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/aio.h>
  23 #include <linux/mm.h>
  24 #include <linux/swap.h>
  25 #include <linux/slab.h>
  26 #include <linux/sysctl.h>
  27 #include <linux/bitmap.h>
  28 #include <linux/signal.h>
  29 #include <linux/printk.h>
  30 #include <linux/proc_fs.h>
  31 #include <linux/security.h>
  32 #include <linux/ctype.h>
  33 #include <linux/kmemcheck.h>
  34 #include <linux/kmemleak.h>
  35 #include <linux/fs.h>
  36 #include <linux/init.h>
  37 #include <linux/kernel.h>
  38 #include <linux/kobject.h>
  39 #include <linux/net.h>
  40 #include <linux/sysrq.h>
  41 #include <linux/highuid.h>
  42 #include <linux/writeback.h>
  43 #include <linux/ratelimit.h>
  44 #include <linux/compaction.h>
  45 #include <linux/hugetlb.h>
  46 #include <linux/initrd.h>
  47 #include <linux/key.h>
  48 #include <linux/times.h>
  49 #include <linux/limits.h>
  50 #include <linux/dcache.h>
  51 #include <linux/dnotify.h>
  52 #include <linux/syscalls.h>
  53 #include <linux/vmstat.h>
  54 #include <linux/nfs_fs.h>
  55 #include <linux/acpi.h>
  56 #include <linux/reboot.h>
  57 #include <linux/ftrace.h>
  58 #include <linux/perf_event.h>
  59 #include <linux/kprobes.h>
  60 #include <linux/pipe_fs_i.h>
  61 #include <linux/oom.h>
  62 #include <linux/kmod.h>
  63 #include <linux/capability.h>
  64 #include <linux/binfmts.h>
  65 #include <linux/sched/sysctl.h>
  66 #include <linux/kexec.h>
  67 #include <linux/bpf.h>
  68 #include <linux/mount.h>
  69
  70 #include <asm/uaccess.h>
  71 #include <asm/processor.h>
  72
  73 #ifdef CONFIG_X86
  74 #include <asm/nmi.h>
  75 #include <asm/stacktrace.h>
  76 #include <asm/io.h>
  77 #endif
  78 #ifdef CONFIG_SPARC
  79 #include <asm/setup.h>
  80 #endif
  81 #ifdef CONFIG_BSD_PROCESS_ACCT
  82 #include <linux/acct.h>
  83 #endif
  84 #ifdef CONFIG_RT_MUTEXES
  85 #include <linux/rtmutex.h>
  86 #endif
  87 #if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_LOCK_STAT)
  88 #include <linux/lockdep.h>
  89 #endif
  90 #ifdef CONFIG_CHR_DEV_SG
  91 #include <scsi/sg.h>
  92 #endif
  93
  94 #ifdef CONFIG_LOCKUP_DETECTOR
  95 #include <linux/nmi.h>
  96 #endif
  97
  98 #if defined(CONFIG_SYSCTL)
  99
 100 /* External variables not in a header file. */
 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
 117 /* Constants used for minimum and  maximum */
 118 #ifdef CONFIG_LOCKUP_DETECTOR
 119 static int sixty = 60;
 120 #endif
 121
 122 static int __maybe_unused neg_one = -1;
 123
 124 static int zero;
 125 static int __maybe_unused one = 1;
 126 static int __maybe_unused two = 2;
 127 static int __maybe_unused four = 4;
 128 static unsigned long one_ul = 1;
 129 static int one_hundred = 100;
 130 #ifdef CONFIG_PRINTK
 131 static int ten_thousand = 10000;
 132 #endif
 133
 134 /* this is needed for the proc_doulongvec_minmax of vm_dirty_bytes */
 135 static unsigned long dirty_bytes_min = 2 * PAGE_SIZE;
 136
 137 /* this is needed for the proc_dointvec_minmax for [fs_]overflow UID and GID */
 138 static int maxolduid = 65535;
 139 static int minolduid;
 140
 141 static int ngroups_max = NGROUPS_MAX;
 142 static const int cap_last_cap = CAP_LAST_CAP;
 143
 144 /*this is needed for proc_doulongvec_minmax of sysctl_hung_task_timeout_secs */
 145 #ifdef CONFIG_DETECT_HUNG_TASK
 146 static unsigned long hung_task_timeout_max = (LONG_MAX/HZ);
 147 #endif
 148
 149 #ifdef CONFIG_INOTIFY_USER
 150 #include <linux/inotify.h>
 151 #endif
 152 #ifdef CONFIG_SPARC
 153 #endif
 154
 155 #ifdef __hppa__
 156 extern int pwrsw_enabled;
 157 #endif
 158
 159 #ifdef CONFIG_SYSCTL_ARCH_UNALIGN_ALLOW
 160 extern int unaligned_enabled;
 161 #endif
 162
 163 #ifdef CONFIG_IA64
 164 extern int unaligned_dump_stack;
 165 #endif
 166
 167 #ifdef CONFIG_SYSCTL_ARCH_UNALIGN_NO_WARN
 168 extern int no_unaligned_warning;
 169 #endif
 170
 171 #ifdef CONFIG_PROC_SYSCTL
 172
 173 #define SYSCTL_WRITES_LEGACY    -1
 174 #define SYSCTL_WRITES_WARN       0
 175 #define SYSCTL_WRITES_STRICT     1
 176
 177 static int sysctl_writes_strict = SYSCTL_WRITES_STRICT;
 178
 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 = CONFIG_MAGIC_SYSRQ_DEFAULT_ENABLE;
 200
 201 static int sysrq_sysctl_handler(struct 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 #endif /* CONFIG_SMP */
 355 #ifdef CONFIG_NUMA_BALANCING
 356         {
 357                 .procname       = "numa_balancing_scan_delay_ms",
 358                 .data           = &sysctl_numa_balancing_scan_delay,
 359                 .maxlen         = sizeof(unsigned int),
 360                 .mode           = 0644,
 361                 .proc_handler   = proc_dointvec,
 362         },
 363         {
 364                 .procname       = "numa_balancing_scan_period_min_ms",
 365                 .data           = &sysctl_numa_balancing_scan_period_min,
 366                 .maxlen         = sizeof(unsigned int),
 367                 .mode           = 0644,
 368                 .proc_handler   = proc_dointvec,
 369         },
 370         {
 371                 .procname       = "numa_balancing_scan_period_max_ms",
 372                 .data           = &sysctl_numa_balancing_scan_period_max,
 373                 .maxlen         = sizeof(unsigned int),
 374                 .mode           = 0644,
 375                 .proc_handler   = proc_dointvec,
 376         },
 377         {
 378                 .procname       = "numa_balancing_scan_size_mb",
 379                 .data           = &sysctl_numa_balancing_scan_size,
 380                 .maxlen         = sizeof(unsigned int),
 381                 .mode           = 0644,
 382                 .proc_handler   = proc_dointvec_minmax,
 383                 .extra1         = &one,
 384         },
 385         {
 386                 .procname       = "numa_balancing",
 387                 .data           = NULL, /* filled in by handler */
 388                 .maxlen         = sizeof(unsigned int),
 389                 .mode           = 0644,
 390                 .proc_handler   = sysctl_numa_balancing,
 391                 .extra1         = &zero,
 392                 .extra2         = &one,
 393         },
 394 #endif /* CONFIG_NUMA_BALANCING */
 395 #endif /* CONFIG_SCHED_DEBUG */
 396         {
 397                 .procname       = "sched_rt_period_us",
 398                 .data           = &sysctl_sched_rt_period,
 399                 .maxlen         = sizeof(unsigned int),
 400                 .mode           = 0644,
 401                 .proc_handler   = sched_rt_handler,
 402         },
 403         {
 404                 .procname       = "sched_rt_runtime_us",
 405                 .data           = &sysctl_sched_rt_runtime,
 406                 .maxlen         = sizeof(int),
 407                 .mode           = 0644,
 408                 .proc_handler   = sched_rt_handler,
 409         },
 410         {
 411                 .procname       = "sched_rr_timeslice_ms",
 412                 .data           = &sched_rr_timeslice,
 413                 .maxlen         = sizeof(int),
 414                 .mode           = 0644,
 415                 .proc_handler   = sched_rr_handler,
 416         },
 417 #ifdef CONFIG_SCHED_AUTOGROUP
 418         {
 419                 .procname       = "sched_autogroup_enabled",
 420                 .data           = &sysctl_sched_autogroup_enabled,
 421                 .maxlen         = sizeof(unsigned int),
 422                 .mode           = 0644,
 423                 .proc_handler   = proc_dointvec_minmax,
 424                 .extra1         = &zero,
 425                 .extra2         = &one,
 426         },
 427 #endif
 428 #ifdef CONFIG_CFS_BANDWIDTH
 429         {
 430                 .procname       = "sched_cfs_bandwidth_slice_us",
 431                 .data           = &sysctl_sched_cfs_bandwidth_slice,
 432                 .maxlen         = sizeof(unsigned int),
 433                 .mode           = 0644,
 434                 .proc_handler   = proc_dointvec_minmax,
 435                 .extra1         = &one,
 436         },
 437 #endif
 438 #ifdef CONFIG_PROVE_LOCKING
 439         {
 440                 .procname       = "prove_locking",
 441                 .data           = &prove_locking,
 442                 .maxlen         = sizeof(int),
 443                 .mode           = 0644,
 444                 .proc_handler   = proc_dointvec,
 445         },
 446 #endif
 447 #ifdef CONFIG_LOCK_STAT
 448         {
 449                 .procname       = "lock_stat",
 450                 .data           = &lock_stat,
 451                 .maxlen         = sizeof(int),
 452                 .mode           = 0644,
 453                 .proc_handler   = proc_dointvec,
 454         },
 455 #endif
 456         {
 457                 .procname       = "panic",
 458                 .data           = &panic_timeout,
 459                 .maxlen         = sizeof(int),
 460                 .mode           = 0644,
 461                 .proc_handler   = proc_dointvec,
 462         },
 463 #ifdef CONFIG_COREDUMP
 464         {
 465                 .procname       = "core_uses_pid",
 466                 .data           = &core_uses_pid,
 467                 .maxlen         = sizeof(int),
 468                 .mode           = 0644,
 469                 .proc_handler   = proc_dointvec,
 470         },
 471         {
 472                 .procname       = "core_pattern",
 473                 .data           = core_pattern,
 474                 .maxlen         = CORENAME_MAX_SIZE,
 475                 .mode           = 0644,
 476                 .proc_handler   = proc_dostring_coredump,
 477         },
 478         {
 479                 .procname       = "core_pipe_limit",
 480                 .data           = &core_pipe_limit,
 481                 .maxlen         = sizeof(unsigned int),
 482                 .mode           = 0644,
 483                 .proc_handler   = proc_dointvec,
 484         },
 485 #endif
 486 #ifdef CONFIG_PROC_SYSCTL
 487         {
 488                 .procname       = "tainted",
 489                 .maxlen         = sizeof(long),
 490                 .mode           = 0644,
 491                 .proc_handler   = proc_taint,
 492         },
 493         {
 494                 .procname       = "sysctl_writes_strict",
 495                 .data           = &sysctl_writes_strict,
 496                 .maxlen         = sizeof(int),
 497                 .mode           = 0644,
 498                 .proc_handler   = proc_dointvec_minmax,
 499                 .extra1         = &neg_one,
 500                 .extra2         = &one,
 501         },
 502 #endif
 503 #ifdef CONFIG_LATENCYTOP
 504         {
 505                 .procname       = "latencytop",
 506                 .data           = &latencytop_enabled,
 507                 .maxlen         = sizeof(int),
 508                 .mode           = 0644,
 509                 .proc_handler   = proc_dointvec,
 510         },
 511 #endif
 512 #ifdef CONFIG_BLK_DEV_INITRD
 513         {
 514                 .procname       = "real-root-dev",
 515                 .data           = &real_root_dev,
 516                 .maxlen         = sizeof(int),
 517                 .mode           = 0644,
 518                 .proc_handler   = proc_dointvec,
 519         },
 520 #endif
 521         {
 522                 .procname       = "print-fatal-signals",
 523                 .data           = &print_fatal_signals,
 524                 .maxlen         = sizeof(int),
 525                 .mode           = 0644,
 526                 .proc_handler   = proc_dointvec,
 527         },
 528 #ifdef CONFIG_SPARC
 529         {
 530                 .procname       = "reboot-cmd",
 531                 .data           = reboot_command,
 532                 .maxlen         = 256,
 533                 .mode           = 0644,
 534                 .proc_handler   = proc_dostring,
 535         },
 536         {
 537                 .procname       = "stop-a",
 538                 .data           = &stop_a_enabled,
 539                 .maxlen         = sizeof (int),
 540                 .mode           = 0644,
 541                 .proc_handler   = proc_dointvec,
 542         },
 543         {
 544                 .procname       = "scons-poweroff",
 545                 .data           = &scons_pwroff,
 546                 .maxlen         = sizeof (int),
 547                 .mode           = 0644,
 548                 .proc_handler   = proc_dointvec,
 549         },
 550 #endif
 551 #ifdef CONFIG_SPARC64
 552         {
 553                 .procname       = "tsb-ratio",
 554                 .data           = &sysctl_tsb_ratio,
 555                 .maxlen         = sizeof (int),
 556                 .mode           = 0644,
 557                 .proc_handler   = proc_dointvec,
 558         },
 559 #endif
 560 #ifdef __hppa__
 561         {
 562                 .procname       = "soft-power",
 563                 .data           = &pwrsw_enabled,
 564                 .maxlen         = sizeof (int),
 565                 .mode           = 0644,
 566                 .proc_handler   = proc_dointvec,
 567         },
 568 #endif
 569 #ifdef CONFIG_SYSCTL_ARCH_UNALIGN_ALLOW
 570         {
 571                 .procname       = "unaligned-trap",
 572                 .data           = &unaligned_enabled,
 573                 .maxlen         = sizeof (int),
 574                 .mode           = 0644,
 575                 .proc_handler   = proc_dointvec,
 576         },
 577 #endif
 578         {
 579                 .procname       = "ctrl-alt-del",
 580                 .data           = &C_A_D,
 581                 .maxlen         = sizeof(int),
 582                 .mode           = 0644,
 583                 .proc_handler   = proc_dointvec,
 584         },
 585 #ifdef CONFIG_FUNCTION_TRACER
 586         {
 587                 .procname       = "ftrace_enabled",
 588                 .data           = &ftrace_enabled,
 589                 .maxlen         = sizeof(int),
 590                 .mode           = 0644,
 591                 .proc_handler   = ftrace_enable_sysctl,
 592         },
 593 #endif
 594 #ifdef CONFIG_STACK_TRACER
 595         {
 596                 .procname       = "stack_tracer_enabled",
 597                 .data           = &stack_tracer_enabled,
 598                 .maxlen         = sizeof(int),
 599                 .mode           = 0644,
 600                 .proc_handler   = stack_trace_sysctl,
 601         },
 602 #endif
 603 #ifdef CONFIG_TRACING
 604         {
 605                 .procname       = "ftrace_dump_on_oops",
 606                 .data           = &ftrace_dump_on_oops,
 607                 .maxlen         = sizeof(int),
 608                 .mode           = 0644,
 609                 .proc_handler   = proc_dointvec,
 610         },
 611         {
 612                 .procname       = "traceoff_on_warning",
 613                 .data           = &__disable_trace_on_warning,
 614                 .maxlen         = sizeof(__disable_trace_on_warning),
 615                 .mode           = 0644,
 616                 .proc_handler   = proc_dointvec,
 617         },
 618         {
 619                 .procname       = "tracepoint_printk",
 620                 .data           = &tracepoint_printk,
 621                 .maxlen         = sizeof(tracepoint_printk),
 622                 .mode           = 0644,
 623                 .proc_handler   = proc_dointvec,
 624         },
 625 #endif
 626 #ifdef CONFIG_KEXEC_CORE
 627         {
 628                 .procname       = "kexec_load_disabled",
 629                 .data           = &kexec_load_disabled,
 630                 .maxlen         = sizeof(int),
 631                 .mode           = 0644,
 632                 /* only handle a transition from default "0" to "1" */
 633                 .proc_handler   = proc_dointvec_minmax,
 634                 .extra1         = &one,
 635                 .extra2         = &one,
 636         },
 637 #endif
 638 #ifdef CONFIG_MODULES
 639         {
 640                 .procname       = "modprobe",
 641                 .data           = &modprobe_path,
 642                 .maxlen         = KMOD_PATH_LEN,
 643                 .mode           = 0644,
 644                 .proc_handler   = proc_dostring,
 645         },
 646         {
 647                 .procname       = "modules_disabled",
 648                 .data           = &modules_disabled,
 649                 .maxlen         = sizeof(int),
 650                 .mode           = 0644,
 651                 /* only handle a transition from default "0" to "1" */
 652                 .proc_handler   = proc_dointvec_minmax,
 653                 .extra1         = &one,
 654                 .extra2         = &one,
 655         },
 656 #endif
 657 #ifdef CONFIG_UEVENT_HELPER
 658         {
 659                 .procname       = "hotplug",
 660                 .data           = &uevent_helper,
 661                 .maxlen         = UEVENT_HELPER_PATH_LEN,
 662                 .mode           = 0644,
 663                 .proc_handler   = proc_dostring,
 664         },
 665 #endif
 666 #ifdef CONFIG_CHR_DEV_SG
 667         {
 668                 .procname       = "sg-big-buff",
 669                 .data           = &sg_big_buff,
 670                 .maxlen         = sizeof (int),
 671                 .mode           = 0444,
 672                 .proc_handler   = proc_dointvec,
 673         },
 674 #endif
 675 #ifdef CONFIG_BSD_PROCESS_ACCT
 676         {
 677                 .procname       = "acct",
 678                 .data           = &acct_parm,
 679                 .maxlen         = 3*sizeof(int),
 680                 .mode           = 0644,
 681                 .proc_handler   = proc_dointvec,
 682         },
 683 #endif
 684 #ifdef CONFIG_MAGIC_SYSRQ
 685         {
 686                 .procname       = "sysrq",
 687                 .data           = &__sysrq_enabled,
 688                 .maxlen         = sizeof (int),
 689                 .mode           = 0644,
 690                 .proc_handler   = sysrq_sysctl_handler,
 691         },
 692 #endif
 693 #ifdef CONFIG_PROC_SYSCTL
 694         {
 695                 .procname       = "cad_pid",
 696                 .data           = NULL,
 697                 .maxlen         = sizeof (int),
 698                 .mode           = 0600,
 699                 .proc_handler   = proc_do_cad_pid,
 700         },
 701 #endif
 702         {
 703                 .procname       = "threads-max",
 704                 .data           = NULL,
 705                 .maxlen         = sizeof(int),
 706                 .mode           = 0644,
 707                 .proc_handler   = sysctl_max_threads,
 708         },
 709         {
 710                 .procname       = "random",
 711                 .mode           = 0555,
 712                 .child          = random_table,
 713         },
 714         {
 715                 .procname       = "usermodehelper",
 716                 .mode           = 0555,
 717                 .child          = usermodehelper_table,
 718         },
 719         {
 720                 .procname       = "overflowuid",
 721                 .data           = &overflowuid,
 722                 .maxlen         = sizeof(int),
 723                 .mode           = 0644,
 724                 .proc_handler   = proc_dointvec_minmax,
 725                 .extra1         = &minolduid,
 726                 .extra2         = &maxolduid,
 727         },
 728         {
 729                 .procname       = "overflowgid",
 730                 .data           = &overflowgid,
 731                 .maxlen         = sizeof(int),
 732                 .mode           = 0644,
 733                 .proc_handler   = proc_dointvec_minmax,
 734                 .extra1         = &minolduid,
 735                 .extra2         = &maxolduid,
 736         },
 737 #ifdef CONFIG_S390
 738 #ifdef CONFIG_MATHEMU
 739         {
 740                 .procname       = "ieee_emulation_warnings",
 741                 .data           = &sysctl_ieee_emulation_warnings,
 742                 .maxlen         = sizeof(int),
 743                 .mode           = 0644,
 744                 .proc_handler   = proc_dointvec,
 745         },
 746 #endif
 747         {
 748                 .procname       = "userprocess_debug",
 749                 .data           = &show_unhandled_signals,
 750                 .maxlen         = sizeof(int),
 751                 .mode           = 0644,
 752                 .proc_handler   = proc_dointvec,
 753         },
 754 #endif
 755         {
 756                 .procname       = "pid_max",
 757                 .data           = &pid_max,
 758                 .maxlen         = sizeof (int),
 759                 .mode           = 0644,
 760                 .proc_handler   = proc_dointvec_minmax,
 761                 .extra1         = &pid_max_min,
 762                 .extra2         = &pid_max_max,
 763         },
 764         {
 765                 .procname       = "panic_on_oops",
 766                 .data           = &panic_on_oops,
 767                 .maxlen         = sizeof(int),
 768                 .mode           = 0644,
 769                 .proc_handler   = proc_dointvec,
 770         },
 771 #if defined CONFIG_PRINTK
 772         {
 773                 .procname       = "printk",
 774                 .data           = &console_loglevel,
 775                 .maxlen         = 4*sizeof(int),
 776                 .mode           = 0644,
 777                 .proc_handler   = proc_dointvec,
 778         },
 779         {
 780                 .procname       = "printk_ratelimit",
 781                 .data           = &printk_ratelimit_state.interval,
 782                 .maxlen         = sizeof(int),
 783                 .mode           = 0644,
 784                 .proc_handler   = proc_dointvec_jiffies,
 785         },
 786         {
 787                 .procname       = "printk_ratelimit_burst",
 788                 .data           = &printk_ratelimit_state.burst,
 789                 .maxlen         = sizeof(int),
 790                 .mode           = 0644,
 791                 .proc_handler   = proc_dointvec,
 792         },
 793         {
 794                 .procname       = "printk_delay",
 795                 .data           = &printk_delay_msec,
 796                 .maxlen         = sizeof(int),
 797                 .mode           = 0644,
 798                 .proc_handler   = proc_dointvec_minmax,
 799                 .extra1         = &zero,
 800                 .extra2         = &ten_thousand,
 801         },
 802         {
 803                 .procname       = "dmesg_restrict",
 804                 .data           = &dmesg_restrict,
 805                 .maxlen         = sizeof(int),
 806                 .mode           = 0644,
 807                 .proc_handler   = proc_dointvec_minmax_sysadmin,
 808                 .extra1         = &zero,
 809                 .extra2         = &one,
 810         },
 811         {
 812                 .procname       = "kptr_restrict",
 813                 .data           = &kptr_restrict,
 814                 .maxlen         = sizeof(int),
 815                 .mode           = 0644,
 816                 .proc_handler   = proc_dointvec_minmax_sysadmin,
 817                 .extra1         = &zero,
 818                 .extra2         = &two,
 819         },
 820 #endif
 821         {
 822                 .procname       = "ngroups_max",
 823                 .data           = &ngroups_max,
 824                 .maxlen         = sizeof (int),
 825                 .mode           = 0444,
 826                 .proc_handler   = proc_dointvec,
 827         },
 828         {
 829                 .procname       = "cap_last_cap",
 830                 .data           = (void *)&cap_last_cap,
 831                 .maxlen         = sizeof(int),
 832                 .mode           = 0444,
 833                 .proc_handler   = proc_dointvec,
 834         },
 835 #if defined(CONFIG_LOCKUP_DETECTOR)
 836         {
 837                 .procname       = "watchdog",
 838                 .data           = &watchdog_user_enabled,
 839                 .maxlen         = sizeof (int),
 840                 .mode           = 0644,
 841                 .proc_handler   = proc_watchdog,
 842                 .extra1         = &zero,
 843                 .extra2         = &one,
 844         },
 845         {
 846                 .procname       = "watchdog_thresh",
 847                 .data           = &watchdog_thresh,
 848                 .maxlen         = sizeof(int),
 849                 .mode           = 0644,
 850                 .proc_handler   = proc_watchdog_thresh,
 851                 .extra1         = &zero,
 852                 .extra2         = &sixty,
 853         },
 854         {
 855                 .procname       = "nmi_watchdog",
 856                 .data           = &nmi_watchdog_enabled,
 857                 .maxlen         = sizeof (int),
 858                 .mode           = 0644,
 859                 .proc_handler   = proc_nmi_watchdog,
 860                 .extra1         = &zero,
 861 #if defined(CONFIG_HAVE_NMI_WATCHDOG) || defined(CONFIG_HARDLOCKUP_DETECTOR)
 862                 .extra2         = &one,
 863 #else
 864                 .extra2         = &zero,
 865 #endif
 866         },
 867         {
 868                 .procname       = "soft_watchdog",
 869                 .data           = &soft_watchdog_enabled,
 870                 .maxlen         = sizeof (int),
 871                 .mode           = 0644,
 872                 .proc_handler   = proc_soft_watchdog,
 873                 .extra1         = &zero,
 874                 .extra2         = &one,
 875         },
 876         {
 877                 .procname       = "watchdog_cpumask",
 878                 .data           = &watchdog_cpumask_bits,
 879                 .maxlen         = NR_CPUS,
 880                 .mode           = 0644,
 881                 .proc_handler   = proc_watchdog_cpumask,
 882         },
 883         {
 884                 .procname       = "softlockup_panic",
 885                 .data           = &softlockup_panic,
 886                 .maxlen         = sizeof(int),
 887                 .mode           = 0644,
 888                 .proc_handler   = proc_dointvec_minmax,
 889                 .extra1         = &zero,
 890                 .extra2         = &one,
 891         },
 892 #ifdef CONFIG_HARDLOCKUP_DETECTOR
 893         {
 894                 .procname       = "hardlockup_panic",
 895                 .data           = &hardlockup_panic,
 896                 .maxlen         = sizeof(int),
 897                 .mode           = 0644,
 898                 .proc_handler   = proc_dointvec_minmax,
 899                 .extra1         = &zero,
 900                 .extra2         = &one,
 901         },
 902 #endif
 903 #ifdef CONFIG_SMP
 904         {
 905                 .procname       = "softlockup_all_cpu_backtrace",
 906                 .data           = &sysctl_softlockup_all_cpu_backtrace,
 907                 .maxlen         = sizeof(int),
 908                 .mode           = 0644,
 909                 .proc_handler   = proc_dointvec_minmax,
 910                 .extra1         = &zero,
 911                 .extra2         = &one,
 912         },
 913         {
 914                 .procname       = "hardlockup_all_cpu_backtrace",
 915                 .data           = &sysctl_hardlockup_all_cpu_backtrace,
 916                 .maxlen         = sizeof(int),
 917                 .mode           = 0644,
 918                 .proc_handler   = proc_dointvec_minmax,
 919                 .extra1         = &zero,
 920                 .extra2         = &one,
 921         },
 922 #endif /* CONFIG_SMP */
 923 #endif
 924 #if defined(CONFIG_X86_LOCAL_APIC) && defined(CONFIG_X86)
 925         {
 926                 .procname       = "unknown_nmi_panic",
 927                 .data           = &unknown_nmi_panic,
 928                 .maxlen         = sizeof (int),
 929                 .mode           = 0644,
 930                 .proc_handler   = proc_dointvec,
 931         },
 932 #endif
 933 #if defined(CONFIG_X86)
 934         {
 935                 .procname       = "panic_on_unrecovered_nmi",
 936                 .data           = &panic_on_unrecovered_nmi,
 937                 .maxlen         = sizeof(int),
 938                 .mode           = 0644,
 939                 .proc_handler   = proc_dointvec,
 940         },
 941         {
 942                 .procname       = "panic_on_io_nmi",
 943                 .data           = &panic_on_io_nmi,
 944                 .maxlen         = sizeof(int),
 945                 .mode           = 0644,
 946                 .proc_handler   = proc_dointvec,
 947         },
 948 #ifdef CONFIG_DEBUG_STACKOVERFLOW
 949         {
 950                 .procname       = "panic_on_stackoverflow",
 951                 .data           = &sysctl_panic_on_stackoverflow,
 952                 .maxlen         = sizeof(int),
 953                 .mode           = 0644,
 954                 .proc_handler   = proc_dointvec,
 955         },
 956 #endif
 957         {
 958                 .procname       = "bootloader_type",
 959                 .data           = &bootloader_type,
 960                 .maxlen         = sizeof (int),
 961                 .mode           = 0444,
 962                 .proc_handler   = proc_dointvec,
 963         },
 964         {
 965                 .procname       = "bootloader_version",
 966                 .data           = &bootloader_version,
 967                 .maxlen         = sizeof (int),
 968                 .mode           = 0444,
 969                 .proc_handler   = proc_dointvec,
 970         },
 971         {
 972                 .procname       = "kstack_depth_to_print",
 973                 .data           = &kstack_depth_to_print,
 974                 .maxlen         = sizeof(int),
 975                 .mode           = 0644,
 976                 .proc_handler   = proc_dointvec,
 977         },
 978         {
 979                 .procname       = "io_delay_type",
 980                 .data           = &io_delay_type,
 981                 .maxlen         = sizeof(int),
 982                 .mode           = 0644,
 983                 .proc_handler   = proc_dointvec,
 984         },
 985 #endif
 986 #if defined(CONFIG_MMU)
 987         {
 988                 .procname       = "randomize_va_space",
 989                 .data           = &randomize_va_space,
 990                 .maxlen         = sizeof(int),
 991                 .mode           = 0644,
 992                 .proc_handler   = proc_dointvec,
 993         },
 994 #endif
 995 #if defined(CONFIG_S390) && defined(CONFIG_SMP)
 996         {
 997                 .procname       = "spin_retry",
 998                 .data           = &spin_retry,
 999                 .maxlen         = sizeof (int),
1000                 .mode           = 0644,
1001                 .proc_handler   = proc_dointvec,
1002         },
1003 #endif
1004 #if     defined(CONFIG_ACPI_SLEEP) && defined(CONFIG_X86)
1005         {
1006                 .procname       = "acpi_video_flags",
1007                 .data           = &acpi_realmode_flags,
1008                 .maxlen         = sizeof (unsigned long),
1009                 .mode           = 0644,
1010                 .proc_handler   = proc_doulongvec_minmax,
1011         },
1012 #endif
1013 #ifdef CONFIG_SYSCTL_ARCH_UNALIGN_NO_WARN
1014         {
1015                 .procname       = "ignore-unaligned-usertrap",
1016                 .data           = &no_unaligned_warning,
1017                 .maxlen         = sizeof (int),
1018                 .mode           = 0644,
1019                 .proc_handler   = proc_dointvec,
1020         },
1021 #endif
1022 #ifdef CONFIG_IA64
1023         {
1024                 .procname       = "unaligned-dump-stack",
1025                 .data           = &unaligned_dump_stack,
1026                 .maxlen         = sizeof (int),
1027                 .mode           = 0644,
1028                 .proc_handler   = proc_dointvec,
1029         },
1030 #endif
1031 #ifdef CONFIG_DETECT_HUNG_TASK
1032         {
1033                 .procname       = "hung_task_panic",
1034                 .data           = &sysctl_hung_task_panic,
1035                 .maxlen         = sizeof(int),
1036                 .mode           = 0644,
1037                 .proc_handler   = proc_dointvec_minmax,
1038                 .extra1         = &zero,
1039                 .extra2         = &one,
1040         },
1041         {
1042                 .procname       = "hung_task_check_count",
1043                 .data           = &sysctl_hung_task_check_count,
1044                 .maxlen         = sizeof(int),
1045                 .mode           = 0644,
1046                 .proc_handler   = proc_dointvec_minmax,
1047                 .extra1         = &zero,
1048         },
1049         {
1050                 .procname       = "hung_task_timeout_secs",
1051                 .data           = &sysctl_hung_task_timeout_secs,
1052                 .maxlen         = sizeof(unsigned long),
1053                 .mode           = 0644,
1054                 .proc_handler   = proc_dohung_task_timeout_secs,
1055                 .extra2         = &hung_task_timeout_max,
1056         },
1057         {
1058                 .procname       = "hung_task_warnings",
1059                 .data           = &sysctl_hung_task_warnings,
1060                 .maxlen         = sizeof(int),
1061                 .mode           = 0644,
1062                 .proc_handler   = proc_dointvec_minmax,
1063                 .extra1         = &neg_one,
1064         },
1065 #endif
1066 #ifdef CONFIG_COMPAT
1067         {
1068                 .procname       = "compat-log",
1069                 .data           = &compat_log,
1070                 .maxlen         = sizeof (int),
1071                 .mode           = 0644,
1072                 .proc_handler   = proc_dointvec,
1073         },
1074 #endif
1075 #ifdef CONFIG_RT_MUTEXES
1076         {
1077                 .procname       = "max_lock_depth",
1078                 .data           = &max_lock_depth,
1079                 .maxlen         = sizeof(int),
1080                 .mode           = 0644,
1081                 .proc_handler   = proc_dointvec,
1082         },
1083 #endif
1084         {
1085                 .procname       = "poweroff_cmd",
1086                 .data           = &poweroff_cmd,
1087                 .maxlen         = POWEROFF_CMD_PATH_LEN,
1088                 .mode           = 0644,
1089                 .proc_handler   = proc_dostring,
1090         },
1091 #ifdef CONFIG_KEYS
1092         {
1093                 .procname       = "keys",
1094                 .mode           = 0555,
1095                 .child          = key_sysctls,
1096         },
1097 #endif
1098 #ifdef CONFIG_PERF_EVENTS
1099         /*
1100          * User-space scripts rely on the existence of this file
1101          * as a feature check for perf_events being enabled.
1102          *
1103          * So it's an ABI, do not remove!
1104          */
1105         {
1106                 .procname       = "perf_event_paranoid",
1107                 .data           = &sysctl_perf_event_paranoid,
1108                 .maxlen         = sizeof(sysctl_perf_event_paranoid),
1109                 .mode           = 0644,
1110                 .proc_handler   = proc_dointvec,
1111         },
1112         {
1113                 .procname       = "perf_event_mlock_kb",
1114                 .data           = &sysctl_perf_event_mlock,
1115                 .maxlen         = sizeof(sysctl_perf_event_mlock),
1116                 .mode           = 0644,
1117                 .proc_handler   = proc_dointvec,
1118         },
1119         {
1120                 .procname       = "perf_event_max_sample_rate",
1121                 .data           = &sysctl_perf_event_sample_rate,
1122                 .maxlen         = sizeof(sysctl_perf_event_sample_rate),
1123                 .mode           = 0644,
1124                 .proc_handler   = perf_proc_update_handler,
1125                 .extra1         = &one,
1126         },
1127         {
1128                 .procname       = "perf_cpu_time_max_percent",
1129                 .data           = &sysctl_perf_cpu_time_max_percent,
1130                 .maxlen         = sizeof(sysctl_perf_cpu_time_max_percent),
1131                 .mode           = 0644,
1132                 .proc_handler   = perf_cpu_time_max_percent_handler,
1133                 .extra1         = &zero,
1134                 .extra2         = &one_hundred,
1135         },
1136 #endif
1137 #ifdef CONFIG_KMEMCHECK
1138         {
1139                 .procname       = "kmemcheck",
1140                 .data           = &kmemcheck_enabled,
1141                 .maxlen         = sizeof(int),
1142                 .mode           = 0644,
1143                 .proc_handler   = proc_dointvec,
1144         },
1145 #endif
1146         {
1147                 .procname       = "panic_on_warn",
1148                 .data           = &panic_on_warn,
1149                 .maxlen         = sizeof(int),
1150                 .mode           = 0644,
1151                 .proc_handler   = proc_dointvec_minmax,
1152                 .extra1         = &zero,
1153                 .extra2         = &one,
1154         },
1155 #if defined(CONFIG_SMP) && defined(CONFIG_NO_HZ_COMMON)
1156         {
1157                 .procname       = "timer_migration",
1158                 .data           = &sysctl_timer_migration,
1159                 .maxlen         = sizeof(unsigned int),
1160                 .mode           = 0644,
1161                 .proc_handler   = timer_migration_handler,
1162                 .extra1         = &zero,
1163                 .extra2         = &one,
1164         },
1165 #endif
1166 #ifdef CONFIG_BPF_SYSCALL
1167         {
1168                 .procname       = "unprivileged_bpf_disabled",
1169                 .data           = &sysctl_unprivileged_bpf_disabled,
1170                 .maxlen         = sizeof(sysctl_unprivileged_bpf_disabled),
1171                 .mode           = 0644,
1172                 /* only handle a transition from default "0" to "1" */
1173                 .proc_handler   = proc_dointvec_minmax,
1174                 .extra1         = &one,
1175                 .extra2         = &one,
1176         },
1177 #endif
1178         { }
1179 };
1180
1181 static struct ctl_table vm_table[] = {
1182         {
1183                 .procname       = "overcommit_memory",
1184                 .data           = &sysctl_overcommit_memory,
1185                 .maxlen         = sizeof(sysctl_overcommit_memory),
1186                 .mode           = 0644,
1187                 .proc_handler   = proc_dointvec_minmax,
1188                 .extra1         = &zero,
1189                 .extra2         = &two,
1190         },
1191         {
1192                 .procname       = "panic_on_oom",
1193                 .data           = &sysctl_panic_on_oom,
1194                 .maxlen         = sizeof(sysctl_panic_on_oom),
1195                 .mode           = 0644,
1196                 .proc_handler   = proc_dointvec_minmax,
1197                 .extra1         = &zero,
1198                 .extra2         = &two,
1199         },
1200         {
1201                 .procname       = "oom_kill_allocating_task",
1202                 .data           = &sysctl_oom_kill_allocating_task,
1203                 .maxlen         = sizeof(sysctl_oom_kill_allocating_task),
1204                 .mode           = 0644,
1205                 .proc_handler   = proc_dointvec,
1206         },
1207         {
1208                 .procname       = "oom_dump_tasks",
1209                 .data           = &sysctl_oom_dump_tasks,
1210                 .maxlen         = sizeof(sysctl_oom_dump_tasks),
1211                 .mode           = 0644,
1212                 .proc_handler   = proc_dointvec,
1213         },
1214         {
1215                 .procname       = "overcommit_ratio",
1216                 .data           = &sysctl_overcommit_ratio,
1217                 .maxlen         = sizeof(sysctl_overcommit_ratio),
1218                 .mode           = 0644,
1219                 .proc_handler   = overcommit_ratio_handler,
1220         },
1221         {
1222                 .procname       = "overcommit_kbytes",
1223                 .data           = &sysctl_overcommit_kbytes,
1224                 .maxlen         = sizeof(sysctl_overcommit_kbytes),
1225                 .mode           = 0644,
1226                 .proc_handler   = overcommit_kbytes_handler,
1227         },
1228         {
1229                 .procname       = "page-cluster",
1230                 .data           = &page_cluster,
1231                 .maxlen         = sizeof(int),
1232                 .mode           = 0644,
1233                 .proc_handler   = proc_dointvec_minmax,
1234                 .extra1         = &zero,
1235         },
1236         {
1237                 .procname       = "dirty_background_ratio",
1238                 .data           = &dirty_background_ratio,
1239                 .maxlen         = sizeof(dirty_background_ratio),
1240                 .mode           = 0644,
1241                 .proc_handler   = dirty_background_ratio_handler,
1242                 .extra1         = &zero,
1243                 .extra2         = &one_hundred,
1244         },
1245         {
1246                 .procname       = "dirty_background_bytes",
1247                 .data           = &dirty_background_bytes,
1248                 .maxlen         = sizeof(dirty_background_bytes),
1249                 .mode           = 0644,
1250                 .proc_handler   = dirty_background_bytes_handler,
1251                 .extra1         = &one_ul,
1252         },
1253         {
1254                 .procname       = "dirty_ratio",
1255                 .data           = &vm_dirty_ratio,
1256                 .maxlen         = sizeof(vm_dirty_ratio),
1257                 .mode           = 0644,
1258                 .proc_handler   = dirty_ratio_handler,
1259                 .extra1         = &zero,
1260                 .extra2         = &one_hundred,
1261         },
1262         {
1263                 .procname       = "dirty_bytes",
1264                 .data           = &vm_dirty_bytes,
1265                 .maxlen         = sizeof(vm_dirty_bytes),
1266                 .mode           = 0644,
1267                 .proc_handler   = dirty_bytes_handler,
1268                 .extra1         = &dirty_bytes_min,
1269         },
1270         {
1271                 .procname       = "dirty_writeback_centisecs",
1272                 .data           = &dirty_writeback_interval,
1273                 .maxlen         = sizeof(dirty_writeback_interval),
1274                 .mode           = 0644,
1275                 .proc_handler   = dirty_writeback_centisecs_handler,
1276         },
1277         {
1278                 .procname       = "dirty_expire_centisecs",
1279                 .data           = &dirty_expire_interval,
1280                 .maxlen         = sizeof(dirty_expire_interval),
1281                 .mode           = 0644,
1282                 .proc_handler   = proc_dointvec_minmax,
1283                 .extra1         = &zero,
1284         },
1285         {
1286                 .procname       = "dirtytime_expire_seconds",
1287                 .data           = &dirtytime_expire_interval,
1288                 .maxlen         = sizeof(dirty_expire_interval),
1289                 .mode           = 0644,
1290                 .proc_handler   = dirtytime_interval_handler,
1291                 .extra1         = &zero,
1292         },
1293         {
1294                 .procname       = "nr_pdflush_threads",
1295                 .mode           = 0444 /* read-only */,
1296                 .proc_handler   = pdflush_proc_obsolete,
1297         },
1298         {
1299                 .procname       = "swappiness",
1300                 .data           = &vm_swappiness,
1301                 .maxlen         = sizeof(vm_swappiness),
1302                 .mode           = 0644,
1303                 .proc_handler   = proc_dointvec_minmax,
1304                 .extra1         = &zero,
1305                 .extra2         = &one_hundred,
1306         },
1307 #ifdef CONFIG_HUGETLB_PAGE
1308         {
1309                 .procname       = "nr_hugepages",
1310                 .data           = NULL,
1311                 .maxlen         = sizeof(unsigned long),
1312                 .mode           = 0644,
1313                 .proc_handler   = hugetlb_sysctl_handler,
1314         },
1315 #ifdef CONFIG_NUMA
1316         {
1317                 .procname       = "nr_hugepages_mempolicy",
1318                 .data           = NULL,
1319                 .maxlen         = sizeof(unsigned long),
1320                 .mode           = 0644,
1321                 .proc_handler   = &hugetlb_mempolicy_sysctl_handler,
1322         },
1323 #endif
1324          {
1325                 .procname       = "hugetlb_shm_group",
1326                 .data           = &sysctl_hugetlb_shm_group,
1327                 .maxlen         = sizeof(gid_t),
1328                 .mode           = 0644,
1329                 .proc_handler   = proc_dointvec,
1330          },
1331          {
1332                 .procname       = "hugepages_treat_as_movable",
1333                 .data           = &hugepages_treat_as_movable,
1334                 .maxlen         = sizeof(int),
1335                 .mode           = 0644,
1336                 .proc_handler   = proc_dointvec,
1337         },
1338         {
1339                 .procname       = "nr_overcommit_hugepages",
1340                 .data           = NULL,
1341                 .maxlen         = sizeof(unsigned long),
1342                 .mode           = 0644,
1343                 .proc_handler   = hugetlb_overcommit_handler,
1344         },
1345 #endif
1346         {
1347                 .procname       = "lowmem_reserve_ratio",
1348                 .data           = &sysctl_lowmem_reserve_ratio,
1349                 .maxlen         = sizeof(sysctl_lowmem_reserve_ratio),
1350                 .mode           = 0644,
1351                 .proc_handler   = lowmem_reserve_ratio_sysctl_handler,
1352         },
1353         {
1354                 .procname       = "drop_caches",
1355                 .data           = &sysctl_drop_caches,
1356                 .maxlen         = sizeof(int),
1357                 .mode           = 0644,
1358                 .proc_handler   = drop_caches_sysctl_handler,
1359                 .extra1         = &one,
1360                 .extra2         = &four,
1361         },
1362 #ifdef CONFIG_COMPACTION
1363         {
1364                 .procname       = "compact_memory",
1365                 .data           = &sysctl_compact_memory,
1366                 .maxlen         = sizeof(int),
1367                 .mode           = 0200,
1368                 .proc_handler   = sysctl_compaction_handler,
1369         },
1370         {
1371                 .procname       = "extfrag_threshold",
1372                 .data           = &sysctl_extfrag_threshold,
1373                 .maxlen         = sizeof(int),
1374                 .mode           = 0644,
1375                 .proc_handler   = sysctl_extfrag_handler,
1376                 .extra1         = &min_extfrag_threshold,
1377                 .extra2         = &max_extfrag_threshold,
1378         },
1379         {
1380                 .procname       = "compact_unevictable_allowed",
1381                 .data           = &sysctl_compact_unevictable_allowed,
1382                 .maxlen         = sizeof(int),
1383                 .mode           = 0644,
1384                 .proc_handler   = proc_dointvec,
1385                 .extra1         = &zero,
1386                 .extra2         = &one,
1387         },
1388
1389 #endif /* CONFIG_COMPACTION */
1390         {
1391                 .procname       = "min_free_kbytes",
1392                 .data           = &min_free_kbytes,
1393                 .maxlen         = sizeof(min_free_kbytes),
1394                 .mode           = 0644,
1395                 .proc_handler   = min_free_kbytes_sysctl_handler,
1396                 .extra1         = &zero,
1397         },
1398         {
1399                 .procname       = "percpu_pagelist_fraction",
1400                 .data           = &percpu_pagelist_fraction,
1401                 .maxlen         = sizeof(percpu_pagelist_fraction),
1402                 .mode           = 0644,
1403                 .proc_handler   = percpu_pagelist_fraction_sysctl_handler,
1404                 .extra1         = &zero,
1405         },
1406 #ifdef CONFIG_MMU
1407         {
1408                 .procname       = "max_map_count",
1409                 .data           = &sysctl_max_map_count,
1410                 .maxlen         = sizeof(sysctl_max_map_count),
1411                 .mode           = 0644,
1412                 .proc_handler   = proc_dointvec_minmax,
1413                 .extra1         = &zero,
1414         },
1415 #else
1416         {
1417                 .procname       = "nr_trim_pages",
1418                 .data           = &sysctl_nr_trim_pages,
1419                 .maxlen         = sizeof(sysctl_nr_trim_pages),
1420                 .mode           = 0644,
1421                 .proc_handler   = proc_dointvec_minmax,
1422                 .extra1         = &zero,
1423         },
1424 #endif
1425         {
1426                 .procname       = "laptop_mode",
1427                 .data           = &laptop_mode,
1428                 .maxlen         = sizeof(laptop_mode),
1429                 .mode           = 0644,
1430                 .proc_handler   = proc_dointvec_jiffies,
1431         },
1432         {
1433                 .procname       = "block_dump",
1434                 .data           = &block_dump,
1435                 .maxlen         = sizeof(block_dump),
1436                 .mode           = 0644,
1437                 .proc_handler   = proc_dointvec,
1438                 .extra1         = &zero,
1439         },
1440         {
1441                 .procname       = "vfs_cache_pressure",
1442                 .data           = &sysctl_vfs_cache_pressure,
1443                 .maxlen         = sizeof(sysctl_vfs_cache_pressure),
1444                 .mode           = 0644,
1445                 .proc_handler   = proc_dointvec,
1446                 .extra1         = &zero,
1447         },
1448 #ifdef HAVE_ARCH_PICK_MMAP_LAYOUT
1449         {
1450                 .procname       = "legacy_va_layout",
1451                 .data           = &sysctl_legacy_va_layout,
1452                 .maxlen         = sizeof(sysctl_legacy_va_layout),
1453                 .mode           = 0644,
1454                 .proc_handler   = proc_dointvec,
1455                 .extra1         = &zero,
1456         },
1457 #endif
1458 #ifdef CONFIG_NUMA
1459         {
1460                 .procname       = "zone_reclaim_mode",
1461                 .data           = &zone_reclaim_mode,
1462                 .maxlen         = sizeof(zone_reclaim_mode),
1463                 .mode           = 0644,
1464                 .proc_handler   = proc_dointvec,
1465                 .extra1         = &zero,
1466         },
1467         {
1468                 .procname       = "min_unmapped_ratio",
1469                 .data           = &sysctl_min_unmapped_ratio,
1470                 .maxlen         = sizeof(sysctl_min_unmapped_ratio),
1471                 .mode           = 0644,
1472                 .proc_handler   = sysctl_min_unmapped_ratio_sysctl_handler,
1473                 .extra1         = &zero,
1474                 .extra2         = &one_hundred,
1475         },
1476         {
1477                 .procname       = "min_slab_ratio",
1478                 .data           = &sysctl_min_slab_ratio,
1479                 .maxlen         = sizeof(sysctl_min_slab_ratio),
1480                 .mode           = 0644,
1481                 .proc_handler   = sysctl_min_slab_ratio_sysctl_handler,
1482                 .extra1         = &zero,
1483                 .extra2         = &one_hundred,
1484         },
1485 #endif
1486 #ifdef CONFIG_SMP
1487         {
1488                 .procname       = "stat_interval",
1489                 .data           = &sysctl_stat_interval,
1490                 .maxlen         = sizeof(sysctl_stat_interval),
1491                 .mode           = 0644,
1492                 .proc_handler   = proc_dointvec_jiffies,
1493         },
1494 #endif
1495 #ifdef CONFIG_MMU
1496         {
1497                 .procname       = "mmap_min_addr",
1498                 .data           = &dac_mmap_min_addr,
1499                 .maxlen         = sizeof(unsigned long),
1500                 .mode           = 0644,
1501                 .proc_handler   = mmap_min_addr_handler,
1502         },
1503 #endif
1504 #ifdef CONFIG_NUMA
1505         {
1506                 .procname       = "numa_zonelist_order",
1507                 .data           = &numa_zonelist_order,
1508                 .maxlen         = NUMA_ZONELIST_ORDER_LEN,
1509                 .mode           = 0644,
1510                 .proc_handler   = numa_zonelist_order_handler,
1511         },
1512 #endif
1513 #if (defined(CONFIG_X86_32) && !defined(CONFIG_UML))|| \
1514    (defined(CONFIG_SUPERH) && defined(CONFIG_VSYSCALL))
1515         {
1516                 .procname       = "vdso_enabled",
1517 #ifdef CONFIG_X86_32
1518                 .data           = &vdso32_enabled,
1519                 .maxlen         = sizeof(vdso32_enabled),
1520 #else
1521                 .data           = &vdso_enabled,
1522                 .maxlen         = sizeof(vdso_enabled),
1523 #endif
1524                 .mode           = 0644,
1525                 .proc_handler   = proc_dointvec,
1526                 .extra1         = &zero,
1527         },
1528 #endif
1529 #ifdef CONFIG_HIGHMEM
1530         {
1531                 .procname       = "highmem_is_dirtyable",
1532                 .data           = &vm_highmem_is_dirtyable,
1533                 .maxlen         = sizeof(vm_highmem_is_dirtyable),
1534                 .mode           = 0644,
1535                 .proc_handler   = proc_dointvec_minmax,
1536                 .extra1         = &zero,
1537                 .extra2         = &one,
1538         },
1539 #endif
1540 #ifdef CONFIG_MEMORY_FAILURE
1541         {
1542                 .procname       = "memory_failure_early_kill",
1543                 .data           = &sysctl_memory_failure_early_kill,
1544                 .maxlen         = sizeof(sysctl_memory_failure_early_kill),
1545                 .mode           = 0644,
1546                 .proc_handler   = proc_dointvec_minmax,
1547                 .extra1         = &zero,
1548                 .extra2         = &one,
1549         },
1550         {
1551                 .procname       = "memory_failure_recovery",
1552                 .data           = &sysctl_memory_failure_recovery,
1553                 .maxlen         = sizeof(sysctl_memory_failure_recovery),
1554                 .mode           = 0644,
1555                 .proc_handler   = proc_dointvec_minmax,
1556                 .extra1         = &zero,
1557                 .extra2         = &one,
1558         },
1559 #endif
1560         {
1561                 .procname       = "user_reserve_kbytes",
1562                 .data           = &sysctl_user_reserve_kbytes,
1563                 .maxlen         = sizeof(sysctl_user_reserve_kbytes),
1564                 .mode           = 0644,
1565                 .proc_handler   = proc_doulongvec_minmax,
1566         },
1567         {
1568                 .procname       = "admin_reserve_kbytes",
1569                 .data           = &sysctl_admin_reserve_kbytes,
1570                 .maxlen         = sizeof(sysctl_admin_reserve_kbytes),
1571                 .mode           = 0644,
1572                 .proc_handler   = proc_doulongvec_minmax,
1573         },
1574         { }
1575 };
1576
1577 static struct ctl_table fs_table[] = {
1578         {
1579                 .procname       = "inode-nr",
1580                 .data           = &inodes_stat,
1581                 .maxlen         = 2*sizeof(long),
1582                 .mode           = 0444,
1583                 .proc_handler   = proc_nr_inodes,
1584         },
1585         {
1586                 .procname       = "inode-state",
1587                 .data           = &inodes_stat,
1588                 .maxlen         = 7*sizeof(long),
1589                 .mode           = 0444,
1590                 .proc_handler   = proc_nr_inodes,
1591         },
1592         {
1593                 .procname       = "file-nr",
1594                 .data           = &files_stat,
1595                 .maxlen         = sizeof(files_stat),
1596                 .mode           = 0444,
1597                 .proc_handler   = proc_nr_files,
1598         },
1599         {
1600                 .procname       = "file-max",
1601                 .data           = &files_stat.max_files,
1602                 .maxlen         = sizeof(files_stat.max_files),
1603                 .mode           = 0644,
1604                 .proc_handler   = proc_doulongvec_minmax,
1605         },
1606         {
1607                 .procname       = "nr_open",
1608                 .data           = &sysctl_nr_open,
1609                 .maxlen         = sizeof(int),
1610                 .mode           = 0644,
1611                 .proc_handler   = proc_dointvec_minmax,
1612                 .extra1         = &sysctl_nr_open_min,
1613                 .extra2         = &sysctl_nr_open_max,
1614         },
1615         {
1616                 .procname       = "dentry-state",
1617                 .data           = &dentry_stat,
1618                 .maxlen         = 6*sizeof(long),
1619                 .mode           = 0444,
1620                 .proc_handler   = proc_nr_dentry,
1621         },
1622         {
1623                 .procname       = "overflowuid",
1624                 .data           = &fs_overflowuid,
1625                 .maxlen         = sizeof(int),
1626                 .mode           = 0644,
1627                 .proc_handler   = proc_dointvec_minmax,
1628                 .extra1         = &minolduid,
1629                 .extra2         = &maxolduid,
1630         },
1631         {
1632                 .procname       = "overflowgid",
1633                 .data           = &fs_overflowgid,
1634                 .maxlen         = sizeof(int),
1635                 .mode           = 0644,
1636                 .proc_handler   = proc_dointvec_minmax,
1637                 .extra1         = &minolduid,
1638                 .extra2         = &maxolduid,
1639         },
1640 #ifdef CONFIG_FILE_LOCKING
1641         {
1642                 .procname       = "leases-enable",
1643                 .data           = &leases_enable,
1644                 .maxlen         = sizeof(int),
1645                 .mode           = 0644,
1646                 .proc_handler   = proc_dointvec,
1647         },
1648 #endif
1649 #ifdef CONFIG_DNOTIFY
1650         {
1651                 .procname       = "dir-notify-enable",
1652                 .data           = &dir_notify_enable,
1653                 .maxlen         = sizeof(int),
1654                 .mode           = 0644,
1655                 .proc_handler   = proc_dointvec,
1656         },
1657 #endif
1658 #ifdef CONFIG_MMU
1659 #ifdef CONFIG_FILE_LOCKING
1660         {
1661                 .procname       = "lease-break-time",
1662                 .data           = &lease_break_time,
1663                 .maxlen         = sizeof(int),
1664                 .mode           = 0644,
1665                 .proc_handler   = proc_dointvec,
1666         },
1667 #endif
1668 #ifdef CONFIG_AIO
1669         {
1670                 .procname       = "aio-nr",
1671                 .data           = &aio_nr,
1672                 .maxlen         = sizeof(aio_nr),
1673                 .mode           = 0444,
1674                 .proc_handler   = proc_doulongvec_minmax,
1675         },
1676         {
1677                 .procname       = "aio-max-nr",
1678                 .data           = &aio_max_nr,
1679                 .maxlen         = sizeof(aio_max_nr),
1680                 .mode           = 0644,
1681                 .proc_handler   = proc_doulongvec_minmax,
1682         },
1683 #endif /* CONFIG_AIO */
1684 #ifdef CONFIG_INOTIFY_USER
1685         {
1686                 .procname       = "inotify",
1687                 .mode           = 0555,
1688                 .child          = inotify_table,
1689         },
1690 #endif
1691 #ifdef CONFIG_EPOLL
1692         {
1693                 .procname       = "epoll",
1694                 .mode           = 0555,
1695                 .child          = epoll_table,
1696         },
1697 #endif
1698 #endif
1699         {
1700                 .procname       = "protected_symlinks",
1701                 .data           = &sysctl_protected_symlinks,
1702                 .maxlen         = sizeof(int),
1703                 .mode           = 0600,
1704                 .proc_handler   = proc_dointvec_minmax,
1705                 .extra1         = &zero,
1706                 .extra2         = &one,
1707         },
1708         {
1709                 .procname       = "protected_hardlinks",
1710                 .data           = &sysctl_protected_hardlinks,
1711                 .maxlen         = sizeof(int),
1712                 .mode           = 0600,
1713                 .proc_handler   = proc_dointvec_minmax,
1714                 .extra1         = &zero,
1715                 .extra2         = &one,
1716         },
1717         {
1718                 .procname       = "suid_dumpable",
1719                 .data           = &suid_dumpable,
1720                 .maxlen         = sizeof(int),
1721                 .mode           = 0644,
1722                 .proc_handler   = proc_dointvec_minmax_coredump,
1723                 .extra1         = &zero,
1724                 .extra2         = &two,
1725         },
1726 #if defined(CONFIG_BINFMT_MISC) || defined(CONFIG_BINFMT_MISC_MODULE)
1727         {
1728                 .procname       = "binfmt_misc",
1729                 .mode           = 0555,
1730                 .child          = sysctl_mount_point,
1731         },
1732 #endif
1733         {
1734                 .procname       = "pipe-max-size",
1735                 .data           = &pipe_max_size,
1736                 .maxlen         = sizeof(int),
1737                 .mode           = 0644,
1738                 .proc_handler   = &pipe_proc_fn,
1739                 .extra1         = &pipe_min_size,
1740         },
1741         {
1742                 .procname       = "pipe-user-pages-hard",
1743                 .data           = &pipe_user_pages_hard,
1744                 .maxlen         = sizeof(pipe_user_pages_hard),
1745                 .mode           = 0644,
1746                 .proc_handler   = proc_doulongvec_minmax,
1747         },
1748         {
1749                 .procname       = "pipe-user-pages-soft",
1750                 .data           = &pipe_user_pages_soft,
1751                 .maxlen         = sizeof(pipe_user_pages_soft),
1752                 .mode           = 0644,
1753                 .proc_handler   = proc_doulongvec_minmax,
1754         },
1755         {
1756                 .procname       = "mount-max",
1757                 .data           = &sysctl_mount_max,
1758                 .maxlen         = sizeof(unsigned int),
1759                 .mode           = 0644,
1760                 .proc_handler   = proc_dointvec_minmax,
1761                 .extra1         = &one,
1762         },
1763         { }
1764 };
1765
1766 static struct ctl_table debug_table[] = {
1767 #ifdef CONFIG_SYSCTL_EXCEPTION_TRACE
1768         {
1769                 .procname       = "exception-trace",
1770                 .data           = &show_unhandled_signals,
1771                 .maxlen         = sizeof(int),
1772                 .mode           = 0644,
1773                 .proc_handler   = proc_dointvec
1774         },
1775 #endif
1776 #if defined(CONFIG_OPTPROBES)
1777         {
1778                 .procname       = "kprobes-optimization",
1779                 .data           = &sysctl_kprobes_optimization,
1780                 .maxlen         = sizeof(int),
1781                 .mode           = 0644,
1782                 .proc_handler   = proc_kprobes_optimization_handler,
1783                 .extra1         = &zero,
1784                 .extra2         = &one,
1785         },
1786 #endif
1787         { }
1788 };
1789
1790 static struct ctl_table dev_table[] = {
1791         { }
1792 };
1793
1794 int __init sysctl_init(void)
1795 {
1796         struct ctl_table_header *hdr;
1797
1798         hdr = register_sysctl_table(sysctl_base_table);
1799         kmemleak_not_leak(hdr);
1800         return 0;
1801 }
1802
1803 #endif /* CONFIG_SYSCTL */
1804
1805 /*
1806  * /proc/sys support
1807  */
1808
1809 #ifdef CONFIG_PROC_SYSCTL
1810
1811 static int _proc_do_string(char *data, int maxlen, int write,
1812                            char __user *buffer,
1813                            size_t *lenp, loff_t *ppos)
1814 {
1815         size_t len;
1816         char __user *p;
1817         char c;
1818
1819         if (!data || !maxlen || !*lenp) {
1820                 *lenp = 0;
1821                 return 0;
1822         }
1823
1824         if (write) {
1825                 if (sysctl_writes_strict == SYSCTL_WRITES_STRICT) {
1826                         /* Only continue writes not past the end of buffer. */
1827                         len = strlen(data);
1828                         if (len > maxlen - 1)
1829                                 len = maxlen - 1;
1830
1831                         if (*ppos > len)
1832                                 return 0;
1833                         len = *ppos;
1834                 } else {
1835                         /* Start writing from beginning of buffer. */
1836                         len = 0;
1837                 }
1838
1839                 *ppos += *lenp;
1840                 p = buffer;
1841                 while ((p - buffer) < *lenp && len < maxlen - 1) {
1842                         if (get_user(c, p++))
1843                                 return -EFAULT;
1844                         if (c == 0 || c == '\n')
1845                                 break;
1846                         data[len++] = c;
1847                 }
1848                 data[len] = 0;
1849         } else {
1850                 len = strlen(data);
1851                 if (len > maxlen)
1852                         len = maxlen;
1853
1854                 if (*ppos > len) {
1855                         *lenp = 0;
1856                         return 0;
1857                 }
1858
1859                 data += *ppos;
1860                 len  -= *ppos;
1861
1862                 if (len > *lenp)
1863                         len = *lenp;
1864                 if (len)
1865                         if (copy_to_user(buffer, data, len))
1866                                 return -EFAULT;
1867                 if (len < *lenp) {
1868                         if (put_user('\n', buffer + len))
1869                                 return -EFAULT;
1870                         len++;
1871                 }
1872                 *lenp = len;
1873                 *ppos += len;
1874         }
1875         return 0;
1876 }
1877
1878 static void warn_sysctl_write(struct ctl_table *table)
1879 {
1880         pr_warn_once("%s wrote to %s when file position was not 0!\n"
1881                 "This will not be supported in the future. To silence this\n"
1882                 "warning, set kernel.sysctl_writes_strict = -1\n",
1883                 current->comm, table->procname);
1884 }
1885
1886 /**
1887  * proc_dostring - read a string sysctl
1888  * @table: the sysctl table
1889  * @write: %TRUE if this is a write to the sysctl file
1890  * @buffer: the user buffer
1891  * @lenp: the size of the user buffer
1892  * @ppos: file position
1893  *
1894  * Reads/writes a string from/to the user buffer. If the kernel
1895  * buffer provided is not large enough to hold the string, the
1896  * string is truncated. The copied string is %NULL-terminated.
1897  * If the string is being read by the user process, it is copied
1898  * and a newline '\n' is added. It is truncated if the buffer is
1899  * not large enough.
1900  *
1901  * Returns 0 on success.
1902  */
1903 int proc_dostring(struct ctl_table *table, int write,
1904                   void __user *buffer, size_t *lenp, loff_t *ppos)
1905 {
1906         if (write && *ppos && sysctl_writes_strict == SYSCTL_WRITES_WARN)
1907                 warn_sysctl_write(table);
1908
1909         return _proc_do_string((char *)(table->data), table->maxlen, write,
1910                                (char __user *)buffer, lenp, ppos);
1911 }
1912
1913 static size_t proc_skip_spaces(char **buf)
1914 {
1915         size_t ret;
1916         char *tmp = skip_spaces(*buf);
1917         ret = tmp - *buf;
1918         *buf = tmp;
1919         return ret;
1920 }
1921
1922 static void proc_skip_char(char **buf, size_t *size, const char v)
1923 {
1924         while (*size) {
1925                 if (**buf != v)
1926                         break;
1927                 (*size)--;
1928                 (*buf)++;
1929         }
1930 }
1931
1932 #define TMPBUFLEN 22
1933 /**
1934  * proc_get_long - reads an ASCII formatted integer from a user buffer
1935  *
1936  * @buf: a kernel buffer
1937  * @size: size of the kernel buffer
1938  * @val: this is where the number will be stored
1939  * @neg: set to %TRUE if number is negative
1940  * @perm_tr: a vector which contains the allowed trailers
1941  * @perm_tr_len: size of the perm_tr vector
1942  * @tr: pointer to store the trailer character
1943  *
1944  * In case of success %0 is returned and @buf and @size are updated with
1945  * the amount of bytes read. If @tr is non-NULL and a trailing
1946  * character exists (size is non-zero after returning from this
1947  * function), @tr is updated with the trailing character.
1948  */
1949 static int proc_get_long(char **buf, size_t *size,
1950                           unsigned long *val, bool *neg,
1951                           const char *perm_tr, unsigned perm_tr_len, char *tr)
1952 {
1953         int len;
1954         char *p, tmp[TMPBUFLEN];
1955
1956         if (!*size)
1957                 return -EINVAL;
1958
1959         len = *size;
1960         if (len > TMPBUFLEN - 1)
1961                 len = TMPBUFLEN - 1;
1962
1963         memcpy(tmp, *buf, len);
1964
1965         tmp[len] = 0;
1966         p = tmp;
1967         if (*p == '-' && *size > 1) {
1968                 *neg = true;
1969                 p++;
1970         } else
1971                 *neg = false;
1972         if (!isdigit(*p))
1973                 return -EINVAL;
1974
1975         *val = simple_strtoul(p, &p, 0);
1976
1977         len = p - tmp;
1978
1979         /* We don't know if the next char is whitespace thus we may accept
1980          * invalid integers (e.g. 1234...a) or two integers instead of one
1981          * (e.g. 123...1). So lets not allow such large numbers. */
1982         if (len == TMPBUFLEN - 1)
1983                 return -EINVAL;
1984
1985         if (len < *size && perm_tr_len && !memchr(perm_tr, *p, perm_tr_len))
1986                 return -EINVAL;
1987
1988         if (tr && (len < *size))
1989                 *tr = *p;
1990
1991         *buf += len;
1992         *size -= len;
1993
1994         return 0;
1995 }
1996
1997 /**
1998  * proc_put_long - converts an integer to a decimal ASCII formatted string
1999  *
2000  * @buf: the user buffer
2001  * @size: the size of the user buffer
2002  * @val: the integer to be converted
2003  * @neg: sign of the number, %TRUE for negative
2004  *
2005  * In case of success %0 is returned and @buf and @size are updated with
2006  * the amount of bytes written.
2007  */
2008 static int proc_put_long(void __user **buf, size_t *size, unsigned long val,
2009                           bool neg)
2010 {
2011         int len;
2012         char tmp[TMPBUFLEN], *p = tmp;
2013
2014         sprintf(p, "%s%lu", neg ? "-" : "", val);
2015         len = strlen(tmp);
2016         if (len > *size)
2017                 len = *size;
2018         if (copy_to_user(*buf, tmp, len))
2019                 return -EFAULT;
2020         *size -= len;
2021         *buf += len;
2022         return 0;
2023 }
2024 #undef TMPBUFLEN
2025
2026 static int proc_put_char(void __user **buf, size_t *size, char c)
2027 {
2028         if (*size) {
2029                 char __user **buffer = (char __user **)buf;
2030                 if (put_user(c, *buffer))
2031                         return -EFAULT;
2032                 (*size)--, (*buffer)++;
2033                 *buf = *buffer;
2034         }
2035         return 0;
2036 }
2037
2038 static int do_proc_dointvec_conv(bool *negp, unsigned long *lvalp,
2039                                  int *valp,
2040                                  int write, void *data)
2041 {
2042         if (write) {
2043                 if (*negp) {
2044                         if (*lvalp > (unsigned long) INT_MAX + 1)
2045                                 return -EINVAL;
2046                         *valp = -*lvalp;
2047                 } else {
2048                         if (*lvalp > (unsigned long) INT_MAX)
2049                                 return -EINVAL;
2050                         *valp = *lvalp;
2051                 }
2052         } else {
2053                 int val = *valp;
2054                 if (val < 0) {
2055                         *negp = true;
2056                         *lvalp = -(unsigned long)val;
2057                 } else {
2058                         *negp = false;
2059                         *lvalp = (unsigned long)val;
2060                 }
2061         }
2062         return 0;
2063 }
2064
2065 static int do_proc_douintvec_conv(bool *negp, unsigned long *lvalp,
2066                                  int *valp,
2067                                  int write, void *data)
2068 {
2069         if (write) {
2070                 if (*negp)
2071                         return -EINVAL;
2072                 if (*lvalp > UINT_MAX)
2073                         return -EINVAL;
2074                 *valp = *lvalp;
2075         } else {
2076                 unsigned int val = *valp;
2077                 *negp = false;
2078                 *lvalp = (unsigned long)val;
2079         }
2080         return 0;
2081 }
2082
2083 static const char proc_wspace_sep[] = { ' ', '\t', '\n' };
2084
2085 static int __do_proc_dointvec(void *tbl_data, struct ctl_table *table,
2086                   int write, void __user *buffer,
2087                   size_t *lenp, loff_t *ppos,
2088                   int (*conv)(bool *negp, unsigned long *lvalp, int *valp,
2089                               int write, void *data),
2090                   void *data)
2091 {
2092         int *i, vleft, first = 1, err = 0;
2093         unsigned long page = 0;
2094         size_t left;
2095         char *kbuf;
2096
2097         if (!tbl_data || !table->maxlen || !*lenp || (*ppos && !write)) {
2098                 *lenp = 0;
2099                 return 0;
2100         }
2101
2102         i = (int *) tbl_data;
2103         vleft = table->maxlen / sizeof(*i);
2104         left = *lenp;
2105
2106         if (!conv)
2107                 conv = do_proc_dointvec_conv;
2108
2109         if (write) {
2110                 if (*ppos) {
2111                         switch (sysctl_writes_strict) {
2112                         case SYSCTL_WRITES_STRICT:
2113                                 goto out;
2114                         case SYSCTL_WRITES_WARN:
2115                                 warn_sysctl_write(table);
2116                                 break;
2117                         default:
2118                                 break;
2119                         }
2120                 }
2121
2122                 if (left > PAGE_SIZE - 1)
2123                         left = PAGE_SIZE - 1;
2124                 page = __get_free_page(GFP_TEMPORARY);
2125                 kbuf = (char *) page;
2126                 if (!kbuf)
2127                         return -ENOMEM;
2128                 if (copy_from_user(kbuf, buffer, left)) {
2129                         err = -EFAULT;
2130                         goto free;
2131                 }
2132                 kbuf[left] = 0;
2133         }
2134
2135         for (; left && vleft--; i++, first=0) {
2136                 unsigned long lval;
2137                 bool neg;
2138
2139                 if (write) {
2140                         left -= proc_skip_spaces(&kbuf);
2141
2142                         if (!left)
2143                                 break;
2144                         err = proc_get_long(&kbuf, &left, &lval, &neg,
2145                                              proc_wspace_sep,
2146                                              sizeof(proc_wspace_sep), NULL);
2147                         if (err)
2148                                 break;
2149                         if (conv(&neg, &lval, i, 1, data)) {
2150                                 err = -EINVAL;
2151                                 break;
2152                         }
2153                 } else {
2154                         if (conv(&neg, &lval, i, 0, data)) {
2155                                 err = -EINVAL;
2156                                 break;
2157                         }
2158                         if (!first)
2159                                 err = proc_put_char(&buffer, &left, '\t');
2160                         if (err)
2161                                 break;
2162                         err = proc_put_long(&buffer, &left, lval, neg);
2163                         if (err)
2164                                 break;
2165                 }
2166         }
2167
2168         if (!write && !first && left && !err)
2169                 err = proc_put_char(&buffer, &left, '\n');
2170         if (write && !err && left)
2171                 left -= proc_skip_spaces(&kbuf);
2172 free:
2173         if (write) {
2174                 free_page(page);
2175                 if (first)
2176                         return err ? : -EINVAL;
2177         }
2178         *lenp -= left;
2179 out:
2180         *ppos += *lenp;
2181         return err;
2182 }
2183
2184 static int do_proc_dointvec(struct ctl_table *table, int write,
2185                   void __user *buffer, size_t *lenp, loff_t *ppos,
2186                   int (*conv)(bool *negp, unsigned long *lvalp, int *valp,
2187                               int write, void *data),
2188                   void *data)
2189 {
2190         return __do_proc_dointvec(table->data, table, write,
2191                         buffer, lenp, ppos, conv, data);
2192 }
2193
2194 /**
2195  * proc_dointvec - read a vector of integers
2196  * @table: the sysctl table
2197  * @write: %TRUE if this is a write to the sysctl file
2198  * @buffer: the user buffer
2199  * @lenp: the size of the user buffer
2200  * @ppos: file position
2201  *
2202  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2203  * values from/to the user buffer, treated as an ASCII string.
2204  *
2205  * Returns 0 on success.
2206  */
2207 int proc_dointvec(struct ctl_table *table, int write,
2208                      void __user *buffer, size_t *lenp, loff_t *ppos)
2209 {
2210         return do_proc_dointvec(table, write, buffer, lenp, ppos, NULL, NULL);
2211 }
2212
2213 /**
2214  * proc_douintvec - read a vector of unsigned integers
2215  * @table: the sysctl table
2216  * @write: %TRUE if this is a write to the sysctl file
2217  * @buffer: the user buffer
2218  * @lenp: the size of the user buffer
2219  * @ppos: file position
2220  *
2221  * Reads/writes up to table->maxlen/sizeof(unsigned int) unsigned integer
2222  * values from/to the user buffer, treated as an ASCII string.
2223  *
2224  * Returns 0 on success.
2225  */
2226 int proc_douintvec(struct ctl_table *table, int write,
2227                      void __user *buffer, size_t *lenp, loff_t *ppos)
2228 {
2229         return do_proc_dointvec(table, write, buffer, lenp, ppos,
2230                                 do_proc_douintvec_conv, NULL);
2231 }
2232
2233 /*
2234  * Taint values can only be increased
2235  * This means we can safely use a temporary.
2236  */
2237 static int proc_taint(struct ctl_table *table, int write,
2238                                void __user *buffer, size_t *lenp, loff_t *ppos)
2239 {
2240         struct ctl_table t;
2241         unsigned long tmptaint = get_taint();
2242         int err;
2243
2244         if (write && !capable(CAP_SYS_ADMIN))
2245                 return -EPERM;
2246
2247         t = *table;
2248         t.data = &tmptaint;
2249         err = proc_doulongvec_minmax(&t, write, buffer, lenp, ppos);
2250         if (err < 0)
2251                 return err;
2252
2253         if (write) {
2254                 /*
2255                  * Poor man's atomic or. Not worth adding a primitive
2256                  * to everyone's atomic.h for this
2257                  */
2258                 int i;
2259                 for (i = 0; i < BITS_PER_LONG && tmptaint >> i; i++) {
2260                         if ((tmptaint >> i) & 1)
2261                                 add_taint(i, LOCKDEP_STILL_OK);
2262                 }
2263         }
2264
2265         return err;
2266 }
2267
2268 #ifdef CONFIG_PRINTK
2269 static int proc_dointvec_minmax_sysadmin(struct ctl_table *table, int write,
2270                                 void __user *buffer, size_t *lenp, loff_t *ppos)
2271 {
2272         if (write && !capable(CAP_SYS_ADMIN))
2273                 return -EPERM;
2274
2275         return proc_dointvec_minmax(table, write, buffer, lenp, ppos);
2276 }
2277 #endif
2278
2279 struct do_proc_dointvec_minmax_conv_param {
2280         int *min;
2281         int *max;
2282 };
2283
2284 static int do_proc_dointvec_minmax_conv(bool *negp, unsigned long *lvalp,
2285                                         int *valp,
2286                                         int write, void *data)
2287 {
2288         struct do_proc_dointvec_minmax_conv_param *param = data;
2289         if (write) {
2290                 int val = *negp ? -*lvalp : *lvalp;
2291                 if ((param->min && *param->min > val) ||
2292                     (param->max && *param->max < val))
2293                         return -EINVAL;
2294                 *valp = val;
2295         } else {
2296                 int val = *valp;
2297                 if (val < 0) {
2298                         *negp = true;
2299                         *lvalp = -(unsigned long)val;
2300                 } else {
2301                         *negp = false;
2302                         *lvalp = (unsigned long)val;
2303                 }
2304         }
2305         return 0;
2306 }
2307
2308 /**
2309  * proc_dointvec_minmax - read a vector of integers with min/max values
2310  * @table: the sysctl table
2311  * @write: %TRUE if this is a write to the sysctl file
2312  * @buffer: the user buffer
2313  * @lenp: the size of the user buffer
2314  * @ppos: file position
2315  *
2316  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2317  * values from/to the user buffer, treated as an ASCII string.
2318  *
2319  * This routine will ensure the values are within the range specified by
2320  * table->extra1 (min) and table->extra2 (max).
2321  *
2322  * Returns 0 on success.
2323  */
2324 int proc_dointvec_minmax(struct ctl_table *table, int write,
2325                   void __user *buffer, size_t *lenp, loff_t *ppos)
2326 {
2327         struct do_proc_dointvec_minmax_conv_param param = {
2328                 .min = (int *) table->extra1,
2329                 .max = (int *) table->extra2,
2330         };
2331         return do_proc_dointvec(table, write, buffer, lenp, ppos,
2332                                 do_proc_dointvec_minmax_conv, &param);
2333 }
2334
2335 static void validate_coredump_safety(void)
2336 {
2337 #ifdef CONFIG_COREDUMP
2338         if (suid_dumpable == SUID_DUMP_ROOT &&
2339             core_pattern[0] != '/' && core_pattern[0] != '|') {
2340                 printk(KERN_WARNING "Unsafe core_pattern used with "\
2341                         "suid_dumpable=2. Pipe handler or fully qualified "\
2342                         "core dump path required.\n");
2343         }
2344 #endif
2345 }
2346
2347 static int proc_dointvec_minmax_coredump(struct ctl_table *table, int write,
2348                 void __user *buffer, size_t *lenp, loff_t *ppos)
2349 {
2350         int error = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
2351         if (!error)
2352                 validate_coredump_safety();
2353         return error;
2354 }
2355
2356 #ifdef CONFIG_COREDUMP
2357 static int proc_dostring_coredump(struct ctl_table *table, int write,
2358                   void __user *buffer, size_t *lenp, loff_t *ppos)
2359 {
2360         int error = proc_dostring(table, write, buffer, lenp, ppos);
2361         if (!error)
2362                 validate_coredump_safety();
2363         return error;
2364 }
2365 #endif
2366
2367 static int __do_proc_doulongvec_minmax(void *data, struct ctl_table *table, int write,
2368                                      void __user *buffer,
2369                                      size_t *lenp, loff_t *ppos,
2370                                      unsigned long convmul,
2371                                      unsigned long convdiv)
2372 {
2373         unsigned long *i, *min, *max;
2374         int vleft, first = 1, err = 0;
2375         unsigned long page = 0;
2376         size_t left;
2377         char *kbuf;
2378
2379         if (!data || !table->maxlen || !*lenp || (*ppos && !write)) {
2380                 *lenp = 0;
2381                 return 0;
2382         }
2383
2384         i = (unsigned long *) data;
2385         min = (unsigned long *) table->extra1;
2386         max = (unsigned long *) table->extra2;
2387         vleft = table->maxlen / sizeof(unsigned long);
2388         left = *lenp;
2389
2390         if (write) {
2391                 if (*ppos) {
2392                         switch (sysctl_writes_strict) {
2393                         case SYSCTL_WRITES_STRICT:
2394                                 goto out;
2395                         case SYSCTL_WRITES_WARN:
2396                                 warn_sysctl_write(table);
2397                                 break;
2398                         default:
2399                                 break;
2400                         }
2401                 }
2402
2403                 if (left > PAGE_SIZE - 1)
2404                         left = PAGE_SIZE - 1;
2405                 page = __get_free_page(GFP_TEMPORARY);
2406                 kbuf = (char *) page;
2407                 if (!kbuf)
2408                         return -ENOMEM;
2409                 if (copy_from_user(kbuf, buffer, left)) {
2410                         err = -EFAULT;
2411                         goto free;
2412                 }
2413                 kbuf[left] = 0;
2414         }
2415
2416         for (; left && vleft--; i++, first = 0) {
2417                 unsigned long val;
2418
2419                 if (write) {
2420                         bool neg;
2421
2422                         left -= proc_skip_spaces(&kbuf);
2423
2424                         err = proc_get_long(&kbuf, &left, &val, &neg,
2425                                              proc_wspace_sep,
2426                                              sizeof(proc_wspace_sep), NULL);
2427                         if (err)
2428                                 break;
2429                         if (neg)
2430                                 continue;
2431                         val = convmul * val / convdiv;
2432                         if ((min && val < *min) || (max && val > *max))
2433                                 continue;
2434                         *i = val;
2435                 } else {
2436                         val = convdiv * (*i) / convmul;
2437                         if (!first) {
2438                                 err = proc_put_char(&buffer, &left, '\t');
2439                                 if (err)
2440                                         break;
2441                         }
2442                         err = proc_put_long(&buffer, &left, val, false);
2443                         if (err)
2444                                 break;
2445                 }
2446         }
2447
2448         if (!write && !first && left && !err)
2449                 err = proc_put_char(&buffer, &left, '\n');
2450         if (write && !err)
2451                 left -= proc_skip_spaces(&kbuf);
2452 free:
2453         if (write) {
2454                 free_page(page);
2455                 if (first)
2456                         return err ? : -EINVAL;
2457         }
2458         *lenp -= left;
2459 out:
2460         *ppos += *lenp;
2461         return err;
2462 }
2463
2464 static int do_proc_doulongvec_minmax(struct ctl_table *table, int write,
2465                                      void __user *buffer,
2466                                      size_t *lenp, loff_t *ppos,
2467                                      unsigned long convmul,
2468                                      unsigned long convdiv)
2469 {
2470         return __do_proc_doulongvec_minmax(table->data, table, write,
2471                         buffer, lenp, ppos, convmul, convdiv);
2472 }
2473
2474 /**
2475  * proc_doulongvec_minmax - read a vector of long integers with min/max values
2476  * @table: the sysctl table
2477  * @write: %TRUE if this is a write to the sysctl file
2478  * @buffer: the user buffer
2479  * @lenp: the size of the user buffer
2480  * @ppos: file position
2481  *
2482  * Reads/writes up to table->maxlen/sizeof(unsigned long) unsigned long
2483  * values from/to the user buffer, treated as an ASCII string.
2484  *
2485  * This routine will ensure the values are within the range specified by
2486  * table->extra1 (min) and table->extra2 (max).
2487  *
2488  * Returns 0 on success.
2489  */
2490 int proc_doulongvec_minmax(struct ctl_table *table, int write,
2491                            void __user *buffer, size_t *lenp, loff_t *ppos)
2492 {
2493     return do_proc_doulongvec_minmax(table, write, buffer, lenp, ppos, 1l, 1l);
2494 }
2495
2496 /**
2497  * proc_doulongvec_ms_jiffies_minmax - read a vector of millisecond values with min/max values
2498  * @table: the sysctl table
2499  * @write: %TRUE if this is a write to the sysctl file
2500  * @buffer: the user buffer
2501  * @lenp: the size of the user buffer
2502  * @ppos: file position
2503  *
2504  * Reads/writes up to table->maxlen/sizeof(unsigned long) unsigned long
2505  * values from/to the user buffer, treated as an ASCII string. The values
2506  * are treated as milliseconds, and converted to jiffies when they are stored.
2507  *
2508  * This routine will ensure the values are within the range specified by
2509  * table->extra1 (min) and table->extra2 (max).
2510  *
2511  * Returns 0 on success.
2512  */
2513 int proc_doulongvec_ms_jiffies_minmax(struct ctl_table *table, int write,
2514                                       void __user *buffer,
2515                                       size_t *lenp, loff_t *ppos)
2516 {
2517     return do_proc_doulongvec_minmax(table, write, buffer,
2518                                      lenp, ppos, HZ, 1000l);
2519 }
2520
2521
2522 static int do_proc_dointvec_jiffies_conv(bool *negp, unsigned long *lvalp,
2523                                          int *valp,
2524                                          int write, void *data)
2525 {
2526         if (write) {
2527                 if (*lvalp > LONG_MAX / HZ)
2528                         return 1;
2529                 *valp = *negp ? -(*lvalp*HZ) : (*lvalp*HZ);
2530         } else {
2531                 int val = *valp;
2532                 unsigned long lval;
2533                 if (val < 0) {
2534                         *negp = true;
2535                         lval = -(unsigned long)val;
2536                 } else {
2537                         *negp = false;
2538                         lval = (unsigned long)val;
2539                 }
2540                 *lvalp = lval / HZ;
2541         }
2542         return 0;
2543 }
2544
2545 static int do_proc_dointvec_userhz_jiffies_conv(bool *negp, unsigned long *lvalp,
2546                                                 int *valp,
2547                                                 int write, void *data)
2548 {
2549         if (write) {
2550                 if (USER_HZ < HZ && *lvalp > (LONG_MAX / HZ) * USER_HZ)
2551                         return 1;
2552                 *valp = clock_t_to_jiffies(*negp ? -*lvalp : *lvalp);
2553         } else {
2554                 int val = *valp;
2555                 unsigned long lval;
2556                 if (val < 0) {
2557                         *negp = true;
2558                         lval = -(unsigned long)val;
2559                 } else {
2560                         *negp = false;
2561                         lval = (unsigned long)val;
2562                 }
2563                 *lvalp = jiffies_to_clock_t(lval);
2564         }
2565         return 0;
2566 }
2567
2568 static int do_proc_dointvec_ms_jiffies_conv(bool *negp, unsigned long *lvalp,
2569                                             int *valp,
2570                                             int write, void *data)
2571 {
2572         if (write) {
2573                 unsigned long jif = msecs_to_jiffies(*negp ? -*lvalp : *lvalp);
2574
2575                 if (jif > INT_MAX)
2576                         return 1;
2577                 *valp = (int)jif;
2578         } else {
2579                 int val = *valp;
2580                 unsigned long lval;
2581                 if (val < 0) {
2582                         *negp = true;
2583                         lval = -(unsigned long)val;
2584                 } else {
2585                         *negp = false;
2586                         lval = (unsigned long)val;
2587                 }
2588                 *lvalp = jiffies_to_msecs(lval);
2589         }
2590         return 0;
2591 }
2592
2593 /**
2594  * proc_dointvec_jiffies - read a vector of integers as seconds
2595  * @table: the sysctl table
2596  * @write: %TRUE if this is a write to the sysctl file
2597  * @buffer: the user buffer
2598  * @lenp: the size of the user buffer
2599  * @ppos: file position
2600  *
2601  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2602  * values from/to the user buffer, treated as an ASCII string.
2603  * The values read are assumed to be in seconds, and are converted into
2604  * jiffies.
2605  *
2606  * Returns 0 on success.
2607  */
2608 int proc_dointvec_jiffies(struct ctl_table *table, int write,
2609                           void __user *buffer, size_t *lenp, loff_t *ppos)
2610 {
2611     return do_proc_dointvec(table,write,buffer,lenp,ppos,
2612                             do_proc_dointvec_jiffies_conv,NULL);
2613 }
2614
2615 /**
2616  * proc_dointvec_userhz_jiffies - read a vector of integers as 1/USER_HZ seconds
2617  * @table: the sysctl table
2618  * @write: %TRUE if this is a write to the sysctl file
2619  * @buffer: the user buffer
2620  * @lenp: the size of the user buffer
2621  * @ppos: pointer to the file position
2622  *
2623  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2624  * values from/to the user buffer, treated as an ASCII string.
2625  * The values read are assumed to be in 1/USER_HZ seconds, and
2626  * are converted into jiffies.
2627  *
2628  * Returns 0 on success.
2629  */
2630 int proc_dointvec_userhz_jiffies(struct ctl_table *table, int write,
2631                                  void __user *buffer, size_t *lenp, loff_t *ppos)
2632 {
2633     return do_proc_dointvec(table,write,buffer,lenp,ppos,
2634                             do_proc_dointvec_userhz_jiffies_conv,NULL);
2635 }
2636
2637 /**
2638  * proc_dointvec_ms_jiffies - read a vector of integers as 1 milliseconds
2639  * @table: the sysctl table
2640  * @write: %TRUE if this is a write to the sysctl file
2641  * @buffer: the user buffer
2642  * @lenp: the size of the user buffer
2643  * @ppos: file position
2644  * @ppos: the current position in the file
2645  *
2646  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2647  * values from/to the user buffer, treated as an ASCII string.
2648  * The values read are assumed to be in 1/1000 seconds, and
2649  * are converted into jiffies.
2650  *
2651  * Returns 0 on success.
2652  */
2653 int proc_dointvec_ms_jiffies(struct ctl_table *table, int write,
2654                              void __user *buffer, size_t *lenp, loff_t *ppos)
2655 {
2656         return do_proc_dointvec(table, write, buffer, lenp, ppos,
2657                                 do_proc_dointvec_ms_jiffies_conv, NULL);
2658 }
2659
2660 static int proc_do_cad_pid(struct ctl_table *table, int write,
2661                            void __user *buffer, size_t *lenp, loff_t *ppos)
2662 {
2663         struct pid *new_pid;
2664         pid_t tmp;
2665         int r;
2666
2667         tmp = pid_vnr(cad_pid);
2668
2669         r = __do_proc_dointvec(&tmp, table, write, buffer,
2670                                lenp, ppos, NULL, NULL);
2671         if (r || !write)
2672                 return r;
2673
2674         new_pid = find_get_pid(tmp);
2675         if (!new_pid)
2676                 return -ESRCH;
2677
2678         put_pid(xchg(&cad_pid, new_pid));
2679         return 0;
2680 }
2681
2682 /**
2683  * proc_do_large_bitmap - read/write from/to a large bitmap
2684  * @table: the sysctl table
2685  * @write: %TRUE if this is a write to the sysctl file
2686  * @buffer: the user buffer
2687  * @lenp: the size of the user buffer
2688  * @ppos: file position
2689  *
2690  * The bitmap is stored at table->data and the bitmap length (in bits)
2691  * in table->maxlen.
2692  *
2693  * We use a range comma separated format (e.g. 1,3-4,10-10) so that
2694  * large bitmaps may be represented in a compact manner. Writing into
2695  * the file will clear the bitmap then update it with the given input.
2696  *
2697  * Returns 0 on success.
2698  */
2699 int proc_do_large_bitmap(struct ctl_table *table, int write,
2700                          void __user *buffer, size_t *lenp, loff_t *ppos)
2701 {
2702         int err = 0;
2703         bool first = 1;
2704         size_t left = *lenp;
2705         unsigned long bitmap_len = table->maxlen;
2706         unsigned long *bitmap = *(unsigned long **) table->data;
2707         unsigned long *tmp_bitmap = NULL;
2708         char tr_a[] = { '-', ',', '\n' }, tr_b[] = { ',', '\n', 0 }, c;
2709
2710         if (!bitmap || !bitmap_len || !left || (*ppos && !write)) {
2711                 *lenp = 0;
2712                 return 0;
2713         }
2714
2715         if (write) {
2716                 unsigned long page = 0;
2717                 char *kbuf;
2718
2719                 if (left > PAGE_SIZE - 1)
2720                         left = PAGE_SIZE - 1;
2721
2722                 page = __get_free_page(GFP_TEMPORARY);
2723                 kbuf = (char *) page;
2724                 if (!kbuf)
2725                         return -ENOMEM;
2726                 if (copy_from_user(kbuf, buffer, left)) {
2727                         free_page(page);
2728                         return -EFAULT;
2729                 }
2730                 kbuf[left] = 0;
2731
2732                 tmp_bitmap = kzalloc(BITS_TO_LONGS(bitmap_len) * sizeof(unsigned long),
2733                                      GFP_KERNEL);
2734                 if (!tmp_bitmap) {
2735                         free_page(page);
2736                         return -ENOMEM;
2737                 }
2738                 proc_skip_char(&kbuf, &left, '\n');
2739                 while (!err && left) {
2740                         unsigned long val_a, val_b;
2741                         bool neg;
2742
2743                         err = proc_get_long(&kbuf, &left, &val_a, &neg, tr_a,
2744                                              sizeof(tr_a), &c);
2745                         if (err)
2746                                 break;
2747                         if (val_a >= bitmap_len || neg) {
2748                                 err = -EINVAL;
2749                                 break;
2750                         }
2751
2752                         val_b = val_a;
2753                         if (left) {
2754                                 kbuf++;
2755                                 left--;
2756                         }
2757
2758                         if (c == '-') {
2759                                 err = proc_get_long(&kbuf, &left, &val_b,
2760                                                      &neg, tr_b, sizeof(tr_b),
2761                                                      &c);
2762                                 if (err)
2763                                         break;
2764                                 if (val_b >= bitmap_len || neg ||
2765                                     val_a > val_b) {
2766                                         err = -EINVAL;
2767                                         break;
2768                                 }
2769                                 if (left) {
2770                                         kbuf++;
2771                                         left--;
2772                                 }
2773                         }
2774
2775                         bitmap_set(tmp_bitmap, val_a, val_b - val_a + 1);
2776                         first = 0;
2777                         proc_skip_char(&kbuf, &left, '\n');
2778                 }
2779                 free_page(page);
2780         } else {
2781                 unsigned long bit_a, bit_b = 0;
2782
2783                 while (left) {
2784                         bit_a = find_next_bit(bitmap, bitmap_len, bit_b);
2785                         if (bit_a >= bitmap_len)
2786                                 break;
2787                         bit_b = find_next_zero_bit(bitmap, bitmap_len,
2788                                                    bit_a + 1) - 1;
2789
2790                         if (!first) {
2791                                 err = proc_put_char(&buffer, &left, ',');
2792                                 if (err)
2793                                         break;
2794                         }
2795                         err = proc_put_long(&buffer, &left, bit_a, false);
2796                         if (err)
2797                                 break;
2798                         if (bit_a != bit_b) {
2799                                 err = proc_put_char(&buffer, &left, '-');
2800                                 if (err)
2801                                         break;
2802                                 err = proc_put_long(&buffer, &left, bit_b, false);
2803                                 if (err)
2804                                         break;
2805                         }
2806
2807                         first = 0; bit_b++;
2808                 }
2809                 if (!err)
2810                         err = proc_put_char(&buffer, &left, '\n');
2811         }
2812
2813         if (!err) {
2814                 if (write) {
2815                         if (*ppos)
2816                                 bitmap_or(bitmap, bitmap, tmp_bitmap, bitmap_len);
2817                         else
2818                                 bitmap_copy(bitmap, tmp_bitmap, bitmap_len);
2819                 }
2820                 kfree(tmp_bitmap);
2821                 *lenp -= left;
2822                 *ppos += *lenp;
2823                 return 0;
2824         } else {
2825                 kfree(tmp_bitmap);
2826                 return err;
2827         }
2828 }
2829
2830 #else /* CONFIG_PROC_SYSCTL */
2831
2832 int proc_dostring(struct ctl_table *table, int write,
2833                   void __user *buffer, size_t *lenp, loff_t *ppos)
2834 {
2835         return -ENOSYS;
2836 }
2837
2838 int proc_dointvec(struct ctl_table *table, int write,
2839                   void __user *buffer, size_t *lenp, loff_t *ppos)
2840 {
2841         return -ENOSYS;
2842 }
2843
2844 int proc_douintvec(struct ctl_table *table, int write,
2845                   void __user *buffer, size_t *lenp, loff_t *ppos)
2846 {
2847         return -ENOSYS;
2848 }
2849
2850 int proc_dointvec_minmax(struct ctl_table *table, int write,
2851                     void __user *buffer, size_t *lenp, loff_t *ppos)
2852 {
2853         return -ENOSYS;
2854 }
2855
2856 int proc_dointvec_jiffies(struct ctl_table *table, int write,
2857                     void __user *buffer, size_t *lenp, loff_t *ppos)
2858 {
2859         return -ENOSYS;
2860 }
2861
2862 int proc_dointvec_userhz_jiffies(struct ctl_table *table, int write,
2863                     void __user *buffer, size_t *lenp, loff_t *ppos)
2864 {
2865         return -ENOSYS;
2866 }
2867
2868 int proc_dointvec_ms_jiffies(struct ctl_table *table, int write,
2869                              void __user *buffer, size_t *lenp, loff_t *ppos)
2870 {
2871         return -ENOSYS;
2872 }
2873
2874 int proc_doulongvec_minmax(struct ctl_table *table, int write,
2875                     void __user *buffer, size_t *lenp, loff_t *ppos)
2876 {
2877         return -ENOSYS;
2878 }
2879
2880 int proc_doulongvec_ms_jiffies_minmax(struct ctl_table *table, int write,
2881                                       void __user *buffer,
2882                                       size_t *lenp, loff_t *ppos)
2883 {
2884     return -ENOSYS;
2885 }
2886
2887
2888 #endif /* CONFIG_PROC_SYSCTL */
2889
2890 /*
2891  * No sense putting this after each symbol definition, twice,
2892  * exception granted :-)
2893  */
2894 EXPORT_SYMBOL(proc_dointvec);
2895 EXPORT_SYMBOL(proc_douintvec);
2896 EXPORT_SYMBOL(proc_dointvec_jiffies);
2897 EXPORT_SYMBOL(proc_dointvec_minmax);
2898 EXPORT_SYMBOL(proc_dointvec_userhz_jiffies);
2899 EXPORT_SYMBOL(proc_dointvec_ms_jiffies);
2900 EXPORT_SYMBOL(proc_dostring);
2901 EXPORT_SYMBOL(proc_doulongvec_minmax);
2902 EXPORT_SYMBOL(proc_doulongvec_ms_jiffies_minmax);