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       = "bootloader_version",
 747                 .data           = &bootloader_version,
 748                 .maxlen         = sizeof (int),
 749                 .mode           = 0444,
 750                 .proc_handler   = &proc_dointvec,
 751         },
 752         {
 753                 .ctl_name       = CTL_UNNUMBERED,
 754                 .procname       = "kstack_depth_to_print",
 755                 .data           = &kstack_depth_to_print,
 756                 .maxlen         = sizeof(int),
 757                 .mode           = 0644,
 758                 .proc_handler   = &proc_dointvec,
 759         },
 760         {
 761                 .ctl_name       = CTL_UNNUMBERED,
 762                 .procname       = "io_delay_type",
 763                 .data           = &io_delay_type,
 764                 .maxlen         = sizeof(int),
 765                 .mode           = 0644,
 766                 .proc_handler   = &proc_dointvec,
 767         },
 768 #endif
 769 #if defined(CONFIG_MMU)
 770         {
 771                 .ctl_name       = KERN_RANDOMIZE,
 772                 .procname       = "randomize_va_space",
 773                 .data           = &randomize_va_space,
 774                 .maxlen         = sizeof(int),
 775                 .mode           = 0644,
 776                 .proc_handler   = &proc_dointvec,
 777         },
 778 #endif
 779 #if defined(CONFIG_S390) && defined(CONFIG_SMP)
 780         {
 781                 .ctl_name       = KERN_SPIN_RETRY,
 782                 .procname       = "spin_retry",
 783                 .data           = &spin_retry,
 784                 .maxlen         = sizeof (int),
 785                 .mode           = 0644,
 786                 .proc_handler   = &proc_dointvec,
 787         },
 788 #endif
 789 #if     defined(CONFIG_ACPI_SLEEP) && defined(CONFIG_X86)
 790         {
 791                 .procname       = "acpi_video_flags",
 792                 .data           = &acpi_realmode_flags,
 793                 .maxlen         = sizeof (unsigned long),
 794                 .mode           = 0644,
 795                 .proc_handler   = &proc_doulongvec_minmax,
 796         },
 797 #endif
 798 #ifdef CONFIG_IA64
 799         {
 800                 .ctl_name       = KERN_IA64_UNALIGNED,
 801                 .procname       = "ignore-unaligned-usertrap",
 802                 .data           = &no_unaligned_warning,
 803                 .maxlen         = sizeof (int),
 804                 .mode           = 0644,
 805                 .proc_handler   = &proc_dointvec,
 806         },
 807         {
 808                 .ctl_name       = CTL_UNNUMBERED,
 809                 .procname       = "unaligned-dump-stack",
 810                 .data           = &unaligned_dump_stack,
 811                 .maxlen         = sizeof (int),
 812                 .mode           = 0644,
 813                 .proc_handler   = &proc_dointvec,
 814         },
 815 #endif
 816 #ifdef CONFIG_DETECT_SOFTLOCKUP
 817         {
 818                 .ctl_name       = CTL_UNNUMBERED,
 819                 .procname       = "softlockup_panic",
 820                 .data           = &softlockup_panic,
 821                 .maxlen         = sizeof(int),
 822                 .mode           = 0644,
 823                 .proc_handler   = &proc_dointvec_minmax,
 824                 .strategy       = &sysctl_intvec,
 825                 .extra1         = &zero,
 826                 .extra2         = &one,
 827         },
 828         {
 829                 .ctl_name       = CTL_UNNUMBERED,
 830                 .procname       = "softlockup_thresh",
 831                 .data           = &softlockup_thresh,
 832                 .maxlen         = sizeof(int),
 833                 .mode           = 0644,
 834                 .proc_handler   = &proc_dosoftlockup_thresh,
 835                 .strategy       = &sysctl_intvec,
 836                 .extra1         = &neg_one,
 837                 .extra2         = &sixty,
 838         },
 839 #endif
 840 #ifdef CONFIG_DETECT_HUNG_TASK
 841         {
 842                 .ctl_name       = CTL_UNNUMBERED,
 843                 .procname       = "hung_task_panic",
 844                 .data           = &sysctl_hung_task_panic,
 845                 .maxlen         = sizeof(int),
 846                 .mode           = 0644,
 847                 .proc_handler   = &proc_dointvec_minmax,
 848                 .strategy       = &sysctl_intvec,
 849                 .extra1         = &zero,
 850                 .extra2         = &one,
 851         },
 852         {
 853                 .ctl_name       = CTL_UNNUMBERED,
 854                 .procname       = "hung_task_check_count",
 855                 .data           = &sysctl_hung_task_check_count,
 856                 .maxlen         = sizeof(unsigned long),
 857                 .mode           = 0644,
 858                 .proc_handler   = &proc_doulongvec_minmax,
 859                 .strategy       = &sysctl_intvec,
 860         },
 861         {
 862                 .ctl_name       = CTL_UNNUMBERED,
 863                 .procname       = "hung_task_timeout_secs",
 864                 .data           = &sysctl_hung_task_timeout_secs,
 865                 .maxlen         = sizeof(unsigned long),
 866                 .mode           = 0644,
 867                 .proc_handler   = &proc_dohung_task_timeout_secs,
 868                 .strategy       = &sysctl_intvec,
 869         },
 870         {
 871                 .ctl_name       = CTL_UNNUMBERED,
 872                 .procname       = "hung_task_warnings",
 873                 .data           = &sysctl_hung_task_warnings,
 874                 .maxlen         = sizeof(unsigned long),
 875                 .mode           = 0644,
 876                 .proc_handler   = &proc_doulongvec_minmax,
 877                 .strategy       = &sysctl_intvec,
 878         },
 879 #endif
 880 #ifdef CONFIG_COMPAT
 881         {
 882                 .ctl_name       = KERN_COMPAT_LOG,
 883                 .procname       = "compat-log",
 884                 .data           = &compat_log,
 885                 .maxlen         = sizeof (int),
 886                 .mode           = 0644,
 887                 .proc_handler   = &proc_dointvec,
 888         },
 889 #endif
 890 #ifdef CONFIG_RT_MUTEXES
 891         {
 892                 .ctl_name       = KERN_MAX_LOCK_DEPTH,
 893                 .procname       = "max_lock_depth",
 894                 .data           = &max_lock_depth,
 895                 .maxlen         = sizeof(int),
 896                 .mode           = 0644,
 897                 .proc_handler   = &proc_dointvec,
 898         },
 899 #endif
 900         {
 901                 .ctl_name       = CTL_UNNUMBERED,
 902                 .procname       = "poweroff_cmd",
 903                 .data           = &poweroff_cmd,
 904                 .maxlen         = POWEROFF_CMD_PATH_LEN,
 905                 .mode           = 0644,
 906                 .proc_handler   = &proc_dostring,
 907                 .strategy       = &sysctl_string,
 908         },
 909 #ifdef CONFIG_KEYS
 910         {
 911                 .ctl_name       = CTL_UNNUMBERED,
 912                 .procname       = "keys",
 913                 .mode           = 0555,
 914                 .child          = key_sysctls,
 915         },
 916 #endif
 917 #ifdef CONFIG_RCU_TORTURE_TEST
 918         {
 919                 .ctl_name       = CTL_UNNUMBERED,
 920                 .procname       = "rcutorture_runnable",
 921                 .data           = &rcutorture_runnable,
 922                 .maxlen         = sizeof(int),
 923                 .mode           = 0644,
 924                 .proc_handler   = &proc_dointvec,
 925         },
 926 #endif
 927 #ifdef CONFIG_SLOW_WORK
 928         {
 929                 .ctl_name       = CTL_UNNUMBERED,
 930                 .procname       = "slow-work",
 931                 .mode           = 0555,
 932                 .child          = slow_work_sysctls,
 933         },
 934 #endif
 935 /*
 936  * NOTE: do not add new entries to this table unless you have read
 937  * Documentation/sysctl/ctl_unnumbered.txt
 938  */
 939         { .ctl_name = 0 }
 940 };
 941
 942 static struct ctl_table vm_table[] = {
 943         {
 944                 .ctl_name       = VM_OVERCOMMIT_MEMORY,
 945                 .procname       = "overcommit_memory",
 946                 .data           = &sysctl_overcommit_memory,
 947                 .maxlen         = sizeof(sysctl_overcommit_memory),
 948                 .mode           = 0644,
 949                 .proc_handler   = &proc_dointvec,
 950         },
 951         {
 952                 .ctl_name       = VM_PANIC_ON_OOM,
 953                 .procname       = "panic_on_oom",
 954                 .data           = &sysctl_panic_on_oom,
 955                 .maxlen         = sizeof(sysctl_panic_on_oom),
 956                 .mode           = 0644,
 957                 .proc_handler   = &proc_dointvec,
 958         },
 959         {
 960                 .ctl_name       = CTL_UNNUMBERED,
 961                 .procname       = "oom_kill_allocating_task",
 962                 .data           = &sysctl_oom_kill_allocating_task,
 963                 .maxlen         = sizeof(sysctl_oom_kill_allocating_task),
 964                 .mode           = 0644,
 965                 .proc_handler   = &proc_dointvec,
 966         },
 967         {
 968                 .ctl_name       = CTL_UNNUMBERED,
 969                 .procname       = "oom_dump_tasks",
 970                 .data           = &sysctl_oom_dump_tasks,
 971                 .maxlen         = sizeof(sysctl_oom_dump_tasks),
 972                 .mode           = 0644,
 973                 .proc_handler   = &proc_dointvec,
 974         },
 975         {
 976                 .ctl_name       = VM_OVERCOMMIT_RATIO,
 977                 .procname       = "overcommit_ratio",
 978                 .data           = &sysctl_overcommit_ratio,
 979                 .maxlen         = sizeof(sysctl_overcommit_ratio),
 980                 .mode           = 0644,
 981                 .proc_handler   = &proc_dointvec,
 982         },
 983         {
 984                 .ctl_name       = VM_PAGE_CLUSTER,
 985                 .procname       = "page-cluster",
 986                 .data           = &page_cluster,
 987                 .maxlen         = sizeof(int),
 988                 .mode           = 0644,
 989                 .proc_handler   = &proc_dointvec,
 990         },
 991         {
 992                 .ctl_name       = VM_DIRTY_BACKGROUND,
 993                 .procname       = "dirty_background_ratio",
 994                 .data           = &dirty_background_ratio,
 995                 .maxlen         = sizeof(dirty_background_ratio),
 996                 .mode           = 0644,
 997                 .proc_handler   = &dirty_background_ratio_handler,
 998                 .strategy       = &sysctl_intvec,
 999                 .extra1         = &zero,
1000                 .extra2         = &one_hundred,
1001         },
1002         {
1003                 .ctl_name       = CTL_UNNUMBERED,
1004                 .procname       = "dirty_background_bytes",
1005                 .data           = &dirty_background_bytes,
1006                 .maxlen         = sizeof(dirty_background_bytes),
1007                 .mode           = 0644,
1008                 .proc_handler   = &dirty_background_bytes_handler,
1009                 .strategy       = &sysctl_intvec,
1010                 .extra1         = &one_ul,
1011         },
1012         {
1013                 .ctl_name       = VM_DIRTY_RATIO,
1014                 .procname       = "dirty_ratio",
1015                 .data           = &vm_dirty_ratio,
1016                 .maxlen         = sizeof(vm_dirty_ratio),
1017                 .mode           = 0644,
1018                 .proc_handler   = &dirty_ratio_handler,
1019                 .strategy       = &sysctl_intvec,
1020                 .extra1         = &zero,
1021                 .extra2         = &one_hundred,
1022         },
1023         {
1024                 .ctl_name       = CTL_UNNUMBERED,
1025                 .procname       = "dirty_bytes",
1026                 .data           = &vm_dirty_bytes,
1027                 .maxlen         = sizeof(vm_dirty_bytes),
1028                 .mode           = 0644,
1029                 .proc_handler   = &dirty_bytes_handler,
1030                 .strategy       = &sysctl_intvec,
1031                 .extra1         = &dirty_bytes_min,
1032         },
1033         {
1034                 .procname       = "dirty_writeback_centisecs",
1035                 .data           = &dirty_writeback_interval,
1036                 .maxlen         = sizeof(dirty_writeback_interval),
1037                 .mode           = 0644,
1038                 .proc_handler   = &dirty_writeback_centisecs_handler,
1039         },
1040         {
1041                 .procname       = "dirty_expire_centisecs",
1042                 .data           = &dirty_expire_interval,
1043                 .maxlen         = sizeof(dirty_expire_interval),
1044                 .mode           = 0644,
1045                 .proc_handler   = &proc_dointvec,
1046         },
1047         {
1048                 .ctl_name       = VM_NR_PDFLUSH_THREADS,
1049                 .procname       = "nr_pdflush_threads",
1050                 .data           = &nr_pdflush_threads,
1051                 .maxlen         = sizeof nr_pdflush_threads,
1052                 .mode           = 0444 /* read-only*/,
1053                 .proc_handler   = &proc_dointvec,
1054         },
1055         {
1056                 .ctl_name       = VM_SWAPPINESS,
1057                 .procname       = "swappiness",
1058                 .data           = &vm_swappiness,
1059                 .maxlen         = sizeof(vm_swappiness),
1060                 .mode           = 0644,
1061                 .proc_handler   = &proc_dointvec_minmax,
1062                 .strategy       = &sysctl_intvec,
1063                 .extra1         = &zero,
1064                 .extra2         = &one_hundred,
1065         },
1066 #ifdef CONFIG_HUGETLB_PAGE
1067          {
1068                 .procname       = "nr_hugepages",
1069                 .data           = NULL,
1070                 .maxlen         = sizeof(unsigned long),
1071                 .mode           = 0644,
1072                 .proc_handler   = &hugetlb_sysctl_handler,
1073                 .extra1         = (void *)&hugetlb_zero,
1074                 .extra2         = (void *)&hugetlb_infinity,
1075          },
1076          {
1077                 .ctl_name       = VM_HUGETLB_GROUP,
1078                 .procname       = "hugetlb_shm_group",
1079                 .data           = &sysctl_hugetlb_shm_group,
1080                 .maxlen         = sizeof(gid_t),
1081                 .mode           = 0644,
1082                 .proc_handler   = &proc_dointvec,
1083          },
1084          {
1085                 .ctl_name       = CTL_UNNUMBERED,
1086                 .procname       = "hugepages_treat_as_movable",
1087                 .data           = &hugepages_treat_as_movable,
1088                 .maxlen         = sizeof(int),
1089                 .mode           = 0644,
1090                 .proc_handler   = &hugetlb_treat_movable_handler,
1091         },
1092         {
1093                 .ctl_name       = CTL_UNNUMBERED,
1094                 .procname       = "nr_overcommit_hugepages",
1095                 .data           = NULL,
1096                 .maxlen         = sizeof(unsigned long),
1097                 .mode           = 0644,
1098                 .proc_handler   = &hugetlb_overcommit_handler,
1099                 .extra1         = (void *)&hugetlb_zero,
1100                 .extra2         = (void *)&hugetlb_infinity,
1101         },
1102 #endif
1103         {
1104                 .ctl_name       = VM_LOWMEM_RESERVE_RATIO,
1105                 .procname       = "lowmem_reserve_ratio",
1106                 .data           = &sysctl_lowmem_reserve_ratio,
1107                 .maxlen         = sizeof(sysctl_lowmem_reserve_ratio),
1108                 .mode           = 0644,
1109                 .proc_handler   = &lowmem_reserve_ratio_sysctl_handler,
1110                 .strategy       = &sysctl_intvec,
1111         },
1112         {
1113                 .ctl_name       = VM_DROP_PAGECACHE,
1114                 .procname       = "drop_caches",
1115                 .data           = &sysctl_drop_caches,
1116                 .maxlen         = sizeof(int),
1117                 .mode           = 0644,
1118                 .proc_handler   = drop_caches_sysctl_handler,
1119                 .strategy       = &sysctl_intvec,
1120         },
1121         {
1122                 .ctl_name       = VM_MIN_FREE_KBYTES,
1123                 .procname       = "min_free_kbytes",
1124                 .data           = &min_free_kbytes,
1125                 .maxlen         = sizeof(min_free_kbytes),
1126                 .mode           = 0644,
1127                 .proc_handler   = &min_free_kbytes_sysctl_handler,
1128                 .strategy       = &sysctl_intvec,
1129                 .extra1         = &zero,
1130         },
1131         {
1132                 .ctl_name       = VM_PERCPU_PAGELIST_FRACTION,
1133                 .procname       = "percpu_pagelist_fraction",
1134                 .data           = &percpu_pagelist_fraction,
1135                 .maxlen         = sizeof(percpu_pagelist_fraction),
1136                 .mode           = 0644,
1137                 .proc_handler   = &percpu_pagelist_fraction_sysctl_handler,
1138                 .strategy       = &sysctl_intvec,
1139                 .extra1         = &min_percpu_pagelist_fract,
1140         },
1141 #ifdef CONFIG_MMU
1142         {
1143                 .ctl_name       = VM_MAX_MAP_COUNT,
1144                 .procname       = "max_map_count",
1145                 .data           = &sysctl_max_map_count,
1146                 .maxlen         = sizeof(sysctl_max_map_count),
1147                 .mode           = 0644,
1148                 .proc_handler   = &proc_dointvec
1149         },
1150 #else
1151         {
1152                 .ctl_name       = CTL_UNNUMBERED,
1153                 .procname       = "nr_trim_pages",
1154                 .data           = &sysctl_nr_trim_pages,
1155                 .maxlen         = sizeof(sysctl_nr_trim_pages),
1156                 .mode           = 0644,
1157                 .proc_handler   = &proc_dointvec_minmax,
1158                 .strategy       = &sysctl_intvec,
1159                 .extra1         = &zero,
1160         },
1161 #endif
1162         {
1163                 .ctl_name       = VM_LAPTOP_MODE,
1164                 .procname       = "laptop_mode",
1165                 .data           = &laptop_mode,
1166                 .maxlen         = sizeof(laptop_mode),
1167                 .mode           = 0644,
1168                 .proc_handler   = &proc_dointvec_jiffies,
1169                 .strategy       = &sysctl_jiffies,
1170         },
1171         {
1172                 .ctl_name       = VM_BLOCK_DUMP,
1173                 .procname       = "block_dump",
1174                 .data           = &block_dump,
1175                 .maxlen         = sizeof(block_dump),
1176                 .mode           = 0644,
1177                 .proc_handler   = &proc_dointvec,
1178                 .strategy       = &sysctl_intvec,
1179                 .extra1         = &zero,
1180         },
1181         {
1182                 .ctl_name       = VM_VFS_CACHE_PRESSURE,
1183                 .procname       = "vfs_cache_pressure",
1184                 .data           = &sysctl_vfs_cache_pressure,
1185                 .maxlen         = sizeof(sysctl_vfs_cache_pressure),
1186                 .mode           = 0644,
1187                 .proc_handler   = &proc_dointvec,
1188                 .strategy       = &sysctl_intvec,
1189                 .extra1         = &zero,
1190         },
1191 #ifdef HAVE_ARCH_PICK_MMAP_LAYOUT
1192         {
1193                 .ctl_name       = VM_LEGACY_VA_LAYOUT,
1194                 .procname       = "legacy_va_layout",
1195                 .data           = &sysctl_legacy_va_layout,
1196                 .maxlen         = sizeof(sysctl_legacy_va_layout),
1197                 .mode           = 0644,
1198                 .proc_handler   = &proc_dointvec,
1199                 .strategy       = &sysctl_intvec,
1200                 .extra1         = &zero,
1201         },
1202 #endif
1203 #ifdef CONFIG_NUMA
1204         {
1205                 .ctl_name       = VM_ZONE_RECLAIM_MODE,
1206                 .procname       = "zone_reclaim_mode",
1207                 .data           = &zone_reclaim_mode,
1208                 .maxlen         = sizeof(zone_reclaim_mode),
1209                 .mode           = 0644,
1210                 .proc_handler   = &proc_dointvec,
1211                 .strategy       = &sysctl_intvec,
1212                 .extra1         = &zero,
1213         },
1214         {
1215                 .ctl_name       = VM_MIN_UNMAPPED,
1216                 .procname       = "min_unmapped_ratio",
1217                 .data           = &sysctl_min_unmapped_ratio,
1218                 .maxlen         = sizeof(sysctl_min_unmapped_ratio),
1219                 .mode           = 0644,
1220                 .proc_handler   = &sysctl_min_unmapped_ratio_sysctl_handler,
1221                 .strategy       = &sysctl_intvec,
1222                 .extra1         = &zero,
1223                 .extra2         = &one_hundred,
1224         },
1225         {
1226                 .ctl_name       = VM_MIN_SLAB,
1227                 .procname       = "min_slab_ratio",
1228                 .data           = &sysctl_min_slab_ratio,
1229                 .maxlen         = sizeof(sysctl_min_slab_ratio),
1230                 .mode           = 0644,
1231                 .proc_handler   = &sysctl_min_slab_ratio_sysctl_handler,
1232                 .strategy       = &sysctl_intvec,
1233                 .extra1         = &zero,
1234                 .extra2         = &one_hundred,
1235         },
1236 #endif
1237 #ifdef CONFIG_SMP
1238         {
1239                 .ctl_name       = CTL_UNNUMBERED,
1240                 .procname       = "stat_interval",
1241                 .data           = &sysctl_stat_interval,
1242                 .maxlen         = sizeof(sysctl_stat_interval),
1243                 .mode           = 0644,
1244                 .proc_handler   = &proc_dointvec_jiffies,
1245                 .strategy       = &sysctl_jiffies,
1246         },
1247 #endif
1248         {
1249                 .ctl_name       = CTL_UNNUMBERED,
1250                 .procname       = "mmap_min_addr",
1251                 .data           = &mmap_min_addr,
1252                 .maxlen         = sizeof(unsigned long),
1253                 .mode           = 0644,
1254                 .proc_handler   = &proc_doulongvec_minmax,
1255         },
1256 #ifdef CONFIG_NUMA
1257         {
1258                 .ctl_name       = CTL_UNNUMBERED,
1259                 .procname       = "numa_zonelist_order",
1260                 .data           = &numa_zonelist_order,
1261                 .maxlen         = NUMA_ZONELIST_ORDER_LEN,
1262                 .mode           = 0644,
1263                 .proc_handler   = &numa_zonelist_order_handler,
1264                 .strategy       = &sysctl_string,
1265         },
1266 #endif
1267 #if (defined(CONFIG_X86_32) && !defined(CONFIG_UML))|| \
1268    (defined(CONFIG_SUPERH) && defined(CONFIG_VSYSCALL))
1269         {
1270                 .ctl_name       = VM_VDSO_ENABLED,
1271                 .procname       = "vdso_enabled",
1272                 .data           = &vdso_enabled,
1273                 .maxlen         = sizeof(vdso_enabled),
1274                 .mode           = 0644,
1275                 .proc_handler   = &proc_dointvec,
1276                 .strategy       = &sysctl_intvec,
1277                 .extra1         = &zero,
1278         },
1279 #endif
1280 #ifdef CONFIG_HIGHMEM
1281         {
1282                 .ctl_name       = CTL_UNNUMBERED,
1283                 .procname       = "highmem_is_dirtyable",
1284                 .data           = &vm_highmem_is_dirtyable,
1285                 .maxlen         = sizeof(vm_highmem_is_dirtyable),
1286                 .mode           = 0644,
1287                 .proc_handler   = &proc_dointvec_minmax,
1288                 .strategy       = &sysctl_intvec,
1289                 .extra1         = &zero,
1290                 .extra2         = &one,
1291         },
1292 #endif
1293 #ifdef CONFIG_UNEVICTABLE_LRU
1294         {
1295                 .ctl_name       = CTL_UNNUMBERED,
1296                 .procname       = "scan_unevictable_pages",
1297                 .data           = &scan_unevictable_pages,
1298                 .maxlen         = sizeof(scan_unevictable_pages),
1299                 .mode           = 0644,
1300                 .proc_handler   = &scan_unevictable_handler,
1301         },
1302 #endif
1303 /*
1304  * NOTE: do not add new entries to this table unless you have read
1305  * Documentation/sysctl/ctl_unnumbered.txt
1306  */
1307         { .ctl_name = 0 }
1308 };
1309
1310 #if defined(CONFIG_BINFMT_MISC) || defined(CONFIG_BINFMT_MISC_MODULE)
1311 static struct ctl_table binfmt_misc_table[] = {
1312         { .ctl_name = 0 }
1313 };
1314 #endif
1315
1316 static struct ctl_table fs_table[] = {
1317         {
1318                 .ctl_name       = FS_NRINODE,
1319                 .procname       = "inode-nr",
1320                 .data           = &inodes_stat,
1321                 .maxlen         = 2*sizeof(int),
1322                 .mode           = 0444,
1323                 .proc_handler   = &proc_dointvec,
1324         },
1325         {
1326                 .ctl_name       = FS_STATINODE,
1327                 .procname       = "inode-state",
1328                 .data           = &inodes_stat,
1329                 .maxlen         = 7*sizeof(int),
1330                 .mode           = 0444,
1331                 .proc_handler   = &proc_dointvec,
1332         },
1333         {
1334                 .procname       = "file-nr",
1335                 .data           = &files_stat,
1336                 .maxlen         = 3*sizeof(int),
1337                 .mode           = 0444,
1338                 .proc_handler   = &proc_nr_files,
1339         },
1340         {
1341                 .ctl_name       = FS_MAXFILE,
1342                 .procname       = "file-max",
1343                 .data           = &files_stat.max_files,
1344                 .maxlen         = sizeof(int),
1345                 .mode           = 0644,
1346                 .proc_handler   = &proc_dointvec,
1347         },
1348         {
1349                 .ctl_name       = CTL_UNNUMBERED,
1350                 .procname       = "nr_open",
1351                 .data           = &sysctl_nr_open,
1352                 .maxlen         = sizeof(int),
1353                 .mode           = 0644,
1354                 .proc_handler   = &proc_dointvec_minmax,
1355                 .extra1         = &sysctl_nr_open_min,
1356                 .extra2         = &sysctl_nr_open_max,
1357         },
1358         {
1359                 .ctl_name       = FS_DENTRY,
1360                 .procname       = "dentry-state",
1361                 .data           = &dentry_stat,
1362                 .maxlen         = 6*sizeof(int),
1363                 .mode           = 0444,
1364                 .proc_handler   = &proc_dointvec,
1365         },
1366         {
1367                 .ctl_name       = FS_OVERFLOWUID,
1368                 .procname       = "overflowuid",
1369                 .data           = &fs_overflowuid,
1370                 .maxlen         = sizeof(int),
1371                 .mode           = 0644,
1372                 .proc_handler   = &proc_dointvec_minmax,
1373                 .strategy       = &sysctl_intvec,
1374                 .extra1         = &minolduid,
1375                 .extra2         = &maxolduid,
1376         },
1377         {
1378                 .ctl_name       = FS_OVERFLOWGID,
1379                 .procname       = "overflowgid",
1380                 .data           = &fs_overflowgid,
1381                 .maxlen         = sizeof(int),
1382                 .mode           = 0644,
1383                 .proc_handler   = &proc_dointvec_minmax,
1384                 .strategy       = &sysctl_intvec,
1385                 .extra1         = &minolduid,
1386                 .extra2         = &maxolduid,
1387         },
1388 #ifdef CONFIG_FILE_LOCKING
1389         {
1390                 .ctl_name       = FS_LEASES,
1391                 .procname       = "leases-enable",
1392                 .data           = &leases_enable,
1393                 .maxlen         = sizeof(int),
1394                 .mode           = 0644,
1395                 .proc_handler   = &proc_dointvec,
1396         },
1397 #endif
1398 #ifdef CONFIG_DNOTIFY
1399         {
1400                 .ctl_name       = FS_DIR_NOTIFY,
1401                 .procname       = "dir-notify-enable",
1402                 .data           = &dir_notify_enable,
1403                 .maxlen         = sizeof(int),
1404                 .mode           = 0644,
1405                 .proc_handler   = &proc_dointvec,
1406         },
1407 #endif
1408 #ifdef CONFIG_MMU
1409 #ifdef CONFIG_FILE_LOCKING
1410         {
1411                 .ctl_name       = FS_LEASE_TIME,
1412                 .procname       = "lease-break-time",
1413                 .data           = &lease_break_time,
1414                 .maxlen         = sizeof(int),
1415                 .mode           = 0644,
1416                 .proc_handler   = &proc_dointvec,
1417         },
1418 #endif
1419 #ifdef CONFIG_AIO
1420         {
1421                 .procname       = "aio-nr",
1422                 .data           = &aio_nr,
1423                 .maxlen         = sizeof(aio_nr),
1424                 .mode           = 0444,
1425                 .proc_handler   = &proc_doulongvec_minmax,
1426         },
1427         {
1428                 .procname       = "aio-max-nr",
1429                 .data           = &aio_max_nr,
1430                 .maxlen         = sizeof(aio_max_nr),
1431                 .mode           = 0644,
1432                 .proc_handler   = &proc_doulongvec_minmax,
1433         },
1434 #endif /* CONFIG_AIO */
1435 #ifdef CONFIG_INOTIFY_USER
1436         {
1437                 .ctl_name       = FS_INOTIFY,
1438                 .procname       = "inotify",
1439                 .mode           = 0555,
1440                 .child          = inotify_table,
1441         },
1442 #endif
1443 #ifdef CONFIG_EPOLL
1444         {
1445                 .procname       = "epoll",
1446                 .mode           = 0555,
1447                 .child          = epoll_table,
1448         },
1449 #endif
1450 #endif
1451         {
1452                 .ctl_name       = KERN_SETUID_DUMPABLE,
1453                 .procname       = "suid_dumpable",
1454                 .data           = &suid_dumpable,
1455                 .maxlen         = sizeof(int),
1456                 .mode           = 0644,
1457                 .proc_handler   = &proc_dointvec_minmax,
1458                 .strategy       = &sysctl_intvec,
1459                 .extra1         = &zero,
1460                 .extra2         = &two,
1461         },
1462 #if defined(CONFIG_BINFMT_MISC) || defined(CONFIG_BINFMT_MISC_MODULE)
1463         {
1464                 .ctl_name       = CTL_UNNUMBERED,
1465                 .procname       = "binfmt_misc",
1466                 .mode           = 0555,
1467                 .child          = binfmt_misc_table,
1468         },
1469 #endif
1470 /*
1471  * NOTE: do not add new entries to this table unless you have read
1472  * Documentation/sysctl/ctl_unnumbered.txt
1473  */
1474         { .ctl_name = 0 }
1475 };
1476
1477 static struct ctl_table debug_table[] = {
1478 #if defined(CONFIG_X86) || defined(CONFIG_PPC)
1479         {
1480                 .ctl_name       = CTL_UNNUMBERED,
1481                 .procname       = "exception-trace",
1482                 .data           = &show_unhandled_signals,
1483                 .maxlen         = sizeof(int),
1484                 .mode           = 0644,
1485                 .proc_handler   = proc_dointvec
1486         },
1487 #endif
1488         { .ctl_name = 0 }
1489 };
1490
1491 static struct ctl_table dev_table[] = {
1492         { .ctl_name = 0 }
1493 };
1494
1495 static DEFINE_SPINLOCK(sysctl_lock);
1496
1497 /* called under sysctl_lock */
1498 static int use_table(struct ctl_table_header *p)
1499 {
1500         if (unlikely(p->unregistering))
1501                 return 0;
1502         p->used++;
1503         return 1;
1504 }
1505
1506 /* called under sysctl_lock */
1507 static void unuse_table(struct ctl_table_header *p)
1508 {
1509         if (!--p->used)
1510                 if (unlikely(p->unregistering))
1511                         complete(p->unregistering);
1512 }
1513
1514 /* called under sysctl_lock, will reacquire if has to wait */
1515 static void start_unregistering(struct ctl_table_header *p)
1516 {
1517         /*
1518          * if p->used is 0, nobody will ever touch that entry again;
1519          * we'll eliminate all paths to it before dropping sysctl_lock
1520          */
1521         if (unlikely(p->used)) {
1522                 struct completion wait;
1523                 init_completion(&wait);
1524                 p->unregistering = &wait;
1525                 spin_unlock(&sysctl_lock);
1526                 wait_for_completion(&wait);
1527                 spin_lock(&sysctl_lock);
1528         } else {
1529                 /* anything non-NULL; we'll never dereference it */
1530                 p->unregistering = ERR_PTR(-EINVAL);
1531         }
1532         /*
1533          * do not remove from the list until nobody holds it; walking the
1534          * list in do_sysctl() relies on that.
1535          */
1536         list_del_init(&p->ctl_entry);
1537 }
1538
1539 void sysctl_head_get(struct ctl_table_header *head)
1540 {
1541         spin_lock(&sysctl_lock);
1542         head->count++;
1543         spin_unlock(&sysctl_lock);
1544 }
1545
1546 void sysctl_head_put(struct ctl_table_header *head)
1547 {
1548         spin_lock(&sysctl_lock);
1549         if (!--head->count)
1550                 kfree(head);
1551         spin_unlock(&sysctl_lock);
1552 }
1553
1554 struct ctl_table_header *sysctl_head_grab(struct ctl_table_header *head)
1555 {
1556         if (!head)
1557                 BUG();
1558         spin_lock(&sysctl_lock);
1559         if (!use_table(head))
1560                 head = ERR_PTR(-ENOENT);
1561         spin_unlock(&sysctl_lock);
1562         return head;
1563 }
1564
1565 void sysctl_head_finish(struct ctl_table_header *head)
1566 {
1567         if (!head)
1568                 return;
1569         spin_lock(&sysctl_lock);
1570         unuse_table(head);
1571         spin_unlock(&sysctl_lock);
1572 }
1573
1574 static struct ctl_table_set *
1575 lookup_header_set(struct ctl_table_root *root, struct nsproxy *namespaces)
1576 {
1577         struct ctl_table_set *set = &root->default_set;
1578         if (root->lookup)
1579                 set = root->lookup(root, namespaces);
1580         return set;
1581 }
1582
1583 static struct list_head *
1584 lookup_header_list(struct ctl_table_root *root, struct nsproxy *namespaces)
1585 {
1586         struct ctl_table_set *set = lookup_header_set(root, namespaces);
1587         return &set->list;
1588 }
1589
1590 struct ctl_table_header *__sysctl_head_next(struct nsproxy *namespaces,
1591                                             struct ctl_table_header *prev)
1592 {
1593         struct ctl_table_root *root;
1594         struct list_head *header_list;
1595         struct ctl_table_header *head;
1596         struct list_head *tmp;
1597
1598         spin_lock(&sysctl_lock);
1599         if (prev) {
1600                 head = prev;
1601                 tmp = &prev->ctl_entry;
1602                 unuse_table(prev);
1603                 goto next;
1604         }
1605         tmp = &root_table_header.ctl_entry;
1606         for (;;) {
1607                 head = list_entry(tmp, struct ctl_table_header, ctl_entry);
1608
1609                 if (!use_table(head))
1610                         goto next;
1611                 spin_unlock(&sysctl_lock);
1612                 return head;
1613         next:
1614                 root = head->root;
1615                 tmp = tmp->next;
1616                 header_list = lookup_header_list(root, namespaces);
1617                 if (tmp != header_list)
1618                         continue;
1619
1620                 do {
1621                         root = list_entry(root->root_list.next,
1622                                         struct ctl_table_root, root_list);
1623                         if (root == &sysctl_table_root)
1624                                 goto out;
1625                         header_list = lookup_header_list(root, namespaces);
1626                 } while (list_empty(header_list));
1627                 tmp = header_list->next;
1628         }
1629 out:
1630         spin_unlock(&sysctl_lock);
1631         return NULL;
1632 }
1633
1634 struct ctl_table_header *sysctl_head_next(struct ctl_table_header *prev)
1635 {
1636         return __sysctl_head_next(current->nsproxy, prev);
1637 }
1638
1639 void register_sysctl_root(struct ctl_table_root *root)
1640 {
1641         spin_lock(&sysctl_lock);
1642         list_add_tail(&root->root_list, &sysctl_table_root.root_list);
1643         spin_unlock(&sysctl_lock);
1644 }
1645
1646 #ifdef CONFIG_SYSCTL_SYSCALL
1647 /* Perform the actual read/write of a sysctl table entry. */
1648 static int do_sysctl_strategy(struct ctl_table_root *root,
1649                         struct ctl_table *table,
1650                         void __user *oldval, size_t __user *oldlenp,
1651                         void __user *newval, size_t newlen)
1652 {
1653         int op = 0, rc;
1654
1655         if (oldval)
1656                 op |= MAY_READ;
1657         if (newval)
1658                 op |= MAY_WRITE;
1659         if (sysctl_perm(root, table, op))
1660                 return -EPERM;
1661
1662         if (table->strategy) {
1663                 rc = table->strategy(table, oldval, oldlenp, newval, newlen);
1664                 if (rc < 0)
1665                         return rc;
1666                 if (rc > 0)
1667                         return 0;
1668         }
1669
1670         /* If there is no strategy routine, or if the strategy returns
1671          * zero, proceed with automatic r/w */
1672         if (table->data && table->maxlen) {
1673                 rc = sysctl_data(table, oldval, oldlenp, newval, newlen);
1674                 if (rc < 0)
1675                         return rc;
1676         }
1677         return 0;
1678 }
1679
1680 static int parse_table(int __user *name, int nlen,
1681                        void __user *oldval, size_t __user *oldlenp,
1682                        void __user *newval, size_t newlen,
1683                        struct ctl_table_root *root,
1684                        struct ctl_table *table)
1685 {
1686         int n;
1687 repeat:
1688         if (!nlen)
1689                 return -ENOTDIR;
1690         if (get_user(n, name))
1691                 return -EFAULT;
1692         for ( ; table->ctl_name || table->procname; table++) {
1693                 if (!table->ctl_name)
1694                         continue;
1695                 if (n == table->ctl_name) {
1696                         int error;
1697                         if (table->child) {
1698                                 if (sysctl_perm(root, table, MAY_EXEC))
1699                                         return -EPERM;
1700                                 name++;
1701                                 nlen--;
1702                                 table = table->child;
1703                                 goto repeat;
1704                         }
1705                         error = do_sysctl_strategy(root, table,
1706                                                    oldval, oldlenp,
1707                                                    newval, newlen);
1708                         return error;
1709                 }
1710         }
1711         return -ENOTDIR;
1712 }
1713
1714 int do_sysctl(int __user *name, int nlen, void __user *oldval, size_t __user *oldlenp,
1715                void __user *newval, size_t newlen)
1716 {
1717         struct ctl_table_header *head;
1718         int error = -ENOTDIR;
1719
1720         if (nlen <= 0 || nlen >= CTL_MAXNAME)
1721                 return -ENOTDIR;
1722         if (oldval) {
1723                 int old_len;
1724                 if (!oldlenp || get_user(old_len, oldlenp))
1725                         return -EFAULT;
1726         }
1727
1728         for (head = sysctl_head_next(NULL); head;
1729                         head = sysctl_head_next(head)) {
1730                 error = parse_table(name, nlen, oldval, oldlenp,
1731                                         newval, newlen,
1732                                         head->root, head->ctl_table);
1733                 if (error != -ENOTDIR) {
1734                         sysctl_head_finish(head);
1735                         break;
1736                 }
1737         }
1738         return error;
1739 }
1740
1741 SYSCALL_DEFINE1(sysctl, struct __sysctl_args __user *, args)
1742 {
1743         struct __sysctl_args tmp;
1744         int error;
1745
1746         if (copy_from_user(&tmp, args, sizeof(tmp)))
1747                 return -EFAULT;
1748
1749         error = deprecated_sysctl_warning(&tmp);
1750         if (error)
1751                 goto out;
1752
1753         lock_kernel();
1754         error = do_sysctl(tmp.name, tmp.nlen, tmp.oldval, tmp.oldlenp,
1755                           tmp.newval, tmp.newlen);
1756         unlock_kernel();
1757 out:
1758         return error;
1759 }
1760 #endif /* CONFIG_SYSCTL_SYSCALL */
1761
1762 /*
1763  * sysctl_perm does NOT grant the superuser all rights automatically, because
1764  * some sysctl variables are readonly even to root.
1765  */
1766
1767 static int test_perm(int mode, int op)
1768 {
1769         if (!current_euid())
1770                 mode >>= 6;
1771         else if (in_egroup_p(0))
1772                 mode >>= 3;
1773         if ((op & ~mode & (MAY_READ|MAY_WRITE|MAY_EXEC)) == 0)
1774                 return 0;
1775         return -EACCES;
1776 }
1777
1778 int sysctl_perm(struct ctl_table_root *root, struct ctl_table *table, int op)
1779 {
1780         int error;
1781         int mode;
1782
1783         error = security_sysctl(table, op & (MAY_READ | MAY_WRITE | MAY_EXEC));
1784         if (error)
1785                 return error;
1786
1787         if (root->permissions)
1788                 mode = root->permissions(root, current->nsproxy, table);
1789         else
1790                 mode = table->mode;
1791
1792         return test_perm(mode, op);
1793 }
1794
1795 static void sysctl_set_parent(struct ctl_table *parent, struct ctl_table *table)
1796 {
1797         for (; table->ctl_name || table->procname; table++) {
1798                 table->parent = parent;
1799                 if (table->child)
1800                         sysctl_set_parent(table, table->child);
1801         }
1802 }
1803
1804 static __init int sysctl_init(void)
1805 {
1806         sysctl_set_parent(NULL, root_table);
1807 #ifdef CONFIG_SYSCTL_SYSCALL_CHECK
1808         {
1809                 int err;
1810                 err = sysctl_check_table(current->nsproxy, root_table);
1811         }
1812 #endif
1813         return 0;
1814 }
1815
1816 core_initcall(sysctl_init);
1817
1818 static struct ctl_table *is_branch_in(struct ctl_table *branch,
1819                                       struct ctl_table *table)
1820 {
1821         struct ctl_table *p;
1822         const char *s = branch->procname;
1823
1824         /* branch should have named subdirectory as its first element */
1825         if (!s || !branch->child)
1826                 return NULL;
1827
1828         /* ... and nothing else */
1829         if (branch[1].procname || branch[1].ctl_name)
1830                 return NULL;
1831
1832         /* table should contain subdirectory with the same name */
1833         for (p = table; p->procname || p->ctl_name; p++) {
1834                 if (!p->child)
1835                         continue;
1836                 if (p->procname && strcmp(p->procname, s) == 0)
1837                         return p;
1838         }
1839         return NULL;
1840 }
1841
1842 /* see if attaching q to p would be an improvement */
1843 static void try_attach(struct ctl_table_header *p, struct ctl_table_header *q)
1844 {
1845         struct ctl_table *to = p->ctl_table, *by = q->ctl_table;
1846         struct ctl_table *next;
1847         int is_better = 0;
1848         int not_in_parent = !p->attached_by;
1849
1850         while ((next = is_branch_in(by, to)) != NULL) {
1851                 if (by == q->attached_by)
1852                         is_better = 1;
1853                 if (to == p->attached_by)
1854                         not_in_parent = 1;
1855                 by = by->child;
1856                 to = next->child;
1857         }
1858
1859         if (is_better && not_in_parent) {
1860                 q->attached_by = by;
1861                 q->attached_to = to;
1862                 q->parent = p;
1863         }
1864 }
1865
1866 /**
1867  * __register_sysctl_paths - register a sysctl hierarchy
1868  * @root: List of sysctl headers to register on
1869  * @namespaces: Data to compute which lists of sysctl entries are visible
1870  * @path: The path to the directory the sysctl table is in.
1871  * @table: the top-level table structure
1872  *
1873  * Register a sysctl table hierarchy. @table should be a filled in ctl_table
1874  * array. A completely 0 filled entry terminates the table.
1875  *
1876  * The members of the &struct ctl_table structure are used as follows:
1877  *
1878  * ctl_name - This is the numeric sysctl value used by sysctl(2). The number
1879  *            must be unique within that level of sysctl
1880  *
1881  * procname - the name of the sysctl file under /proc/sys. Set to %NULL to not
1882  *            enter a sysctl file
1883  *
1884  * data - a pointer to data for use by proc_handler
1885  *
1886  * maxlen - the maximum size in bytes of the data
1887  *
1888  * mode - the file permissions for the /proc/sys file, and for sysctl(2)
1889  *
1890  * child - a pointer to the child sysctl table if this entry is a directory, or
1891  *         %NULL.
1892  *
1893  * proc_handler - the text handler routine (described below)
1894  *
1895  * strategy - the strategy routine (described below)
1896  *
1897  * de - for internal use by the sysctl routines
1898  *
1899  * extra1, extra2 - extra pointers usable by the proc handler routines
1900  *
1901  * Leaf nodes in the sysctl tree will be represented by a single file
1902  * under /proc; non-leaf nodes will be represented by directories.
1903  *
1904  * sysctl(2) can automatically manage read and write requests through
1905  * the sysctl table.  The data and maxlen fields of the ctl_table
1906  * struct enable minimal validation of the values being written to be
1907  * performed, and the mode field allows minimal authentication.
1908  *
1909  * More sophisticated management can be enabled by the provision of a
1910  * strategy routine with the table entry.  This will be called before
1911  * any automatic read or write of the data is performed.
1912  *
1913  * The strategy routine may return
1914  *
1915  * < 0 - Error occurred (error is passed to user process)
1916  *
1917  * 0   - OK - proceed with automatic read or write.
1918  *
1919  * > 0 - OK - read or write has been done by the strategy routine, so
1920  *       return immediately.
1921  *
1922  * There must be a proc_handler routine for any terminal nodes
1923  * mirrored under /proc/sys (non-terminals are handled by a built-in
1924  * directory handler).  Several default handlers are available to
1925  * cover common cases -
1926  *
1927  * proc_dostring(), proc_dointvec(), proc_dointvec_jiffies(),
1928  * proc_dointvec_userhz_jiffies(), proc_dointvec_minmax(),
1929  * proc_doulongvec_ms_jiffies_minmax(), proc_doulongvec_minmax()
1930  *
1931  * It is the handler's job to read the input buffer from user memory
1932  * and process it. The handler should return 0 on success.
1933  *
1934  * This routine returns %NULL on a failure to register, and a pointer
1935  * to the table header on success.
1936  */
1937 struct ctl_table_header *__register_sysctl_paths(
1938         struct ctl_table_root *root,
1939         struct nsproxy *namespaces,
1940         const struct ctl_path *path, struct ctl_table *table)
1941 {
1942         struct ctl_table_header *header;
1943         struct ctl_table *new, **prevp;
1944         unsigned int n, npath;
1945         struct ctl_table_set *set;
1946
1947         /* Count the path components */
1948         for (npath = 0; path[npath].ctl_name || path[npath].procname; ++npath)
1949                 ;
1950
1951         /*
1952          * For each path component, allocate a 2-element ctl_table array.
1953          * The first array element will be filled with the sysctl entry
1954          * for this, the second will be the sentinel (ctl_name == 0).
1955          *
1956          * We allocate everything in one go so that we don't have to
1957          * worry about freeing additional memory in unregister_sysctl_table.
1958          */
1959         header = kzalloc(sizeof(struct ctl_table_header) +
1960                          (2 * npath * sizeof(struct ctl_table)), GFP_KERNEL);
1961         if (!header)
1962                 return NULL;
1963
1964         new = (struct ctl_table *) (header + 1);
1965
1966         /* Now connect the dots */
1967         prevp = &header->ctl_table;
1968         for (n = 0; n < npath; ++n, ++path) {
1969                 /* Copy the procname */
1970                 new->procname = path->procname;
1971                 new->ctl_name = path->ctl_name;
1972                 new->mode     = 0555;
1973
1974                 *prevp = new;
1975                 prevp = &new->child;
1976
1977                 new += 2;
1978         }
1979         *prevp = table;
1980         header->ctl_table_arg = table;
1981
1982         INIT_LIST_HEAD(&header->ctl_entry);
1983         header->used = 0;
1984         header->unregistering = NULL;
1985         header->root = root;
1986         sysctl_set_parent(NULL, header->ctl_table);
1987         header->count = 1;
1988 #ifdef CONFIG_SYSCTL_SYSCALL_CHECK
1989         if (sysctl_check_table(namespaces, header->ctl_table)) {
1990                 kfree(header);
1991                 return NULL;
1992         }
1993 #endif
1994         spin_lock(&sysctl_lock);
1995         header->set = lookup_header_set(root, namespaces);
1996         header->attached_by = header->ctl_table;
1997         header->attached_to = root_table;
1998         header->parent = &root_table_header;
1999         for (set = header->set; set; set = set->parent) {
2000                 struct ctl_table_header *p;
2001                 list_for_each_entry(p, &set->list, ctl_entry) {
2002                         if (p->unregistering)
2003                                 continue;
2004                         try_attach(p, header);
2005                 }
2006         }
2007         header->parent->count++;
2008         list_add_tail(&header->ctl_entry, &header->set->list);
2009         spin_unlock(&sysctl_lock);
2010
2011         return header;
2012 }
2013
2014 /**
2015  * register_sysctl_table_path - register a sysctl table hierarchy
2016  * @path: The path to the directory the sysctl table is in.
2017  * @table: the top-level table structure
2018  *
2019  * Register a sysctl table hierarchy. @table should be a filled in ctl_table
2020  * array. A completely 0 filled entry terminates the table.
2021  *
2022  * See __register_sysctl_paths for more details.
2023  */
2024 struct ctl_table_header *register_sysctl_paths(const struct ctl_path *path,
2025                                                 struct ctl_table *table)
2026 {
2027         return __register_sysctl_paths(&sysctl_table_root, current->nsproxy,
2028                                         path, table);
2029 }
2030
2031 /**
2032  * register_sysctl_table - register a sysctl table hierarchy
2033  * @table: the top-level table structure
2034  *
2035  * Register a sysctl table hierarchy. @table should be a filled in ctl_table
2036  * array. A completely 0 filled entry terminates the table.
2037  *
2038  * See register_sysctl_paths for more details.
2039  */
2040 struct ctl_table_header *register_sysctl_table(struct ctl_table *table)
2041 {
2042         static const struct ctl_path null_path[] = { {} };
2043
2044         return register_sysctl_paths(null_path, table);
2045 }
2046
2047 /**
2048  * unregister_sysctl_table - unregister a sysctl table hierarchy
2049  * @header: the header returned from register_sysctl_table
2050  *
2051  * Unregisters the sysctl table and all children. proc entries may not
2052  * actually be removed until they are no longer used by anyone.
2053  */
2054 void unregister_sysctl_table(struct ctl_table_header * header)
2055 {
2056         might_sleep();
2057
2058         if (header == NULL)
2059                 return;
2060
2061         spin_lock(&sysctl_lock);
2062         start_unregistering(header);
2063         if (!--header->parent->count) {
2064                 WARN_ON(1);
2065                 kfree(header->parent);
2066         }
2067         if (!--header->count)
2068                 kfree(header);
2069         spin_unlock(&sysctl_lock);
2070 }
2071
2072 int sysctl_is_seen(struct ctl_table_header *p)
2073 {
2074         struct ctl_table_set *set = p->set;
2075         int res;
2076         spin_lock(&sysctl_lock);
2077         if (p->unregistering)
2078                 res = 0;
2079         else if (!set->is_seen)
2080                 res = 1;
2081         else
2082                 res = set->is_seen(set);
2083         spin_unlock(&sysctl_lock);
2084         return res;
2085 }
2086
2087 void setup_sysctl_set(struct ctl_table_set *p,
2088         struct ctl_table_set *parent,
2089         int (*is_seen)(struct ctl_table_set *))
2090 {
2091         INIT_LIST_HEAD(&p->list);
2092         p->parent = parent ? parent : &sysctl_table_root.default_set;
2093         p->is_seen = is_seen;
2094 }
2095
2096 #else /* !CONFIG_SYSCTL */
2097 struct ctl_table_header *register_sysctl_table(struct ctl_table * table)
2098 {
2099         return NULL;
2100 }
2101
2102 struct ctl_table_header *register_sysctl_paths(const struct ctl_path *path,
2103                                                     struct ctl_table *table)
2104 {
2105         return NULL;
2106 }
2107
2108 void unregister_sysctl_table(struct ctl_table_header * table)
2109 {
2110 }
2111
2112 void setup_sysctl_set(struct ctl_table_set *p,
2113         struct ctl_table_set *parent,
2114         int (*is_seen)(struct ctl_table_set *))
2115 {
2116 }
2117
2118 void sysctl_head_put(struct ctl_table_header *head)
2119 {
2120 }
2121
2122 #endif /* CONFIG_SYSCTL */
2123
2124 /*
2125  * /proc/sys support
2126  */
2127
2128 #ifdef CONFIG_PROC_SYSCTL
2129
2130 static int _proc_do_string(void* data, int maxlen, int write,
2131                            struct file *filp, void __user *buffer,
2132                            size_t *lenp, loff_t *ppos)
2133 {
2134         size_t len;
2135         char __user *p;
2136         char c;
2137
2138         if (!data || !maxlen || !*lenp) {
2139                 *lenp = 0;
2140                 return 0;
2141         }
2142
2143         if (write) {
2144                 len = 0;
2145                 p = buffer;
2146                 while (len < *lenp) {
2147                         if (get_user(c, p++))
2148                                 return -EFAULT;
2149                         if (c == 0 || c == '\n')
2150                                 break;
2151                         len++;
2152                 }
2153                 if (len >= maxlen)
2154                         len = maxlen-1;
2155                 if(copy_from_user(data, buffer, len))
2156                         return -EFAULT;
2157                 ((char *) data)[len] = 0;
2158                 *ppos += *lenp;
2159         } else {
2160                 len = strlen(data);
2161                 if (len > maxlen)
2162                         len = maxlen;
2163
2164                 if (*ppos > len) {
2165                         *lenp = 0;
2166                         return 0;
2167                 }
2168
2169                 data += *ppos;
2170                 len  -= *ppos;
2171
2172                 if (len > *lenp)
2173                         len = *lenp;
2174                 if (len)
2175                         if(copy_to_user(buffer, data, len))
2176                                 return -EFAULT;
2177                 if (len < *lenp) {
2178                         if(put_user('\n', ((char __user *) buffer) + len))
2179                                 return -EFAULT;
2180                         len++;
2181                 }
2182                 *lenp = len;
2183                 *ppos += len;
2184         }
2185         return 0;
2186 }
2187
2188 /**
2189  * proc_dostring - read a string sysctl
2190  * @table: the sysctl table
2191  * @write: %TRUE if this is a write to the sysctl file
2192  * @filp: the file structure
2193  * @buffer: the user buffer
2194  * @lenp: the size of the user buffer
2195  * @ppos: file position
2196  *
2197  * Reads/writes a string from/to the user buffer. If the kernel
2198  * buffer provided is not large enough to hold the string, the
2199  * string is truncated. The copied string is %NULL-terminated.
2200  * If the string is being read by the user process, it is copied
2201  * and a newline '\n' is added. It is truncated if the buffer is
2202  * not large enough.
2203  *
2204  * Returns 0 on success.
2205  */
2206 int proc_dostring(struct ctl_table *table, int write, struct file *filp,
2207                   void __user *buffer, size_t *lenp, loff_t *ppos)
2208 {
2209         return _proc_do_string(table->data, table->maxlen, write, filp,
2210                                buffer, lenp, ppos);
2211 }
2212
2213
2214 static int do_proc_dointvec_conv(int *negp, unsigned long *lvalp,
2215                                  int *valp,
2216                                  int write, void *data)
2217 {
2218         if (write) {
2219                 *valp = *negp ? -*lvalp : *lvalp;
2220         } else {
2221                 int val = *valp;
2222                 if (val < 0) {
2223                         *negp = -1;
2224                         *lvalp = (unsigned long)-val;
2225                 } else {
2226                         *negp = 0;
2227                         *lvalp = (unsigned long)val;
2228                 }
2229         }
2230         return 0;
2231 }
2232
2233 static int __do_proc_dointvec(void *tbl_data, struct ctl_table *table,
2234                   int write, struct file *filp, void __user *buffer,
2235                   size_t *lenp, loff_t *ppos,
2236                   int (*conv)(int *negp, unsigned long *lvalp, int *valp,
2237                               int write, void *data),
2238                   void *data)
2239 {
2240 #define TMPBUFLEN 21
2241         int *i, vleft, first=1, neg, val;
2242         unsigned long lval;
2243         size_t left, len;
2244
2245         char buf[TMPBUFLEN], *p;
2246         char __user *s = buffer;
2247
2248         if (!tbl_data || !table->maxlen || !*lenp ||
2249             (*ppos && !write)) {
2250                 *lenp = 0;
2251                 return 0;
2252         }
2253
2254         i = (int *) tbl_data;
2255         vleft = table->maxlen / sizeof(*i);
2256         left = *lenp;
2257
2258         if (!conv)
2259                 conv = do_proc_dointvec_conv;
2260
2261         for (; left && vleft--; i++, first=0) {
2262                 if (write) {
2263                         while (left) {
2264                                 char c;
2265                                 if (get_user(c, s))
2266                                         return -EFAULT;
2267                                 if (!isspace(c))
2268                                         break;
2269                                 left--;
2270                                 s++;
2271                         }
2272                         if (!left)
2273                                 break;
2274                         neg = 0;
2275                         len = left;
2276                         if (len > sizeof(buf) - 1)
2277                                 len = sizeof(buf) - 1;
2278                         if (copy_from_user(buf, s, len))
2279                                 return -EFAULT;
2280                         buf[len] = 0;
2281                         p = buf;
2282                         if (*p == '-' && left > 1) {
2283                                 neg = 1;
2284                                 p++;
2285                         }
2286                         if (*p < '0' || *p > '9')
2287                                 break;
2288
2289                         lval = simple_strtoul(p, &p, 0);
2290
2291                         len = p-buf;
2292                         if ((len < left) && *p && !isspace(*p))
2293                                 break;
2294                         if (neg)
2295                                 val = -val;
2296                         s += len;
2297                         left -= len;
2298
2299                         if (conv(&neg, &lval, i, 1, data))
2300                                 break;
2301                 } else {
2302                         p = buf;
2303                         if (!first)
2304                                 *p++ = '\t';
2305
2306                         if (conv(&neg, &lval, i, 0, data))
2307                                 break;
2308
2309                         sprintf(p, "%s%lu", neg ? "-" : "", lval);
2310                         len = strlen(buf);
2311                         if (len > left)
2312                                 len = left;
2313                         if(copy_to_user(s, buf, len))
2314                                 return -EFAULT;
2315                         left -= len;
2316                         s += len;
2317                 }
2318         }
2319
2320         if (!write && !first && left) {
2321                 if(put_user('\n', s))
2322                         return -EFAULT;
2323                 left--, s++;
2324         }
2325         if (write) {
2326                 while (left) {
2327                         char c;
2328                         if (get_user(c, s++))
2329                                 return -EFAULT;
2330                         if (!isspace(c))
2331                                 break;
2332                         left--;
2333                 }
2334         }
2335         if (write && first)
2336                 return -EINVAL;
2337         *lenp -= left;
2338         *ppos += *lenp;
2339         return 0;
2340 #undef TMPBUFLEN
2341 }
2342
2343 static int do_proc_dointvec(struct ctl_table *table, int write, struct file *filp,
2344                   void __user *buffer, size_t *lenp, loff_t *ppos,
2345                   int (*conv)(int *negp, unsigned long *lvalp, int *valp,
2346                               int write, void *data),
2347                   void *data)
2348 {
2349         return __do_proc_dointvec(table->data, table, write, filp,
2350                         buffer, lenp, ppos, conv, data);
2351 }
2352
2353 /**
2354  * proc_dointvec - read a vector of integers
2355  * @table: the sysctl table
2356  * @write: %TRUE if this is a write to the sysctl file
2357  * @filp: the file structure
2358  * @buffer: the user buffer
2359  * @lenp: the size of the user buffer
2360  * @ppos: file position
2361  *
2362  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2363  * values from/to the user buffer, treated as an ASCII string.
2364  *
2365  * Returns 0 on success.
2366  */
2367 int proc_dointvec(struct ctl_table *table, int write, struct file *filp,
2368                      void __user *buffer, size_t *lenp, loff_t *ppos)
2369 {
2370     return do_proc_dointvec(table,write,filp,buffer,lenp,ppos,
2371                             NULL,NULL);
2372 }
2373
2374 /*
2375  * Taint values can only be increased
2376  * This means we can safely use a temporary.
2377  */
2378 static int proc_taint(struct ctl_table *table, int write, struct file *filp,
2379                                void __user *buffer, size_t *lenp, loff_t *ppos)
2380 {
2381         struct ctl_table t;
2382         unsigned long tmptaint = get_taint();
2383         int err;
2384
2385         if (write && !capable(CAP_SYS_ADMIN))
2386                 return -EPERM;
2387
2388         t = *table;
2389         t.data = &tmptaint;
2390         err = proc_doulongvec_minmax(&t, write, filp, buffer, lenp, ppos);
2391         if (err < 0)
2392                 return err;
2393
2394         if (write) {
2395                 /*
2396                  * Poor man's atomic or. Not worth adding a primitive
2397                  * to everyone's atomic.h for this
2398                  */
2399                 int i;
2400                 for (i = 0; i < BITS_PER_LONG && tmptaint >> i; i++) {
2401                         if ((tmptaint >> i) & 1)
2402                                 add_taint(i);
2403                 }
2404         }
2405
2406         return err;
2407 }
2408
2409 struct do_proc_dointvec_minmax_conv_param {
2410         int *min;
2411         int *max;
2412 };
2413
2414 static int do_proc_dointvec_minmax_conv(int *negp, unsigned long *lvalp,
2415                                         int *valp,
2416                                         int write, void *data)
2417 {
2418         struct do_proc_dointvec_minmax_conv_param *param = data;
2419         if (write) {
2420                 int val = *negp ? -*lvalp : *lvalp;
2421                 if ((param->min && *param->min > val) ||
2422                     (param->max && *param->max < val))
2423                         return -EINVAL;
2424                 *valp = val;
2425         } else {
2426                 int val = *valp;
2427                 if (val < 0) {
2428                         *negp = -1;
2429                         *lvalp = (unsigned long)-val;
2430                 } else {
2431                         *negp = 0;
2432                         *lvalp = (unsigned long)val;
2433                 }
2434         }
2435         return 0;
2436 }
2437
2438 /**
2439  * proc_dointvec_minmax - read a vector of integers with min/max values
2440  * @table: the sysctl table
2441  * @write: %TRUE if this is a write to the sysctl file
2442  * @filp: the file structure
2443  * @buffer: the user buffer
2444  * @lenp: the size of the user buffer
2445  * @ppos: file position
2446  *
2447  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2448  * values from/to the user buffer, treated as an ASCII string.
2449  *
2450  * This routine will ensure the values are within the range specified by
2451  * table->extra1 (min) and table->extra2 (max).
2452  *
2453  * Returns 0 on success.
2454  */
2455 int proc_dointvec_minmax(struct ctl_table *table, int write, struct file *filp,
2456                   void __user *buffer, size_t *lenp, loff_t *ppos)
2457 {
2458         struct do_proc_dointvec_minmax_conv_param param = {
2459                 .min = (int *) table->extra1,
2460                 .max = (int *) table->extra2,
2461         };
2462         return do_proc_dointvec(table, write, filp, buffer, lenp, ppos,
2463                                 do_proc_dointvec_minmax_conv, &param);
2464 }
2465
2466 static int __do_proc_doulongvec_minmax(void *data, struct ctl_table *table, int write,
2467                                      struct file *filp,
2468                                      void __user *buffer,
2469                                      size_t *lenp, loff_t *ppos,
2470                                      unsigned long convmul,
2471                                      unsigned long convdiv)
2472 {
2473 #define TMPBUFLEN 21
2474         unsigned long *i, *min, *max, val;
2475         int vleft, first=1, neg;
2476         size_t len, left;
2477         char buf[TMPBUFLEN], *p;
2478         char __user *s = buffer;
2479
2480         if (!data || !table->maxlen || !*lenp ||
2481             (*ppos && !write)) {
2482                 *lenp = 0;
2483                 return 0;
2484         }
2485
2486         i = (unsigned long *) data;
2487         min = (unsigned long *) table->extra1;
2488         max = (unsigned long *) table->extra2;
2489         vleft = table->maxlen / sizeof(unsigned long);
2490         left = *lenp;
2491
2492         for (; left && vleft--; i++, min++, max++, first=0) {
2493                 if (write) {
2494                         while (left) {
2495                                 char c;
2496                                 if (get_user(c, s))
2497                                         return -EFAULT;
2498                                 if (!isspace(c))
2499                                         break;
2500                                 left--;
2501                                 s++;
2502                         }
2503                         if (!left)
2504                                 break;
2505                         neg = 0;
2506                         len = left;
2507                         if (len > TMPBUFLEN-1)
2508                                 len = TMPBUFLEN-1;
2509                         if (copy_from_user(buf, s, len))
2510                                 return -EFAULT;
2511                         buf[len] = 0;
2512                         p = buf;
2513                         if (*p == '-' && left > 1) {
2514                                 neg = 1;
2515                                 p++;
2516                         }
2517                         if (*p < '0' || *p > '9')
2518                                 break;
2519                         val = simple_strtoul(p, &p, 0) * convmul / convdiv ;
2520                         len = p-buf;
2521                         if ((len < left) && *p && !isspace(*p))
2522                                 break;
2523                         if (neg)
2524                                 val = -val;
2525                         s += len;
2526                         left -= len;
2527
2528                         if(neg)
2529                                 continue;
2530                         if ((min && val < *min) || (max && val > *max))
2531                                 continue;
2532                         *i = val;
2533                 } else {
2534                         p = buf;
2535                         if (!first)
2536                                 *p++ = '\t';
2537                         sprintf(p, "%lu", convdiv * (*i) / convmul);
2538                         len = strlen(buf);
2539                         if (len > left)
2540                                 len = left;
2541                         if(copy_to_user(s, buf, len))
2542                                 return -EFAULT;
2543                         left -= len;
2544                         s += len;
2545                 }
2546         }
2547
2548         if (!write && !first && left) {
2549                 if(put_user('\n', s))
2550                         return -EFAULT;
2551                 left--, s++;
2552         }
2553         if (write) {
2554                 while (left) {
2555                         char c;
2556                         if (get_user(c, s++))
2557                                 return -EFAULT;
2558                         if (!isspace(c))
2559                                 break;
2560                         left--;
2561                 }
2562         }
2563         if (write && first)
2564                 return -EINVAL;
2565         *lenp -= left;
2566         *ppos += *lenp;
2567         return 0;
2568 #undef TMPBUFLEN
2569 }
2570
2571 static int do_proc_doulongvec_minmax(struct ctl_table *table, int write,
2572                                      struct file *filp,
2573                                      void __user *buffer,
2574                                      size_t *lenp, loff_t *ppos,
2575                                      unsigned long convmul,
2576                                      unsigned long convdiv)
2577 {
2578         return __do_proc_doulongvec_minmax(table->data, table, write,
2579                         filp, buffer, lenp, ppos, convmul, convdiv);
2580 }
2581
2582 /**
2583  * proc_doulongvec_minmax - read a vector of long integers with min/max values
2584  * @table: the sysctl table
2585  * @write: %TRUE if this is a write to the sysctl file
2586  * @filp: the file structure
2587  * @buffer: the user buffer
2588  * @lenp: the size of the user buffer
2589  * @ppos: file position
2590  *
2591  * Reads/writes up to table->maxlen/sizeof(unsigned long) unsigned long
2592  * values from/to the user buffer, treated as an ASCII string.
2593  *
2594  * This routine will ensure the values are within the range specified by
2595  * table->extra1 (min) and table->extra2 (max).
2596  *
2597  * Returns 0 on success.
2598  */
2599 int proc_doulongvec_minmax(struct ctl_table *table, int write, struct file *filp,
2600                            void __user *buffer, size_t *lenp, loff_t *ppos)
2601 {
2602     return do_proc_doulongvec_minmax(table, write, filp, buffer, lenp, ppos, 1l, 1l);
2603 }
2604
2605 /**
2606  * proc_doulongvec_ms_jiffies_minmax - read a vector of millisecond values with min/max values
2607  * @table: the sysctl table
2608  * @write: %TRUE if this is a write to the sysctl file
2609  * @filp: the file structure
2610  * @buffer: the user buffer
2611  * @lenp: the size of the user buffer
2612  * @ppos: file position
2613  *
2614  * Reads/writes up to table->maxlen/sizeof(unsigned long) unsigned long
2615  * values from/to the user buffer, treated as an ASCII string. The values
2616  * are treated as milliseconds, and converted to jiffies when they are stored.
2617  *
2618  * This routine will ensure the values are within the range specified by
2619  * table->extra1 (min) and table->extra2 (max).
2620  *
2621  * Returns 0 on success.
2622  */
2623 int proc_doulongvec_ms_jiffies_minmax(struct ctl_table *table, int write,
2624                                       struct file *filp,
2625                                       void __user *buffer,
2626                                       size_t *lenp, loff_t *ppos)
2627 {
2628     return do_proc_doulongvec_minmax(table, write, filp, buffer,
2629                                      lenp, ppos, HZ, 1000l);
2630 }
2631
2632
2633 static int do_proc_dointvec_jiffies_conv(int *negp, unsigned long *lvalp,
2634                                          int *valp,
2635                                          int write, void *data)
2636 {
2637         if (write) {
2638                 if (*lvalp > LONG_MAX / HZ)
2639                         return 1;
2640                 *valp = *negp ? -(*lvalp*HZ) : (*lvalp*HZ);
2641         } else {
2642                 int val = *valp;
2643                 unsigned long lval;
2644                 if (val < 0) {
2645                         *negp = -1;
2646                         lval = (unsigned long)-val;
2647                 } else {
2648                         *negp = 0;
2649                         lval = (unsigned long)val;
2650                 }
2651                 *lvalp = lval / HZ;
2652         }
2653         return 0;
2654 }
2655
2656 static int do_proc_dointvec_userhz_jiffies_conv(int *negp, unsigned long *lvalp,
2657                                                 int *valp,
2658                                                 int write, void *data)
2659 {
2660         if (write) {
2661                 if (USER_HZ < HZ && *lvalp > (LONG_MAX / HZ) * USER_HZ)
2662                         return 1;
2663                 *valp = clock_t_to_jiffies(*negp ? -*lvalp : *lvalp);
2664         } else {
2665                 int val = *valp;
2666                 unsigned long lval;
2667                 if (val < 0) {
2668                         *negp = -1;
2669                         lval = (unsigned long)-val;
2670                 } else {
2671                         *negp = 0;
2672                         lval = (unsigned long)val;
2673                 }
2674                 *lvalp = jiffies_to_clock_t(lval);
2675         }
2676         return 0;
2677 }
2678
2679 static int do_proc_dointvec_ms_jiffies_conv(int *negp, unsigned long *lvalp,
2680                                             int *valp,
2681                                             int write, void *data)
2682 {
2683         if (write) {
2684                 *valp = msecs_to_jiffies(*negp ? -*lvalp : *lvalp);
2685         } else {
2686                 int val = *valp;
2687                 unsigned long lval;
2688                 if (val < 0) {
2689                         *negp = -1;
2690                         lval = (unsigned long)-val;
2691                 } else {
2692                         *negp = 0;
2693                         lval = (unsigned long)val;
2694                 }
2695                 *lvalp = jiffies_to_msecs(lval);
2696         }
2697         return 0;
2698 }
2699
2700 /**
2701  * proc_dointvec_jiffies - read a vector of integers as seconds
2702  * @table: the sysctl table
2703  * @write: %TRUE if this is a write to the sysctl file
2704  * @filp: the file structure
2705  * @buffer: the user buffer
2706  * @lenp: the size of the user buffer
2707  * @ppos: file position
2708  *
2709  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2710  * values from/to the user buffer, treated as an ASCII string.
2711  * The values read are assumed to be in seconds, and are converted into
2712  * jiffies.
2713  *
2714  * Returns 0 on success.
2715  */
2716 int proc_dointvec_jiffies(struct ctl_table *table, int write, struct file *filp,
2717                           void __user *buffer, size_t *lenp, loff_t *ppos)
2718 {
2719     return do_proc_dointvec(table,write,filp,buffer,lenp,ppos,
2720                             do_proc_dointvec_jiffies_conv,NULL);
2721 }
2722
2723 /**
2724  * proc_dointvec_userhz_jiffies - read a vector of integers as 1/USER_HZ seconds
2725  * @table: the sysctl table
2726  * @write: %TRUE if this is a write to the sysctl file
2727  * @filp: the file structure
2728  * @buffer: the user buffer
2729  * @lenp: the size of the user buffer
2730  * @ppos: pointer to the file position
2731  *
2732  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2733  * values from/to the user buffer, treated as an ASCII string.
2734  * The values read are assumed to be in 1/USER_HZ seconds, and
2735  * are converted into jiffies.
2736  *
2737  * Returns 0 on success.
2738  */
2739 int proc_dointvec_userhz_jiffies(struct ctl_table *table, int write, struct file *filp,
2740                                  void __user *buffer, size_t *lenp, loff_t *ppos)
2741 {
2742     return do_proc_dointvec(table,write,filp,buffer,lenp,ppos,
2743                             do_proc_dointvec_userhz_jiffies_conv,NULL);
2744 }
2745
2746 /**
2747  * proc_dointvec_ms_jiffies - read a vector of integers as 1 milliseconds
2748  * @table: the sysctl table
2749  * @write: %TRUE if this is a write to the sysctl file
2750  * @filp: the file structure
2751  * @buffer: the user buffer
2752  * @lenp: the size of the user buffer
2753  * @ppos: file position
2754  * @ppos: the current position in the file
2755  *
2756  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2757  * values from/to the user buffer, treated as an ASCII string.
2758  * The values read are assumed to be in 1/1000 seconds, and
2759  * are converted into jiffies.
2760  *
2761  * Returns 0 on success.
2762  */
2763 int proc_dointvec_ms_jiffies(struct ctl_table *table, int write, struct file *filp,
2764                              void __user *buffer, size_t *lenp, loff_t *ppos)
2765 {
2766         return do_proc_dointvec(table, write, filp, buffer, lenp, ppos,
2767                                 do_proc_dointvec_ms_jiffies_conv, NULL);
2768 }
2769
2770 static int proc_do_cad_pid(struct ctl_table *table, int write, struct file *filp,
2771                            void __user *buffer, size_t *lenp, loff_t *ppos)
2772 {
2773         struct pid *new_pid;
2774         pid_t tmp;
2775         int r;
2776
2777         tmp = pid_vnr(cad_pid);
2778
2779         r = __do_proc_dointvec(&tmp, table, write, filp, buffer,
2780                                lenp, ppos, NULL, NULL);
2781         if (r || !write)
2782                 return r;
2783
2784         new_pid = find_get_pid(tmp);
2785         if (!new_pid)
2786                 return -ESRCH;
2787
2788         put_pid(xchg(&cad_pid, new_pid));
2789         return 0;
2790 }
2791
2792 #else /* CONFIG_PROC_FS */
2793
2794 int proc_dostring(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(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_minmax(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_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_userhz_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_dointvec_ms_jiffies(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_minmax(struct ctl_table *table, int write, struct file *filp,
2831                     void __user *buffer, size_t *lenp, loff_t *ppos)
2832 {
2833         return -ENOSYS;
2834 }
2835
2836 int proc_doulongvec_ms_jiffies_minmax(struct ctl_table *table, int write,
2837                                       struct file *filp,
2838                                       void __user *buffer,
2839                                       size_t *lenp, loff_t *ppos)
2840 {
2841     return -ENOSYS;
2842 }
2843
2844
2845 #endif /* CONFIG_PROC_FS */
2846
2847
2848 #ifdef CONFIG_SYSCTL_SYSCALL
2849 /*
2850  * General sysctl support routines
2851  */
2852
2853 /* The generic sysctl data routine (used if no strategy routine supplied) */
2854 int sysctl_data(struct ctl_table *table,
2855                 void __user *oldval, size_t __user *oldlenp,
2856                 void __user *newval, size_t newlen)
2857 {
2858         size_t len;
2859
2860         /* Get out of I don't have a variable */
2861         if (!table->data || !table->maxlen)
2862                 return -ENOTDIR;
2863
2864         if (oldval && oldlenp) {
2865                 if (get_user(len, oldlenp))
2866                         return -EFAULT;
2867                 if (len) {
2868                         if (len > table->maxlen)
2869                                 len = table->maxlen;
2870                         if (copy_to_user(oldval, table->data, len))
2871                                 return -EFAULT;
2872                         if (put_user(len, oldlenp))
2873                                 return -EFAULT;
2874                 }
2875         }
2876
2877         if (newval && newlen) {
2878                 if (newlen > table->maxlen)
2879                         newlen = table->maxlen;
2880
2881                 if (copy_from_user(table->data, newval, newlen))
2882                         return -EFAULT;
2883         }
2884         return 1;
2885 }
2886
2887 /* The generic string strategy routine: */
2888 int sysctl_string(struct ctl_table *table,
2889                   void __user *oldval, size_t __user *oldlenp,
2890                   void __user *newval, size_t newlen)
2891 {
2892         if (!table->data || !table->maxlen)
2893                 return -ENOTDIR;
2894
2895         if (oldval && oldlenp) {
2896                 size_t bufsize;
2897                 if (get_user(bufsize, oldlenp))
2898                         return -EFAULT;
2899                 if (bufsize) {
2900                         size_t len = strlen(table->data), copied;
2901
2902                         /* This shouldn't trigger for a well-formed sysctl */
2903                         if (len > table->maxlen)
2904                                 len = table->maxlen;
2905
2906                         /* Copy up to a max of bufsize-1 bytes of the string */
2907                         copied = (len >= bufsize) ? bufsize - 1 : len;
2908
2909                         if (copy_to_user(oldval, table->data, copied) ||
2910                             put_user(0, (char __user *)(oldval + copied)))
2911                                 return -EFAULT;
2912                         if (put_user(len, oldlenp))
2913                                 return -EFAULT;
2914                 }
2915         }
2916         if (newval && newlen) {
2917                 size_t len = newlen;
2918                 if (len > table->maxlen)
2919                         len = table->maxlen;
2920                 if(copy_from_user(table->data, newval, len))
2921                         return -EFAULT;
2922                 if (len == table->maxlen)
2923                         len--;
2924                 ((char *) table->data)[len] = 0;
2925         }
2926         return 1;
2927 }
2928
2929 /*
2930  * This function makes sure that all of the integers in the vector
2931  * are between the minimum and maximum values given in the arrays
2932  * table->extra1 and table->extra2, respectively.
2933  */
2934 int sysctl_intvec(struct ctl_table *table,
2935                 void __user *oldval, size_t __user *oldlenp,
2936                 void __user *newval, size_t newlen)
2937 {
2938
2939         if (newval && newlen) {
2940                 int __user *vec = (int __user *) newval;
2941                 int *min = (int *) table->extra1;
2942                 int *max = (int *) table->extra2;
2943                 size_t length;
2944                 int i;
2945
2946                 if (newlen % sizeof(int) != 0)
2947                         return -EINVAL;
2948
2949                 if (!table->extra1 && !table->extra2)
2950                         return 0;
2951
2952                 if (newlen > table->maxlen)
2953                         newlen = table->maxlen;
2954                 length = newlen / sizeof(int);
2955
2956                 for (i = 0; i < length; i++) {
2957                         int value;
2958                         if (get_user(value, vec + i))
2959                                 return -EFAULT;
2960                         if (min && value < min[i])
2961                                 return -EINVAL;
2962                         if (max && value > max[i])
2963                                 return -EINVAL;
2964                 }
2965         }
2966         return 0;
2967 }
2968
2969 /* Strategy function to convert jiffies to seconds */
2970 int sysctl_jiffies(struct ctl_table *table,
2971                 void __user *oldval, size_t __user *oldlenp,
2972                 void __user *newval, size_t newlen)
2973 {
2974         if (oldval && oldlenp) {
2975                 size_t olen;
2976
2977                 if (get_user(olen, oldlenp))
2978                         return -EFAULT;
2979                 if (olen) {
2980                         int val;
2981
2982                         if (olen < sizeof(int))
2983                                 return -EINVAL;
2984
2985                         val = *(int *)(table->data) / HZ;
2986                         if (put_user(val, (int __user *)oldval))
2987                                 return -EFAULT;
2988                         if (put_user(sizeof(int), oldlenp))
2989                                 return -EFAULT;
2990                 }
2991         }
2992         if (newval && newlen) {
2993                 int new;
2994                 if (newlen != sizeof(int))
2995                         return -EINVAL;
2996                 if (get_user(new, (int __user *)newval))
2997                         return -EFAULT;
2998                 *(int *)(table->data) = new*HZ;
2999         }
3000         return 1;
3001 }
3002
3003 /* Strategy function to convert jiffies to seconds */
3004 int sysctl_ms_jiffies(struct ctl_table *table,
3005                 void __user *oldval, size_t __user *oldlenp,
3006                 void __user *newval, size_t newlen)
3007 {
3008         if (oldval && oldlenp) {
3009                 size_t olen;
3010
3011                 if (get_user(olen, oldlenp))
3012                         return -EFAULT;
3013                 if (olen) {
3014                         int val;
3015
3016                         if (olen < sizeof(int))
3017                                 return -EINVAL;
3018
3019                         val = jiffies_to_msecs(*(int *)(table->data));
3020                         if (put_user(val, (int __user *)oldval))
3021                                 return -EFAULT;
3022                         if (put_user(sizeof(int), oldlenp))
3023                                 return -EFAULT;
3024                 }
3025         }
3026         if (newval && newlen) {
3027                 int new;
3028                 if (newlen != sizeof(int))
3029                         return -EINVAL;
3030                 if (get_user(new, (int __user *)newval))
3031                         return -EFAULT;
3032                 *(int *)(table->data) = msecs_to_jiffies(new);
3033         }
3034         return 1;
3035 }
3036
3037
3038
3039 #else /* CONFIG_SYSCTL_SYSCALL */
3040
3041
3042 SYSCALL_DEFINE1(sysctl, struct __sysctl_args __user *, args)
3043 {
3044         struct __sysctl_args tmp;
3045         int error;
3046
3047         if (copy_from_user(&tmp, args, sizeof(tmp)))
3048                 return -EFAULT;
3049
3050         error = deprecated_sysctl_warning(&tmp);
3051
3052         /* If no error reading the parameters then just -ENOSYS ... */
3053         if (!error)
3054                 error = -ENOSYS;
3055
3056         return error;
3057 }
3058
3059 int sysctl_data(struct ctl_table *table,
3060                   void __user *oldval, size_t __user *oldlenp,
3061                   void __user *newval, size_t newlen)
3062 {
3063         return -ENOSYS;
3064 }
3065
3066 int sysctl_string(struct ctl_table *table,
3067                   void __user *oldval, size_t __user *oldlenp,
3068                   void __user *newval, size_t newlen)
3069 {
3070         return -ENOSYS;
3071 }
3072
3073 int sysctl_intvec(struct ctl_table *table,
3074                 void __user *oldval, size_t __user *oldlenp,
3075                 void __user *newval, size_t newlen)
3076 {
3077         return -ENOSYS;
3078 }
3079
3080 int sysctl_jiffies(struct ctl_table *table,
3081                 void __user *oldval, size_t __user *oldlenp,
3082                 void __user *newval, size_t newlen)
3083 {
3084         return -ENOSYS;
3085 }
3086
3087 int sysctl_ms_jiffies(struct ctl_table *table,
3088                 void __user *oldval, size_t __user *oldlenp,
3089                 void __user *newval, size_t newlen)
3090 {
3091         return -ENOSYS;
3092 }
3093
3094 #endif /* CONFIG_SYSCTL_SYSCALL */
3095
3096 static int deprecated_sysctl_warning(struct __sysctl_args *args)
3097 {
3098         static int msg_count;
3099         int name[CTL_MAXNAME];
3100         int i;
3101
3102         /* Check args->nlen. */
3103         if (args->nlen < 0 || args->nlen > CTL_MAXNAME)
3104                 return -ENOTDIR;
3105
3106         /* Read in the sysctl name for better debug message logging */
3107         for (i = 0; i < args->nlen; i++)
3108                 if (get_user(name[i], args->name + i))
3109                         return -EFAULT;
3110
3111         /* Ignore accesses to kernel.version */
3112         if ((args->nlen == 2) && (name[0] == CTL_KERN) && (name[1] == KERN_VERSION))
3113                 return 0;
3114
3115         if (msg_count < 5) {
3116                 msg_count++;
3117                 printk(KERN_INFO
3118                         "warning: process `%s' used the deprecated sysctl "
3119                         "system call with ", current->comm);
3120                 for (i = 0; i < args->nlen; i++)
3121                         printk("%d.", name[i]);
3122                 printk("\n");
3123         }
3124         return 0;
3125 }
3126
3127 /*
3128  * No sense putting this after each symbol definition, twice,
3129  * exception granted :-)
3130  */
3131 EXPORT_SYMBOL(proc_dointvec);
3132 EXPORT_SYMBOL(proc_dointvec_jiffies);
3133 EXPORT_SYMBOL(proc_dointvec_minmax);
3134 EXPORT_SYMBOL(proc_dointvec_userhz_jiffies);
3135 EXPORT_SYMBOL(proc_dointvec_ms_jiffies);
3136 EXPORT_SYMBOL(proc_dostring);
3137 EXPORT_SYMBOL(proc_doulongvec_minmax);
3138 EXPORT_SYMBOL(proc_doulongvec_ms_jiffies_minmax);
3139 EXPORT_SYMBOL(register_sysctl_table);
3140 EXPORT_SYMBOL(register_sysctl_paths);
3141 EXPORT_SYMBOL(sysctl_intvec);
3142 EXPORT_SYMBOL(sysctl_jiffies);
3143 EXPORT_SYMBOL(sysctl_ms_jiffies);
3144 EXPORT_SYMBOL(sysctl_string);
3145 EXPORT_SYMBOL(sysctl_data);
3146 EXPORT_SYMBOL(unregister_sysctl_table);