git.oblomov.eu Git - linux-2.6/blob - kernel/sysctl.c

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