kernel/sysctl.c

   1 /*
   2  * sysctl.c: General linux system control interface
   3  *
   4  * Begun 24 March 1995, Stephen Tweedie
   5  * Added /proc support, Dec 1995
   6  * Added bdflush entry and intvec min/max checking, 2/23/96, Tom Dyas.
   7  * Added hooks for /proc/sys/net (minor, minor patch), 96/4/1, Mike Shaver.
   8  * Added kernel/java-{interpreter,appletviewer}, 96/5/10, Mike Shaver.
   9  * Dynamic registration fixes, Stephen Tweedie.
  10  * Added kswapd-interval, ctrl-alt-del, printk stuff, 1/8/97, Chris Horn.
  11  * Made sysctl support optional via CONFIG_SYSCTL, 1/10/97, Chris
  12  *  Horn.
  13  * Added proc_doulongvec_ms_jiffies_minmax, 09/08/99, Carlos H. Bauer.
  14  * Added proc_doulongvec_minmax, 09/08/99, Carlos H. Bauer.
  15  * Changed linked lists to use list.h instead of lists.h, 02/24/00, Bill
  16  *  Wendling.
  17  * The list_for_each() macro wasn't appropriate for the sysctl loop.
  18  *  Removed it and replaced it with older style, 03/23/00, Bill Wendling
  19  */
  20
  21 #include <linux/module.h>
  22 #include <linux/mm.h>
  23 #include <linux/swap.h>
  24 #include <linux/slab.h>
  25 #include <linux/sysctl.h>
  26 #include <linux/bitmap.h>
  27 #include <linux/signal.h>
  28 #include <linux/printk.h>
  29 #include <linux/proc_fs.h>
  30 #include <linux/security.h>
  31 #include <linux/ctype.h>
  32 #include <linux/kmemcheck.h>
  33 #include <linux/kmemleak.h>
  34 #include <linux/fs.h>
  35 #include <linux/init.h>
  36 #include <linux/kernel.h>
  37 #include <linux/kobject.h>
  38 #include <linux/net.h>
  39 #include <linux/sysrq.h>
  40 #include <linux/highuid.h>
  41 #include <linux/writeback.h>
  42 #include <linux/ratelimit.h>
  43 #include <linux/compaction.h>
  44 #include <linux/hugetlb.h>
  45 #include <linux/initrd.h>
  46 #include <linux/key.h>
  47 #include <linux/times.h>
  48 #include <linux/limits.h>
  49 #include <linux/dcache.h>
  50 #include <linux/dnotify.h>
  51 #include <linux/syscalls.h>
  52 #include <linux/vmstat.h>
  53 #include <linux/nfs_fs.h>
  54 #include <linux/acpi.h>
  55 #include <linux/reboot.h>
  56 #include <linux/ftrace.h>
  57 #include <linux/perf_event.h>
  58 #include <linux/kprobes.h>
  59 #include <linux/pipe_fs_i.h>
  60 #include <linux/oom.h>
  61 #include <linux/kmod.h>
  62 #include <linux/capability.h>
  63 #include <linux/binfmts.h>
  64 #include <linux/sched/sysctl.h>
  65
  66 #include <asm/uaccess.h>
  67 #include <asm/processor.h>
  68
  69 #ifdef CONFIG_X86
  70 #include <asm/nmi.h>
  71 #include <asm/stacktrace.h>
  72 #include <asm/io.h>
  73 #endif
  74 #ifdef CONFIG_SPARC
  75 #include <asm/setup.h>
  76 #endif
  77 #ifdef CONFIG_BSD_PROCESS_ACCT
  78 #include <linux/acct.h>
  79 #endif
  80 #ifdef CONFIG_RT_MUTEXES
  81 #include <linux/rtmutex.h>
  82 #endif
  83 #if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_LOCK_STAT)
  84 #include <linux/lockdep.h>
  85 #endif
  86 #ifdef CONFIG_CHR_DEV_SG
  87 #include <scsi/sg.h>
  88 #endif
  89
  90 #ifdef CONFIG_LOCKUP_DETECTOR
  91 #include <linux/nmi.h>
  92 #endif
  93
  94
  95 #if defined(CONFIG_SYSCTL)
  96
  97 /* External variables not in a header file. */
  98 extern int sysctl_overcommit_memory;
  99 extern int sysctl_overcommit_ratio;
 100 extern int max_threads;
 101 extern int suid_dumpable;
 102 #ifdef CONFIG_COREDUMP
 103 extern int core_uses_pid;
 104 extern char core_pattern[];
 105 extern unsigned int core_pipe_limit;
 106 #endif
 107 extern int pid_max;
 108 extern int pid_max_min, pid_max_max;
 109 extern int percpu_pagelist_fraction;
 110 extern int compat_log;
 111 extern int latencytop_enabled;
 112 extern int sysctl_nr_open_min, sysctl_nr_open_max;
 113 #ifndef CONFIG_MMU
 114 extern int sysctl_nr_trim_pages;
 115 #endif
 116 #ifdef CONFIG_BLOCK
 117 extern int blk_iopoll_enabled;
 118 #endif
 119
 120 /* Constants used for minimum and  maximum */
 121 #ifdef CONFIG_LOCKUP_DETECTOR
 122 static int sixty = 60;
 123 #endif
 124
 125 static int zero;
 126 static int __maybe_unused one = 1;
 127 static int __maybe_unused two = 2;
 128 static int __maybe_unused three = 3;
 129 static unsigned long one_ul = 1;
 130 static int one_hundred = 100;
 131 #ifdef CONFIG_PRINTK
 132 static int ten_thousand = 10000;
 133 #endif
 134
 135 /* this is needed for the proc_doulongvec_minmax of vm_dirty_bytes */
 136 static unsigned long dirty_bytes_min = 2 * PAGE_SIZE;
 137
 138 /* this is needed for the proc_dointvec_minmax for [fs_]overflow UID and GID */
 139 static int maxolduid = 65535;
 140 static int minolduid;
 141 static int min_percpu_pagelist_fract = 8;
 142
 143 static int ngroups_max = NGROUPS_MAX;
 144 static const int cap_last_cap = CAP_LAST_CAP;
 145
 146 #ifdef CONFIG_INOTIFY_USER
 147 #include <linux/inotify.h>
 148 #endif
 149 #ifdef CONFIG_SPARC
 150 #endif
 151
 152 #ifdef CONFIG_SPARC64
 153 extern int sysctl_tsb_ratio;
 154 #endif
 155
 156 #ifdef __hppa__
 157 extern int pwrsw_enabled;
 158 #endif
 159
 160 #ifdef CONFIG_SYSCTL_ARCH_UNALIGN_ALLOW
 161 extern int unaligned_enabled;
 162 #endif
 163
 164 #ifdef CONFIG_IA64
 165 extern int unaligned_dump_stack;
 166 #endif
 167
 168 #ifdef CONFIG_SYSCTL_ARCH_UNALIGN_NO_WARN
 169 extern int no_unaligned_warning;
 170 #endif
 171
 172 #ifdef CONFIG_PROC_SYSCTL
 173 static int proc_do_cad_pid(struct ctl_table *table, int write,
 174                   void __user *buffer, size_t *lenp, loff_t *ppos);
 175 static int proc_taint(struct ctl_table *table, int write,
 176                                void __user *buffer, size_t *lenp, loff_t *ppos);
 177 #endif
 178
 179 #ifdef CONFIG_PRINTK
 180 static int proc_dointvec_minmax_sysadmin(struct ctl_table *table, int write,
 181                                 void __user *buffer, size_t *lenp, loff_t *ppos);
 182 #endif
 183
 184 static int proc_dointvec_minmax_coredump(struct ctl_table *table, int write,
 185                 void __user *buffer, size_t *lenp, loff_t *ppos);
 186 #ifdef CONFIG_COREDUMP
 187 static int proc_dostring_coredump(struct ctl_table *table, int write,
 188                 void __user *buffer, size_t *lenp, loff_t *ppos);
 189 #endif
 190
 191 #ifdef CONFIG_MAGIC_SYSRQ
 192 /* Note: sysrq code uses it's own private copy */
 193 static int __sysrq_enabled = CONFIG_MAGIC_SYSRQ_DEFAULT_ENABLE;
 194
 195 static int sysrq_sysctl_handler(ctl_table *table, int write,
 196                                 void __user *buffer, size_t *lenp,
 197                                 loff_t *ppos)
 198 {
 199         int error;
 200
 201         error = proc_dointvec(table, write, buffer, lenp, ppos);
 202         if (error)
 203                 return error;
 204
 205         if (write)
 206                 sysrq_toggle_support(__sysrq_enabled);
 207
 208         return 0;
 209 }
 210
 211 #endif
 212
 213 static struct ctl_table kern_table[];
 214 static struct ctl_table vm_table[];
 215 static struct ctl_table fs_table[];
 216 static struct ctl_table debug_table[];
 217 static struct ctl_table dev_table[];
 218 extern struct ctl_table random_table[];
 219 #ifdef CONFIG_EPOLL
 220 extern struct ctl_table epoll_table[];
 221 #endif
 222
 223 #ifdef HAVE_ARCH_PICK_MMAP_LAYOUT
 224 int sysctl_legacy_va_layout;
 225 #endif
 226
 227 /* The default sysctl tables: */
 228
 229 static struct ctl_table sysctl_base_table[] = {
 230         {
 231                 .procname       = "kernel",
 232                 .mode           = 0555,
 233                 .child          = kern_table,
 234         },
 235         {
 236                 .procname       = "vm",
 237                 .mode           = 0555,
 238                 .child          = vm_table,
 239         },
 240         {
 241                 .procname       = "fs",
 242                 .mode           = 0555,
 243                 .child          = fs_table,
 244         },
 245         {
 246                 .procname       = "debug",
 247                 .mode           = 0555,
 248                 .child          = debug_table,
 249         },
 250         {
 251                 .procname       = "dev",
 252                 .mode           = 0555,
 253                 .child          = dev_table,
 254         },
 255         { }
 256 };
 257
 258 #ifdef CONFIG_SCHED_DEBUG
 259 static int min_sched_granularity_ns = 100000;           /* 100 usecs */
 260 static int max_sched_granularity_ns = NSEC_PER_SEC;     /* 1 second */
 261 static int min_wakeup_granularity_ns;                   /* 0 usecs */
 262 static int max_wakeup_granularity_ns = NSEC_PER_SEC;    /* 1 second */
 263 #ifdef CONFIG_SMP
 264 static int min_sched_tunable_scaling = SCHED_TUNABLESCALING_NONE;
 265 static int max_sched_tunable_scaling = SCHED_TUNABLESCALING_END-1;
 266 #endif /* CONFIG_SMP */
 267 #endif /* CONFIG_SCHED_DEBUG */
 268
 269 #ifdef CONFIG_COMPACTION
 270 static int min_extfrag_threshold;
 271 static int max_extfrag_threshold = 1000;
 272 #endif
 273
 274 static struct ctl_table kern_table[] = {
 275         {
 276                 .procname       = "sched_child_runs_first",
 277                 .data           = &sysctl_sched_child_runs_first,
 278                 .maxlen         = sizeof(unsigned int),
 279                 .mode           = 0644,
 280                 .proc_handler   = proc_dointvec,
 281         },
 282 #ifdef CONFIG_SCHED_DEBUG
 283         {
 284                 .procname       = "sched_min_granularity_ns",
 285                 .data           = &sysctl_sched_min_granularity,
 286                 .maxlen         = sizeof(unsigned int),
 287                 .mode           = 0644,
 288                 .proc_handler   = sched_proc_update_handler,
 289                 .extra1         = &min_sched_granularity_ns,
 290                 .extra2         = &max_sched_granularity_ns,
 291         },
 292         {
 293                 .procname       = "sched_latency_ns",
 294                 .data           = &sysctl_sched_latency,
 295                 .maxlen         = sizeof(unsigned int),
 296                 .mode           = 0644,
 297                 .proc_handler   = sched_proc_update_handler,
 298                 .extra1         = &min_sched_granularity_ns,
 299                 .extra2         = &max_sched_granularity_ns,
 300         },
 301         {
 302                 .procname       = "sched_wakeup_granularity_ns",
 303                 .data           = &sysctl_sched_wakeup_granularity,
 304                 .maxlen         = sizeof(unsigned int),
 305                 .mode           = 0644,
 306                 .proc_handler   = sched_proc_update_handler,
 307                 .extra1         = &min_wakeup_granularity_ns,
 308                 .extra2         = &max_wakeup_granularity_ns,
 309         },
 310 #ifdef CONFIG_SMP
 311         {
 312                 .procname       = "sched_tunable_scaling",
 313                 .data           = &sysctl_sched_tunable_scaling,
 314                 .maxlen         = sizeof(enum sched_tunable_scaling),
 315                 .mode           = 0644,
 316                 .proc_handler   = sched_proc_update_handler,
 317                 .extra1         = &min_sched_tunable_scaling,
 318                 .extra2         = &max_sched_tunable_scaling,
 319         },
 320         {
 321                 .procname       = "sched_migration_cost_ns",
 322                 .data           = &sysctl_sched_migration_cost,
 323                 .maxlen         = sizeof(unsigned int),
 324                 .mode           = 0644,
 325                 .proc_handler   = proc_dointvec,
 326         },
 327         {
 328                 .procname       = "sched_nr_migrate",
 329                 .data           = &sysctl_sched_nr_migrate,
 330                 .maxlen         = sizeof(unsigned int),
 331                 .mode           = 0644,
 332                 .proc_handler   = proc_dointvec,
 333         },
 334         {
 335                 .procname       = "sched_time_avg_ms",
 336                 .data           = &sysctl_sched_time_avg,
 337                 .maxlen         = sizeof(unsigned int),
 338                 .mode           = 0644,
 339                 .proc_handler   = proc_dointvec,
 340         },
 341         {
 342                 .procname       = "sched_shares_window_ns",
 343                 .data           = &sysctl_sched_shares_window,
 344                 .maxlen         = sizeof(unsigned int),
 345                 .mode           = 0644,
 346                 .proc_handler   = proc_dointvec,
 347         },
 348         {
 349                 .procname       = "timer_migration",
 350                 .data           = &sysctl_timer_migration,
 351                 .maxlen         = sizeof(unsigned int),
 352                 .mode           = 0644,
 353                 .proc_handler   = proc_dointvec_minmax,
 354                 .extra1         = &zero,
 355                 .extra2         = &one,
 356         },
 357 #endif /* CONFIG_SMP */
 358 #ifdef CONFIG_NUMA_BALANCING
 359         {
 360                 .procname       = "numa_balancing_scan_delay_ms",
 361                 .data           = &sysctl_numa_balancing_scan_delay,
 362                 .maxlen         = sizeof(unsigned int),
 363                 .mode           = 0644,
 364                 .proc_handler   = proc_dointvec,
 365         },
 366         {
 367                 .procname       = "numa_balancing_scan_period_min_ms",
 368                 .data           = &sysctl_numa_balancing_scan_period_min,
 369                 .maxlen         = sizeof(unsigned int),
 370                 .mode           = 0644,
 371                 .proc_handler   = proc_dointvec,
 372         },
 373         {
 374                 .procname       = "numa_balancing_scan_period_max_ms",
 375                 .data           = &sysctl_numa_balancing_scan_period_max,
 376                 .maxlen         = sizeof(unsigned int),
 377                 .mode           = 0644,
 378                 .proc_handler   = proc_dointvec,
 379         },
 380         {
 381                 .procname       = "numa_balancing_scan_size_mb",
 382                 .data           = &sysctl_numa_balancing_scan_size,
 383                 .maxlen         = sizeof(unsigned int),
 384                 .mode           = 0644,
 385                 .proc_handler   = proc_dointvec,
 386         },
 387         {
 388                 .procname       = "numa_balancing_migrate_deferred",
 389                 .data           = &sysctl_numa_balancing_migrate_deferred,
 390                 .maxlen         = sizeof(unsigned int),
 391                 .mode           = 0644,
 392                 .proc_handler   = proc_dointvec,
 393         },
 394 #endif /* CONFIG_NUMA_BALANCING */
 395 #endif /* CONFIG_SCHED_DEBUG */
 396         {
 397                 .procname       = "sched_rt_period_us",
 398                 .data           = &sysctl_sched_rt_period,
 399                 .maxlen         = sizeof(unsigned int),
 400                 .mode           = 0644,
 401                 .proc_handler   = sched_rt_handler,
 402         },
 403         {
 404                 .procname       = "sched_rt_runtime_us",
 405                 .data           = &sysctl_sched_rt_runtime,
 406                 .maxlen         = sizeof(int),
 407                 .mode           = 0644,
 408                 .proc_handler   = sched_rt_handler,
 409         },
 410         {
 411                 .procname       = "sched_rr_timeslice_ms",
 412                 .data           = &sched_rr_timeslice,
 413                 .maxlen         = sizeof(int),
 414                 .mode           = 0644,
 415                 .proc_handler   = sched_rr_handler,
 416         },
 417 #ifdef CONFIG_SCHED_AUTOGROUP
 418         {
 419                 .procname       = "sched_autogroup_enabled",
 420                 .data           = &sysctl_sched_autogroup_enabled,
 421                 .maxlen         = sizeof(unsigned int),
 422                 .mode           = 0644,
 423                 .proc_handler   = proc_dointvec_minmax,
 424                 .extra1         = &zero,
 425                 .extra2         = &one,
 426         },
 427 #endif
 428 #ifdef CONFIG_CFS_BANDWIDTH
 429         {
 430                 .procname       = "sched_cfs_bandwidth_slice_us",
 431                 .data           = &sysctl_sched_cfs_bandwidth_slice,
 432                 .maxlen         = sizeof(unsigned int),
 433                 .mode           = 0644,
 434                 .proc_handler   = proc_dointvec_minmax,
 435                 .extra1         = &one,
 436         },
 437 #endif
 438 #ifdef CONFIG_PROVE_LOCKING
 439         {
 440                 .procname       = "prove_locking",
 441                 .data           = &prove_locking,
 442                 .maxlen         = sizeof(int),
 443                 .mode           = 0644,
 444                 .proc_handler   = proc_dointvec,
 445         },
 446 #endif
 447 #ifdef CONFIG_LOCK_STAT
 448         {
 449                 .procname       = "lock_stat",
 450                 .data           = &lock_stat,
 451                 .maxlen         = sizeof(int),
 452                 .mode           = 0644,
 453                 .proc_handler   = proc_dointvec,
 454         },
 455 #endif
 456         {
 457                 .procname       = "panic",
 458                 .data           = &panic_timeout,
 459                 .maxlen         = sizeof(int),
 460                 .mode           = 0644,
 461                 .proc_handler   = proc_dointvec,
 462         },
 463 #ifdef CONFIG_COREDUMP
 464         {
 465                 .procname       = "core_uses_pid",
 466                 .data           = &core_uses_pid,
 467                 .maxlen         = sizeof(int),
 468                 .mode           = 0644,
 469                 .proc_handler   = proc_dointvec,
 470         },
 471         {
 472                 .procname       = "core_pattern",
 473                 .data           = core_pattern,
 474                 .maxlen         = CORENAME_MAX_SIZE,
 475                 .mode           = 0644,
 476                 .proc_handler   = proc_dostring_coredump,
 477         },
 478         {
 479                 .procname       = "core_pipe_limit",
 480                 .data           = &core_pipe_limit,
 481                 .maxlen         = sizeof(unsigned int),
 482                 .mode           = 0644,
 483                 .proc_handler   = proc_dointvec,
 484         },
 485 #endif
 486 #ifdef CONFIG_PROC_SYSCTL
 487         {
 488                 .procname       = "tainted",
 489                 .maxlen         = sizeof(long),
 490                 .mode           = 0644,
 491                 .proc_handler   = proc_taint,
 492         },
 493 #endif
 494 #ifdef CONFIG_LATENCYTOP
 495         {
 496                 .procname       = "latencytop",
 497                 .data           = &latencytop_enabled,
 498                 .maxlen         = sizeof(int),
 499                 .mode           = 0644,
 500                 .proc_handler   = proc_dointvec,
 501         },
 502 #endif
 503 #ifdef CONFIG_BLK_DEV_INITRD
 504         {
 505                 .procname       = "real-root-dev",
 506                 .data           = &real_root_dev,
 507                 .maxlen         = sizeof(int),
 508                 .mode           = 0644,
 509                 .proc_handler   = proc_dointvec,
 510         },
 511 #endif
 512         {
 513                 .procname       = "print-fatal-signals",
 514                 .data           = &print_fatal_signals,
 515                 .maxlen         = sizeof(int),
 516                 .mode           = 0644,
 517                 .proc_handler   = proc_dointvec,
 518         },
 519 #ifdef CONFIG_SPARC
 520         {
 521                 .procname       = "reboot-cmd",
 522                 .data           = reboot_command,
 523                 .maxlen         = 256,
 524                 .mode           = 0644,
 525                 .proc_handler   = proc_dostring,
 526         },
 527         {
 528                 .procname       = "stop-a",
 529                 .data           = &stop_a_enabled,
 530                 .maxlen         = sizeof (int),
 531                 .mode           = 0644,
 532                 .proc_handler   = proc_dointvec,
 533         },
 534         {
 535                 .procname       = "scons-poweroff",
 536                 .data           = &scons_pwroff,
 537                 .maxlen         = sizeof (int),
 538                 .mode           = 0644,
 539                 .proc_handler   = proc_dointvec,
 540         },
 541 #endif
 542 #ifdef CONFIG_SPARC64
 543         {
 544                 .procname       = "tsb-ratio",
 545                 .data           = &sysctl_tsb_ratio,
 546                 .maxlen         = sizeof (int),
 547                 .mode           = 0644,
 548                 .proc_handler   = proc_dointvec,
 549         },
 550 #endif
 551 #ifdef __hppa__
 552         {
 553                 .procname       = "soft-power",
 554                 .data           = &pwrsw_enabled,
 555                 .maxlen         = sizeof (int),
 556                 .mode           = 0644,
 557                 .proc_handler   = proc_dointvec,
 558         },
 559 #endif
 560 #ifdef CONFIG_SYSCTL_ARCH_UNALIGN_ALLOW
 561         {
 562                 .procname       = "unaligned-trap",
 563                 .data           = &unaligned_enabled,
 564                 .maxlen         = sizeof (int),
 565                 .mode           = 0644,
 566                 .proc_handler   = proc_dointvec,
 567         },
 568 #endif
 569         {
 570                 .procname       = "ctrl-alt-del",
 571                 .data           = &C_A_D,
 572                 .maxlen         = sizeof(int),
 573                 .mode           = 0644,
 574                 .proc_handler   = proc_dointvec,
 575         },
 576 #ifdef CONFIG_FUNCTION_TRACER
 577         {
 578                 .procname       = "ftrace_enabled",
 579                 .data           = &ftrace_enabled,
 580                 .maxlen         = sizeof(int),
 581                 .mode           = 0644,
 582                 .proc_handler   = ftrace_enable_sysctl,
 583         },
 584 #endif
 585 #ifdef CONFIG_STACK_TRACER
 586         {
 587                 .procname       = "stack_tracer_enabled",
 588                 .data           = &stack_tracer_enabled,
 589                 .maxlen         = sizeof(int),
 590                 .mode           = 0644,
 591                 .proc_handler   = stack_trace_sysctl,
 592         },
 593 #endif
 594 #ifdef CONFIG_TRACING
 595         {
 596                 .procname       = "ftrace_dump_on_oops",
 597                 .data           = &ftrace_dump_on_oops,
 598                 .maxlen         = sizeof(int),
 599                 .mode           = 0644,
 600                 .proc_handler   = proc_dointvec,
 601         },
 602         {
 603                 .procname       = "traceoff_on_warning",
 604                 .data           = &__disable_trace_on_warning,
 605                 .maxlen         = sizeof(__disable_trace_on_warning),
 606                 .mode           = 0644,
 607                 .proc_handler   = proc_dointvec,
 608         },
 609 #endif
 610 #ifdef CONFIG_MODULES
 611         {
 612                 .procname       = "modprobe",
 613                 .data           = &modprobe_path,
 614                 .maxlen         = KMOD_PATH_LEN,
 615                 .mode           = 0644,
 616                 .proc_handler   = proc_dostring,
 617         },
 618         {
 619                 .procname       = "modules_disabled",
 620                 .data           = &modules_disabled,
 621                 .maxlen         = sizeof(int),
 622                 .mode           = 0644,
 623                 /* only handle a transition from default "0" to "1" */
 624                 .proc_handler   = proc_dointvec_minmax,
 625                 .extra1         = &one,
 626                 .extra2         = &one,
 627         },
 628 #endif
 629
 630         {
 631                 .procname       = "hotplug",
 632                 .data           = &uevent_helper,
 633                 .maxlen         = UEVENT_HELPER_PATH_LEN,
 634                 .mode           = 0644,
 635                 .proc_handler   = proc_dostring,
 636         },
 637
 638 #ifdef CONFIG_CHR_DEV_SG
 639         {
 640                 .procname       = "sg-big-buff",
 641                 .data           = &sg_big_buff,
 642                 .maxlen         = sizeof (int),
 643                 .mode           = 0444,
 644                 .proc_handler   = proc_dointvec,
 645         },
 646 #endif
 647 #ifdef CONFIG_BSD_PROCESS_ACCT
 648         {
 649                 .procname       = "acct",
 650                 .data           = &acct_parm,
 651                 .maxlen         = 3*sizeof(int),
 652                 .mode           = 0644,
 653                 .proc_handler   = proc_dointvec,
 654         },
 655 #endif
 656 #ifdef CONFIG_MAGIC_SYSRQ
 657         {
 658                 .procname       = "sysrq",
 659                 .data           = &__sysrq_enabled,
 660                 .maxlen         = sizeof (int),
 661                 .mode           = 0644,
 662                 .proc_handler   = sysrq_sysctl_handler,
 663         },
 664 #endif
 665 #ifdef CONFIG_PROC_SYSCTL
 666         {
 667                 .procname       = "cad_pid",
 668                 .data           = NULL,
 669                 .maxlen         = sizeof (int),
 670                 .mode           = 0600,
 671                 .proc_handler   = proc_do_cad_pid,
 672         },
 673 #endif
 674         {
 675                 .procname       = "threads-max",
 676                 .data           = &max_threads,
 677                 .maxlen         = sizeof(int),
 678                 .mode           = 0644,
 679                 .proc_handler   = proc_dointvec,
 680         },
 681         {
 682                 .procname       = "random",
 683                 .mode           = 0555,
 684                 .child          = random_table,
 685         },
 686         {
 687                 .procname       = "usermodehelper",
 688                 .mode           = 0555,
 689                 .child          = usermodehelper_table,
 690         },
 691         {
 692                 .procname       = "overflowuid",
 693                 .data           = &overflowuid,
 694                 .maxlen         = sizeof(int),
 695                 .mode           = 0644,
 696                 .proc_handler   = proc_dointvec_minmax,
 697                 .extra1         = &minolduid,
 698                 .extra2         = &maxolduid,
 699         },
 700         {
 701                 .procname       = "overflowgid",
 702                 .data           = &overflowgid,
 703                 .maxlen         = sizeof(int),
 704                 .mode           = 0644,
 705                 .proc_handler   = proc_dointvec_minmax,
 706                 .extra1         = &minolduid,
 707                 .extra2         = &maxolduid,
 708         },
 709 #ifdef CONFIG_S390
 710 #ifdef CONFIG_MATHEMU
 711         {
 712                 .procname       = "ieee_emulation_warnings",
 713                 .data           = &sysctl_ieee_emulation_warnings,
 714                 .maxlen         = sizeof(int),
 715                 .mode           = 0644,
 716                 .proc_handler   = proc_dointvec,
 717         },
 718 #endif
 719         {
 720                 .procname       = "userprocess_debug",
 721                 .data           = &show_unhandled_signals,
 722                 .maxlen         = sizeof(int),
 723                 .mode           = 0644,
 724                 .proc_handler   = proc_dointvec,
 725         },
 726 #endif
 727         {
 728                 .procname       = "pid_max",
 729                 .data           = &pid_max,
 730                 .maxlen         = sizeof (int),
 731                 .mode           = 0644,
 732                 .proc_handler   = proc_dointvec_minmax,
 733                 .extra1         = &pid_max_min,
 734                 .extra2         = &pid_max_max,
 735         },
 736         {
 737                 .procname       = "panic_on_oops",
 738                 .data           = &panic_on_oops,
 739                 .maxlen         = sizeof(int),
 740                 .mode           = 0644,
 741                 .proc_handler   = proc_dointvec,
 742         },
 743 #if defined CONFIG_PRINTK
 744         {
 745                 .procname       = "printk",
 746                 .data           = &console_loglevel,
 747                 .maxlen         = 4*sizeof(int),
 748                 .mode           = 0644,
 749                 .proc_handler   = proc_dointvec,
 750         },
 751         {
 752                 .procname       = "printk_ratelimit",
 753                 .data           = &printk_ratelimit_state.interval,
 754                 .maxlen         = sizeof(int),
 755                 .mode           = 0644,
 756                 .proc_handler   = proc_dointvec_jiffies,
 757         },
 758         {
 759                 .procname       = "printk_ratelimit_burst",
 760                 .data           = &printk_ratelimit_state.burst,
 761                 .maxlen         = sizeof(int),
 762                 .mode           = 0644,
 763                 .proc_handler   = proc_dointvec,
 764         },
 765         {
 766                 .procname       = "printk_delay",
 767                 .data           = &printk_delay_msec,
 768                 .maxlen         = sizeof(int),
 769                 .mode           = 0644,
 770                 .proc_handler   = proc_dointvec_minmax,
 771                 .extra1         = &zero,
 772                 .extra2         = &ten_thousand,
 773         },
 774         {
 775                 .procname       = "dmesg_restrict",
 776                 .data           = &dmesg_restrict,
 777                 .maxlen         = sizeof(int),
 778                 .mode           = 0644,
 779                 .proc_handler   = proc_dointvec_minmax_sysadmin,
 780                 .extra1         = &zero,
 781                 .extra2         = &one,
 782         },
 783         {
 784                 .procname       = "kptr_restrict",
 785                 .data           = &kptr_restrict,
 786                 .maxlen         = sizeof(int),
 787                 .mode           = 0644,
 788                 .proc_handler   = proc_dointvec_minmax_sysadmin,
 789                 .extra1         = &zero,
 790                 .extra2         = &two,
 791         },
 792 #endif
 793         {
 794                 .procname       = "ngroups_max",
 795                 .data           = &ngroups_max,
 796                 .maxlen         = sizeof (int),
 797                 .mode           = 0444,
 798                 .proc_handler   = proc_dointvec,
 799         },
 800         {
 801                 .procname       = "cap_last_cap",
 802                 .data           = (void *)&cap_last_cap,
 803                 .maxlen         = sizeof(int),
 804                 .mode           = 0444,
 805                 .proc_handler   = proc_dointvec,
 806         },
 807 #if defined(CONFIG_LOCKUP_DETECTOR)
 808         {
 809                 .procname       = "watchdog",
 810                 .data           = &watchdog_user_enabled,
 811                 .maxlen         = sizeof (int),
 812                 .mode           = 0644,
 813                 .proc_handler   = proc_dowatchdog,
 814                 .extra1         = &zero,
 815                 .extra2         = &one,
 816         },
 817         {
 818                 .procname       = "watchdog_thresh",
 819                 .data           = &watchdog_thresh,
 820                 .maxlen         = sizeof(int),
 821                 .mode           = 0644,
 822                 .proc_handler   = proc_dowatchdog,
 823                 .extra1         = &zero,
 824                 .extra2         = &sixty,
 825         },
 826         {
 827                 .procname       = "softlockup_panic",
 828                 .data           = &softlockup_panic,
 829                 .maxlen         = sizeof(int),
 830                 .mode           = 0644,
 831                 .proc_handler   = proc_dointvec_minmax,
 832                 .extra1         = &zero,
 833                 .extra2         = &one,
 834         },
 835         {
 836                 .procname       = "nmi_watchdog",
 837                 .data           = &watchdog_user_enabled,
 838                 .maxlen         = sizeof (int),
 839                 .mode           = 0644,
 840                 .proc_handler   = proc_dowatchdog,
 841                 .extra1         = &zero,
 842                 .extra2         = &one,
 843         },
 844 #endif
 845 #if defined(CONFIG_X86_LOCAL_APIC) && defined(CONFIG_X86)
 846         {
 847                 .procname       = "unknown_nmi_panic",
 848                 .data           = &unknown_nmi_panic,
 849                 .maxlen         = sizeof (int),
 850                 .mode           = 0644,
 851                 .proc_handler   = proc_dointvec,
 852         },
 853 #endif
 854 #if defined(CONFIG_X86)
 855         {
 856                 .procname       = "panic_on_unrecovered_nmi",
 857                 .data           = &panic_on_unrecovered_nmi,
 858                 .maxlen         = sizeof(int),
 859                 .mode           = 0644,
 860                 .proc_handler   = proc_dointvec,
 861         },
 862         {
 863                 .procname       = "panic_on_io_nmi",
 864                 .data           = &panic_on_io_nmi,
 865                 .maxlen         = sizeof(int),
 866                 .mode           = 0644,
 867                 .proc_handler   = proc_dointvec,
 868         },
 869 #ifdef CONFIG_DEBUG_STACKOVERFLOW
 870         {
 871                 .procname       = "panic_on_stackoverflow",
 872                 .data           = &sysctl_panic_on_stackoverflow,
 873                 .maxlen         = sizeof(int),
 874                 .mode           = 0644,
 875                 .proc_handler   = proc_dointvec,
 876         },
 877 #endif
 878         {
 879                 .procname       = "bootloader_type",
 880                 .data           = &bootloader_type,
 881                 .maxlen         = sizeof (int),
 882                 .mode           = 0444,
 883                 .proc_handler   = proc_dointvec,
 884         },
 885         {
 886                 .procname       = "bootloader_version",
 887                 .data           = &bootloader_version,
 888                 .maxlen         = sizeof (int),
 889                 .mode           = 0444,
 890                 .proc_handler   = proc_dointvec,
 891         },
 892         {
 893                 .procname       = "kstack_depth_to_print",
 894                 .data           = &kstack_depth_to_print,
 895                 .maxlen         = sizeof(int),
 896                 .mode           = 0644,
 897                 .proc_handler   = proc_dointvec,
 898         },
 899         {
 900                 .procname       = "io_delay_type",
 901                 .data           = &io_delay_type,
 902                 .maxlen         = sizeof(int),
 903                 .mode           = 0644,
 904                 .proc_handler   = proc_dointvec,
 905         },
 906 #endif
 907 #if defined(CONFIG_MMU)
 908         {
 909                 .procname       = "randomize_va_space",
 910                 .data           = &randomize_va_space,
 911                 .maxlen         = sizeof(int),
 912                 .mode           = 0644,
 913                 .proc_handler   = proc_dointvec,
 914         },
 915 #endif
 916 #if defined(CONFIG_S390) && defined(CONFIG_SMP)
 917         {
 918                 .procname       = "spin_retry",
 919                 .data           = &spin_retry,
 920                 .maxlen         = sizeof (int),
 921                 .mode           = 0644,
 922                 .proc_handler   = proc_dointvec,
 923         },
 924 #endif
 925 #if     defined(CONFIG_ACPI_SLEEP) && defined(CONFIG_X86)
 926         {
 927                 .procname       = "acpi_video_flags",
 928                 .data           = &acpi_realmode_flags,
 929                 .maxlen         = sizeof (unsigned long),
 930                 .mode           = 0644,
 931                 .proc_handler   = proc_doulongvec_minmax,
 932         },
 933 #endif
 934 #ifdef CONFIG_SYSCTL_ARCH_UNALIGN_NO_WARN
 935         {
 936                 .procname       = "ignore-unaligned-usertrap",
 937                 .data           = &no_unaligned_warning,
 938                 .maxlen         = sizeof (int),
 939                 .mode           = 0644,
 940                 .proc_handler   = proc_dointvec,
 941         },
 942 #endif
 943 #ifdef CONFIG_IA64
 944         {
 945                 .procname       = "unaligned-dump-stack",
 946                 .data           = &unaligned_dump_stack,
 947                 .maxlen         = sizeof (int),
 948                 .mode           = 0644,
 949                 .proc_handler   = proc_dointvec,
 950         },
 951 #endif
 952 #ifdef CONFIG_DETECT_HUNG_TASK
 953         {
 954                 .procname       = "hung_task_panic",
 955                 .data           = &sysctl_hung_task_panic,
 956                 .maxlen         = sizeof(int),
 957                 .mode           = 0644,
 958                 .proc_handler   = proc_dointvec_minmax,
 959                 .extra1         = &zero,
 960                 .extra2         = &one,
 961         },
 962         {
 963                 .procname       = "hung_task_check_count",
 964                 .data           = &sysctl_hung_task_check_count,
 965                 .maxlen         = sizeof(int),
 966                 .mode           = 0644,
 967                 .proc_handler   = proc_dointvec_minmax,
 968                 .extra1         = &zero,
 969         },
 970         {
 971                 .procname       = "hung_task_timeout_secs",
 972                 .data           = &sysctl_hung_task_timeout_secs,
 973                 .maxlen         = sizeof(unsigned long),
 974                 .mode           = 0644,
 975                 .proc_handler   = proc_dohung_task_timeout_secs,
 976         },
 977         {
 978                 .procname       = "hung_task_warnings",
 979                 .data           = &sysctl_hung_task_warnings,
 980                 .maxlen         = sizeof(unsigned long),
 981                 .mode           = 0644,
 982                 .proc_handler   = proc_doulongvec_minmax,
 983         },
 984 #endif
 985 #ifdef CONFIG_COMPAT
 986         {
 987                 .procname       = "compat-log",
 988                 .data           = &compat_log,
 989                 .maxlen         = sizeof (int),
 990                 .mode           = 0644,
 991                 .proc_handler   = proc_dointvec,
 992         },
 993 #endif
 994 #ifdef CONFIG_RT_MUTEXES
 995         {
 996                 .procname       = "max_lock_depth",
 997                 .data           = &max_lock_depth,
 998                 .maxlen         = sizeof(int),
 999                 .mode           = 0644,
1000                 .proc_handler   = proc_dointvec,
1001         },
1002 #endif
1003         {
1004                 .procname       = "poweroff_cmd",
1005                 .data           = &poweroff_cmd,
1006                 .maxlen         = POWEROFF_CMD_PATH_LEN,
1007                 .mode           = 0644,
1008                 .proc_handler   = proc_dostring,
1009         },
1010 #ifdef CONFIG_KEYS
1011         {
1012                 .procname       = "keys",
1013                 .mode           = 0555,
1014                 .child          = key_sysctls,
1015         },
1016 #endif
1017 #ifdef CONFIG_RCU_TORTURE_TEST
1018         {
1019                 .procname       = "rcutorture_runnable",
1020                 .data           = &rcutorture_runnable,
1021                 .maxlen         = sizeof(int),
1022                 .mode           = 0644,
1023                 .proc_handler   = proc_dointvec,
1024         },
1025 #endif
1026 #ifdef CONFIG_PERF_EVENTS
1027         /*
1028          * User-space scripts rely on the existence of this file
1029          * as a feature check for perf_events being enabled.
1030          *
1031          * So it's an ABI, do not remove!
1032          */
1033         {
1034                 .procname       = "perf_event_paranoid",
1035                 .data           = &sysctl_perf_event_paranoid,
1036                 .maxlen         = sizeof(sysctl_perf_event_paranoid),
1037                 .mode           = 0644,
1038                 .proc_handler   = proc_dointvec,
1039         },
1040         {
1041                 .procname       = "perf_event_mlock_kb",
1042                 .data           = &sysctl_perf_event_mlock,
1043                 .maxlen         = sizeof(sysctl_perf_event_mlock),
1044                 .mode           = 0644,
1045                 .proc_handler   = proc_dointvec,
1046         },
1047         {
1048                 .procname       = "perf_event_max_sample_rate",
1049                 .data           = &sysctl_perf_event_sample_rate,
1050                 .maxlen         = sizeof(sysctl_perf_event_sample_rate),
1051                 .mode           = 0644,
1052                 .proc_handler   = perf_proc_update_handler,
1053                 .extra1         = &one,
1054         },
1055         {
1056                 .procname       = "perf_cpu_time_max_percent",
1057                 .data           = &sysctl_perf_cpu_time_max_percent,
1058                 .maxlen         = sizeof(sysctl_perf_cpu_time_max_percent),
1059                 .mode           = 0644,
1060                 .proc_handler   = perf_cpu_time_max_percent_handler,
1061                 .extra1         = &zero,
1062                 .extra2         = &one_hundred,
1063         },
1064 #endif
1065 #ifdef CONFIG_KMEMCHECK
1066         {
1067                 .procname       = "kmemcheck",
1068                 .data           = &kmemcheck_enabled,
1069                 .maxlen         = sizeof(int),
1070                 .mode           = 0644,
1071                 .proc_handler   = proc_dointvec,
1072         },
1073 #endif
1074 #ifdef CONFIG_BLOCK
1075         {
1076                 .procname       = "blk_iopoll",
1077                 .data           = &blk_iopoll_enabled,
1078                 .maxlen         = sizeof(int),
1079                 .mode           = 0644,
1080                 .proc_handler   = proc_dointvec,
1081         },
1082 #endif
1083         { }
1084 };
1085
1086 static struct ctl_table vm_table[] = {
1087         {
1088                 .procname       = "overcommit_memory",
1089                 .data           = &sysctl_overcommit_memory,
1090                 .maxlen         = sizeof(sysctl_overcommit_memory),
1091                 .mode           = 0644,
1092                 .proc_handler   = proc_dointvec_minmax,
1093                 .extra1         = &zero,
1094                 .extra2         = &two,
1095         },
1096         {
1097                 .procname       = "panic_on_oom",
1098                 .data           = &sysctl_panic_on_oom,
1099                 .maxlen         = sizeof(sysctl_panic_on_oom),
1100                 .mode           = 0644,
1101                 .proc_handler   = proc_dointvec_minmax,
1102                 .extra1         = &zero,
1103                 .extra2         = &two,
1104         },
1105         {
1106                 .procname       = "oom_kill_allocating_task",
1107                 .data           = &sysctl_oom_kill_allocating_task,
1108                 .maxlen         = sizeof(sysctl_oom_kill_allocating_task),
1109                 .mode           = 0644,
1110                 .proc_handler   = proc_dointvec,
1111         },
1112         {
1113                 .procname       = "oom_dump_tasks",
1114                 .data           = &sysctl_oom_dump_tasks,
1115                 .maxlen         = sizeof(sysctl_oom_dump_tasks),
1116                 .mode           = 0644,
1117                 .proc_handler   = proc_dointvec,
1118         },
1119         {
1120                 .procname       = "overcommit_ratio",
1121                 .data           = &sysctl_overcommit_ratio,
1122                 .maxlen         = sizeof(sysctl_overcommit_ratio),
1123                 .mode           = 0644,
1124                 .proc_handler   = proc_dointvec,
1125         },
1126         {
1127                 .procname       = "page-cluster",
1128                 .data           = &page_cluster,
1129                 .maxlen         = sizeof(int),
1130                 .mode           = 0644,
1131                 .proc_handler   = proc_dointvec_minmax,
1132                 .extra1         = &zero,
1133         },
1134         {
1135                 .procname       = "dirty_background_ratio",
1136                 .data           = &dirty_background_ratio,
1137                 .maxlen         = sizeof(dirty_background_ratio),
1138                 .mode           = 0644,
1139                 .proc_handler   = dirty_background_ratio_handler,
1140                 .extra1         = &zero,
1141                 .extra2         = &one_hundred,
1142         },
1143         {
1144                 .procname       = "dirty_background_bytes",
1145                 .data           = &dirty_background_bytes,
1146                 .maxlen         = sizeof(dirty_background_bytes),
1147                 .mode           = 0644,
1148                 .proc_handler   = dirty_background_bytes_handler,
1149                 .extra1         = &one_ul,
1150         },
1151         {
1152                 .procname       = "dirty_ratio",
1153                 .data           = &vm_dirty_ratio,
1154                 .maxlen         = sizeof(vm_dirty_ratio),
1155                 .mode           = 0644,
1156                 .proc_handler   = dirty_ratio_handler,
1157                 .extra1         = &zero,
1158                 .extra2         = &one_hundred,
1159         },
1160         {
1161                 .procname       = "dirty_bytes",
1162                 .data           = &vm_dirty_bytes,
1163                 .maxlen         = sizeof(vm_dirty_bytes),
1164                 .mode           = 0644,
1165                 .proc_handler   = dirty_bytes_handler,
1166                 .extra1         = &dirty_bytes_min,
1167         },
1168         {
1169                 .procname       = "dirty_writeback_centisecs",
1170                 .data           = &dirty_writeback_interval,
1171                 .maxlen         = sizeof(dirty_writeback_interval),
1172                 .mode           = 0644,
1173                 .proc_handler   = dirty_writeback_centisecs_handler,
1174         },
1175         {
1176                 .procname       = "dirty_expire_centisecs",
1177                 .data           = &dirty_expire_interval,
1178                 .maxlen         = sizeof(dirty_expire_interval),
1179                 .mode           = 0644,
1180                 .proc_handler   = proc_dointvec_minmax,
1181                 .extra1         = &zero,
1182         },
1183         {
1184                 .procname       = "nr_pdflush_threads",
1185                 .mode           = 0444 /* read-only */,
1186                 .proc_handler   = pdflush_proc_obsolete,
1187         },
1188         {
1189                 .procname       = "swappiness",
1190                 .data           = &vm_swappiness,
1191                 .maxlen         = sizeof(vm_swappiness),
1192                 .mode           = 0644,
1193                 .proc_handler   = proc_dointvec_minmax,
1194                 .extra1         = &zero,
1195                 .extra2         = &one_hundred,
1196         },
1197 #ifdef CONFIG_HUGETLB_PAGE
1198         {
1199                 .procname       = "nr_hugepages",
1200                 .data           = NULL,
1201                 .maxlen         = sizeof(unsigned long),
1202                 .mode           = 0644,
1203                 .proc_handler   = hugetlb_sysctl_handler,
1204                 .extra1         = (void *)&hugetlb_zero,
1205                 .extra2         = (void *)&hugetlb_infinity,
1206         },
1207 #ifdef CONFIG_NUMA
1208         {
1209                 .procname       = "nr_hugepages_mempolicy",
1210                 .data           = NULL,
1211                 .maxlen         = sizeof(unsigned long),
1212                 .mode           = 0644,
1213                 .proc_handler   = &hugetlb_mempolicy_sysctl_handler,
1214                 .extra1         = (void *)&hugetlb_zero,
1215                 .extra2         = (void *)&hugetlb_infinity,
1216         },
1217 #endif
1218          {
1219                 .procname       = "hugetlb_shm_group",
1220                 .data           = &sysctl_hugetlb_shm_group,
1221                 .maxlen         = sizeof(gid_t),
1222                 .mode           = 0644,
1223                 .proc_handler   = proc_dointvec,
1224          },
1225          {
1226                 .procname       = "hugepages_treat_as_movable",
1227                 .data           = &hugepages_treat_as_movable,
1228                 .maxlen         = sizeof(int),
1229                 .mode           = 0644,
1230                 .proc_handler   = proc_dointvec,
1231         },
1232         {
1233                 .procname       = "nr_overcommit_hugepages",
1234                 .data           = NULL,
1235                 .maxlen         = sizeof(unsigned long),
1236                 .mode           = 0644,
1237                 .proc_handler   = hugetlb_overcommit_handler,
1238                 .extra1         = (void *)&hugetlb_zero,
1239                 .extra2         = (void *)&hugetlb_infinity,
1240         },
1241 #endif
1242         {
1243                 .procname       = "lowmem_reserve_ratio",
1244                 .data           = &sysctl_lowmem_reserve_ratio,
1245                 .maxlen         = sizeof(sysctl_lowmem_reserve_ratio),
1246                 .mode           = 0644,
1247                 .proc_handler   = lowmem_reserve_ratio_sysctl_handler,
1248         },
1249         {
1250                 .procname       = "drop_caches",
1251                 .data           = &sysctl_drop_caches,
1252                 .maxlen         = sizeof(int),
1253                 .mode           = 0644,
1254                 .proc_handler   = drop_caches_sysctl_handler,
1255                 .extra1         = &one,
1256                 .extra2         = &three,
1257         },
1258 #ifdef CONFIG_COMPACTION
1259         {
1260                 .procname       = "compact_memory",
1261                 .data           = &sysctl_compact_memory,
1262                 .maxlen         = sizeof(int),
1263                 .mode           = 0200,
1264                 .proc_handler   = sysctl_compaction_handler,
1265         },
1266         {
1267                 .procname       = "extfrag_threshold",
1268                 .data           = &sysctl_extfrag_threshold,
1269                 .maxlen         = sizeof(int),
1270                 .mode           = 0644,
1271                 .proc_handler   = sysctl_extfrag_handler,
1272                 .extra1         = &min_extfrag_threshold,
1273                 .extra2         = &max_extfrag_threshold,
1274         },
1275
1276 #endif /* CONFIG_COMPACTION */
1277         {
1278                 .procname       = "min_free_kbytes",
1279                 .data           = &min_free_kbytes,
1280                 .maxlen         = sizeof(min_free_kbytes),
1281                 .mode           = 0644,
1282                 .proc_handler   = min_free_kbytes_sysctl_handler,
1283                 .extra1         = &zero,
1284         },
1285         {
1286                 .procname       = "percpu_pagelist_fraction",
1287                 .data           = &percpu_pagelist_fraction,
1288                 .maxlen         = sizeof(percpu_pagelist_fraction),
1289                 .mode           = 0644,
1290                 .proc_handler   = percpu_pagelist_fraction_sysctl_handler,
1291                 .extra1         = &min_percpu_pagelist_fract,
1292         },
1293 #ifdef CONFIG_MMU
1294         {
1295                 .procname       = "max_map_count",
1296                 .data           = &sysctl_max_map_count,
1297                 .maxlen         = sizeof(sysctl_max_map_count),
1298                 .mode           = 0644,
1299                 .proc_handler   = proc_dointvec_minmax,
1300                 .extra1         = &zero,
1301         },
1302 #else
1303         {
1304                 .procname       = "nr_trim_pages",
1305                 .data           = &sysctl_nr_trim_pages,
1306                 .maxlen         = sizeof(sysctl_nr_trim_pages),
1307                 .mode           = 0644,
1308                 .proc_handler   = proc_dointvec_minmax,
1309                 .extra1         = &zero,
1310         },
1311 #endif
1312         {
1313                 .procname       = "laptop_mode",
1314                 .data           = &laptop_mode,
1315                 .maxlen         = sizeof(laptop_mode),
1316                 .mode           = 0644,
1317                 .proc_handler   = proc_dointvec_jiffies,
1318         },
1319         {
1320                 .procname       = "block_dump",
1321                 .data           = &block_dump,
1322                 .maxlen         = sizeof(block_dump),
1323                 .mode           = 0644,
1324                 .proc_handler   = proc_dointvec,
1325                 .extra1         = &zero,
1326         },
1327         {
1328                 .procname       = "vfs_cache_pressure",
1329                 .data           = &sysctl_vfs_cache_pressure,
1330                 .maxlen         = sizeof(sysctl_vfs_cache_pressure),
1331                 .mode           = 0644,
1332                 .proc_handler   = proc_dointvec,
1333                 .extra1         = &zero,
1334         },
1335 #ifdef HAVE_ARCH_PICK_MMAP_LAYOUT
1336         {
1337                 .procname       = "legacy_va_layout",
1338                 .data           = &sysctl_legacy_va_layout,
1339                 .maxlen         = sizeof(sysctl_legacy_va_layout),
1340                 .mode           = 0644,
1341                 .proc_handler   = proc_dointvec,
1342                 .extra1         = &zero,
1343         },
1344 #endif
1345 #ifdef CONFIG_NUMA
1346         {
1347                 .procname       = "zone_reclaim_mode",
1348                 .data           = &zone_reclaim_mode,
1349                 .maxlen         = sizeof(zone_reclaim_mode),
1350                 .mode           = 0644,
1351                 .proc_handler   = proc_dointvec,
1352                 .extra1         = &zero,
1353         },
1354         {
1355                 .procname       = "min_unmapped_ratio",
1356                 .data           = &sysctl_min_unmapped_ratio,
1357                 .maxlen         = sizeof(sysctl_min_unmapped_ratio),
1358                 .mode           = 0644,
1359                 .proc_handler   = sysctl_min_unmapped_ratio_sysctl_handler,
1360                 .extra1         = &zero,
1361                 .extra2         = &one_hundred,
1362         },
1363         {
1364                 .procname       = "min_slab_ratio",
1365                 .data           = &sysctl_min_slab_ratio,
1366                 .maxlen         = sizeof(sysctl_min_slab_ratio),
1367                 .mode           = 0644,
1368                 .proc_handler   = sysctl_min_slab_ratio_sysctl_handler,
1369                 .extra1         = &zero,
1370                 .extra2         = &one_hundred,
1371         },
1372 #endif
1373 #ifdef CONFIG_SMP
1374         {
1375                 .procname       = "stat_interval",
1376                 .data           = &sysctl_stat_interval,
1377                 .maxlen         = sizeof(sysctl_stat_interval),
1378                 .mode           = 0644,
1379                 .proc_handler   = proc_dointvec_jiffies,
1380         },
1381 #endif
1382 #ifdef CONFIG_MMU
1383         {
1384                 .procname       = "mmap_min_addr",
1385                 .data           = &dac_mmap_min_addr,
1386                 .maxlen         = sizeof(unsigned long),
1387                 .mode           = 0644,
1388                 .proc_handler   = mmap_min_addr_handler,
1389         },
1390 #endif
1391 #ifdef CONFIG_NUMA
1392         {
1393                 .procname       = "numa_zonelist_order",
1394                 .data           = &numa_zonelist_order,
1395                 .maxlen         = NUMA_ZONELIST_ORDER_LEN,
1396                 .mode           = 0644,
1397                 .proc_handler   = numa_zonelist_order_handler,
1398         },
1399 #endif
1400 #if (defined(CONFIG_X86_32) && !defined(CONFIG_UML))|| \
1401    (defined(CONFIG_SUPERH) && defined(CONFIG_VSYSCALL))
1402         {
1403                 .procname       = "vdso_enabled",
1404                 .data           = &vdso_enabled,
1405                 .maxlen         = sizeof(vdso_enabled),
1406                 .mode           = 0644,
1407                 .proc_handler   = proc_dointvec,
1408                 .extra1         = &zero,
1409         },
1410 #endif
1411 #ifdef CONFIG_HIGHMEM
1412         {
1413                 .procname       = "highmem_is_dirtyable",
1414                 .data           = &vm_highmem_is_dirtyable,
1415                 .maxlen         = sizeof(vm_highmem_is_dirtyable),
1416                 .mode           = 0644,
1417                 .proc_handler   = proc_dointvec_minmax,
1418                 .extra1         = &zero,
1419                 .extra2         = &one,
1420         },
1421 #endif
1422         {
1423                 .procname       = "scan_unevictable_pages",
1424                 .data           = &scan_unevictable_pages,
1425                 .maxlen         = sizeof(scan_unevictable_pages),
1426                 .mode           = 0644,
1427                 .proc_handler   = scan_unevictable_handler,
1428         },
1429 #ifdef CONFIG_MEMORY_FAILURE
1430         {
1431                 .procname       = "memory_failure_early_kill",
1432                 .data           = &sysctl_memory_failure_early_kill,
1433                 .maxlen         = sizeof(sysctl_memory_failure_early_kill),
1434                 .mode           = 0644,
1435                 .proc_handler   = proc_dointvec_minmax,
1436                 .extra1         = &zero,
1437                 .extra2         = &one,
1438         },
1439         {
1440                 .procname       = "memory_failure_recovery",
1441                 .data           = &sysctl_memory_failure_recovery,
1442                 .maxlen         = sizeof(sysctl_memory_failure_recovery),
1443                 .mode           = 0644,
1444                 .proc_handler   = proc_dointvec_minmax,
1445                 .extra1         = &zero,
1446                 .extra2         = &one,
1447         },
1448 #endif
1449         {
1450                 .procname       = "user_reserve_kbytes",
1451                 .data           = &sysctl_user_reserve_kbytes,
1452                 .maxlen         = sizeof(sysctl_user_reserve_kbytes),
1453                 .mode           = 0644,
1454                 .proc_handler   = proc_doulongvec_minmax,
1455         },
1456         {
1457                 .procname       = "admin_reserve_kbytes",
1458                 .data           = &sysctl_admin_reserve_kbytes,
1459                 .maxlen         = sizeof(sysctl_admin_reserve_kbytes),
1460                 .mode           = 0644,
1461                 .proc_handler   = proc_doulongvec_minmax,
1462         },
1463         { }
1464 };
1465
1466 #if defined(CONFIG_BINFMT_MISC) || defined(CONFIG_BINFMT_MISC_MODULE)
1467 static struct ctl_table binfmt_misc_table[] = {
1468         { }
1469 };
1470 #endif
1471
1472 static struct ctl_table fs_table[] = {
1473         {
1474                 .procname       = "inode-nr",
1475                 .data           = &inodes_stat,
1476                 .maxlen         = 2*sizeof(long),
1477                 .mode           = 0444,
1478                 .proc_handler   = proc_nr_inodes,
1479         },
1480         {
1481                 .procname       = "inode-state",
1482                 .data           = &inodes_stat,
1483                 .maxlen         = 7*sizeof(long),
1484                 .mode           = 0444,
1485                 .proc_handler   = proc_nr_inodes,
1486         },
1487         {
1488                 .procname       = "file-nr",
1489                 .data           = &files_stat,
1490                 .maxlen         = sizeof(files_stat),
1491                 .mode           = 0444,
1492                 .proc_handler   = proc_nr_files,
1493         },
1494         {
1495                 .procname       = "file-max",
1496                 .data           = &files_stat.max_files,
1497                 .maxlen         = sizeof(files_stat.max_files),
1498                 .mode           = 0644,
1499                 .proc_handler   = proc_doulongvec_minmax,
1500         },
1501         {
1502                 .procname       = "nr_open",
1503                 .data           = &sysctl_nr_open,
1504                 .maxlen         = sizeof(int),
1505                 .mode           = 0644,
1506                 .proc_handler   = proc_dointvec_minmax,
1507                 .extra1         = &sysctl_nr_open_min,
1508                 .extra2         = &sysctl_nr_open_max,
1509         },
1510         {
1511                 .procname       = "dentry-state",
1512                 .data           = &dentry_stat,
1513                 .maxlen         = 6*sizeof(long),
1514                 .mode           = 0444,
1515                 .proc_handler   = proc_nr_dentry,
1516         },
1517         {
1518                 .procname       = "overflowuid",
1519                 .data           = &fs_overflowuid,
1520                 .maxlen         = sizeof(int),
1521                 .mode           = 0644,
1522                 .proc_handler   = proc_dointvec_minmax,
1523                 .extra1         = &minolduid,
1524                 .extra2         = &maxolduid,
1525         },
1526         {
1527                 .procname       = "overflowgid",
1528                 .data           = &fs_overflowgid,
1529                 .maxlen         = sizeof(int),
1530                 .mode           = 0644,
1531                 .proc_handler   = proc_dointvec_minmax,
1532                 .extra1         = &minolduid,
1533                 .extra2         = &maxolduid,
1534         },
1535 #ifdef CONFIG_FILE_LOCKING
1536         {
1537                 .procname       = "leases-enable",
1538                 .data           = &leases_enable,
1539                 .maxlen         = sizeof(int),
1540                 .mode           = 0644,
1541                 .proc_handler   = proc_dointvec,
1542         },
1543 #endif
1544 #ifdef CONFIG_DNOTIFY
1545         {
1546                 .procname       = "dir-notify-enable",
1547                 .data           = &dir_notify_enable,
1548                 .maxlen         = sizeof(int),
1549                 .mode           = 0644,
1550                 .proc_handler   = proc_dointvec,
1551         },
1552 #endif
1553 #ifdef CONFIG_MMU
1554 #ifdef CONFIG_FILE_LOCKING
1555         {
1556                 .procname       = "lease-break-time",
1557                 .data           = &lease_break_time,
1558                 .maxlen         = sizeof(int),
1559                 .mode           = 0644,
1560                 .proc_handler   = proc_dointvec,
1561         },
1562 #endif
1563 #ifdef CONFIG_AIO
1564         {
1565                 .procname       = "aio-nr",
1566                 .data           = &aio_nr,
1567                 .maxlen         = sizeof(aio_nr),
1568                 .mode           = 0444,
1569                 .proc_handler   = proc_doulongvec_minmax,
1570         },
1571         {
1572                 .procname       = "aio-max-nr",
1573                 .data           = &aio_max_nr,
1574                 .maxlen         = sizeof(aio_max_nr),
1575                 .mode           = 0644,
1576                 .proc_handler   = proc_doulongvec_minmax,
1577         },
1578 #endif /* CONFIG_AIO */
1579 #ifdef CONFIG_INOTIFY_USER
1580         {
1581                 .procname       = "inotify",
1582                 .mode           = 0555,
1583                 .child          = inotify_table,
1584         },
1585 #endif
1586 #ifdef CONFIG_EPOLL
1587         {
1588                 .procname       = "epoll",
1589                 .mode           = 0555,
1590                 .child          = epoll_table,
1591         },
1592 #endif
1593 #endif
1594         {
1595                 .procname       = "protected_symlinks",
1596                 .data           = &sysctl_protected_symlinks,
1597                 .maxlen         = sizeof(int),
1598                 .mode           = 0600,
1599                 .proc_handler   = proc_dointvec_minmax,
1600                 .extra1         = &zero,
1601                 .extra2         = &one,
1602         },
1603         {
1604                 .procname       = "protected_hardlinks",
1605                 .data           = &sysctl_protected_hardlinks,
1606                 .maxlen         = sizeof(int),
1607                 .mode           = 0600,
1608                 .proc_handler   = proc_dointvec_minmax,
1609                 .extra1         = &zero,
1610                 .extra2         = &one,
1611         },
1612         {
1613                 .procname       = "suid_dumpable",
1614                 .data           = &suid_dumpable,
1615                 .maxlen         = sizeof(int),
1616                 .mode           = 0644,
1617                 .proc_handler   = proc_dointvec_minmax_coredump,
1618                 .extra1         = &zero,
1619                 .extra2         = &two,
1620         },
1621 #if defined(CONFIG_BINFMT_MISC) || defined(CONFIG_BINFMT_MISC_MODULE)
1622         {
1623                 .procname       = "binfmt_misc",
1624                 .mode           = 0555,
1625                 .child          = binfmt_misc_table,
1626         },
1627 #endif
1628         {
1629                 .procname       = "pipe-max-size",
1630                 .data           = &pipe_max_size,
1631                 .maxlen         = sizeof(int),
1632                 .mode           = 0644,
1633                 .proc_handler   = &pipe_proc_fn,
1634                 .extra1         = &pipe_min_size,
1635         },
1636         { }
1637 };
1638
1639 static struct ctl_table debug_table[] = {
1640 #ifdef CONFIG_SYSCTL_EXCEPTION_TRACE
1641         {
1642                 .procname       = "exception-trace",
1643                 .data           = &show_unhandled_signals,
1644                 .maxlen         = sizeof(int),
1645                 .mode           = 0644,
1646                 .proc_handler   = proc_dointvec
1647         },
1648 #endif
1649 #if defined(CONFIG_OPTPROBES)
1650         {
1651                 .procname       = "kprobes-optimization",
1652                 .data           = &sysctl_kprobes_optimization,
1653                 .maxlen         = sizeof(int),
1654                 .mode           = 0644,
1655                 .proc_handler   = proc_kprobes_optimization_handler,
1656                 .extra1         = &zero,
1657                 .extra2         = &one,
1658         },
1659 #endif
1660         { }
1661 };
1662
1663 static struct ctl_table dev_table[] = {
1664         { }
1665 };
1666
1667 int __init sysctl_init(void)
1668 {
1669         struct ctl_table_header *hdr;
1670
1671         hdr = register_sysctl_table(sysctl_base_table);
1672         kmemleak_not_leak(hdr);
1673         return 0;
1674 }
1675
1676 #endif /* CONFIG_SYSCTL */
1677
1678 /*
1679  * /proc/sys support
1680  */
1681
1682 #ifdef CONFIG_PROC_SYSCTL
1683
1684 static int _proc_do_string(void* data, int maxlen, int write,
1685                            void __user *buffer,
1686                            size_t *lenp, loff_t *ppos)
1687 {
1688         size_t len;
1689         char __user *p;
1690         char c;
1691
1692         if (!data || !maxlen || !*lenp) {
1693                 *lenp = 0;
1694                 return 0;
1695         }
1696
1697         if (write) {
1698                 len = 0;
1699                 p = buffer;
1700                 while (len < *lenp) {
1701                         if (get_user(c, p++))
1702                                 return -EFAULT;
1703                         if (c == 0 || c == '\n')
1704                                 break;
1705                         len++;
1706                 }
1707                 if (len >= maxlen)
1708                         len = maxlen-1;
1709                 if(copy_from_user(data, buffer, len))
1710                         return -EFAULT;
1711                 ((char *) data)[len] = 0;
1712                 *ppos += *lenp;
1713         } else {
1714                 len = strlen(data);
1715                 if (len > maxlen)
1716                         len = maxlen;
1717
1718                 if (*ppos > len) {
1719                         *lenp = 0;
1720                         return 0;
1721                 }
1722
1723                 data += *ppos;
1724                 len  -= *ppos;
1725
1726                 if (len > *lenp)
1727                         len = *lenp;
1728                 if (len)
1729                         if(copy_to_user(buffer, data, len))
1730                                 return -EFAULT;
1731                 if (len < *lenp) {
1732                         if(put_user('\n', ((char __user *) buffer) + len))
1733                                 return -EFAULT;
1734                         len++;
1735                 }
1736                 *lenp = len;
1737                 *ppos += len;
1738         }
1739         return 0;
1740 }
1741
1742 /**
1743  * proc_dostring - read a string sysctl
1744  * @table: the sysctl table
1745  * @write: %TRUE if this is a write to the sysctl file
1746  * @buffer: the user buffer
1747  * @lenp: the size of the user buffer
1748  * @ppos: file position
1749  *
1750  * Reads/writes a string from/to the user buffer. If the kernel
1751  * buffer provided is not large enough to hold the string, the
1752  * string is truncated. The copied string is %NULL-terminated.
1753  * If the string is being read by the user process, it is copied
1754  * and a newline '\n' is added. It is truncated if the buffer is
1755  * not large enough.
1756  *
1757  * Returns 0 on success.
1758  */
1759 int proc_dostring(struct ctl_table *table, int write,
1760                   void __user *buffer, size_t *lenp, loff_t *ppos)
1761 {
1762         return _proc_do_string(table->data, table->maxlen, write,
1763                                buffer, lenp, ppos);
1764 }
1765
1766 static size_t proc_skip_spaces(char **buf)
1767 {
1768         size_t ret;
1769         char *tmp = skip_spaces(*buf);
1770         ret = tmp - *buf;
1771         *buf = tmp;
1772         return ret;
1773 }
1774
1775 static void proc_skip_char(char **buf, size_t *size, const char v)
1776 {
1777         while (*size) {
1778                 if (**buf != v)
1779                         break;
1780                 (*size)--;
1781                 (*buf)++;
1782         }
1783 }
1784
1785 #define TMPBUFLEN 22
1786 /**
1787  * proc_get_long - reads an ASCII formatted integer from a user buffer
1788  *
1789  * @buf: a kernel buffer
1790  * @size: size of the kernel buffer
1791  * @val: this is where the number will be stored
1792  * @neg: set to %TRUE if number is negative
1793  * @perm_tr: a vector which contains the allowed trailers
1794  * @perm_tr_len: size of the perm_tr vector
1795  * @tr: pointer to store the trailer character
1796  *
1797  * In case of success %0 is returned and @buf and @size are updated with
1798  * the amount of bytes read. If @tr is non-NULL and a trailing
1799  * character exists (size is non-zero after returning from this
1800  * function), @tr is updated with the trailing character.
1801  */
1802 static int proc_get_long(char **buf, size_t *size,
1803                           unsigned long *val, bool *neg,
1804                           const char *perm_tr, unsigned perm_tr_len, char *tr)
1805 {
1806         int len;
1807         char *p, tmp[TMPBUFLEN];
1808
1809         if (!*size)
1810                 return -EINVAL;
1811
1812         len = *size;
1813         if (len > TMPBUFLEN - 1)
1814                 len = TMPBUFLEN - 1;
1815
1816         memcpy(tmp, *buf, len);
1817
1818         tmp[len] = 0;
1819         p = tmp;
1820         if (*p == '-' && *size > 1) {
1821                 *neg = true;
1822                 p++;
1823         } else
1824                 *neg = false;
1825         if (!isdigit(*p))
1826                 return -EINVAL;
1827
1828         *val = simple_strtoul(p, &p, 0);
1829
1830         len = p - tmp;
1831
1832         /* We don't know if the next char is whitespace thus we may accept
1833          * invalid integers (e.g. 1234...a) or two integers instead of one
1834          * (e.g. 123...1). So lets not allow such large numbers. */
1835         if (len == TMPBUFLEN - 1)
1836                 return -EINVAL;
1837
1838         if (len < *size && perm_tr_len && !memchr(perm_tr, *p, perm_tr_len))
1839                 return -EINVAL;
1840
1841         if (tr && (len < *size))
1842                 *tr = *p;
1843
1844         *buf += len;
1845         *size -= len;
1846
1847         return 0;
1848 }
1849
1850 /**
1851  * proc_put_long - converts an integer to a decimal ASCII formatted string
1852  *
1853  * @buf: the user buffer
1854  * @size: the size of the user buffer
1855  * @val: the integer to be converted
1856  * @neg: sign of the number, %TRUE for negative
1857  *
1858  * In case of success %0 is returned and @buf and @size are updated with
1859  * the amount of bytes written.
1860  */
1861 static int proc_put_long(void __user **buf, size_t *size, unsigned long val,
1862                           bool neg)
1863 {
1864         int len;
1865         char tmp[TMPBUFLEN], *p = tmp;
1866
1867         sprintf(p, "%s%lu", neg ? "-" : "", val);
1868         len = strlen(tmp);
1869         if (len > *size)
1870                 len = *size;
1871         if (copy_to_user(*buf, tmp, len))
1872                 return -EFAULT;
1873         *size -= len;
1874         *buf += len;
1875         return 0;
1876 }
1877 #undef TMPBUFLEN
1878
1879 static int proc_put_char(void __user **buf, size_t *size, char c)
1880 {
1881         if (*size) {
1882                 char __user **buffer = (char __user **)buf;
1883                 if (put_user(c, *buffer))
1884                         return -EFAULT;
1885                 (*size)--, (*buffer)++;
1886                 *buf = *buffer;
1887         }
1888         return 0;
1889 }
1890
1891 static int do_proc_dointvec_conv(bool *negp, unsigned long *lvalp,
1892                                  int *valp,
1893                                  int write, void *data)
1894 {
1895         if (write) {
1896                 *valp = *negp ? -*lvalp : *lvalp;
1897         } else {
1898                 int val = *valp;
1899                 if (val < 0) {
1900                         *negp = true;
1901                         *lvalp = (unsigned long)-val;
1902                 } else {
1903                         *negp = false;
1904                         *lvalp = (unsigned long)val;
1905                 }
1906         }
1907         return 0;
1908 }
1909
1910 static const char proc_wspace_sep[] = { ' ', '\t', '\n' };
1911
1912 static int __do_proc_dointvec(void *tbl_data, struct ctl_table *table,
1913                   int write, void __user *buffer,
1914                   size_t *lenp, loff_t *ppos,
1915                   int (*conv)(bool *negp, unsigned long *lvalp, int *valp,
1916                               int write, void *data),
1917                   void *data)
1918 {
1919         int *i, vleft, first = 1, err = 0;
1920         unsigned long page = 0;
1921         size_t left;
1922         char *kbuf;
1923
1924         if (!tbl_data || !table->maxlen || !*lenp || (*ppos && !write)) {
1925                 *lenp = 0;
1926                 return 0;
1927         }
1928
1929         i = (int *) tbl_data;
1930         vleft = table->maxlen / sizeof(*i);
1931         left = *lenp;
1932
1933         if (!conv)
1934                 conv = do_proc_dointvec_conv;
1935
1936         if (write) {
1937                 if (left > PAGE_SIZE - 1)
1938                         left = PAGE_SIZE - 1;
1939                 page = __get_free_page(GFP_TEMPORARY);
1940                 kbuf = (char *) page;
1941                 if (!kbuf)
1942                         return -ENOMEM;
1943                 if (copy_from_user(kbuf, buffer, left)) {
1944                         err = -EFAULT;
1945                         goto free;
1946                 }
1947                 kbuf[left] = 0;
1948         }
1949
1950         for (; left && vleft--; i++, first=0) {
1951                 unsigned long lval;
1952                 bool neg;
1953
1954                 if (write) {
1955                         left -= proc_skip_spaces(&kbuf);
1956
1957                         if (!left)
1958                                 break;
1959                         err = proc_get_long(&kbuf, &left, &lval, &neg,
1960                                              proc_wspace_sep,
1961                                              sizeof(proc_wspace_sep), NULL);
1962                         if (err)
1963                                 break;
1964                         if (conv(&neg, &lval, i, 1, data)) {
1965                                 err = -EINVAL;
1966                                 break;
1967                         }
1968                 } else {
1969                         if (conv(&neg, &lval, i, 0, data)) {
1970                                 err = -EINVAL;
1971                                 break;
1972                         }
1973                         if (!first)
1974                                 err = proc_put_char(&buffer, &left, '\t');
1975                         if (err)
1976                                 break;
1977                         err = proc_put_long(&buffer, &left, lval, neg);
1978                         if (err)
1979                                 break;
1980                 }
1981         }
1982
1983         if (!write && !first && left && !err)
1984                 err = proc_put_char(&buffer, &left, '\n');
1985         if (write && !err && left)
1986                 left -= proc_skip_spaces(&kbuf);
1987 free:
1988         if (write) {
1989                 free_page(page);
1990                 if (first)
1991                         return err ? : -EINVAL;
1992         }
1993         *lenp -= left;
1994         *ppos += *lenp;
1995         return err;
1996 }
1997
1998 static int do_proc_dointvec(struct ctl_table *table, int write,
1999                   void __user *buffer, size_t *lenp, loff_t *ppos,
2000                   int (*conv)(bool *negp, unsigned long *lvalp, int *valp,
2001                               int write, void *data),
2002                   void *data)
2003 {
2004         return __do_proc_dointvec(table->data, table, write,
2005                         buffer, lenp, ppos, conv, data);
2006 }
2007
2008 /**
2009  * proc_dointvec - read a vector of integers
2010  * @table: the sysctl table
2011  * @write: %TRUE if this is a write to the sysctl file
2012  * @buffer: the user buffer
2013  * @lenp: the size of the user buffer
2014  * @ppos: file position
2015  *
2016  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2017  * values from/to the user buffer, treated as an ASCII string.
2018  *
2019  * Returns 0 on success.
2020  */
2021 int proc_dointvec(struct ctl_table *table, int write,
2022                      void __user *buffer, size_t *lenp, loff_t *ppos)
2023 {
2024     return do_proc_dointvec(table,write,buffer,lenp,ppos,
2025                             NULL,NULL);
2026 }
2027
2028 /*
2029  * Taint values can only be increased
2030  * This means we can safely use a temporary.
2031  */
2032 static int proc_taint(struct ctl_table *table, int write,
2033                                void __user *buffer, size_t *lenp, loff_t *ppos)
2034 {
2035         struct ctl_table t;
2036         unsigned long tmptaint = get_taint();
2037         int err;
2038
2039         if (write && !capable(CAP_SYS_ADMIN))
2040                 return -EPERM;
2041
2042         t = *table;
2043         t.data = &tmptaint;
2044         err = proc_doulongvec_minmax(&t, write, buffer, lenp, ppos);
2045         if (err < 0)
2046                 return err;
2047
2048         if (write) {
2049                 /*
2050                  * Poor man's atomic or. Not worth adding a primitive
2051                  * to everyone's atomic.h for this
2052                  */
2053                 int i;
2054                 for (i = 0; i < BITS_PER_LONG && tmptaint >> i; i++) {
2055                         if ((tmptaint >> i) & 1)
2056                                 add_taint(i, LOCKDEP_STILL_OK);
2057                 }
2058         }
2059
2060         return err;
2061 }
2062
2063 #ifdef CONFIG_PRINTK
2064 static int proc_dointvec_minmax_sysadmin(struct ctl_table *table, int write,
2065                                 void __user *buffer, size_t *lenp, loff_t *ppos)
2066 {
2067         if (write && !capable(CAP_SYS_ADMIN))
2068                 return -EPERM;
2069
2070         return proc_dointvec_minmax(table, write, buffer, lenp, ppos);
2071 }
2072 #endif
2073
2074 struct do_proc_dointvec_minmax_conv_param {
2075         int *min;
2076         int *max;
2077 };
2078
2079 static int do_proc_dointvec_minmax_conv(bool *negp, unsigned long *lvalp,
2080                                         int *valp,
2081                                         int write, void *data)
2082 {
2083         struct do_proc_dointvec_minmax_conv_param *param = data;
2084         if (write) {
2085                 int val = *negp ? -*lvalp : *lvalp;
2086                 if ((param->min && *param->min > val) ||
2087                     (param->max && *param->max < val))
2088                         return -EINVAL;
2089                 *valp = val;
2090         } else {
2091                 int val = *valp;
2092                 if (val < 0) {
2093                         *negp = true;
2094                         *lvalp = (unsigned long)-val;
2095                 } else {
2096                         *negp = false;
2097                         *lvalp = (unsigned long)val;
2098                 }
2099         }
2100         return 0;
2101 }
2102
2103 /**
2104  * proc_dointvec_minmax - read a vector of integers with min/max values
2105  * @table: the sysctl table
2106  * @write: %TRUE if this is a write to the sysctl file
2107  * @buffer: the user buffer
2108  * @lenp: the size of the user buffer
2109  * @ppos: file position
2110  *
2111  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2112  * values from/to the user buffer, treated as an ASCII string.
2113  *
2114  * This routine will ensure the values are within the range specified by
2115  * table->extra1 (min) and table->extra2 (max).
2116  *
2117  * Returns 0 on success.
2118  */
2119 int proc_dointvec_minmax(struct ctl_table *table, int write,
2120                   void __user *buffer, size_t *lenp, loff_t *ppos)
2121 {
2122         struct do_proc_dointvec_minmax_conv_param param = {
2123                 .min = (int *) table->extra1,
2124                 .max = (int *) table->extra2,
2125         };
2126         return do_proc_dointvec(table, write, buffer, lenp, ppos,
2127                                 do_proc_dointvec_minmax_conv, &param);
2128 }
2129
2130 static void validate_coredump_safety(void)
2131 {
2132 #ifdef CONFIG_COREDUMP
2133         if (suid_dumpable == SUID_DUMP_ROOT &&
2134             core_pattern[0] != '/' && core_pattern[0] != '|') {
2135                 printk(KERN_WARNING "Unsafe core_pattern used with "\
2136                         "suid_dumpable=2. Pipe handler or fully qualified "\
2137                         "core dump path required.\n");
2138         }
2139 #endif
2140 }
2141
2142 static int proc_dointvec_minmax_coredump(struct ctl_table *table, int write,
2143                 void __user *buffer, size_t *lenp, loff_t *ppos)
2144 {
2145         int error = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
2146         if (!error)
2147                 validate_coredump_safety();
2148         return error;
2149 }
2150
2151 #ifdef CONFIG_COREDUMP
2152 static int proc_dostring_coredump(struct ctl_table *table, int write,
2153                   void __user *buffer, size_t *lenp, loff_t *ppos)
2154 {
2155         int error = proc_dostring(table, write, buffer, lenp, ppos);
2156         if (!error)
2157                 validate_coredump_safety();
2158         return error;
2159 }
2160 #endif
2161
2162 static int __do_proc_doulongvec_minmax(void *data, struct ctl_table *table, int write,
2163                                      void __user *buffer,
2164                                      size_t *lenp, loff_t *ppos,
2165                                      unsigned long convmul,
2166                                      unsigned long convdiv)
2167 {
2168         unsigned long *i, *min, *max;
2169         int vleft, first = 1, err = 0;
2170         unsigned long page = 0;
2171         size_t left;
2172         char *kbuf;
2173
2174         if (!data || !table->maxlen || !*lenp || (*ppos && !write)) {
2175                 *lenp = 0;
2176                 return 0;
2177         }
2178
2179         i = (unsigned long *) data;
2180         min = (unsigned long *) table->extra1;
2181         max = (unsigned long *) table->extra2;
2182         vleft = table->maxlen / sizeof(unsigned long);
2183         left = *lenp;
2184
2185         if (write) {
2186                 if (left > PAGE_SIZE - 1)
2187                         left = PAGE_SIZE - 1;
2188                 page = __get_free_page(GFP_TEMPORARY);
2189                 kbuf = (char *) page;
2190                 if (!kbuf)
2191                         return -ENOMEM;
2192                 if (copy_from_user(kbuf, buffer, left)) {
2193                         err = -EFAULT;
2194                         goto free;
2195                 }
2196                 kbuf[left] = 0;
2197         }
2198
2199         for (; left && vleft--; i++, first = 0) {
2200                 unsigned long val;
2201
2202                 if (write) {
2203                         bool neg;
2204
2205                         left -= proc_skip_spaces(&kbuf);
2206
2207                         err = proc_get_long(&kbuf, &left, &val, &neg,
2208                                              proc_wspace_sep,
2209                                              sizeof(proc_wspace_sep), NULL);
2210                         if (err)
2211                                 break;
2212                         if (neg)
2213                                 continue;
2214                         if ((min && val < *min) || (max && val > *max))
2215                                 continue;
2216                         *i = val;
2217                 } else {
2218                         val = convdiv * (*i) / convmul;
2219                         if (!first) {
2220                                 err = proc_put_char(&buffer, &left, '\t');
2221                                 if (err)
2222                                         break;
2223                         }
2224                         err = proc_put_long(&buffer, &left, val, false);
2225                         if (err)
2226                                 break;
2227                 }
2228         }
2229
2230         if (!write && !first && left && !err)
2231                 err = proc_put_char(&buffer, &left, '\n');
2232         if (write && !err)
2233                 left -= proc_skip_spaces(&kbuf);
2234 free:
2235         if (write) {
2236                 free_page(page);
2237                 if (first)
2238                         return err ? : -EINVAL;
2239         }
2240         *lenp -= left;
2241         *ppos += *lenp;
2242         return err;
2243 }
2244
2245 static int do_proc_doulongvec_minmax(struct ctl_table *table, int write,
2246                                      void __user *buffer,
2247                                      size_t *lenp, loff_t *ppos,
2248                                      unsigned long convmul,
2249                                      unsigned long convdiv)
2250 {
2251         return __do_proc_doulongvec_minmax(table->data, table, write,
2252                         buffer, lenp, ppos, convmul, convdiv);
2253 }
2254
2255 /**
2256  * proc_doulongvec_minmax - read a vector of long integers with min/max values
2257  * @table: the sysctl table
2258  * @write: %TRUE if this is a write to the sysctl file
2259  * @buffer: the user buffer
2260  * @lenp: the size of the user buffer
2261  * @ppos: file position
2262  *
2263  * Reads/writes up to table->maxlen/sizeof(unsigned long) unsigned long
2264  * values from/to the user buffer, treated as an ASCII string.
2265  *
2266  * This routine will ensure the values are within the range specified by
2267  * table->extra1 (min) and table->extra2 (max).
2268  *
2269  * Returns 0 on success.
2270  */
2271 int proc_doulongvec_minmax(struct ctl_table *table, int write,
2272                            void __user *buffer, size_t *lenp, loff_t *ppos)
2273 {
2274     return do_proc_doulongvec_minmax(table, write, buffer, lenp, ppos, 1l, 1l);
2275 }
2276
2277 /**
2278  * proc_doulongvec_ms_jiffies_minmax - read a vector of millisecond values with min/max values
2279  * @table: the sysctl table
2280  * @write: %TRUE if this is a write to the sysctl file
2281  * @buffer: the user buffer
2282  * @lenp: the size of the user buffer
2283  * @ppos: file position
2284  *
2285  * Reads/writes up to table->maxlen/sizeof(unsigned long) unsigned long
2286  * values from/to the user buffer, treated as an ASCII string. The values
2287  * are treated as milliseconds, and converted to jiffies when they are stored.
2288  *
2289  * This routine will ensure the values are within the range specified by
2290  * table->extra1 (min) and table->extra2 (max).
2291  *
2292  * Returns 0 on success.
2293  */
2294 int proc_doulongvec_ms_jiffies_minmax(struct ctl_table *table, int write,
2295                                       void __user *buffer,
2296                                       size_t *lenp, loff_t *ppos)
2297 {
2298     return do_proc_doulongvec_minmax(table, write, buffer,
2299                                      lenp, ppos, HZ, 1000l);
2300 }
2301
2302
2303 static int do_proc_dointvec_jiffies_conv(bool *negp, unsigned long *lvalp,
2304                                          int *valp,
2305                                          int write, void *data)
2306 {
2307         if (write) {
2308                 if (*lvalp > LONG_MAX / HZ)
2309                         return 1;
2310                 *valp = *negp ? -(*lvalp*HZ) : (*lvalp*HZ);
2311         } else {
2312                 int val = *valp;
2313                 unsigned long lval;
2314                 if (val < 0) {
2315                         *negp = true;
2316                         lval = (unsigned long)-val;
2317                 } else {
2318                         *negp = false;
2319                         lval = (unsigned long)val;
2320                 }
2321                 *lvalp = lval / HZ;
2322         }
2323         return 0;
2324 }
2325
2326 static int do_proc_dointvec_userhz_jiffies_conv(bool *negp, unsigned long *lvalp,
2327                                                 int *valp,
2328                                                 int write, void *data)
2329 {
2330         if (write) {
2331                 if (USER_HZ < HZ && *lvalp > (LONG_MAX / HZ) * USER_HZ)
2332                         return 1;
2333                 *valp = clock_t_to_jiffies(*negp ? -*lvalp : *lvalp);
2334         } else {
2335                 int val = *valp;
2336                 unsigned long lval;
2337                 if (val < 0) {
2338                         *negp = true;
2339                         lval = (unsigned long)-val;
2340                 } else {
2341                         *negp = false;
2342                         lval = (unsigned long)val;
2343                 }
2344                 *lvalp = jiffies_to_clock_t(lval);
2345         }
2346         return 0;
2347 }
2348
2349 static int do_proc_dointvec_ms_jiffies_conv(bool *negp, unsigned long *lvalp,
2350                                             int *valp,
2351                                             int write, void *data)
2352 {
2353         if (write) {
2354                 unsigned long jif = msecs_to_jiffies(*negp ? -*lvalp : *lvalp);
2355
2356                 if (jif > INT_MAX)
2357                         return 1;
2358                 *valp = (int)jif;
2359         } else {
2360                 int val = *valp;
2361                 unsigned long lval;
2362                 if (val < 0) {
2363                         *negp = true;
2364                         lval = (unsigned long)-val;
2365                 } else {
2366                         *negp = false;
2367                         lval = (unsigned long)val;
2368                 }
2369                 *lvalp = jiffies_to_msecs(lval);
2370         }
2371         return 0;
2372 }
2373
2374 /**
2375  * proc_dointvec_jiffies - read a vector of integers as seconds
2376  * @table: the sysctl table
2377  * @write: %TRUE if this is a write to the sysctl file
2378  * @buffer: the user buffer
2379  * @lenp: the size of the user buffer
2380  * @ppos: file position
2381  *
2382  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2383  * values from/to the user buffer, treated as an ASCII string.
2384  * The values read are assumed to be in seconds, and are converted into
2385  * jiffies.
2386  *
2387  * Returns 0 on success.
2388  */
2389 int proc_dointvec_jiffies(struct ctl_table *table, int write,
2390                           void __user *buffer, size_t *lenp, loff_t *ppos)
2391 {
2392     return do_proc_dointvec(table,write,buffer,lenp,ppos,
2393                             do_proc_dointvec_jiffies_conv,NULL);
2394 }
2395
2396 /**
2397  * proc_dointvec_userhz_jiffies - read a vector of integers as 1/USER_HZ seconds
2398  * @table: the sysctl table
2399  * @write: %TRUE if this is a write to the sysctl file
2400  * @buffer: the user buffer
2401  * @lenp: the size of the user buffer
2402  * @ppos: pointer to the file position
2403  *
2404  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2405  * values from/to the user buffer, treated as an ASCII string.
2406  * The values read are assumed to be in 1/USER_HZ seconds, and
2407  * are converted into jiffies.
2408  *
2409  * Returns 0 on success.
2410  */
2411 int proc_dointvec_userhz_jiffies(struct ctl_table *table, int write,
2412                                  void __user *buffer, size_t *lenp, loff_t *ppos)
2413 {
2414     return do_proc_dointvec(table,write,buffer,lenp,ppos,
2415                             do_proc_dointvec_userhz_jiffies_conv,NULL);
2416 }
2417
2418 /**
2419  * proc_dointvec_ms_jiffies - read a vector of integers as 1 milliseconds
2420  * @table: the sysctl table
2421  * @write: %TRUE if this is a write to the sysctl file
2422  * @buffer: the user buffer
2423  * @lenp: the size of the user buffer
2424  * @ppos: file position
2425  * @ppos: the current position in the file
2426  *
2427  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2428  * values from/to the user buffer, treated as an ASCII string.
2429  * The values read are assumed to be in 1/1000 seconds, and
2430  * are converted into jiffies.
2431  *
2432  * Returns 0 on success.
2433  */
2434 int proc_dointvec_ms_jiffies(struct ctl_table *table, int write,
2435                              void __user *buffer, size_t *lenp, loff_t *ppos)
2436 {
2437         return do_proc_dointvec(table, write, buffer, lenp, ppos,
2438                                 do_proc_dointvec_ms_jiffies_conv, NULL);
2439 }
2440
2441 static int proc_do_cad_pid(struct ctl_table *table, int write,
2442                            void __user *buffer, size_t *lenp, loff_t *ppos)
2443 {
2444         struct pid *new_pid;
2445         pid_t tmp;
2446         int r;
2447
2448         tmp = pid_vnr(cad_pid);
2449
2450         r = __do_proc_dointvec(&tmp, table, write, buffer,
2451                                lenp, ppos, NULL, NULL);
2452         if (r || !write)
2453                 return r;
2454
2455         new_pid = find_get_pid(tmp);
2456         if (!new_pid)
2457                 return -ESRCH;
2458
2459         put_pid(xchg(&cad_pid, new_pid));
2460         return 0;
2461 }
2462
2463 /**
2464  * proc_do_large_bitmap - read/write from/to a large bitmap
2465  * @table: the sysctl table
2466  * @write: %TRUE if this is a write to the sysctl file
2467  * @buffer: the user buffer
2468  * @lenp: the size of the user buffer
2469  * @ppos: file position
2470  *
2471  * The bitmap is stored at table->data and the bitmap length (in bits)
2472  * in table->maxlen.
2473  *
2474  * We use a range comma separated format (e.g. 1,3-4,10-10) so that
2475  * large bitmaps may be represented in a compact manner. Writing into
2476  * the file will clear the bitmap then update it with the given input.
2477  *
2478  * Returns 0 on success.
2479  */
2480 int proc_do_large_bitmap(struct ctl_table *table, int write,
2481                          void __user *buffer, size_t *lenp, loff_t *ppos)
2482 {
2483         int err = 0;
2484         bool first = 1;
2485         size_t left = *lenp;
2486         unsigned long bitmap_len = table->maxlen;
2487         unsigned long *bitmap = (unsigned long *) table->data;
2488         unsigned long *tmp_bitmap = NULL;
2489         char tr_a[] = { '-', ',', '\n' }, tr_b[] = { ',', '\n', 0 }, c;
2490
2491         if (!bitmap_len || !left || (*ppos && !write)) {
2492                 *lenp = 0;
2493                 return 0;
2494         }
2495
2496         if (write) {
2497                 unsigned long page = 0;
2498                 char *kbuf;
2499
2500                 if (left > PAGE_SIZE - 1)
2501                         left = PAGE_SIZE - 1;
2502
2503                 page = __get_free_page(GFP_TEMPORARY);
2504                 kbuf = (char *) page;
2505                 if (!kbuf)
2506                         return -ENOMEM;
2507                 if (copy_from_user(kbuf, buffer, left)) {
2508                         free_page(page);
2509                         return -EFAULT;
2510                 }
2511                 kbuf[left] = 0;
2512
2513                 tmp_bitmap = kzalloc(BITS_TO_LONGS(bitmap_len) * sizeof(unsigned long),
2514                                      GFP_KERNEL);
2515                 if (!tmp_bitmap) {
2516                         free_page(page);
2517                         return -ENOMEM;
2518                 }
2519                 proc_skip_char(&kbuf, &left, '\n');
2520                 while (!err && left) {
2521                         unsigned long val_a, val_b;
2522                         bool neg;
2523
2524                         err = proc_get_long(&kbuf, &left, &val_a, &neg, tr_a,
2525                                              sizeof(tr_a), &c);
2526                         if (err)
2527                                 break;
2528                         if (val_a >= bitmap_len || neg) {
2529                                 err = -EINVAL;
2530                                 break;
2531                         }
2532
2533                         val_b = val_a;
2534                         if (left) {
2535                                 kbuf++;
2536                                 left--;
2537                         }
2538
2539                         if (c == '-') {
2540                                 err = proc_get_long(&kbuf, &left, &val_b,
2541                                                      &neg, tr_b, sizeof(tr_b),
2542                                                      &c);
2543                                 if (err)
2544                                         break;
2545                                 if (val_b >= bitmap_len || neg ||
2546                                     val_a > val_b) {
2547                                         err = -EINVAL;
2548                                         break;
2549                                 }
2550                                 if (left) {
2551                                         kbuf++;
2552                                         left--;
2553                                 }
2554                         }
2555
2556                         bitmap_set(tmp_bitmap, val_a, val_b - val_a + 1);
2557                         first = 0;
2558                         proc_skip_char(&kbuf, &left, '\n');
2559                 }
2560                 free_page(page);
2561         } else {
2562                 unsigned long bit_a, bit_b = 0;
2563
2564                 while (left) {
2565                         bit_a = find_next_bit(bitmap, bitmap_len, bit_b);
2566                         if (bit_a >= bitmap_len)
2567                                 break;
2568                         bit_b = find_next_zero_bit(bitmap, bitmap_len,
2569                                                    bit_a + 1) - 1;
2570
2571                         if (!first) {
2572                                 err = proc_put_char(&buffer, &left, ',');
2573                                 if (err)
2574                                         break;
2575                         }
2576                         err = proc_put_long(&buffer, &left, bit_a, false);
2577                         if (err)
2578                                 break;
2579                         if (bit_a != bit_b) {
2580                                 err = proc_put_char(&buffer, &left, '-');
2581                                 if (err)
2582                                         break;
2583                                 err = proc_put_long(&buffer, &left, bit_b, false);
2584                                 if (err)
2585                                         break;
2586                         }
2587
2588                         first = 0; bit_b++;
2589                 }
2590                 if (!err)
2591                         err = proc_put_char(&buffer, &left, '\n');
2592         }
2593
2594         if (!err) {
2595                 if (write) {
2596                         if (*ppos)
2597                                 bitmap_or(bitmap, bitmap, tmp_bitmap, bitmap_len);
2598                         else
2599                                 bitmap_copy(bitmap, tmp_bitmap, bitmap_len);
2600                 }
2601                 kfree(tmp_bitmap);
2602                 *lenp -= left;
2603                 *ppos += *lenp;
2604                 return 0;
2605         } else {
2606                 kfree(tmp_bitmap);
2607                 return err;
2608         }
2609 }
2610
2611 #else /* CONFIG_PROC_SYSCTL */
2612
2613 int proc_dostring(struct ctl_table *table, int write,
2614                   void __user *buffer, size_t *lenp, loff_t *ppos)
2615 {
2616         return -ENOSYS;
2617 }
2618
2619 int proc_dointvec(struct ctl_table *table, int write,
2620                   void __user *buffer, size_t *lenp, loff_t *ppos)
2621 {
2622         return -ENOSYS;
2623 }
2624
2625 int proc_dointvec_minmax(struct ctl_table *table, int write,
2626                     void __user *buffer, size_t *lenp, loff_t *ppos)
2627 {
2628         return -ENOSYS;
2629 }
2630
2631 int proc_dointvec_jiffies(struct ctl_table *table, int write,
2632                     void __user *buffer, size_t *lenp, loff_t *ppos)
2633 {
2634         return -ENOSYS;
2635 }
2636
2637 int proc_dointvec_userhz_jiffies(struct ctl_table *table, int write,
2638                     void __user *buffer, size_t *lenp, loff_t *ppos)
2639 {
2640         return -ENOSYS;
2641 }
2642
2643 int proc_dointvec_ms_jiffies(struct ctl_table *table, int write,
2644                              void __user *buffer, size_t *lenp, loff_t *ppos)
2645 {
2646         return -ENOSYS;
2647 }
2648
2649 int proc_doulongvec_minmax(struct ctl_table *table, int write,
2650                     void __user *buffer, size_t *lenp, loff_t *ppos)
2651 {
2652         return -ENOSYS;
2653 }
2654
2655 int proc_doulongvec_ms_jiffies_minmax(struct ctl_table *table, int write,
2656                                       void __user *buffer,
2657                                       size_t *lenp, loff_t *ppos)
2658 {
2659     return -ENOSYS;
2660 }
2661
2662
2663 #endif /* CONFIG_PROC_SYSCTL */
2664
2665 /*
2666  * No sense putting this after each symbol definition, twice,
2667  * exception granted :-)
2668  */
2669 EXPORT_SYMBOL(proc_dointvec);
2670 EXPORT_SYMBOL(proc_dointvec_jiffies);
2671 EXPORT_SYMBOL(proc_dointvec_minmax);
2672 EXPORT_SYMBOL(proc_dointvec_userhz_jiffies);
2673 EXPORT_SYMBOL(proc_dointvec_ms_jiffies);
2674 EXPORT_SYMBOL(proc_dostring);
2675 EXPORT_SYMBOL(proc_doulongvec_minmax);
2676 EXPORT_SYMBOL(proc_doulongvec_ms_jiffies_minmax);