kernel/sysctl.c

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