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

   1 /*
   2  * sysctl.c: General linux system control interface
   3  *
   4  * Begun 24 March 1995, Stephen Tweedie
   5  * Added /proc support, Dec 1995
   6  * Added bdflush entry and intvec min/max checking, 2/23/96, Tom Dyas.
   7  * Added hooks for /proc/sys/net (minor, minor patch), 96/4/1, Mike Shaver.
   8  * Added kernel/java-{interpreter,appletviewer}, 96/5/10, Mike Shaver.
   9  * Dynamic registration fixes, Stephen Tweedie.
  10  * Added kswapd-interval, ctrl-alt-del, printk stuff, 1/8/97, Chris Horn.
  11  * Made sysctl support optional via CONFIG_SYSCTL, 1/10/97, Chris
  12  *  Horn.
  13  * Added proc_doulongvec_ms_jiffies_minmax, 09/08/99, Carlos H. Bauer.
  14  * Added proc_doulongvec_minmax, 09/08/99, Carlos H. Bauer.
  15  * Changed linked lists to use list.h instead of lists.h, 02/24/00, Bill
  16  *  Wendling.
  17  * The list_for_each() macro wasn't appropriate for the sysctl loop.
  18  *  Removed it and replaced it with older style, 03/23/00, Bill Wendling
  19  */
  20
  21 #include <linux/module.h>
  22 #include <linux/mm.h>
  23 #include <linux/swap.h>
  24 #include <linux/slab.h>
  25 #include <linux/sysctl.h>
  26 #include <linux/proc_fs.h>
  27 #include <linux/capability.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/security.h>
  41 #include <linux/initrd.h>
  42 #include <linux/times.h>
  43 #include <linux/limits.h>
  44 #include <linux/dcache.h>
  45 #include <linux/syscalls.h>
  46 #include <linux/nfs_fs.h>
  47 #include <linux/acpi.h>
  48 #include <linux/reboot.h>
  49
  50 #include <asm/uaccess.h>
  51 #include <asm/processor.h>
  52
  53 #ifdef CONFIG_X86
  54 #include <asm/nmi.h>
  55 #include <asm/stacktrace.h>
  56 #endif
  57
  58 #if defined(CONFIG_SYSCTL)
  59
  60 /* External variables not in a header file. */
  61 extern int C_A_D;
  62 extern int print_fatal_signals;
  63 extern int sysctl_overcommit_memory;
  64 extern int sysctl_overcommit_ratio;
  65 extern int sysctl_panic_on_oom;
  66 extern int sysctl_oom_kill_allocating_task;
  67 extern int max_threads;
  68 extern int core_uses_pid;
  69 extern int suid_dumpable;
  70 extern char core_pattern[];
  71 extern int pid_max;
  72 extern int min_free_kbytes;
  73 extern int printk_ratelimit_jiffies;
  74 extern int printk_ratelimit_burst;
  75 extern int pid_max_min, pid_max_max;
  76 extern int sysctl_drop_caches;
  77 extern int percpu_pagelist_fraction;
  78 extern int compat_log;
  79 extern int maps_protect;
  80 extern int sysctl_stat_interval;
  81 extern int audit_argv_kb;
  82
  83 /* Constants used for minimum and  maximum */
  84 #ifdef CONFIG_DETECT_SOFTLOCKUP
  85 static int one = 1;
  86 static int sixty = 60;
  87 #endif
  88
  89 #ifdef CONFIG_MMU
  90 static int two = 2;
  91 #endif
  92
  93 static int zero;
  94 static int one_hundred = 100;
  95
  96 /* this is needed for the proc_dointvec_minmax for [fs_]overflow UID and GID */
  97 static int maxolduid = 65535;
  98 static int minolduid;
  99 static int min_percpu_pagelist_fract = 8;
 100
 101 static int ngroups_max = NGROUPS_MAX;
 102
 103 #ifdef CONFIG_KMOD
 104 extern char modprobe_path[];
 105 #endif
 106 #ifdef CONFIG_CHR_DEV_SG
 107 extern int sg_big_buff;
 108 #endif
 109
 110 #ifdef __sparc__
 111 extern char reboot_command [];
 112 extern int stop_a_enabled;
 113 extern int scons_pwroff;
 114 #endif
 115
 116 #ifdef __hppa__
 117 extern int pwrsw_enabled;
 118 extern int unaligned_enabled;
 119 #endif
 120
 121 #ifdef CONFIG_S390
 122 #ifdef CONFIG_MATHEMU
 123 extern int sysctl_ieee_emulation_warnings;
 124 #endif
 125 extern int sysctl_userprocess_debug;
 126 extern int spin_retry;
 127 #endif
 128
 129 extern int sysctl_hz_timer;
 130
 131 #ifdef CONFIG_BSD_PROCESS_ACCT
 132 extern int acct_parm[];
 133 #endif
 134
 135 #ifdef CONFIG_IA64
 136 extern int no_unaligned_warning;
 137 #endif
 138
 139 #ifdef CONFIG_RT_MUTEXES
 140 extern int max_lock_depth;
 141 #endif
 142
 143 #ifdef CONFIG_SYSCTL_SYSCALL
 144 static int parse_table(int __user *, int, void __user *, size_t __user *,
 145                 void __user *, size_t, struct ctl_table *);
 146 #endif
 147
 148
 149 #ifdef CONFIG_PROC_SYSCTL
 150 static int proc_do_cad_pid(struct ctl_table *table, int write, struct file *filp,
 151                   void __user *buffer, size_t *lenp, loff_t *ppos);
 152 static int proc_dointvec_taint(struct ctl_table *table, int write, struct file *filp,
 153                                void __user *buffer, size_t *lenp, loff_t *ppos);
 154 #endif
 155
 156 static struct ctl_table root_table[];
 157 static struct ctl_table_header root_table_header =
 158         { root_table, LIST_HEAD_INIT(root_table_header.ctl_entry) };
 159
 160 static struct ctl_table kern_table[];
 161 static struct ctl_table vm_table[];
 162 static struct ctl_table fs_table[];
 163 static struct ctl_table debug_table[];
 164 static struct ctl_table dev_table[];
 165 extern struct ctl_table random_table[];
 166 #ifdef CONFIG_UNIX98_PTYS
 167 extern struct ctl_table pty_table[];
 168 #endif
 169 #ifdef CONFIG_INOTIFY_USER
 170 extern struct ctl_table inotify_table[];
 171 #endif
 172
 173 #ifdef HAVE_ARCH_PICK_MMAP_LAYOUT
 174 int sysctl_legacy_va_layout;
 175 #endif
 176
 177 extern int prove_locking;
 178 extern int lock_stat;
 179
 180 /* The default sysctl tables: */
 181
 182 static struct ctl_table root_table[] = {
 183         {
 184                 .ctl_name       = CTL_KERN,
 185                 .procname       = "kernel",
 186                 .mode           = 0555,
 187                 .child          = kern_table,
 188         },
 189         {
 190                 .ctl_name       = CTL_VM,
 191                 .procname       = "vm",
 192                 .mode           = 0555,
 193                 .child          = vm_table,
 194         },
 195 #ifdef CONFIG_NET
 196         {
 197                 .ctl_name       = CTL_NET,
 198                 .procname       = "net",
 199                 .mode           = 0555,
 200                 .child          = net_table,
 201         },
 202 #endif
 203         {
 204                 .ctl_name       = CTL_FS,
 205                 .procname       = "fs",
 206                 .mode           = 0555,
 207                 .child          = fs_table,
 208         },
 209         {
 210                 .ctl_name       = CTL_DEBUG,
 211                 .procname       = "debug",
 212                 .mode           = 0555,
 213                 .child          = debug_table,
 214         },
 215         {
 216                 .ctl_name       = CTL_DEV,
 217                 .procname       = "dev",
 218                 .mode           = 0555,
 219                 .child          = dev_table,
 220         },
 221 /*
 222  * NOTE: do not add new entries to this table unless you have read
 223  * Documentation/sysctl/ctl_unnumbered.txt
 224  */
 225         { .ctl_name = 0 }
 226 };
 227
 228 #ifdef CONFIG_SCHED_DEBUG
 229 static unsigned long min_sched_granularity_ns = 100000;         /* 100 usecs */
 230 static unsigned long max_sched_granularity_ns = 1000000000;     /* 1 second */
 231 static unsigned long min_wakeup_granularity_ns;                 /* 0 usecs */
 232 static unsigned long max_wakeup_granularity_ns = 1000000000;    /* 1 second */
 233 #endif
 234
 235 static struct ctl_table kern_table[] = {
 236 #ifdef CONFIG_SCHED_DEBUG
 237         {
 238                 .ctl_name       = CTL_UNNUMBERED,
 239                 .procname       = "sched_nr_latency",
 240                 .data           = &sysctl_sched_nr_latency,
 241                 .maxlen         = sizeof(unsigned int),
 242                 .mode           = 0644,
 243                 .proc_handler   = &proc_dointvec,
 244         },
 245         {
 246                 .ctl_name       = CTL_UNNUMBERED,
 247                 .procname       = "sched_latency_ns",
 248                 .data           = &sysctl_sched_latency,
 249                 .maxlen         = sizeof(unsigned int),
 250                 .mode           = 0644,
 251                 .proc_handler   = &proc_dointvec_minmax,
 252                 .strategy       = &sysctl_intvec,
 253                 .extra1         = &min_sched_granularity_ns,
 254                 .extra2         = &max_sched_granularity_ns,
 255         },
 256         {
 257                 .ctl_name       = CTL_UNNUMBERED,
 258                 .procname       = "sched_wakeup_granularity_ns",
 259                 .data           = &sysctl_sched_wakeup_granularity,
 260                 .maxlen         = sizeof(unsigned int),
 261                 .mode           = 0644,
 262                 .proc_handler   = &proc_dointvec_minmax,
 263                 .strategy       = &sysctl_intvec,
 264                 .extra1         = &min_wakeup_granularity_ns,
 265                 .extra2         = &max_wakeup_granularity_ns,
 266         },
 267         {
 268                 .ctl_name       = CTL_UNNUMBERED,
 269                 .procname       = "sched_batch_wakeup_granularity_ns",
 270                 .data           = &sysctl_sched_batch_wakeup_granularity,
 271                 .maxlen         = sizeof(unsigned int),
 272                 .mode           = 0644,
 273                 .proc_handler   = &proc_dointvec_minmax,
 274                 .strategy       = &sysctl_intvec,
 275                 .extra1         = &min_wakeup_granularity_ns,
 276                 .extra2         = &max_wakeup_granularity_ns,
 277         },
 278         {
 279                 .ctl_name       = CTL_UNNUMBERED,
 280                 .procname       = "sched_child_runs_first",
 281                 .data           = &sysctl_sched_child_runs_first,
 282                 .maxlen         = sizeof(unsigned int),
 283                 .mode           = 0644,
 284                 .proc_handler   = &proc_dointvec,
 285         },
 286         {
 287                 .ctl_name       = CTL_UNNUMBERED,
 288                 .procname       = "sched_features",
 289                 .data           = &sysctl_sched_features,
 290                 .maxlen         = sizeof(unsigned int),
 291                 .mode           = 0644,
 292                 .proc_handler   = &proc_dointvec,
 293         },
 294         {
 295                 .ctl_name       = CTL_UNNUMBERED,
 296                 .procname       = "sched_migration_cost",
 297                 .data           = &sysctl_sched_migration_cost,
 298                 .maxlen         = sizeof(unsigned int),
 299                 .mode           = 0644,
 300                 .proc_handler   = &proc_dointvec,
 301         },
 302 #endif
 303         {
 304                 .ctl_name       = CTL_UNNUMBERED,
 305                 .procname       = "sched_compat_yield",
 306                 .data           = &sysctl_sched_compat_yield,
 307                 .maxlen         = sizeof(unsigned int),
 308                 .mode           = 0644,
 309                 .proc_handler   = &proc_dointvec,
 310         },
 311 #ifdef CONFIG_PROVE_LOCKING
 312         {
 313                 .ctl_name       = CTL_UNNUMBERED,
 314                 .procname       = "prove_locking",
 315                 .data           = &prove_locking,
 316                 .maxlen         = sizeof(int),
 317                 .mode           = 0644,
 318                 .proc_handler   = &proc_dointvec,
 319         },
 320 #endif
 321 #ifdef CONFIG_LOCK_STAT
 322         {
 323                 .ctl_name       = CTL_UNNUMBERED,
 324                 .procname       = "lock_stat",
 325                 .data           = &lock_stat,
 326                 .maxlen         = sizeof(int),
 327                 .mode           = 0644,
 328                 .proc_handler   = &proc_dointvec,
 329         },
 330 #endif
 331         {
 332                 .ctl_name       = KERN_PANIC,
 333                 .procname       = "panic",
 334                 .data           = &panic_timeout,
 335                 .maxlen         = sizeof(int),
 336                 .mode           = 0644,
 337                 .proc_handler   = &proc_dointvec,
 338         },
 339         {
 340                 .ctl_name       = KERN_CORE_USES_PID,
 341                 .procname       = "core_uses_pid",
 342                 .data           = &core_uses_pid,
 343                 .maxlen         = sizeof(int),
 344                 .mode           = 0644,
 345                 .proc_handler   = &proc_dointvec,
 346         },
 347 #ifdef CONFIG_AUDITSYSCALL
 348         {
 349                 .ctl_name       = CTL_UNNUMBERED,
 350                 .procname       = "audit_argv_kb",
 351                 .data           = &audit_argv_kb,
 352                 .maxlen         = sizeof(int),
 353                 .mode           = 0644,
 354                 .proc_handler   = &proc_dointvec,
 355         },
 356 #endif
 357         {
 358                 .ctl_name       = KERN_CORE_PATTERN,
 359                 .procname       = "core_pattern",
 360                 .data           = core_pattern,
 361                 .maxlen         = CORENAME_MAX_SIZE,
 362                 .mode           = 0644,
 363                 .proc_handler   = &proc_dostring,
 364                 .strategy       = &sysctl_string,
 365         },
 366 #ifdef CONFIG_PROC_SYSCTL
 367         {
 368                 .procname       = "tainted",
 369                 .data           = &tainted,
 370                 .maxlen         = sizeof(int),
 371                 .mode           = 0644,
 372                 .proc_handler   = &proc_dointvec_taint,
 373         },
 374 #endif
 375         {
 376                 .procname       = "cap-bound",
 377                 .data           = &cap_bset,
 378                 .maxlen         = sizeof(kernel_cap_t),
 379                 .mode           = 0600,
 380                 .proc_handler   = &proc_dointvec_bset,
 381         },
 382 #ifdef CONFIG_BLK_DEV_INITRD
 383         {
 384                 .ctl_name       = KERN_REALROOTDEV,
 385                 .procname       = "real-root-dev",
 386                 .data           = &real_root_dev,
 387                 .maxlen         = sizeof(int),
 388                 .mode           = 0644,
 389                 .proc_handler   = &proc_dointvec,
 390         },
 391 #endif
 392         {
 393                 .ctl_name       = CTL_UNNUMBERED,
 394                 .procname       = "print-fatal-signals",
 395                 .data           = &print_fatal_signals,
 396                 .maxlen         = sizeof(int),
 397                 .mode           = 0644,
 398                 .proc_handler   = &proc_dointvec,
 399         },
 400 #ifdef __sparc__
 401         {
 402                 .ctl_name       = KERN_SPARC_REBOOT,
 403                 .procname       = "reboot-cmd",
 404                 .data           = reboot_command,
 405                 .maxlen         = 256,
 406                 .mode           = 0644,
 407                 .proc_handler   = &proc_dostring,
 408                 .strategy       = &sysctl_string,
 409         },
 410         {
 411                 .ctl_name       = KERN_SPARC_STOP_A,
 412                 .procname       = "stop-a",
 413                 .data           = &stop_a_enabled,
 414                 .maxlen         = sizeof (int),
 415                 .mode           = 0644,
 416                 .proc_handler   = &proc_dointvec,
 417         },
 418         {
 419                 .ctl_name       = KERN_SPARC_SCONS_PWROFF,
 420                 .procname       = "scons-poweroff",
 421                 .data           = &scons_pwroff,
 422                 .maxlen         = sizeof (int),
 423                 .mode           = 0644,
 424                 .proc_handler   = &proc_dointvec,
 425         },
 426 #endif
 427 #ifdef __hppa__
 428         {
 429                 .ctl_name       = KERN_HPPA_PWRSW,
 430                 .procname       = "soft-power",
 431                 .data           = &pwrsw_enabled,
 432                 .maxlen         = sizeof (int),
 433                 .mode           = 0644,
 434                 .proc_handler   = &proc_dointvec,
 435         },
 436         {
 437                 .ctl_name       = KERN_HPPA_UNALIGNED,
 438                 .procname       = "unaligned-trap",
 439                 .data           = &unaligned_enabled,
 440                 .maxlen         = sizeof (int),
 441                 .mode           = 0644,
 442                 .proc_handler   = &proc_dointvec,
 443         },
 444 #endif
 445         {
 446                 .ctl_name       = KERN_CTLALTDEL,
 447                 .procname       = "ctrl-alt-del",
 448                 .data           = &C_A_D,
 449                 .maxlen         = sizeof(int),
 450                 .mode           = 0644,
 451                 .proc_handler   = &proc_dointvec,
 452         },
 453         {
 454                 .ctl_name       = KERN_PRINTK,
 455                 .procname       = "printk",
 456                 .data           = &console_loglevel,
 457                 .maxlen         = 4*sizeof(int),
 458                 .mode           = 0644,
 459                 .proc_handler   = &proc_dointvec,
 460         },
 461 #ifdef CONFIG_KMOD
 462         {
 463                 .ctl_name       = KERN_MODPROBE,
 464                 .procname       = "modprobe",
 465                 .data           = &modprobe_path,
 466                 .maxlen         = KMOD_PATH_LEN,
 467                 .mode           = 0644,
 468                 .proc_handler   = &proc_dostring,
 469                 .strategy       = &sysctl_string,
 470         },
 471 #endif
 472 #if defined(CONFIG_HOTPLUG) && defined(CONFIG_NET)
 473         {
 474                 .ctl_name       = KERN_HOTPLUG,
 475                 .procname       = "hotplug",
 476                 .data           = &uevent_helper,
 477                 .maxlen         = UEVENT_HELPER_PATH_LEN,
 478                 .mode           = 0644,
 479                 .proc_handler   = &proc_dostring,
 480                 .strategy       = &sysctl_string,
 481         },
 482 #endif
 483 #ifdef CONFIG_CHR_DEV_SG
 484         {
 485                 .ctl_name       = KERN_SG_BIG_BUFF,
 486                 .procname       = "sg-big-buff",
 487                 .data           = &sg_big_buff,
 488                 .maxlen         = sizeof (int),
 489                 .mode           = 0444,
 490                 .proc_handler   = &proc_dointvec,
 491         },
 492 #endif
 493 #ifdef CONFIG_BSD_PROCESS_ACCT
 494         {
 495                 .ctl_name       = KERN_ACCT,
 496                 .procname       = "acct",
 497                 .data           = &acct_parm,
 498                 .maxlen         = 3*sizeof(int),
 499                 .mode           = 0644,
 500                 .proc_handler   = &proc_dointvec,
 501         },
 502 #endif
 503 #ifdef CONFIG_MAGIC_SYSRQ
 504         {
 505                 .ctl_name       = KERN_SYSRQ,
 506                 .procname       = "sysrq",
 507                 .data           = &__sysrq_enabled,
 508                 .maxlen         = sizeof (int),
 509                 .mode           = 0644,
 510                 .proc_handler   = &proc_dointvec,
 511         },
 512 #endif
 513 #ifdef CONFIG_PROC_SYSCTL
 514         {
 515                 .ctl_name       = KERN_CADPID,
 516                 .procname       = "cad_pid",
 517                 .data           = NULL,
 518                 .maxlen         = sizeof (int),
 519                 .mode           = 0600,
 520                 .proc_handler   = &proc_do_cad_pid,
 521         },
 522 #endif
 523         {
 524                 .ctl_name       = KERN_MAX_THREADS,
 525                 .procname       = "threads-max",
 526                 .data           = &max_threads,
 527                 .maxlen         = sizeof(int),
 528                 .mode           = 0644,
 529                 .proc_handler   = &proc_dointvec,
 530         },
 531         {
 532                 .ctl_name       = KERN_RANDOM,
 533                 .procname       = "random",
 534                 .mode           = 0555,
 535                 .child          = random_table,
 536         },
 537 #ifdef CONFIG_UNIX98_PTYS
 538         {
 539                 .ctl_name       = KERN_PTY,
 540                 .procname       = "pty",
 541                 .mode           = 0555,
 542                 .child          = pty_table,
 543         },
 544 #endif
 545         {
 546                 .ctl_name       = KERN_OVERFLOWUID,
 547                 .procname       = "overflowuid",
 548                 .data           = &overflowuid,
 549                 .maxlen         = sizeof(int),
 550                 .mode           = 0644,
 551                 .proc_handler   = &proc_dointvec_minmax,
 552                 .strategy       = &sysctl_intvec,
 553                 .extra1         = &minolduid,
 554                 .extra2         = &maxolduid,
 555         },
 556         {
 557                 .ctl_name       = KERN_OVERFLOWGID,
 558                 .procname       = "overflowgid",
 559                 .data           = &overflowgid,
 560                 .maxlen         = sizeof(int),
 561                 .mode           = 0644,
 562                 .proc_handler   = &proc_dointvec_minmax,
 563                 .strategy       = &sysctl_intvec,
 564                 .extra1         = &minolduid,
 565                 .extra2         = &maxolduid,
 566         },
 567 #ifdef CONFIG_S390
 568 #ifdef CONFIG_MATHEMU
 569         {
 570                 .ctl_name       = KERN_IEEE_EMULATION_WARNINGS,
 571                 .procname       = "ieee_emulation_warnings",
 572                 .data           = &sysctl_ieee_emulation_warnings,
 573                 .maxlen         = sizeof(int),
 574                 .mode           = 0644,
 575                 .proc_handler   = &proc_dointvec,
 576         },
 577 #endif
 578 #ifdef CONFIG_NO_IDLE_HZ
 579         {
 580                 .ctl_name       = KERN_HZ_TIMER,
 581                 .procname       = "hz_timer",
 582                 .data           = &sysctl_hz_timer,
 583                 .maxlen         = sizeof(int),
 584                 .mode           = 0644,
 585                 .proc_handler   = &proc_dointvec,
 586         },
 587 #endif
 588         {
 589                 .ctl_name       = KERN_S390_USER_DEBUG_LOGGING,
 590                 .procname       = "userprocess_debug",
 591                 .data           = &sysctl_userprocess_debug,
 592                 .maxlen         = sizeof(int),
 593                 .mode           = 0644,
 594                 .proc_handler   = &proc_dointvec,
 595         },
 596 #endif
 597         {
 598                 .ctl_name       = KERN_PIDMAX,
 599                 .procname       = "pid_max",
 600                 .data           = &pid_max,
 601                 .maxlen         = sizeof (int),
 602                 .mode           = 0644,
 603                 .proc_handler   = &proc_dointvec_minmax,
 604                 .strategy       = sysctl_intvec,
 605                 .extra1         = &pid_max_min,
 606                 .extra2         = &pid_max_max,
 607         },
 608         {
 609                 .ctl_name       = KERN_PANIC_ON_OOPS,
 610                 .procname       = "panic_on_oops",
 611                 .data           = &panic_on_oops,
 612                 .maxlen         = sizeof(int),
 613                 .mode           = 0644,
 614                 .proc_handler   = &proc_dointvec,
 615         },
 616         {
 617                 .ctl_name       = KERN_PRINTK_RATELIMIT,
 618                 .procname       = "printk_ratelimit",
 619                 .data           = &printk_ratelimit_jiffies,
 620                 .maxlen         = sizeof(int),
 621                 .mode           = 0644,
 622                 .proc_handler   = &proc_dointvec_jiffies,
 623                 .strategy       = &sysctl_jiffies,
 624         },
 625         {
 626                 .ctl_name       = KERN_PRINTK_RATELIMIT_BURST,
 627                 .procname       = "printk_ratelimit_burst",
 628                 .data           = &printk_ratelimit_burst,
 629                 .maxlen         = sizeof(int),
 630                 .mode           = 0644,
 631                 .proc_handler   = &proc_dointvec,
 632         },
 633         {
 634                 .ctl_name       = KERN_NGROUPS_MAX,
 635                 .procname       = "ngroups_max",
 636                 .data           = &ngroups_max,
 637                 .maxlen         = sizeof (int),
 638                 .mode           = 0444,
 639                 .proc_handler   = &proc_dointvec,
 640         },
 641 #if defined(CONFIG_X86_LOCAL_APIC) && defined(CONFIG_X86)
 642         {
 643                 .ctl_name       = KERN_UNKNOWN_NMI_PANIC,
 644                 .procname       = "unknown_nmi_panic",
 645                 .data           = &unknown_nmi_panic,
 646                 .maxlen         = sizeof (int),
 647                 .mode           = 0644,
 648                 .proc_handler   = &proc_dointvec,
 649         },
 650         {
 651                 .procname       = "nmi_watchdog",
 652                 .data           = &nmi_watchdog_enabled,
 653                 .maxlen         = sizeof (int),
 654                 .mode           = 0644,
 655                 .proc_handler   = &proc_nmi_enabled,
 656         },
 657 #endif
 658 #if defined(CONFIG_X86)
 659         {
 660                 .ctl_name       = KERN_PANIC_ON_NMI,
 661                 .procname       = "panic_on_unrecovered_nmi",
 662                 .data           = &panic_on_unrecovered_nmi,
 663                 .maxlen         = sizeof(int),
 664                 .mode           = 0644,
 665                 .proc_handler   = &proc_dointvec,
 666         },
 667         {
 668                 .ctl_name       = KERN_BOOTLOADER_TYPE,
 669                 .procname       = "bootloader_type",
 670                 .data           = &bootloader_type,
 671                 .maxlen         = sizeof (int),
 672                 .mode           = 0444,
 673                 .proc_handler   = &proc_dointvec,
 674         },
 675         {
 676                 .ctl_name       = CTL_UNNUMBERED,
 677                 .procname       = "kstack_depth_to_print",
 678                 .data           = &kstack_depth_to_print,
 679                 .maxlen         = sizeof(int),
 680                 .mode           = 0644,
 681                 .proc_handler   = &proc_dointvec,
 682         },
 683 #endif
 684 #if defined(CONFIG_MMU)
 685         {
 686                 .ctl_name       = KERN_RANDOMIZE,
 687                 .procname       = "randomize_va_space",
 688                 .data           = &randomize_va_space,
 689                 .maxlen         = sizeof(int),
 690                 .mode           = 0644,
 691                 .proc_handler   = &proc_dointvec,
 692         },
 693 #endif
 694 #if defined(CONFIG_S390) && defined(CONFIG_SMP)
 695         {
 696                 .ctl_name       = KERN_SPIN_RETRY,
 697                 .procname       = "spin_retry",
 698                 .data           = &spin_retry,
 699                 .maxlen         = sizeof (int),
 700                 .mode           = 0644,
 701                 .proc_handler   = &proc_dointvec,
 702         },
 703 #endif
 704 #if     defined(CONFIG_ACPI_SLEEP) && defined(CONFIG_X86)
 705         {
 706                 .procname       = "acpi_video_flags",
 707                 .data           = &acpi_realmode_flags,
 708                 .maxlen         = sizeof (unsigned long),
 709                 .mode           = 0644,
 710                 .proc_handler   = &proc_doulongvec_minmax,
 711         },
 712 #endif
 713 #ifdef CONFIG_IA64
 714         {
 715                 .ctl_name       = KERN_IA64_UNALIGNED,
 716                 .procname       = "ignore-unaligned-usertrap",
 717                 .data           = &no_unaligned_warning,
 718                 .maxlen         = sizeof (int),
 719                 .mode           = 0644,
 720                 .proc_handler   = &proc_dointvec,
 721         },
 722 #endif
 723 #ifdef CONFIG_DETECT_SOFTLOCKUP
 724         {
 725                 .ctl_name       = CTL_UNNUMBERED,
 726                 .procname       = "softlockup_thresh",
 727                 .data           = &softlockup_thresh,
 728                 .maxlen         = sizeof(int),
 729                 .mode           = 0644,
 730                 .proc_handler   = &proc_dointvec_minmax,
 731                 .strategy       = &sysctl_intvec,
 732                 .extra1         = &one,
 733                 .extra2         = &sixty,
 734         },
 735 #endif
 736 #ifdef CONFIG_COMPAT
 737         {
 738                 .ctl_name       = KERN_COMPAT_LOG,
 739                 .procname       = "compat-log",
 740                 .data           = &compat_log,
 741                 .maxlen         = sizeof (int),
 742                 .mode           = 0644,
 743                 .proc_handler   = &proc_dointvec,
 744         },
 745 #endif
 746 #ifdef CONFIG_RT_MUTEXES
 747         {
 748                 .ctl_name       = KERN_MAX_LOCK_DEPTH,
 749                 .procname       = "max_lock_depth",
 750                 .data           = &max_lock_depth,
 751                 .maxlen         = sizeof(int),
 752                 .mode           = 0644,
 753                 .proc_handler   = &proc_dointvec,
 754         },
 755 #endif
 756 #ifdef CONFIG_PROC_FS
 757         {
 758                 .ctl_name       = CTL_UNNUMBERED,
 759                 .procname       = "maps_protect",
 760                 .data           = &maps_protect,
 761                 .maxlen         = sizeof(int),
 762                 .mode           = 0644,
 763                 .proc_handler   = &proc_dointvec,
 764         },
 765 #endif
 766         {
 767                 .ctl_name       = CTL_UNNUMBERED,
 768                 .procname       = "poweroff_cmd",
 769                 .data           = &poweroff_cmd,
 770                 .maxlen         = POWEROFF_CMD_PATH_LEN,
 771                 .mode           = 0644,
 772                 .proc_handler   = &proc_dostring,
 773                 .strategy       = &sysctl_string,
 774         },
 775 /*
 776  * NOTE: do not add new entries to this table unless you have read
 777  * Documentation/sysctl/ctl_unnumbered.txt
 778  */
 779         { .ctl_name = 0 }
 780 };
 781
 782 static struct ctl_table vm_table[] = {
 783         {
 784                 .ctl_name       = VM_OVERCOMMIT_MEMORY,
 785                 .procname       = "overcommit_memory",
 786                 .data           = &sysctl_overcommit_memory,
 787                 .maxlen         = sizeof(sysctl_overcommit_memory),
 788                 .mode           = 0644,
 789                 .proc_handler   = &proc_dointvec,
 790         },
 791         {
 792                 .ctl_name       = VM_PANIC_ON_OOM,
 793                 .procname       = "panic_on_oom",
 794                 .data           = &sysctl_panic_on_oom,
 795                 .maxlen         = sizeof(sysctl_panic_on_oom),
 796                 .mode           = 0644,
 797                 .proc_handler   = &proc_dointvec,
 798         },
 799         {
 800                 .ctl_name       = CTL_UNNUMBERED,
 801                 .procname       = "oom_kill_allocating_task",
 802                 .data           = &sysctl_oom_kill_allocating_task,
 803                 .maxlen         = sizeof(sysctl_oom_kill_allocating_task),
 804                 .mode           = 0644,
 805                 .proc_handler   = &proc_dointvec,
 806         },
 807         {
 808                 .ctl_name       = VM_OVERCOMMIT_RATIO,
 809                 .procname       = "overcommit_ratio",
 810                 .data           = &sysctl_overcommit_ratio,
 811                 .maxlen         = sizeof(sysctl_overcommit_ratio),
 812                 .mode           = 0644,
 813                 .proc_handler   = &proc_dointvec,
 814         },
 815         {
 816                 .ctl_name       = VM_PAGE_CLUSTER,
 817                 .procname       = "page-cluster",
 818                 .data           = &page_cluster,
 819                 .maxlen         = sizeof(int),
 820                 .mode           = 0644,
 821                 .proc_handler   = &proc_dointvec,
 822         },
 823         {
 824                 .ctl_name       = VM_DIRTY_BACKGROUND,
 825                 .procname       = "dirty_background_ratio",
 826                 .data           = &dirty_background_ratio,
 827                 .maxlen         = sizeof(dirty_background_ratio),
 828                 .mode           = 0644,
 829                 .proc_handler   = &proc_dointvec_minmax,
 830                 .strategy       = &sysctl_intvec,
 831                 .extra1         = &zero,
 832                 .extra2         = &one_hundred,
 833         },
 834         {
 835                 .ctl_name       = VM_DIRTY_RATIO,
 836                 .procname       = "dirty_ratio",
 837                 .data           = &vm_dirty_ratio,
 838                 .maxlen         = sizeof(vm_dirty_ratio),
 839                 .mode           = 0644,
 840                 .proc_handler   = &dirty_ratio_handler,
 841                 .strategy       = &sysctl_intvec,
 842                 .extra1         = &zero,
 843                 .extra2         = &one_hundred,
 844         },
 845         {
 846                 .procname       = "dirty_writeback_centisecs",
 847                 .data           = &dirty_writeback_interval,
 848                 .maxlen         = sizeof(dirty_writeback_interval),
 849                 .mode           = 0644,
 850                 .proc_handler   = &dirty_writeback_centisecs_handler,
 851         },
 852         {
 853                 .procname       = "dirty_expire_centisecs",
 854                 .data           = &dirty_expire_interval,
 855                 .maxlen         = sizeof(dirty_expire_interval),
 856                 .mode           = 0644,
 857                 .proc_handler   = &proc_dointvec_userhz_jiffies,
 858         },
 859         {
 860                 .ctl_name       = VM_NR_PDFLUSH_THREADS,
 861                 .procname       = "nr_pdflush_threads",
 862                 .data           = &nr_pdflush_threads,
 863                 .maxlen         = sizeof nr_pdflush_threads,
 864                 .mode           = 0444 /* read-only*/,
 865                 .proc_handler   = &proc_dointvec,
 866         },
 867         {
 868                 .ctl_name       = VM_SWAPPINESS,
 869                 .procname       = "swappiness",
 870                 .data           = &vm_swappiness,
 871                 .maxlen         = sizeof(vm_swappiness),
 872                 .mode           = 0644,
 873                 .proc_handler   = &proc_dointvec_minmax,
 874                 .strategy       = &sysctl_intvec,
 875                 .extra1         = &zero,
 876                 .extra2         = &one_hundred,
 877         },
 878 #ifdef CONFIG_HUGETLB_PAGE
 879          {
 880                 .procname       = "nr_hugepages",
 881                 .data           = &max_huge_pages,
 882                 .maxlen         = sizeof(unsigned long),
 883                 .mode           = 0644,
 884                 .proc_handler   = &hugetlb_sysctl_handler,
 885                 .extra1         = (void *)&hugetlb_zero,
 886                 .extra2         = (void *)&hugetlb_infinity,
 887          },
 888          {
 889                 .ctl_name       = VM_HUGETLB_GROUP,
 890                 .procname       = "hugetlb_shm_group",
 891                 .data           = &sysctl_hugetlb_shm_group,
 892                 .maxlen         = sizeof(gid_t),
 893                 .mode           = 0644,
 894                 .proc_handler   = &proc_dointvec,
 895          },
 896          {
 897                 .ctl_name       = CTL_UNNUMBERED,
 898                 .procname       = "hugepages_treat_as_movable",
 899                 .data           = &hugepages_treat_as_movable,
 900                 .maxlen         = sizeof(int),
 901                 .mode           = 0644,
 902                 .proc_handler   = &hugetlb_treat_movable_handler,
 903         },
 904         {
 905                 .ctl_name       = CTL_UNNUMBERED,
 906                 .procname       = "hugetlb_dynamic_pool",
 907                 .data           = &hugetlb_dynamic_pool,
 908                 .maxlen         = sizeof(hugetlb_dynamic_pool),
 909                 .mode           = 0644,
 910                 .proc_handler   = &proc_dointvec,
 911         },
 912 #endif
 913         {
 914                 .ctl_name       = VM_LOWMEM_RESERVE_RATIO,
 915                 .procname       = "lowmem_reserve_ratio",
 916                 .data           = &sysctl_lowmem_reserve_ratio,
 917                 .maxlen         = sizeof(sysctl_lowmem_reserve_ratio),
 918                 .mode           = 0644,
 919                 .proc_handler   = &lowmem_reserve_ratio_sysctl_handler,
 920                 .strategy       = &sysctl_intvec,
 921         },
 922         {
 923                 .ctl_name       = VM_DROP_PAGECACHE,
 924                 .procname       = "drop_caches",
 925                 .data           = &sysctl_drop_caches,
 926                 .maxlen         = sizeof(int),
 927                 .mode           = 0644,
 928                 .proc_handler   = drop_caches_sysctl_handler,
 929                 .strategy       = &sysctl_intvec,
 930         },
 931         {
 932                 .ctl_name       = VM_MIN_FREE_KBYTES,
 933                 .procname       = "min_free_kbytes",
 934                 .data           = &min_free_kbytes,
 935                 .maxlen         = sizeof(min_free_kbytes),
 936                 .mode           = 0644,
 937                 .proc_handler   = &min_free_kbytes_sysctl_handler,
 938                 .strategy       = &sysctl_intvec,
 939                 .extra1         = &zero,
 940         },
 941         {
 942                 .ctl_name       = VM_PERCPU_PAGELIST_FRACTION,
 943                 .procname       = "percpu_pagelist_fraction",
 944                 .data           = &percpu_pagelist_fraction,
 945                 .maxlen         = sizeof(percpu_pagelist_fraction),
 946                 .mode           = 0644,
 947                 .proc_handler   = &percpu_pagelist_fraction_sysctl_handler,
 948                 .strategy       = &sysctl_intvec,
 949                 .extra1         = &min_percpu_pagelist_fract,
 950         },
 951 #ifdef CONFIG_MMU
 952         {
 953                 .ctl_name       = VM_MAX_MAP_COUNT,
 954                 .procname       = "max_map_count",
 955                 .data           = &sysctl_max_map_count,
 956                 .maxlen         = sizeof(sysctl_max_map_count),
 957                 .mode           = 0644,
 958                 .proc_handler   = &proc_dointvec
 959         },
 960 #endif
 961         {
 962                 .ctl_name       = VM_LAPTOP_MODE,
 963                 .procname       = "laptop_mode",
 964                 .data           = &laptop_mode,
 965                 .maxlen         = sizeof(laptop_mode),
 966                 .mode           = 0644,
 967                 .proc_handler   = &proc_dointvec_jiffies,
 968                 .strategy       = &sysctl_jiffies,
 969         },
 970         {
 971                 .ctl_name       = VM_BLOCK_DUMP,
 972                 .procname       = "block_dump",
 973                 .data           = &block_dump,
 974                 .maxlen         = sizeof(block_dump),
 975                 .mode           = 0644,
 976                 .proc_handler   = &proc_dointvec,
 977                 .strategy       = &sysctl_intvec,
 978                 .extra1         = &zero,
 979         },
 980         {
 981                 .ctl_name       = VM_VFS_CACHE_PRESSURE,
 982                 .procname       = "vfs_cache_pressure",
 983                 .data           = &sysctl_vfs_cache_pressure,
 984                 .maxlen         = sizeof(sysctl_vfs_cache_pressure),
 985                 .mode           = 0644,
 986                 .proc_handler   = &proc_dointvec,
 987                 .strategy       = &sysctl_intvec,
 988                 .extra1         = &zero,
 989         },
 990 #ifdef HAVE_ARCH_PICK_MMAP_LAYOUT
 991         {
 992                 .ctl_name       = VM_LEGACY_VA_LAYOUT,
 993                 .procname       = "legacy_va_layout",
 994                 .data           = &sysctl_legacy_va_layout,
 995                 .maxlen         = sizeof(sysctl_legacy_va_layout),
 996                 .mode           = 0644,
 997                 .proc_handler   = &proc_dointvec,
 998                 .strategy       = &sysctl_intvec,
 999                 .extra1         = &zero,
1000         },
1001 #endif
1002 #ifdef CONFIG_NUMA
1003         {
1004                 .ctl_name       = VM_ZONE_RECLAIM_MODE,
1005                 .procname       = "zone_reclaim_mode",
1006                 .data           = &zone_reclaim_mode,
1007                 .maxlen         = sizeof(zone_reclaim_mode),
1008                 .mode           = 0644,
1009                 .proc_handler   = &proc_dointvec,
1010                 .strategy       = &sysctl_intvec,
1011                 .extra1         = &zero,
1012         },
1013         {
1014                 .ctl_name       = VM_MIN_UNMAPPED,
1015                 .procname       = "min_unmapped_ratio",
1016                 .data           = &sysctl_min_unmapped_ratio,
1017                 .maxlen         = sizeof(sysctl_min_unmapped_ratio),
1018                 .mode           = 0644,
1019                 .proc_handler   = &sysctl_min_unmapped_ratio_sysctl_handler,
1020                 .strategy       = &sysctl_intvec,
1021                 .extra1         = &zero,
1022                 .extra2         = &one_hundred,
1023         },
1024         {
1025                 .ctl_name       = VM_MIN_SLAB,
1026                 .procname       = "min_slab_ratio",
1027                 .data           = &sysctl_min_slab_ratio,
1028                 .maxlen         = sizeof(sysctl_min_slab_ratio),
1029                 .mode           = 0644,
1030                 .proc_handler   = &sysctl_min_slab_ratio_sysctl_handler,
1031                 .strategy       = &sysctl_intvec,
1032                 .extra1         = &zero,
1033                 .extra2         = &one_hundred,
1034         },
1035 #endif
1036 #ifdef CONFIG_SMP
1037         {
1038                 .ctl_name       = CTL_UNNUMBERED,
1039                 .procname       = "stat_interval",
1040                 .data           = &sysctl_stat_interval,
1041                 .maxlen         = sizeof(sysctl_stat_interval),
1042                 .mode           = 0644,
1043                 .proc_handler   = &proc_dointvec_jiffies,
1044                 .strategy       = &sysctl_jiffies,
1045         },
1046 #endif
1047 #ifdef CONFIG_SECURITY
1048         {
1049                 .ctl_name       = CTL_UNNUMBERED,
1050                 .procname       = "mmap_min_addr",
1051                 .data           = &mmap_min_addr,
1052                 .maxlen         = sizeof(unsigned long),
1053                 .mode           = 0644,
1054                 .proc_handler   = &proc_doulongvec_minmax,
1055         },
1056 #endif
1057 #ifdef CONFIG_NUMA
1058         {
1059                 .ctl_name       = CTL_UNNUMBERED,
1060                 .procname       = "numa_zonelist_order",
1061                 .data           = &numa_zonelist_order,
1062                 .maxlen         = NUMA_ZONELIST_ORDER_LEN,
1063                 .mode           = 0644,
1064                 .proc_handler   = &numa_zonelist_order_handler,
1065                 .strategy       = &sysctl_string,
1066         },
1067 #endif
1068 #if (defined(CONFIG_X86_32) && !defined(CONFIG_UML))|| \
1069    (defined(CONFIG_SUPERH) && defined(CONFIG_VSYSCALL))
1070         {
1071                 .ctl_name       = VM_VDSO_ENABLED,
1072                 .procname       = "vdso_enabled",
1073                 .data           = &vdso_enabled,
1074                 .maxlen         = sizeof(vdso_enabled),
1075                 .mode           = 0644,
1076                 .proc_handler   = &proc_dointvec,
1077                 .strategy       = &sysctl_intvec,
1078                 .extra1         = &zero,
1079         },
1080 #endif
1081 /*
1082  * NOTE: do not add new entries to this table unless you have read
1083  * Documentation/sysctl/ctl_unnumbered.txt
1084  */
1085         { .ctl_name = 0 }
1086 };
1087
1088 #if defined(CONFIG_BINFMT_MISC) || defined(CONFIG_BINFMT_MISC_MODULE)
1089 static struct ctl_table binfmt_misc_table[] = {
1090         { .ctl_name = 0 }
1091 };
1092 #endif
1093
1094 static struct ctl_table fs_table[] = {
1095         {
1096                 .ctl_name       = FS_NRINODE,
1097                 .procname       = "inode-nr",
1098                 .data           = &inodes_stat,
1099                 .maxlen         = 2*sizeof(int),
1100                 .mode           = 0444,
1101                 .proc_handler   = &proc_dointvec,
1102         },
1103         {
1104                 .ctl_name       = FS_STATINODE,
1105                 .procname       = "inode-state",
1106                 .data           = &inodes_stat,
1107                 .maxlen         = 7*sizeof(int),
1108                 .mode           = 0444,
1109                 .proc_handler   = &proc_dointvec,
1110         },
1111         {
1112                 .procname       = "file-nr",
1113                 .data           = &files_stat,
1114                 .maxlen         = 3*sizeof(int),
1115                 .mode           = 0444,
1116                 .proc_handler   = &proc_nr_files,
1117         },
1118         {
1119                 .ctl_name       = FS_MAXFILE,
1120                 .procname       = "file-max",
1121                 .data           = &files_stat.max_files,
1122                 .maxlen         = sizeof(int),
1123                 .mode           = 0644,
1124                 .proc_handler   = &proc_dointvec,
1125         },
1126         {
1127                 .ctl_name       = FS_DENTRY,
1128                 .procname       = "dentry-state",
1129                 .data           = &dentry_stat,
1130                 .maxlen         = 6*sizeof(int),
1131                 .mode           = 0444,
1132                 .proc_handler   = &proc_dointvec,
1133         },
1134         {
1135                 .ctl_name       = FS_OVERFLOWUID,
1136                 .procname       = "overflowuid",
1137                 .data           = &fs_overflowuid,
1138                 .maxlen         = sizeof(int),
1139                 .mode           = 0644,
1140                 .proc_handler   = &proc_dointvec_minmax,
1141                 .strategy       = &sysctl_intvec,
1142                 .extra1         = &minolduid,
1143                 .extra2         = &maxolduid,
1144         },
1145         {
1146                 .ctl_name       = FS_OVERFLOWGID,
1147                 .procname       = "overflowgid",
1148                 .data           = &fs_overflowgid,
1149                 .maxlen         = sizeof(int),
1150                 .mode           = 0644,
1151                 .proc_handler   = &proc_dointvec_minmax,
1152                 .strategy       = &sysctl_intvec,
1153                 .extra1         = &minolduid,
1154                 .extra2         = &maxolduid,
1155         },
1156         {
1157                 .ctl_name       = FS_LEASES,
1158                 .procname       = "leases-enable",
1159                 .data           = &leases_enable,
1160                 .maxlen         = sizeof(int),
1161                 .mode           = 0644,
1162                 .proc_handler   = &proc_dointvec,
1163         },
1164 #ifdef CONFIG_DNOTIFY
1165         {
1166                 .ctl_name       = FS_DIR_NOTIFY,
1167                 .procname       = "dir-notify-enable",
1168                 .data           = &dir_notify_enable,
1169                 .maxlen         = sizeof(int),
1170                 .mode           = 0644,
1171                 .proc_handler   = &proc_dointvec,
1172         },
1173 #endif
1174 #ifdef CONFIG_MMU
1175         {
1176                 .ctl_name       = FS_LEASE_TIME,
1177                 .procname       = "lease-break-time",
1178                 .data           = &lease_break_time,
1179                 .maxlen         = sizeof(int),
1180                 .mode           = 0644,
1181                 .proc_handler   = &proc_dointvec_minmax,
1182                 .strategy       = &sysctl_intvec,
1183                 .extra1         = &zero,
1184                 .extra2         = &two,
1185         },
1186         {
1187                 .procname       = "aio-nr",
1188                 .data           = &aio_nr,
1189                 .maxlen         = sizeof(aio_nr),
1190                 .mode           = 0444,
1191                 .proc_handler   = &proc_doulongvec_minmax,
1192         },
1193         {
1194                 .procname       = "aio-max-nr",
1195                 .data           = &aio_max_nr,
1196                 .maxlen         = sizeof(aio_max_nr),
1197                 .mode           = 0644,
1198                 .proc_handler   = &proc_doulongvec_minmax,
1199         },
1200 #ifdef CONFIG_INOTIFY_USER
1201         {
1202                 .ctl_name       = FS_INOTIFY,
1203                 .procname       = "inotify",
1204                 .mode           = 0555,
1205                 .child          = inotify_table,
1206         },
1207 #endif
1208 #endif
1209         {
1210                 .ctl_name       = KERN_SETUID_DUMPABLE,
1211                 .procname       = "suid_dumpable",
1212                 .data           = &suid_dumpable,
1213                 .maxlen         = sizeof(int),
1214                 .mode           = 0644,
1215                 .proc_handler   = &proc_dointvec,
1216         },
1217 #if defined(CONFIG_BINFMT_MISC) || defined(CONFIG_BINFMT_MISC_MODULE)
1218         {
1219                 .ctl_name       = CTL_UNNUMBERED,
1220                 .procname       = "binfmt_misc",
1221                 .mode           = 0555,
1222                 .child          = binfmt_misc_table,
1223         },
1224 #endif
1225 /*
1226  * NOTE: do not add new entries to this table unless you have read
1227  * Documentation/sysctl/ctl_unnumbered.txt
1228  */
1229         { .ctl_name = 0 }
1230 };
1231
1232 static struct ctl_table debug_table[] = {
1233 #if defined(CONFIG_X86) || defined(CONFIG_PPC)
1234         {
1235                 .ctl_name       = CTL_UNNUMBERED,
1236                 .procname       = "exception-trace",
1237                 .data           = &show_unhandled_signals,
1238                 .maxlen         = sizeof(int),
1239                 .mode           = 0644,
1240                 .proc_handler   = proc_dointvec
1241         },
1242 #endif
1243         { .ctl_name = 0 }
1244 };
1245
1246 static struct ctl_table dev_table[] = {
1247         { .ctl_name = 0 }
1248 };
1249
1250 static DEFINE_SPINLOCK(sysctl_lock);
1251
1252 /* called under sysctl_lock */
1253 static int use_table(struct ctl_table_header *p)
1254 {
1255         if (unlikely(p->unregistering))
1256                 return 0;
1257         p->used++;
1258         return 1;
1259 }
1260
1261 /* called under sysctl_lock */
1262 static void unuse_table(struct ctl_table_header *p)
1263 {
1264         if (!--p->used)
1265                 if (unlikely(p->unregistering))
1266                         complete(p->unregistering);
1267 }
1268
1269 /* called under sysctl_lock, will reacquire if has to wait */
1270 static void start_unregistering(struct ctl_table_header *p)
1271 {
1272         /*
1273          * if p->used is 0, nobody will ever touch that entry again;
1274          * we'll eliminate all paths to it before dropping sysctl_lock
1275          */
1276         if (unlikely(p->used)) {
1277                 struct completion wait;
1278                 init_completion(&wait);
1279                 p->unregistering = &wait;
1280                 spin_unlock(&sysctl_lock);
1281                 wait_for_completion(&wait);
1282                 spin_lock(&sysctl_lock);
1283         }
1284         /*
1285          * do not remove from the list until nobody holds it; walking the
1286          * list in do_sysctl() relies on that.
1287          */
1288         list_del_init(&p->ctl_entry);
1289 }
1290
1291 void sysctl_head_finish(struct ctl_table_header *head)
1292 {
1293         if (!head)
1294                 return;
1295         spin_lock(&sysctl_lock);
1296         unuse_table(head);
1297         spin_unlock(&sysctl_lock);
1298 }
1299
1300 struct ctl_table_header *sysctl_head_next(struct ctl_table_header *prev)
1301 {
1302         struct ctl_table_header *head;
1303         struct list_head *tmp;
1304         spin_lock(&sysctl_lock);
1305         if (prev) {
1306                 tmp = &prev->ctl_entry;
1307                 unuse_table(prev);
1308                 goto next;
1309         }
1310         tmp = &root_table_header.ctl_entry;
1311         for (;;) {
1312                 head = list_entry(tmp, struct ctl_table_header, ctl_entry);
1313
1314                 if (!use_table(head))
1315                         goto next;
1316                 spin_unlock(&sysctl_lock);
1317                 return head;
1318         next:
1319                 tmp = tmp->next;
1320                 if (tmp == &root_table_header.ctl_entry)
1321                         break;
1322         }
1323         spin_unlock(&sysctl_lock);
1324         return NULL;
1325 }
1326
1327 #ifdef CONFIG_SYSCTL_SYSCALL
1328 int do_sysctl(int __user *name, int nlen, void __user *oldval, size_t __user *oldlenp,
1329                void __user *newval, size_t newlen)
1330 {
1331         struct ctl_table_header *head;
1332         int error = -ENOTDIR;
1333
1334         if (nlen <= 0 || nlen >= CTL_MAXNAME)
1335                 return -ENOTDIR;
1336         if (oldval) {
1337                 int old_len;
1338                 if (!oldlenp || get_user(old_len, oldlenp))
1339                         return -EFAULT;
1340         }
1341
1342         for (head = sysctl_head_next(NULL); head;
1343                         head = sysctl_head_next(head)) {
1344                 error = parse_table(name, nlen, oldval, oldlenp,
1345                                         newval, newlen, head->ctl_table);
1346                 if (error != -ENOTDIR) {
1347                         sysctl_head_finish(head);
1348                         break;
1349                 }
1350         }
1351         return error;
1352 }
1353
1354 asmlinkage long sys_sysctl(struct __sysctl_args __user *args)
1355 {
1356         struct __sysctl_args tmp;
1357         int error;
1358
1359         if (copy_from_user(&tmp, args, sizeof(tmp)))
1360                 return -EFAULT;
1361
1362         lock_kernel();
1363         error = do_sysctl(tmp.name, tmp.nlen, tmp.oldval, tmp.oldlenp,
1364                           tmp.newval, tmp.newlen);
1365         unlock_kernel();
1366         return error;
1367 }
1368 #endif /* CONFIG_SYSCTL_SYSCALL */
1369
1370 /*
1371  * sysctl_perm does NOT grant the superuser all rights automatically, because
1372  * some sysctl variables are readonly even to root.
1373  */
1374
1375 static int test_perm(int mode, int op)
1376 {
1377         if (!current->euid)
1378                 mode >>= 6;
1379         else if (in_egroup_p(0))
1380                 mode >>= 3;
1381         if ((mode & op & 0007) == op)
1382                 return 0;
1383         return -EACCES;
1384 }
1385
1386 int sysctl_perm(struct ctl_table *table, int op)
1387 {
1388         int error;
1389         error = security_sysctl(table, op);
1390         if (error)
1391                 return error;
1392         return test_perm(table->mode, op);
1393 }
1394
1395 #ifdef CONFIG_SYSCTL_SYSCALL
1396 static int parse_table(int __user *name, int nlen,
1397                        void __user *oldval, size_t __user *oldlenp,
1398                        void __user *newval, size_t newlen,
1399                        struct ctl_table *table)
1400 {
1401         int n;
1402 repeat:
1403         if (!nlen)
1404                 return -ENOTDIR;
1405         if (get_user(n, name))
1406                 return -EFAULT;
1407         for ( ; table->ctl_name || table->procname; table++) {
1408                 if (!table->ctl_name)
1409                         continue;
1410                 if (n == table->ctl_name) {
1411                         int error;
1412                         if (table->child) {
1413                                 if (sysctl_perm(table, 001))
1414                                         return -EPERM;
1415                                 name++;
1416                                 nlen--;
1417                                 table = table->child;
1418                                 goto repeat;
1419                         }
1420                         error = do_sysctl_strategy(table, name, nlen,
1421                                                    oldval, oldlenp,
1422                                                    newval, newlen);
1423                         return error;
1424                 }
1425         }
1426         return -ENOTDIR;
1427 }
1428
1429 /* Perform the actual read/write of a sysctl table entry. */
1430 int do_sysctl_strategy (struct ctl_table *table,
1431                         int __user *name, int nlen,
1432                         void __user *oldval, size_t __user *oldlenp,
1433                         void __user *newval, size_t newlen)
1434 {
1435         int op = 0, rc;
1436
1437         if (oldval)
1438                 op |= 004;
1439         if (newval)
1440                 op |= 002;
1441         if (sysctl_perm(table, op))
1442                 return -EPERM;
1443
1444         if (table->strategy) {
1445                 rc = table->strategy(table, name, nlen, oldval, oldlenp,
1446                                      newval, newlen);
1447                 if (rc < 0)
1448                         return rc;
1449                 if (rc > 0)
1450                         return 0;
1451         }
1452
1453         /* If there is no strategy routine, or if the strategy returns
1454          * zero, proceed with automatic r/w */
1455         if (table->data && table->maxlen) {
1456                 rc = sysctl_data(table, name, nlen, oldval, oldlenp,
1457                                  newval, newlen);
1458                 if (rc < 0)
1459                         return rc;
1460         }
1461         return 0;
1462 }
1463 #endif /* CONFIG_SYSCTL_SYSCALL */
1464
1465 static void sysctl_set_parent(struct ctl_table *parent, struct ctl_table *table)
1466 {
1467         for (; table->ctl_name || table->procname; table++) {
1468                 table->parent = parent;
1469                 if (table->child)
1470                         sysctl_set_parent(table, table->child);
1471         }
1472 }
1473
1474 static __init int sysctl_init(void)
1475 {
1476         sysctl_set_parent(NULL, root_table);
1477         return 0;
1478 }
1479
1480 core_initcall(sysctl_init);
1481
1482 /**
1483  * register_sysctl_table - register a sysctl hierarchy
1484  * @table: the top-level table structure
1485  *
1486  * Register a sysctl table hierarchy. @table should be a filled in ctl_table
1487  * array. An entry with a ctl_name of 0 terminates the table.
1488  *
1489  * The members of the &struct ctl_table structure are used as follows:
1490  *
1491  * ctl_name - This is the numeric sysctl value used by sysctl(2). The number
1492  *            must be unique within that level of sysctl
1493  *
1494  * procname - the name of the sysctl file under /proc/sys. Set to %NULL to not
1495  *            enter a sysctl file
1496  *
1497  * data - a pointer to data for use by proc_handler
1498  *
1499  * maxlen - the maximum size in bytes of the data
1500  *
1501  * mode - the file permissions for the /proc/sys file, and for sysctl(2)
1502  *
1503  * child - a pointer to the child sysctl table if this entry is a directory, or
1504  *         %NULL.
1505  *
1506  * proc_handler - the text handler routine (described below)
1507  *
1508  * strategy - the strategy routine (described below)
1509  *
1510  * de - for internal use by the sysctl routines
1511  *
1512  * extra1, extra2 - extra pointers usable by the proc handler routines
1513  *
1514  * Leaf nodes in the sysctl tree will be represented by a single file
1515  * under /proc; non-leaf nodes will be represented by directories.
1516  *
1517  * sysctl(2) can automatically manage read and write requests through
1518  * the sysctl table.  The data and maxlen fields of the ctl_table
1519  * struct enable minimal validation of the values being written to be
1520  * performed, and the mode field allows minimal authentication.
1521  *
1522  * More sophisticated management can be enabled by the provision of a
1523  * strategy routine with the table entry.  This will be called before
1524  * any automatic read or write of the data is performed.
1525  *
1526  * The strategy routine may return
1527  *
1528  * < 0 - Error occurred (error is passed to user process)
1529  *
1530  * 0   - OK - proceed with automatic read or write.
1531  *
1532  * > 0 - OK - read or write has been done by the strategy routine, so
1533  *       return immediately.
1534  *
1535  * There must be a proc_handler routine for any terminal nodes
1536  * mirrored under /proc/sys (non-terminals are handled by a built-in
1537  * directory handler).  Several default handlers are available to
1538  * cover common cases -
1539  *
1540  * proc_dostring(), proc_dointvec(), proc_dointvec_jiffies(),
1541  * proc_dointvec_userhz_jiffies(), proc_dointvec_minmax(),
1542  * proc_doulongvec_ms_jiffies_minmax(), proc_doulongvec_minmax()
1543  *
1544  * It is the handler's job to read the input buffer from user memory
1545  * and process it. The handler should return 0 on success.
1546  *
1547  * This routine returns %NULL on a failure to register, and a pointer
1548  * to the table header on success.
1549  */
1550 struct ctl_table_header *register_sysctl_table(struct ctl_table * table)
1551 {
1552         struct ctl_table_header *tmp;
1553         tmp = kmalloc(sizeof(struct ctl_table_header), GFP_KERNEL);
1554         if (!tmp)
1555                 return NULL;
1556         tmp->ctl_table = table;
1557         INIT_LIST_HEAD(&tmp->ctl_entry);
1558         tmp->used = 0;
1559         tmp->unregistering = NULL;
1560         sysctl_set_parent(NULL, table);
1561         spin_lock(&sysctl_lock);
1562         list_add_tail(&tmp->ctl_entry, &root_table_header.ctl_entry);
1563         spin_unlock(&sysctl_lock);
1564         return tmp;
1565 }
1566
1567 /**
1568  * unregister_sysctl_table - unregister a sysctl table hierarchy
1569  * @header: the header returned from register_sysctl_table
1570  *
1571  * Unregisters the sysctl table and all children. proc entries may not
1572  * actually be removed until they are no longer used by anyone.
1573  */
1574 void unregister_sysctl_table(struct ctl_table_header * header)
1575 {
1576         might_sleep();
1577         spin_lock(&sysctl_lock);
1578         start_unregistering(header);
1579         spin_unlock(&sysctl_lock);
1580         kfree(header);
1581 }
1582
1583 #else /* !CONFIG_SYSCTL */
1584 struct ctl_table_header *register_sysctl_table(struct ctl_table * table)
1585 {
1586         return NULL;
1587 }
1588
1589 void unregister_sysctl_table(struct ctl_table_header * table)
1590 {
1591 }
1592
1593 #endif /* CONFIG_SYSCTL */
1594
1595 /*
1596  * /proc/sys support
1597  */
1598
1599 #ifdef CONFIG_PROC_SYSCTL
1600
1601 static int _proc_do_string(void* data, int maxlen, int write,
1602                            struct file *filp, void __user *buffer,
1603                            size_t *lenp, loff_t *ppos)
1604 {
1605         size_t len;
1606         char __user *p;
1607         char c;
1608
1609         if (!data || !maxlen || !*lenp) {
1610                 *lenp = 0;
1611                 return 0;
1612         }
1613
1614         if (write) {
1615                 len = 0;
1616                 p = buffer;
1617                 while (len < *lenp) {
1618                         if (get_user(c, p++))
1619                                 return -EFAULT;
1620                         if (c == 0 || c == '\n')
1621                                 break;
1622                         len++;
1623                 }
1624                 if (len >= maxlen)
1625                         len = maxlen-1;
1626                 if(copy_from_user(data, buffer, len))
1627                         return -EFAULT;
1628                 ((char *) data)[len] = 0;
1629                 *ppos += *lenp;
1630         } else {
1631                 len = strlen(data);
1632                 if (len > maxlen)
1633                         len = maxlen;
1634
1635                 if (*ppos > len) {
1636                         *lenp = 0;
1637                         return 0;
1638                 }
1639
1640                 data += *ppos;
1641                 len  -= *ppos;
1642
1643                 if (len > *lenp)
1644                         len = *lenp;
1645                 if (len)
1646                         if(copy_to_user(buffer, data, len))
1647                                 return -EFAULT;
1648                 if (len < *lenp) {
1649                         if(put_user('\n', ((char __user *) buffer) + len))
1650                                 return -EFAULT;
1651                         len++;
1652                 }
1653                 *lenp = len;
1654                 *ppos += len;
1655         }
1656         return 0;
1657 }
1658
1659 /**
1660  * proc_dostring - read a string sysctl
1661  * @table: the sysctl table
1662  * @write: %TRUE if this is a write to the sysctl file
1663  * @filp: the file structure
1664  * @buffer: the user buffer
1665  * @lenp: the size of the user buffer
1666  * @ppos: file position
1667  *
1668  * Reads/writes a string from/to the user buffer. If the kernel
1669  * buffer provided is not large enough to hold the string, the
1670  * string is truncated. The copied string is %NULL-terminated.
1671  * If the string is being read by the user process, it is copied
1672  * and a newline '\n' is added. It is truncated if the buffer is
1673  * not large enough.
1674  *
1675  * Returns 0 on success.
1676  */
1677 int proc_dostring(struct ctl_table *table, int write, struct file *filp,
1678                   void __user *buffer, size_t *lenp, loff_t *ppos)
1679 {
1680         return _proc_do_string(table->data, table->maxlen, write, filp,
1681                                buffer, lenp, ppos);
1682 }
1683
1684
1685 static int do_proc_dointvec_conv(int *negp, unsigned long *lvalp,
1686                                  int *valp,
1687                                  int write, void *data)
1688 {
1689         if (write) {
1690                 *valp = *negp ? -*lvalp : *lvalp;
1691         } else {
1692                 int val = *valp;
1693                 if (val < 0) {
1694                         *negp = -1;
1695                         *lvalp = (unsigned long)-val;
1696                 } else {
1697                         *negp = 0;
1698                         *lvalp = (unsigned long)val;
1699                 }
1700         }
1701         return 0;
1702 }
1703
1704 static int __do_proc_dointvec(void *tbl_data, struct ctl_table *table,
1705                   int write, struct file *filp, void __user *buffer,
1706                   size_t *lenp, loff_t *ppos,
1707                   int (*conv)(int *negp, unsigned long *lvalp, int *valp,
1708                               int write, void *data),
1709                   void *data)
1710 {
1711 #define TMPBUFLEN 21
1712         int *i, vleft, first=1, neg, val;
1713         unsigned long lval;
1714         size_t left, len;
1715
1716         char buf[TMPBUFLEN], *p;
1717         char __user *s = buffer;
1718
1719         if (!tbl_data || !table->maxlen || !*lenp ||
1720             (*ppos && !write)) {
1721                 *lenp = 0;
1722                 return 0;
1723         }
1724
1725         i = (int *) tbl_data;
1726         vleft = table->maxlen / sizeof(*i);
1727         left = *lenp;
1728
1729         if (!conv)
1730                 conv = do_proc_dointvec_conv;
1731
1732         for (; left && vleft--; i++, first=0) {
1733                 if (write) {
1734                         while (left) {
1735                                 char c;
1736                                 if (get_user(c, s))
1737                                         return -EFAULT;
1738                                 if (!isspace(c))
1739                                         break;
1740                                 left--;
1741                                 s++;
1742                         }
1743                         if (!left)
1744                                 break;
1745                         neg = 0;
1746                         len = left;
1747                         if (len > sizeof(buf) - 1)
1748                                 len = sizeof(buf) - 1;
1749                         if (copy_from_user(buf, s, len))
1750                                 return -EFAULT;
1751                         buf[len] = 0;
1752                         p = buf;
1753                         if (*p == '-' && left > 1) {
1754                                 neg = 1;
1755                                 p++;
1756                         }
1757                         if (*p < '0' || *p > '9')
1758                                 break;
1759
1760                         lval = simple_strtoul(p, &p, 0);
1761
1762                         len = p-buf;
1763                         if ((len < left) && *p && !isspace(*p))
1764                                 break;
1765                         if (neg)
1766                                 val = -val;
1767                         s += len;
1768                         left -= len;
1769
1770                         if (conv(&neg, &lval, i, 1, data))
1771                                 break;
1772                 } else {
1773                         p = buf;
1774                         if (!first)
1775                                 *p++ = '\t';
1776
1777                         if (conv(&neg, &lval, i, 0, data))
1778                                 break;
1779
1780                         sprintf(p, "%s%lu", neg ? "-" : "", lval);
1781                         len = strlen(buf);
1782                         if (len > left)
1783                                 len = left;
1784                         if(copy_to_user(s, buf, len))
1785                                 return -EFAULT;
1786                         left -= len;
1787                         s += len;
1788                 }
1789         }
1790
1791         if (!write && !first && left) {
1792                 if(put_user('\n', s))
1793                         return -EFAULT;
1794                 left--, s++;
1795         }
1796         if (write) {
1797                 while (left) {
1798                         char c;
1799                         if (get_user(c, s++))
1800                                 return -EFAULT;
1801                         if (!isspace(c))
1802                                 break;
1803                         left--;
1804                 }
1805         }
1806         if (write && first)
1807                 return -EINVAL;
1808         *lenp -= left;
1809         *ppos += *lenp;
1810         return 0;
1811 #undef TMPBUFLEN
1812 }
1813
1814 static int do_proc_dointvec(struct ctl_table *table, int write, struct file *filp,
1815                   void __user *buffer, size_t *lenp, loff_t *ppos,
1816                   int (*conv)(int *negp, unsigned long *lvalp, int *valp,
1817                               int write, void *data),
1818                   void *data)
1819 {
1820         return __do_proc_dointvec(table->data, table, write, filp,
1821                         buffer, lenp, ppos, conv, data);
1822 }
1823
1824 /**
1825  * proc_dointvec - read a vector of integers
1826  * @table: the sysctl table
1827  * @write: %TRUE if this is a write to the sysctl file
1828  * @filp: the file structure
1829  * @buffer: the user buffer
1830  * @lenp: the size of the user buffer
1831  * @ppos: file position
1832  *
1833  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
1834  * values from/to the user buffer, treated as an ASCII string.
1835  *
1836  * Returns 0 on success.
1837  */
1838 int proc_dointvec(struct ctl_table *table, int write, struct file *filp,
1839                      void __user *buffer, size_t *lenp, loff_t *ppos)
1840 {
1841     return do_proc_dointvec(table,write,filp,buffer,lenp,ppos,
1842                             NULL,NULL);
1843 }
1844
1845 #define OP_SET  0
1846 #define OP_AND  1
1847 #define OP_OR   2
1848
1849 static int do_proc_dointvec_bset_conv(int *negp, unsigned long *lvalp,
1850                                       int *valp,
1851                                       int write, void *data)
1852 {
1853         int op = *(int *)data;
1854         if (write) {
1855                 int val = *negp ? -*lvalp : *lvalp;
1856                 switch(op) {
1857                 case OP_SET:    *valp = val; break;
1858                 case OP_AND:    *valp &= val; break;
1859                 case OP_OR:     *valp |= val; break;
1860                 }
1861         } else {
1862                 int val = *valp;
1863                 if (val < 0) {
1864                         *negp = -1;
1865                         *lvalp = (unsigned long)-val;
1866                 } else {
1867                         *negp = 0;
1868                         *lvalp = (unsigned long)val;
1869                 }
1870         }
1871         return 0;
1872 }
1873
1874 /*
1875  *      init may raise the set.
1876  */
1877
1878 int proc_dointvec_bset(struct ctl_table *table, int write, struct file *filp,
1879                         void __user *buffer, size_t *lenp, loff_t *ppos)
1880 {
1881         int op;
1882
1883         if (write && !capable(CAP_SYS_MODULE)) {
1884                 return -EPERM;
1885         }
1886
1887         op = is_init(current) ? OP_SET : OP_AND;
1888         return do_proc_dointvec(table,write,filp,buffer,lenp,ppos,
1889                                 do_proc_dointvec_bset_conv,&op);
1890 }
1891
1892 /*
1893  *      Taint values can only be increased
1894  */
1895 static int proc_dointvec_taint(struct ctl_table *table, int write, struct file *filp,
1896                                void __user *buffer, size_t *lenp, loff_t *ppos)
1897 {
1898         int op;
1899
1900         if (write && !capable(CAP_SYS_ADMIN))
1901                 return -EPERM;
1902
1903         op = OP_OR;
1904         return do_proc_dointvec(table,write,filp,buffer,lenp,ppos,
1905                                 do_proc_dointvec_bset_conv,&op);
1906 }
1907
1908 struct do_proc_dointvec_minmax_conv_param {
1909         int *min;
1910         int *max;
1911 };
1912
1913 static int do_proc_dointvec_minmax_conv(int *negp, unsigned long *lvalp,
1914                                         int *valp,
1915                                         int write, void *data)
1916 {
1917         struct do_proc_dointvec_minmax_conv_param *param = data;
1918         if (write) {
1919                 int val = *negp ? -*lvalp : *lvalp;
1920                 if ((param->min && *param->min > val) ||
1921                     (param->max && *param->max < val))
1922                         return -EINVAL;
1923                 *valp = val;
1924         } else {
1925                 int val = *valp;
1926                 if (val < 0) {
1927                         *negp = -1;
1928                         *lvalp = (unsigned long)-val;
1929                 } else {
1930                         *negp = 0;
1931                         *lvalp = (unsigned long)val;
1932                 }
1933         }
1934         return 0;
1935 }
1936
1937 /**
1938  * proc_dointvec_minmax - read a vector of integers with min/max values
1939  * @table: the sysctl table
1940  * @write: %TRUE if this is a write to the sysctl file
1941  * @filp: the file structure
1942  * @buffer: the user buffer
1943  * @lenp: the size of the user buffer
1944  * @ppos: file position
1945  *
1946  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
1947  * values from/to the user buffer, treated as an ASCII string.
1948  *
1949  * This routine will ensure the values are within the range specified by
1950  * table->extra1 (min) and table->extra2 (max).
1951  *
1952  * Returns 0 on success.
1953  */
1954 int proc_dointvec_minmax(struct ctl_table *table, int write, struct file *filp,
1955                   void __user *buffer, size_t *lenp, loff_t *ppos)
1956 {
1957         struct do_proc_dointvec_minmax_conv_param param = {
1958                 .min = (int *) table->extra1,
1959                 .max = (int *) table->extra2,
1960         };
1961         return do_proc_dointvec(table, write, filp, buffer, lenp, ppos,
1962                                 do_proc_dointvec_minmax_conv, &param);
1963 }
1964
1965 static int __do_proc_doulongvec_minmax(void *data, struct ctl_table *table, int write,
1966                                      struct file *filp,
1967                                      void __user *buffer,
1968                                      size_t *lenp, loff_t *ppos,
1969                                      unsigned long convmul,
1970                                      unsigned long convdiv)
1971 {
1972 #define TMPBUFLEN 21
1973         unsigned long *i, *min, *max, val;
1974         int vleft, first=1, neg;
1975         size_t len, left;
1976         char buf[TMPBUFLEN], *p;
1977         char __user *s = buffer;
1978
1979         if (!data || !table->maxlen || !*lenp ||
1980             (*ppos && !write)) {
1981                 *lenp = 0;
1982                 return 0;
1983         }
1984
1985         i = (unsigned long *) data;
1986         min = (unsigned long *) table->extra1;
1987         max = (unsigned long *) table->extra2;
1988         vleft = table->maxlen / sizeof(unsigned long);
1989         left = *lenp;
1990
1991         for (; left && vleft--; i++, min++, max++, first=0) {
1992                 if (write) {
1993                         while (left) {
1994                                 char c;
1995                                 if (get_user(c, s))
1996                                         return -EFAULT;
1997                                 if (!isspace(c))
1998                                         break;
1999                                 left--;
2000                                 s++;
2001                         }
2002                         if (!left)
2003                                 break;
2004                         neg = 0;
2005                         len = left;
2006                         if (len > TMPBUFLEN-1)
2007                                 len = TMPBUFLEN-1;
2008                         if (copy_from_user(buf, s, len))
2009                                 return -EFAULT;
2010                         buf[len] = 0;
2011                         p = buf;
2012                         if (*p == '-' && left > 1) {
2013                                 neg = 1;
2014                                 p++;
2015                         }
2016                         if (*p < '0' || *p > '9')
2017                                 break;
2018                         val = simple_strtoul(p, &p, 0) * convmul / convdiv ;
2019                         len = p-buf;
2020                         if ((len < left) && *p && !isspace(*p))
2021                                 break;
2022                         if (neg)
2023                                 val = -val;
2024                         s += len;
2025                         left -= len;
2026
2027                         if(neg)
2028                                 continue;
2029                         if ((min && val < *min) || (max && val > *max))
2030                                 continue;
2031                         *i = val;
2032                 } else {
2033                         p = buf;
2034                         if (!first)
2035                                 *p++ = '\t';
2036                         sprintf(p, "%lu", convdiv * (*i) / convmul);
2037                         len = strlen(buf);
2038                         if (len > left)
2039                                 len = left;
2040                         if(copy_to_user(s, buf, len))
2041                                 return -EFAULT;
2042                         left -= len;
2043                         s += len;
2044                 }
2045         }
2046
2047         if (!write && !first && left) {
2048                 if(put_user('\n', s))
2049                         return -EFAULT;
2050                 left--, s++;
2051         }
2052         if (write) {
2053                 while (left) {
2054                         char c;
2055                         if (get_user(c, s++))
2056                                 return -EFAULT;
2057                         if (!isspace(c))
2058                                 break;
2059                         left--;
2060                 }
2061         }
2062         if (write && first)
2063                 return -EINVAL;
2064         *lenp -= left;
2065         *ppos += *lenp;
2066         return 0;
2067 #undef TMPBUFLEN
2068 }
2069
2070 static int do_proc_doulongvec_minmax(struct ctl_table *table, int write,
2071                                      struct file *filp,
2072                                      void __user *buffer,
2073                                      size_t *lenp, loff_t *ppos,
2074                                      unsigned long convmul,
2075                                      unsigned long convdiv)
2076 {
2077         return __do_proc_doulongvec_minmax(table->data, table, write,
2078                         filp, buffer, lenp, ppos, convmul, convdiv);
2079 }
2080
2081 /**
2082  * proc_doulongvec_minmax - read a vector of long integers with min/max values
2083  * @table: the sysctl table
2084  * @write: %TRUE if this is a write to the sysctl file
2085  * @filp: the file structure
2086  * @buffer: the user buffer
2087  * @lenp: the size of the user buffer
2088  * @ppos: file position
2089  *
2090  * Reads/writes up to table->maxlen/sizeof(unsigned long) unsigned long
2091  * values from/to the user buffer, treated as an ASCII string.
2092  *
2093  * This routine will ensure the values are within the range specified by
2094  * table->extra1 (min) and table->extra2 (max).
2095  *
2096  * Returns 0 on success.
2097  */
2098 int proc_doulongvec_minmax(struct ctl_table *table, int write, struct file *filp,
2099                            void __user *buffer, size_t *lenp, loff_t *ppos)
2100 {
2101     return do_proc_doulongvec_minmax(table, write, filp, buffer, lenp, ppos, 1l, 1l);
2102 }
2103
2104 /**
2105  * proc_doulongvec_ms_jiffies_minmax - read a vector of millisecond values with min/max values
2106  * @table: the sysctl table
2107  * @write: %TRUE if this is a write to the sysctl file
2108  * @filp: the file structure
2109  * @buffer: the user buffer
2110  * @lenp: the size of the user buffer
2111  * @ppos: file position
2112  *
2113  * Reads/writes up to table->maxlen/sizeof(unsigned long) unsigned long
2114  * values from/to the user buffer, treated as an ASCII string. The values
2115  * are treated as milliseconds, and converted to jiffies when they are stored.
2116  *
2117  * This routine will ensure the values are within the range specified by
2118  * table->extra1 (min) and table->extra2 (max).
2119  *
2120  * Returns 0 on success.
2121  */
2122 int proc_doulongvec_ms_jiffies_minmax(struct ctl_table *table, int write,
2123                                       struct file *filp,
2124                                       void __user *buffer,
2125                                       size_t *lenp, loff_t *ppos)
2126 {
2127     return do_proc_doulongvec_minmax(table, write, filp, buffer,
2128                                      lenp, ppos, HZ, 1000l);
2129 }
2130
2131
2132 static int do_proc_dointvec_jiffies_conv(int *negp, unsigned long *lvalp,
2133                                          int *valp,
2134                                          int write, void *data)
2135 {
2136         if (write) {
2137                 if (*lvalp > LONG_MAX / HZ)
2138                         return 1;
2139                 *valp = *negp ? -(*lvalp*HZ) : (*lvalp*HZ);
2140         } else {
2141                 int val = *valp;
2142                 unsigned long lval;
2143                 if (val < 0) {
2144                         *negp = -1;
2145                         lval = (unsigned long)-val;
2146                 } else {
2147                         *negp = 0;
2148                         lval = (unsigned long)val;
2149                 }
2150                 *lvalp = lval / HZ;
2151         }
2152         return 0;
2153 }
2154
2155 static int do_proc_dointvec_userhz_jiffies_conv(int *negp, unsigned long *lvalp,
2156                                                 int *valp,
2157                                                 int write, void *data)
2158 {
2159         if (write) {
2160                 if (USER_HZ < HZ && *lvalp > (LONG_MAX / HZ) * USER_HZ)
2161                         return 1;
2162                 *valp = clock_t_to_jiffies(*negp ? -*lvalp : *lvalp);
2163         } else {
2164                 int val = *valp;
2165                 unsigned long lval;
2166                 if (val < 0) {
2167                         *negp = -1;
2168                         lval = (unsigned long)-val;
2169                 } else {
2170                         *negp = 0;
2171                         lval = (unsigned long)val;
2172                 }
2173                 *lvalp = jiffies_to_clock_t(lval);
2174         }
2175         return 0;
2176 }
2177
2178 static int do_proc_dointvec_ms_jiffies_conv(int *negp, unsigned long *lvalp,
2179                                             int *valp,
2180                                             int write, void *data)
2181 {
2182         if (write) {
2183                 *valp = msecs_to_jiffies(*negp ? -*lvalp : *lvalp);
2184         } else {
2185                 int val = *valp;
2186                 unsigned long lval;
2187                 if (val < 0) {
2188                         *negp = -1;
2189                         lval = (unsigned long)-val;
2190                 } else {
2191                         *negp = 0;
2192                         lval = (unsigned long)val;
2193                 }
2194                 *lvalp = jiffies_to_msecs(lval);
2195         }
2196         return 0;
2197 }
2198
2199 /**
2200  * proc_dointvec_jiffies - read a vector of integers as seconds
2201  * @table: the sysctl table
2202  * @write: %TRUE if this is a write to the sysctl file
2203  * @filp: the file structure
2204  * @buffer: the user buffer
2205  * @lenp: the size of the user buffer
2206  * @ppos: file position
2207  *
2208  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2209  * values from/to the user buffer, treated as an ASCII string.
2210  * The values read are assumed to be in seconds, and are converted into
2211  * jiffies.
2212  *
2213  * Returns 0 on success.
2214  */
2215 int proc_dointvec_jiffies(struct ctl_table *table, int write, struct file *filp,
2216                           void __user *buffer, size_t *lenp, loff_t *ppos)
2217 {
2218     return do_proc_dointvec(table,write,filp,buffer,lenp,ppos,
2219                             do_proc_dointvec_jiffies_conv,NULL);
2220 }
2221
2222 /**
2223  * proc_dointvec_userhz_jiffies - read a vector of integers as 1/USER_HZ seconds
2224  * @table: the sysctl table
2225  * @write: %TRUE if this is a write to the sysctl file
2226  * @filp: the file structure
2227  * @buffer: the user buffer
2228  * @lenp: the size of the user buffer
2229  * @ppos: pointer to the file position
2230  *
2231  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2232  * values from/to the user buffer, treated as an ASCII string.
2233  * The values read are assumed to be in 1/USER_HZ seconds, and
2234  * are converted into jiffies.
2235  *
2236  * Returns 0 on success.
2237  */
2238 int proc_dointvec_userhz_jiffies(struct ctl_table *table, int write, struct file *filp,
2239                                  void __user *buffer, size_t *lenp, loff_t *ppos)
2240 {
2241     return do_proc_dointvec(table,write,filp,buffer,lenp,ppos,
2242                             do_proc_dointvec_userhz_jiffies_conv,NULL);
2243 }
2244
2245 /**
2246  * proc_dointvec_ms_jiffies - read a vector of integers as 1 milliseconds
2247  * @table: the sysctl table
2248  * @write: %TRUE if this is a write to the sysctl file
2249  * @filp: the file structure
2250  * @buffer: the user buffer
2251  * @lenp: the size of the user buffer
2252  * @ppos: file position
2253  * @ppos: the current position in the file
2254  *
2255  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2256  * values from/to the user buffer, treated as an ASCII string.
2257  * The values read are assumed to be in 1/1000 seconds, and
2258  * are converted into jiffies.
2259  *
2260  * Returns 0 on success.
2261  */
2262 int proc_dointvec_ms_jiffies(struct ctl_table *table, int write, struct file *filp,
2263                              void __user *buffer, size_t *lenp, loff_t *ppos)
2264 {
2265         return do_proc_dointvec(table, write, filp, buffer, lenp, ppos,
2266                                 do_proc_dointvec_ms_jiffies_conv, NULL);
2267 }
2268
2269 static int proc_do_cad_pid(struct ctl_table *table, int write, struct file *filp,
2270                            void __user *buffer, size_t *lenp, loff_t *ppos)
2271 {
2272         struct pid *new_pid;
2273         pid_t tmp;
2274         int r;
2275
2276         tmp = pid_nr(cad_pid);
2277
2278         r = __do_proc_dointvec(&tmp, table, write, filp, buffer,
2279                                lenp, ppos, NULL, NULL);
2280         if (r || !write)
2281                 return r;
2282
2283         new_pid = find_get_pid(tmp);
2284         if (!new_pid)
2285                 return -ESRCH;
2286
2287         put_pid(xchg(&cad_pid, new_pid));
2288         return 0;
2289 }
2290
2291 #else /* CONFIG_PROC_FS */
2292
2293 int proc_dostring(struct ctl_table *table, int write, struct file *filp,
2294                   void __user *buffer, size_t *lenp, loff_t *ppos)
2295 {
2296         return -ENOSYS;
2297 }
2298
2299 int proc_dointvec(struct ctl_table *table, int write, struct file *filp,
2300                   void __user *buffer, size_t *lenp, loff_t *ppos)
2301 {
2302         return -ENOSYS;
2303 }
2304
2305 int proc_dointvec_bset(struct ctl_table *table, int write, struct file *filp,
2306                         void __user *buffer, size_t *lenp, loff_t *ppos)
2307 {
2308         return -ENOSYS;
2309 }
2310
2311 int proc_dointvec_minmax(struct ctl_table *table, int write, struct file *filp,
2312                     void __user *buffer, size_t *lenp, loff_t *ppos)
2313 {
2314         return -ENOSYS;
2315 }
2316
2317 int proc_dointvec_jiffies(struct ctl_table *table, int write, struct file *filp,
2318                     void __user *buffer, size_t *lenp, loff_t *ppos)
2319 {
2320         return -ENOSYS;
2321 }
2322
2323 int proc_dointvec_userhz_jiffies(struct ctl_table *table, int write, struct file *filp,
2324                     void __user *buffer, size_t *lenp, loff_t *ppos)
2325 {
2326         return -ENOSYS;
2327 }
2328
2329 int proc_dointvec_ms_jiffies(struct ctl_table *table, int write, struct file *filp,
2330                              void __user *buffer, size_t *lenp, loff_t *ppos)
2331 {
2332         return -ENOSYS;
2333 }
2334
2335 int proc_doulongvec_minmax(struct ctl_table *table, int write, struct file *filp,
2336                     void __user *buffer, size_t *lenp, loff_t *ppos)
2337 {
2338         return -ENOSYS;
2339 }
2340
2341 int proc_doulongvec_ms_jiffies_minmax(struct ctl_table *table, int write,
2342                                       struct file *filp,
2343                                       void __user *buffer,
2344                                       size_t *lenp, loff_t *ppos)
2345 {
2346     return -ENOSYS;
2347 }
2348
2349
2350 #endif /* CONFIG_PROC_FS */
2351
2352
2353 #ifdef CONFIG_SYSCTL_SYSCALL
2354 /*
2355  * General sysctl support routines
2356  */
2357
2358 /* The generic sysctl data routine (used if no strategy routine supplied) */
2359 int sysctl_data(struct ctl_table *table, int __user *name, int nlen,
2360                 void __user *oldval, size_t __user *oldlenp,
2361                 void __user *newval, size_t newlen)
2362 {
2363         size_t len;
2364
2365         /* Get out of I don't have a variable */
2366         if (!table->data || !table->maxlen)
2367                 return -ENOTDIR;
2368
2369         if (oldval && oldlenp) {
2370                 if (get_user(len, oldlenp))
2371                         return -EFAULT;
2372                 if (len) {
2373                         if (len > table->maxlen)
2374                                 len = table->maxlen;
2375                         if (copy_to_user(oldval, table->data, len))
2376                                 return -EFAULT;
2377                         if (put_user(len, oldlenp))
2378                                 return -EFAULT;
2379                 }
2380         }
2381
2382         if (newval && newlen) {
2383                 if (newlen > table->maxlen)
2384                         newlen = table->maxlen;
2385
2386                 if (copy_from_user(table->data, newval, newlen))
2387                         return -EFAULT;
2388         }
2389         return 1;
2390 }
2391
2392 /* The generic string strategy routine: */
2393 int sysctl_string(struct ctl_table *table, int __user *name, int nlen,
2394                   void __user *oldval, size_t __user *oldlenp,
2395                   void __user *newval, size_t newlen)
2396 {
2397         if (!table->data || !table->maxlen)
2398                 return -ENOTDIR;
2399
2400         if (oldval && oldlenp) {
2401                 size_t bufsize;
2402                 if (get_user(bufsize, oldlenp))
2403                         return -EFAULT;
2404                 if (bufsize) {
2405                         size_t len = strlen(table->data), copied;
2406
2407                         /* This shouldn't trigger for a well-formed sysctl */
2408                         if (len > table->maxlen)
2409                                 len = table->maxlen;
2410
2411                         /* Copy up to a max of bufsize-1 bytes of the string */
2412                         copied = (len >= bufsize) ? bufsize - 1 : len;
2413
2414                         if (copy_to_user(oldval, table->data, copied) ||
2415                             put_user(0, (char __user *)(oldval + copied)))
2416                                 return -EFAULT;
2417                         if (put_user(len, oldlenp))
2418                                 return -EFAULT;
2419                 }
2420         }
2421         if (newval && newlen) {
2422                 size_t len = newlen;
2423                 if (len > table->maxlen)
2424                         len = table->maxlen;
2425                 if(copy_from_user(table->data, newval, len))
2426                         return -EFAULT;
2427                 if (len == table->maxlen)
2428                         len--;
2429                 ((char *) table->data)[len] = 0;
2430         }
2431         return 1;
2432 }
2433
2434 /*
2435  * This function makes sure that all of the integers in the vector
2436  * are between the minimum and maximum values given in the arrays
2437  * table->extra1 and table->extra2, respectively.
2438  */
2439 int sysctl_intvec(struct ctl_table *table, int __user *name, int nlen,
2440                 void __user *oldval, size_t __user *oldlenp,
2441                 void __user *newval, size_t newlen)
2442 {
2443
2444         if (newval && newlen) {
2445                 int __user *vec = (int __user *) newval;
2446                 int *min = (int *) table->extra1;
2447                 int *max = (int *) table->extra2;
2448                 size_t length;
2449                 int i;
2450
2451                 if (newlen % sizeof(int) != 0)
2452                         return -EINVAL;
2453
2454                 if (!table->extra1 && !table->extra2)
2455                         return 0;
2456
2457                 if (newlen > table->maxlen)
2458                         newlen = table->maxlen;
2459                 length = newlen / sizeof(int);
2460
2461                 for (i = 0; i < length; i++) {
2462                         int value;
2463                         if (get_user(value, vec + i))
2464                                 return -EFAULT;
2465                         if (min && value < min[i])
2466                                 return -EINVAL;
2467                         if (max && value > max[i])
2468                                 return -EINVAL;
2469                 }
2470         }
2471         return 0;
2472 }
2473
2474 /* Strategy function to convert jiffies to seconds */
2475 int sysctl_jiffies(struct ctl_table *table, int __user *name, int nlen,
2476                 void __user *oldval, size_t __user *oldlenp,
2477                 void __user *newval, size_t newlen)
2478 {
2479         if (oldval && oldlenp) {
2480                 size_t olen;
2481
2482                 if (get_user(olen, oldlenp))
2483                         return -EFAULT;
2484                 if (olen) {
2485                         int val;
2486
2487                         if (olen < sizeof(int))
2488                                 return -EINVAL;
2489
2490                         val = *(int *)(table->data) / HZ;
2491                         if (put_user(val, (int __user *)oldval))
2492                                 return -EFAULT;
2493                         if (put_user(sizeof(int), oldlenp))
2494                                 return -EFAULT;
2495                 }
2496         }
2497         if (newval && newlen) {
2498                 int new;
2499                 if (newlen != sizeof(int))
2500                         return -EINVAL;
2501                 if (get_user(new, (int __user *)newval))
2502                         return -EFAULT;
2503                 *(int *)(table->data) = new*HZ;
2504         }
2505         return 1;
2506 }
2507
2508 /* Strategy function to convert jiffies to seconds */
2509 int sysctl_ms_jiffies(struct ctl_table *table, int __user *name, int nlen,
2510                 void __user *oldval, size_t __user *oldlenp,
2511                 void __user *newval, size_t newlen)
2512 {
2513         if (oldval && oldlenp) {
2514                 size_t olen;
2515
2516                 if (get_user(olen, oldlenp))
2517                         return -EFAULT;
2518                 if (olen) {
2519                         int val;
2520
2521                         if (olen < sizeof(int))
2522                                 return -EINVAL;
2523
2524                         val = jiffies_to_msecs(*(int *)(table->data));
2525                         if (put_user(val, (int __user *)oldval))
2526                                 return -EFAULT;
2527                         if (put_user(sizeof(int), oldlenp))
2528                                 return -EFAULT;
2529                 }
2530         }
2531         if (newval && newlen) {
2532                 int new;
2533                 if (newlen != sizeof(int))
2534                         return -EINVAL;
2535                 if (get_user(new, (int __user *)newval))
2536                         return -EFAULT;
2537                 *(int *)(table->data) = msecs_to_jiffies(new);
2538         }
2539         return 1;
2540 }
2541
2542
2543
2544 #else /* CONFIG_SYSCTL_SYSCALL */
2545
2546
2547 asmlinkage long sys_sysctl(struct __sysctl_args __user *args)
2548 {
2549         static int msg_count;
2550         struct __sysctl_args tmp;
2551         int name[CTL_MAXNAME];
2552         int i;
2553
2554         /* Read in the sysctl name for better debug message logging */
2555         if (copy_from_user(&tmp, args, sizeof(tmp)))
2556                 return -EFAULT;
2557         if (tmp.nlen <= 0 || tmp.nlen >= CTL_MAXNAME)
2558                 return -ENOTDIR;
2559         for (i = 0; i < tmp.nlen; i++)
2560                 if (get_user(name[i], tmp.name + i))
2561                         return -EFAULT;
2562
2563         /* Ignore accesses to kernel.version */
2564         if ((tmp.nlen == 2) && (name[0] == CTL_KERN) && (name[1] == KERN_VERSION))
2565                 goto out;
2566
2567         if (msg_count < 5) {
2568                 msg_count++;
2569                 printk(KERN_INFO
2570                         "warning: process `%s' used the removed sysctl "
2571                         "system call with ", current->comm);
2572                 for (i = 0; i < tmp.nlen; i++)
2573                         printk("%d.", name[i]);
2574                 printk("\n");
2575         }
2576 out:
2577         return -ENOSYS;
2578 }
2579
2580 int sysctl_data(struct ctl_table *table, int __user *name, int nlen,
2581                   void __user *oldval, size_t __user *oldlenp,
2582                   void __user *newval, size_t newlen)
2583 {
2584         return -ENOSYS;
2585 }
2586
2587 int sysctl_string(struct ctl_table *table, int __user *name, int nlen,
2588                   void __user *oldval, size_t __user *oldlenp,
2589                   void __user *newval, size_t newlen)
2590 {
2591         return -ENOSYS;
2592 }
2593
2594 int sysctl_intvec(struct ctl_table *table, int __user *name, int nlen,
2595                 void __user *oldval, size_t __user *oldlenp,
2596                 void __user *newval, size_t newlen)
2597 {
2598         return -ENOSYS;
2599 }
2600
2601 int sysctl_jiffies(struct ctl_table *table, int __user *name, int nlen,
2602                 void __user *oldval, size_t __user *oldlenp,
2603                 void __user *newval, size_t newlen)
2604 {
2605         return -ENOSYS;
2606 }
2607
2608 int sysctl_ms_jiffies(struct ctl_table *table, int __user *name, int nlen,
2609                 void __user *oldval, size_t __user *oldlenp,
2610                 void __user *newval, size_t newlen)
2611 {
2612         return -ENOSYS;
2613 }
2614
2615 #endif /* CONFIG_SYSCTL_SYSCALL */
2616
2617 /*
2618  * No sense putting this after each symbol definition, twice,
2619  * exception granted :-)
2620  */
2621 EXPORT_SYMBOL(proc_dointvec);
2622 EXPORT_SYMBOL(proc_dointvec_jiffies);
2623 EXPORT_SYMBOL(proc_dointvec_minmax);
2624 EXPORT_SYMBOL(proc_dointvec_userhz_jiffies);
2625 EXPORT_SYMBOL(proc_dointvec_ms_jiffies);
2626 EXPORT_SYMBOL(proc_dostring);
2627 EXPORT_SYMBOL(proc_doulongvec_minmax);
2628 EXPORT_SYMBOL(proc_doulongvec_ms_jiffies_minmax);
2629 EXPORT_SYMBOL(register_sysctl_table);
2630 EXPORT_SYMBOL(sysctl_intvec);
2631 EXPORT_SYMBOL(sysctl_jiffies);
2632 EXPORT_SYMBOL(sysctl_ms_jiffies);
2633 EXPORT_SYMBOL(sysctl_string);
2634 EXPORT_SYMBOL(sysctl_data);
2635 EXPORT_SYMBOL(unregister_sysctl_table);