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

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