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