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

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