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