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

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