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