Pileus Git - ~andy/linux/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 #endif
 924 /*
 925  * NOTE: do not add new entries to this table unless you have read
 926  * Documentation/sysctl/ctl_unnumbered.txt
 927  */
 928         { .ctl_name = 0 }
 929 };
 930
 931 static struct ctl_table vm_table[] = {
 932         {
 933                 .ctl_name       = VM_OVERCOMMIT_MEMORY,
 934                 .procname       = "overcommit_memory",
 935                 .data           = &sysctl_overcommit_memory,
 936                 .maxlen         = sizeof(sysctl_overcommit_memory),
 937                 .mode           = 0644,
 938                 .proc_handler   = &proc_dointvec,
 939         },
 940         {
 941                 .ctl_name       = VM_PANIC_ON_OOM,
 942                 .procname       = "panic_on_oom",
 943                 .data           = &sysctl_panic_on_oom,
 944                 .maxlen         = sizeof(sysctl_panic_on_oom),
 945                 .mode           = 0644,
 946                 .proc_handler   = &proc_dointvec,
 947         },
 948         {
 949                 .ctl_name       = CTL_UNNUMBERED,
 950                 .procname       = "oom_kill_allocating_task",
 951                 .data           = &sysctl_oom_kill_allocating_task,
 952                 .maxlen         = sizeof(sysctl_oom_kill_allocating_task),
 953                 .mode           = 0644,
 954                 .proc_handler   = &proc_dointvec,
 955         },
 956         {
 957                 .ctl_name       = CTL_UNNUMBERED,
 958                 .procname       = "oom_dump_tasks",
 959                 .data           = &sysctl_oom_dump_tasks,
 960                 .maxlen         = sizeof(sysctl_oom_dump_tasks),
 961                 .mode           = 0644,
 962                 .proc_handler   = &proc_dointvec,
 963         },
 964         {
 965                 .ctl_name       = VM_OVERCOMMIT_RATIO,
 966                 .procname       = "overcommit_ratio",
 967                 .data           = &sysctl_overcommit_ratio,
 968                 .maxlen         = sizeof(sysctl_overcommit_ratio),
 969                 .mode           = 0644,
 970                 .proc_handler   = &proc_dointvec,
 971         },
 972         {
 973                 .ctl_name       = VM_PAGE_CLUSTER,
 974                 .procname       = "page-cluster",
 975                 .data           = &page_cluster,
 976                 .maxlen         = sizeof(int),
 977                 .mode           = 0644,
 978                 .proc_handler   = &proc_dointvec,
 979         },
 980         {
 981                 .ctl_name       = VM_DIRTY_BACKGROUND,
 982                 .procname       = "dirty_background_ratio",
 983                 .data           = &dirty_background_ratio,
 984                 .maxlen         = sizeof(dirty_background_ratio),
 985                 .mode           = 0644,
 986                 .proc_handler   = &dirty_background_ratio_handler,
 987                 .strategy       = &sysctl_intvec,
 988                 .extra1         = &zero,
 989                 .extra2         = &one_hundred,
 990         },
 991         {
 992                 .ctl_name       = CTL_UNNUMBERED,
 993                 .procname       = "dirty_background_bytes",
 994                 .data           = &dirty_background_bytes,
 995                 .maxlen         = sizeof(dirty_background_bytes),
 996                 .mode           = 0644,
 997                 .proc_handler   = &dirty_background_bytes_handler,
 998                 .strategy       = &sysctl_intvec,
 999                 .extra1         = &one_ul,
1000         },
1001         {
1002                 .ctl_name       = VM_DIRTY_RATIO,
1003                 .procname       = "dirty_ratio",
1004                 .data           = &vm_dirty_ratio,
1005                 .maxlen         = sizeof(vm_dirty_ratio),
1006                 .mode           = 0644,
1007                 .proc_handler   = &dirty_ratio_handler,
1008                 .strategy       = &sysctl_intvec,
1009                 .extra1         = &zero,
1010                 .extra2         = &one_hundred,
1011         },
1012         {
1013                 .ctl_name       = CTL_UNNUMBERED,
1014                 .procname       = "dirty_bytes",
1015                 .data           = &vm_dirty_bytes,
1016                 .maxlen         = sizeof(vm_dirty_bytes),
1017                 .mode           = 0644,
1018                 .proc_handler   = &dirty_bytes_handler,
1019                 .strategy       = &sysctl_intvec,
1020                 .extra1         = &one_ul,
1021         },
1022         {
1023                 .procname       = "dirty_writeback_centisecs",
1024                 .data           = &dirty_writeback_interval,
1025                 .maxlen         = sizeof(dirty_writeback_interval),
1026                 .mode           = 0644,
1027                 .proc_handler   = &dirty_writeback_centisecs_handler,
1028         },
1029         {
1030                 .procname       = "dirty_expire_centisecs",
1031                 .data           = &dirty_expire_interval,
1032                 .maxlen         = sizeof(dirty_expire_interval),
1033                 .mode           = 0644,
1034                 .proc_handler   = &proc_dointvec,
1035         },
1036         {
1037                 .ctl_name       = VM_NR_PDFLUSH_THREADS,
1038                 .procname       = "nr_pdflush_threads",
1039                 .data           = &nr_pdflush_threads,
1040                 .maxlen         = sizeof nr_pdflush_threads,
1041                 .mode           = 0444 /* read-only*/,
1042                 .proc_handler   = &proc_dointvec,
1043         },
1044         {
1045                 .ctl_name       = CTL_UNNUMBERED,
1046                 .procname       = "nr_pdflush_threads_min",
1047                 .data           = &nr_pdflush_threads_min,
1048                 .maxlen         = sizeof nr_pdflush_threads_min,
1049                 .mode           = 0644 /* read-write */,
1050                 .proc_handler   = &proc_dointvec_minmax,
1051                 .strategy       = &sysctl_intvec,
1052                 .extra1         = &one,
1053                 .extra2         = &nr_pdflush_threads_max,
1054         },
1055         {
1056                 .ctl_name       = CTL_UNNUMBERED,
1057                 .procname       = "nr_pdflush_threads_max",
1058                 .data           = &nr_pdflush_threads_max,
1059                 .maxlen         = sizeof nr_pdflush_threads_max,
1060                 .mode           = 0644 /* read-write */,
1061                 .proc_handler   = &proc_dointvec_minmax,
1062                 .strategy       = &sysctl_intvec,
1063                 .extra1         = &nr_pdflush_threads_min,
1064                 .extra2         = &one_thousand,
1065         },
1066         {
1067                 .ctl_name       = VM_SWAPPINESS,
1068                 .procname       = "swappiness",
1069                 .data           = &vm_swappiness,
1070                 .maxlen         = sizeof(vm_swappiness),
1071                 .mode           = 0644,
1072                 .proc_handler   = &proc_dointvec_minmax,
1073                 .strategy       = &sysctl_intvec,
1074                 .extra1         = &zero,
1075                 .extra2         = &one_hundred,
1076         },
1077 #ifdef CONFIG_HUGETLB_PAGE
1078          {
1079                 .procname       = "nr_hugepages",
1080                 .data           = NULL,
1081                 .maxlen         = sizeof(unsigned long),
1082                 .mode           = 0644,
1083                 .proc_handler   = &hugetlb_sysctl_handler,
1084                 .extra1         = (void *)&hugetlb_zero,
1085                 .extra2         = (void *)&hugetlb_infinity,
1086          },
1087          {
1088                 .ctl_name       = VM_HUGETLB_GROUP,
1089                 .procname       = "hugetlb_shm_group",
1090                 .data           = &sysctl_hugetlb_shm_group,
1091                 .maxlen         = sizeof(gid_t),
1092                 .mode           = 0644,
1093                 .proc_handler   = &proc_dointvec,
1094          },
1095          {
1096                 .ctl_name       = CTL_UNNUMBERED,
1097                 .procname       = "hugepages_treat_as_movable",
1098                 .data           = &hugepages_treat_as_movable,
1099                 .maxlen         = sizeof(int),
1100                 .mode           = 0644,
1101                 .proc_handler   = &hugetlb_treat_movable_handler,
1102         },
1103         {
1104                 .ctl_name       = CTL_UNNUMBERED,
1105                 .procname       = "nr_overcommit_hugepages",
1106                 .data           = NULL,
1107                 .maxlen         = sizeof(unsigned long),
1108                 .mode           = 0644,
1109                 .proc_handler   = &hugetlb_overcommit_handler,
1110                 .extra1         = (void *)&hugetlb_zero,
1111                 .extra2         = (void *)&hugetlb_infinity,
1112         },
1113 #endif
1114         {
1115                 .ctl_name       = VM_LOWMEM_RESERVE_RATIO,
1116                 .procname       = "lowmem_reserve_ratio",
1117                 .data           = &sysctl_lowmem_reserve_ratio,
1118                 .maxlen         = sizeof(sysctl_lowmem_reserve_ratio),
1119                 .mode           = 0644,
1120                 .proc_handler   = &lowmem_reserve_ratio_sysctl_handler,
1121                 .strategy       = &sysctl_intvec,
1122         },
1123         {
1124                 .ctl_name       = VM_DROP_PAGECACHE,
1125                 .procname       = "drop_caches",
1126                 .data           = &sysctl_drop_caches,
1127                 .maxlen         = sizeof(int),
1128                 .mode           = 0644,
1129                 .proc_handler   = drop_caches_sysctl_handler,
1130                 .strategy       = &sysctl_intvec,
1131         },
1132         {
1133                 .ctl_name       = VM_MIN_FREE_KBYTES,
1134                 .procname       = "min_free_kbytes",
1135                 .data           = &min_free_kbytes,
1136                 .maxlen         = sizeof(min_free_kbytes),
1137                 .mode           = 0644,
1138                 .proc_handler   = &min_free_kbytes_sysctl_handler,
1139                 .strategy       = &sysctl_intvec,
1140                 .extra1         = &zero,
1141         },
1142         {
1143                 .ctl_name       = VM_PERCPU_PAGELIST_FRACTION,
1144                 .procname       = "percpu_pagelist_fraction",
1145                 .data           = &percpu_pagelist_fraction,
1146                 .maxlen         = sizeof(percpu_pagelist_fraction),
1147                 .mode           = 0644,
1148                 .proc_handler   = &percpu_pagelist_fraction_sysctl_handler,
1149                 .strategy       = &sysctl_intvec,
1150                 .extra1         = &min_percpu_pagelist_fract,
1151         },
1152 #ifdef CONFIG_MMU
1153         {
1154                 .ctl_name       = VM_MAX_MAP_COUNT,
1155                 .procname       = "max_map_count",
1156                 .data           = &sysctl_max_map_count,
1157                 .maxlen         = sizeof(sysctl_max_map_count),
1158                 .mode           = 0644,
1159                 .proc_handler   = &proc_dointvec
1160         },
1161 #else
1162         {
1163                 .ctl_name       = CTL_UNNUMBERED,
1164                 .procname       = "nr_trim_pages",
1165                 .data           = &sysctl_nr_trim_pages,
1166                 .maxlen         = sizeof(sysctl_nr_trim_pages),
1167                 .mode           = 0644,
1168                 .proc_handler   = &proc_dointvec_minmax,
1169                 .strategy       = &sysctl_intvec,
1170                 .extra1         = &zero,
1171         },
1172 #endif
1173         {
1174                 .ctl_name       = VM_LAPTOP_MODE,
1175                 .procname       = "laptop_mode",
1176                 .data           = &laptop_mode,
1177                 .maxlen         = sizeof(laptop_mode),
1178                 .mode           = 0644,
1179                 .proc_handler   = &proc_dointvec_jiffies,
1180                 .strategy       = &sysctl_jiffies,
1181         },
1182         {
1183                 .ctl_name       = VM_BLOCK_DUMP,
1184                 .procname       = "block_dump",
1185                 .data           = &block_dump,
1186                 .maxlen         = sizeof(block_dump),
1187                 .mode           = 0644,
1188                 .proc_handler   = &proc_dointvec,
1189                 .strategy       = &sysctl_intvec,
1190                 .extra1         = &zero,
1191         },
1192         {
1193                 .ctl_name       = VM_VFS_CACHE_PRESSURE,
1194                 .procname       = "vfs_cache_pressure",
1195                 .data           = &sysctl_vfs_cache_pressure,
1196                 .maxlen         = sizeof(sysctl_vfs_cache_pressure),
1197                 .mode           = 0644,
1198                 .proc_handler   = &proc_dointvec,
1199                 .strategy       = &sysctl_intvec,
1200                 .extra1         = &zero,
1201         },
1202 #ifdef HAVE_ARCH_PICK_MMAP_LAYOUT
1203         {
1204                 .ctl_name       = VM_LEGACY_VA_LAYOUT,
1205                 .procname       = "legacy_va_layout",
1206                 .data           = &sysctl_legacy_va_layout,
1207                 .maxlen         = sizeof(sysctl_legacy_va_layout),
1208                 .mode           = 0644,
1209                 .proc_handler   = &proc_dointvec,
1210                 .strategy       = &sysctl_intvec,
1211                 .extra1         = &zero,
1212         },
1213 #endif
1214 #ifdef CONFIG_NUMA
1215         {
1216                 .ctl_name       = VM_ZONE_RECLAIM_MODE,
1217                 .procname       = "zone_reclaim_mode",
1218                 .data           = &zone_reclaim_mode,
1219                 .maxlen         = sizeof(zone_reclaim_mode),
1220                 .mode           = 0644,
1221                 .proc_handler   = &proc_dointvec,
1222                 .strategy       = &sysctl_intvec,
1223                 .extra1         = &zero,
1224         },
1225         {
1226                 .ctl_name       = VM_MIN_UNMAPPED,
1227                 .procname       = "min_unmapped_ratio",
1228                 .data           = &sysctl_min_unmapped_ratio,
1229                 .maxlen         = sizeof(sysctl_min_unmapped_ratio),
1230                 .mode           = 0644,
1231                 .proc_handler   = &sysctl_min_unmapped_ratio_sysctl_handler,
1232                 .strategy       = &sysctl_intvec,
1233                 .extra1         = &zero,
1234                 .extra2         = &one_hundred,
1235         },
1236         {
1237                 .ctl_name       = VM_MIN_SLAB,
1238                 .procname       = "min_slab_ratio",
1239                 .data           = &sysctl_min_slab_ratio,
1240                 .maxlen         = sizeof(sysctl_min_slab_ratio),
1241                 .mode           = 0644,
1242                 .proc_handler   = &sysctl_min_slab_ratio_sysctl_handler,
1243                 .strategy       = &sysctl_intvec,
1244                 .extra1         = &zero,
1245                 .extra2         = &one_hundred,
1246         },
1247 #endif
1248 #ifdef CONFIG_SMP
1249         {
1250                 .ctl_name       = CTL_UNNUMBERED,
1251                 .procname       = "stat_interval",
1252                 .data           = &sysctl_stat_interval,
1253                 .maxlen         = sizeof(sysctl_stat_interval),
1254                 .mode           = 0644,
1255                 .proc_handler   = &proc_dointvec_jiffies,
1256                 .strategy       = &sysctl_jiffies,
1257         },
1258 #endif
1259 #ifdef CONFIG_SECURITY
1260         {
1261                 .ctl_name       = CTL_UNNUMBERED,
1262                 .procname       = "mmap_min_addr",
1263                 .data           = &mmap_min_addr,
1264                 .maxlen         = sizeof(unsigned long),
1265                 .mode           = 0644,
1266                 .proc_handler   = &proc_doulongvec_minmax,
1267         },
1268 #endif
1269 #ifdef CONFIG_NUMA
1270         {
1271                 .ctl_name       = CTL_UNNUMBERED,
1272                 .procname       = "numa_zonelist_order",
1273                 .data           = &numa_zonelist_order,
1274                 .maxlen         = NUMA_ZONELIST_ORDER_LEN,
1275                 .mode           = 0644,
1276                 .proc_handler   = &numa_zonelist_order_handler,
1277                 .strategy       = &sysctl_string,
1278         },
1279 #endif
1280 #if (defined(CONFIG_X86_32) && !defined(CONFIG_UML))|| \
1281    (defined(CONFIG_SUPERH) && defined(CONFIG_VSYSCALL))
1282         {
1283                 .ctl_name       = VM_VDSO_ENABLED,
1284                 .procname       = "vdso_enabled",
1285                 .data           = &vdso_enabled,
1286                 .maxlen         = sizeof(vdso_enabled),
1287                 .mode           = 0644,
1288                 .proc_handler   = &proc_dointvec,
1289                 .strategy       = &sysctl_intvec,
1290                 .extra1         = &zero,
1291         },
1292 #endif
1293 #ifdef CONFIG_HIGHMEM
1294         {
1295                 .ctl_name       = CTL_UNNUMBERED,
1296                 .procname       = "highmem_is_dirtyable",
1297                 .data           = &vm_highmem_is_dirtyable,
1298                 .maxlen         = sizeof(vm_highmem_is_dirtyable),
1299                 .mode           = 0644,
1300                 .proc_handler   = &proc_dointvec_minmax,
1301                 .strategy       = &sysctl_intvec,
1302                 .extra1         = &zero,
1303                 .extra2         = &one,
1304         },
1305 #endif
1306 #ifdef CONFIG_UNEVICTABLE_LRU
1307         {
1308                 .ctl_name       = CTL_UNNUMBERED,
1309                 .procname       = "scan_unevictable_pages",
1310                 .data           = &scan_unevictable_pages,
1311                 .maxlen         = sizeof(scan_unevictable_pages),
1312                 .mode           = 0644,
1313                 .proc_handler   = &scan_unevictable_handler,
1314         },
1315 #endif
1316 /*
1317  * NOTE: do not add new entries to this table unless you have read
1318  * Documentation/sysctl/ctl_unnumbered.txt
1319  */
1320         { .ctl_name = 0 }
1321 };
1322
1323 #if defined(CONFIG_BINFMT_MISC) || defined(CONFIG_BINFMT_MISC_MODULE)
1324 static struct ctl_table binfmt_misc_table[] = {
1325         { .ctl_name = 0 }
1326 };
1327 #endif
1328
1329 static struct ctl_table fs_table[] = {
1330         {
1331                 .ctl_name       = FS_NRINODE,
1332                 .procname       = "inode-nr",
1333                 .data           = &inodes_stat,
1334                 .maxlen         = 2*sizeof(int),
1335                 .mode           = 0444,
1336                 .proc_handler   = &proc_dointvec,
1337         },
1338         {
1339                 .ctl_name       = FS_STATINODE,
1340                 .procname       = "inode-state",
1341                 .data           = &inodes_stat,
1342                 .maxlen         = 7*sizeof(int),
1343                 .mode           = 0444,
1344                 .proc_handler   = &proc_dointvec,
1345         },
1346         {
1347                 .procname       = "file-nr",
1348                 .data           = &files_stat,
1349                 .maxlen         = 3*sizeof(int),
1350                 .mode           = 0444,
1351                 .proc_handler   = &proc_nr_files,
1352         },
1353         {
1354                 .ctl_name       = FS_MAXFILE,
1355                 .procname       = "file-max",
1356                 .data           = &files_stat.max_files,
1357                 .maxlen         = sizeof(int),
1358                 .mode           = 0644,
1359                 .proc_handler   = &proc_dointvec,
1360         },
1361         {
1362                 .ctl_name       = CTL_UNNUMBERED,
1363                 .procname       = "nr_open",
1364                 .data           = &sysctl_nr_open,
1365                 .maxlen         = sizeof(int),
1366                 .mode           = 0644,
1367                 .proc_handler   = &proc_dointvec_minmax,
1368                 .extra1         = &sysctl_nr_open_min,
1369                 .extra2         = &sysctl_nr_open_max,
1370         },
1371         {
1372                 .ctl_name       = FS_DENTRY,
1373                 .procname       = "dentry-state",
1374                 .data           = &dentry_stat,
1375                 .maxlen         = 6*sizeof(int),
1376                 .mode           = 0444,
1377                 .proc_handler   = &proc_dointvec,
1378         },
1379         {
1380                 .ctl_name       = FS_OVERFLOWUID,
1381                 .procname       = "overflowuid",
1382                 .data           = &fs_overflowuid,
1383                 .maxlen         = sizeof(int),
1384                 .mode           = 0644,
1385                 .proc_handler   = &proc_dointvec_minmax,
1386                 .strategy       = &sysctl_intvec,
1387                 .extra1         = &minolduid,
1388                 .extra2         = &maxolduid,
1389         },
1390         {
1391                 .ctl_name       = FS_OVERFLOWGID,
1392                 .procname       = "overflowgid",
1393                 .data           = &fs_overflowgid,
1394                 .maxlen         = sizeof(int),
1395                 .mode           = 0644,
1396                 .proc_handler   = &proc_dointvec_minmax,
1397                 .strategy       = &sysctl_intvec,
1398                 .extra1         = &minolduid,
1399                 .extra2         = &maxolduid,
1400         },
1401 #ifdef CONFIG_FILE_LOCKING
1402         {
1403                 .ctl_name       = FS_LEASES,
1404                 .procname       = "leases-enable",
1405                 .data           = &leases_enable,
1406                 .maxlen         = sizeof(int),
1407                 .mode           = 0644,
1408                 .proc_handler   = &proc_dointvec,
1409         },
1410 #endif
1411 #ifdef CONFIG_DNOTIFY
1412         {
1413                 .ctl_name       = FS_DIR_NOTIFY,
1414                 .procname       = "dir-notify-enable",
1415                 .data           = &dir_notify_enable,
1416                 .maxlen         = sizeof(int),
1417                 .mode           = 0644,
1418                 .proc_handler   = &proc_dointvec,
1419         },
1420 #endif
1421 #ifdef CONFIG_MMU
1422 #ifdef CONFIG_FILE_LOCKING
1423         {
1424                 .ctl_name       = FS_LEASE_TIME,
1425                 .procname       = "lease-break-time",
1426                 .data           = &lease_break_time,
1427                 .maxlen         = sizeof(int),
1428                 .mode           = 0644,
1429                 .proc_handler   = &proc_dointvec,
1430         },
1431 #endif
1432 #ifdef CONFIG_AIO
1433         {
1434                 .procname       = "aio-nr",
1435                 .data           = &aio_nr,
1436                 .maxlen         = sizeof(aio_nr),
1437                 .mode           = 0444,
1438                 .proc_handler   = &proc_doulongvec_minmax,
1439         },
1440         {
1441                 .procname       = "aio-max-nr",
1442                 .data           = &aio_max_nr,
1443                 .maxlen         = sizeof(aio_max_nr),
1444                 .mode           = 0644,
1445                 .proc_handler   = &proc_doulongvec_minmax,
1446         },
1447 #endif /* CONFIG_AIO */
1448 #ifdef CONFIG_INOTIFY_USER
1449         {
1450                 .ctl_name       = FS_INOTIFY,
1451                 .procname       = "inotify",
1452                 .mode           = 0555,
1453                 .child          = inotify_table,
1454         },
1455 #endif
1456 #ifdef CONFIG_EPOLL
1457         {
1458                 .procname       = "epoll",
1459                 .mode           = 0555,
1460                 .child          = epoll_table,
1461         },
1462 #endif
1463 #endif
1464         {
1465                 .ctl_name       = KERN_SETUID_DUMPABLE,
1466                 .procname       = "suid_dumpable",
1467                 .data           = &suid_dumpable,
1468                 .maxlen         = sizeof(int),
1469                 .mode           = 0644,
1470                 .proc_handler   = &proc_dointvec_minmax,
1471                 .strategy       = &sysctl_intvec,
1472                 .extra1         = &zero,
1473                 .extra2         = &two,
1474         },
1475 #if defined(CONFIG_BINFMT_MISC) || defined(CONFIG_BINFMT_MISC_MODULE)
1476         {
1477                 .ctl_name       = CTL_UNNUMBERED,
1478                 .procname       = "binfmt_misc",
1479                 .mode           = 0555,
1480                 .child          = binfmt_misc_table,
1481         },
1482 #endif
1483 /*
1484  * NOTE: do not add new entries to this table unless you have read
1485  * Documentation/sysctl/ctl_unnumbered.txt
1486  */
1487         { .ctl_name = 0 }
1488 };
1489
1490 static struct ctl_table debug_table[] = {
1491 #if defined(CONFIG_X86) || defined(CONFIG_PPC)
1492         {
1493                 .ctl_name       = CTL_UNNUMBERED,
1494                 .procname       = "exception-trace",
1495                 .data           = &show_unhandled_signals,
1496                 .maxlen         = sizeof(int),
1497                 .mode           = 0644,
1498                 .proc_handler   = proc_dointvec
1499         },
1500 #endif
1501         { .ctl_name = 0 }
1502 };
1503
1504 static struct ctl_table dev_table[] = {
1505         { .ctl_name = 0 }
1506 };
1507
1508 static DEFINE_SPINLOCK(sysctl_lock);
1509
1510 /* called under sysctl_lock */
1511 static int use_table(struct ctl_table_header *p)
1512 {
1513         if (unlikely(p->unregistering))
1514                 return 0;
1515         p->used++;
1516         return 1;
1517 }
1518
1519 /* called under sysctl_lock */
1520 static void unuse_table(struct ctl_table_header *p)
1521 {
1522         if (!--p->used)
1523                 if (unlikely(p->unregistering))
1524                         complete(p->unregistering);
1525 }
1526
1527 /* called under sysctl_lock, will reacquire if has to wait */
1528 static void start_unregistering(struct ctl_table_header *p)
1529 {
1530         /*
1531          * if p->used is 0, nobody will ever touch that entry again;
1532          * we'll eliminate all paths to it before dropping sysctl_lock
1533          */
1534         if (unlikely(p->used)) {
1535                 struct completion wait;
1536                 init_completion(&wait);
1537                 p->unregistering = &wait;
1538                 spin_unlock(&sysctl_lock);
1539                 wait_for_completion(&wait);
1540                 spin_lock(&sysctl_lock);
1541         } else {
1542                 /* anything non-NULL; we'll never dereference it */
1543                 p->unregistering = ERR_PTR(-EINVAL);
1544         }
1545         /*
1546          * do not remove from the list until nobody holds it; walking the
1547          * list in do_sysctl() relies on that.
1548          */
1549         list_del_init(&p->ctl_entry);
1550 }
1551
1552 void sysctl_head_get(struct ctl_table_header *head)
1553 {
1554         spin_lock(&sysctl_lock);
1555         head->count++;
1556         spin_unlock(&sysctl_lock);
1557 }
1558
1559 void sysctl_head_put(struct ctl_table_header *head)
1560 {
1561         spin_lock(&sysctl_lock);
1562         if (!--head->count)
1563                 kfree(head);
1564         spin_unlock(&sysctl_lock);
1565 }
1566
1567 struct ctl_table_header *sysctl_head_grab(struct ctl_table_header *head)
1568 {
1569         if (!head)
1570                 BUG();
1571         spin_lock(&sysctl_lock);
1572         if (!use_table(head))
1573                 head = ERR_PTR(-ENOENT);
1574         spin_unlock(&sysctl_lock);
1575         return head;
1576 }
1577
1578 void sysctl_head_finish(struct ctl_table_header *head)
1579 {
1580         if (!head)
1581                 return;
1582         spin_lock(&sysctl_lock);
1583         unuse_table(head);
1584         spin_unlock(&sysctl_lock);
1585 }
1586
1587 static struct ctl_table_set *
1588 lookup_header_set(struct ctl_table_root *root, struct nsproxy *namespaces)
1589 {
1590         struct ctl_table_set *set = &root->default_set;
1591         if (root->lookup)
1592                 set = root->lookup(root, namespaces);
1593         return set;
1594 }
1595
1596 static struct list_head *
1597 lookup_header_list(struct ctl_table_root *root, struct nsproxy *namespaces)
1598 {
1599         struct ctl_table_set *set = lookup_header_set(root, namespaces);
1600         return &set->list;
1601 }
1602
1603 struct ctl_table_header *__sysctl_head_next(struct nsproxy *namespaces,
1604                                             struct ctl_table_header *prev)
1605 {
1606         struct ctl_table_root *root;
1607         struct list_head *header_list;
1608         struct ctl_table_header *head;
1609         struct list_head *tmp;
1610
1611         spin_lock(&sysctl_lock);
1612         if (prev) {
1613                 head = prev;
1614                 tmp = &prev->ctl_entry;
1615                 unuse_table(prev);
1616                 goto next;
1617         }
1618         tmp = &root_table_header.ctl_entry;
1619         for (;;) {
1620                 head = list_entry(tmp, struct ctl_table_header, ctl_entry);
1621
1622                 if (!use_table(head))
1623                         goto next;
1624                 spin_unlock(&sysctl_lock);
1625                 return head;
1626         next:
1627                 root = head->root;
1628                 tmp = tmp->next;
1629                 header_list = lookup_header_list(root, namespaces);
1630                 if (tmp != header_list)
1631                         continue;
1632
1633                 do {
1634                         root = list_entry(root->root_list.next,
1635                                         struct ctl_table_root, root_list);
1636                         if (root == &sysctl_table_root)
1637                                 goto out;
1638                         header_list = lookup_header_list(root, namespaces);
1639                 } while (list_empty(header_list));
1640                 tmp = header_list->next;
1641         }
1642 out:
1643         spin_unlock(&sysctl_lock);
1644         return NULL;
1645 }
1646
1647 struct ctl_table_header *sysctl_head_next(struct ctl_table_header *prev)
1648 {
1649         return __sysctl_head_next(current->nsproxy, prev);
1650 }
1651
1652 void register_sysctl_root(struct ctl_table_root *root)
1653 {
1654         spin_lock(&sysctl_lock);
1655         list_add_tail(&root->root_list, &sysctl_table_root.root_list);
1656         spin_unlock(&sysctl_lock);
1657 }
1658
1659 #ifdef CONFIG_SYSCTL_SYSCALL
1660 /* Perform the actual read/write of a sysctl table entry. */
1661 static int do_sysctl_strategy(struct ctl_table_root *root,
1662                         struct ctl_table *table,
1663                         void __user *oldval, size_t __user *oldlenp,
1664                         void __user *newval, size_t newlen)
1665 {
1666         int op = 0, rc;
1667
1668         if (oldval)
1669                 op |= MAY_READ;
1670         if (newval)
1671                 op |= MAY_WRITE;
1672         if (sysctl_perm(root, table, op))
1673                 return -EPERM;
1674
1675         if (table->strategy) {
1676                 rc = table->strategy(table, oldval, oldlenp, newval, newlen);
1677                 if (rc < 0)
1678                         return rc;
1679                 if (rc > 0)
1680                         return 0;
1681         }
1682
1683         /* If there is no strategy routine, or if the strategy returns
1684          * zero, proceed with automatic r/w */
1685         if (table->data && table->maxlen) {
1686                 rc = sysctl_data(table, oldval, oldlenp, newval, newlen);
1687                 if (rc < 0)
1688                         return rc;
1689         }
1690         return 0;
1691 }
1692
1693 static int parse_table(int __user *name, int nlen,
1694                        void __user *oldval, size_t __user *oldlenp,
1695                        void __user *newval, size_t newlen,
1696                        struct ctl_table_root *root,
1697                        struct ctl_table *table)
1698 {
1699         int n;
1700 repeat:
1701         if (!nlen)
1702                 return -ENOTDIR;
1703         if (get_user(n, name))
1704                 return -EFAULT;
1705         for ( ; table->ctl_name || table->procname; table++) {
1706                 if (!table->ctl_name)
1707                         continue;
1708                 if (n == table->ctl_name) {
1709                         int error;
1710                         if (table->child) {
1711                                 if (sysctl_perm(root, table, MAY_EXEC))
1712                                         return -EPERM;
1713                                 name++;
1714                                 nlen--;
1715                                 table = table->child;
1716                                 goto repeat;
1717                         }
1718                         error = do_sysctl_strategy(root, table,
1719                                                    oldval, oldlenp,
1720                                                    newval, newlen);
1721                         return error;
1722                 }
1723         }
1724         return -ENOTDIR;
1725 }
1726
1727 int do_sysctl(int __user *name, int nlen, void __user *oldval, size_t __user *oldlenp,
1728                void __user *newval, size_t newlen)
1729 {
1730         struct ctl_table_header *head;
1731         int error = -ENOTDIR;
1732
1733         if (nlen <= 0 || nlen >= CTL_MAXNAME)
1734                 return -ENOTDIR;
1735         if (oldval) {
1736                 int old_len;
1737                 if (!oldlenp || get_user(old_len, oldlenp))
1738                         return -EFAULT;
1739         }
1740
1741         for (head = sysctl_head_next(NULL); head;
1742                         head = sysctl_head_next(head)) {
1743                 error = parse_table(name, nlen, oldval, oldlenp,
1744                                         newval, newlen,
1745                                         head->root, head->ctl_table);
1746                 if (error != -ENOTDIR) {
1747                         sysctl_head_finish(head);
1748                         break;
1749                 }
1750         }
1751         return error;
1752 }
1753
1754 SYSCALL_DEFINE1(sysctl, struct __sysctl_args __user *, args)
1755 {
1756         struct __sysctl_args tmp;
1757         int error;
1758
1759         if (copy_from_user(&tmp, args, sizeof(tmp)))
1760                 return -EFAULT;
1761
1762         error = deprecated_sysctl_warning(&tmp);
1763         if (error)
1764                 goto out;
1765
1766         lock_kernel();
1767         error = do_sysctl(tmp.name, tmp.nlen, tmp.oldval, tmp.oldlenp,
1768                           tmp.newval, tmp.newlen);
1769         unlock_kernel();
1770 out:
1771         return error;
1772 }
1773 #endif /* CONFIG_SYSCTL_SYSCALL */
1774
1775 /*
1776  * sysctl_perm does NOT grant the superuser all rights automatically, because
1777  * some sysctl variables are readonly even to root.
1778  */
1779
1780 static int test_perm(int mode, int op)
1781 {
1782         if (!current_euid())
1783                 mode >>= 6;
1784         else if (in_egroup_p(0))
1785                 mode >>= 3;
1786         if ((op & ~mode & (MAY_READ|MAY_WRITE|MAY_EXEC)) == 0)
1787                 return 0;
1788         return -EACCES;
1789 }
1790
1791 int sysctl_perm(struct ctl_table_root *root, struct ctl_table *table, int op)
1792 {
1793         int error;
1794         int mode;
1795
1796         error = security_sysctl(table, op & (MAY_READ | MAY_WRITE | MAY_EXEC));
1797         if (error)
1798                 return error;
1799
1800         if (root->permissions)
1801                 mode = root->permissions(root, current->nsproxy, table);
1802         else
1803                 mode = table->mode;
1804
1805         return test_perm(mode, op);
1806 }
1807
1808 static void sysctl_set_parent(struct ctl_table *parent, struct ctl_table *table)
1809 {
1810         for (; table->ctl_name || table->procname; table++) {
1811                 table->parent = parent;
1812                 if (table->child)
1813                         sysctl_set_parent(table, table->child);
1814         }
1815 }
1816
1817 static __init int sysctl_init(void)
1818 {
1819         sysctl_set_parent(NULL, root_table);
1820 #ifdef CONFIG_SYSCTL_SYSCALL_CHECK
1821         {
1822                 int err;
1823                 err = sysctl_check_table(current->nsproxy, root_table);
1824         }
1825 #endif
1826         return 0;
1827 }
1828
1829 core_initcall(sysctl_init);
1830
1831 static struct ctl_table *is_branch_in(struct ctl_table *branch,
1832                                       struct ctl_table *table)
1833 {
1834         struct ctl_table *p;
1835         const char *s = branch->procname;
1836
1837         /* branch should have named subdirectory as its first element */
1838         if (!s || !branch->child)
1839                 return NULL;
1840
1841         /* ... and nothing else */
1842         if (branch[1].procname || branch[1].ctl_name)
1843                 return NULL;
1844
1845         /* table should contain subdirectory with the same name */
1846         for (p = table; p->procname || p->ctl_name; p++) {
1847                 if (!p->child)
1848                         continue;
1849                 if (p->procname && strcmp(p->procname, s) == 0)
1850                         return p;
1851         }
1852         return NULL;
1853 }
1854
1855 /* see if attaching q to p would be an improvement */
1856 static void try_attach(struct ctl_table_header *p, struct ctl_table_header *q)
1857 {
1858         struct ctl_table *to = p->ctl_table, *by = q->ctl_table;
1859         struct ctl_table *next;
1860         int is_better = 0;
1861         int not_in_parent = !p->attached_by;
1862
1863         while ((next = is_branch_in(by, to)) != NULL) {
1864                 if (by == q->attached_by)
1865                         is_better = 1;
1866                 if (to == p->attached_by)
1867                         not_in_parent = 1;
1868                 by = by->child;
1869                 to = next->child;
1870         }
1871
1872         if (is_better && not_in_parent) {
1873                 q->attached_by = by;
1874                 q->attached_to = to;
1875                 q->parent = p;
1876         }
1877 }
1878
1879 /**
1880  * __register_sysctl_paths - register a sysctl hierarchy
1881  * @root: List of sysctl headers to register on
1882  * @namespaces: Data to compute which lists of sysctl entries are visible
1883  * @path: The path to the directory the sysctl table is in.
1884  * @table: the top-level table structure
1885  *
1886  * Register a sysctl table hierarchy. @table should be a filled in ctl_table
1887  * array. A completely 0 filled entry terminates the table.
1888  *
1889  * The members of the &struct ctl_table structure are used as follows:
1890  *
1891  * ctl_name - This is the numeric sysctl value used by sysctl(2). The number
1892  *            must be unique within that level of sysctl
1893  *
1894  * procname - the name of the sysctl file under /proc/sys. Set to %NULL to not
1895  *            enter a sysctl file
1896  *
1897  * data - a pointer to data for use by proc_handler
1898  *
1899  * maxlen - the maximum size in bytes of the data
1900  *
1901  * mode - the file permissions for the /proc/sys file, and for sysctl(2)
1902  *
1903  * child - a pointer to the child sysctl table if this entry is a directory, or
1904  *         %NULL.
1905  *
1906  * proc_handler - the text handler routine (described below)
1907  *
1908  * strategy - the strategy routine (described below)
1909  *
1910  * de - for internal use by the sysctl routines
1911  *
1912  * extra1, extra2 - extra pointers usable by the proc handler routines
1913  *
1914  * Leaf nodes in the sysctl tree will be represented by a single file
1915  * under /proc; non-leaf nodes will be represented by directories.
1916  *
1917  * sysctl(2) can automatically manage read and write requests through
1918  * the sysctl table.  The data and maxlen fields of the ctl_table
1919  * struct enable minimal validation of the values being written to be
1920  * performed, and the mode field allows minimal authentication.
1921  *
1922  * More sophisticated management can be enabled by the provision of a
1923  * strategy routine with the table entry.  This will be called before
1924  * any automatic read or write of the data is performed.
1925  *
1926  * The strategy routine may return
1927  *
1928  * < 0 - Error occurred (error is passed to user process)
1929  *
1930  * 0   - OK - proceed with automatic read or write.
1931  *
1932  * > 0 - OK - read or write has been done by the strategy routine, so
1933  *       return immediately.
1934  *
1935  * There must be a proc_handler routine for any terminal nodes
1936  * mirrored under /proc/sys (non-terminals are handled by a built-in
1937  * directory handler).  Several default handlers are available to
1938  * cover common cases -
1939  *
1940  * proc_dostring(), proc_dointvec(), proc_dointvec_jiffies(),
1941  * proc_dointvec_userhz_jiffies(), proc_dointvec_minmax(),
1942  * proc_doulongvec_ms_jiffies_minmax(), proc_doulongvec_minmax()
1943  *
1944  * It is the handler's job to read the input buffer from user memory
1945  * and process it. The handler should return 0 on success.
1946  *
1947  * This routine returns %NULL on a failure to register, and a pointer
1948  * to the table header on success.
1949  */
1950 struct ctl_table_header *__register_sysctl_paths(
1951         struct ctl_table_root *root,
1952         struct nsproxy *namespaces,
1953         const struct ctl_path *path, struct ctl_table *table)
1954 {
1955         struct ctl_table_header *header;
1956         struct ctl_table *new, **prevp;
1957         unsigned int n, npath;
1958         struct ctl_table_set *set;
1959
1960         /* Count the path components */
1961         for (npath = 0; path[npath].ctl_name || path[npath].procname; ++npath)
1962                 ;
1963
1964         /*
1965          * For each path component, allocate a 2-element ctl_table array.
1966          * The first array element will be filled with the sysctl entry
1967          * for this, the second will be the sentinel (ctl_name == 0).
1968          *
1969          * We allocate everything in one go so that we don't have to
1970          * worry about freeing additional memory in unregister_sysctl_table.
1971          */
1972         header = kzalloc(sizeof(struct ctl_table_header) +
1973                          (2 * npath * sizeof(struct ctl_table)), GFP_KERNEL);
1974         if (!header)
1975                 return NULL;
1976
1977         new = (struct ctl_table *) (header + 1);
1978
1979         /* Now connect the dots */
1980         prevp = &header->ctl_table;
1981         for (n = 0; n < npath; ++n, ++path) {
1982                 /* Copy the procname */
1983                 new->procname = path->procname;
1984                 new->ctl_name = path->ctl_name;
1985                 new->mode     = 0555;
1986
1987                 *prevp = new;
1988                 prevp = &new->child;
1989
1990                 new += 2;
1991         }
1992         *prevp = table;
1993         header->ctl_table_arg = table;
1994
1995         INIT_LIST_HEAD(&header->ctl_entry);
1996         header->used = 0;
1997         header->unregistering = NULL;
1998         header->root = root;
1999         sysctl_set_parent(NULL, header->ctl_table);
2000         header->count = 1;
2001 #ifdef CONFIG_SYSCTL_SYSCALL_CHECK
2002         if (sysctl_check_table(namespaces, header->ctl_table)) {
2003                 kfree(header);
2004                 return NULL;
2005         }
2006 #endif
2007         spin_lock(&sysctl_lock);
2008         header->set = lookup_header_set(root, namespaces);
2009         header->attached_by = header->ctl_table;
2010         header->attached_to = root_table;
2011         header->parent = &root_table_header;
2012         for (set = header->set; set; set = set->parent) {
2013                 struct ctl_table_header *p;
2014                 list_for_each_entry(p, &set->list, ctl_entry) {
2015                         if (p->unregistering)
2016                                 continue;
2017                         try_attach(p, header);
2018                 }
2019         }
2020         header->parent->count++;
2021         list_add_tail(&header->ctl_entry, &header->set->list);
2022         spin_unlock(&sysctl_lock);
2023
2024         return header;
2025 }
2026
2027 /**
2028  * register_sysctl_table_path - register a sysctl table hierarchy
2029  * @path: The path to the directory the sysctl table is in.
2030  * @table: the top-level table structure
2031  *
2032  * Register a sysctl table hierarchy. @table should be a filled in ctl_table
2033  * array. A completely 0 filled entry terminates the table.
2034  *
2035  * See __register_sysctl_paths for more details.
2036  */
2037 struct ctl_table_header *register_sysctl_paths(const struct ctl_path *path,
2038                                                 struct ctl_table *table)
2039 {
2040         return __register_sysctl_paths(&sysctl_table_root, current->nsproxy,
2041                                         path, table);
2042 }
2043
2044 /**
2045  * register_sysctl_table - register a sysctl table hierarchy
2046  * @table: the top-level table structure
2047  *
2048  * Register a sysctl table hierarchy. @table should be a filled in ctl_table
2049  * array. A completely 0 filled entry terminates the table.
2050  *
2051  * See register_sysctl_paths for more details.
2052  */
2053 struct ctl_table_header *register_sysctl_table(struct ctl_table *table)
2054 {
2055         static const struct ctl_path null_path[] = { {} };
2056
2057         return register_sysctl_paths(null_path, table);
2058 }
2059
2060 /**
2061  * unregister_sysctl_table - unregister a sysctl table hierarchy
2062  * @header: the header returned from register_sysctl_table
2063  *
2064  * Unregisters the sysctl table and all children. proc entries may not
2065  * actually be removed until they are no longer used by anyone.
2066  */
2067 void unregister_sysctl_table(struct ctl_table_header * header)
2068 {
2069         might_sleep();
2070
2071         if (header == NULL)
2072                 return;
2073
2074         spin_lock(&sysctl_lock);
2075         start_unregistering(header);
2076         if (!--header->parent->count) {
2077                 WARN_ON(1);
2078                 kfree(header->parent);
2079         }
2080         if (!--header->count)
2081                 kfree(header);
2082         spin_unlock(&sysctl_lock);
2083 }
2084
2085 int sysctl_is_seen(struct ctl_table_header *p)
2086 {
2087         struct ctl_table_set *set = p->set;
2088         int res;
2089         spin_lock(&sysctl_lock);
2090         if (p->unregistering)
2091                 res = 0;
2092         else if (!set->is_seen)
2093                 res = 1;
2094         else
2095                 res = set->is_seen(set);
2096         spin_unlock(&sysctl_lock);
2097         return res;
2098 }
2099
2100 void setup_sysctl_set(struct ctl_table_set *p,
2101         struct ctl_table_set *parent,
2102         int (*is_seen)(struct ctl_table_set *))
2103 {
2104         INIT_LIST_HEAD(&p->list);
2105         p->parent = parent ? parent : &sysctl_table_root.default_set;
2106         p->is_seen = is_seen;
2107 }
2108
2109 #else /* !CONFIG_SYSCTL */
2110 struct ctl_table_header *register_sysctl_table(struct ctl_table * table)
2111 {
2112         return NULL;
2113 }
2114
2115 struct ctl_table_header *register_sysctl_paths(const struct ctl_path *path,
2116                                                     struct ctl_table *table)
2117 {
2118         return NULL;
2119 }
2120
2121 void unregister_sysctl_table(struct ctl_table_header * table)
2122 {
2123 }
2124
2125 void setup_sysctl_set(struct ctl_table_set *p,
2126         struct ctl_table_set *parent,
2127         int (*is_seen)(struct ctl_table_set *))
2128 {
2129 }
2130
2131 void sysctl_head_put(struct ctl_table_header *head)
2132 {
2133 }
2134
2135 #endif /* CONFIG_SYSCTL */
2136
2137 /*
2138  * /proc/sys support
2139  */
2140
2141 #ifdef CONFIG_PROC_SYSCTL
2142
2143 static int _proc_do_string(void* data, int maxlen, int write,
2144                            struct file *filp, void __user *buffer,
2145                            size_t *lenp, loff_t *ppos)
2146 {
2147         size_t len;
2148         char __user *p;
2149         char c;
2150
2151         if (!data || !maxlen || !*lenp) {
2152                 *lenp = 0;
2153                 return 0;
2154         }
2155
2156         if (write) {
2157                 len = 0;
2158                 p = buffer;
2159                 while (len < *lenp) {
2160                         if (get_user(c, p++))
2161                                 return -EFAULT;
2162                         if (c == 0 || c == '\n')
2163                                 break;
2164                         len++;
2165                 }
2166                 if (len >= maxlen)
2167                         len = maxlen-1;
2168                 if(copy_from_user(data, buffer, len))
2169                         return -EFAULT;
2170                 ((char *) data)[len] = 0;
2171                 *ppos += *lenp;
2172         } else {
2173                 len = strlen(data);
2174                 if (len > maxlen)
2175                         len = maxlen;
2176
2177                 if (*ppos > len) {
2178                         *lenp = 0;
2179                         return 0;
2180                 }
2181
2182                 data += *ppos;
2183                 len  -= *ppos;
2184
2185                 if (len > *lenp)
2186                         len = *lenp;
2187                 if (len)
2188                         if(copy_to_user(buffer, data, len))
2189                                 return -EFAULT;
2190                 if (len < *lenp) {
2191                         if(put_user('\n', ((char __user *) buffer) + len))
2192                                 return -EFAULT;
2193                         len++;
2194                 }
2195                 *lenp = len;
2196                 *ppos += len;
2197         }
2198         return 0;
2199 }
2200
2201 /**
2202  * proc_dostring - read a string sysctl
2203  * @table: the sysctl table
2204  * @write: %TRUE if this is a write to the sysctl file
2205  * @filp: the file structure
2206  * @buffer: the user buffer
2207  * @lenp: the size of the user buffer
2208  * @ppos: file position
2209  *
2210  * Reads/writes a string from/to the user buffer. If the kernel
2211  * buffer provided is not large enough to hold the string, the
2212  * string is truncated. The copied string is %NULL-terminated.
2213  * If the string is being read by the user process, it is copied
2214  * and a newline '\n' is added. It is truncated if the buffer is
2215  * not large enough.
2216  *
2217  * Returns 0 on success.
2218  */
2219 int proc_dostring(struct ctl_table *table, int write, struct file *filp,
2220                   void __user *buffer, size_t *lenp, loff_t *ppos)
2221 {
2222         return _proc_do_string(table->data, table->maxlen, write, filp,
2223                                buffer, lenp, ppos);
2224 }
2225
2226
2227 static int do_proc_dointvec_conv(int *negp, unsigned long *lvalp,
2228                                  int *valp,
2229                                  int write, void *data)
2230 {
2231         if (write) {
2232                 *valp = *negp ? -*lvalp : *lvalp;
2233         } else {
2234                 int val = *valp;
2235                 if (val < 0) {
2236                         *negp = -1;
2237                         *lvalp = (unsigned long)-val;
2238                 } else {
2239                         *negp = 0;
2240                         *lvalp = (unsigned long)val;
2241                 }
2242         }
2243         return 0;
2244 }
2245
2246 static int __do_proc_dointvec(void *tbl_data, struct ctl_table *table,
2247                   int write, struct file *filp, void __user *buffer,
2248                   size_t *lenp, loff_t *ppos,
2249                   int (*conv)(int *negp, unsigned long *lvalp, int *valp,
2250                               int write, void *data),
2251                   void *data)
2252 {
2253 #define TMPBUFLEN 21
2254         int *i, vleft, first=1, neg, val;
2255         unsigned long lval;
2256         size_t left, len;
2257
2258         char buf[TMPBUFLEN], *p;
2259         char __user *s = buffer;
2260
2261         if (!tbl_data || !table->maxlen || !*lenp ||
2262             (*ppos && !write)) {
2263                 *lenp = 0;
2264                 return 0;
2265         }
2266
2267         i = (int *) tbl_data;
2268         vleft = table->maxlen / sizeof(*i);
2269         left = *lenp;
2270
2271         if (!conv)
2272                 conv = do_proc_dointvec_conv;
2273
2274         for (; left && vleft--; i++, first=0) {
2275                 if (write) {
2276                         while (left) {
2277                                 char c;
2278                                 if (get_user(c, s))
2279                                         return -EFAULT;
2280                                 if (!isspace(c))
2281                                         break;
2282                                 left--;
2283                                 s++;
2284                         }
2285                         if (!left)
2286                                 break;
2287                         neg = 0;
2288                         len = left;
2289                         if (len > sizeof(buf) - 1)
2290                                 len = sizeof(buf) - 1;
2291                         if (copy_from_user(buf, s, len))
2292                                 return -EFAULT;
2293                         buf[len] = 0;
2294                         p = buf;
2295                         if (*p == '-' && left > 1) {
2296                                 neg = 1;
2297                                 p++;
2298                         }
2299                         if (*p < '0' || *p > '9')
2300                                 break;
2301
2302                         lval = simple_strtoul(p, &p, 0);
2303
2304                         len = p-buf;
2305                         if ((len < left) && *p && !isspace(*p))
2306                                 break;
2307                         if (neg)
2308                                 val = -val;
2309                         s += len;
2310                         left -= len;
2311
2312                         if (conv(&neg, &lval, i, 1, data))
2313                                 break;
2314                 } else {
2315                         p = buf;
2316                         if (!first)
2317                                 *p++ = '\t';
2318
2319                         if (conv(&neg, &lval, i, 0, data))
2320                                 break;
2321
2322                         sprintf(p, "%s%lu", neg ? "-" : "", lval);
2323                         len = strlen(buf);
2324                         if (len > left)
2325                                 len = left;
2326                         if(copy_to_user(s, buf, len))
2327                                 return -EFAULT;
2328                         left -= len;
2329                         s += len;
2330                 }
2331         }
2332
2333         if (!write && !first && left) {
2334                 if(put_user('\n', s))
2335                         return -EFAULT;
2336                 left--, s++;
2337         }
2338         if (write) {
2339                 while (left) {
2340                         char c;
2341                         if (get_user(c, s++))
2342                                 return -EFAULT;
2343                         if (!isspace(c))
2344                                 break;
2345                         left--;
2346                 }
2347         }
2348         if (write && first)
2349                 return -EINVAL;
2350         *lenp -= left;
2351         *ppos += *lenp;
2352         return 0;
2353 #undef TMPBUFLEN
2354 }
2355
2356 static int do_proc_dointvec(struct ctl_table *table, int write, struct file *filp,
2357                   void __user *buffer, size_t *lenp, loff_t *ppos,
2358                   int (*conv)(int *negp, unsigned long *lvalp, int *valp,
2359                               int write, void *data),
2360                   void *data)
2361 {
2362         return __do_proc_dointvec(table->data, table, write, filp,
2363                         buffer, lenp, ppos, conv, data);
2364 }
2365
2366 /**
2367  * proc_dointvec - read a vector of integers
2368  * @table: the sysctl table
2369  * @write: %TRUE if this is a write to the sysctl file
2370  * @filp: the file structure
2371  * @buffer: the user buffer
2372  * @lenp: the size of the user buffer
2373  * @ppos: file position
2374  *
2375  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2376  * values from/to the user buffer, treated as an ASCII string.
2377  *
2378  * Returns 0 on success.
2379  */
2380 int proc_dointvec(struct ctl_table *table, int write, struct file *filp,
2381                      void __user *buffer, size_t *lenp, loff_t *ppos)
2382 {
2383     return do_proc_dointvec(table,write,filp,buffer,lenp,ppos,
2384                             NULL,NULL);
2385 }
2386
2387 /*
2388  * Taint values can only be increased
2389  * This means we can safely use a temporary.
2390  */
2391 static int proc_taint(struct ctl_table *table, int write, struct file *filp,
2392                                void __user *buffer, size_t *lenp, loff_t *ppos)
2393 {
2394         struct ctl_table t;
2395         unsigned long tmptaint = get_taint();
2396         int err;
2397
2398         if (write && !capable(CAP_SYS_ADMIN))
2399                 return -EPERM;
2400
2401         t = *table;
2402         t.data = &tmptaint;
2403         err = proc_doulongvec_minmax(&t, write, filp, buffer, lenp, ppos);
2404         if (err < 0)
2405                 return err;
2406
2407         if (write) {
2408                 /*
2409                  * Poor man's atomic or. Not worth adding a primitive
2410                  * to everyone's atomic.h for this
2411                  */
2412                 int i;
2413                 for (i = 0; i < BITS_PER_LONG && tmptaint >> i; i++) {
2414                         if ((tmptaint >> i) & 1)
2415                                 add_taint(i);
2416                 }
2417         }
2418
2419         return err;
2420 }
2421
2422 struct do_proc_dointvec_minmax_conv_param {
2423         int *min;
2424         int *max;
2425 };
2426
2427 static int do_proc_dointvec_minmax_conv(int *negp, unsigned long *lvalp,
2428                                         int *valp,
2429                                         int write, void *data)
2430 {
2431         struct do_proc_dointvec_minmax_conv_param *param = data;
2432         if (write) {
2433                 int val = *negp ? -*lvalp : *lvalp;
2434                 if ((param->min && *param->min > val) ||
2435                     (param->max && *param->max < val))
2436                         return -EINVAL;
2437                 *valp = val;
2438         } else {
2439                 int val = *valp;
2440                 if (val < 0) {
2441                         *negp = -1;
2442                         *lvalp = (unsigned long)-val;
2443                 } else {
2444                         *negp = 0;
2445                         *lvalp = (unsigned long)val;
2446                 }
2447         }
2448         return 0;
2449 }
2450
2451 /**
2452  * proc_dointvec_minmax - read a vector of integers with min/max values
2453  * @table: the sysctl table
2454  * @write: %TRUE if this is a write to the sysctl file
2455  * @filp: the file structure
2456  * @buffer: the user buffer
2457  * @lenp: the size of the user buffer
2458  * @ppos: file position
2459  *
2460  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2461  * values from/to the user buffer, treated as an ASCII string.
2462  *
2463  * This routine will ensure the values are within the range specified by
2464  * table->extra1 (min) and table->extra2 (max).
2465  *
2466  * Returns 0 on success.
2467  */
2468 int proc_dointvec_minmax(struct ctl_table *table, int write, struct file *filp,
2469                   void __user *buffer, size_t *lenp, loff_t *ppos)
2470 {
2471         struct do_proc_dointvec_minmax_conv_param param = {
2472                 .min = (int *) table->extra1,
2473                 .max = (int *) table->extra2,
2474         };
2475         return do_proc_dointvec(table, write, filp, buffer, lenp, ppos,
2476                                 do_proc_dointvec_minmax_conv, &param);
2477 }
2478
2479 static int __do_proc_doulongvec_minmax(void *data, struct ctl_table *table, int write,
2480                                      struct file *filp,
2481                                      void __user *buffer,
2482                                      size_t *lenp, loff_t *ppos,
2483                                      unsigned long convmul,
2484                                      unsigned long convdiv)
2485 {
2486 #define TMPBUFLEN 21
2487         unsigned long *i, *min, *max, val;
2488         int vleft, first=1, neg;
2489         size_t len, left;
2490         char buf[TMPBUFLEN], *p;
2491         char __user *s = buffer;
2492
2493         if (!data || !table->maxlen || !*lenp ||
2494             (*ppos && !write)) {
2495                 *lenp = 0;
2496                 return 0;
2497         }
2498
2499         i = (unsigned long *) data;
2500         min = (unsigned long *) table->extra1;
2501         max = (unsigned long *) table->extra2;
2502         vleft = table->maxlen / sizeof(unsigned long);
2503         left = *lenp;
2504
2505         for (; left && vleft--; i++, min++, max++, first=0) {
2506                 if (write) {
2507                         while (left) {
2508                                 char c;
2509                                 if (get_user(c, s))
2510                                         return -EFAULT;
2511                                 if (!isspace(c))
2512                                         break;
2513                                 left--;
2514                                 s++;
2515                         }
2516                         if (!left)
2517                                 break;
2518                         neg = 0;
2519                         len = left;
2520                         if (len > TMPBUFLEN-1)
2521                                 len = TMPBUFLEN-1;
2522                         if (copy_from_user(buf, s, len))
2523                                 return -EFAULT;
2524                         buf[len] = 0;
2525                         p = buf;
2526                         if (*p == '-' && left > 1) {
2527                                 neg = 1;
2528                                 p++;
2529                         }
2530                         if (*p < '0' || *p > '9')
2531                                 break;
2532                         val = simple_strtoul(p, &p, 0) * convmul / convdiv ;
2533                         len = p-buf;
2534                         if ((len < left) && *p && !isspace(*p))
2535                                 break;
2536                         if (neg)
2537                                 val = -val;
2538                         s += len;
2539                         left -= len;
2540
2541                         if(neg)
2542                                 continue;
2543                         if ((min && val < *min) || (max && val > *max))
2544                                 continue;
2545                         *i = val;
2546                 } else {
2547                         p = buf;
2548                         if (!first)
2549                                 *p++ = '\t';
2550                         sprintf(p, "%lu", convdiv * (*i) / convmul);
2551                         len = strlen(buf);
2552                         if (len > left)
2553                                 len = left;
2554                         if(copy_to_user(s, buf, len))
2555                                 return -EFAULT;
2556                         left -= len;
2557                         s += len;
2558                 }
2559         }
2560
2561         if (!write && !first && left) {
2562                 if(put_user('\n', s))
2563                         return -EFAULT;
2564                 left--, s++;
2565         }
2566         if (write) {
2567                 while (left) {
2568                         char c;
2569                         if (get_user(c, s++))
2570                                 return -EFAULT;
2571                         if (!isspace(c))
2572                                 break;
2573                         left--;
2574                 }
2575         }
2576         if (write && first)
2577                 return -EINVAL;
2578         *lenp -= left;
2579         *ppos += *lenp;
2580         return 0;
2581 #undef TMPBUFLEN
2582 }
2583
2584 static int do_proc_doulongvec_minmax(struct ctl_table *table, int write,
2585                                      struct file *filp,
2586                                      void __user *buffer,
2587                                      size_t *lenp, loff_t *ppos,
2588                                      unsigned long convmul,
2589                                      unsigned long convdiv)
2590 {
2591         return __do_proc_doulongvec_minmax(table->data, table, write,
2592                         filp, buffer, lenp, ppos, convmul, convdiv);
2593 }
2594
2595 /**
2596  * proc_doulongvec_minmax - read a vector of long integers with min/max values
2597  * @table: the sysctl table
2598  * @write: %TRUE if this is a write to the sysctl file
2599  * @filp: the file structure
2600  * @buffer: the user buffer
2601  * @lenp: the size of the user buffer
2602  * @ppos: file position
2603  *
2604  * Reads/writes up to table->maxlen/sizeof(unsigned long) unsigned long
2605  * values from/to the user buffer, treated as an ASCII string.
2606  *
2607  * This routine will ensure the values are within the range specified by
2608  * table->extra1 (min) and table->extra2 (max).
2609  *
2610  * Returns 0 on success.
2611  */
2612 int proc_doulongvec_minmax(struct ctl_table *table, int write, struct file *filp,
2613                            void __user *buffer, size_t *lenp, loff_t *ppos)
2614 {
2615     return do_proc_doulongvec_minmax(table, write, filp, buffer, lenp, ppos, 1l, 1l);
2616 }
2617
2618 /**
2619  * proc_doulongvec_ms_jiffies_minmax - read a vector of millisecond values with min/max values
2620  * @table: the sysctl table
2621  * @write: %TRUE if this is a write to the sysctl file
2622  * @filp: the file structure
2623  * @buffer: the user buffer
2624  * @lenp: the size of the user buffer
2625  * @ppos: file position
2626  *
2627  * Reads/writes up to table->maxlen/sizeof(unsigned long) unsigned long
2628  * values from/to the user buffer, treated as an ASCII string. The values
2629  * are treated as milliseconds, and converted to jiffies when they are stored.
2630  *
2631  * This routine will ensure the values are within the range specified by
2632  * table->extra1 (min) and table->extra2 (max).
2633  *
2634  * Returns 0 on success.
2635  */
2636 int proc_doulongvec_ms_jiffies_minmax(struct ctl_table *table, int write,
2637                                       struct file *filp,
2638                                       void __user *buffer,
2639                                       size_t *lenp, loff_t *ppos)
2640 {
2641     return do_proc_doulongvec_minmax(table, write, filp, buffer,
2642                                      lenp, ppos, HZ, 1000l);
2643 }
2644
2645
2646 static int do_proc_dointvec_jiffies_conv(int *negp, unsigned long *lvalp,
2647                                          int *valp,
2648                                          int write, void *data)
2649 {
2650         if (write) {
2651                 if (*lvalp > LONG_MAX / HZ)
2652                         return 1;
2653                 *valp = *negp ? -(*lvalp*HZ) : (*lvalp*HZ);
2654         } else {
2655                 int val = *valp;
2656                 unsigned long lval;
2657                 if (val < 0) {
2658                         *negp = -1;
2659                         lval = (unsigned long)-val;
2660                 } else {
2661                         *negp = 0;
2662                         lval = (unsigned long)val;
2663                 }
2664                 *lvalp = lval / HZ;
2665         }
2666         return 0;
2667 }
2668
2669 static int do_proc_dointvec_userhz_jiffies_conv(int *negp, unsigned long *lvalp,
2670                                                 int *valp,
2671                                                 int write, void *data)
2672 {
2673         if (write) {
2674                 if (USER_HZ < HZ && *lvalp > (LONG_MAX / HZ) * USER_HZ)
2675                         return 1;
2676                 *valp = clock_t_to_jiffies(*negp ? -*lvalp : *lvalp);
2677         } else {
2678                 int val = *valp;
2679                 unsigned long lval;
2680                 if (val < 0) {
2681                         *negp = -1;
2682                         lval = (unsigned long)-val;
2683                 } else {
2684                         *negp = 0;
2685                         lval = (unsigned long)val;
2686                 }
2687                 *lvalp = jiffies_to_clock_t(lval);
2688         }
2689         return 0;
2690 }
2691
2692 static int do_proc_dointvec_ms_jiffies_conv(int *negp, unsigned long *lvalp,
2693                                             int *valp,
2694                                             int write, void *data)
2695 {
2696         if (write) {
2697                 *valp = msecs_to_jiffies(*negp ? -*lvalp : *lvalp);
2698         } else {
2699                 int val = *valp;
2700                 unsigned long lval;
2701                 if (val < 0) {
2702                         *negp = -1;
2703                         lval = (unsigned long)-val;
2704                 } else {
2705                         *negp = 0;
2706                         lval = (unsigned long)val;
2707                 }
2708                 *lvalp = jiffies_to_msecs(lval);
2709         }
2710         return 0;
2711 }
2712
2713 /**
2714  * proc_dointvec_jiffies - read a vector of integers as seconds
2715  * @table: the sysctl table
2716  * @write: %TRUE if this is a write to the sysctl file
2717  * @filp: the file structure
2718  * @buffer: the user buffer
2719  * @lenp: the size of the user buffer
2720  * @ppos: file position
2721  *
2722  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2723  * values from/to the user buffer, treated as an ASCII string.
2724  * The values read are assumed to be in seconds, and are converted into
2725  * jiffies.
2726  *
2727  * Returns 0 on success.
2728  */
2729 int proc_dointvec_jiffies(struct ctl_table *table, int write, struct file *filp,
2730                           void __user *buffer, size_t *lenp, loff_t *ppos)
2731 {
2732     return do_proc_dointvec(table,write,filp,buffer,lenp,ppos,
2733                             do_proc_dointvec_jiffies_conv,NULL);
2734 }
2735
2736 /**
2737  * proc_dointvec_userhz_jiffies - read a vector of integers as 1/USER_HZ seconds
2738  * @table: the sysctl table
2739  * @write: %TRUE if this is a write to the sysctl file
2740  * @filp: the file structure
2741  * @buffer: the user buffer
2742  * @lenp: the size of the user buffer
2743  * @ppos: pointer to the file position
2744  *
2745  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2746  * values from/to the user buffer, treated as an ASCII string.
2747  * The values read are assumed to be in 1/USER_HZ seconds, and
2748  * are converted into jiffies.
2749  *
2750  * Returns 0 on success.
2751  */
2752 int proc_dointvec_userhz_jiffies(struct ctl_table *table, int write, struct file *filp,
2753                                  void __user *buffer, size_t *lenp, loff_t *ppos)
2754 {
2755     return do_proc_dointvec(table,write,filp,buffer,lenp,ppos,
2756                             do_proc_dointvec_userhz_jiffies_conv,NULL);
2757 }
2758
2759 /**
2760  * proc_dointvec_ms_jiffies - read a vector of integers as 1 milliseconds
2761  * @table: the sysctl table
2762  * @write: %TRUE if this is a write to the sysctl file
2763  * @filp: the file structure
2764  * @buffer: the user buffer
2765  * @lenp: the size of the user buffer
2766  * @ppos: file position
2767  * @ppos: the current position in the file
2768  *
2769  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2770  * values from/to the user buffer, treated as an ASCII string.
2771  * The values read are assumed to be in 1/1000 seconds, and
2772  * are converted into jiffies.
2773  *
2774  * Returns 0 on success.
2775  */
2776 int proc_dointvec_ms_jiffies(struct ctl_table *table, int write, struct file *filp,
2777                              void __user *buffer, size_t *lenp, loff_t *ppos)
2778 {
2779         return do_proc_dointvec(table, write, filp, buffer, lenp, ppos,
2780                                 do_proc_dointvec_ms_jiffies_conv, NULL);
2781 }
2782
2783 static int proc_do_cad_pid(struct ctl_table *table, int write, struct file *filp,
2784                            void __user *buffer, size_t *lenp, loff_t *ppos)
2785 {
2786         struct pid *new_pid;
2787         pid_t tmp;
2788         int r;
2789
2790         tmp = pid_vnr(cad_pid);
2791
2792         r = __do_proc_dointvec(&tmp, table, write, filp, buffer,
2793                                lenp, ppos, NULL, NULL);
2794         if (r || !write)
2795                 return r;
2796
2797         new_pid = find_get_pid(tmp);
2798         if (!new_pid)
2799                 return -ESRCH;
2800
2801         put_pid(xchg(&cad_pid, new_pid));
2802         return 0;
2803 }
2804
2805 #else /* CONFIG_PROC_FS */
2806
2807 int proc_dostring(struct ctl_table *table, int write, struct file *filp,
2808                   void __user *buffer, size_t *lenp, loff_t *ppos)
2809 {
2810         return -ENOSYS;
2811 }
2812
2813 int proc_dointvec(struct ctl_table *table, int write, struct file *filp,
2814                   void __user *buffer, size_t *lenp, loff_t *ppos)
2815 {
2816         return -ENOSYS;
2817 }
2818
2819 int proc_dointvec_minmax(struct ctl_table *table, int write, struct file *filp,
2820                     void __user *buffer, size_t *lenp, loff_t *ppos)
2821 {
2822         return -ENOSYS;
2823 }
2824
2825 int proc_dointvec_jiffies(struct ctl_table *table, int write, struct file *filp,
2826                     void __user *buffer, size_t *lenp, loff_t *ppos)
2827 {
2828         return -ENOSYS;
2829 }
2830
2831 int proc_dointvec_userhz_jiffies(struct ctl_table *table, int write, struct file *filp,
2832                     void __user *buffer, size_t *lenp, loff_t *ppos)
2833 {
2834         return -ENOSYS;
2835 }
2836
2837 int proc_dointvec_ms_jiffies(struct ctl_table *table, int write, struct file *filp,
2838                              void __user *buffer, size_t *lenp, loff_t *ppos)
2839 {
2840         return -ENOSYS;
2841 }
2842
2843 int proc_doulongvec_minmax(struct ctl_table *table, int write, struct file *filp,
2844                     void __user *buffer, size_t *lenp, loff_t *ppos)
2845 {
2846         return -ENOSYS;
2847 }
2848
2849 int proc_doulongvec_ms_jiffies_minmax(struct ctl_table *table, int write,
2850                                       struct file *filp,
2851                                       void __user *buffer,
2852                                       size_t *lenp, loff_t *ppos)
2853 {
2854     return -ENOSYS;
2855 }
2856
2857
2858 #endif /* CONFIG_PROC_FS */
2859
2860
2861 #ifdef CONFIG_SYSCTL_SYSCALL
2862 /*
2863  * General sysctl support routines
2864  */
2865
2866 /* The generic sysctl data routine (used if no strategy routine supplied) */
2867 int sysctl_data(struct ctl_table *table,
2868                 void __user *oldval, size_t __user *oldlenp,
2869                 void __user *newval, size_t newlen)
2870 {
2871         size_t len;
2872
2873         /* Get out of I don't have a variable */
2874         if (!table->data || !table->maxlen)
2875                 return -ENOTDIR;
2876
2877         if (oldval && oldlenp) {
2878                 if (get_user(len, oldlenp))
2879                         return -EFAULT;
2880                 if (len) {
2881                         if (len > table->maxlen)
2882                                 len = table->maxlen;
2883                         if (copy_to_user(oldval, table->data, len))
2884                                 return -EFAULT;
2885                         if (put_user(len, oldlenp))
2886                                 return -EFAULT;
2887                 }
2888         }
2889
2890         if (newval && newlen) {
2891                 if (newlen > table->maxlen)
2892                         newlen = table->maxlen;
2893
2894                 if (copy_from_user(table->data, newval, newlen))
2895                         return -EFAULT;
2896         }
2897         return 1;
2898 }
2899
2900 /* The generic string strategy routine: */
2901 int sysctl_string(struct ctl_table *table,
2902                   void __user *oldval, size_t __user *oldlenp,
2903                   void __user *newval, size_t newlen)
2904 {
2905         if (!table->data || !table->maxlen)
2906                 return -ENOTDIR;
2907
2908         if (oldval && oldlenp) {
2909                 size_t bufsize;
2910                 if (get_user(bufsize, oldlenp))
2911                         return -EFAULT;
2912                 if (bufsize) {
2913                         size_t len = strlen(table->data), copied;
2914
2915                         /* This shouldn't trigger for a well-formed sysctl */
2916                         if (len > table->maxlen)
2917                                 len = table->maxlen;
2918
2919                         /* Copy up to a max of bufsize-1 bytes of the string */
2920                         copied = (len >= bufsize) ? bufsize - 1 : len;
2921
2922                         if (copy_to_user(oldval, table->data, copied) ||
2923                             put_user(0, (char __user *)(oldval + copied)))
2924                                 return -EFAULT;
2925                         if (put_user(len, oldlenp))
2926                                 return -EFAULT;
2927                 }
2928         }
2929         if (newval && newlen) {
2930                 size_t len = newlen;
2931                 if (len > table->maxlen)
2932                         len = table->maxlen;
2933                 if(copy_from_user(table->data, newval, len))
2934                         return -EFAULT;
2935                 if (len == table->maxlen)
2936                         len--;
2937                 ((char *) table->data)[len] = 0;
2938         }
2939         return 1;
2940 }
2941
2942 /*
2943  * This function makes sure that all of the integers in the vector
2944  * are between the minimum and maximum values given in the arrays
2945  * table->extra1 and table->extra2, respectively.
2946  */
2947 int sysctl_intvec(struct ctl_table *table,
2948                 void __user *oldval, size_t __user *oldlenp,
2949                 void __user *newval, size_t newlen)
2950 {
2951
2952         if (newval && newlen) {
2953                 int __user *vec = (int __user *) newval;
2954                 int *min = (int *) table->extra1;
2955                 int *max = (int *) table->extra2;
2956                 size_t length;
2957                 int i;
2958
2959                 if (newlen % sizeof(int) != 0)
2960                         return -EINVAL;
2961
2962                 if (!table->extra1 && !table->extra2)
2963                         return 0;
2964
2965                 if (newlen > table->maxlen)
2966                         newlen = table->maxlen;
2967                 length = newlen / sizeof(int);
2968
2969                 for (i = 0; i < length; i++) {
2970                         int value;
2971                         if (get_user(value, vec + i))
2972                                 return -EFAULT;
2973                         if (min && value < min[i])
2974                                 return -EINVAL;
2975                         if (max && value > max[i])
2976                                 return -EINVAL;
2977                 }
2978         }
2979         return 0;
2980 }
2981
2982 /* Strategy function to convert jiffies to seconds */
2983 int sysctl_jiffies(struct ctl_table *table,
2984                 void __user *oldval, size_t __user *oldlenp,
2985                 void __user *newval, size_t newlen)
2986 {
2987         if (oldval && oldlenp) {
2988                 size_t olen;
2989
2990                 if (get_user(olen, oldlenp))
2991                         return -EFAULT;
2992                 if (olen) {
2993                         int val;
2994
2995                         if (olen < sizeof(int))
2996                                 return -EINVAL;
2997
2998                         val = *(int *)(table->data) / HZ;
2999                         if (put_user(val, (int __user *)oldval))
3000                                 return -EFAULT;
3001                         if (put_user(sizeof(int), oldlenp))
3002                                 return -EFAULT;
3003                 }
3004         }
3005         if (newval && newlen) {
3006                 int new;
3007                 if (newlen != sizeof(int))
3008                         return -EINVAL;
3009                 if (get_user(new, (int __user *)newval))
3010                         return -EFAULT;
3011                 *(int *)(table->data) = new*HZ;
3012         }
3013         return 1;
3014 }
3015
3016 /* Strategy function to convert jiffies to seconds */
3017 int sysctl_ms_jiffies(struct ctl_table *table,
3018                 void __user *oldval, size_t __user *oldlenp,
3019                 void __user *newval, size_t newlen)
3020 {
3021         if (oldval && oldlenp) {
3022                 size_t olen;
3023
3024                 if (get_user(olen, oldlenp))
3025                         return -EFAULT;
3026                 if (olen) {
3027                         int val;
3028
3029                         if (olen < sizeof(int))
3030                                 return -EINVAL;
3031
3032                         val = jiffies_to_msecs(*(int *)(table->data));
3033                         if (put_user(val, (int __user *)oldval))
3034                                 return -EFAULT;
3035                         if (put_user(sizeof(int), oldlenp))
3036                                 return -EFAULT;
3037                 }
3038         }
3039         if (newval && newlen) {
3040                 int new;
3041                 if (newlen != sizeof(int))
3042                         return -EINVAL;
3043                 if (get_user(new, (int __user *)newval))
3044                         return -EFAULT;
3045                 *(int *)(table->data) = msecs_to_jiffies(new);
3046         }
3047         return 1;
3048 }
3049
3050
3051
3052 #else /* CONFIG_SYSCTL_SYSCALL */
3053
3054
3055 SYSCALL_DEFINE1(sysctl, struct __sysctl_args __user *, args)
3056 {
3057         struct __sysctl_args tmp;
3058         int error;
3059
3060         if (copy_from_user(&tmp, args, sizeof(tmp)))
3061                 return -EFAULT;
3062
3063         error = deprecated_sysctl_warning(&tmp);
3064
3065         /* If no error reading the parameters then just -ENOSYS ... */
3066         if (!error)
3067                 error = -ENOSYS;
3068
3069         return error;
3070 }
3071
3072 int sysctl_data(struct ctl_table *table,
3073                   void __user *oldval, size_t __user *oldlenp,
3074                   void __user *newval, size_t newlen)
3075 {
3076         return -ENOSYS;
3077 }
3078
3079 int sysctl_string(struct ctl_table *table,
3080                   void __user *oldval, size_t __user *oldlenp,
3081                   void __user *newval, size_t newlen)
3082 {
3083         return -ENOSYS;
3084 }
3085
3086 int sysctl_intvec(struct ctl_table *table,
3087                 void __user *oldval, size_t __user *oldlenp,
3088                 void __user *newval, size_t newlen)
3089 {
3090         return -ENOSYS;
3091 }
3092
3093 int sysctl_jiffies(struct ctl_table *table,
3094                 void __user *oldval, size_t __user *oldlenp,
3095                 void __user *newval, size_t newlen)
3096 {
3097         return -ENOSYS;
3098 }
3099
3100 int sysctl_ms_jiffies(struct ctl_table *table,
3101                 void __user *oldval, size_t __user *oldlenp,
3102                 void __user *newval, size_t newlen)
3103 {
3104         return -ENOSYS;
3105 }
3106
3107 #endif /* CONFIG_SYSCTL_SYSCALL */
3108
3109 static int deprecated_sysctl_warning(struct __sysctl_args *args)
3110 {
3111         static int msg_count;
3112         int name[CTL_MAXNAME];
3113         int i;
3114
3115         /* Check args->nlen. */
3116         if (args->nlen < 0 || args->nlen > CTL_MAXNAME)
3117                 return -ENOTDIR;
3118
3119         /* Read in the sysctl name for better debug message logging */
3120         for (i = 0; i < args->nlen; i++)
3121                 if (get_user(name[i], args->name + i))
3122                         return -EFAULT;
3123
3124         /* Ignore accesses to kernel.version */
3125         if ((args->nlen == 2) && (name[0] == CTL_KERN) && (name[1] == KERN_VERSION))
3126                 return 0;
3127
3128         if (msg_count < 5) {
3129                 msg_count++;
3130                 printk(KERN_INFO
3131                         "warning: process `%s' used the deprecated sysctl "
3132                         "system call with ", current->comm);
3133                 for (i = 0; i < args->nlen; i++)
3134                         printk("%d.", name[i]);
3135                 printk("\n");
3136         }
3137         return 0;
3138 }
3139
3140 /*
3141  * No sense putting this after each symbol definition, twice,
3142  * exception granted :-)
3143  */
3144 EXPORT_SYMBOL(proc_dointvec);
3145 EXPORT_SYMBOL(proc_dointvec_jiffies);
3146 EXPORT_SYMBOL(proc_dointvec_minmax);
3147 EXPORT_SYMBOL(proc_dointvec_userhz_jiffies);
3148 EXPORT_SYMBOL(proc_dointvec_ms_jiffies);
3149 EXPORT_SYMBOL(proc_dostring);
3150 EXPORT_SYMBOL(proc_doulongvec_minmax);
3151 EXPORT_SYMBOL(proc_doulongvec_ms_jiffies_minmax);
3152 EXPORT_SYMBOL(register_sysctl_table);
3153 EXPORT_SYMBOL(register_sysctl_paths);
3154 EXPORT_SYMBOL(sysctl_intvec);
3155 EXPORT_SYMBOL(sysctl_jiffies);
3156 EXPORT_SYMBOL(sysctl_ms_jiffies);
3157 EXPORT_SYMBOL(sysctl_string);
3158 EXPORT_SYMBOL(sysctl_data);
3159 EXPORT_SYMBOL(unregister_sysctl_table);