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

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