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

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