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/config.h>
  22 #include <linux/module.h>
  23 #include <linux/mm.h>
  24 #include <linux/swap.h>
  25 #include <linux/slab.h>
  26 #include <linux/sysctl.h>
  27 #include <linux/proc_fs.h>
  28 #include <linux/capability.h>
  29 #include <linux/ctype.h>
  30 #include <linux/utsname.h>
  31 #include <linux/capability.h>
  32 #include <linux/smp_lock.h>
  33 #include <linux/init.h>
  34 #include <linux/kernel.h>
  35 #include <linux/kobject.h>
  36 #include <linux/net.h>
  37 #include <linux/sysrq.h>
  38 #include <linux/highuid.h>
  39 #include <linux/writeback.h>
  40 #include <linux/hugetlb.h>
  41 #include <linux/security.h>
  42 #include <linux/initrd.h>
  43 #include <linux/times.h>
  44 #include <linux/limits.h>
  45 #include <linux/dcache.h>
  46 #include <linux/syscalls.h>
  47 #include <linux/nfs_fs.h>
  48 #include <linux/acpi.h>
  49 #include <linux/vserver/cvirt.h>
  50
  51 #include <asm/uaccess.h>
  52 #include <asm/processor.h>
  53
  54 extern int proc_nr_files(ctl_table *table, int write, struct file *filp,
  55                      void __user *buffer, size_t *lenp, loff_t *ppos);
  56
  57 #if defined(CONFIG_SYSCTL)
  58
  59 /* External variables not in a header file. */
  60 extern int C_A_D;
  61 extern int sysctl_overcommit_memory;
  62 extern int sysctl_overcommit_ratio;
  63 extern int max_threads;
  64 extern int sysrq_enabled;
  65 extern int core_uses_pid;
  66 extern int suid_dumpable;
  67 extern char core_pattern[];
  68 extern int cad_pid;
  69 extern int pid_max;
  70 extern int min_free_kbytes;
  71 extern int printk_ratelimit_jiffies;
  72 extern int printk_ratelimit_burst;
  73 extern int pid_max_min, pid_max_max;
  74 extern int sysctl_drop_caches;
  75 extern int percpu_pagelist_fraction;
  76
  77 #if defined(CONFIG_X86_LOCAL_APIC) && defined(CONFIG_X86)
  78 int unknown_nmi_panic;
  79 extern int proc_unknown_nmi_panic(ctl_table *, int, struct file *,
  80                                   void __user *, size_t *, loff_t *);
  81 #endif
  82
  83 extern unsigned int vdso_enabled, vdso_populate;
  84
  85 int exec_shield = (1<<3) | (1<<0);
  86 /* exec_shield is a bitmask:
  87           0: off; vdso at STACK_TOP, 1 page below TASK_SIZE
  88    (1<<0) 1: on [also on if !=0]
  89    (1<<1) 2: noexecstack by default
  90    (1<<2) 4: vdso just below .text of main (unless too low)
  91    (1<<3) 8: vdso just below .text of PT_INTERP (unless too low)
  92 Yes, vdso placement is overloaded here; but exec_shield off
  93 is a strong incentive to place vdso at STACK_TOP, so the bit
  94 for vdso just below .text comes along for the ride.
  95 */
  96
  97 static int __init setup_exec_shield(char *str)
  98 {
  99         get_option (&str, &exec_shield);
 100
 101         return 1;
 102 }
 103
 104 __setup("exec-shield=", setup_exec_shield);
 105
 106 /* this is needed for the proc_dointvec_minmax for [fs_]overflow UID and GID */
 107 static int maxolduid = 65535;
 108 static int minolduid;
 109 static int min_percpu_pagelist_fract = 8;
 110
 111 static int ngroups_max = NGROUPS_MAX;
 112
 113 #ifdef CONFIG_KMOD
 114 extern char modprobe_path[];
 115 #endif
 116 extern char vshelper_path[];
 117 #ifdef CONFIG_CHR_DEV_SG
 118 extern int sg_big_buff;
 119 #endif
 120 #ifdef CONFIG_SYSVIPC
 121 extern size_t shm_ctlmax;
 122 extern size_t shm_ctlall;
 123 extern int shm_ctlmni;
 124 extern int msg_ctlmax;
 125 extern int msg_ctlmnb;
 126 extern int msg_ctlmni;
 127 extern int sem_ctls[];
 128 #endif
 129
 130 #ifdef __sparc__
 131 extern char reboot_command [];
 132 extern int stop_a_enabled;
 133 extern int scons_pwroff;
 134 #endif
 135
 136 #ifdef __hppa__
 137 extern int pwrsw_enabled;
 138 extern int unaligned_enabled;
 139 #endif
 140
 141 #ifdef CONFIG_S390
 142 #ifdef CONFIG_MATHEMU
 143 extern int sysctl_ieee_emulation_warnings;
 144 #endif
 145 extern int sysctl_userprocess_debug;
 146 extern int spin_retry;
 147 #endif
 148
 149 extern int sysctl_hz_timer;
 150
 151 #ifdef CONFIG_BSD_PROCESS_ACCT
 152 extern int acct_parm[];
 153 #endif
 154
 155 #ifdef CONFIG_IA64
 156 extern int no_unaligned_warning;
 157 #endif
 158
 159 static int parse_table(int __user *, int, void __user *, size_t __user *, void __user *, size_t,
 160                        ctl_table *, void **);
 161 static int proc_doutsstring(ctl_table *table, int write, struct file *filp,
 162                   void __user *buffer, size_t *lenp, loff_t *ppos);
 163
 164 static ctl_table root_table[];
 165 static struct ctl_table_header root_table_header =
 166         { root_table, LIST_HEAD_INIT(root_table_header.ctl_entry) };
 167
 168 static ctl_table kern_table[];
 169 static ctl_table vm_table[];
 170 static ctl_table proc_table[];
 171 static ctl_table fs_table[];
 172 static ctl_table debug_table[];
 173 static ctl_table dev_table[];
 174 extern ctl_table random_table[];
 175 #ifdef CONFIG_UNIX98_PTYS
 176 extern ctl_table pty_table[];
 177 #endif
 178 #ifdef CONFIG_INOTIFY
 179 extern ctl_table inotify_table[];
 180 #endif
 181
 182 #ifdef HAVE_ARCH_PICK_MMAP_LAYOUT
 183 int sysctl_legacy_va_layout;
 184 #endif
 185
 186 /* /proc declarations: */
 187
 188 #ifdef CONFIG_PROC_FS
 189
 190 static ssize_t proc_readsys(struct file *, char __user *, size_t, loff_t *);
 191 static ssize_t proc_writesys(struct file *, const char __user *, size_t, loff_t *);
 192 static int proc_opensys(struct inode *, struct file *);
 193
 194 struct file_operations proc_sys_file_operations = {
 195         .open           = proc_opensys,
 196         .read           = proc_readsys,
 197         .write          = proc_writesys,
 198 };
 199
 200 extern struct proc_dir_entry *proc_sys_root;
 201
 202 static void register_proc_table(ctl_table *, struct proc_dir_entry *, void *);
 203 static void unregister_proc_table(ctl_table *, struct proc_dir_entry *);
 204 #endif
 205
 206 /* The default sysctl tables: */
 207
 208 static ctl_table root_table[] = {
 209         {
 210                 .ctl_name       = CTL_KERN,
 211                 .procname       = "kernel",
 212                 .mode           = 0555,
 213                 .child          = kern_table,
 214         },
 215         {
 216                 .ctl_name       = CTL_VM,
 217                 .procname       = "vm",
 218                 .mode           = 0555,
 219                 .child          = vm_table,
 220         },
 221 #ifdef CONFIG_NET
 222         {
 223                 .ctl_name       = CTL_NET,
 224                 .procname       = "net",
 225                 .mode           = 0555,
 226                 .child          = net_table,
 227         },
 228 #endif
 229         {
 230                 .ctl_name       = CTL_PROC,
 231                 .procname       = "proc",
 232                 .mode           = 0555,
 233                 .child          = proc_table,
 234         },
 235         {
 236                 .ctl_name       = CTL_FS,
 237                 .procname       = "fs",
 238                 .mode           = 0555,
 239                 .child          = fs_table,
 240         },
 241         {
 242                 .ctl_name       = CTL_DEBUG,
 243                 .procname       = "debug",
 244                 .mode           = 0555,
 245                 .child          = debug_table,
 246         },
 247         {
 248                 .ctl_name       = CTL_DEV,
 249                 .procname       = "dev",
 250                 .mode           = 0555,
 251                 .child          = dev_table,
 252         },
 253
 254         { .ctl_name = 0 }
 255 };
 256
 257 static ctl_table kern_table[] = {
 258         {
 259                 .ctl_name       = KERN_OSTYPE,
 260                 .procname       = "ostype",
 261                 .data           = system_utsname.sysname,
 262                 .maxlen         = sizeof(system_utsname.sysname),
 263                 .mode           = 0444,
 264                 .proc_handler   = &proc_doutsstring,
 265                 .virt_handler   = &vx_uts_virt_handler,
 266                 .strategy       = &sysctl_string,
 267         },
 268         {
 269                 .ctl_name       = KERN_OSRELEASE,
 270                 .procname       = "osrelease",
 271                 .data           = system_utsname.release,
 272                 .maxlen         = sizeof(system_utsname.release),
 273                 .mode           = 0444,
 274                 .proc_handler   = &proc_doutsstring,
 275                 .virt_handler   = &vx_uts_virt_handler,
 276                 .strategy       = &sysctl_string,
 277         },
 278         {
 279                 .ctl_name       = KERN_VERSION,
 280                 .procname       = "version",
 281                 .data           = system_utsname.version,
 282                 .maxlen         = sizeof(system_utsname.version),
 283                 .mode           = 0444,
 284                 .proc_handler   = &proc_doutsstring,
 285                 .virt_handler   = &vx_uts_virt_handler,
 286                 .strategy       = &sysctl_string,
 287         },
 288         {
 289                 .ctl_name       = KERN_NODENAME,
 290                 .procname       = "hostname",
 291                 .data           = system_utsname.nodename,
 292                 .maxlen         = sizeof(system_utsname.nodename),
 293                 .mode           = 0644,
 294                 .proc_handler   = &proc_doutsstring,
 295                 .virt_handler   = &vx_uts_virt_handler,
 296                 .strategy       = &sysctl_string,
 297         },
 298         {
 299                 .ctl_name       = KERN_DOMAINNAME,
 300                 .procname       = "domainname",
 301                 .data           = system_utsname.domainname,
 302                 .maxlen         = sizeof(system_utsname.domainname),
 303                 .mode           = 0644,
 304                 .proc_handler   = &proc_doutsstring,
 305                 .virt_handler   = &vx_uts_virt_handler,
 306                 .strategy       = &sysctl_string,
 307         },
 308         {
 309                 .ctl_name       = KERN_PANIC,
 310                 .procname       = "panic",
 311                 .data           = &panic_timeout,
 312                 .maxlen         = sizeof(int),
 313                 .mode           = 0644,
 314                 .proc_handler   = &proc_dointvec,
 315         },
 316         {
 317                 .ctl_name       = KERN_EXEC_SHIELD,
 318                 .procname       = "exec-shield",
 319                 .data           = &exec_shield,
 320                 .maxlen         = sizeof(int),
 321                 .mode           = 0644,
 322                 .proc_handler   = &proc_dointvec,
 323         },
 324         {
 325                 .ctl_name       = KERN_PRINT_FATAL,
 326                 .procname       = "print-fatal-signals",
 327                 .data           = &print_fatal_signals,
 328                 .maxlen         = sizeof(int),
 329                 .mode           = 0644,
 330                 .proc_handler   = &proc_dointvec,
 331         },
 332 #ifdef __i386__
 333         {
 334                 .ctl_name       = KERN_VDSO,
 335                 .procname       = "vdso",
 336                 .data           = &vdso_enabled,
 337                 .maxlen         = sizeof(int),
 338                 .mode           = 0644,
 339                 .proc_handler   = &proc_dointvec,
 340         },
 341         {
 342                 .ctl_name       = KERN_VDSO,
 343                 .procname       = "vdso_populate",
 344                 .data           = &vdso_populate,
 345                 .maxlen         = sizeof(int),
 346                 .mode           = 0644,
 347                 .proc_handler   = &proc_dointvec,
 348         },
 349 #endif
 350         {
 351                 .ctl_name       = KERN_CORE_USES_PID,
 352                 .procname       = "core_uses_pid",
 353                 .data           = &core_uses_pid,
 354                 .maxlen         = sizeof(int),
 355                 .mode           = 0644,
 356                 .proc_handler   = &proc_dointvec,
 357         },
 358         {
 359                 .ctl_name       = KERN_CORE_PATTERN,
 360                 .procname       = "core_pattern",
 361                 .data           = core_pattern,
 362                 .maxlen         = 64,
 363                 .mode           = 0644,
 364                 .proc_handler   = &proc_dostring,
 365                 .strategy       = &sysctl_string,
 366         },
 367         {
 368                 .ctl_name       = KERN_TAINTED,
 369                 .procname       = "tainted",
 370                 .data           = &tainted,
 371                 .maxlen         = sizeof(int),
 372                 .mode           = 0444,
 373                 .proc_handler   = &proc_dointvec,
 374         },
 375         {
 376                 .ctl_name       = KERN_CAP_BSET,
 377                 .procname       = "cap-bound",
 378                 .data           = &cap_bset,
 379                 .maxlen         = sizeof(kernel_cap_t),
 380                 .mode           = 0600,
 381                 .proc_handler   = &proc_dointvec_bset,
 382         },
 383 #ifdef CONFIG_BLK_DEV_INITRD
 384         {
 385                 .ctl_name       = KERN_REALROOTDEV,
 386                 .procname       = "real-root-dev",
 387                 .data           = &real_root_dev,
 388                 .maxlen         = sizeof(int),
 389                 .mode           = 0644,
 390                 .proc_handler   = &proc_dointvec,
 391         },
 392 #endif
 393 #ifdef __sparc__
 394         {
 395                 .ctl_name       = KERN_SPARC_REBOOT,
 396                 .procname       = "reboot-cmd",
 397                 .data           = reboot_command,
 398                 .maxlen         = 256,
 399                 .mode           = 0644,
 400                 .proc_handler   = &proc_dostring,
 401                 .strategy       = &sysctl_string,
 402         },
 403         {
 404                 .ctl_name       = KERN_SPARC_STOP_A,
 405                 .procname       = "stop-a",
 406                 .data           = &stop_a_enabled,
 407                 .maxlen         = sizeof (int),
 408                 .mode           = 0644,
 409                 .proc_handler   = &proc_dointvec,
 410         },
 411         {
 412                 .ctl_name       = KERN_SPARC_SCONS_PWROFF,
 413                 .procname       = "scons-poweroff",
 414                 .data           = &scons_pwroff,
 415                 .maxlen         = sizeof (int),
 416                 .mode           = 0644,
 417                 .proc_handler   = &proc_dointvec,
 418         },
 419 #endif
 420 #ifdef __hppa__
 421         {
 422                 .ctl_name       = KERN_HPPA_PWRSW,
 423                 .procname       = "soft-power",
 424                 .data           = &pwrsw_enabled,
 425                 .maxlen         = sizeof (int),
 426                 .mode           = 0644,
 427                 .proc_handler   = &proc_dointvec,
 428         },
 429         {
 430                 .ctl_name       = KERN_HPPA_UNALIGNED,
 431                 .procname       = "unaligned-trap",
 432                 .data           = &unaligned_enabled,
 433                 .maxlen         = sizeof (int),
 434                 .mode           = 0644,
 435                 .proc_handler   = &proc_dointvec,
 436         },
 437 #endif
 438         {
 439                 .ctl_name       = KERN_CTLALTDEL,
 440                 .procname       = "ctrl-alt-del",
 441                 .data           = &C_A_D,
 442                 .maxlen         = sizeof(int),
 443                 .mode           = 0644,
 444                 .proc_handler   = &proc_dointvec,
 445         },
 446         {
 447                 .ctl_name       = KERN_PRINTK,
 448                 .procname       = "printk",
 449                 .data           = &console_loglevel,
 450                 .maxlen         = 4*sizeof(int),
 451                 .mode           = 0644,
 452                 .proc_handler   = &proc_dointvec,
 453         },
 454 #ifdef CONFIG_KMOD
 455         {
 456                 .ctl_name       = KERN_MODPROBE,
 457                 .procname       = "modprobe",
 458                 .data           = &modprobe_path,
 459                 .maxlen         = KMOD_PATH_LEN,
 460                 .mode           = 0644,
 461                 .proc_handler   = &proc_dostring,
 462                 .strategy       = &sysctl_string,
 463         },
 464 #endif
 465 #ifdef CONFIG_HOTPLUG
 466         {
 467                 .ctl_name       = KERN_HOTPLUG,
 468                 .procname       = "hotplug",
 469                 .data           = &uevent_helper,
 470                 .maxlen         = UEVENT_HELPER_PATH_LEN,
 471                 .mode           = 0644,
 472                 .proc_handler   = &proc_dostring,
 473                 .strategy       = &sysctl_string,
 474         },
 475 #endif
 476         {
 477                 .ctl_name       = KERN_VSHELPER,
 478                 .procname       = "vshelper",
 479                 .data           = &vshelper_path,
 480                 .maxlen         = 256,
 481                 .mode           = 0644,
 482                 .proc_handler   = &proc_dostring,
 483                 .strategy       = &sysctl_string,
 484         },
 485 #ifdef CONFIG_CHR_DEV_SG
 486         {
 487                 .ctl_name       = KERN_SG_BIG_BUFF,
 488                 .procname       = "sg-big-buff",
 489                 .data           = &sg_big_buff,
 490                 .maxlen         = sizeof (int),
 491                 .mode           = 0444,
 492                 .proc_handler   = &proc_dointvec,
 493         },
 494 #endif
 495 #ifdef CONFIG_BSD_PROCESS_ACCT
 496         {
 497                 .ctl_name       = KERN_ACCT,
 498                 .procname       = "acct",
 499                 .data           = &acct_parm,
 500                 .maxlen         = 3*sizeof(int),
 501                 .mode           = 0644,
 502                 .proc_handler   = &proc_dointvec,
 503         },
 504 #endif
 505 #ifdef CONFIG_SYSVIPC
 506         {
 507                 .ctl_name       = KERN_SHMMAX,
 508                 .procname       = "shmmax",
 509                 .data           = &shm_ctlmax,
 510                 .maxlen         = sizeof (size_t),
 511                 .mode           = 0644,
 512                 .proc_handler   = &proc_doulongvec_minmax,
 513         },
 514         {
 515                 .ctl_name       = KERN_SHMALL,
 516                 .procname       = "shmall",
 517                 .data           = &shm_ctlall,
 518                 .maxlen         = sizeof (size_t),
 519                 .mode           = 0644,
 520                 .proc_handler   = &proc_doulongvec_minmax,
 521         },
 522         {
 523                 .ctl_name       = KERN_SHMMNI,
 524                 .procname       = "shmmni",
 525                 .data           = &shm_ctlmni,
 526                 .maxlen         = sizeof (int),
 527                 .mode           = 0644,
 528                 .proc_handler   = &proc_dointvec,
 529         },
 530         {
 531                 .ctl_name       = KERN_MSGMAX,
 532                 .procname       = "msgmax",
 533                 .data           = &msg_ctlmax,
 534                 .maxlen         = sizeof (int),
 535                 .mode           = 0644,
 536                 .proc_handler   = &proc_dointvec,
 537         },
 538         {
 539                 .ctl_name       = KERN_MSGMNI,
 540                 .procname       = "msgmni",
 541                 .data           = &msg_ctlmni,
 542                 .maxlen         = sizeof (int),
 543                 .mode           = 0644,
 544                 .proc_handler   = &proc_dointvec,
 545         },
 546         {
 547                 .ctl_name       = KERN_MSGMNB,
 548                 .procname       =  "msgmnb",
 549                 .data           = &msg_ctlmnb,
 550                 .maxlen         = sizeof (int),
 551                 .mode           = 0644,
 552                 .proc_handler   = &proc_dointvec,
 553         },
 554         {
 555                 .ctl_name       = KERN_SEM,
 556                 .procname       = "sem",
 557                 .data           = &sem_ctls,
 558                 .maxlen         = 4*sizeof (int),
 559                 .mode           = 0644,
 560                 .proc_handler   = &proc_dointvec,
 561         },
 562 #endif
 563 #ifdef CONFIG_MAGIC_SYSRQ
 564         {
 565                 .ctl_name       = KERN_SYSRQ,
 566                 .procname       = "sysrq",
 567                 .data           = &sysrq_enabled,
 568                 .maxlen         = sizeof (int),
 569                 .mode           = 0644,
 570                 .proc_handler   = &proc_dointvec,
 571         },
 572 #endif
 573         {
 574                 .ctl_name       = KERN_CADPID,
 575                 .procname       = "cad_pid",
 576                 .data           = &cad_pid,
 577                 .maxlen         = sizeof (int),
 578                 .mode           = 0600,
 579                 .proc_handler   = &proc_dointvec,
 580         },
 581         {
 582                 .ctl_name       = KERN_MAX_THREADS,
 583                 .procname       = "threads-max",
 584                 .data           = &max_threads,
 585                 .maxlen         = sizeof(int),
 586                 .mode           = 0644,
 587                 .proc_handler   = &proc_dointvec,
 588         },
 589         {
 590                 .ctl_name       = KERN_RANDOM,
 591                 .procname       = "random",
 592                 .mode           = 0555,
 593                 .child          = random_table,
 594         },
 595 #ifdef CONFIG_UNIX98_PTYS
 596         {
 597                 .ctl_name       = KERN_PTY,
 598                 .procname       = "pty",
 599                 .mode           = 0555,
 600                 .child          = pty_table,
 601         },
 602 #endif
 603         {
 604                 .ctl_name       = KERN_OVERFLOWUID,
 605                 .procname       = "overflowuid",
 606                 .data           = &overflowuid,
 607                 .maxlen         = sizeof(int),
 608                 .mode           = 0644,
 609                 .proc_handler   = &proc_dointvec_minmax,
 610                 .strategy       = &sysctl_intvec,
 611                 .extra1         = &minolduid,
 612                 .extra2         = &maxolduid,
 613         },
 614         {
 615                 .ctl_name       = KERN_OVERFLOWGID,
 616                 .procname       = "overflowgid",
 617                 .data           = &overflowgid,
 618                 .maxlen         = sizeof(int),
 619                 .mode           = 0644,
 620                 .proc_handler   = &proc_dointvec_minmax,
 621                 .strategy       = &sysctl_intvec,
 622                 .extra1         = &minolduid,
 623                 .extra2         = &maxolduid,
 624         },
 625 #ifdef CONFIG_S390
 626 #ifdef CONFIG_MATHEMU
 627         {
 628                 .ctl_name       = KERN_IEEE_EMULATION_WARNINGS,
 629                 .procname       = "ieee_emulation_warnings",
 630                 .data           = &sysctl_ieee_emulation_warnings,
 631                 .maxlen         = sizeof(int),
 632                 .mode           = 0644,
 633                 .proc_handler   = &proc_dointvec,
 634         },
 635 #endif
 636 #ifdef CONFIG_NO_IDLE_HZ
 637         {
 638                 .ctl_name       = KERN_HZ_TIMER,
 639                 .procname       = "hz_timer",
 640                 .data           = &sysctl_hz_timer,
 641                 .maxlen         = sizeof(int),
 642                 .mode           = 0644,
 643                 .proc_handler   = &proc_dointvec,
 644         },
 645 #endif
 646         {
 647                 .ctl_name       = KERN_S390_USER_DEBUG_LOGGING,
 648                 .procname       = "userprocess_debug",
 649                 .data           = &sysctl_userprocess_debug,
 650                 .maxlen         = sizeof(int),
 651                 .mode           = 0644,
 652                 .proc_handler   = &proc_dointvec,
 653         },
 654 #endif
 655         {
 656                 .ctl_name       = KERN_PIDMAX,
 657                 .procname       = "pid_max",
 658                 .data           = &pid_max,
 659                 .maxlen         = sizeof (int),
 660                 .mode           = 0644,
 661                 .proc_handler   = &proc_dointvec_minmax,
 662                 .strategy       = sysctl_intvec,
 663                 .extra1         = &pid_max_min,
 664                 .extra2         = &pid_max_max,
 665         },
 666         {
 667                 .ctl_name       = KERN_PANIC_ON_OOPS,
 668                 .procname       = "panic_on_oops",
 669                 .data           = &panic_on_oops,
 670                 .maxlen         = sizeof(int),
 671                 .mode           = 0644,
 672                 .proc_handler   = &proc_dointvec,
 673         },
 674         {
 675                 .ctl_name       = KERN_PRINTK_RATELIMIT,
 676                 .procname       = "printk_ratelimit",
 677                 .data           = &printk_ratelimit_jiffies,
 678                 .maxlen         = sizeof(int),
 679                 .mode           = 0644,
 680                 .proc_handler   = &proc_dointvec_jiffies,
 681                 .strategy       = &sysctl_jiffies,
 682         },
 683         {
 684                 .ctl_name       = KERN_PRINTK_RATELIMIT_BURST,
 685                 .procname       = "printk_ratelimit_burst",
 686                 .data           = &printk_ratelimit_burst,
 687                 .maxlen         = sizeof(int),
 688                 .mode           = 0644,
 689                 .proc_handler   = &proc_dointvec,
 690         },
 691         {
 692                 .ctl_name       = KERN_NGROUPS_MAX,
 693                 .procname       = "ngroups_max",
 694                 .data           = &ngroups_max,
 695                 .maxlen         = sizeof (int),
 696                 .mode           = 0444,
 697                 .proc_handler   = &proc_dointvec,
 698         },
 699 #if defined(CONFIG_X86_LOCAL_APIC) && defined(CONFIG_X86)
 700         {
 701                 .ctl_name       = KERN_UNKNOWN_NMI_PANIC,
 702                 .procname       = "unknown_nmi_panic",
 703                 .data           = &unknown_nmi_panic,
 704                 .maxlen         = sizeof (int),
 705                 .mode           = 0644,
 706                 .proc_handler   = &proc_unknown_nmi_panic,
 707         },
 708 #endif
 709 #if defined(CONFIG_X86)
 710         {
 711                 .ctl_name       = KERN_BOOTLOADER_TYPE,
 712                 .procname       = "bootloader_type",
 713                 .data           = &bootloader_type,
 714                 .maxlen         = sizeof (int),
 715                 .mode           = 0444,
 716                 .proc_handler   = &proc_dointvec,
 717         },
 718 #endif
 719 #if defined(CONFIG_MMU)
 720         {
 721                 .ctl_name       = KERN_RANDOMIZE,
 722                 .procname       = "randomize_va_space",
 723                 .data           = &randomize_va_space,
 724                 .maxlen         = sizeof(int),
 725                 .mode           = 0644,
 726                 .proc_handler   = &proc_dointvec,
 727         },
 728 #endif
 729 #if defined(CONFIG_S390) && defined(CONFIG_SMP)
 730         {
 731                 .ctl_name       = KERN_SPIN_RETRY,
 732                 .procname       = "spin_retry",
 733                 .data           = &spin_retry,
 734                 .maxlen         = sizeof (int),
 735                 .mode           = 0644,
 736                 .proc_handler   = &proc_dointvec,
 737         },
 738 #endif
 739 #ifdef CONFIG_ACPI_SLEEP
 740         {
 741                 .ctl_name       = KERN_ACPI_VIDEO_FLAGS,
 742                 .procname       = "acpi_video_flags",
 743                 .data           = &acpi_video_flags,
 744                 .maxlen         = sizeof (unsigned long),
 745                 .mode           = 0644,
 746                 .proc_handler   = &proc_doulongvec_minmax,
 747         },
 748 #endif
 749 #ifdef CONFIG_IA64
 750         {
 751                 .ctl_name       = KERN_IA64_UNALIGNED,
 752                 .procname       = "ignore-unaligned-usertrap",
 753                 .data           = &no_unaligned_warning,
 754                 .maxlen         = sizeof (int),
 755                 .mode           = 0644,
 756                 .proc_handler   = &proc_dointvec,
 757         },
 758 #endif
 759         { .ctl_name = 0 }
 760 };
 761
 762 /* Constants for minimum and maximum testing in vm_table.
 763    We use these as one-element integer vectors. */
 764 static int zero;
 765 static int one_hundred = 100;
 766
 767
 768 static ctl_table vm_table[] = {
 769         {
 770                 .ctl_name       = VM_OVERCOMMIT_MEMORY,
 771                 .procname       = "overcommit_memory",
 772                 .data           = &sysctl_overcommit_memory,
 773                 .maxlen         = sizeof(sysctl_overcommit_memory),
 774                 .mode           = 0644,
 775                 .proc_handler   = &proc_dointvec,
 776         },
 777         {
 778                 .ctl_name       = VM_OVERCOMMIT_RATIO,
 779                 .procname       = "overcommit_ratio",
 780                 .data           = &sysctl_overcommit_ratio,
 781                 .maxlen         = sizeof(sysctl_overcommit_ratio),
 782                 .mode           = 0644,
 783                 .proc_handler   = &proc_dointvec,
 784         },
 785         {
 786                 .ctl_name       = VM_PAGE_CLUSTER,
 787                 .procname       = "page-cluster",
 788                 .data           = &page_cluster,
 789                 .maxlen         = sizeof(int),
 790                 .mode           = 0644,
 791                 .proc_handler   = &proc_dointvec,
 792         },
 793         {
 794                 .ctl_name       = VM_DIRTY_BACKGROUND,
 795                 .procname       = "dirty_background_ratio",
 796                 .data           = &dirty_background_ratio,
 797                 .maxlen         = sizeof(dirty_background_ratio),
 798                 .mode           = 0644,
 799                 .proc_handler   = &proc_dointvec_minmax,
 800                 .strategy       = &sysctl_intvec,
 801                 .extra1         = &zero,
 802                 .extra2         = &one_hundred,
 803         },
 804         {
 805                 .ctl_name       = VM_DIRTY_RATIO,
 806                 .procname       = "dirty_ratio",
 807                 .data           = &vm_dirty_ratio,
 808                 .maxlen         = sizeof(vm_dirty_ratio),
 809                 .mode           = 0644,
 810                 .proc_handler   = &proc_dointvec_minmax,
 811                 .strategy       = &sysctl_intvec,
 812                 .extra1         = &zero,
 813                 .extra2         = &one_hundred,
 814         },
 815         {
 816                 .ctl_name       = VM_DIRTY_WB_CS,
 817                 .procname       = "dirty_writeback_centisecs",
 818                 .data           = &dirty_writeback_interval,
 819                 .maxlen         = sizeof(dirty_writeback_interval),
 820                 .mode           = 0644,
 821                 .proc_handler   = &dirty_writeback_centisecs_handler,
 822         },
 823         {
 824                 .ctl_name       = VM_DIRTY_EXPIRE_CS,
 825                 .procname       = "dirty_expire_centisecs",
 826                 .data           = &dirty_expire_interval,
 827                 .maxlen         = sizeof(dirty_expire_interval),
 828                 .mode           = 0644,
 829                 .proc_handler   = &proc_dointvec_userhz_jiffies,
 830         },
 831         {
 832                 .ctl_name       = VM_NR_PDFLUSH_THREADS,
 833                 .procname       = "nr_pdflush_threads",
 834                 .data           = &nr_pdflush_threads,
 835                 .maxlen         = sizeof nr_pdflush_threads,
 836                 .mode           = 0444 /* read-only*/,
 837                 .proc_handler   = &proc_dointvec,
 838         },
 839         {
 840                 .ctl_name       = VM_SWAPPINESS,
 841                 .procname       = "swappiness",
 842                 .data           = &vm_swappiness,
 843                 .maxlen         = sizeof(vm_swappiness),
 844                 .mode           = 0644,
 845                 .proc_handler   = &proc_dointvec_minmax,
 846                 .strategy       = &sysctl_intvec,
 847                 .extra1         = &zero,
 848                 .extra2         = &one_hundred,
 849         },
 850 #ifdef CONFIG_HUGETLB_PAGE
 851          {
 852                 .ctl_name       = VM_HUGETLB_PAGES,
 853                 .procname       = "nr_hugepages",
 854                 .data           = &max_huge_pages,
 855                 .maxlen         = sizeof(unsigned long),
 856                 .mode           = 0644,
 857                 .proc_handler   = &hugetlb_sysctl_handler,
 858                 .extra1         = (void *)&hugetlb_zero,
 859                 .extra2         = (void *)&hugetlb_infinity,
 860          },
 861          {
 862                 .ctl_name       = VM_HUGETLB_GROUP,
 863                 .procname       = "hugetlb_shm_group",
 864                 .data           = &sysctl_hugetlb_shm_group,
 865                 .maxlen         = sizeof(gid_t),
 866                 .mode           = 0644,
 867                 .proc_handler   = &proc_dointvec,
 868          },
 869 #endif
 870         {
 871                 .ctl_name       = VM_LOWMEM_RESERVE_RATIO,
 872                 .procname       = "lowmem_reserve_ratio",
 873                 .data           = &sysctl_lowmem_reserve_ratio,
 874                 .maxlen         = sizeof(sysctl_lowmem_reserve_ratio),
 875                 .mode           = 0644,
 876                 .proc_handler   = &lowmem_reserve_ratio_sysctl_handler,
 877                 .strategy       = &sysctl_intvec,
 878         },
 879         {
 880                 .ctl_name       = VM_DROP_PAGECACHE,
 881                 .procname       = "drop_caches",
 882                 .data           = &sysctl_drop_caches,
 883                 .maxlen         = sizeof(int),
 884                 .mode           = 0644,
 885                 .proc_handler   = drop_caches_sysctl_handler,
 886                 .strategy       = &sysctl_intvec,
 887         },
 888         {
 889                 .ctl_name       = VM_MIN_FREE_KBYTES,
 890                 .procname       = "min_free_kbytes",
 891                 .data           = &min_free_kbytes,
 892                 .maxlen         = sizeof(min_free_kbytes),
 893                 .mode           = 0644,
 894                 .proc_handler   = &min_free_kbytes_sysctl_handler,
 895                 .strategy       = &sysctl_intvec,
 896                 .extra1         = &zero,
 897         },
 898         {
 899                 .ctl_name       = VM_PERCPU_PAGELIST_FRACTION,
 900                 .procname       = "percpu_pagelist_fraction",
 901                 .data           = &percpu_pagelist_fraction,
 902                 .maxlen         = sizeof(percpu_pagelist_fraction),
 903                 .mode           = 0644,
 904                 .proc_handler   = &percpu_pagelist_fraction_sysctl_handler,
 905                 .strategy       = &sysctl_intvec,
 906                 .extra1         = &min_percpu_pagelist_fract,
 907         },
 908 #ifdef CONFIG_MMU
 909         {
 910                 .ctl_name       = VM_MAX_MAP_COUNT,
 911                 .procname       = "max_map_count",
 912                 .data           = &sysctl_max_map_count,
 913                 .maxlen         = sizeof(sysctl_max_map_count),
 914                 .mode           = 0644,
 915                 .proc_handler   = &proc_dointvec
 916         },
 917 #endif
 918         {
 919                 .ctl_name       = VM_LAPTOP_MODE,
 920                 .procname       = "laptop_mode",
 921                 .data           = &laptop_mode,
 922                 .maxlen         = sizeof(laptop_mode),
 923                 .mode           = 0644,
 924                 .proc_handler   = &proc_dointvec_jiffies,
 925                 .strategy       = &sysctl_jiffies,
 926         },
 927         {
 928                 .ctl_name       = VM_BLOCK_DUMP,
 929                 .procname       = "block_dump",
 930                 .data           = &block_dump,
 931                 .maxlen         = sizeof(block_dump),
 932                 .mode           = 0644,
 933                 .proc_handler   = &proc_dointvec,
 934                 .strategy       = &sysctl_intvec,
 935                 .extra1         = &zero,
 936         },
 937         {
 938                 .ctl_name       = VM_VFS_CACHE_PRESSURE,
 939                 .procname       = "vfs_cache_pressure",
 940                 .data           = &sysctl_vfs_cache_pressure,
 941                 .maxlen         = sizeof(sysctl_vfs_cache_pressure),
 942                 .mode           = 0644,
 943                 .proc_handler   = &proc_dointvec,
 944                 .strategy       = &sysctl_intvec,
 945                 .extra1         = &zero,
 946         },
 947 #ifdef HAVE_ARCH_PICK_MMAP_LAYOUT
 948         {
 949                 .ctl_name       = VM_LEGACY_VA_LAYOUT,
 950                 .procname       = "legacy_va_layout",
 951                 .data           = &sysctl_legacy_va_layout,
 952                 .maxlen         = sizeof(sysctl_legacy_va_layout),
 953                 .mode           = 0644,
 954                 .proc_handler   = &proc_dointvec,
 955                 .strategy       = &sysctl_intvec,
 956                 .extra1         = &zero,
 957         },
 958 #endif
 959 #ifdef CONFIG_SWAP
 960         {
 961                 .ctl_name       = VM_SWAP_TOKEN_TIMEOUT,
 962                 .procname       = "swap_token_timeout",
 963                 .data           = &swap_token_default_timeout,
 964                 .maxlen         = sizeof(swap_token_default_timeout),
 965                 .mode           = 0644,
 966                 .proc_handler   = &proc_dointvec_jiffies,
 967                 .strategy       = &sysctl_jiffies,
 968         },
 969 #endif
 970 #ifdef CONFIG_NUMA
 971         {
 972                 .ctl_name       = VM_ZONE_RECLAIM_MODE,
 973                 .procname       = "zone_reclaim_mode",
 974                 .data           = &zone_reclaim_mode,
 975                 .maxlen         = sizeof(zone_reclaim_mode),
 976                 .mode           = 0644,
 977                 .proc_handler   = &proc_dointvec,
 978                 .strategy       = &sysctl_intvec,
 979                 .extra1         = &zero,
 980         },
 981         {
 982                 .ctl_name       = VM_ZONE_RECLAIM_INTERVAL,
 983                 .procname       = "zone_reclaim_interval",
 984                 .data           = &zone_reclaim_interval,
 985                 .maxlen         = sizeof(zone_reclaim_interval),
 986                 .mode           = 0644,
 987                 .proc_handler   = &proc_dointvec_jiffies,
 988                 .strategy       = &sysctl_jiffies,
 989         },
 990 #endif
 991         { .ctl_name = 0 }
 992 };
 993
 994 static ctl_table proc_table[] = {
 995         { .ctl_name = 0 }
 996 };
 997
 998 static ctl_table fs_table[] = {
 999         {
1000                 .ctl_name       = FS_NRINODE,
1001                 .procname       = "inode-nr",
1002                 .data           = &inodes_stat,
1003                 .maxlen         = 2*sizeof(int),
1004                 .mode           = 0444,
1005                 .proc_handler   = &proc_dointvec,
1006         },
1007         {
1008                 .ctl_name       = FS_STATINODE,
1009                 .procname       = "inode-state",
1010                 .data           = &inodes_stat,
1011                 .maxlen         = 7*sizeof(int),
1012                 .mode           = 0444,
1013                 .proc_handler   = &proc_dointvec,
1014         },
1015         {
1016                 .ctl_name       = FS_NRFILE,
1017                 .procname       = "file-nr",
1018                 .data           = &files_stat,
1019                 .maxlen         = 3*sizeof(int),
1020                 .mode           = 0444,
1021                 .proc_handler   = &proc_nr_files,
1022         },
1023         {
1024                 .ctl_name       = FS_MAXFILE,
1025                 .procname       = "file-max",
1026                 .data           = &files_stat.max_files,
1027                 .maxlen         = sizeof(int),
1028                 .mode           = 0644,
1029                 .proc_handler   = &proc_dointvec,
1030         },
1031         {
1032                 .ctl_name       = FS_DENTRY,
1033                 .procname       = "dentry-state",
1034                 .data           = &dentry_stat,
1035                 .maxlen         = 6*sizeof(int),
1036                 .mode           = 0444,
1037                 .proc_handler   = &proc_dointvec,
1038         },
1039         {
1040                 .ctl_name       = FS_OVERFLOWUID,
1041                 .procname       = "overflowuid",
1042                 .data           = &fs_overflowuid,
1043                 .maxlen         = sizeof(int),
1044                 .mode           = 0644,
1045                 .proc_handler   = &proc_dointvec_minmax,
1046                 .strategy       = &sysctl_intvec,
1047                 .extra1         = &minolduid,
1048                 .extra2         = &maxolduid,
1049         },
1050         {
1051                 .ctl_name       = FS_OVERFLOWGID,
1052                 .procname       = "overflowgid",
1053                 .data           = &fs_overflowgid,
1054                 .maxlen         = sizeof(int),
1055                 .mode           = 0644,
1056                 .proc_handler   = &proc_dointvec_minmax,
1057                 .strategy       = &sysctl_intvec,
1058                 .extra1         = &minolduid,
1059                 .extra2         = &maxolduid,
1060         },
1061         {
1062                 .ctl_name       = FS_LEASES,
1063                 .procname       = "leases-enable",
1064                 .data           = &leases_enable,
1065                 .maxlen         = sizeof(int),
1066                 .mode           = 0644,
1067                 .proc_handler   = &proc_dointvec,
1068         },
1069 #ifdef CONFIG_DNOTIFY
1070         {
1071                 .ctl_name       = FS_DIR_NOTIFY,
1072                 .procname       = "dir-notify-enable",
1073                 .data           = &dir_notify_enable,
1074                 .maxlen         = sizeof(int),
1075                 .mode           = 0644,
1076                 .proc_handler   = &proc_dointvec,
1077         },
1078 #endif
1079 #ifdef CONFIG_MMU
1080         {
1081                 .ctl_name       = FS_LEASE_TIME,
1082                 .procname       = "lease-break-time",
1083                 .data           = &lease_break_time,
1084                 .maxlen         = sizeof(int),
1085                 .mode           = 0644,
1086                 .proc_handler   = &proc_dointvec,
1087         },
1088         {
1089                 .ctl_name       = FS_AIO_NR,
1090                 .procname       = "aio-nr",
1091                 .data           = &aio_nr,
1092                 .maxlen         = sizeof(aio_nr),
1093                 .mode           = 0444,
1094                 .proc_handler   = &proc_doulongvec_minmax,
1095         },
1096         {
1097                 .ctl_name       = FS_AIO_MAX_NR,
1098                 .procname       = "aio-max-nr",
1099                 .data           = &aio_max_nr,
1100                 .maxlen         = sizeof(aio_max_nr),
1101                 .mode           = 0644,
1102                 .proc_handler   = &proc_doulongvec_minmax,
1103         },
1104 #ifdef CONFIG_INOTIFY
1105         {
1106                 .ctl_name       = FS_INOTIFY,
1107                 .procname       = "inotify",
1108                 .mode           = 0555,
1109                 .child          = inotify_table,
1110         },
1111 #endif
1112 #endif
1113         {
1114                 .ctl_name       = KERN_SETUID_DUMPABLE,
1115                 .procname       = "suid_dumpable",
1116                 .data           = &suid_dumpable,
1117                 .maxlen         = sizeof(int),
1118                 .mode           = 0644,
1119                 .proc_handler   = &proc_dointvec,
1120         },
1121         { .ctl_name = 0 }
1122 };
1123
1124 static ctl_table debug_table[] = {
1125         { .ctl_name = 0 }
1126 };
1127
1128 static ctl_table dev_table[] = {
1129         { .ctl_name = 0 }
1130 };
1131
1132 extern void init_irq_proc (void);
1133
1134 static DEFINE_SPINLOCK(sysctl_lock);
1135
1136 /* called under sysctl_lock */
1137 static int use_table(struct ctl_table_header *p)
1138 {
1139         if (unlikely(p->unregistering))
1140                 return 0;
1141         p->used++;
1142         return 1;
1143 }
1144
1145 /* called under sysctl_lock */
1146 static void unuse_table(struct ctl_table_header *p)
1147 {
1148         if (!--p->used)
1149                 if (unlikely(p->unregistering))
1150                         complete(p->unregistering);
1151 }
1152
1153 /* called under sysctl_lock, will reacquire if has to wait */
1154 static void start_unregistering(struct ctl_table_header *p)
1155 {
1156         /*
1157          * if p->used is 0, nobody will ever touch that entry again;
1158          * we'll eliminate all paths to it before dropping sysctl_lock
1159          */
1160         if (unlikely(p->used)) {
1161                 struct completion wait;
1162                 init_completion(&wait);
1163                 p->unregistering = &wait;
1164                 spin_unlock(&sysctl_lock);
1165                 wait_for_completion(&wait);
1166                 spin_lock(&sysctl_lock);
1167         }
1168         /*
1169          * do not remove from the list until nobody holds it; walking the
1170          * list in do_sysctl() relies on that.
1171          */
1172         list_del_init(&p->ctl_entry);
1173 }
1174
1175 void __init sysctl_init(void)
1176 {
1177 #ifdef CONFIG_PROC_FS
1178         register_proc_table(root_table, proc_sys_root, &root_table_header);
1179         init_irq_proc();
1180 #endif
1181 }
1182
1183 int do_sysctl(int __user *name, int nlen, void __user *oldval, size_t __user *oldlenp,
1184                void __user *newval, size_t newlen)
1185 {
1186         struct list_head *tmp;
1187         int error = -ENOTDIR;
1188
1189         if (nlen <= 0 || nlen >= CTL_MAXNAME)
1190                 return -ENOTDIR;
1191         if (oldval) {
1192                 int old_len;
1193                 if (!oldlenp || get_user(old_len, oldlenp))
1194                         return -EFAULT;
1195         }
1196         spin_lock(&sysctl_lock);
1197         tmp = &root_table_header.ctl_entry;
1198         do {
1199                 struct ctl_table_header *head =
1200                         list_entry(tmp, struct ctl_table_header, ctl_entry);
1201                 void *context = NULL;
1202
1203                 if (!use_table(head))
1204                         continue;
1205
1206                 spin_unlock(&sysctl_lock);
1207
1208                 error = parse_table(name, nlen, oldval, oldlenp,
1209                                         newval, newlen, head->ctl_table,
1210                                         &context);
1211                 kfree(context);
1212
1213                 spin_lock(&sysctl_lock);
1214                 unuse_table(head);
1215                 if (error != -ENOTDIR)
1216                         break;
1217         } while ((tmp = tmp->next) != &root_table_header.ctl_entry);
1218         spin_unlock(&sysctl_lock);
1219         return error;
1220 }
1221
1222 asmlinkage long sys_sysctl(struct __sysctl_args __user *args)
1223 {
1224         struct __sysctl_args tmp;
1225         int error;
1226
1227         if (copy_from_user(&tmp, args, sizeof(tmp)))
1228                 return -EFAULT;
1229
1230         lock_kernel();
1231         error = do_sysctl(tmp.name, tmp.nlen, tmp.oldval, tmp.oldlenp,
1232                           tmp.newval, tmp.newlen);
1233         unlock_kernel();
1234         return error;
1235 }
1236
1237 /*
1238  * ctl_perm does NOT grant the superuser all rights automatically, because
1239  * some sysctl variables are readonly even to root.
1240  */
1241
1242 static int test_perm(int mode, int op)
1243 {
1244         if (!current->euid)
1245                 mode >>= 6;
1246         else if (in_egroup_p(0))
1247                 mode >>= 3;
1248         if ((mode & op & 0007) == op)
1249                 return 0;
1250         return -EACCES;
1251 }
1252
1253 static inline int ctl_perm(ctl_table *table, int op)
1254 {
1255         int error;
1256         error = security_sysctl(table, op);
1257         if (error)
1258                 return error;
1259         return test_perm(table->mode, op);
1260 }
1261
1262 static int parse_table(int __user *name, int nlen,
1263                        void __user *oldval, size_t __user *oldlenp,
1264                        void __user *newval, size_t newlen,
1265                        ctl_table *table, void **context)
1266 {
1267         int n;
1268 repeat:
1269         if (!nlen)
1270                 return -ENOTDIR;
1271         if (get_user(n, name))
1272                 return -EFAULT;
1273         for ( ; table->ctl_name; table++) {
1274                 if (n == table->ctl_name || table->ctl_name == CTL_ANY) {
1275                         int error;
1276                         if (table->child) {
1277                                 if (ctl_perm(table, 001))
1278                                         return -EPERM;
1279                                 if (table->strategy) {
1280                                         error = table->strategy(
1281                                                 table, name, nlen,
1282                                                 oldval, oldlenp,
1283                                                 newval, newlen, context);
1284                                         if (error)
1285                                                 return error;
1286                                 }
1287                                 name++;
1288                                 nlen--;
1289                                 table = table->child;
1290                                 goto repeat;
1291                         }
1292                         error = do_sysctl_strategy(table, name, nlen,
1293                                                    oldval, oldlenp,
1294                                                    newval, newlen, context);
1295                         return error;
1296                 }
1297         }
1298         return -ENOTDIR;
1299 }
1300
1301 /* Perform the actual read/write of a sysctl table entry. */
1302 int do_sysctl_strategy (ctl_table *table,
1303                         int __user *name, int nlen,
1304                         void __user *oldval, size_t __user *oldlenp,
1305                         void __user *newval, size_t newlen, void **context)
1306 {
1307         int op = 0, rc;
1308         size_t len;
1309
1310         if (oldval)
1311                 op |= 004;
1312         if (newval)
1313                 op |= 002;
1314         if (ctl_perm(table, op))
1315                 return -EPERM;
1316
1317         if (table->strategy) {
1318                 rc = table->strategy(table, name, nlen, oldval, oldlenp,
1319                                      newval, newlen, context);
1320                 if (rc < 0)
1321                         return rc;
1322                 if (rc > 0)
1323                         return 0;
1324         }
1325
1326         /* If there is no strategy routine, or if the strategy returns
1327          * zero, proceed with automatic r/w */
1328         if (table->data && table->maxlen) {
1329                 if (oldval && oldlenp) {
1330                         if (get_user(len, oldlenp))
1331                                 return -EFAULT;
1332                         if (len) {
1333                                 if (len > table->maxlen)
1334                                         len = table->maxlen;
1335                                 if(copy_to_user(oldval, table->data, len))
1336                                         return -EFAULT;
1337                                 if(put_user(len, oldlenp))
1338                                         return -EFAULT;
1339                         }
1340                 }
1341                 if (newval && newlen) {
1342                         len = newlen;
1343                         if (len > table->maxlen)
1344                                 len = table->maxlen;
1345                         if(copy_from_user(table->data, newval, len))
1346                                 return -EFAULT;
1347                 }
1348         }
1349         return 0;
1350 }
1351
1352 /**
1353  * register_sysctl_table - register a sysctl hierarchy
1354  * @table: the top-level table structure
1355  * @insert_at_head: whether the entry should be inserted in front or at the end
1356  *
1357  * Register a sysctl table hierarchy. @table should be a filled in ctl_table
1358  * array. An entry with a ctl_name of 0 terminates the table.
1359  *
1360  * The members of the &ctl_table structure are used as follows:
1361  *
1362  * ctl_name - This is the numeric sysctl value used by sysctl(2). The number
1363  *            must be unique within that level of sysctl
1364  *
1365  * procname - the name of the sysctl file under /proc/sys. Set to %NULL to not
1366  *            enter a sysctl file
1367  *
1368  * data - a pointer to data for use by proc_handler
1369  *
1370  * maxlen - the maximum size in bytes of the data
1371  *
1372  * mode - the file permissions for the /proc/sys file, and for sysctl(2)
1373  *
1374  * child - a pointer to the child sysctl table if this entry is a directory, or
1375  *         %NULL.
1376  *
1377  * proc_handler - the text handler routine (described below)
1378  *
1379  * strategy - the strategy routine (described below)
1380  *
1381  * de - for internal use by the sysctl routines
1382  *
1383  * extra1, extra2 - extra pointers usable by the proc handler routines
1384  *
1385  * Leaf nodes in the sysctl tree will be represented by a single file
1386  * under /proc; non-leaf nodes will be represented by directories.
1387  *
1388  * sysctl(2) can automatically manage read and write requests through
1389  * the sysctl table.  The data and maxlen fields of the ctl_table
1390  * struct enable minimal validation of the values being written to be
1391  * performed, and the mode field allows minimal authentication.
1392  *
1393  * More sophisticated management can be enabled by the provision of a
1394  * strategy routine with the table entry.  This will be called before
1395  * any automatic read or write of the data is performed.
1396  *
1397  * The strategy routine may return
1398  *
1399  * < 0 - Error occurred (error is passed to user process)
1400  *
1401  * 0   - OK - proceed with automatic read or write.
1402  *
1403  * > 0 - OK - read or write has been done by the strategy routine, so
1404  *       return immediately.
1405  *
1406  * There must be a proc_handler routine for any terminal nodes
1407  * mirrored under /proc/sys (non-terminals are handled by a built-in
1408  * directory handler).  Several default handlers are available to
1409  * cover common cases -
1410  *
1411  * proc_dostring(), proc_dointvec(), proc_dointvec_jiffies(),
1412  * proc_dointvec_userhz_jiffies(), proc_dointvec_minmax(),
1413  * proc_doulongvec_ms_jiffies_minmax(), proc_doulongvec_minmax()
1414  *
1415  * It is the handler's job to read the input buffer from user memory
1416  * and process it. The handler should return 0 on success.
1417  *
1418  * This routine returns %NULL on a failure to register, and a pointer
1419  * to the table header on success.
1420  */
1421 struct ctl_table_header *register_sysctl_table(ctl_table * table,
1422                                                int insert_at_head)
1423 {
1424         struct ctl_table_header *tmp;
1425         tmp = kmalloc(sizeof(struct ctl_table_header), GFP_KERNEL);
1426         if (!tmp)
1427                 return NULL;
1428         tmp->ctl_table = table;
1429         INIT_LIST_HEAD(&tmp->ctl_entry);
1430         tmp->used = 0;
1431         tmp->unregistering = NULL;
1432         spin_lock(&sysctl_lock);
1433         if (insert_at_head)
1434                 list_add(&tmp->ctl_entry, &root_table_header.ctl_entry);
1435         else
1436                 list_add_tail(&tmp->ctl_entry, &root_table_header.ctl_entry);
1437         spin_unlock(&sysctl_lock);
1438 #ifdef CONFIG_PROC_FS
1439         register_proc_table(table, proc_sys_root, tmp);
1440 #endif
1441         return tmp;
1442 }
1443
1444 /**
1445  * unregister_sysctl_table - unregister a sysctl table hierarchy
1446  * @header: the header returned from register_sysctl_table
1447  *
1448  * Unregisters the sysctl table and all children. proc entries may not
1449  * actually be removed until they are no longer used by anyone.
1450  */
1451 void unregister_sysctl_table(struct ctl_table_header * header)
1452 {
1453         might_sleep();
1454         spin_lock(&sysctl_lock);
1455         start_unregistering(header);
1456 #ifdef CONFIG_PROC_FS
1457         unregister_proc_table(header->ctl_table, proc_sys_root);
1458 #endif
1459         spin_unlock(&sysctl_lock);
1460         kfree(header);
1461 }
1462
1463 /*
1464  * /proc/sys support
1465  */
1466
1467 #ifdef CONFIG_PROC_FS
1468
1469 /* Scan the sysctl entries in table and add them all into /proc */
1470 static void register_proc_table(ctl_table * table, struct proc_dir_entry *root, void *set)
1471 {
1472         struct proc_dir_entry *de;
1473         int len;
1474         mode_t mode;
1475
1476         for (; table->ctl_name; table++) {
1477                 /* Can't do anything without a proc name. */
1478                 if (!table->procname)
1479                         continue;
1480                 /* Maybe we can't do anything with it... */
1481                 if (!table->proc_handler && !table->child) {
1482                         printk(KERN_WARNING "SYSCTL: Can't register %s\n",
1483                                 table->procname);
1484                         continue;
1485                 }
1486
1487                 len = strlen(table->procname);
1488                 mode = table->mode;
1489
1490                 de = NULL;
1491                 if (table->proc_handler)
1492                         mode |= S_IFREG;
1493                 else {
1494                         mode |= S_IFDIR;
1495                         for (de = root->subdir; de; de = de->next) {
1496                                 if (proc_match(len, table->procname, de))
1497                                         break;
1498                         }
1499                         /* If the subdir exists already, de is non-NULL */
1500                 }
1501
1502                 if (!de) {
1503                         de = create_proc_entry(table->procname, mode, root);
1504                         if (!de)
1505                                 continue;
1506                         de->set = set;
1507                         de->data = (void *) table;
1508                         if (table->proc_handler)
1509                                 de->proc_fops = &proc_sys_file_operations;
1510                 }
1511                 table->de = de;
1512                 if (de->mode & S_IFDIR)
1513                         register_proc_table(table->child, de, set);
1514         }
1515 }
1516
1517 /*
1518  * Unregister a /proc sysctl table and any subdirectories.
1519  */
1520 static void unregister_proc_table(ctl_table * table, struct proc_dir_entry *root)
1521 {
1522         struct proc_dir_entry *de;
1523         for (; table->ctl_name; table++) {
1524                 if (!(de = table->de))
1525                         continue;
1526                 if (de->mode & S_IFDIR) {
1527                         if (!table->child) {
1528                                 printk (KERN_ALERT "Help - malformed sysctl tree on free\n");
1529                                 continue;
1530                         }
1531                         unregister_proc_table(table->child, de);
1532
1533                         /* Don't unregister directories which still have entries.. */
1534                         if (de->subdir)
1535                                 continue;
1536                 }
1537
1538                 /*
1539                  * In any case, mark the entry as goner; we'll keep it
1540                  * around if it's busy, but we'll know to do nothing with
1541                  * its fields.  We are under sysctl_lock here.
1542                  */
1543                 de->data = NULL;
1544
1545                 /* Don't unregister proc entries that are still being used.. */
1546                 if (atomic_read(&de->count))
1547                         continue;
1548
1549                 table->de = NULL;
1550                 remove_proc_entry(table->procname, root);
1551         }
1552 }
1553
1554 static ssize_t do_rw_proc(int write, struct file * file, char __user * buf,
1555                           size_t count, loff_t *ppos)
1556 {
1557         int op;
1558         struct proc_dir_entry *de = PDE(file->f_dentry->d_inode);
1559         struct ctl_table *table;
1560         size_t res;
1561         ssize_t error = -ENOTDIR;
1562
1563         spin_lock(&sysctl_lock);
1564         if (de && de->data && use_table(de->set)) {
1565                 /*
1566                  * at that point we know that sysctl was not unregistered
1567                  * and won't be until we finish
1568                  */
1569                 spin_unlock(&sysctl_lock);
1570                 table = (struct ctl_table *) de->data;
1571                 if (!table || !table->proc_handler)
1572                         goto out;
1573                 error = -EPERM;
1574                 op = (write ? 002 : 004);
1575                 if (ctl_perm(table, op))
1576                         goto out;
1577
1578                 /* careful: calling conventions are nasty here */
1579                 res = count;
1580                 error = (*table->proc_handler)(table, write, file,
1581                                                 buf, &res, ppos);
1582                 if (!error)
1583                         error = res;
1584         out:
1585                 spin_lock(&sysctl_lock);
1586                 unuse_table(de->set);
1587         }
1588         spin_unlock(&sysctl_lock);
1589         return error;
1590 }
1591
1592 static int proc_opensys(struct inode *inode, struct file *file)
1593 {
1594         if (file->f_mode & FMODE_WRITE) {
1595                 /*
1596                  * sysctl entries that are not writable,
1597                  * are _NOT_ writable, capabilities or not.
1598                  */
1599                 if (!(inode->i_mode & S_IWUSR))
1600                         return -EPERM;
1601         }
1602
1603         return 0;
1604 }
1605
1606 static ssize_t proc_readsys(struct file * file, char __user * buf,
1607                             size_t count, loff_t *ppos)
1608 {
1609         return do_rw_proc(0, file, buf, count, ppos);
1610 }
1611
1612 static ssize_t proc_writesys(struct file * file, const char __user * buf,
1613                              size_t count, loff_t *ppos)
1614 {
1615         return do_rw_proc(1, file, (char __user *) buf, count, ppos);
1616 }
1617
1618 /**
1619  * proc_dostring - read a string sysctl
1620  * @table: the sysctl table
1621  * @write: %TRUE if this is a write to the sysctl file
1622  * @filp: the file structure
1623  * @buffer: the user buffer
1624  * @lenp: the size of the user buffer
1625  * @ppos: file position
1626  *
1627  * Reads/writes a string from/to the user buffer. If the kernel
1628  * buffer provided is not large enough to hold the string, the
1629  * string is truncated. The copied string is %NULL-terminated.
1630  * If the string is being read by the user process, it is copied
1631  * and a newline '\n' is added. It is truncated if the buffer is
1632  * not large enough.
1633  *
1634  * Returns 0 on success.
1635  */
1636 int proc_dostring(ctl_table *table, int write, struct file *filp,
1637                   void __user *buffer, size_t *lenp, loff_t *ppos)
1638 {
1639         size_t len, maxlen;
1640         char __user *p;
1641         char c;
1642         void *data;
1643
1644         data = table->data;
1645         maxlen = table->maxlen;
1646
1647         if (!data || !maxlen || !*lenp || (*ppos && !write))
1648                 return (*lenp = 0);
1649
1650         if (table->virt_handler)
1651                 table->virt_handler(table, write, filp->f_xid, &data, &maxlen);
1652
1653         if (write) {
1654                 len = 0;
1655                 p = buffer;
1656                 while (len < *lenp) {
1657                         if (get_user(c, p++))
1658                                 return -EFAULT;
1659                         if (c == 0 || c == '\n')
1660                                 break;
1661                         len++;
1662                 }
1663                 if (len >= maxlen)
1664                         len = maxlen-1;
1665                 if(copy_from_user(data, buffer, len))
1666                         return -EFAULT;
1667                 ((char *) data)[len] = 0;
1668                 *ppos += *lenp;
1669         } else {
1670                 len = strlen(data);
1671                 if (len > maxlen)
1672                         len = maxlen;
1673                 if (len > *lenp)
1674                         len = *lenp;
1675                 if (len)
1676                         if(copy_to_user(buffer, data, len))
1677                                 return -EFAULT;
1678                 if (len < *lenp) {
1679                         if(put_user('\n', ((char __user *) buffer) + len))
1680                                 return -EFAULT;
1681                         len++;
1682                 }
1683                 *lenp = len;
1684                 *ppos += len;
1685         }
1686         return 0;
1687 }
1688
1689 /*
1690  *      Special case of dostring for the UTS structure. This has locks
1691  *      to observe. Should this be in kernel/sys.c ????
1692  */
1693
1694 static int proc_doutsstring(ctl_table *table, int write, struct file *filp,
1695                   void __user *buffer, size_t *lenp, loff_t *ppos)
1696 {
1697         int r;
1698
1699         if (!write) {
1700                 down_read(&uts_sem);
1701                 r=proc_dostring(table,0,filp,buffer,lenp, ppos);
1702                 up_read(&uts_sem);
1703         } else {
1704                 down_write(&uts_sem);
1705                 r=proc_dostring(table,1,filp,buffer,lenp, ppos);
1706                 up_write(&uts_sem);
1707         }
1708         return r;
1709 }
1710
1711 static int do_proc_dointvec_conv(int *negp, unsigned long *lvalp,
1712                                  int *valp,
1713                                  int write, void *data)
1714 {
1715         if (write) {
1716                 *valp = *negp ? -*lvalp : *lvalp;
1717         } else {
1718                 int val = *valp;
1719                 if (val < 0) {
1720                         *negp = -1;
1721                         *lvalp = (unsigned long)-val;
1722                 } else {
1723                         *negp = 0;
1724                         *lvalp = (unsigned long)val;
1725                 }
1726         }
1727         return 0;
1728 }
1729
1730 static int do_proc_dointvec(ctl_table *table, int write, struct file *filp,
1731                   void __user *buffer, size_t *lenp, loff_t *ppos,
1732                   int (*conv)(int *negp, unsigned long *lvalp, int *valp,
1733                               int write, void *data),
1734                   void *data)
1735 {
1736 #define TMPBUFLEN 21
1737         int *i, vleft, first=1, neg, val;
1738         unsigned long lval;
1739         size_t left, len;
1740
1741         char buf[TMPBUFLEN], *p;
1742         char __user *s = buffer;
1743
1744         if (!table->data || !table->maxlen || !*lenp ||
1745             (*ppos && !write)) {
1746                 *lenp = 0;
1747                 return 0;
1748         }
1749
1750         i = (int *) table->data;
1751         vleft = table->maxlen / sizeof(*i);
1752         left = *lenp;
1753
1754         if (!conv)
1755                 conv = do_proc_dointvec_conv;
1756
1757         for (; left && vleft--; i++, first=0) {
1758                 if (write) {
1759                         while (left) {
1760                                 char c;
1761                                 if (get_user(c, s))
1762                                         return -EFAULT;
1763                                 if (!isspace(c))
1764                                         break;
1765                                 left--;
1766                                 s++;
1767                         }
1768                         if (!left)
1769                                 break;
1770                         neg = 0;
1771                         len = left;
1772                         if (len > sizeof(buf) - 1)
1773                                 len = sizeof(buf) - 1;
1774                         if (copy_from_user(buf, s, len))
1775                                 return -EFAULT;
1776                         buf[len] = 0;
1777                         p = buf;
1778                         if (*p == '-' && left > 1) {
1779                                 neg = 1;
1780                                 left--, p++;
1781                         }
1782                         if (*p < '0' || *p > '9')
1783                                 break;
1784
1785                         lval = simple_strtoul(p, &p, 0);
1786
1787                         len = p-buf;
1788                         if ((len < left) && *p && !isspace(*p))
1789                                 break;
1790                         if (neg)
1791                                 val = -val;
1792                         s += len;
1793                         left -= len;
1794
1795                         if (conv(&neg, &lval, i, 1, data))
1796                                 break;
1797                 } else {
1798                         p = buf;
1799                         if (!first)
1800                                 *p++ = '\t';
1801
1802                         if (conv(&neg, &lval, i, 0, data))
1803                                 break;
1804
1805                         sprintf(p, "%s%lu", neg ? "-" : "", lval);
1806                         len = strlen(buf);
1807                         if (len > left)
1808                                 len = left;
1809                         if(copy_to_user(s, buf, len))
1810                                 return -EFAULT;
1811                         left -= len;
1812                         s += len;
1813                 }
1814         }
1815
1816         if (!write && !first && left) {
1817                 if(put_user('\n', s))
1818                         return -EFAULT;
1819                 left--, s++;
1820         }
1821         if (write) {
1822                 while (left) {
1823                         char c;
1824                         if (get_user(c, s++))
1825                                 return -EFAULT;
1826                         if (!isspace(c))
1827                                 break;
1828                         left--;
1829                 }
1830         }
1831         if (write && first)
1832                 return -EINVAL;
1833         *lenp -= left;
1834         *ppos += *lenp;
1835         return 0;
1836 #undef TMPBUFLEN
1837 }
1838
1839 /**
1840  * proc_dointvec - read a vector of integers
1841  * @table: the sysctl table
1842  * @write: %TRUE if this is a write to the sysctl file
1843  * @filp: the file structure
1844  * @buffer: the user buffer
1845  * @lenp: the size of the user buffer
1846  * @ppos: file position
1847  *
1848  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
1849  * values from/to the user buffer, treated as an ASCII string.
1850  *
1851  * Returns 0 on success.
1852  */
1853 int proc_dointvec(ctl_table *table, int write, struct file *filp,
1854                      void __user *buffer, size_t *lenp, loff_t *ppos)
1855 {
1856     return do_proc_dointvec(table,write,filp,buffer,lenp,ppos,
1857                             NULL,NULL);
1858 }
1859
1860 #define OP_SET  0
1861 #define OP_AND  1
1862 #define OP_OR   2
1863 #define OP_MAX  3
1864 #define OP_MIN  4
1865
1866 static int do_proc_dointvec_bset_conv(int *negp, unsigned long *lvalp,
1867                                       int *valp,
1868                                       int write, void *data)
1869 {
1870         int op = *(int *)data;
1871         if (write) {
1872                 int val = *negp ? -*lvalp : *lvalp;
1873                 switch(op) {
1874                 case OP_SET:    *valp = val; break;
1875                 case OP_AND:    *valp &= val; break;
1876                 case OP_OR:     *valp |= val; break;
1877                 case OP_MAX:    if(*valp < val)
1878                                         *valp = val;
1879                                 break;
1880                 case OP_MIN:    if(*valp > val)
1881                                 *valp = val;
1882                                 break;
1883                 }
1884         } else {
1885                 int val = *valp;
1886                 if (val < 0) {
1887                         *negp = -1;
1888                         *lvalp = (unsigned long)-val;
1889                 } else {
1890                         *negp = 0;
1891                         *lvalp = (unsigned long)val;
1892                 }
1893         }
1894         return 0;
1895 }
1896
1897 /*
1898  *      init may raise the set.
1899  */
1900
1901 int proc_dointvec_bset(ctl_table *table, int write, struct file *filp,
1902                         void __user *buffer, size_t *lenp, loff_t *ppos)
1903 {
1904         int op;
1905
1906         if (!capable(CAP_SYS_MODULE)) {
1907                 return -EPERM;
1908         }
1909
1910         op = (current->pid == 1) ? OP_SET : OP_AND;
1911         return do_proc_dointvec(table,write,filp,buffer,lenp,ppos,
1912                                 do_proc_dointvec_bset_conv,&op);
1913 }
1914
1915 struct do_proc_dointvec_minmax_conv_param {
1916         int *min;
1917         int *max;
1918 };
1919
1920 static int do_proc_dointvec_minmax_conv(int *negp, unsigned long *lvalp,
1921                                         int *valp,
1922                                         int write, void *data)
1923 {
1924         struct do_proc_dointvec_minmax_conv_param *param = data;
1925         if (write) {
1926                 int val = *negp ? -*lvalp : *lvalp;
1927                 if ((param->min && *param->min > val) ||
1928                     (param->max && *param->max < val))
1929                         return -EINVAL;
1930                 *valp = val;
1931         } else {
1932                 int val = *valp;
1933                 if (val < 0) {
1934                         *negp = -1;
1935                         *lvalp = (unsigned long)-val;
1936                 } else {
1937                         *negp = 0;
1938                         *lvalp = (unsigned long)val;
1939                 }
1940         }
1941         return 0;
1942 }
1943
1944 /**
1945  * proc_dointvec_minmax - read a vector of integers with min/max values
1946  * @table: the sysctl table
1947  * @write: %TRUE if this is a write to the sysctl file
1948  * @filp: the file structure
1949  * @buffer: the user buffer
1950  * @lenp: the size of the user buffer
1951  * @ppos: file position
1952  *
1953  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
1954  * values from/to the user buffer, treated as an ASCII string.
1955  *
1956  * This routine will ensure the values are within the range specified by
1957  * table->extra1 (min) and table->extra2 (max).
1958  *
1959  * Returns 0 on success.
1960  */
1961 int proc_dointvec_minmax(ctl_table *table, int write, struct file *filp,
1962                   void __user *buffer, size_t *lenp, loff_t *ppos)
1963 {
1964         struct do_proc_dointvec_minmax_conv_param param = {
1965                 .min = (int *) table->extra1,
1966                 .max = (int *) table->extra2,
1967         };
1968         return do_proc_dointvec(table, write, filp, buffer, lenp, ppos,
1969                                 do_proc_dointvec_minmax_conv, &param);
1970 }
1971
1972 static int do_proc_doulongvec_minmax(ctl_table *table, int write,
1973                                      struct file *filp,
1974                                      void __user *buffer,
1975                                      size_t *lenp, loff_t *ppos,
1976                                      unsigned long convmul,
1977                                      unsigned long convdiv)
1978 {
1979 #define TMPBUFLEN 21
1980         unsigned long *i, *min, *max, val;
1981         int vleft, first=1, neg;
1982         size_t len, left;
1983         char buf[TMPBUFLEN], *p;
1984         char __user *s = buffer;
1985
1986         if (!table->data || !table->maxlen || !*lenp ||
1987             (*ppos && !write)) {
1988                 *lenp = 0;
1989                 return 0;
1990         }
1991
1992         i = (unsigned long *) table->data;
1993         min = (unsigned long *) table->extra1;
1994         max = (unsigned long *) table->extra2;
1995         vleft = table->maxlen / sizeof(unsigned long);
1996         left = *lenp;
1997
1998         for (; left && vleft--; i++, min++, max++, first=0) {
1999                 if (write) {
2000                         while (left) {
2001                                 char c;
2002                                 if (get_user(c, s))
2003                                         return -EFAULT;
2004                                 if (!isspace(c))
2005                                         break;
2006                                 left--;
2007                                 s++;
2008                         }
2009                         if (!left)
2010                                 break;
2011                         neg = 0;
2012                         len = left;
2013                         if (len > TMPBUFLEN-1)
2014                                 len = TMPBUFLEN-1;
2015                         if (copy_from_user(buf, s, len))
2016                                 return -EFAULT;
2017                         buf[len] = 0;
2018                         p = buf;
2019                         if (*p == '-' && left > 1) {
2020                                 neg = 1;
2021                                 left--, p++;
2022                         }
2023                         if (*p < '0' || *p > '9')
2024                                 break;
2025                         val = simple_strtoul(p, &p, 0) * convmul / convdiv ;
2026                         len = p-buf;
2027                         if ((len < left) && *p && !isspace(*p))
2028                                 break;
2029                         if (neg)
2030                                 val = -val;
2031                         s += len;
2032                         left -= len;
2033
2034                         if(neg)
2035                                 continue;
2036                         if ((min && val < *min) || (max && val > *max))
2037                                 continue;
2038                         *i = val;
2039                 } else {
2040                         p = buf;
2041                         if (!first)
2042                                 *p++ = '\t';
2043                         sprintf(p, "%lu", convdiv * (*i) / convmul);
2044                         len = strlen(buf);
2045                         if (len > left)
2046                                 len = left;
2047                         if(copy_to_user(s, buf, len))
2048                                 return -EFAULT;
2049                         left -= len;
2050                         s += len;
2051                 }
2052         }
2053
2054         if (!write && !first && left) {
2055                 if(put_user('\n', s))
2056                         return -EFAULT;
2057                 left--, s++;
2058         }
2059         if (write) {
2060                 while (left) {
2061                         char c;
2062                         if (get_user(c, s++))
2063                                 return -EFAULT;
2064                         if (!isspace(c))
2065                                 break;
2066                         left--;
2067                 }
2068         }
2069         if (write && first)
2070                 return -EINVAL;
2071         *lenp -= left;
2072         *ppos += *lenp;
2073         return 0;
2074 #undef TMPBUFLEN
2075 }
2076
2077 /**
2078  * proc_doulongvec_minmax - read a vector of long integers with min/max values
2079  * @table: the sysctl table
2080  * @write: %TRUE if this is a write to the sysctl file
2081  * @filp: the file structure
2082  * @buffer: the user buffer
2083  * @lenp: the size of the user buffer
2084  * @ppos: file position
2085  *
2086  * Reads/writes up to table->maxlen/sizeof(unsigned long) unsigned long
2087  * values from/to the user buffer, treated as an ASCII string.
2088  *
2089  * This routine will ensure the values are within the range specified by
2090  * table->extra1 (min) and table->extra2 (max).
2091  *
2092  * Returns 0 on success.
2093  */
2094 int proc_doulongvec_minmax(ctl_table *table, int write, struct file *filp,
2095                            void __user *buffer, size_t *lenp, loff_t *ppos)
2096 {
2097     return do_proc_doulongvec_minmax(table, write, filp, buffer, lenp, ppos, 1l, 1l);
2098 }
2099
2100 /**
2101  * proc_doulongvec_ms_jiffies_minmax - read a vector of millisecond values with min/max values
2102  * @table: the sysctl table
2103  * @write: %TRUE if this is a write to the sysctl file
2104  * @filp: the file structure
2105  * @buffer: the user buffer
2106  * @lenp: the size of the user buffer
2107  * @ppos: file position
2108  *
2109  * Reads/writes up to table->maxlen/sizeof(unsigned long) unsigned long
2110  * values from/to the user buffer, treated as an ASCII string. The values
2111  * are treated as milliseconds, and converted to jiffies when they are stored.
2112  *
2113  * This routine will ensure the values are within the range specified by
2114  * table->extra1 (min) and table->extra2 (max).
2115  *
2116  * Returns 0 on success.
2117  */
2118 int proc_doulongvec_ms_jiffies_minmax(ctl_table *table, int write,
2119                                       struct file *filp,
2120                                       void __user *buffer,
2121                                       size_t *lenp, loff_t *ppos)
2122 {
2123     return do_proc_doulongvec_minmax(table, write, filp, buffer,
2124                                      lenp, ppos, HZ, 1000l);
2125 }
2126
2127
2128 static int do_proc_dointvec_jiffies_conv(int *negp, unsigned long *lvalp,
2129                                          int *valp,
2130                                          int write, void *data)
2131 {
2132         if (write) {
2133                 if (*lvalp > LONG_MAX / HZ)
2134                         return 1;
2135                 *valp = *negp ? -(*lvalp*HZ) : (*lvalp*HZ);
2136         } else {
2137                 int val = *valp;
2138                 unsigned long lval;
2139                 if (val < 0) {
2140                         *negp = -1;
2141                         lval = (unsigned long)-val;
2142                 } else {
2143                         *negp = 0;
2144                         lval = (unsigned long)val;
2145                 }
2146                 *lvalp = lval / HZ;
2147         }
2148         return 0;
2149 }
2150
2151 static int do_proc_dointvec_userhz_jiffies_conv(int *negp, unsigned long *lvalp,
2152                                                 int *valp,
2153                                                 int write, void *data)
2154 {
2155         if (write) {
2156                 if (USER_HZ < HZ && *lvalp > (LONG_MAX / HZ) * USER_HZ)
2157                         return 1;
2158                 *valp = clock_t_to_jiffies(*negp ? -*lvalp : *lvalp);
2159         } else {
2160                 int val = *valp;
2161                 unsigned long lval;
2162                 if (val < 0) {
2163                         *negp = -1;
2164                         lval = (unsigned long)-val;
2165                 } else {
2166                         *negp = 0;
2167                         lval = (unsigned long)val;
2168                 }
2169                 *lvalp = jiffies_to_clock_t(lval);
2170         }
2171         return 0;
2172 }
2173
2174 static int do_proc_dointvec_ms_jiffies_conv(int *negp, unsigned long *lvalp,
2175                                             int *valp,
2176                                             int write, void *data)
2177 {
2178         if (write) {
2179                 *valp = msecs_to_jiffies(*negp ? -*lvalp : *lvalp);
2180         } else {
2181                 int val = *valp;
2182                 unsigned long lval;
2183                 if (val < 0) {
2184                         *negp = -1;
2185                         lval = (unsigned long)-val;
2186                 } else {
2187                         *negp = 0;
2188                         lval = (unsigned long)val;
2189                 }
2190                 *lvalp = jiffies_to_msecs(lval);
2191         }
2192         return 0;
2193 }
2194
2195 /**
2196  * proc_dointvec_jiffies - read a vector of integers as seconds
2197  * @table: the sysctl table
2198  * @write: %TRUE if this is a write to the sysctl file
2199  * @filp: the file structure
2200  * @buffer: the user buffer
2201  * @lenp: the size of the user buffer
2202  * @ppos: file position
2203  *
2204  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2205  * values from/to the user buffer, treated as an ASCII string.
2206  * The values read are assumed to be in seconds, and are converted into
2207  * jiffies.
2208  *
2209  * Returns 0 on success.
2210  */
2211 int proc_dointvec_jiffies(ctl_table *table, int write, struct file *filp,
2212                           void __user *buffer, size_t *lenp, loff_t *ppos)
2213 {
2214     return do_proc_dointvec(table,write,filp,buffer,lenp,ppos,
2215                             do_proc_dointvec_jiffies_conv,NULL);
2216 }
2217
2218 /**
2219  * proc_dointvec_userhz_jiffies - read a vector of integers as 1/USER_HZ seconds
2220  * @table: the sysctl table
2221  * @write: %TRUE if this is a write to the sysctl file
2222  * @filp: the file structure
2223  * @buffer: the user buffer
2224  * @lenp: the size of the user buffer
2225  * @ppos: pointer to the file position
2226  *
2227  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2228  * values from/to the user buffer, treated as an ASCII string.
2229  * The values read are assumed to be in 1/USER_HZ seconds, and
2230  * are converted into jiffies.
2231  *
2232  * Returns 0 on success.
2233  */
2234 int proc_dointvec_userhz_jiffies(ctl_table *table, int write, struct file *filp,
2235                                  void __user *buffer, size_t *lenp, loff_t *ppos)
2236 {
2237     return do_proc_dointvec(table,write,filp,buffer,lenp,ppos,
2238                             do_proc_dointvec_userhz_jiffies_conv,NULL);
2239 }
2240
2241 /**
2242  * proc_dointvec_ms_jiffies - read a vector of integers as 1 milliseconds
2243  * @table: the sysctl table
2244  * @write: %TRUE if this is a write to the sysctl file
2245  * @filp: the file structure
2246  * @buffer: the user buffer
2247  * @lenp: the size of the user buffer
2248  * @ppos: file position
2249  * @ppos: the current position in the file
2250  *
2251  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2252  * values from/to the user buffer, treated as an ASCII string.
2253  * The values read are assumed to be in 1/1000 seconds, and
2254  * are converted into jiffies.
2255  *
2256  * Returns 0 on success.
2257  */
2258 int proc_dointvec_ms_jiffies(ctl_table *table, int write, struct file *filp,
2259                              void __user *buffer, size_t *lenp, loff_t *ppos)
2260 {
2261         return do_proc_dointvec(table, write, filp, buffer, lenp, ppos,
2262                                 do_proc_dointvec_ms_jiffies_conv, NULL);
2263 }
2264
2265 #else /* CONFIG_PROC_FS */
2266
2267 int proc_dostring(ctl_table *table, int write, struct file *filp,
2268                   void __user *buffer, size_t *lenp, loff_t *ppos)
2269 {
2270         return -ENOSYS;
2271 }
2272
2273 static int proc_doutsstring(ctl_table *table, int write, struct file *filp,
2274                             void __user *buffer, size_t *lenp, loff_t *ppos)
2275 {
2276         return -ENOSYS;
2277 }
2278
2279 int proc_dointvec(ctl_table *table, int write, struct file *filp,
2280                   void __user *buffer, size_t *lenp, loff_t *ppos)
2281 {
2282         return -ENOSYS;
2283 }
2284
2285 int proc_dointvec_bset(ctl_table *table, int write, struct file *filp,
2286                         void __user *buffer, size_t *lenp, loff_t *ppos)
2287 {
2288         return -ENOSYS;
2289 }
2290
2291 int proc_dointvec_minmax(ctl_table *table, int write, struct file *filp,
2292                     void __user *buffer, size_t *lenp, loff_t *ppos)
2293 {
2294         return -ENOSYS;
2295 }
2296
2297 int proc_dointvec_jiffies(ctl_table *table, int write, struct file *filp,
2298                     void __user *buffer, size_t *lenp, loff_t *ppos)
2299 {
2300         return -ENOSYS;
2301 }
2302
2303 int proc_dointvec_userhz_jiffies(ctl_table *table, int write, struct file *filp,
2304                     void __user *buffer, size_t *lenp, loff_t *ppos)
2305 {
2306         return -ENOSYS;
2307 }
2308
2309 int proc_dointvec_ms_jiffies(ctl_table *table, int write, struct file *filp,
2310                              void __user *buffer, size_t *lenp, loff_t *ppos)
2311 {
2312         return -ENOSYS;
2313 }
2314
2315 int proc_doulongvec_minmax(ctl_table *table, int write, struct file *filp,
2316                     void __user *buffer, size_t *lenp, loff_t *ppos)
2317 {
2318         return -ENOSYS;
2319 }
2320
2321 int proc_doulongvec_ms_jiffies_minmax(ctl_table *table, int write,
2322                                       struct file *filp,
2323                                       void __user *buffer,
2324                                       size_t *lenp, loff_t *ppos)
2325 {
2326     return -ENOSYS;
2327 }
2328
2329
2330 #endif /* CONFIG_PROC_FS */
2331
2332
2333 /*
2334  * General sysctl support routines
2335  */
2336
2337 /* The generic string strategy routine: */
2338 int sysctl_string(ctl_table *table, int __user *name, int nlen,
2339                   void __user *oldval, size_t __user *oldlenp,
2340                   void __user *newval, size_t newlen, void **context)
2341 {
2342         if (!table->data || !table->maxlen)
2343                 return -ENOTDIR;
2344
2345         if (oldval && oldlenp) {
2346                 size_t bufsize;
2347                 if (get_user(bufsize, oldlenp))
2348                         return -EFAULT;
2349                 if (bufsize) {
2350                         size_t len = strlen(table->data), copied;
2351
2352                         /* This shouldn't trigger for a well-formed sysctl */
2353                         if (len > table->maxlen)
2354                                 len = table->maxlen;
2355
2356                         /* Copy up to a max of bufsize-1 bytes of the string */
2357                         copied = (len >= bufsize) ? bufsize - 1 : len;
2358
2359                         if (copy_to_user(oldval, table->data, copied) ||
2360                             put_user(0, (char __user *)(oldval + copied)))
2361                                 return -EFAULT;
2362                         if (put_user(len, oldlenp))
2363                                 return -EFAULT;
2364                 }
2365         }
2366         if (newval && newlen) {
2367                 size_t len = newlen;
2368                 if (len > table->maxlen)
2369                         len = table->maxlen;
2370                 if(copy_from_user(table->data, newval, len))
2371                         return -EFAULT;
2372                 if (len == table->maxlen)
2373                         len--;
2374                 ((char *) table->data)[len] = 0;
2375         }
2376         return 1;
2377 }
2378
2379 /*
2380  * This function makes sure that all of the integers in the vector
2381  * are between the minimum and maximum values given in the arrays
2382  * table->extra1 and table->extra2, respectively.
2383  */
2384 int sysctl_intvec(ctl_table *table, int __user *name, int nlen,
2385                 void __user *oldval, size_t __user *oldlenp,
2386                 void __user *newval, size_t newlen, void **context)
2387 {
2388
2389         if (newval && newlen) {
2390                 int __user *vec = (int __user *) newval;
2391                 int *min = (int *) table->extra1;
2392                 int *max = (int *) table->extra2;
2393                 size_t length;
2394                 int i;
2395
2396                 if (newlen % sizeof(int) != 0)
2397                         return -EINVAL;
2398
2399                 if (!table->extra1 && !table->extra2)
2400                         return 0;
2401
2402                 if (newlen > table->maxlen)
2403                         newlen = table->maxlen;
2404                 length = newlen / sizeof(int);
2405
2406                 for (i = 0; i < length; i++) {
2407                         int value;
2408                         if (get_user(value, vec + i))
2409                                 return -EFAULT;
2410                         if (min && value < min[i])
2411                                 return -EINVAL;
2412                         if (max && value > max[i])
2413                                 return -EINVAL;
2414                 }
2415         }
2416         return 0;
2417 }
2418
2419 /* Strategy function to convert jiffies to seconds */
2420 int sysctl_jiffies(ctl_table *table, int __user *name, int nlen,
2421                 void __user *oldval, size_t __user *oldlenp,
2422                 void __user *newval, size_t newlen, void **context)
2423 {
2424         if (oldval) {
2425                 size_t olen;
2426                 if (oldlenp) {
2427                         if (get_user(olen, oldlenp))
2428                                 return -EFAULT;
2429                         if (olen!=sizeof(int))
2430                                 return -EINVAL;
2431                 }
2432                 if (put_user(*(int *)(table->data)/HZ, (int __user *)oldval) ||
2433                     (oldlenp && put_user(sizeof(int),oldlenp)))
2434                         return -EFAULT;
2435         }
2436         if (newval && newlen) {
2437                 int new;
2438                 if (newlen != sizeof(int))
2439                         return -EINVAL;
2440                 if (get_user(new, (int __user *)newval))
2441                         return -EFAULT;
2442                 *(int *)(table->data) = new*HZ;
2443         }
2444         return 1;
2445 }
2446
2447 /* Strategy function to convert jiffies to seconds */
2448 int sysctl_ms_jiffies(ctl_table *table, int __user *name, int nlen,
2449                 void __user *oldval, size_t __user *oldlenp,
2450                 void __user *newval, size_t newlen, void **context)
2451 {
2452         if (oldval) {
2453                 size_t olen;
2454                 if (oldlenp) {
2455                         if (get_user(olen, oldlenp))
2456                                 return -EFAULT;
2457                         if (olen!=sizeof(int))
2458                                 return -EINVAL;
2459                 }
2460                 if (put_user(jiffies_to_msecs(*(int *)(table->data)), (int __user *)oldval) ||
2461                     (oldlenp && put_user(sizeof(int),oldlenp)))
2462                         return -EFAULT;
2463         }
2464         if (newval && newlen) {
2465                 int new;
2466                 if (newlen != sizeof(int))
2467                         return -EINVAL;
2468                 if (get_user(new, (int __user *)newval))
2469                         return -EFAULT;
2470                 *(int *)(table->data) = msecs_to_jiffies(new);
2471         }
2472         return 1;
2473 }
2474
2475 #else /* CONFIG_SYSCTL */
2476
2477
2478 asmlinkage long sys_sysctl(struct __sysctl_args __user *args)
2479 {
2480         return -ENOSYS;
2481 }
2482
2483 int sysctl_string(ctl_table *table, int __user *name, int nlen,
2484                   void __user *oldval, size_t __user *oldlenp,
2485                   void __user *newval, size_t newlen, void **context)
2486 {
2487         return -ENOSYS;
2488 }
2489
2490 int sysctl_intvec(ctl_table *table, int __user *name, int nlen,
2491                 void __user *oldval, size_t __user *oldlenp,
2492                 void __user *newval, size_t newlen, void **context)
2493 {
2494         return -ENOSYS;
2495 }
2496
2497 int sysctl_jiffies(ctl_table *table, int __user *name, int nlen,
2498                 void __user *oldval, size_t __user *oldlenp,
2499                 void __user *newval, size_t newlen, void **context)
2500 {
2501         return -ENOSYS;
2502 }
2503
2504 int sysctl_ms_jiffies(ctl_table *table, int __user *name, int nlen,
2505                 void __user *oldval, size_t __user *oldlenp,
2506                 void __user *newval, size_t newlen, void **context)
2507 {
2508         return -ENOSYS;
2509 }
2510
2511 int proc_dostring(ctl_table *table, int write, struct file *filp,
2512                   void __user *buffer, size_t *lenp, loff_t *ppos)
2513 {
2514         return -ENOSYS;
2515 }
2516
2517 int proc_dointvec(ctl_table *table, int write, struct file *filp,
2518                   void __user *buffer, size_t *lenp, loff_t *ppos)
2519 {
2520         return -ENOSYS;
2521 }
2522
2523 int proc_dointvec_bset(ctl_table *table, int write, struct file *filp,
2524                         void __user *buffer, size_t *lenp, loff_t *ppos)
2525 {
2526         return -ENOSYS;
2527 }
2528
2529 int proc_dointvec_minmax(ctl_table *table, int write, struct file *filp,
2530                     void __user *buffer, size_t *lenp, loff_t *ppos)
2531 {
2532         return -ENOSYS;
2533 }
2534
2535 int proc_dointvec_jiffies(ctl_table *table, int write, struct file *filp,
2536                           void __user *buffer, size_t *lenp, loff_t *ppos)
2537 {
2538         return -ENOSYS;
2539 }
2540
2541 int proc_dointvec_userhz_jiffies(ctl_table *table, int write, struct file *filp,
2542                           void __user *buffer, size_t *lenp, loff_t *ppos)
2543 {
2544         return -ENOSYS;
2545 }
2546
2547 int proc_dointvec_ms_jiffies(ctl_table *table, int write, struct file *filp,
2548                              void __user *buffer, size_t *lenp, loff_t *ppos)
2549 {
2550         return -ENOSYS;
2551 }
2552
2553 int proc_doulongvec_minmax(ctl_table *table, int write, struct file *filp,
2554                     void __user *buffer, size_t *lenp, loff_t *ppos)
2555 {
2556         return -ENOSYS;
2557 }
2558
2559 int proc_doulongvec_ms_jiffies_minmax(ctl_table *table, int write,
2560                                       struct file *filp,
2561                                       void __user *buffer,
2562                                       size_t *lenp, loff_t *ppos)
2563 {
2564     return -ENOSYS;
2565 }
2566
2567 struct ctl_table_header * register_sysctl_table(ctl_table * table,
2568                                                 int insert_at_head)
2569 {
2570         return NULL;
2571 }
2572
2573 void unregister_sysctl_table(struct ctl_table_header * table)
2574 {
2575 }
2576
2577 #endif /* CONFIG_SYSCTL */
2578
2579 /*
2580  * No sense putting this after each symbol definition, twice,
2581  * exception granted :-)
2582  */
2583 EXPORT_SYMBOL(proc_dointvec);
2584 EXPORT_SYMBOL(proc_dointvec_jiffies);
2585 EXPORT_SYMBOL(proc_dointvec_minmax);
2586 EXPORT_SYMBOL(proc_dointvec_userhz_jiffies);
2587 EXPORT_SYMBOL(proc_dointvec_ms_jiffies);
2588 EXPORT_SYMBOL(proc_dostring);
2589 EXPORT_SYMBOL(proc_doulongvec_minmax);
2590 EXPORT_SYMBOL(proc_doulongvec_ms_jiffies_minmax);
2591 EXPORT_SYMBOL(register_sysctl_table);
2592 EXPORT_SYMBOL(sysctl_intvec);
2593 EXPORT_SYMBOL(sysctl_jiffies);
2594 EXPORT_SYMBOL(sysctl_ms_jiffies);
2595 EXPORT_SYMBOL(sysctl_string);
2596 EXPORT_SYMBOL(unregister_sysctl_table);