doulongvec_minmax of vm_dirty_bytes */
 114 static unsigned long dirty_bytes_min = 2 * PAGE_SIZE;
 115
 116 /* this is needed for the proc_dointvec_minmax for [fs_]overflow UID and GID */
 117 static int maxolduid = 65535;
 118 static int minolduid;
 119 static int min_percpu_pagelist_fract = 8;
 120
 121 static int ngroups_max = NGROUPS_MAX;
 122
 123 #ifdef CONFIG_MODULES
 124 extern char modprobe_path[];
 125 extern int modules_disabled;
 126 #endif
 127 #ifdef CONFIG_CHR_DEV_SG
 128 extern int sg_big_buff;
 129 #endif
 130
 131 #ifdef CONFIG_SPARC
 132 #include <asm/system.h>
 133 #endif
 134
 135 #ifdef CONFIG_SPARC64
 136 extern int sysctl_tsb_ratio;
 137 #endif
 138
 139 #ifdef __hppa__
 140 extern int pwrsw_enabled;
 141 extern int unaligned_enabled;
 142 #endif
 143
 144 #ifdef CONFIG_S390
 145 #ifdef CONFIG_MATHEMU
 146 extern int sysctl_ieee_emulation_warnings;
 147 #endif
 148 extern int sysctl_userprocess_debug;
 149 extern int spin_retry;
 150 #endif
 151
 152 #ifdef CONFIG_BSD_PROCESS_ACCT
 153 extern int acct_parm[];
 154 #endif
 155
 156 #ifdef CONFIG_IA64
 157 extern int no_unaligned_warning;
 158 extern int unaligned_dump_stack;
 159 #endif
 160
 161 #ifdef CONFIG_RT_MUTEXES
 162 extern int max_lock_depth;
 163 #endif
 164
 165 #ifdef CONFIG_PROC_SYSCTL
 166 static int proc_do_cad_pid(struct ctl_table *table, int write, struct file *filp,
 167                   void __user *buffer, size_t *lenp, loff_t *ppos);
 168 static int proc_taint(struct ctl_table *table, int write, struct file *filp,
 169                                void __user *buffer, size_t *lenp, loff_t *ppos);
 170 #endif
 171
 172 static struct ctl_table root_table[];
 173 static struct ctl_table_root sysctl_table_root;
 174 static struct ctl_table_header root_table_header = {
 175         .count = 1,
 176         .ctl_table = root_table,
 177         .ctl_entry = LIST_HEAD_INIT(sysctl_table_root.default_set.list),
 178         .root = &sysctl_table_root,
 179         .set = &sysctl_table_root.default_set,
 180 };
 181 static struct ctl_table_root sysctl_table_root = {
 182         .root_list = LIST_HEAD_INIT(sysctl_table_root.root_list),
 183         .default_set.list = LIST_HEAD_INIT(root_table_header.ctl_entry),
 184 };
 185
 186 static struct ctl_table kern_table[];
 187 static struct ctl_table vm_table[];
 188 static struct ctl_table fs_table[];
 189 static struct ctl_table debug_table[];
 190 static struct ctl_table dev_table[];
 191 extern struct ctl_table random_table[];
 192 #ifdef CONFIG_INOTIFY_USER
 193 extern struct ctl_table inotify_table[];
 194 #endif
 195 #ifdef CONFIG_EPOLL
 196 extern struct ctl_table epoll_table[];
 197 #endif
 198
 199 #ifdef HAVE_ARCH_PICK_MMAP_LAYOUT
 200 int sysctl_legacy_va_layout;
 201 #endif
 202
 203 extern int prove_locking;
 204 extern int lock_stat;
 205
 206 /* The default sysctl tables: */
 207
 208 static struct ctl_table root_table[] = {
 209         {
 210                 .ctl_name       = CTL_KERN,
 211                 .procname       = "kernel",
 212                 .mode           = 0555,
 213                 .child          = kern_table,
 214         },
 215         {
 216                 .ctl_name       = CTL_VM,
 217                 .procname       = "vm",
 218                 .mode           = 0555,
 219                 .child          = vm_table,
 220         },
 221         {
 222                 .ctl_name       = CTL_FS,
 223                 .procname       = "fs",
 224                 .mode           = 0555,
 225                 .child          = fs_table,
 226         },
 227         {
 228                 .ctl_name       = CTL_DEBUG,
 229                 .procname       = "debug",
 230                 .mode           = 0555,
 231                 .child          = debug_table,
 232         },
 233         {
 234                 .ctl_name       = CTL_DEV,
 235                 .procname       = "dev",
 236                 .mode           = 0555,
 237                 .child          = dev_table,
 238         },
 239 /*
 240  * NOTE: do not add new entries to this table unless you have read
 241  * Documentation/sysctl/ctl_unnumbered.txt
 242  */
 243         { .ctl_name = 0 }
 244 };
 245
 246 #ifdef CONFIG_SCHED_DEBUG
 247 static int min_sched_granularity_ns = 100000;           /* 100 usecs */
 248 static int max_sched_granularity_ns = NSEC_PER_SEC;     /* 1 second */
 249 static int min_wakeup_granularity_ns;                   /* 0 usecs */
 250 static int max_wakeup_granularity_ns = NSEC_PER_SEC;    /* 1 second */
 251 #endif
 252
 253 static struct ctl_table kern_table[] = {
 254         {
 255                 .ctl_name       = CTL_UNNUMBERED,
 256                 .procname       = "sched_child_runs_first",
 257                 .data           = &sysctl_sched_child_runs_first,
 258                 .maxlen         = sizeof(unsigned int),
 259                 .mode           = 0644,
 260                 .proc_handler   = &proc_dointvec,
 261         },
 262 #ifdef CONFIG_SCHED_DEBUG
 263         {
 264                 .ctl_name       = CTL_UNNUMBERED,
 265                 .procname       = "sched_min_granularity_ns",
 266                 .data           = &sysctl_sched_min_granularity,
 267                 .maxlen         = sizeof(unsigned int),
 268                 .mode           = 0644,
 269                 .proc_handler   = &sched_nr_latency_handler,
 270                 .strategy       = &sysctl_intvec,
 271                 .extra1         = &min_sched_granularity_ns,
 272                 .extra2         = &max_sched_granularity_ns,
 273         },
 274         {
 275                 .ctl_name       = CTL_UNNUMBERED,
 276                 .procname       = "sched_latency_ns",
 277                 .data           = &sysctl_sched_latency,
 278                 .maxlen         = sizeof(unsigned int),
 279                 .mode           = 0644,
 280                 .proc_handler   = &sched_nr_latency_handler,
 281                 .strategy       = &sysctl_intvec,
 282                 .extra1         = &min_sched_granularity_ns,
 283                 .extra2         = &max_sched_granularity_ns,
 284         },
 285         {
 286                 .ctl_name       = CTL_UNNUMBERED,
 287                 .procname       = "sched_wakeup_granularity_ns",
 288                 .data           = &sysctl_sched_wakeup_granularity,
 289                 .maxlen         = sizeof(unsigned int),
 290                 .mode           = 0644,
 291                 .proc_handler   = &proc_dointvec_minmax,
 292                 .strategy       = &sysctl_intvec,
 293                 .extra1         = &min_wakeup_granularity_ns,
 294                 .extra2         = &max_wakeup_granularity_ns,
 295         },
 296         {
 297                 .ctl_name       = CTL_UNNUMBERED,
 298                 .procname       = "sched_shares_ratelimit",
 299                 .data           = &sysctl_sched_shares_ratelimit,
 300                 .maxlen         = sizeof(unsigned int),
 301                 .mode           = 0644,
 302                 .proc_handler   = &proc_dointvec,
 303         },
 304         {
 305                 .ctl_name       = CTL_UNNUMBERED,
 306                 .procname       = "sched_shares_thresh",
 307                 .data           = &sysctl_sched_shares_thresh,
 308                 .maxlen         = sizeof(unsigned int),
 309                 .mode           = 0644,
 310                 .proc_handler   = &proc_dointvec_minmax,
 311                 .strategy       = &sysctl_intvec,
 312                 .extra1         = &zero,
 313         },
 314         {
 315                 .ctl_name       = CTL_UNNUMBERED,
 316                 .procname       = "sched_features",
 317                 .data           = &sysctl_sched_features,
 318                 .maxlen         = sizeof(unsigned int),
 319                 .mode           = 0644,
 320                 .proc_handler   = &proc_dointvec,
 321         },
 322         {
 323                 .ctl_name       = CTL_UNNUMBERED,
 324                 .procname       = "sched_migration_cost",
 325                 .data           = &sysctl_sched_migration_cost,
 326                 .maxlen         = sizeof(unsigned int),
 327                 .mode           = 0644,
 328                 .proc_handler   = &proc_dointvec,
 329         },
 330         {
 331                 .ctl_name       = CTL_UNNUMBERED,
 332                 .procname       = "sched_nr_migrate",
 333                 .data           = &sysctl_sched_nr_migrate,
 334                 .maxlen         = sizeof(unsigned int),
 335                 .mode           = 0644,
 336                 .proc_handler   = &proc_dointvec,
 337         },
 338         {
 339                 .ctl_name       = CTL_UNNUMBERED,
 340                 .procname       = "sched_time_avg",
 341                 .data           = &sysctl_sched_time_avg,
 342                 .maxlen         = sizeof(unsigned int),
 343                 .mode           = 0644,
 344                 .proc_handler   = &proc_dointvec,
 345         },
 346         {
 347                 .ctl_name       = CTL_UNNUMBERED,
 348                 .procname       = "timer_migration",
 349                 .data           = &sysctl_timer_migration,
 350                 .maxlen         = sizeof(unsigned int),
 351                 .mode           = 0644,
 352                 .proc_handler   = &proc_dointvec_minmax,
 353                 .strategy       = &sysctl_intvec,
 354                 .extra1         = &zero,
 355                 .extra2         = &one,
 356         },
 357 #endif
 358         {
 359                 .ctl_name       = CTL_UNNUMBERED,
 360                 .procname       = "sched_rt_period_us",
 361                 .data           = &sysctl_sched_rt_period,
 362                 .maxlen         = sizeof(unsigned int),
 363                 .mode           = 0644,
 364                 .proc_handler   = &sched_rt_handler,
 365         },
 366         {
 367                 .ctl_name       = CTL_UNNUMBERED,
 368                 .procname       = "sched_rt_runtime_us",
 369                 .data           = &sysctl_sched_rt_runtime,
 370                 .maxlen         = sizeof(int),
 371                 .mode           = 0644,
 372                 .proc_handler   = &sched_rt_handler,
 373         },
 374         {
 375                 .ctl_name       = CTL_UNNUMBERED,
 376                 .procname       = "sched_compat_yield",
 377                 .data           = &sysctl_sched_compat_yield,
 378                 .maxlen         = sizeof(unsigned int),
 379                 .mode           = 0644,
 380                 .proc_handler   = &proc_dointvec,
 381         },
 382 #ifdef CONFIG_PROVE_LOCKING
 383         {
 384                 .ctl_name       = CTL_UNNUMBERED,
 385                 .procname       = "prove_locking",
 386                 .data           = &prove_locking,
 387                 .maxlen         = sizeof(int),
 388                 .mode           = 0644,
 389                 .proc_handler   = &proc_dointvec,
 390         },
 391 #endif
 392 #ifdef CONFIG_LOCK_STAT
 393         {
 394                 .ctl_name       = CTL_UNNUMBERED,
 395                 .procname       = "lock_stat",
 396                 .data           = &lock_stat,
 397                 .maxlen         = sizeof(int),
 398                 .mode           = 0644,
 399                 .proc_handler   = &proc_dointvec,
 400         },
 401 #endif
 402         {
 403                 .ctl_name       = KERN_PANIC,
 404                 .procname       = "panic",
 405                 .data           = &panic_timeout,
 406                 .maxlen         = sizeof(int),
 407                 .mode           = 0644,
 408                 .proc_handler   = &proc_dointvec,
 409         },
 410         {
 411                 .ctl_name       = KERN_CORE_USES_PID,
 412                 .procname       = "core_uses_pid",
 413                 .data           = &core_uses_pid,
 414                 .maxlen         = sizeof(int),
 415                 .mode           = 0644,
 416                 .proc_handler   = &proc_dointvec,
 417         },
 418         {
 419                 .ctl_name       = KERN_CORE_PATTERN,
 420                 .procname       = "core_pattern",
 421                 .data           = core_pattern,
 422                 .maxlen         = CORENAME_MAX_SIZE,
 423                 .mode           = 0644,
 424                 .proc_handler   = &proc_dostring,
 425                 .strategy       = &sysctl_string,
 426         },
 427         {
 428                 .ctl_name       = CTL_UNNUMBERED,
 429                 .procname       = "core_pipe_limit",
 430                 .data           = &core_pipe_limit,
 431                 .maxlen         = sizeof(unsigned int),
 432                 .mode           = 0644,
 433                 .proc_handler   = &proc_dointvec,
 434         },
 435 #ifdef CONFIG_PROC_SYSCTL
 436         {
 437                 .procname       = "tainted",
 438                 .maxlen         = sizeof(long),
 439                 .mode           = 0644,
 440                 .proc_handler   = &proc_taint,
 441         },
 442 #endif
 443 #ifdef CONFIG_LATENCYTOP
 444         {
 445                 .procname       = "latencytop",
 446                 .data           = &latencytop_enabled,
 447                 .maxlen         = sizeof(int),
 448                 .mode           = 0644,
 449                 .proc_handler   = &proc_dointvec,
 450         },
 451 #endif
 452 #ifdef CONFIG_BLK_DEV_INITRD
 453         {
 454                 .ctl_name       = KERN_REALROOTDEV,
 455                 .procname       = "real-root-dev",
 456                 .data           = &real_root_dev,
 457                 .maxlen         = sizeof(int),
 458                 .mode           = 0644,
 459                 .proc_handler   = &proc_dointvec,
 460         },
 461 #endif
 462         {
 463                 .ctl_name       = CTL_UNNUMBERED,
 464                 .procname       = "print-fatal-signals",
 465                 .data           = &print_fatal_signals,
 466                 .maxlen         = sizeof(int),
 467                 .mode           = 0644,
 468                 .proc_handler   = &proc_dointvec,
 469         },
 470 #ifdef CONFIG_SPARC
 471         {
 472                 .ctl_name       = KERN_SPARC_REBOOT,
 473                 .procname       = "reboot-cmd",
 474                 .data           = reboot_command,
 475                 .maxlen         = 256,
 476                 .mode           = 0644,
 477                 .proc_handler   = &proc_dostring,
 478                 .strategy       = &sysctl_string,
 479         },
 480         {
 481                 .ctl_name       = KERN_SPARC_STOP_A,
 482                 .procname       = "stop-a",
 483                 .data           = &stop_a_enabled,
 484                 .maxlen         = sizeof (int),
 485                 .mode           = 0644,
 486                 .proc_handler   = &proc_dointvec,
 487         },
 488         {
 489                 .ctl_name       = KERN_SPARC_SCONS_PWROFF,
 490                 .procname       = "scons-poweroff",
 491                 .data           = &scons_pwroff,
 492                 .maxlen         = sizeof (int),
 493                 .mode           = 0644,
 494                 .proc_handler   = &proc_dointvec,
 495         },
 496 #endif
 497 #ifdef CONFIG_SPARC64
 498         {
 499                 .ctl_name       = CTL_UNNUMBERED,
 500                 .procname       = "tsb-ratio",
 501                 .data           = &sysctl_tsb_ratio,
 502                 .maxlen         = sizeof (int),
 503                 .mode           = 0644,
 504                 .proc_handler   = &proc_dointvec,
 505         },
 506 #endif
 507 #ifdef __hppa__
 508         {
 509                 .ctl_name       = KERN_HPPA_PWRSW,
 510                 .procname       = "soft-power",
 511                 .data           = &pwrsw_enabled,
 512                 .maxlen         = sizeof (int),
 513                 .mode           = 0644,
 514                 .proc_handler   = &proc_dointvec,
 515         },
 516         {
 517                 .ctl_name       = KERN_HPPA_UNALIGNED,
 518                 .procname       = "unaligned-trap",
 519                 .data           = &unaligned_enabled,
 520                 .maxlen         = sizeof (int),
 521                 .mode           = 0644,
 522                 .proc_handler   = &proc_dointvec,
 523         },
 524 #endif
 525         {
 526                 .ctl_name       = KERN_CTLALTDEL,
 527                 .procname       = "ctrl-alt-del",
 528                 .data           = &C_A_D,
 529                 .maxlen         = sizeof(int),
 530                 .mode           = 0644,
 531                 .proc_handler   = &proc_dointvec,
 532         },
 533 #ifdef CONFIG_FUNCTION_TRACER
 534         {
 535                 .ctl_name       = CTL_UNNUMBERED,
 536                 .procname       = "ftrace_enabled",
 537                 .data           = &ftrace_enabled,
 538                 .maxlen         = sizeof(int),
 539                 .mode           = 0644,
 540                 .proc_handler   = &ftrace_enable_sysctl,
 541         },
 542 #endif
 543 #ifdef CONFIG_STACK_TRACER
 544         {
 545                 .ctl_name       = CTL_UNNUMBERED,
 546                 .procname       = "stack_tracer_enabled",
 547                 .data           = &stack_tracer_enabled,
 548                 .maxlen         = sizeof(int),
 549                 .mode           = 0644,
 550                 .proc_handler   = &stack_trace_sysctl,
 551         },
 552 #endif
 553 #ifdef CONFIG_TRACING
 554         {
 555                 .ctl_name       = CTL_UNNUMBERED,
 556                 .procname       = "ftrace_dump_on_oops",
 557                 .data           = &ftrace_dump_on_oops,
 558                 .maxlen         = sizeof(int),
 559                 .mode           = 0644,
 560                 .proc_handler   = &proc_dointvec,
 561         },
 562 #endif
 563 #ifdef CONFIG_MODULES
 564         {
 565                 .ctl_name       = KERN_MODPROBE,
 566                 .procname       = "modprobe",
 567                 .data           = &modprobe_path,
 568                 .maxlen         = KMOD_PATH_LEN,
 569                 .mode           = 0644,
 570                 .proc_handler   = &proc_dostring,
 571                 .strategy       = &sysctl_string,
 572         },
 573         {
 574                 .ctl_name       = CTL_UNNUMBERED,
 575                 .procname       = "modules_disabled",
 576                 .data           = &modules_disabled,
 577                 .maxlen         = sizeof(int),
 578                 .mode           = 0644,
 579                 /* only handle a transition from default "0" to "1" */
 580                 .proc_handler   = &proc_dointvec_minmax,
 581                 .extra1         = &one,
 582                 .extra2         = &one,
 583         },
 584 #endif
 585 #if defined(CONFIG_HOTPLUG) && defined(CONFIG_NET)
 586         {
 587                 .ctl_name       = KERN_HOTPLUG,
 588                 .procname       = "hotplug",
 589                 .data           = &uevent_helper,
 590                 .maxlen         = UEVENT_HELPER_PATH_LEN,
 591                 .mode           = 0644,
 592                 .proc_handler   = &proc_dostring,
 593                 .strategy       = &sysctl_string,
 594         },
 595 #endif
 596 #ifdef CONFIG_CHR_DEV_SG
 597         {
 598                 .ctl_name       = KERN_SG_BIG_BUFF,
 599                 .procname       = "sg-big-buff",
 600                 .data           = &sg_big_buff,
 601                 .maxlen         = sizeof (int),
 602                 .mode           = 0444,
 603                 .proc_handler   = &proc_dointvec,
 604         },
 605 #endif
 606 #ifdef CONFIG_BSD_PROCESS_ACCT
 607         {
 608                 .ctl_name       = KERN_ACCT,
 609                 .procname       = "acct",
 610                 .data           = &acct_parm,
 611                 .maxlen         = 3*sizeof(int),
 612                 .mode           = 0644,
 613                 .proc_handler   = &proc_dointvec,
 614         },
 615 #endif
 616 #ifdef CONFIG_MAGIC_SYSRQ
 617         {
 618                 .ctl_name       = KERN_SYSRQ,
 619                 .procname       = "sysrq",
 620                 .data           = &__sysrq_enabled,
 621                 .maxlen         = sizeof (int),
 622                 .mode           = 0644,
 623                 .proc_handler   = &proc_dointvec,
 624         },
 625 #endif
 626 #ifdef CONFIG_PROC_SYSCTL
 627         {
 628                 .procname       = "cad_pid",
 629                 .data           = NULL,
 630                 .maxlen         = sizeof (int),
 631                 .mode           = 0600,
 632                 .proc_handler   = &proc_do_cad_pid,
 633         },
 634 #endif
 635         {
 636                 .ctl_name       = KERN_MAX_THREADS,
 637                 .procname       = "threads-max",
 638                 .data           = &max_threads,
 639                 .maxlen         = sizeof(int),
 640                 .mode           = 0644,
 641                 .proc_handler   = &proc_dointvec,
 642         },
 643         {
 644                 .ctl_name       = KERN_RANDOM,
 645                 .procname       = "random",
 646                 .mode           = 0555,
 647                 .child          = random_table,
 648         },
 649         {
 650                 .ctl_name       = KERN_OVERFLOWUID,
 651                 .procname       = "overflowuid",
 652                 .data           = &overflowuid,
 653                 .maxlen         = sizeof(int),
 654                 .mode           = 0644,
 655                 .proc_handler   = &proc_dointvec_minmax,
 656                 .strategy       = &sysctl_intvec,
 657                 .extra1         = &minolduid,
 658                 .extra2         = &maxolduid,
 659         },
 660         {
 661                 .ctl_name       = KERN_OVERFLOWGID,
 662                 .procname       = "overflowgid",
 663                 .data           = &overflowgid,
 664                 .maxlen         = sizeof(int),
 665                 .mode           = 0644,
 666                 .proc_handler   = &proc_dointvec_minmax,
 667                 .strategy       = &sysctl_intvec,
 668                 .extra1         = &minolduid,
 669                 .extra2         = &maxolduid,
 670         },
 671 #ifdef CONFIG_S390
 672 #ifdef CONFIG_MATHEMU
 673         {
 674                 .ctl_name       = KERN_IEEE_EMULATION_WARNINGS,
 675                 .procname       = "ieee_emulation_warnings",
 676                 .data           = &sysctl_ieee_emulation_warnings,
 677                 .maxlen         = sizeof(int),
 678                 .mode           = 0644,
 679                 .proc_handler   = &proc_dointvec,
 680         },
 681 #endif
 682         {
 683                 .ctl_name       = KERN_S390_USER_DEBUG_LOGGING,
 684                 .procname       = "userprocess_debug",
 685                 .data           = &sysctl_userprocess_debug,
 686                 .maxlen         = sizeof(int),
 687                 .mode           = 0644,
 688                 .proc_handler   = &proc_dointvec,
 689         },
 690 #endif
 691         {
 692                 .ctl_name       = KERN_PIDMAX,
 693                 .procname       = "pid_max",
 694                 .data           = &pid_max,
 695                 .maxlen         = sizeof (int),
 696                 .mode           = 0644,
 697                 .proc_handler   = &proc_dointvec_minmax,
 698                 .strategy       = sysctl_intvec,
 699                 .extra1         = &pid_max_min,
 700                 .extra2         = &pid_max_max,
 701         },
 702         {
 703                 .ctl_name       = KERN_PANIC_ON_OOPS,
 704                 .procname       = "panic_on_oops",
 705                 .data           = &panic_on_oops,
 706                 .maxlen         = sizeof(int),
 707                 .mode           = 0644,
 708                 .proc_handler   = &proc_dointvec,
 709         },
 710 #if defined CONFIG_PRINTK
 711         {
 712                 .ctl_name       = KERN_PRINTK,
 713                 .procname       = "printk",
 714                 .data           = &console_loglevel,
 715                 .maxlen         = 4*sizeof(int),
 716                 .mode           = 0644,
 717                 .proc_handler   = &proc_dointvec,
 718         },
 719         {
 720                 .ctl_name       = KERN_PRINTK_RATELIMIT,
 721                 .procname       = "printk_ratelimit",
 722                 .data           = &printk_ratelimit_state.interval,
 723                 .maxlen         = sizeof(int),
 724                 .mode           = 0644,
 725                 .proc_handler   = &proc_dointvec_jiffies,
 726                 .strategy       = &sysctl_jiffies,
 727         },
 728         {
 729                 .ctl_name       = KERN_PRINTK_RATELIMIT_BURST,
 730                 .procname       = "printk_ratelimit_burst",
 731                 .data           = &printk_ratelimit_state.burst,
 732                 .maxlen         = sizeof(int),
 733                 .mode           = 0644,
 734                 .proc_handler   = &proc_dointvec,
 735         },
 736         {
 737                 .ctl_name       = CTL_UNNUMBERED,
 738                 .procname       = "printk_delay",
 739                 .data           = &printk_delay_msec,
 740                 .maxlen         = sizeof(int),
 741                 .mode           = 0644,
 742                 .proc_handler   = &proc_dointvec_minmax,
 743                 .strategy       = &sysctl_intvec,
 744                 .extra1         = &zero,
 745                 .extra2         = &ten_thousand,
 746         },
 747 #endif
 748         {
 749                 .ctl_name       = KERN_NGROUPS_MAX,
 750                 .procname       = "ngroups_max",
 751                 .data           = &ngroups_max,
 752                 .maxlen         = sizeof (int),
 753                 .mode           = 0444,
 754                 .proc_handler   = &proc_dointvec,
 755         },
 756 #if defined(CONFIG_X86_LOCAL_APIC) && defined(CONFIG_X86)
 757         {
 758                 .ctl_name       = KERN_UNKNOWN_NMI_PANIC,
 759                 .procname       = "unknown_nmi_panic",
 760                 .data           = &unknown_nmi_panic,
 761                 .maxlen         = sizeof (int),
 762                 .mode           = 0644,
 763                 .proc_handler   = &proc_dointvec,
 764         },
 765         {
 766                 .procname       = "nmi_watchdog",
 767                 .data           = &nmi_watchdog_enabled,
 768                 .maxlen         = sizeof (int),
 769                 .mode           = 0644,
 770                 .proc_handler   = &proc_nmi_enabled,
 771         },
 772 #endif
 773 #if defined(CONFIG_X86)
 774         {
 775                 .ctl_name       = KERN_PANIC_ON_NMI,
 776                 .procname       = "panic_on_unrecovered_nmi",
 777                 .data           = &panic_on_unrecovered_nmi,
 778                 .maxlen         = sizeof(int),
 779                 .mode           = 0644,
 780                 .proc_handler   = &proc_dointvec,
 781         },
 782         {
 783                 .ctl_name       = CTL_UNNUMBERED,
 784                 .procname       = "panic_on_io_nmi",
 785                 .data           = &panic_on_io_nmi,
 786                 .maxlen         = sizeof(int),
 787                 .mode           = 0644,
 788                 .proc_handler   = &proc_dointvec,
 789         },
 790         {
 791                 .ctl_name       = KERN_BOOTLOADER_TYPE,
 792                 .procname       = "bootloader_type",
 793                 .data           = &bootloader_type,
 794                 .maxlen         = sizeof (int),
 795                 .mode           = 0444,
 796                 .proc_handler   = &proc_dointvec,
 797         },
 798         {
 799                 .ctl_name       = CTL_UNNUMBERED,
 800                 .procname       = "bootloader_version",
 801                 .data           = &bootloader_version,
 802                 .maxlen         = sizeof (int),
 803                 .mode           = 0444,
 804                 .proc_handler   = &proc_dointvec,
 805         },
 806         {
 807                 .ctl_name       = CTL_UNNUMBERED,
 808                 .procname       = "kstack_depth_to_print",
 809                 .data           = &kstack_depth_to_print,
 810                 .maxlen         = sizeof(int),
 811                 .mode           = 0644,
 812                 .proc_handler   = &proc_dointvec,
 813         },
 814         {
 815                 .ctl_name       = CTL_UNNUMBERED,
 816                 .procname       = "io_delay_type",
 817                 .data           = &io_delay_type,
 818                 .maxlen         = sizeof(int),
 819                 .mode           = 0644,
 820                 .proc_handler   = &proc_dointvec,
 821         },
 822 #endif
 823 #if defined(CONFIG_MMU)
 824         {
 825                 .ctl_name       = KERN_RANDOMIZE,
 826                 .procname       = "randomize_va_space",
 827                 .data           = &randomize_va_space,
 828                 .maxlen         = sizeof(int),
 829                 .mode           = 0644,
 830                 .proc_handler   = &proc_dointvec,
 831         },
 832 #endif
 833 #if defined(CONFIG_S390) && defined(CONFIG_SMP)
 834         {
 835                 .ctl_name       = KERN_SPIN_RETRY,
 836                 .procname       = "spin_retry",
 837                 .data           = &spin_retry,
 838                 .maxlen         = sizeof (int),
 839                 .mode           = 0644,
 840                 .proc_handler   = &proc_dointvec,
 841         },
 842 #endif
 843 #if     defined(CONFIG_ACPI_SLEEP) && defined(CONFIG_X86)
 844         {
 845                 .procname       = "acpi_video_flags",
 846                 .data           = &acpi_realmode_flags,
 847                 .maxlen         = sizeof (unsigned long),
 848                 .mode           = 0644,
 849                 .proc_handler   = &proc_doulongvec_minmax,
 850         },
 851 #endif
 852 #ifdef CONFIG_IA64
 853         {
 854                 .ctl_name       = KERN_IA64_UNALIGNED,
 855                 .procname       = "ignore-unaligned-usertrap",
 856                 .data           = &no_unaligned_warning,
 857                 .maxlen         = sizeof (int),
 858                 .mode           = 0644,
 859                 .proc_handler   = &proc_dointvec,
 860         },
 861         {
 862                 .ctl_name       = CTL_UNNUMBERED,
 863                 .procname       = "unaligned-dump-stack",
 864                 .data           = &unaligned_dump_stack,
 865                 .maxlen         = sizeof (int),
 866                 .mode           = 0644,
 867                 .proc_handler   = &proc_dointvec,
 868         },
 869 #endif
 870 #ifdef CONFIG_DETECT_SOFTLOCKUP
 871         {
 872                 .ctl_name       = CTL_UNNUMBERED,
 873                 .procname       = "softlockup_panic",
 874                 .data           = &softlockup_panic,
 875                 .maxlen         = sizeof(int),
 876                 .mode           = 0644,
 877                 .proc_handler   = &proc_dointvec_minmax,
 878                 .strategy       = &sysctl_intvec,
 879                 .extra1         = &zero,
 880                 .extra2         = &one,
 881         },
 882         {
 883                 .ctl_name       = CTL_UNNUMBERED,
 884                 .procname       = "softlockup_thresh",
 885                 .data           = &softlockup_thresh,
 886                 .maxlen         = sizeof(int),
 887                 .mode           = 0644,
 888                 .proc_handler   = &proc_dosoftlockup_thresh,
 889                 .strategy       = &sysctl_intvec,
 890                 .extra1         = &neg_one,
 891                 .extra2         = &sixty,
 892         },
 893 #endif
 894 #ifdef CONFIG_DETECT_HUNG_TASK
 895         {
 896                 .ctl_name       = CTL_UNNUMBERED,
 897                 .procname       = "hung_task_panic",
 898                 .data           = &sysctl_hung_task_panic,
 899                 .maxlen         = sizeof(int),
 900                 .mode           = 0644,
 901                 .proc_handler   = &proc_dointvec_minmax,
 902                 .strategy       = &sysctl_intvec,
 903                 .extra1         = &zero,
 904                 .extra2         = &one,
 905         },
 906         {
 907                 .ctl_name       = CTL_UNNUMBERED,
 908                 .procname       = "hung_task_check_count",
 909                 .data           = &sysctl_hung_task_check_count,
 910                 .maxlen         = sizeof(unsigned long),
 911                 .mode           = 0644,
 912                 .proc_handler   = &proc_doulongvec_minmax,
 913                 .strategy       = &sysctl_intvec,
 914         },
 915         {
 916                 .ctl_name       = CTL_UNNUMBERED,
 917                 .procname       = "hung_task_timeout_secs",
 918                 .data           = &sysctl_hung_task_timeout_secs,
 919                 .maxlen         = sizeof(unsigned long),
 920                 .mode           = 0644,
 921                 .proc_handler   = &proc_dohung_task_timeout_secs,
 922                 .strategy       = &sysctl_intvec,
 923         },
 924         {
 925                 .ctl_name       = CTL_UNNUMBERED,
 926                 .procname       = "hung_task_warnings",
 927                 .data           = &sysctl_hung_task_warnings,
 928                 .maxlen         = sizeof(unsigned long),
 929                 .mode           = 0644,
 930                 .proc_handler   = &proc_doulongvec_minmax,
 931                 .strategy       = &sysctl_intvec,
 932         },
 933 #endif
 934 #ifdef CONFIG_COMPAT
 935         {
 936                 .ctl_name       = KERN_COMPAT_LOG,
 937                 .procname       = "compat-log",
 938                 .data           = &compat_log,
 939                 .maxlen         = sizeof (int),
 940                 .mode           = 0644,
 941                 .proc_handler   = &proc_dointvec,
 942         },
 943 #endif
 944 #ifdef CONFIG_RT_MUTEXES
 945         {
 946                 .ctl_name       = KERN_MAX_LOCK_DEPTH,
 947                 .procname       = "max_lock_depth",
 948                 .data           = &max_lock_depth,
 949                 .maxlen         = sizeof(int),
 950                 .mode           = 0644,
 951                 .proc_handler   = &proc_dointvec,
 952         },
 953 #endif
 954         {
 955                 .ctl_name       = CTL_UNNUMBERED,
 956                 .procname       = "poweroff_cmd",
 957                 .data           = &poweroff_cmd,
 958                 .maxlen         = POWEROFF_CMD_PATH_LEN,
 959                 .mode           = 0644,
 960                 .proc_handler   = &proc_dostring,
 961                 .strategy       = &sysctl_string,
 962         },
 963 #ifdef CONFIG_KEYS
 964         {
 965                 .ctl_name       = CTL_UNNUMBERED,
 966                 .procname       = "keys",
 967                 .mode           = 0555,
 968                 .child          = key_sysctls,
 969         },
 970 #endif
 971 #ifdef CONFIG_RCU_TORTURE_TEST
 972         {
 973                 .ctl_name       = CTL_UNNUMBERED,
 974                 .procname       = "rcutorture_runnable",
 975                 .data           = &rcutorture_runnable,
 976                 .maxlen         = sizeof(int),
 977                 .mode           = 0644,
 978                 .proc_handler   = &proc_dointvec,
 979         },
 980 #endif
 981 #ifdef CONFIG_SLOW_WORK
 982         {
 983                 .ctl_name       = CTL_UNNUMBERED,
 984                 .procname       = "slow-work",
 985                 .mode           = 0555,
 986                 .child          = slow_work_sysctls,
 987         },
 988 #endif
 989 #ifdef CONFIG_PERF_EVENTS
 990         {
 991                 .ctl_name       = CTL_UNNUMBERED,
 992                 .procname       = "perf_event_paranoid",
 993                 .data           = &sysctl_perf_event_paranoid,
 994                 .maxlen         = sizeof(sysctl_perf_event_paranoid),
 995                 .mode           = 0644,
 996                 .proc_handler   = &proc_dointvec,
 997         },
 998         {
 999                 .ctl_name       = CTL_UNNUMBERED,
1000                 .procname       = "perf_event_mlock_kb",
1001                 .data           = &sysctl_perf_event_mlock,
1002                 .maxlen         = sizeof(sysctl_perf_event_mlock),
1003                 .mode           = 0644,
1004                 .proc_handler   = &proc_dointvec,
1005         },
1006         {
1007                 .ctl_name       = CTL_UNNUMBERED,
1008                 .procname       = "perf_event_max_sample_rate",
1009                 .data           = &sysctl_perf_event_sample_rate,
1010                 .maxlen         = sizeof(sysctl_perf_event_sample_rate),
1011                 .mode           = 0644,
1012                 .proc_handler   = &proc_dointvec,
1013         },
1014 #endif
1015 #ifdef CONFIG_KMEMCHECK
1016         {
1017                 .ctl_name       = CTL_UNNUMBERED,
1018                 .procname       = "kmemcheck",
1019                 .data           = &kmemcheck_enabled,
1020                 .maxlen         = sizeof(int),
1021                 .mode           = 0644,
1022                 .proc_handler   = &proc_dointvec,
1023         },
1024 #endif
1025 #ifdef CONFIG_BLOCK
1026         {
1027                 .ctl_name       = CTL_UNNUMBERED,
1028                 .procname       = "blk_iopoll",
1029                 .data           = &blk_iopoll_enabled,
1030                 .maxlen         = sizeof(int),
1031                 .mode           = 0644,
1032                 .proc_handler   = &proc_dointvec,
1033         },
1034 #endif
1035 /*
1036  * NOTE: do not add new entries to this table unless you have read
1037  * Documentation/sysctl/ctl_unnumbered.txt
1038  */
1039         { .ctl_name = 0 }
1040 };
1041
1042 static struct ctl_table vm_table[] = {
1043         {
1044                 .ctl_name       = VM_OVERCOMMIT_MEMORY,
1045                 .procname       = "overcommit_memory",
1046                 .data           = &sysctl_overcommit_memory,
1047                 .maxlen         = sizeof(sysctl_overcommit_memory),
1048                 .mode           = 0644,
1049                 .proc_handler   = &proc_dointvec,
1050         },
1051         {
1052                 .ctl_name       = VM_PANIC_ON_OOM,
1053                 .procname       = "panic_on_oom",
1054                 .data           = &sysctl_panic_on_oom,
1055                 .maxlen         = sizeof(sysctl_panic_on_oom),
1056                 .mode           = 0644,
1057                 .proc_handler   = &proc_dointvec,
1058         },
1059         {
1060                 .ctl_name       = CTL_UNNUMBERED,
1061                 .procname       = "oom_kill_allocating_task",
1062                 .data           = &sysctl_oom_kill_allocating_task,
1063                 .maxlen         = sizeof(sysctl_oom_kill_allocating_task),
1064                 .mode           = 0644,
1065                 .proc_handler   = &proc_dointvec,
1066         },
1067         {
1068                 .ctl_name       = CTL_UNNUMBERED,
1069                 .procname       = "oom_dump_tasks",
1070                 .data           = &sysctl_oom_dump_tasks,
1071                 .maxlen         = sizeof(sysctl_oom_dump_tasks),
1072                 .mode           = 0644,
1073                 .proc_handler   = &proc_dointvec,
1074         },
1075         {
1076                 .ctl_name       = VM_OVERCOMMIT_RATIO,
1077                 .procname       = "overcommit_ratio",
1078                 .data           = &sysctl_overcommit_ratio,
1079                 .maxlen         = sizeof(sysctl_overcommit_ratio),
1080                 .mode           = 0644,
1081                 .proc_handler   = &proc_dointvec,
1082         },
1083         {
1084                 .ctl_name       = VM_PAGE_CLUSTER,
1085                 .procname       = "page-cluster",
1086                 .data           = &page_cluster,
1087                 .maxlen         = sizeof(int),
1088                 .mode           = 0644,
1089                 .proc_handler   = &proc_dointvec,
1090         },
1091         {
1092                 .ctl_name       = VM_DIRTY_BACKGROUND,
1093                 .procname       = "dirty_background_ratio",
1094                 .data           = &dirty_background_ratio,
1095                 .maxlen         = sizeof(dirty_background_ratio),
1096                 .mode           = 0644,
1097                 .proc_handler   = &dirty_background_ratio_handler,
1098                 .strategy       = &sysctl_intvec,
1099                 .extra1         = &zero,
1100                 .extra2         = &one_hundred,
1101         },
1102         {
1103                 .ctl_name       = CTL_UNNUMBERED,
1104                 .procname       = "dirty_background_bytes",
1105                 .data           = &dirty_background_bytes,
1106                 .maxlen         = sizeof(dirty_background_bytes),
1107                 .mode           = 0644,
1108                 .proc_handler   = &dirty_background_bytes_handler,
1109                 .strategy       = &sysctl_intvec,
1110                 .extra1         = &one_ul,
1111         },
1112         {
1113                 .ctl_name       = VM_DIRTY_RATIO,
1114                 .procname       = "dirty_ratio",
1115                 .data           = &vm_dirty_ratio,
1116                 .maxlen         = sizeof(vm_dirty_ratio),
1117                 .mode           = 0644,
1118                 .proc_handler   = &dirty_ratio_handler,
1119                 .strategy       = &sysctl_intvec,
1120                 .extra1         = &zero,
1121                 .extra2         = &one_hundred,
1122         },
1123         {
1124                 .ctl_name       = CTL_UNNUMBERED,
1125                 .procname       = "dirty_bytes",
1126                 .data           = &vm_dirty_bytes,
1127                 .maxlen         = sizeof(vm_dirty_bytes),
1128                 .mode           = 0644,
1129                 .proc_handler   = &dirty_bytes_handler,
1130                 .strategy       = &sysctl_intvec,
1131                 .extra1         = &dirty_bytes_min,
1132         },
1133         {
1134                 .procname       = "dirty_writeback_centisecs",
1135                 .data           = &dirty_writeback_interval,
1136                 .maxlen         = sizeof(dirty_writeback_interval),
1137                 .mode           = 0644,
1138                 .proc_handler   = &dirty_writeback_centisecs_handler,
1139         },
1140         {
1141                 .procname       = "dirty_expire_centisecs",
1142                 .data           = &dirty_expire_interval,
1143                 .maxlen         = sizeof(dirty_expire_interval),
1144                 .mode           = 0644,
1145                 .proc_handler   = &proc_dointvec,
1146         },
1147         {
1148                 .ctl_name       = VM_NR_PDFLUSH_THREADS,
1149                 .procname       = "nr_pdflush_threads",
1150                 .data           = &nr_pdflush_threads,
1151                 .maxlen         = sizeof nr_pdflush_threads,
1152                 .mode           = 0444 /* read-only*/,
1153                 .proc_handler   = &proc_dointvec,
1154         },
1155         {
1156                 .ctl_name       = VM_SWAPPINESS,
1157                 .procname       = "swappiness",
1158                 .data           = &vm_swappiness,
1159                 .maxlen         = sizeof(vm_swappiness),
1160                 .mode           = 0644,
1161                 .proc_handler   = &proc_dointvec_minmax,
1162                 .strategy       = &sysctl_intvec,
1163                 .extra1         = &zero,
1164                 .extra2         = &one_hundred,
1165         },
1166 #ifdef CONFIG_HUGETLB_PAGE
1167          {
1168                 .procname       = "nr_hugepages",
1169                 .data           = NULL,
1170                 .maxlen         = sizeof(unsigned long),
1171                 .mode           = 0644,
1172                 .proc_handler   = &hugetlb_sysctl_handler,
1173                 .extra1         = (void *)&hugetlb_zero,
1174                 .extra2         = (void *)&hugetlb_infinity,
1175          },
1176          {
1177                 .ctl_name       = VM_HUGETLB_GROUP,
1178                 .procname       = "hugetlb_shm_group",
1179                 .data           = &sysctl_hugetlb_shm_group,
1180                 .maxlen         = sizeof(gid_t),
1181                 .mode           = 0644,
1182                 .proc_handler   = &proc_dointvec,
1183          },
1184          {
1185                 .ctl_name       = CTL_UNNUMBERED,
1186                 .procname       = "hugepages_treat_as_movable",
1187                 .data           = &hugepages_treat_as_movable,
1188                 .maxlen         = sizeof(int),
1189                 .mode           = 0644,
1190                 .proc_handler   = &hugetlb_treat_movable_handler,
1191         },
1192         {
1193                 .ctl_name       = CTL_UNNUMBERED,
1194                 .procname       = "nr_overcommit_hugepages",
1195                 .data           = NULL,
1196                 .maxlen         = sizeof(unsigned long),
1197                 .mode           = 0644,
1198                 .proc_handler   = &hugetlb_overcommit_handler,
1199                 .extra1         = (void *)&hugetlb_zero,
1200                 .extra2         = (void *)&hugetlb_infinity,
1201         },
1202 #endif
1203         {
1204                 .ctl_name       = VM_LOWMEM_RESERVE_RATIO,
1205                 .procname       = "lowmem_reserve_ratio",
1206                 .data           = &sysctl_lowmem_reserve_ratio,
1207                 .maxlen         = sizeof(sysctl_lowmem_reserve_ratio),
1208                 .mode           = 0644,
1209                 .proc_handler   = &lowmem_reserve_ratio_sysctl_handler,
1210                 .strategy       = &sysctl_intvec,
1211         },
1212         {
1213                 .ctl_name       = VM_DROP_PAGECACHE,
1214                 .procname       = "drop_caches",
1215                 .data           = &sysctl_drop_caches,
1216                 .maxlen         = sizeof(int),
1217                 .mode           = 0644,
1218                 .proc_handler   = drop_caches_sysctl_handler,
1219                 .strategy       = &sysctl_intvec,
1220         },
1221         {
1222                 .ctl_name       = VM_MIN_FREE_KBYTES,
1223                 .procname       = "min_free_kbytes",
1224                 .data           = &min_free_kbytes,
1225                 .maxlen         = sizeof(min_free_kbytes),
1226                 .mode           = 0644,
1227                 .proc_handler   = &min_free_kbytes_sysctl_handler,
1228                 .strategy       = &sysctl_intvec,
1229                 .extra1         = &zero,
1230         },
1231         {
1232                 .ctl_name       = VM_PERCPU_PAGELIST_FRACTION,
1233                 .procname       = "percpu_pagelist_fraction",
1234                 .data           = &percpu_pagelist_fraction,
1235                 .maxlen         = sizeof(percpu_pagelist_fraction),
1236                 .mode           = 0644,
1237                 .proc_handler   = &percpu_pagelist_fraction_sysctl_handler,
1238                 .strategy       = &sysctl_intvec,
1239                 .extra1         = &min_percpu_pagelist_fract,
1240         },
1241 #ifdef CONFIG_MMU
1242         {
1243                 .ctl_name       = VM_MAX_MAP_COUNT,
1244                 .procname       = "max_map_count",
1245                 .data           = &sysctl_max_map_count,
1246                 .maxlen         = sizeof(sysctl_max_map_count),
1247                 .mode           = 0644,
1248                 .proc_handler   = &proc_dointvec
1249         },
1250 #else
1251         {
1252                 .ctl_name       = CTL_UNNUMBERED,
1253                 .procname       = "nr_trim_pages",
1254                 .data           = &sysctl_nr_trim_pages,
1255                 .maxlen         = sizeof(sysctl_nr_trim_pages),
1256                 .mode           = 0644,
1257                 .proc_handler   = &proc_dointvec_minmax,
1258                 .strategy       = &sysctl_intvec,
1259                 .extra1         = &zero,
1260         },
1261 #endif
1262         {
1263                 .ctl_name       = VM_LAPTOP_MODE,
1264                 .procname       = "laptop_mode",
1265                 .data           = &laptop_mode,
1266                 .maxlen         = sizeof(laptop_mode),
1267                 .mode           = 0644,
1268                 .proc_handler   = &proc_dointvec_jiffies,
1269                 .strategy       = &sysctl_jiffies,
1270         },
1271         {
1272                 .ctl_name       = VM_BLOCK_DUMP,
1273                 .procname       = "block_dump",
1274                 .data           = &block_dump,
1275                 .maxlen         = sizeof(block_dump),
1276                 .mode           = 0644,
1277                 .proc_handler   = &proc_dointvec,
1278                 .strategy       = &sysctl_intvec,
1279                 .extra1         = &zero,
1280         },
1281         {
1282                 .ctl_name       = VM_VFS_CACHE_PRESSURE,
1283                 .procname       = "vfs_cache_pressure",
1284                 .data           = &sysctl_vfs_cache_pressure,
1285                 .maxlen         = sizeof(sysctl_vfs_cache_pressure),
1286                 .mode           = 0644,
1287                 .proc_handler   = &proc_dointvec,
1288                 .strategy       = &sysctl_intvec,
1289                 .extra1         = &zero,
1290         },
1291 #ifdef HAVE_ARCH_PICK_MMAP_LAYOUT
1292         {
1293                 .ctl_name       = VM_LEGACY_VA_LAYOUT,
1294                 .procname       = "legacy_va_layout",
1295                 .data           = &sysctl_legacy_va_layout,
1296                 .maxlen         = sizeof(sysctl_legacy_va_layout),
1297                 .mode           = 0644,
1298                 .proc_handler   = &proc_dointvec,
1299                 .strategy       = &sysctl_intvec,
1300                 .extra1         = &zero,
1301         },
1302 #endif
1303 #ifdef CONFIG_NUMA
1304         {
1305                 .ctl_name       = VM_ZONE_RECLAIM_MODE,
1306                 .procname       = "zone_reclaim_mode",
1307                 .data           = &zone_reclaim_mode,
1308                 .maxlen         = sizeof(zone_reclaim_mode),
1309                 .mode           = 0644,
1310                 .proc_handler   = &proc_dointvec,
1311                 .strategy       = &sysctl_intvec,
1312                 .extra1         = &zero,
1313         },
1314         {
1315                 .ctl_name       = VM_MIN_UNMAPPED,
1316                 .procname       = "min_unmapped_ratio",
1317                 .data           = &sysctl_min_unmapped_ratio,
1318                 .maxlen         = sizeof(sysctl_min_unmapped_ratio),
1319                 .mode           = 0644,
1320                 .proc_handler   = &sysctl_min_unmapped_ratio_sysctl_handler,
1321                 .strategy       = &sysctl_intvec,
1322                 .extra1         = &zero,
1323                 .extra2         = &one_hundred,
1324         },
1325         {
1326                 .ctl_name       = VM_MIN_SLAB,
1327                 .procname       = "min_slab_ratio",
1328                 .data           = &sysctl_min_slab_ratio,
1329                 .maxlen         = sizeof(sysctl_min_slab_ratio),
1330                 .mode           = 0644,
1331                 .proc_handler   = &sysctl_min_slab_ratio_sysctl_handler,
1332                 .strategy       = &sysctl_intvec,
1333                 .extra1         = &zero,
1334                 .extra2         = &one_hundred,
1335         },
1336 #endif
1337 #ifdef CONFIG_SMP
1338         {
1339                 .ctl_name       = CTL_UNNUMBERED,
1340                 .procname       = "stat_interval",
1341                 .data           = &sysctl_stat_interval,
1342                 .maxlen         = sizeof(sysctl_stat_interval),
1343                 .mode           = 0644,
1344                 .proc_handler   = &proc_dointvec_jiffies,
1345                 .strategy       = &sysctl_jiffies,
1346         },
1347 #endif
1348         {
1349                 .ctl_name       = CTL_UNNUMBERED,
1350                 .procname       = "mmap_min_addr",
1351                 .data           = &dac_mmap_min_addr,
1352                 .maxlen         = sizeof(unsigned long),
1353                 .mode           = 0644,
1354                 .proc_handler   = &mmap_min_addr_handler,
1355         },
1356 #ifdef CONFIG_NUMA
1357         {
1358                 .ctl_name       = CTL_UNNUMBERED,
1359                 .procname       = "numa_zonelist_order",
1360                 .data           = &numa_zonelist_order,
1361                 .maxlen         = NUMA_ZONELIST_ORDER_LEN,
1362                 .mode           = 0644,
1363                 .proc_handler   = &numa_zonelist_order_handler,
1364                 .strategy       = &sysctl_string,
1365         },
1366 #endif
1367 #if (defined(CONFIG_X86_32) && !defined(CONFIG_UML))|| \
1368    (defined(CONFIG_SUPERH) && defined(CONFIG_VSYSCALL))
1369         {
1370                 .ctl_name       = VM_VDSO_ENABLED,
1371                 .procname       = "vdso_enabled",
1372                 .data           = &vdso_enabled,
1373                 .maxlen         = sizeof(vdso_enabled),
1374                 .mode           = 0644,
1375                 .proc_handler   = &proc_dointvec,
1376                 .strategy       = &sysctl_intvec,
1377                 .extra1         = &zero,
1378         },
1379 #endif
1380 #ifdef CONFIG_HIGHMEM
1381         {
1382                 .ctl_name       = CTL_UNNUMBERED,
1383                 .procname       = "highmem_is_dirtyable",
1384                 .data           = &vm_highmem_is_dirtyable,
1385                 .maxlen         = sizeof(vm_highmem_is_dirtyable),
1386                 .mode           = 0644,
1387                 .proc_handler   = &proc_dointvec_minmax,
1388                 .strategy       = &sysctl_intvec,
1389                 .extra1         = &zero,
1390                 .extra2         = &one,
1391         },
1392 #endif
1393         {
1394                 .ctl_name       = CTL_UNNUMBERED,
1395                 .procname       = "scan_unevictable_pages",
1396                 .data           = &scan_unevictable_pages,
1397                 .maxlen         = sizeof(scan_unevictable_pages),
1398                 .mode           = 0644,
1399                 .proc_handler   = &scan_unevictable_handler,
1400         },
1401 /*
1402  * NOTE: do not add new entries to this table unless you have read
1403  * Documentation/sysctl/ctl_unnumbered.txt
1404  */
1405         { .ctl_name = 0 }
1406 };
1407
1408 #if defined(CONFIG_BINFMT_MISC) || defined(CONFIG_BINFMT_MISC_MODULE)
1409 static struct ctl_table binfmt_misc_table[] = {
1410         { .ctl_name = 0 }
1411 };
1412 #endif
1413
1414 static struct ctl_table fs_table[] = {
1415         {
1416                 .ctl_name       = FS_NRINODE,
1417                 .procname       = "inode-nr",
1418                 .data           = &inodes_stat,
1419                 .maxlen         = 2*sizeof(int),
1420                 .mode           = 0444,
1421                 .proc_handler   = &proc_dointvec,
1422         },
1423         {
1424                 .ctl_name       = FS_STATINODE,
1425                 .procname       = "inode-state",
1426                 .data           = &inodes_stat,
1427                 .maxlen         = 7*sizeof(int),
1428                 .mode           = 0444,
1429                 .proc_handler   = &proc_dointvec,
1430         },
1431         {
1432                 .procname       = "file-nr",
1433                 .data           = &files_stat,
1434                 .maxlen         = 3*sizeof(int),
1435                 .mode           = 0444,
1436                 .proc_handler   = &proc_nr_files,
1437         },
1438         {
1439                 .ctl_name       = FS_MAXFILE,
1440                 .procname       = "file-max",
1441                 .data           = &files_stat.max_files,
1442                 .maxlen         = sizeof(int),
1443                 .mode           = 0644,
1444                 .proc_handler   = &proc_dointvec,
1445         },
1446         {
1447                 .ctl_name       = CTL_UNNUMBERED,
1448                 .procname       = "nr_open",
1449                 .data           = &sysctl_nr_open,
1450                 .maxlen         = sizeof(int),
1451                 .mode           = 0644,
1452                 .proc_handler   = &proc_dointvec_minmax,
1453                 .extra1         = &sysctl_nr_open_min,
1454                 .extra2         = &sysctl_nr_open_max,
1455         },
1456         {
1457                 .ctl_name       = FS_DENTRY,
1458                 .procname       = "dentry-state",
1459                 .data           = &dentry_stat,
1460                 .maxlen         = 6*sizeof(int),
1461                 .mode           = 0444,
1462                 .proc_handler   = &proc_dointvec,
1463         },
1464         {
1465                 .ctl_name       = FS_OVERFLOWUID,
1466                 .procname       = "overflowuid",
1467                 .data           = &fs_overflowuid,
1468                 .maxlen         = sizeof(int),
1469                 .mode           = 0644,
1470                 .proc_handler   = &proc_dointvec_minmax,
1471                 .strategy       = &sysctl_intvec,
1472                 .extra1         = &minolduid,
1473                 .extra2         = &maxolduid,
1474         },
1475         {
1476                 .ctl_name       = FS_OVERFLOWGID,
1477                 .procname       = "overflowgid",
1478                 .data           = &fs_overflowgid,
1479                 .maxlen         = sizeof(int),
1480                 .mode           = 0644,
1481                 .proc_handler   = &proc_dointvec_minmax,
1482                 .strategy       = &sysctl_intvec,
1483                 .extra1         = &minolduid,
1484                 .extra2         = &maxolduid,
1485         },
1486 #ifdef CONFIG_FILE_LOCKING
1487         {
1488                 .ctl_name       = FS_LEASES,
1489                 .procname       = "leases-enable",
1490                 .data           = &leases_enable,
1491                 .maxlen         = sizeof(int),
1492                 .mode           = 0644,
1493                 .proc_handler   = &proc_dointvec,
1494         },
1495 #endif
1496 #ifdef CONFIG_DNOTIFY
1497         {
1498                 .ctl_name       = FS_DIR_NOTIFY,
1499                 .procname       = "dir-notify-enable",
1500                 .data           = &dir_notify_enable,
1501                 .maxlen         = sizeof(int),
1502                 .mode           = 0644,
1503                 .proc_handler   = &proc_dointvec,
1504         },
1505 #endif
1506 #ifdef CONFIG_MMU
1507 #ifdef CONFIG_FILE_LOCKING
1508         {
1509                 .ctl_name       = FS_LEASE_TIME,
1510                 .procname       = "lease-break-time",
1511                 .data           = &lease_break_time,
1512                 .maxlen         = sizeof(int),
1513                 .mode           = 0644,
1514                 .proc_handler   = &proc_dointvec,
1515         },
1516 #endif
1517 #ifdef CONFIG_AIO
1518         {
1519                 .procname       = "aio-nr",
1520                 .data           = &aio_nr,
1521                 .maxlen         = sizeof(aio_nr),
1522                 .mode           = 0444,
1523                 .proc_handler   = &proc_doulongvec_minmax,
1524         },
1525         {
1526                 .procname       = "aio-max-nr",
1527                 .data           = &aio_max_nr,
1528                 .maxlen         = sizeof(aio_max_nr),
1529                 .mode           = 0644,
1530                 .proc_handler   = &proc_doulongvec_minmax,
1531         },
1532 #endif /* CONFIG_AIO */
1533 #ifdef CONFIG_INOTIFY_USER
1534         {
1535                 .ctl_name       = FS_INOTIFY,
1536                 .procname       = "inotify",
1537                 .mode           = 0555,
1538                 .child          = inotify_table,
1539         },
1540 #endif
1541 #ifdef CONFIG_EPOLL
1542         {
1543                 .procname       = "epoll",
1544                 .mode           = 0555,
1545                 .child          = epoll_table,
1546         },
1547 #endif
1548 #endif
1549         {
1550                 .ctl_name       = KERN_SETUID_DUMPABLE,
1551                 .procname       = "suid_dumpable",
1552                 .data           = &suid_dumpable,
1553                 .maxlen         = sizeof(int),
1554                 .mode           = 0644,
1555                 .proc_handler   = &proc_dointvec_minmax,
1556                 .strategy       = &sysctl_intvec,
1557                 .extra1         = &zero,
1558                 .extra2         = &two,
1559         },
1560 #if defined(CONFIG_BINFMT_MISC) || defined(CONFIG_BINFMT_MISC_MODULE)
1561         {
1562                 .ctl_name       = CTL_UNNUMBERED,
1563                 .procname       = "binfmt_misc",
1564                 .mode           = 0555,
1565                 .child          = binfmt_misc_table,
1566         },
1567 #endif
1568 /*
1569  * NOTE: do not add new entries to this table unless you have read
1570  * Documentation/sysctl/ctl_unnumbered.txt
1571  */
1572         { .ctl_name = 0 }
1573 };
1574
1575 static struct ctl_table debug_table[] = {
1576 #if defined(CONFIG_X86) || defined(CONFIG_PPC)
1577         {
1578                 .ctl_name       = CTL_UNNUMBERED,
1579                 .procname       = "exception-trace",
1580                 .data           = &show_unhandled_signals,
1581                 .maxlen         = sizeof(int),
1582                 .mode           = 0644,
1583                 .proc_handler   = proc_dointvec
1584         },
1585 #endif
1586         { .ctl_name = 0 }
1587 };
1588
1589 static struct ctl_table dev_table[] = {
1590         { .ctl_name = 0 }
1591 };
1592
1593 static DEFINE_SPINLOCK(sysctl_lock);
1594
1595 /* called under sysctl_lock */
1596 static int use_table(struct ctl_table_header *p)
1597 {
1598         if (unlikely(p->unregistering))
1599                 return 0;
1600         p->used++;
1601         return 1;
1602 }
1603
1604 /* called under sysctl_lock */
1605 static void unuse_table(struct ctl_table_header *p)
1606 {
1607         if (!--p->used)
1608                 if (unlikely(p->unregistering))
1609                         complete(p->unregistering);
1610 }
1611
1612 /* called under sysctl_lock, will reacquire if has to wait */
1613 static void start_unregistering(struct ctl_table_header *p)
1614 {
1615         /*
1616          * if p->used is 0, nobody will ever touch that entry again;
1617          * we'll eliminate all paths to it before dropping sysctl_lock
1618          */
1619         if (unlikely(p->used)) {
1620                 struct completion wait;
1621                 init_completion(&wait);
1622                 p->unregistering = &wait;
1623                 spin_unlock(&sysctl_lock);
1624                 wait_for_completion(&wait);
1625                 spin_lock(&sysctl_lock);
1626         } else {
1627                 /* anything non-NULL; we'll never dereference it */
1628                 p->unregistering = ERR_PTR(-EINVAL);
1629         }
1630         /*
1631          * do not remove from the list until nobody holds it; walking the
1632          * list in do_sysctl() relies on that.
1633          */
1634         list_del_init(&p->ctl_entry);
1635 }
1636
1637 void sysctl_head_get(struct ctl_table_header *head)
1638 {
1639         spin_lock(&sysctl_lock);
1640         head->count++;
1641         spin_unlock(&sysctl_lock);
1642 }
1643
1644 void sysctl_head_put(struct ctl_table_header *head)
1645 {
1646         spin_lock(&sysctl_lock);
1647         if (!--head->count)
1648                 kfree(head);
1649         spin_unlock(&sysctl_lock);
1650 }
1651
1652 struct ctl_table_header *sysctl_head_grab(struct ctl_table_header *head)
1653 {
1654         if (!head)
1655                 BUG();
1656         spin_lock(&sysctl_lock);
1657         if (!use_table(head))
1658                 head = ERR_PTR(-ENOENT);
1659         spin_unlock(&sysctl_lock);
1660         return head;
1661 }
1662
1663 void sysctl_head_finish(struct ctl_table_header *head)
1664 {
1665         if (!head)
1666                 return;
1667         spin_lock(&sysctl_lock);
1668         unuse_table(head);
1669         spin_unlock(&sysctl_lock);
1670 }
1671
1672 static struct ctl_table_set *
1673 lookup_header_set(struct ctl_table_root *root, struct nsproxy *namespaces)
1674 {
1675         struct ctl_table_set *set = &root->default_set;
1676         if (root->lookup)
1677                 set = root->lookup(root, namespaces);
1678         return set;
1679 }
1680
1681 static struct list_head *
1682 lookup_header_list(struct ctl_table_root *root, struct nsproxy *namespaces)
1683 {
1684         struct ctl_table_set *set = lookup_header_set(root, namespaces);
1685         return &set->list;
1686 }
1687
1688 struct ctl_table_header *__sysctl_head_next(struct nsproxy *namespaces,
1689                                             struct ctl_table_header *prev)
1690 {
1691         struct ctl_table_root *root;
1692         struct list_head *header_list;
1693         struct ctl_table_header *head;
1694         struct list_head *tmp;
1695
1696         spin_lock(&sysctl_lock);
1697         if (prev) {
1698                 head = prev;
1699                 tmp = &prev->ctl_entry;
1700                 unuse_table(prev);
1701                 goto next;
1702         }
1703         tmp = &root_table_header.ctl_entry;
1704         for (;;) {
1705                 head = list_entry(tmp, struct ctl_table_header, ctl_entry);
1706
1707                 if (!use_table(head))
1708                         goto next;
1709                 spin_unlock(&sysctl_lock);
1710                 return head;
1711         next:
1712                 root = head->root;
1713                 tmp = tmp->next;
1714                 header_list = lookup_header_list(root, namespaces);
1715                 if (tmp != header_list)
1716                         continue;
1717
1718                 do {
1719                         root = list_entry(root->root_list.next,
1720                                         struct ctl_table_root, root_list);
1721                         if (root == &sysctl_table_root)
1722                                 goto out;
1723                         header_list = lookup_header_list(root, namespaces);
1724                 } while (list_empty(header_list));
1725                 tmp = header_list->next;
1726         }
1727 out:
1728         spin_unlock(&sysctl_lock);
1729         return NULL;
1730 }
1731
1732 struct ctl_table_header *sysctl_head_next(struct ctl_table_header *prev)
1733 {
1734         return __sysctl_head_next(current->nsproxy, prev);
1735 }
1736
1737 void register_sysctl_root(struct ctl_table_root *root)
1738 {
1739         spin_lock(&sysctl_lock);
1740         list_add_tail(&root->root_list, &sysctl_table_root.root_list);
1741         spin_unlock(&sysctl_lock);
1742 }
1743
1744 #ifdef CONFIG_SYSCTL_SYSCALL
1745 /* Perform the actual read/write of a sysctl table entry. */
1746 static int do_sysctl_strategy(struct ctl_table_root *root,
1747                         struct ctl_table *table,
1748                         void __user *oldval, size_t __user *oldlenp,
1749                         void __user *newval, size_t newlen)
1750 {
1751         int op = 0, rc;
1752
1753         if (oldval)
1754                 op |= MAY_READ;
1755         if (newval)
1756                 op |= MAY_WRITE;
1757         if (sysctl_perm(root, table, op))
1758                 return -EPERM;
1759
1760         if (table->strategy) {
1761                 rc = table->strategy(table, oldval, oldlenp, newval, newlen);
1762                 if (rc < 0)
1763                         return rc;
1764                 if (rc > 0)
1765                         return 0;
1766         }
1767
1768         /* If there is no strategy routine, or if the strategy returns
1769          * zero, proceed with automatic r/w */
1770         if (table->data && table->maxlen) {
1771                 rc = sysctl_data(table, oldval, oldlenp, newval, newlen);
1772                 if (rc < 0)
1773                         return rc;
1774         }
1775         return 0;
1776 }
1777
1778 static int parse_table(int __user *name, int nlen,
1779                        void __user *oldval, size_t __user *oldlenp,
1780                        void __user *newval, size_t newlen,
1781                        struct ctl_table_root *root,
1782                        struct ctl_table *table)
1783 {
1784         int n;
1785 repeat:
1786         if (!nlen)
1787                 return -ENOTDIR;
1788         if (get_user(n, name))
1789                 return -EFAULT;
1790         for ( ; table->ctl_name || table->procname; table++) {
1791                 if (!table->ctl_name)
1792                         continue;
1793                 if (n == table->ctl_name) {
1794                         int error;
1795                         if (table->child) {
1796                                 if (sysctl_perm(root, table, MAY_EXEC))
1797                                         return -EPERM;
1798                                 name++;
1799                                 nlen--;
1800                                 table = table->child;
1801                                 goto repeat;
1802                         }
1803                         error = do_sysctl_strategy(root, table,
1804                                                    oldval, oldlenp,
1805                                                    newval, newlen);
1806                         return error;
1807                 }
1808         }
1809         return -ENOTDIR;
1810 }
1811
1812 int do_sysctl(int __user *name, int nlen, void __user *oldval, size_t __user *oldlenp,
1813                void __user *newval, size_t newlen)
1814 {
1815         struct ctl_table_header *head;
1816         int error = -ENOTDIR;
1817
1818         if (nlen <= 0 || nlen >= CTL_MAXNAME)
1819                 return -ENOTDIR;
1820         if (oldval) {
1821                 int old_len;
1822                 if (!oldlenp || get_user(old_len, oldlenp))
1823                         return -EFAULT;
1824         }
1825
1826         for (head = sysctl_head_next(NULL); head;
1827                         head = sysctl_head_next(head)) {
1828                 error = parse_table(name, nlen, oldval, oldlenp,
1829                                         newval, newlen,
1830                                         head->root, head->ctl_table);
1831                 if (error != -ENOTDIR) {
1832                         sysctl_head_finish(head);
1833                         break;
1834                 }
1835         }
1836         return error;
1837 }
1838
1839 SYSCALL_DEFINE1(sysctl, struct __sysctl_args __user *, args)
1840 {
1841         struct __sysctl_args tmp;
1842         int error;
1843
1844         if (copy_from_user(&tmp, args, sizeof(tmp)))
1845                 return -EFAULT;
1846
1847         error = deprecated_sysctl_warning(&tmp);
1848         if (error)
1849                 goto out;
1850
1851         lock_kernel();
1852         error = do_sysctl(tmp.name, tmp.nlen, tmp.oldval, tmp.oldlenp,
1853                           tmp.newval, tmp.newlen);
1854         unlock_kernel();
1855 out:
1856         return error;
1857 }
1858 #endif /* CONFIG_SYSCTL_SYSCALL */
1859
1860 /*
1861  * sysctl_perm does NOT grant the superuser all rights automatically, because
1862  * some sysctl variables are readonly even to root.
1863  */
1864
1865 static int test_perm(int mode, int op)
1866 {
1867         if (!current_euid())
1868                 mode >>= 6;
1869         else if (in_egroup_p(0))
1870                 mode >>= 3;
1871         if ((op & ~mode & (MAY_READ|MAY_WRITE|MAY_EXEC)) == 0)
1872                 return 0;
1873         return -EACCES;
1874 }
1875
1876 int sysctl_perm(struct ctl_table_root *root, struct ctl_table *table, int op)
1877 {
1878         int error;
1879         int mode;
1880
1881         error = security_sysctl(table, op & (MAY_READ | MAY_WRITE | MAY_EXEC));
1882         if (error)
1883                 return error;
1884
1885         if (root->permissions)
1886                 mode = root->permissions(root, current->nsproxy, table);
1887         else
1888                 mode = table->mode;
1889
1890         return test_perm(mode, op);
1891 }
1892
1893 static void sysctl_set_parent(struct ctl_table *parent, struct ctl_table *table)
1894 {
1895         for (; table->ctl_name || table->procname; table++) {
1896                 table->parent = parent;
1897                 if (table->child)
1898                         sysctl_set_parent(table, table->child);
1899         }
1900 }
1901
1902 static __init int sysctl_init(void)
1903 {
1904         sysctl_set_parent(NULL, root_table);
1905 #ifdef CONFIG_SYSCTL_SYSCALL_CHECK
1906         {
1907                 int err;
1908                 err = sysctl_check_table(current->nsproxy, root_table);
1909         }
1910 #endif
1911         return 0;
1912 }
1913
1914 core_initcall(sysctl_init);
1915
1916 static struct ctl_table *is_branch_in(struct ctl_table *branch,
1917                                       struct ctl_table *table)
1918 {
1919         struct ctl_table *p;
1920         const char *s = branch->procname;
1921
1922         /* branch should have named subdirectory as its first element */
1923         if (!s || !branch->child)
1924                 return NULL;
1925
1926         /* ... and nothing else */
1927         if (branch[1].procname || branch[1].ctl_name)
1928                 return NULL;
1929
1930         /* table should contain subdirectory with the same name */
1931         for (p = table; p->procname || p->ctl_name; p++) {
1932                 if (!p->child)
1933                         continue;
1934                 if (p->procname && strcmp(p->procname, s) == 0)
1935                         return p;
1936         }
1937         return NULL;
1938 }
1939
1940 /* see if attaching q to p would be an improvement */
1941 static void try_attach(struct ctl_table_header *p, struct ctl_table_header *q)
1942 {
1943         struct ctl_table *to = p->ctl_table, *by = q->ctl_table;
1944         struct ctl_table *next;
1945         int is_better = 0;
1946         int not_in_parent = !p->attached_by;
1947
1948         while ((next = is_branch_in(by, to)) != NULL) {
1949                 if (by == q->attached_by)
1950                         is_better = 1;
1951                 if (to == p->attached_by)
1952                         not_in_parent = 1;
1953                 by = by->child;
1954                 to = next->child;
1955         }
1956
1957         if (is_better && not_in_parent) {
1958                 q->attached_by = by;
1959                 q->attached_to = to;
1960                 q->parent = p;
1961         }
1962 }
1963
1964 /**
1965  * __register_sysctl_paths - register a sysctl hierarchy
1966  * @root: List of sysctl headers to register on
1967  * @namespaces: Data to compute which lists of sysctl entries are visible
1968  * @path: The path to the directory the sysctl table is in.
1969  * @table: the top-level table structure
1970  *
1971  * Register a sysctl table hierarchy. @table should be a filled in ctl_table
1972  * array. A completely 0 filled entry terminates the table.
1973  *
1974  * The members of the &struct ctl_table structure are used as follows:
1975  *
1976  * ctl_name - This is the numeric sysctl value used by sysctl(2). The number
1977  *            must be unique within that level of sysctl
1978  *
1979  * procname - the name of the sysctl file under /proc/sys. Set to %NULL to not
1980  *            enter a sysctl file
1981  *
1982  * data - a pointer to data for use by proc_handler
1983  *
1984  * maxlen - the maximum size in bytes of the data
1985  *
1986  * mode - the file permissions for the /proc/sys file, and for sysctl(2)
1987  *
1988  * child - a pointer to the child sysctl table if this entry is a directory, or
1989  *         %NULL.
1990  *
1991  * proc_handler - the text handler routine (described below)
1992  *
1993  * strategy - the strategy routine (described below)
1994  *
1995  * de - for internal use by the sysctl routines
1996  *
1997  * extra1, extra2 - extra pointers usable by the proc handler routines
1998  *
1999  * Leaf nodes in the sysctl tree will be represented by a single file
2000  * under /proc; non-leaf nodes will be represented by directories.
2001  *
2002  * sysctl(2) can automatically manage read and write requests through
2003  * the sysctl table.  The data and maxlen fields of the ctl_table
2004  * struct enable minimal validation of the values being written to be
2005  * performed, and the mode field allows minimal authentication.
2006  *
2007  * More sophisticated management can be enabled by the provision of a
2008  * strategy routine with the table entry.  This will be called before
2009  * any automatic read or write of the data is performed.
2010  *
2011  * The strategy routine may return
2012  *
2013  * < 0 - Error occurred (error is passed to user process)
2014  *
2015  * 0   - OK - proceed with automatic read or write.
2016  *
2017  * > 0 - OK - read or write has been done by the strategy routine, so
2018  *       return immediately.
2019  *
2020  * There must be a proc_handler routine for any terminal nodes
2021  * mirrored under /proc/sys (non-terminals are handled by a built-in
2022  * directory handler).  Several default handlers are available to
2023  * cover common cases -
2024  *
2025  * proc_dostring(), proc_dointvec(), proc_dointvec_jiffies(),
2026  * proc_dointvec_userhz_jiffies(), proc_dointvec_minmax(),
2027  * proc_doulongvec_ms_jiffies_minmax(), proc_doulongvec_minmax()
2028  *
2029  * It is the handler's job to read the input buffer from user memory
2030  * and process it. The handler should return 0 on success.
2031  *
2032  * This routine returns %NULL on a failure to register, and a pointer
2033  * to the table header on success.
2034  */
2035 struct ctl_table_header *__register_sysctl_paths(
2036         struct ctl_table_root *root,
2037         struct nsproxy *namespaces,
2038         const struct ctl_path *path, struct ctl_table *table)
2039 {
2040         struct ctl_table_header *header;
2041         struct ctl_table *new, **prevp;
2042         unsigned int n, npath;
2043         struct ctl_table_set *set;
2044
2045         /* Count the path components */
2046         for (npath = 0; path[npath].ctl_name || path[npath].procname; ++npath)
2047                 ;
2048
2049         /*
2050          * For each path component, allocate a 2-element ctl_table array.
2051          * The first array element will be filled with the sysctl entry
2052          * for this, the second will be the sentinel (ctl_name == 0).
2053          *
2054          * We allocate everything in one go so that we don't have to
2055          * worry about freeing additional memory in unregister_sysctl_table.
2056          */
2057         header = kzalloc(sizeof(struct ctl_table_header) +
2058                          (2 * npath * sizeof(struct ctl_table)), GFP_KERNEL);
2059         if (!header)
2060                 return NULL;
2061
2062         new = (struct ctl_table *) (header + 1);
2063
2064         /* Now connect the dots */
2065         prevp = &header->ctl_table;
2066         for (n = 0; n < npath; ++n, ++path) {
2067                 /* Copy the procname */
2068                 new->procname = path->procname;
2069                 new->ctl_name = path->ctl_name;
2070                 new->mode     = 0555;
2071
2072                 *prevp = new;
2073                 prevp = &new->child;
2074
2075                 new += 2;
2076         }
2077         *prevp = table;
2078         header->ctl_table_arg = table;
2079
2080         INIT_LIST_HEAD(&header->ctl_entry);
2081         header->used = 0;
2082         header->unregistering = NULL;
2083         header->root = root;
2084         sysctl_set_parent(NULL, header->ctl_table);
2085         header->count = 1;
2086 #ifdef CONFIG_SYSCTL_SYSCALL_CHECK
2087         if (sysctl_check_table(namespaces, header->ctl_table)) {
2088                 kfree(header);
2089                 return NULL;
2090         }
2091 #endif
2092         spin_lock(&sysctl_lock);
2093         header->set = lookup_header_set(root, namespaces);
2094         header->attached_by = header->ctl_table;
2095         header->attached_to = root_table;
2096         header->parent = &root_table_header;
2097         for (set = header->set; set; set = set->parent) {
2098                 struct ctl_table_header *p;
2099                 list_for_each_entry(p, &set->list, ctl_entry) {
2100                         if (p->unregistering)
2101                                 continue;
2102                         try_attach(p, header);
2103                 }
2104         }
2105         header->parent->count++;
2106         list_add_tail(&header->ctl_entry, &header->set->list);
2107         spin_unlock(&sysctl_lock);
2108
2109         return header;
2110 }
2111
2112 /**
2113  * register_sysctl_table_path - register a sysctl table hierarchy
2114  * @path: The path to the directory the sysctl table is in.
2115  * @table: the top-level table structure
2116  *
2117  * Register a sysctl table hierarchy. @table should be a filled in ctl_table
2118  * array. A completely 0 filled entry terminates the table.
2119  *
2120  * See __register_sysctl_paths for more details.
2121  */
2122 struct ctl_table_header *register_sysctl_paths(const struct ctl_path *path,
2123                                                 struct ctl_table *table)
2124 {
2125         return __register_sysctl_paths(&sysctl_table_root, current->nsproxy,
2126                                         path, table);
2127 }
2128
2129 /**
2130  * register_sysctl_table - register a sysctl table hierarchy
2131  * @table: the top-level table structure
2132  *
2133  * Register a sysctl table hierarchy. @table should be a filled in ctl_table
2134  * array. A completely 0 filled entry terminates the table.
2135  *
2136  * See register_sysctl_paths for more details.
2137  */
2138 struct ctl_table_header *register_sysctl_table(struct ctl_table *table)
2139 {
2140         static const struct ctl_path null_path[] = { {} };
2141
2142         return register_sysctl_paths(null_path, table);
2143 }
2144
2145 /**
2146  * unregister_sysctl_table - unregister a sysctl table hierarchy
2147  * @header: the header returned from register_sysctl_table
2148  *
2149  * Unregisters the sysctl table and all children. proc entries may not
2150  * actually be removed until they are no longer used by anyone.
2151  */
2152 void unregister_sysctl_table(struct ctl_table_header * header)
2153 {
2154         might_sleep();
2155
2156         if (header == NULL)
2157                 return;
2158
2159         spin_lock(&sysctl_lock);
2160         start_unregistering(header);
2161         if (!--header->parent->count) {
2162                 WARN_ON(1);
2163                 kfree(header->parent);
2164         }
2165         if (!--header->count)
2166                 kfree(header);
2167         spin_unlock(&sysctl_lock);
2168 }
2169
2170 int sysctl_is_seen(struct ctl_table_header *p)
2171 {
2172         struct ctl_table_set *set = p->set;
2173         int res;
2174         spin_lock(&sysctl_lock);
2175         if (p->unregistering)
2176                 res = 0;
2177         else if (!set->is_seen)
2178                 res = 1;
2179         else
2180                 res = set->is_seen(set);
2181         spin_unlock(&sysctl_lock);
2182         return res;
2183 }
2184
2185 void setup_sysctl_set(struct ctl_table_set *p,
2186         struct ctl_table_set *parent,
2187         int (*is_seen)(struct ctl_table_set *))
2188 {
2189         INIT_LIST_HEAD(&p->list);
2190         p->parent = parent ? parent : &sysctl_table_root.default_set;
2191         p->is_seen = is_seen;
2192 }
2193
2194 #else /* !CONFIG_SYSCTL */
2195 struct ctl_table_header *register_sysctl_table(struct ctl_table * table)
2196 {
2197         return NULL;
2198 }
2199
2200 struct ctl_table_header *register_sysctl_paths(const struct ctl_path *path,
2201                                                     struct ctl_table *table)
2202 {
2203         return NULL;
2204 }
2205
2206 void unregister_sysctl_table(struct ctl_table_header * table)
2207 {
2208 }
2209
2210 void setup_sysctl_set(struct ctl_table_set *p,
2211         struct ctl_table_set *parent,
2212         int (*is_seen)(struct ctl_table_set *))
2213 {
2214 }
2215
2216 void sysctl_head_put(struct ctl_table_header *head)
2217 {
2218 }
2219
2220 #endif /* CONFIG_SYSCTL */
2221
2222 /*
2223  * /proc/sys support
2224  */
2225
2226 #ifdef CONFIG_PROC_SYSCTL
2227
2228 static int _proc_do_string(void* data, int maxlen, int write,
2229                            struct file *filp, void __user *buffer,
2230                            size_t *lenp, loff_t *ppos)
2231 {
2232         size_t len;
2233         char __user *p;
2234         char c;
2235
2236         if (!data || !maxlen || !*lenp) {
2237                 *lenp = 0;
2238                 return 0;
2239         }
2240
2241         if (write) {
2242                 len = 0;
2243                 p = buffer;
2244                 while (len < *lenp) {
2245                         if (get_user(c, p++))
2246                                 return -EFAULT;
2247                         if (c == 0 || c == '\n')
2248                                 break;
2249                         len++;
2250                 }
2251                 if (len >= maxlen)
2252                         len = maxlen-1;
2253                 if(copy_from_user(data, buffer, len))
2254                         return -EFAULT;
2255                 ((char *) data)[len] = 0;
2256                 *ppos += *lenp;
2257         } else {
2258                 len = strlen(data);
2259                 if (len > maxlen)
2260                         len = maxlen;
2261
2262                 if (*ppos > len) {
2263                         *lenp = 0;
2264                         return 0;
2265                 }
2266
2267                 data += *ppos;
2268                 len  -= *ppos;
2269
2270                 if (len > *lenp)
2271                         len = *lenp;
2272                 if (len)
2273                         if(copy_to_user(buffer, data, len))
2274                                 return -EFAULT;
2275                 if (len < *lenp) {
2276                         if(put_user('\n', ((char __user *) buffer) + len))
2277                                 return -EFAULT;
2278                         len++;
2279                 }
2280                 *lenp = len;
2281                 *ppos += len;
2282         }
2283         return 0;
2284 }
2285
2286 /**
2287  * proc_dostring - read a string sysctl
2288  * @table: the sysctl table
2289  * @write: %TRUE if this is a write to the sysctl file
2290  * @filp: the file structure
2291  * @buffer: the user buffer
2292  * @lenp: the size of the user buffer
2293  * @ppos: file position
2294  *
2295  * Reads/writes a string from/to the user buffer. If the kernel
2296  * buffer provided is not large enough to hold the string, the
2297  * string is truncated. The copied string is %NULL-terminated.
2298  * If the string is being read by the user process, it is copied
2299  * and a newline '\n' is added. It is truncated if the buffer is
2300  * not large enough.
2301  *
2302  * Returns 0 on success.
2303  */
2304 int proc_dostring(struct ctl_table *table, int write, struct file *filp,
2305                   void __user *buffer, size_t *lenp, loff_t *ppos)
2306 {
2307         return _proc_do_string(table->data, table->maxlen, write, filp,
2308                                buffer, lenp, ppos);
2309 }
2310
2311
2312 static int do_proc_dointvec_conv(int *negp, unsigned long *lvalp,
2313                                  int *valp,
2314                                  int write, void *data)
2315 {
2316         if (write) {
2317                 *valp = *negp ? -*lvalp : *lvalp;
2318         } else {
2319                 int val = *valp;
2320                 if (val < 0) {
2321                         *negp = -1;
2322                         *lvalp = (unsigned long)-val;
2323                 } else {
2324                         *negp = 0;
2325                         *lvalp = (unsigned long)val;
2326                 }
2327         }
2328         return 0;
2329 }
2330
2331 static int __do_proc_dointvec(void *tbl_data, struct ctl_table *table,
2332                   int write, struct file *filp, void __user *buffer,
2333                   size_t *lenp, loff_t *ppos,
2334                   int (*conv)(int *negp, unsigned long *lvalp, int *valp,
2335                               int write, void *data),
2336                   void *data)
2337 {
2338 #define TMPBUFLEN 21
2339         int *i, vleft, first = 1, neg;
2340         unsigned long lval;
2341         size_t left, len;
2342
2343         char buf[TMPBUFLEN], *p;
2344         char __user *s = buffer;
2345
2346         if (!tbl_data || !table->maxlen || !*lenp ||
2347             (*ppos && !write)) {
2348                 *lenp = 0;
2349                 return 0;
2350         }
2351
2352         i = (int *) tbl_data;
2353         vleft = table->maxlen / sizeof(*i);
2354         left = *lenp;
2355
2356         if (!conv)
2357                 conv = do_proc_dointvec_conv;
2358
2359         for (; left && vleft--; i++, first=0) {
2360                 if (write) {
2361                         while (left) {
2362                                 char c;
2363                                 if (get_user(c, s))
2364                                         return -EFAULT;
2365                                 if (!isspace(c))
2366                                         break;
2367                                 left--;
2368                                 s++;
2369                         }
2370                         if (!left)
2371                                 break;
2372                         neg = 0;
2373                         len = left;
2374                         if (len > sizeof(buf) - 1)
2375                                 len = sizeof(buf) - 1;
2376                         if (copy_from_user(buf, s, len))
2377                                 return -EFAULT;
2378                         buf[len] = 0;
2379                         p = buf;
2380                         if (*p == '-' && left > 1) {
2381                                 neg = 1;
2382                                 p++;
2383                         }
2384                         if (*p < '0' || *p > '9')
2385                                 break;
2386
2387                         lval = simple_strtoul(p, &p, 0);
2388
2389                         len = p-buf;
2390                         if ((len < left) && *p && !isspace(*p))
2391                                 break;
2392                         s += len;
2393                         left -= len;
2394
2395                         if (conv(&neg, &lval, i, 1, data))
2396                                 break;
2397                 } else {
2398                         p = buf;
2399                         if (!first)
2400                                 *p++ = '\t';
2401
2402                         if (conv(&neg, &lval, i, 0, data))
2403                                 break;
2404
2405                         sprintf(p, "%s%lu", neg ? "-" : "", lval);
2406                         len = strlen(buf);
2407                         if (len > left)
2408                                 len = left;
2409                         if(copy_to_user(s, buf, len))
2410                                 return -EFAULT;
2411                         left -= len;
2412                         s += len;
2413                 }
2414         }
2415
2416         if (!write && !first && left) {
2417                 if(put_user('\n', s))
2418                         return -EFAULT;
2419                 left--, s++;
2420         }
2421         if (write) {
2422                 while (left) {
2423                         char c;
2424                         if (get_user(c, s++))
2425                                 return -EFAULT;
2426                         if (!isspace(c))
2427                                 break;
2428                         left--;
2429                 }
2430         }
2431         if (write && first)
2432                 return -EINVAL;
2433         *lenp -= left;
2434         *ppos += *lenp;
2435         return 0;
2436 #undef TMPBUFLEN
2437 }
2438
2439 static int do_proc_dointvec(struct ctl_table *table, int write, struct file *filp,
2440                   void __user *buffer, size_t *lenp, loff_t *ppos,
2441                   int (*conv)(int *negp, unsigned long *lvalp, int *valp,
2442                               int write, void *data),
2443                   void *data)
2444 {
2445         return __do_proc_dointvec(table->data, table, write, filp,
2446                         buffer, lenp, ppos, conv, data);
2447 }
2448
2449 /**
2450  * proc_dointvec - read a vector of integers
2451  * @table: the sysctl table
2452  * @write: %TRUE if this is a write to the sysctl file
2453  * @filp: the file structure
2454  * @buffer: the user buffer
2455  * @lenp: the size of the user buffer
2456  * @ppos: file position
2457  *
2458  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2459  * values from/to the user buffer, treated as an ASCII string.
2460  *
2461  * Returns 0 on success.
2462  */
2463 int proc_dointvec(struct ctl_table *table, int write, struct file *filp,
2464                      void __user *buffer, size_t *lenp, loff_t *ppos)
2465 {
2466     return do_proc_dointvec(table,write,filp,buffer,lenp,ppos,
2467                             NULL,NULL);
2468 }
2469
2470 /*
2471  * Taint values can only be increased
2472  * This means we can safely use a temporary.
2473  */
2474 static int proc_taint(struct ctl_table *table, int write, struct file *filp,
2475                                void __user *buffer, size_t *lenp, loff_t *ppos)
2476 {
2477         struct ctl_table t;
2478         unsigned long tmptaint = get_taint();
2479         int err;
2480
2481         if (write && !capable(CAP_SYS_ADMIN))
2482                 return -EPERM;
2483
2484         t = *table;
2485         t.data = &tmptaint;
2486         err = proc_doulongvec_minmax(&t, write, filp, buffer, lenp, ppos);
2487         if (err < 0)
2488                 return err;
2489
2490         if (write) {
2491                 /*
2492                  * Poor man's atomic or. Not worth adding a primitive
2493                  * to everyone's atomic.h for this
2494                  */
2495                 int i;
2496                 for (i = 0; i < BITS_PER_LONG && tmptaint >> i; i++) {
2497                         if ((tmptaint >> i) & 1)
2498                                 add_taint(i);
2499                 }
2500         }
2501
2502         return err;
2503 }
2504
2505 struct do_proc_dointvec_minmax_conv_param {
2506         int *min;
2507         int *max;
2508 };
2509
2510 static int do_proc_dointvec_minmax_conv(int *negp, unsigned long *lvalp,
2511                                         int *valp,
2512                                         int write, void *data)
2513 {
2514         struct do_proc_dointvec_minmax_conv_param *param = data;
2515         if (write) {
2516                 int val = *negp ? -*lvalp : *lvalp;
2517                 if ((param->min && *param->min > val) ||
2518                     (param->max && *param->max < val))
2519                         return -EINVAL;
2520                 *valp = val;
2521         } else {
2522                 int val = *valp;
2523                 if (val < 0) {
2524                         *negp = -1;
2525                         *lvalp = (unsigned long)-val;
2526                 } else {
2527                         *negp = 0;
2528                         *lvalp = (unsigned long)val;
2529                 }
2530         }
2531         return 0;
2532 }
2533
2534 /**
2535  * proc_dointvec_minmax - read a vector of integers with min/max values
2536  * @table: the sysctl table
2537  * @write: %TRUE if this is a write to the sysctl file
2538  * @filp: the file structure
2539  * @buffer: the user buffer
2540  * @lenp: the size of the user buffer
2541  * @ppos: file position
2542  *
2543  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2544  * values from/to the user buffer, treated as an ASCII string.
2545  *
2546  * This routine will ensure the values are within the range specified by
2547  * table->extra1 (min) and table->extra2 (max).
2548  *
2549  * Returns 0 on success.
2550  */
2551 int proc_dointvec_minmax(struct ctl_table *table, int write, struct file *filp,
2552                   void __user *buffer, size_t *lenp, loff_t *ppos)
2553 {
2554         struct do_proc_dointvec_minmax_conv_param param = {
2555                 .min = (int *) table->extra1,
2556                 .max = (int *) table->extra2,
2557         };
2558         return do_proc_dointvec(table, write, filp, buffer, lenp, ppos,
2559                                 do_proc_dointvec_minmax_conv, &param);
2560 }
2561
2562 static int __do_proc_doulongvec_minmax(void *data, struct ctl_table *table, int write,
2563                                      struct file *filp,
2564                                      void __user *buffer,
2565                                      size_t *lenp, loff_t *ppos,
2566                                      unsigned long convmul,
2567                                      unsigned long convdiv)
2568 {
2569 #define TMPBUFLEN 21
2570         unsigned long *i, *min, *max, val;
2571         int vleft, first=1, neg;
2572         size_t len, left;
2573         char buf[TMPBUFLEN], *p;
2574         char __user *s = buffer;
2575
2576         if (!data || !table->maxlen || !*lenp ||
2577             (*ppos && !write)) {
2578                 *lenp = 0;
2579                 return 0;
2580         }
2581
2582         i = (unsigned long *) data;
2583         min = (unsigned long *) table->extra1;
2584         max = (unsigned long *) table->extra2;
2585         vleft = table->maxlen / sizeof(unsigned long);
2586         left = *lenp;
2587
2588         for (; left && vleft--; i++, min++, max++, first=0) {
2589                 if (write) {
2590                         while (left) {
2591                                 char c;
2592                                 if (get_user(c, s))
2593                                         return -EFAULT;
2594                                 if (!isspace(c))
2595                                         break;
2596                                 left--;
2597                                 s++;
2598                         }
2599                         if (!left)
2600                                 break;
2601                         neg = 0;
2602                         len = left;
2603                         if (len > TMPBUFLEN-1)
2604                                 len = TMPBUFLEN-1;
2605                         if (copy_from_user(buf, s, len))
2606                                 return -EFAULT;
2607                         buf[len] = 0;
2608                         p = buf;
2609                         if (*p == '-' && left > 1) {
2610                                 neg = 1;
2611                                 p++;
2612                         }
2613                         if (*p < '0' || *p > '9')
2614                                 break;
2615                         val = simple_strtoul(p, &p, 0) * convmul / convdiv ;
2616                         len = p-buf;
2617                         if ((len < left) && *p && !isspace(*p))
2618                                 break;
2619                         if (neg)
2620                                 val = -val;
2621                         s += len;
2622                         left -= len;
2623
2624                         if(neg)
2625                                 continue;
2626                         if ((min && val < *min) || (max && val > *max))
2627                                 continue;
2628                         *i = val;
2629                 } else {
2630                         p = buf;
2631                         if (!first)
2632                                 *p++ = '\t';
2633                         sprintf(p, "%lu", convdiv * (*i) / convmul);
2634                         len = strlen(buf);
2635                         if (len > left)
2636                                 len = left;
2637                         if(copy_to_user(s, buf, len))
2638                                 return -EFAULT;
2639                         left -= len;
2640                         s += len;
2641                 }
2642         }
2643
2644         if (!write && !first && left) {
2645                 if(put_user('\n', s))
2646                         return -EFAULT;
2647                 left--, s++;
2648         }
2649         if (write) {
2650                 while (left) {
2651                         char c;
2652                         if (get_user(c, s++))
2653                                 return -EFAULT;
2654                         if (!isspace(c))
2655                                 break;
2656                         left--;
2657                 }
2658         }
2659         if (write && first)
2660                 return -EINVAL;
2661         *lenp -= left;
2662         *ppos += *lenp;
2663         return 0;
2664 #undef TMPBUFLEN
2665 }
2666
2667 static int do_proc_doulongvec_minmax(struct ctl_table *table, int write,
2668                                      struct file *filp,
2669                                      void __user *buffer,
2670                                      size_t *lenp, loff_t *ppos,
2671                                      unsigned long convmul,
2672                                      unsigned long convdiv)
2673 {
2674         return __do_proc_doulongvec_minmax(table->data, table, write,
2675                         filp, buffer, lenp, ppos, convmul, convdiv);
2676 }
2677
2678 /**
2679  * proc_doulongvec_minmax - read a vector of long integers with min/max values
2680  * @table: the sysctl table
2681  * @write: %TRUE if this is a write to the sysctl file
2682  * @filp: the file structure
2683  * @buffer: the user buffer
2684  * @lenp: the size of the user buffer
2685  * @ppos: file position
2686  *
2687  * Reads/writes up to table->maxlen/sizeof(unsigned long) unsigned long
2688  * values from/to the user buffer, treated as an ASCII string.
2689  *
2690  * This routine will ensure the values are within the range specified by
2691  * table->extra1 (min) and table->extra2 (max).
2692  *
2693  * Returns 0 on success.
2694  */
2695 int proc_doulongvec_minmax(struct ctl_table *table, int write, struct file *filp,
2696                            void __user *buffer, size_t *lenp, loff_t *ppos)
2697 {
2698     return do_proc_doulongvec_minmax(table, write, filp, buffer, lenp, ppos, 1l, 1l);
2699 }
2700
2701 /**
2702  * proc_doulongvec_ms_jiffies_minmax - read a vector of millisecond values with min/max values
2703  * @table: the sysctl table
2704  * @write: %TRUE if this is a write to the sysctl file
2705  * @filp: the file structure
2706  * @buffer: the user buffer
2707  * @lenp: the size of the user buffer
2708  * @ppos: file position
2709  *
2710  * Reads/writes up to table->maxlen/sizeof(unsigned long) unsigned long
2711  * values from/to the user buffer, treated as an ASCII string. The values
2712  * are treated as milliseconds, and converted to jiffies when they are stored.
2713  *
2714  * This routine will ensure the values are within the range specified by
2715  * table->extra1 (min) and table->extra2 (max).
2716  *
2717  * Returns 0 on success.
2718  */
2719 int proc_doulongvec_ms_jiffies_minmax(struct ctl_table *table, int write,
2720                                       struct file *filp,
2721                                       void __user *buffer,
2722                                       size_t *lenp, loff_t *ppos)
2723 {
2724     return do_proc_doulongvec_minmax(table, write, filp, buffer,
2725                                      lenp, ppos, HZ, 1000l);
2726 }
2727
2728
2729 static int do_proc_dointvec_jiffies_conv(int *negp, unsigned long *lvalp,
2730                                          int *valp,
2731                                          int write, void *data)
2732 {
2733         if (write) {
2734                 if (*lvalp > LONG_MAX / HZ)
2735                         return 1;
2736                 *valp = *negp ? -(*lvalp*HZ) : (*lvalp*HZ);
2737         } else {
2738                 int val = *valp;
2739                 unsigned long lval;
2740                 if (val < 0) {
2741                         *negp = -1;
2742                         lval = (unsigned long)-val;
2743                 } else {
2744                         *negp = 0;
2745                         lval = (unsigned long)val;
2746                 }
2747                 *lvalp = lval / HZ;
2748         }
2749         return 0;
2750 }
2751
2752 static int do_proc_dointvec_userhz_jiffies_conv(int *negp, unsigned long *lvalp,
2753                                                 int *valp,
2754                                                 int write, void *data)
2755 {
2756         if (write) {
2757                 if (USER_HZ < HZ && *lvalp > (LONG_MAX / HZ) * USER_HZ)
2758                         return 1;
2759                 *valp = clock_t_to_jiffies(*negp ? -*lvalp : *lvalp);
2760         } else {
2761                 int val = *valp;
2762                 unsigned long lval;
2763                 if (val < 0) {
2764                         *negp = -1;
2765                         lval = (unsigned long)-val;
2766                 } else {
2767                         *negp = 0;
2768                         lval = (unsigned long)val;
2769                 }
2770                 *lvalp = jiffies_to_clock_t(lval);
2771         }
2772         return 0;
2773 }
2774
2775 static int do_proc_dointvec_ms_jiffies_conv(int *negp, unsigned long *lvalp,
2776                                             int *valp,
2777                                             int write, void *data)
2778 {
2779         if (write) {
2780                 *valp = msecs_to_jiffies(*negp ? -*lvalp : *lvalp);
2781         } else {
2782                 int val = *valp;
2783                 unsigned long lval;
2784                 if (val < 0) {
2785                         *negp = -1;
2786                         lval = (unsigned long)-val;
2787                 } else {
2788                         *negp = 0;
2789                         lval = (unsigned long)val;
2790                 }
2791                 *lvalp = jiffies_to_msecs(lval);
2792         }
2793         return 0;
2794 }
2795
2796 /**
2797  * proc_dointvec_jiffies - read a vector of integers as seconds
2798  * @table: the sysctl table
2799  * @write: %TRUE if this is a write to the sysctl file
2800  * @filp: the file structure
2801  * @buffer: the user buffer
2802  * @lenp: the size of the user buffer
2803  * @ppos: file position
2804  *
2805  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2806  * values from/to the user buffer, treated as an ASCII string.
2807  * The values read are assumed to be in seconds, and are converted into
2808  * jiffies.
2809  *
2810  * Returns 0 on success.
2811  */
2812 int proc_dointvec_jiffies(struct ctl_table *table, int write, struct file *filp,
2813                           void __user *buffer, size_t *lenp, loff_t *ppos)
2814 {
2815     return do_proc_dointvec(table,write,filp,buffer,lenp,ppos,
2816                             do_proc_dointvec_jiffies_conv,NULL);
2817 }
2818
2819 /**
2820  * proc_dointvec_userhz_jiffies - read a vector of integers as 1/USER_HZ seconds
2821  * @table: the sysctl table
2822  * @write: %TRUE if this is a write to the sysctl file
2823  * @filp: the file structure
2824  * @buffer: the user buffer
2825  * @lenp: the size of the user buffer
2826  * @ppos: pointer to the file position
2827  *
2828  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2829  * values from/to the user buffer, treated as an ASCII string.
2830  * The values read are assumed to be in 1/USER_HZ seconds, and
2831  * are converted into jiffies.
2832  *
2833  * Returns 0 on success.
2834  */
2835 int proc_dointvec_userhz_jiffies(struct ctl_table *table, int write, struct file *filp,
2836                                  void __user *buffer, size_t *lenp, loff_t *ppos)
2837 {
2838     return do_proc_dointvec(table,write,filp,buffer,lenp,ppos,
2839                             do_proc_dointvec_userhz_jiffies_conv,NULL);
2840 }
2841
2842 /**
2843  * proc_dointvec_ms_jiffies - read a vector of integers as 1 milliseconds
2844  * @table: the sysctl table
2845  * @write: %TRUE if this is a write to the sysctl file
2846  * @filp: the file structure
2847  * @buffer: the user buffer
2848  * @lenp: the size of the user buffer
2849  * @ppos: file position
2850  * @ppos: the current position in the file
2851  *
2852  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2853  * values from/to the user buffer, treated as an ASCII string.
2854  * The values read are assumed to be in 1/1000 seconds, and
2855  * are converted into jiffies.
2856  *
2857  * Returns 0 on success.
2858  */
2859 int proc_dointvec_ms_jiffies(struct ctl_table *table, int write, struct file *filp,
2860                              void __user *buffer, size_t *lenp, loff_t *ppos)
2861 {
2862         return do_proc_dointvec(table, write, filp, buffer, lenp, ppos,
2863                                 do_proc_dointvec_ms_jiffies_conv, NULL);
2864 }
2865
2866 static int proc_do_cad_pid(struct ctl_table *table, int write, struct file *filp,
2867                            void __user *buffer, size_t *lenp, loff_t *ppos)
2868 {
2869         struct pid *new_pid;
2870         pid_t tmp;
2871         int r;
2872
2873         tmp = pid_vnr(cad_pid);
2874
2875         r = __do_proc_dointvec(&tmp, table, write, filp, buffer,
2876                                lenp, ppos, NULL, NULL);
2877         if (r || !write)
2878                 return r;
2879
2880         new_pid = find_get_pid(tmp);
2881         if (!new_pid)
2882                 return -ESRCH;
2883
2884         put_pid(xchg(&cad_pid, new_pid));
2885         return 0;
2886 }
2887
2888 #else /* CONFIG_PROC_FS */
2889
2890 int proc_dostring(struct ctl_table *table, int write, struct file *filp,
2891                   void __user *buffer, size_t *lenp, loff_t *ppos)
2892 {
2893         return -ENOSYS;
2894 }
2895
2896 int proc_dointvec(struct ctl_table *table, int write, struct file *filp,
2897                   void __user *buffer, size_t *lenp, loff_t *ppos)
2898 {
2899         return -ENOSYS;
2900 }
2901
2902 int proc_dointvec_minmax(struct ctl_table *table, int write, struct file *filp,
2903                     void __user *buffer, size_t *lenp, loff_t *ppos)
2904 {
2905         return -ENOSYS;
2906 }
2907
2908 int proc_dointvec_jiffies(struct ctl_table *table, int write, struct file *filp,
2909                     void __user *buffer, size_t *lenp, loff_t *ppos)
2910 {
2911         return -ENOSYS;
2912 }
2913
2914 int proc_dointvec_userhz_jiffies(struct ctl_table *table, int write, struct file *filp,
2915                     void __user *buffer, size_t *lenp, loff_t *ppos)
2916 {
2917         return -ENOSYS;
2918 }
2919
2920 int proc_dointvec_ms_jiffies(struct ctl_table *table, int write, struct file *filp,
2921                              void __user *buffer, size_t *lenp, loff_t *ppos)
2922 {
2923         return -ENOSYS;
2924 }
2925
2926 int proc_doulongvec_minmax(struct ctl_table *table, int write, struct file *filp,
2927                     void __user *buffer, size_t *lenp, loff_t *ppos)
2928 {
2929         return -ENOSYS;
2930 }
2931
2932 int proc_doulongvec_ms_jiffies_minmax(struct ctl_table *table, int write,
2933                                       struct file *filp,
2934                                       void __user *buffer,
2935                                       size_t *lenp, loff_t *ppos)
2936 {
2937     return -ENOSYS;
2938 }
2939
2940
2941 #endif /* CONFIG_PROC_FS */
2942
2943
2944 #ifdef CONFIG_SYSCTL_SYSCALL
2945 /*
2946  * General sysctl support routines
2947  */
2948
2949 /* The generic sysctl data routine (used if no strategy routine supplied) */
2950 int sysctl_data(struct ctl_table *table,
2951                 void __user *oldval, size_t __user *oldlenp,
2952                 void __user *newval, size_t newlen)
2953 {
2954         size_t len;
2955
2956         /* Get out of I don't have a variable */
2957         if (!table->data || !table->maxlen)
2958                 return -ENOTDIR;
2959
2960         if (oldval && oldlenp) {
2961                 if (get_user(len, oldlenp))
2962                         return -EFAULT;
2963                 if (len) {
2964                         if (len > table->maxlen)
2965                                 len = table->maxlen;
2966                         if (copy_to_user(oldval, table->data, len))
2967                                 return -EFAULT;
2968                         if (put_user(len, oldlenp))
2969                                 return -EFAULT;
2970                 }
2971         }
2972
2973         if (newval && newlen) {
2974                 if (newlen > table->maxlen)
2975                         newlen = table->maxlen;
2976
2977                 if (copy_from_user(table->data, newval, newlen))
2978                         return -EFAULT;
2979         }
2980         return 1;
2981 }
2982
2983 /* The generic string strategy routine: */
2984 int sysctl_string(struct ctl_table *table,
2985                   void __user *oldval, size_t __user *oldlenp,
2986                   void __user *newval, size_t newlen)
2987 {
2988         if (!table->data || !table->maxlen)
2989                 return -ENOTDIR;
2990
2991         if (oldval && oldlenp) {
2992                 size_t bufsize;
2993                 if (get_user(bufsize, oldlenp))
2994                         return -EFAULT;
2995                 if (bufsize) {
2996                         size_t len = strlen(table->data), copied;
2997
2998                         /* This shouldn't trigger for a well-formed sysctl */
2999                         if (len > table->maxlen)
3000                                 len = table->maxlen;
3001
3002                         /* Copy up to a max of bufsize-1 bytes of the string */
3003                         copied = (len >= bufsize) ? bufsize - 1 : len;
3004
3005                         if (copy_to_user(oldval, table->data, copied) ||
3006                             put_user(0, (char __user *)(oldval + copied)))
3007                                 return -EFAULT;
3008                         if (put_user(len, oldlenp))
3009                                 return -EFAULT;
3010                 }
3011         }
3012         if (newval && newlen) {
3013                 size_t len = newlen;
3014                 if (len > table->maxlen)
3015                         len = table->maxlen;
3016                 if(copy_from_user(table->data, newval, len))
3017                         return -EFAULT;
3018                 if (len == table->maxlen)
3019                         len--;
3020                 ((char *) table->data)[len] = 0;
3021         }
3022         return 1;
3023 }
3024
3025 /*
3026  * This function makes sure that all of the integers in the vector
3027  * are between the minimum and maximum values given in the arrays
3028  * table->extra1 and table->extra2, respectively.
3029  */
3030 int sysctl_intvec(struct ctl_table *table,
3031                 void __user *oldval, size_t __user *oldlenp,
3032                 void __user *newval, size_t newlen)
3033 {
3034
3035         if (newval && newlen) {
3036                 int __user *vec = (int __user *) newval;
3037                 int *min = (int *) table->extra1;
3038                 int *max = (int *) table->extra2;
3039                 size_t length;
3040                 int i;
3041
3042                 if (newlen % sizeof(int) != 0)
3043                         return -EINVAL;
3044
3045                 if (!table->extra1 && !table->extra2)
3046                         return 0;
3047
3048                 if (newlen > table->maxlen)
3049                         newlen = table->maxlen;
3050                 length = newlen / sizeof(int);
3051
3052                 for (i = 0; i < length; i++) {
3053                         int value;
3054                         if (get_user(value, vec + i))
3055                                 return -EFAULT;
3056                         if (min && value < min[i])
3057                                 return -EINVAL;
3058                         if (max && value > max[i])
3059                                 return -EINVAL;
3060                 }
3061         }
3062         return 0;
3063 }
3064
3065 /* Strategy function to convert jiffies to seconds */
3066 int sysctl_jiffies(struct ctl_table *table,
3067                 void __user *oldval, size_t __user *oldlenp,
3068                 void __user *newval, size_t newlen)
3069 {
3070         if (oldval && oldlenp) {
3071                 size_t olen;
3072
3073                 if (get_user(olen, oldlenp))
3074                         return -EFAULT;
3075                 if (olen) {
3076                         int val;
3077
3078                         if (olen < sizeof(int))
3079                                 return -EINVAL;
3080
3081                         val = *(int *)(table->data) / HZ;
3082                         if (put_user(val, (int __user *)oldval))
3083                                 return -EFAULT;
3084                         if (put_user(sizeof(int), oldlenp))
3085                                 return -EFAULT;
3086                 }
3087         }
3088         if (newval && newlen) {
3089                 int new;
3090                 if (newlen != sizeof(int))
3091                         return -EINVAL;
3092                 if (get_user(new, (int __user *)newval))
3093                         return -EFAULT;
3094                 *(int *)(table->data) = new*HZ;
3095         }
3096         return 1;
3097 }
3098
3099 /* Strategy function to convert jiffies to seconds */
3100 int sysctl_ms_jiffies(struct ctl_table *table,
3101                 void __user *oldval, size_t __user *oldlenp,
3102                 void __user *newval, size_t newlen)
3103 {
3104         if (oldval && oldlenp) {
3105                 size_t olen;
3106
3107                 if (get_user(olen, oldlenp))
3108                         return -EFAULT;
3109                 if (olen) {
3110                         int val;
3111
3112                         if (olen < sizeof(int))
3113                                 return -EINVAL;
3114
3115                         val = jiffies_to_msecs(*(int *)(table->data));
3116                         if (put_user(val, (int __user *)oldval))
3117                                 return -EFAULT;
3118                         if (put_user(sizeof(int), oldlenp))
3119                                 return -EFAULT;
3120                 }
3121         }
3122         if (newval && newlen) {
3123                 int new;
3124                 if (newlen != sizeof(int))
3125                         return -EINVAL;
3126                 if (get_user(new, (int __user *)newval))
3127                         return -EFAULT;
3128                 *(int *)(table->data) = msecs_to_jiffies(new);
3129         }
3130         return 1;
3131 }
3132
3133
3134
3135 #else /* CONFIG_SYSCTL_SYSCALL */
3136
3137
3138 SYSCALL_DEFINE1(sysctl, struct __sysctl_args __user *, args)
3139 {
3140         struct __sysctl_args tmp;
3141         int error;
3142
3143         if (copy_from_user(&tmp, args, sizeof(tmp)))
3144                 return -EFAULT;
3145
3146         error = deprecated_sysctl_warning(&tmp);
3147
3148         /* If no error reading the parameters then just -ENOSYS ... */
3149         if (!error)
3150                 error = -ENOSYS;
3151
3152         return error;
3153 }
3154
3155 int sysctl_data(struct ctl_table *table,
3156                   void __user *oldval, size_t __user *oldlenp,
3157                   void __user *newval, size_t newlen)
3158 {
3159         return -ENOSYS;
3160 }
3161
3162 int sysctl_string(struct ctl_table *table,
3163                   void __user *oldval, size_t __user *oldlenp,
3164                   void __user *newval, size_t newlen)
3165 {
3166         return -ENOSYS;
3167 }
3168
3169 int sysctl_intvec(struct ctl_table *table,
3170                 void __user *oldval, size_t __user *oldlenp,
3171                 void __user *newval, size_t newlen)
3172 {
3173         return -ENOSYS;
3174 }
3175
3176 int sysctl_jiffies(struct ctl_table *table,
3177                 void __user *oldval, size_t __user *oldlenp,
3178                 void __user *newval, size_t newlen)
3179 {
3180         return -ENOSYS;
3181 }
3182
3183 int sysctl_ms_jiffies(struct ctl_table *table,
3184                 void __user *oldval, size_t __user *oldlenp,
3185                 void __user *newval, size_t newlen)
3186 {
3187         return -ENOSYS;
3188 }
3189
3190 #endif /* CONFIG_SYSCTL_SYSCALL */
3191
3192 static int deprecated_sysctl_warning(struct __sysctl_args *args)
3193 {
3194         static int msg_count;
3195         int name[CTL_MAXNAME];
3196         int i;
3197
3198         /* Check args->nlen. */
3199         if (args->nlen < 0 || args->nlen > CTL_MAXNAME)
3200                 return -ENOTDIR;
3201
3202         /* Read in the sysctl name for better debug message logging */
3203         for (i = 0; i < args->nlen; i++)
3204                 if (get_user(name[i], args->name + i))
3205                         return -EFAULT;
3206
3207         /* Ignore accesses to kernel.version */
3208         if ((args->nlen == 2) && (name[0] == CTL_KERN) && (name[1] == KERN_VERSION))
3209                 return 0;
3210
3211         if (msg_count < 5) {
3212                 msg_count++;
3213                 printk(KERN_INFO
3214                         "warning: process `%s' used the deprecated sysctl "
3215                         "system call with ", current->comm);
3216                 for (i = 0; i < args->nlen; i++)
3217                         printk("%d.", name[i]);
3218                 printk("\n");
3219         }
3220         return 0;
3221 }
3222
3223 /*
3224  * No sense putting this after each symbol definition, twice,
3225  * exception granted :-)
3226  */
3227 EXPORT_SYMBOL(proc_dointvec);
3228 EXPORT_SYMBOL(proc_dointvec_jiffies);
3229 EXPORT_SYMBOL(proc_dointvec_minmax);
3230 EXPORT_SYMBOL(proc_dointvec_userhz_jiffies);
3231 EXPORT_SYMBOL(proc_dointvec_ms_jiffies);
3232 EXPORT_SYMBOL(proc_dostring);
3233 EXPORT_SYMBOL(proc_doulongvec_minmax);
3234 EXPORT_SYMBOL(proc_doulongvec_ms_jiffies_minmax);
3235 EXPORT_SYMBOL(register_sysctl_table);
3236 EXPORT_SYMBOL(register_sysctl_paths);
3237 EXPORT_SYMBOL(sysctl_intvec);
3238 EXPORT_SYMBOL(sysctl_jiffies);
3239 EXPORT_SYMBOL(sysctl_ms_jiffies);
3240 EXPORT_SYMBOL(sysctl_string);
3241 EXPORT_SYMBOL(sysctl_data);
3242 EXPORT_SYMBOL(unregister_sysctl_table);