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
  69 #include <asm/uaccess.h>
  70 #include <asm/processor.h>
  71
  72 #ifdef CONFIG_X86
  73 #include <asm/nmi.h>
  74 #include <asm/stacktrace.h>
  75 #include <asm/io.h>
  76 #endif
  77 #ifdef CONFIG_SPARC
  78 #include <asm/setup.h>
  79 #endif
  80 #ifdef CONFIG_BSD_PROCESS_ACCT
  81 #include <linux/acct.h>
  82 #endif
  83 #ifdef CONFIG_RT_MUTEXES
  84 #include <linux/rtmutex.h>
  85 #endif
  86 #if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_LOCK_STAT)
  87 #include <linux/lockdep.h>
  88 #endif
  89 #ifdef CONFIG_CHR_DEV_SG
  90 #include <scsi/sg.h>
  91 #endif
  92
  93 #ifdef CONFIG_LOCKUP_DETECTOR
  94 #include <linux/nmi.h>
  95 #endif
  96
  97 #if defined(CONFIG_SYSCTL)
  98
  99 /* External variables not in a header file. */
 100 extern int suid_dumpable;
 101 #ifdef CONFIG_COREDUMP
 102 extern int core_uses_pid;
 103 extern char core_pattern[];
 104 extern unsigned int core_pipe_limit;
 105 #endif
 106 extern int pid_max;
 107 extern int pid_max_min, pid_max_max;
 108 extern int percpu_pagelist_fraction;
 109 extern int compat_log;
 110 extern int latencytop_enabled;
 111 extern int sysctl_nr_open_min, sysctl_nr_open_max;
 112 #ifndef CONFIG_MMU
 113 extern int sysctl_nr_trim_pages;
 114 #endif
 115
 116 /* Constants used for minimum and  maximum */
 117 #ifdef CONFIG_LOCKUP_DETECTOR
 118 static int sixty = 60;
 119 #endif
 120
 121 static int __maybe_unused neg_one = -1;
 122
 123 static int zero;
 124 static int __maybe_unused one = 1;
 125 static int __maybe_unused two = 2;
 126 static int __maybe_unused four = 4;
 127 static unsigned long one_ul = 1;
 128 static int one_hundred = 100;
 129 #ifdef CONFIG_PRINTK
 130 static int ten_thousand = 10000;
 131 #endif
 132
 133 /* this is needed for the proc_doulongvec_minmax of vm_dirty_bytes */
 134 static unsigned long dirty_bytes_min = 2 * PAGE_SIZE;
 135
 136 /* this is needed for the proc_dointvec_minmax for [fs_]overflow UID and GID */
 137 static int maxolduid = 65535;
 138 static int minolduid;
 139
 140 static int ngroups_max = NGROUPS_MAX;
 141 static const int cap_last_cap = CAP_LAST_CAP;
 142
 143 /*this is needed for proc_doulongvec_minmax of sysctl_hung_task_timeout_secs */
 144 #ifdef CONFIG_DETECT_HUNG_TASK
 145 static unsigned long hung_task_timeout_max = (LONG_MAX/HZ);
 146 #endif
 147
 148 #ifdef CONFIG_INOTIFY_USER
 149 #include <linux/inotify.h>
 150 #endif
 151 #ifdef CONFIG_SPARC
 152 #endif
 153
 154 #ifdef __hppa__
 155 extern int pwrsw_enabled;
 156 #endif
 157
 158 #ifdef CONFIG_SYSCTL_ARCH_UNALIGN_ALLOW
 159 extern int unaligned_enabled;
 160 #endif
 161
 162 #ifdef CONFIG_IA64
 163 extern int unaligned_dump_stack;
 164 #endif
 165
 166 #ifdef CONFIG_SYSCTL_ARCH_UNALIGN_NO_WARN
 167 extern int no_unaligned_warning;
 168 #endif
 169
 170 #ifdef CONFIG_PROC_SYSCTL
 171
 172 #define SYSCTL_WRITES_LEGACY    -1
 173 #define SYSCTL_WRITES_WARN       0
 174 #define SYSCTL_WRITES_STRICT     1
 175
 176 static int sysctl_writes_strict = SYSCTL_WRITES_STRICT;
 177
 178 static int proc_do_cad_pid(struct ctl_table *table, int write,
 179                   void __user *buffer, size_t *lenp, loff_t *ppos);
 180 static int proc_taint(struct ctl_table *table, int write,
 181                                void __user *buffer, size_t *lenp, loff_t *ppos);
 182 #endif
 183
 184 #ifdef CONFIG_PRINTK
 185 static int proc_dointvec_minmax_sysadmin(struct ctl_table *table, int write,
 186                                 void __user *buffer, size_t *lenp, loff_t *ppos);
 187 #endif
 188
 189 static int proc_dointvec_minmax_coredump(struct ctl_table *table, int write,
 190                 void __user *buffer, size_t *lenp, loff_t *ppos);
 191 #ifdef CONFIG_COREDUMP
 192 static int proc_dostring_coredump(struct ctl_table *table, int write,
 193                 void __user *buffer, size_t *lenp, loff_t *ppos);
 194 #endif
 195
 196 #ifdef CONFIG_MAGIC_SYSRQ
 197 /* Note: sysrq code uses it's own private copy */
 198 static int __sysrq_enabled = CONFIG_MAGIC_SYSRQ_DEFAULT_ENABLE;
 199
 200 static int sysrq_sysctl_handler(struct ctl_table *table, int write,
 201                                 void __user *buffer, size_t *lenp,
 202                                 loff_t *ppos)
 203 {
 204         int error;
 205
 206         error = proc_dointvec(table, write, buffer, lenp, ppos);
 207         if (error)
 208                 return error;
 209
 210         if (write)
 211                 sysrq_toggle_support(__sysrq_enabled);
 212
 213         return 0;
 214 }
 215
 216 #endif
 217
 218 static struct ctl_table kern_table[];
 219 static struct ctl_table vm_table[];
 220 static struct ctl_table fs_table[];
 221 static struct ctl_table debug_table[];
 222 static struct ctl_table dev_table[];
 223 extern struct ctl_table random_table[];
 224 #ifdef CONFIG_EPOLL
 225 extern struct ctl_table epoll_table[];
 226 #endif
 227
 228 #ifdef HAVE_ARCH_PICK_MMAP_LAYOUT
 229 int sysctl_legacy_va_layout;
 230 #endif
 231
 232 /* The default sysctl tables: */
 233
 234 static struct ctl_table sysctl_base_table[] = {
 235         {
 236                 .procname       = "kernel",
 237                 .mode           = 0555,
 238                 .child          = kern_table,
 239         },
 240         {
 241                 .procname       = "vm",
 242                 .mode           = 0555,
 243                 .child          = vm_table,
 244         },
 245         {
 246                 .procname       = "fs",
 247                 .mode           = 0555,
 248                 .child          = fs_table,
 249         },
 250         {
 251                 .procname       = "debug",
 252                 .mode           = 0555,
 253                 .child          = debug_table,
 254         },
 255         {
 256                 .procname       = "dev",
 257                 .mode           = 0555,
 258                 .child          = dev_table,
 259         },
 260         { }
 261 };
 262
 263 #ifdef CONFIG_SCHED_DEBUG
 264 static int min_sched_granularity_ns = 100000;           /* 100 usecs */
 265 static int max_sched_granularity_ns = NSEC_PER_SEC;     /* 1 second */
 266 static int min_wakeup_granularity_ns;                   /* 0 usecs */
 267 static int max_wakeup_granularity_ns = NSEC_PER_SEC;    /* 1 second */
 268 #ifdef CONFIG_SMP
 269 static int min_sched_tunable_scaling = SCHED_TUNABLESCALING_NONE;
 270 static int max_sched_tunable_scaling = SCHED_TUNABLESCALING_END-1;
 271 #endif /* CONFIG_SMP */
 272 #endif /* CONFIG_SCHED_DEBUG */
 273
 274 #ifdef CONFIG_COMPACTION
 275 static int min_extfrag_threshold;
 276 static int max_extfrag_threshold = 1000;
 277 #endif
 278
 279 static struct ctl_table kern_table[] = {
 280         {
 281                 .procname       = "sched_child_runs_first",
 282                 .data           = &sysctl_sched_child_runs_first,
 283                 .maxlen         = sizeof(unsigned int),
 284                 .mode           = 0644,
 285                 .proc_handler   = proc_dointvec,
 286         },
 287 #ifdef CONFIG_SCHED_DEBUG
 288         {
 289                 .procname       = "sched_min_granularity_ns",
 290                 .data           = &sysctl_sched_min_granularity,
 291                 .maxlen         = sizeof(unsigned int),
 292                 .mode           = 0644,
 293                 .proc_handler   = sched_proc_update_handler,
 294                 .extra1         = &min_sched_granularity_ns,
 295                 .extra2         = &max_sched_granularity_ns,
 296         },
 297         {
 298                 .procname       = "sched_latency_ns",
 299                 .data           = &sysctl_sched_latency,
 300                 .maxlen         = sizeof(unsigned int),
 301                 .mode           = 0644,
 302                 .proc_handler   = sched_proc_update_handler,
 303                 .extra1         = &min_sched_granularity_ns,
 304                 .extra2         = &max_sched_granularity_ns,
 305         },
 306         {
 307                 .procname       = "sched_wakeup_granularity_ns",
 308                 .data           = &sysctl_sched_wakeup_granularity,
 309                 .maxlen         = sizeof(unsigned int),
 310                 .mode           = 0644,
 311                 .proc_handler   = sched_proc_update_handler,
 312                 .extra1         = &min_wakeup_granularity_ns,
 313                 .extra2         = &max_wakeup_granularity_ns,
 314         },
 315 #ifdef CONFIG_SMP
 316         {
 317                 .procname       = "sched_tunable_scaling",
 318                 .data           = &sysctl_sched_tunable_scaling,
 319                 .maxlen         = sizeof(enum sched_tunable_scaling),
 320                 .mode           = 0644,
 321                 .proc_handler   = sched_proc_update_handler,
 322                 .extra1         = &min_sched_tunable_scaling,
 323                 .extra2         = &max_sched_tunable_scaling,
 324         },
 325         {
 326                 .procname       = "sched_migration_cost_ns",
 327                 .data           = &sysctl_sched_migration_cost,
 328                 .maxlen         = sizeof(unsigned int),
 329                 .mode           = 0644,
 330                 .proc_handler   = proc_dointvec,
 331         },
 332         {
 333                 .procname       = "sched_nr_migrate",
 334                 .data           = &sysctl_sched_nr_migrate,
 335                 .maxlen         = sizeof(unsigned int),
 336                 .mode           = 0644,
 337                 .proc_handler   = proc_dointvec,
 338         },
 339         {
 340                 .procname       = "sched_time_avg_ms",
 341                 .data           = &sysctl_sched_time_avg,
 342                 .maxlen         = sizeof(unsigned int),
 343                 .mode           = 0644,
 344                 .proc_handler   = proc_dointvec,
 345         },
 346         {
 347                 .procname       = "sched_shares_window_ns",
 348                 .data           = &sysctl_sched_shares_window,
 349                 .maxlen         = sizeof(unsigned int),
 350                 .mode           = 0644,
 351                 .proc_handler   = proc_dointvec,
 352         },
 353 #ifdef CONFIG_SCHEDSTATS
 354         {
 355                 .procname       = "sched_schedstats",
 356                 .data           = NULL,
 357                 .maxlen         = sizeof(unsigned int),
 358                 .mode           = 0644,
 359                 .proc_handler   = sysctl_schedstats,
 360                 .extra1         = &zero,
 361                 .extra2         = &one,
 362         },
 363 #endif /* CONFIG_SCHEDSTATS */
 364 #endif /* CONFIG_SMP */
 365 #ifdef CONFIG_NUMA_BALANCING
 366         {
 367                 .procname       = "numa_balancing_scan_delay_ms",
 368                 .data           = &sysctl_numa_balancing_scan_delay,
 369                 .maxlen         = sizeof(unsigned int),
 370                 .mode           = 0644,
 371                 .proc_handler   = proc_dointvec,
 372         },
 373         {
 374                 .procname       = "numa_balancing_scan_period_min_ms",
 375                 .data           = &sysctl_numa_balancing_scan_period_min,
 376                 .maxlen         = sizeof(unsigned int),
 377                 .mode           = 0644,
 378                 .proc_handler   = proc_dointvec,
 379         },
 380         {
 381                 .procname       = "numa_balancing_scan_period_max_ms",
 382                 .data           = &sysctl_numa_balancing_scan_period_max,
 383                 .maxlen         = sizeof(unsigned int),
 384                 .mode           = 0644,
 385                 .proc_handler   = proc_dointvec,
 386         },
 387         {
 388                 .procname       = "numa_balancing_scan_size_mb",
 389                 .data           = &sysctl_numa_balancing_scan_size,
 390                 .maxlen         = sizeof(unsigned int),
 391                 .mode           = 0644,
 392                 .proc_handler   = proc_dointvec_minmax,
 393                 .extra1         = &one,
 394         },
 395         {
 396                 .procname       = "numa_balancing",
 397                 .data           = NULL, /* filled in by handler */
 398                 .maxlen         = sizeof(unsigned int),
 399                 .mode           = 0644,
 400                 .proc_handler   = sysctl_numa_balancing,
 401                 .extra1         = &zero,
 402                 .extra2         = &one,
 403         },
 404 #endif /* CONFIG_NUMA_BALANCING */
 405 #endif /* CONFIG_SCHED_DEBUG */
 406         {
 407                 .procname       = "sched_rt_period_us",
 408                 .data           = &sysctl_sched_rt_period,
 409                 .maxlen         = sizeof(unsigned int),
 410                 .mode           = 0644,
 411                 .proc_handler   = sched_rt_handler,
 412         },
 413         {
 414                 .procname       = "sched_rt_runtime_us",
 415                 .data           = &sysctl_sched_rt_runtime,
 416                 .maxlen         = sizeof(int),
 417                 .mode           = 0644,
 418                 .proc_handler   = sched_rt_handler,
 419         },
 420         {
 421                 .procname       = "sched_rr_timeslice_ms",
 422                 .data           = &sched_rr_timeslice,
 423                 .maxlen         = sizeof(int),
 424                 .mode           = 0644,
 425                 .proc_handler   = sched_rr_handler,
 426         },
 427 #ifdef CONFIG_SCHED_AUTOGROUP
 428         {
 429                 .procname       = "sched_autogroup_enabled",
 430                 .data           = &sysctl_sched_autogroup_enabled,
 431                 .maxlen         = sizeof(unsigned int),
 432                 .mode           = 0644,
 433                 .proc_handler   = proc_dointvec_minmax,
 434                 .extra1         = &zero,
 435                 .extra2         = &one,
 436         },
 437 #endif
 438 #ifdef CONFIG_CFS_BANDWIDTH
 439         {
 440                 .procname       = "sched_cfs_bandwidth_slice_us",
 441                 .data           = &sysctl_sched_cfs_bandwidth_slice,
 442                 .maxlen         = sizeof(unsigned int),
 443                 .mode           = 0644,
 444                 .proc_handler   = proc_dointvec_minmax,
 445                 .extra1         = &one,
 446         },
 447 #endif
 448 #ifdef CONFIG_PROVE_LOCKING
 449         {
 450                 .procname       = "prove_locking",
 451                 .data           = &prove_locking,
 452                 .maxlen         = sizeof(int),
 453                 .mode           = 0644,
 454                 .proc_handler   = proc_dointvec,
 455         },
 456 #endif
 457 #ifdef CONFIG_LOCK_STAT
 458         {
 459                 .procname       = "lock_stat",
 460                 .data           = &lock_stat,
 461                 .maxlen         = sizeof(int),
 462                 .mode           = 0644,
 463                 .proc_handler   = proc_dointvec,
 464         },
 465 #endif
 466         {
 467                 .procname       = "panic",
 468                 .data           = &panic_timeout,
 469                 .maxlen         = sizeof(int),
 470                 .mode           = 0644,
 471                 .proc_handler   = proc_dointvec,
 472         },
 473 #ifdef CONFIG_COREDUMP
 474         {
 475                 .procname       = "core_uses_pid",
 476                 .data           = &core_uses_pid,
 477                 .maxlen         = sizeof(int),
 478                 .mode           = 0644,
 479                 .proc_handler   = proc_dointvec,
 480         },
 481         {
 482                 .procname       = "core_pattern",
 483                 .data           = core_pattern,
 484                 .maxlen         = CORENAME_MAX_SIZE,
 485                 .mode           = 0644,
 486                 .proc_handler   = proc_dostring_coredump,
 487         },
 488         {
 489                 .procname       = "core_pipe_limit",
 490                 .data           = &core_pipe_limit,
 491                 .maxlen         = sizeof(unsigned int),
 492                 .mode           = 0644,
 493                 .proc_handler   = proc_dointvec,
 494         },
 495 #endif
 496 #ifdef CONFIG_PROC_SYSCTL
 497         {
 498                 .procname       = "tainted",
 499                 .maxlen         = sizeof(long),
 500                 .mode           = 0644,
 501                 .proc_handler   = proc_taint,
 502         },
 503         {
 504                 .procname       = "sysctl_writes_strict",
 505                 .data           = &sysctl_writes_strict,
 506                 .maxlen         = sizeof(int),
 507                 .mode           = 0644,
 508                 .proc_handler   = proc_dointvec_minmax,
 509                 .extra1         = &neg_one,
 510                 .extra2         = &one,
 511         },
 512 #endif
 513 #ifdef CONFIG_LATENCYTOP
 514         {
 515                 .procname       = "latencytop",
 516                 .data           = &latencytop_enabled,
 517                 .maxlen         = sizeof(int),
 518                 .mode           = 0644,
 519                 .proc_handler   = sysctl_latencytop,
 520         },
 521 #endif
 522 #ifdef CONFIG_BLK_DEV_INITRD
 523         {
 524                 .procname       = "real-root-dev",
 525                 .data           = &real_root_dev,
 526                 .maxlen         = sizeof(int),
 527                 .mode           = 0644,
 528                 .proc_handler   = proc_dointvec,
 529         },
 530 #endif
 531         {
 532                 .procname       = "print-fatal-signals",
 533                 .data           = &print_fatal_signals,
 534                 .maxlen         = sizeof(int),
 535                 .mode           = 0644,
 536                 .proc_handler   = proc_dointvec,
 537         },
 538 #ifdef CONFIG_SPARC
 539         {
 540                 .procname       = "reboot-cmd",
 541                 .data           = reboot_command,
 542                 .maxlen         = 256,
 543                 .mode           = 0644,
 544                 .proc_handler   = proc_dostring,
 545         },
 546         {
 547                 .procname       = "stop-a",
 548                 .data           = &stop_a_enabled,
 549                 .maxlen         = sizeof (int),
 550                 .mode           = 0644,
 551                 .proc_handler   = proc_dointvec,
 552         },
 553         {
 554                 .procname       = "scons-poweroff",
 555                 .data           = &scons_pwroff,
 556                 .maxlen         = sizeof (int),
 557                 .mode           = 0644,
 558                 .proc_handler   = proc_dointvec,
 559         },
 560 #endif
 561 #ifdef CONFIG_SPARC64
 562         {
 563                 .procname       = "tsb-ratio",
 564                 .data           = &sysctl_tsb_ratio,
 565                 .maxlen         = sizeof (int),
 566                 .mode           = 0644,
 567                 .proc_handler   = proc_dointvec,
 568         },
 569 #endif
 570 #ifdef __hppa__
 571         {
 572                 .procname       = "soft-power",
 573                 .data           = &pwrsw_enabled,
 574                 .maxlen         = sizeof (int),
 575                 .mode           = 0644,
 576                 .proc_handler   = proc_dointvec,
 577         },
 578 #endif
 579 #ifdef CONFIG_SYSCTL_ARCH_UNALIGN_ALLOW
 580         {
 581                 .procname       = "unaligned-trap",
 582                 .data           = &unaligned_enabled,
 583                 .maxlen         = sizeof (int),
 584                 .mode           = 0644,
 585                 .proc_handler   = proc_dointvec,
 586         },
 587 #endif
 588         {
 589                 .procname       = "ctrl-alt-del",
 590                 .data           = &C_A_D,
 591                 .maxlen         = sizeof(int),
 592                 .mode           = 0644,
 593                 .proc_handler   = proc_dointvec,
 594         },
 595 #ifdef CONFIG_FUNCTION_TRACER
 596         {
 597                 .procname       = "ftrace_enabled",
 598                 .data           = &ftrace_enabled,
 599                 .maxlen         = sizeof(int),
 600                 .mode           = 0644,
 601                 .proc_handler   = ftrace_enable_sysctl,
 602         },
 603 #endif
 604 #ifdef CONFIG_STACK_TRACER
 605         {
 606                 .procname       = "stack_tracer_enabled",
 607                 .data           = &stack_tracer_enabled,
 608                 .maxlen         = sizeof(int),
 609                 .mode           = 0644,
 610                 .proc_handler   = stack_trace_sysctl,
 611         },
 612 #endif
 613 #ifdef CONFIG_TRACING
 614         {
 615                 .procname       = "ftrace_dump_on_oops",
 616                 .data           = &ftrace_dump_on_oops,
 617                 .maxlen         = sizeof(int),
 618                 .mode           = 0644,
 619                 .proc_handler   = proc_dointvec,
 620         },
 621         {
 622                 .procname       = "traceoff_on_warning",
 623                 .data           = &__disable_trace_on_warning,
 624                 .maxlen         = sizeof(__disable_trace_on_warning),
 625                 .mode           = 0644,
 626                 .proc_handler   = proc_dointvec,
 627         },
 628         {
 629                 .procname       = "tracepoint_printk",
 630                 .data           = &tracepoint_printk,
 631                 .maxlen         = sizeof(tracepoint_printk),
 632                 .mode           = 0644,
 633                 .proc_handler   = proc_dointvec,
 634         },
 635 #endif
 636 #ifdef CONFIG_KEXEC_CORE
 637         {
 638                 .procname       = "kexec_load_disabled",
 639                 .data           = &kexec_load_disabled,
 640                 .maxlen         = sizeof(int),
 641                 .mode           = 0644,
 642                 /* only handle a transition from default "0" to "1" */
 643                 .proc_handler   = proc_dointvec_minmax,
 644                 .extra1         = &one,
 645                 .extra2         = &one,
 646         },
 647 #endif
 648 #ifdef CONFIG_MODULES
 649         {
 650                 .procname       = "modprobe",
 651                 .data           = &modprobe_path,
 652                 .maxlen         = KMOD_PATH_LEN,
 653                 .mode           = 0644,
 654                 .proc_handler   = proc_dostring,
 655         },
 656         {
 657                 .procname       = "modules_disabled",
 658                 .data           = &modules_disabled,
 659                 .maxlen         = sizeof(int),
 660                 .mode           = 0644,
 661                 /* only handle a transition from default "0" to "1" */
 662                 .proc_handler   = proc_dointvec_minmax,
 663                 .extra1         = &one,
 664                 .extra2         = &one,
 665         },
 666 #endif
 667 #ifdef CONFIG_UEVENT_HELPER
 668         {
 669                 .procname       = "hotplug",
 670                 .data           = &uevent_helper,
 671                 .maxlen         = UEVENT_HELPER_PATH_LEN,
 672                 .mode           = 0644,
 673                 .proc_handler   = proc_dostring,
 674         },
 675 #endif
 676 #ifdef CONFIG_CHR_DEV_SG
 677         {
 678                 .procname       = "sg-big-buff",
 679                 .data           = &sg_big_buff,
 680                 .maxlen         = sizeof (int),
 681                 .mode           = 0444,
 682                 .proc_handler   = proc_dointvec,
 683         },
 684 #endif
 685 #ifdef CONFIG_BSD_PROCESS_ACCT
 686         {
 687                 .procname       = "acct",
 688                 .data           = &acct_parm,
 689                 .maxlen         = 3*sizeof(int),
 690                 .mode           = 0644,
 691                 .proc_handler   = proc_dointvec,
 692         },
 693 #endif
 694 #ifdef CONFIG_MAGIC_SYSRQ
 695         {
 696                 .procname       = "sysrq",
 697                 .data           = &__sysrq_enabled,
 698                 .maxlen         = sizeof (int),
 699                 .mode           = 0644,
 700                 .proc_handler   = sysrq_sysctl_handler,
 701         },
 702 #endif
 703 #ifdef CONFIG_PROC_SYSCTL
 704         {
 705                 .procname       = "cad_pid",
 706                 .data           = NULL,
 707                 .maxlen         = sizeof (int),
 708                 .mode           = 0600,
 709                 .proc_handler   = proc_do_cad_pid,
 710         },
 711 #endif
 712         {
 713                 .procname       = "threads-max",
 714                 .data           = NULL,
 715                 .maxlen         = sizeof(int),
 716                 .mode           = 0644,
 717                 .proc_handler   = sysctl_max_threads,
 718         },
 719         {
 720                 .procname       = "random",
 721                 .mode           = 0555,
 722                 .child          = random_table,
 723         },
 724         {
 725                 .procname       = "usermodehelper",
 726                 .mode           = 0555,
 727                 .child          = usermodehelper_table,
 728         },
 729         {
 730                 .procname       = "overflowuid",
 731                 .data           = &overflowuid,
 732                 .maxlen         = sizeof(int),
 733                 .mode           = 0644,
 734                 .proc_handler   = proc_dointvec_minmax,
 735                 .extra1         = &minolduid,
 736                 .extra2         = &maxolduid,
 737         },
 738         {
 739                 .procname       = "overflowgid",
 740                 .data           = &overflowgid,
 741                 .maxlen         = sizeof(int),
 742                 .mode           = 0644,
 743                 .proc_handler   = proc_dointvec_minmax,
 744                 .extra1         = &minolduid,
 745                 .extra2         = &maxolduid,
 746         },
 747 #ifdef CONFIG_S390
 748 #ifdef CONFIG_MATHEMU
 749         {
 750                 .procname       = "ieee_emulation_warnings",
 751                 .data           = &sysctl_ieee_emulation_warnings,
 752                 .maxlen         = sizeof(int),
 753                 .mode           = 0644,
 754                 .proc_handler   = proc_dointvec,
 755         },
 756 #endif
 757         {
 758                 .procname       = "userprocess_debug",
 759                 .data           = &show_unhandled_signals,
 760                 .maxlen         = sizeof(int),
 761                 .mode           = 0644,
 762                 .proc_handler   = proc_dointvec,
 763         },
 764 #endif
 765         {
 766                 .procname       = "pid_max",
 767                 .data           = &pid_max,
 768                 .maxlen         = sizeof (int),
 769                 .mode           = 0644,
 770                 .proc_handler   = proc_dointvec_minmax,
 771                 .extra1         = &pid_max_min,
 772                 .extra2         = &pid_max_max,
 773         },
 774         {
 775                 .procname       = "panic_on_oops",
 776                 .data           = &panic_on_oops,
 777                 .maxlen         = sizeof(int),
 778                 .mode           = 0644,
 779                 .proc_handler   = proc_dointvec,
 780         },
 781 #if defined CONFIG_PRINTK
 782         {
 783                 .procname       = "printk",
 784                 .data           = &console_loglevel,
 785                 .maxlen         = 4*sizeof(int),
 786                 .mode           = 0644,
 787                 .proc_handler   = proc_dointvec,
 788         },
 789         {
 790                 .procname       = "printk_ratelimit",
 791                 .data           = &printk_ratelimit_state.interval,
 792                 .maxlen         = sizeof(int),
 793                 .mode           = 0644,
 794                 .proc_handler   = proc_dointvec_jiffies,
 795         },
 796         {
 797                 .procname       = "printk_ratelimit_burst",
 798                 .data           = &printk_ratelimit_state.burst,
 799                 .maxlen         = sizeof(int),
 800                 .mode           = 0644,
 801                 .proc_handler   = proc_dointvec,
 802         },
 803         {
 804                 .procname       = "printk_delay",
 805                 .data           = &printk_delay_msec,
 806                 .maxlen         = sizeof(int),
 807                 .mode           = 0644,
 808                 .proc_handler   = proc_dointvec_minmax,
 809                 .extra1         = &zero,
 810                 .extra2         = &ten_thousand,
 811         },
 812         {
 813                 .procname       = "dmesg_restrict",
 814                 .data           = &dmesg_restrict,
 815                 .maxlen         = sizeof(int),
 816                 .mode           = 0644,
 817                 .proc_handler   = proc_dointvec_minmax_sysadmin,
 818                 .extra1         = &zero,
 819                 .extra2         = &one,
 820         },
 821         {
 822                 .procname       = "kptr_restrict",
 823                 .data           = &kptr_restrict,
 824                 .maxlen         = sizeof(int),
 825                 .mode           = 0644,
 826                 .proc_handler   = proc_dointvec_minmax_sysadmin,
 827                 .extra1         = &zero,
 828                 .extra2         = &two,
 829         },
 830 #endif
 831         {
 832                 .procname       = "ngroups_max",
 833                 .data           = &ngroups_max,
 834                 .maxlen         = sizeof (int),
 835                 .mode           = 0444,
 836                 .proc_handler   = proc_dointvec,
 837         },
 838         {
 839                 .procname       = "cap_last_cap",
 840                 .data           = (void *)&cap_last_cap,
 841                 .maxlen         = sizeof(int),
 842                 .mode           = 0444,
 843                 .proc_handler   = proc_dointvec,
 844         },
 845 #if defined(CONFIG_LOCKUP_DETECTOR)
 846         {
 847                 .procname       = "watchdog",
 848                 .data           = &watchdog_user_enabled,
 849                 .maxlen         = sizeof (int),
 850                 .mode           = 0644,
 851                 .proc_handler   = proc_watchdog,
 852                 .extra1         = &zero,
 853                 .extra2         = &one,
 854         },
 855         {
 856                 .procname       = "watchdog_thresh",
 857                 .data           = &watchdog_thresh,
 858                 .maxlen         = sizeof(int),
 859                 .mode           = 0644,
 860                 .proc_handler   = proc_watchdog_thresh,
 861                 .extra1         = &zero,
 862                 .extra2         = &sixty,
 863         },
 864         {
 865                 .procname       = "nmi_watchdog",
 866                 .data           = &nmi_watchdog_enabled,
 867                 .maxlen         = sizeof (int),
 868                 .mode           = 0644,
 869                 .proc_handler   = proc_nmi_watchdog,
 870                 .extra1         = &zero,
 871 #if defined(CONFIG_HAVE_NMI_WATCHDOG) || defined(CONFIG_HARDLOCKUP_DETECTOR)
 872                 .extra2         = &one,
 873 #else
 874                 .extra2         = &zero,
 875 #endif
 876         },
 877         {
 878                 .procname       = "soft_watchdog",
 879                 .data           = &soft_watchdog_enabled,
 880                 .maxlen         = sizeof (int),
 881                 .mode           = 0644,
 882                 .proc_handler   = proc_soft_watchdog,
 883                 .extra1         = &zero,
 884                 .extra2         = &one,
 885         },
 886         {
 887                 .procname       = "watchdog_cpumask",
 888                 .data           = &watchdog_cpumask_bits,
 889                 .maxlen         = NR_CPUS,
 890                 .mode           = 0644,
 891                 .proc_handler   = proc_watchdog_cpumask,
 892         },
 893         {
 894                 .procname       = "softlockup_panic",
 895                 .data           = &softlockup_panic,
 896                 .maxlen         = sizeof(int),
 897                 .mode           = 0644,
 898                 .proc_handler   = proc_dointvec_minmax,
 899                 .extra1         = &zero,
 900                 .extra2         = &one,
 901         },
 902 #ifdef CONFIG_HARDLOCKUP_DETECTOR
 903         {
 904                 .procname       = "hardlockup_panic",
 905                 .data           = &hardlockup_panic,
 906                 .maxlen         = sizeof(int),
 907                 .mode           = 0644,
 908                 .proc_handler   = proc_dointvec_minmax,
 909                 .extra1         = &zero,
 910                 .extra2         = &one,
 911         },
 912 #endif
 913 #ifdef CONFIG_SMP
 914         {
 915                 .procname       = "softlockup_all_cpu_backtrace",
 916                 .data           = &sysctl_softlockup_all_cpu_backtrace,
 917                 .maxlen         = sizeof(int),
 918                 .mode           = 0644,
 919                 .proc_handler   = proc_dointvec_minmax,
 920                 .extra1         = &zero,
 921                 .extra2         = &one,
 922         },
 923         {
 924                 .procname       = "hardlockup_all_cpu_backtrace",
 925                 .data           = &sysctl_hardlockup_all_cpu_backtrace,
 926                 .maxlen         = sizeof(int),
 927                 .mode           = 0644,
 928                 .proc_handler   = proc_dointvec_minmax,
 929                 .extra1         = &zero,
 930                 .extra2         = &one,
 931         },
 932 #endif /* CONFIG_SMP */
 933 #endif
 934 #if defined(CONFIG_X86_LOCAL_APIC) && defined(CONFIG_X86)
 935         {
 936                 .procname       = "unknown_nmi_panic",
 937                 .data           = &unknown_nmi_panic,
 938                 .maxlen         = sizeof (int),
 939                 .mode           = 0644,
 940                 .proc_handler   = proc_dointvec,
 941         },
 942 #endif
 943 #if defined(CONFIG_X86)
 944         {
 945                 .procname       = "panic_on_unrecovered_nmi",
 946                 .data           = &panic_on_unrecovered_nmi,
 947                 .maxlen         = sizeof(int),
 948                 .mode           = 0644,
 949                 .proc_handler   = proc_dointvec,
 950         },
 951         {
 952                 .procname       = "panic_on_io_nmi",
 953                 .data           = &panic_on_io_nmi,
 954                 .maxlen         = sizeof(int),
 955                 .mode           = 0644,
 956                 .proc_handler   = proc_dointvec,
 957         },
 958 #ifdef CONFIG_DEBUG_STACKOVERFLOW
 959         {
 960                 .procname       = "panic_on_stackoverflow",
 961                 .data           = &sysctl_panic_on_stackoverflow,
 962                 .maxlen         = sizeof(int),
 963                 .mode           = 0644,
 964                 .proc_handler   = proc_dointvec,
 965         },
 966 #endif
 967         {
 968                 .procname       = "bootloader_type",
 969                 .data           = &bootloader_type,
 970                 .maxlen         = sizeof (int),
 971                 .mode           = 0444,
 972                 .proc_handler   = proc_dointvec,
 973         },
 974         {
 975                 .procname       = "bootloader_version",
 976                 .data           = &bootloader_version,
 977                 .maxlen         = sizeof (int),
 978                 .mode           = 0444,
 979                 .proc_handler   = proc_dointvec,
 980         },
 981         {
 982                 .procname       = "kstack_depth_to_print",
 983                 .data           = &kstack_depth_to_print,
 984                 .maxlen         = sizeof(int),
 985                 .mode           = 0644,
 986                 .proc_handler   = proc_dointvec,
 987         },
 988         {
 989                 .procname       = "io_delay_type",
 990                 .data           = &io_delay_type,
 991                 .maxlen         = sizeof(int),
 992                 .mode           = 0644,
 993                 .proc_handler   = proc_dointvec,
 994         },
 995 #endif
 996 #if defined(CONFIG_MMU)
 997         {
 998                 .procname       = "randomize_va_space",
 999                 .data           = &randomize_va_space,
1000                 .maxlen         = sizeof(int),
1001                 .mode           = 0644,
1002                 .proc_handler   = proc_dointvec,
1003         },
1004 #endif
1005 #if defined(CONFIG_S390) && defined(CONFIG_SMP)
1006         {
1007                 .procname       = "spin_retry",
1008                 .data           = &spin_retry,
1009                 .maxlen         = sizeof (int),
1010                 .mode           = 0644,
1011                 .proc_handler   = proc_dointvec,
1012         },
1013 #endif
1014 #if     defined(CONFIG_ACPI_SLEEP) && defined(CONFIG_X86)
1015         {
1016                 .procname       = "acpi_video_flags",
1017                 .data           = &acpi_realmode_flags,
1018                 .maxlen         = sizeof (unsigned long),
1019                 .mode           = 0644,
1020                 .proc_handler   = proc_doulongvec_minmax,
1021         },
1022 #endif
1023 #ifdef CONFIG_SYSCTL_ARCH_UNALIGN_NO_WARN
1024         {
1025                 .procname       = "ignore-unaligned-usertrap",
1026                 .data           = &no_unaligned_warning,
1027                 .maxlen         = sizeof (int),
1028                 .mode           = 0644,
1029                 .proc_handler   = proc_dointvec,
1030         },
1031 #endif
1032 #ifdef CONFIG_IA64
1033         {
1034                 .procname       = "unaligned-dump-stack",
1035                 .data           = &unaligned_dump_stack,
1036                 .maxlen         = sizeof (int),
1037                 .mode           = 0644,
1038                 .proc_handler   = proc_dointvec,
1039         },
1040 #endif
1041 #ifdef CONFIG_DETECT_HUNG_TASK
1042         {
1043                 .procname       = "hung_task_panic",
1044                 .data           = &sysctl_hung_task_panic,
1045                 .maxlen         = sizeof(int),
1046                 .mode           = 0644,
1047                 .proc_handler   = proc_dointvec_minmax,
1048                 .extra1         = &zero,
1049                 .extra2         = &one,
1050         },
1051         {
1052                 .procname       = "hung_task_check_count",
1053                 .data           = &sysctl_hung_task_check_count,
1054                 .maxlen         = sizeof(int),
1055                 .mode           = 0644,
1056                 .proc_handler   = proc_dointvec_minmax,
1057                 .extra1         = &zero,
1058         },
1059         {
1060                 .procname       = "hung_task_timeout_secs",
1061                 .data           = &sysctl_hung_task_timeout_secs,
1062                 .maxlen         = sizeof(unsigned long),
1063                 .mode           = 0644,
1064                 .proc_handler   = proc_dohung_task_timeout_secs,
1065                 .extra2         = &hung_task_timeout_max,
1066         },
1067         {
1068                 .procname       = "hung_task_warnings",
1069                 .data           = &sysctl_hung_task_warnings,
1070                 .maxlen         = sizeof(int),
1071                 .mode           = 0644,
1072                 .proc_handler   = proc_dointvec_minmax,
1073                 .extra1         = &neg_one,
1074         },
1075 #endif
1076 #ifdef CONFIG_COMPAT
1077         {
1078                 .procname       = "compat-log",
1079                 .data           = &compat_log,
1080                 .maxlen         = sizeof (int),
1081                 .mode           = 0644,
1082                 .proc_handler   = proc_dointvec,
1083         },
1084 #endif
1085 #ifdef CONFIG_RT_MUTEXES
1086         {
1087                 .procname       = "max_lock_depth",
1088                 .data           = &max_lock_depth,
1089                 .maxlen         = sizeof(int),
1090                 .mode           = 0644,
1091                 .proc_handler   = proc_dointvec,
1092         },
1093 #endif
1094         {
1095                 .procname       = "poweroff_cmd",
1096                 .data           = &poweroff_cmd,
1097                 .maxlen         = POWEROFF_CMD_PATH_LEN,
1098                 .mode           = 0644,
1099                 .proc_handler   = proc_dostring,
1100         },
1101 #ifdef CONFIG_KEYS
1102         {
1103                 .procname       = "keys",
1104                 .mode           = 0555,
1105                 .child          = key_sysctls,
1106         },
1107 #endif
1108 #ifdef CONFIG_PERF_EVENTS
1109         /*
1110          * User-space scripts rely on the existence of this file
1111          * as a feature check for perf_events being enabled.
1112          *
1113          * So it's an ABI, do not remove!
1114          */
1115         {
1116                 .procname       = "perf_event_paranoid",
1117                 .data           = &sysctl_perf_event_paranoid,
1118                 .maxlen         = sizeof(sysctl_perf_event_paranoid),
1119                 .mode           = 0644,
1120                 .proc_handler   = proc_dointvec,
1121         },
1122         {
1123                 .procname       = "perf_event_mlock_kb",
1124                 .data           = &sysctl_perf_event_mlock,
1125                 .maxlen         = sizeof(sysctl_perf_event_mlock),
1126                 .mode           = 0644,
1127                 .proc_handler   = proc_dointvec,
1128         },
1129         {
1130                 .procname       = "perf_event_max_sample_rate",
1131                 .data           = &sysctl_perf_event_sample_rate,
1132                 .maxlen         = sizeof(sysctl_perf_event_sample_rate),
1133                 .mode           = 0644,
1134                 .proc_handler   = perf_proc_update_handler,
1135                 .extra1         = &one,
1136         },
1137         {
1138                 .procname       = "perf_cpu_time_max_percent",
1139                 .data           = &sysctl_perf_cpu_time_max_percent,
1140                 .maxlen         = sizeof(sysctl_perf_cpu_time_max_percent),
1141                 .mode           = 0644,
1142                 .proc_handler   = perf_cpu_time_max_percent_handler,
1143                 .extra1         = &zero,
1144                 .extra2         = &one_hundred,
1145         },
1146 #endif
1147 #ifdef CONFIG_KMEMCHECK
1148         {
1149                 .procname       = "kmemcheck",
1150                 .data           = &kmemcheck_enabled,
1151                 .maxlen         = sizeof(int),
1152                 .mode           = 0644,
1153                 .proc_handler   = proc_dointvec,
1154         },
1155 #endif
1156         {
1157                 .procname       = "panic_on_warn",
1158                 .data           = &panic_on_warn,
1159                 .maxlen         = sizeof(int),
1160                 .mode           = 0644,
1161                 .proc_handler   = proc_dointvec_minmax,
1162                 .extra1         = &zero,
1163                 .extra2         = &one,
1164         },
1165 #if defined(CONFIG_SMP) && defined(CONFIG_NO_HZ_COMMON)
1166         {
1167                 .procname       = "timer_migration",
1168                 .data           = &sysctl_timer_migration,
1169                 .maxlen         = sizeof(unsigned int),
1170                 .mode           = 0644,
1171                 .proc_handler   = timer_migration_handler,
1172         },
1173 #endif
1174 #ifdef CONFIG_BPF_SYSCALL
1175         {
1176                 .procname       = "unprivileged_bpf_disabled",
1177                 .data           = &sysctl_unprivileged_bpf_disabled,
1178                 .maxlen         = sizeof(sysctl_unprivileged_bpf_disabled),
1179                 .mode           = 0644,
1180                 /* only handle a transition from default "0" to "1" */
1181                 .proc_handler   = proc_dointvec_minmax,
1182                 .extra1         = &one,
1183                 .extra2         = &one,
1184         },
1185 #endif
1186         { }
1187 };
1188
1189 static struct ctl_table vm_table[] = {
1190         {
1191                 .procname       = "overcommit_memory",
1192                 .data           = &sysctl_overcommit_memory,
1193                 .maxlen         = sizeof(sysctl_overcommit_memory),
1194                 .mode           = 0644,
1195                 .proc_handler   = proc_dointvec_minmax,
1196                 .extra1         = &zero,
1197                 .extra2         = &two,
1198         },
1199         {
1200                 .procname       = "panic_on_oom",
1201                 .data           = &sysctl_panic_on_oom,
1202                 .maxlen         = sizeof(sysctl_panic_on_oom),
1203                 .mode           = 0644,
1204                 .proc_handler   = proc_dointvec_minmax,
1205                 .extra1         = &zero,
1206                 .extra2         = &two,
1207         },
1208         {
1209                 .procname       = "oom_kill_allocating_task",
1210                 .data           = &sysctl_oom_kill_allocating_task,
1211                 .maxlen         = sizeof(sysctl_oom_kill_allocating_task),
1212                 .mode           = 0644,
1213                 .proc_handler   = proc_dointvec,
1214         },
1215         {
1216                 .procname       = "oom_dump_tasks",
1217                 .data           = &sysctl_oom_dump_tasks,
1218                 .maxlen         = sizeof(sysctl_oom_dump_tasks),
1219                 .mode           = 0644,
1220                 .proc_handler   = proc_dointvec,
1221         },
1222         {
1223                 .procname       = "overcommit_ratio",
1224                 .data           = &sysctl_overcommit_ratio,
1225                 .maxlen         = sizeof(sysctl_overcommit_ratio),
1226                 .mode           = 0644,
1227                 .proc_handler   = overcommit_ratio_handler,
1228         },
1229         {
1230                 .procname       = "overcommit_kbytes",
1231                 .data           = &sysctl_overcommit_kbytes,
1232                 .maxlen         = sizeof(sysctl_overcommit_kbytes),
1233                 .mode           = 0644,
1234                 .proc_handler   = overcommit_kbytes_handler,
1235         },
1236         {
1237                 .procname       = "page-cluster",
1238                 .data           = &page_cluster,
1239                 .maxlen         = sizeof(int),
1240                 .mode           = 0644,
1241                 .proc_handler   = proc_dointvec_minmax,
1242                 .extra1         = &zero,
1243         },
1244         {
1245                 .procname       = "dirty_background_ratio",
1246                 .data           = &dirty_background_ratio,
1247                 .maxlen         = sizeof(dirty_background_ratio),
1248                 .mode           = 0644,
1249                 .proc_handler   = dirty_background_ratio_handler,
1250                 .extra1         = &zero,
1251                 .extra2         = &one_hundred,
1252         },
1253         {
1254                 .procname       = "dirty_background_bytes",
1255                 .data           = &dirty_background_bytes,
1256                 .maxlen         = sizeof(dirty_background_bytes),
1257                 .mode           = 0644,
1258                 .proc_handler   = dirty_background_bytes_handler,
1259                 .extra1         = &one_ul,
1260         },
1261         {
1262                 .procname       = "dirty_ratio",
1263                 .data           = &vm_dirty_ratio,
1264                 .maxlen         = sizeof(vm_dirty_ratio),
1265                 .mode           = 0644,
1266                 .proc_handler   = dirty_ratio_handler,
1267                 .extra1         = &zero,
1268                 .extra2         = &one_hundred,
1269         },
1270         {
1271                 .procname       = "dirty_bytes",
1272                 .data           = &vm_dirty_bytes,
1273                 .maxlen         = sizeof(vm_dirty_bytes),
1274                 .mode           = 0644,
1275                 .proc_handler   = dirty_bytes_handler,
1276                 .extra1         = &dirty_bytes_min,
1277         },
1278         {
1279                 .procname       = "dirty_writeback_centisecs",
1280                 .data           = &dirty_writeback_interval,
1281                 .maxlen         = sizeof(dirty_writeback_interval),
1282                 .mode           = 0644,
1283                 .proc_handler   = dirty_writeback_centisecs_handler,
1284         },
1285         {
1286                 .procname       = "dirty_expire_centisecs",
1287                 .data           = &dirty_expire_interval,
1288                 .maxlen         = sizeof(dirty_expire_interval),
1289                 .mode           = 0644,
1290                 .proc_handler   = proc_dointvec_minmax,
1291                 .extra1         = &zero,
1292         },
1293         {
1294                 .procname       = "dirtytime_expire_seconds",
1295                 .data           = &dirtytime_expire_interval,
1296                 .maxlen         = sizeof(dirty_expire_interval),
1297                 .mode           = 0644,
1298                 .proc_handler   = dirtytime_interval_handler,
1299                 .extra1         = &zero,
1300         },
1301         {
1302                 .procname       = "nr_pdflush_threads",
1303                 .mode           = 0444 /* read-only */,
1304                 .proc_handler   = pdflush_proc_obsolete,
1305         },
1306         {
1307                 .procname       = "swappiness",
1308                 .data           = &vm_swappiness,
1309                 .maxlen         = sizeof(vm_swappiness),
1310                 .mode           = 0644,
1311                 .proc_handler   = proc_dointvec_minmax,
1312                 .extra1         = &zero,
1313                 .extra2         = &one_hundred,
1314         },
1315 #ifdef CONFIG_HUGETLB_PAGE
1316         {
1317                 .procname       = "nr_hugepages",
1318                 .data           = NULL,
1319                 .maxlen         = sizeof(unsigned long),
1320                 .mode           = 0644,
1321                 .proc_handler   = hugetlb_sysctl_handler,
1322         },
1323 #ifdef CONFIG_NUMA
1324         {
1325                 .procname       = "nr_hugepages_mempolicy",
1326                 .data           = NULL,
1327                 .maxlen         = sizeof(unsigned long),
1328                 .mode           = 0644,
1329                 .proc_handler   = &hugetlb_mempolicy_sysctl_handler,
1330         },
1331 #endif
1332          {
1333                 .procname       = "hugetlb_shm_group",
1334                 .data           = &sysctl_hugetlb_shm_group,
1335                 .maxlen         = sizeof(gid_t),
1336                 .mode           = 0644,
1337                 .proc_handler   = proc_dointvec,
1338          },
1339          {
1340                 .procname       = "hugepages_treat_as_movable",
1341                 .data           = &hugepages_treat_as_movable,
1342                 .maxlen         = sizeof(int),
1343                 .mode           = 0644,
1344                 .proc_handler   = proc_dointvec,
1345         },
1346         {
1347                 .procname       = "nr_overcommit_hugepages",
1348                 .data           = NULL,
1349                 .maxlen         = sizeof(unsigned long),
1350                 .mode           = 0644,
1351                 .proc_handler   = hugetlb_overcommit_handler,
1352         },
1353 #endif
1354         {
1355                 .procname       = "lowmem_reserve_ratio",
1356                 .data           = &sysctl_lowmem_reserve_ratio,
1357                 .maxlen         = sizeof(sysctl_lowmem_reserve_ratio),
1358                 .mode           = 0644,
1359                 .proc_handler   = lowmem_reserve_ratio_sysctl_handler,
1360         },
1361         {
1362                 .procname       = "drop_caches",
1363                 .data           = &sysctl_drop_caches,
1364                 .maxlen         = sizeof(int),
1365                 .mode           = 0644,
1366                 .proc_handler   = drop_caches_sysctl_handler,
1367                 .extra1         = &one,
1368                 .extra2         = &four,
1369         },
1370 #ifdef CONFIG_COMPACTION
1371         {
1372                 .procname       = "compact_memory",
1373                 .data           = &sysctl_compact_memory,
1374                 .maxlen         = sizeof(int),
1375                 .mode           = 0200,
1376                 .proc_handler   = sysctl_compaction_handler,
1377         },
1378         {
1379                 .procname       = "extfrag_threshold",
1380                 .data           = &sysctl_extfrag_threshold,
1381                 .maxlen         = sizeof(int),
1382                 .mode           = 0644,
1383                 .proc_handler   = sysctl_extfrag_handler,
1384                 .extra1         = &min_extfrag_threshold,
1385                 .extra2         = &max_extfrag_threshold,
1386         },
1387         {
1388                 .procname       = "compact_unevictable_allowed",
1389                 .data           = &sysctl_compact_unevictable_allowed,
1390                 .maxlen         = sizeof(int),
1391                 .mode           = 0644,
1392                 .proc_handler   = proc_dointvec,
1393                 .extra1         = &zero,
1394                 .extra2         = &one,
1395         },
1396
1397 #endif /* CONFIG_COMPACTION */
1398         {
1399                 .procname       = "min_free_kbytes",
1400                 .data           = &min_free_kbytes,
1401                 .maxlen         = sizeof(min_free_kbytes),
1402                 .mode           = 0644,
1403                 .proc_handler   = min_free_kbytes_sysctl_handler,
1404                 .extra1         = &zero,
1405         },
1406         {
1407                 .procname       = "percpu_pagelist_fraction",
1408                 .data           = &percpu_pagelist_fraction,
1409                 .maxlen         = sizeof(percpu_pagelist_fraction),
1410                 .mode           = 0644,
1411                 .proc_handler   = percpu_pagelist_fraction_sysctl_handler,
1412                 .extra1         = &zero,
1413         },
1414 #ifdef CONFIG_MMU
1415         {
1416                 .procname       = "max_map_count",
1417                 .data           = &sysctl_max_map_count,
1418                 .maxlen         = sizeof(sysctl_max_map_count),
1419                 .mode           = 0644,
1420                 .proc_handler   = proc_dointvec_minmax,
1421                 .extra1         = &zero,
1422         },
1423 #else
1424         {
1425                 .procname       = "nr_trim_pages",
1426                 .data           = &sysctl_nr_trim_pages,
1427                 .maxlen         = sizeof(sysctl_nr_trim_pages),
1428                 .mode           = 0644,
1429                 .proc_handler   = proc_dointvec_minmax,
1430                 .extra1         = &zero,
1431         },
1432 #endif
1433         {
1434                 .procname       = "laptop_mode",
1435                 .data           = &laptop_mode,
1436                 .maxlen         = sizeof(laptop_mode),
1437                 .mode           = 0644,
1438                 .proc_handler   = proc_dointvec_jiffies,
1439         },
1440         {
1441                 .procname       = "block_dump",
1442                 .data           = &block_dump,
1443                 .maxlen         = sizeof(block_dump),
1444                 .mode           = 0644,
1445                 .proc_handler   = proc_dointvec,
1446                 .extra1         = &zero,
1447         },
1448         {
1449                 .procname       = "vfs_cache_pressure",
1450                 .data           = &sysctl_vfs_cache_pressure,
1451                 .maxlen         = sizeof(sysctl_vfs_cache_pressure),
1452                 .mode           = 0644,
1453                 .proc_handler   = proc_dointvec,
1454                 .extra1         = &zero,
1455         },
1456 #ifdef HAVE_ARCH_PICK_MMAP_LAYOUT
1457         {
1458                 .procname       = "legacy_va_layout",
1459                 .data           = &sysctl_legacy_va_layout,
1460                 .maxlen         = sizeof(sysctl_legacy_va_layout),
1461                 .mode           = 0644,
1462                 .proc_handler   = proc_dointvec,
1463                 .extra1         = &zero,
1464         },
1465 #endif
1466 #ifdef CONFIG_NUMA
1467         {
1468                 .procname       = "zone_reclaim_mode",
1469                 .data           = &zone_reclaim_mode,
1470                 .maxlen         = sizeof(zone_reclaim_mode),
1471                 .mode           = 0644,
1472                 .proc_handler   = proc_dointvec,
1473                 .extra1         = &zero,
1474         },
1475         {
1476                 .procname       = "min_unmapped_ratio",
1477                 .data           = &sysctl_min_unmapped_ratio,
1478                 .maxlen         = sizeof(sysctl_min_unmapped_ratio),
1479                 .mode           = 0644,
1480                 .proc_handler   = sysctl_min_unmapped_ratio_sysctl_handler,
1481                 .extra1         = &zero,
1482                 .extra2         = &one_hundred,
1483         },
1484         {
1485                 .procname       = "min_slab_ratio",
1486                 .data           = &sysctl_min_slab_ratio,
1487                 .maxlen         = sizeof(sysctl_min_slab_ratio),
1488                 .mode           = 0644,
1489                 .proc_handler   = sysctl_min_slab_ratio_sysctl_handler,
1490                 .extra1         = &zero,
1491                 .extra2         = &one_hundred,
1492         },
1493 #endif
1494 #ifdef CONFIG_SMP
1495         {
1496                 .procname       = "stat_interval",
1497                 .data           = &sysctl_stat_interval,
1498                 .maxlen         = sizeof(sysctl_stat_interval),
1499                 .mode           = 0644,
1500                 .proc_handler   = proc_dointvec_jiffies,
1501         },
1502 #endif
1503 #ifdef CONFIG_MMU
1504         {
1505                 .procname       = "mmap_min_addr",
1506                 .data           = &dac_mmap_min_addr,
1507                 .maxlen         = sizeof(unsigned long),
1508                 .mode           = 0644,
1509                 .proc_handler   = mmap_min_addr_handler,
1510         },
1511 #endif
1512 #ifdef CONFIG_NUMA
1513         {
1514                 .procname       = "numa_zonelist_order",
1515                 .data           = &numa_zonelist_order,
1516                 .maxlen         = NUMA_ZONELIST_ORDER_LEN,
1517                 .mode           = 0644,
1518                 .proc_handler   = numa_zonelist_order_handler,
1519         },
1520 #endif
1521 #if (defined(CONFIG_X86_32) && !defined(CONFIG_UML))|| \
1522    (defined(CONFIG_SUPERH) && defined(CONFIG_VSYSCALL))
1523         {
1524                 .procname       = "vdso_enabled",
1525 #ifdef CONFIG_X86_32
1526                 .data           = &vdso32_enabled,
1527                 .maxlen         = sizeof(vdso32_enabled),
1528 #else
1529                 .data           = &vdso_enabled,
1530                 .maxlen         = sizeof(vdso_enabled),
1531 #endif
1532                 .mode           = 0644,
1533                 .proc_handler   = proc_dointvec,
1534                 .extra1         = &zero,
1535         },
1536 #endif
1537 #ifdef CONFIG_HIGHMEM
1538         {
1539                 .procname       = "highmem_is_dirtyable",
1540                 .data           = &vm_highmem_is_dirtyable,
1541                 .maxlen         = sizeof(vm_highmem_is_dirtyable),
1542                 .mode           = 0644,
1543                 .proc_handler   = proc_dointvec_minmax,
1544                 .extra1         = &zero,
1545                 .extra2         = &one,
1546         },
1547 #endif
1548 #ifdef CONFIG_MEMORY_FAILURE
1549         {
1550                 .procname       = "memory_failure_early_kill",
1551                 .data           = &sysctl_memory_failure_early_kill,
1552                 .maxlen         = sizeof(sysctl_memory_failure_early_kill),
1553                 .mode           = 0644,
1554                 .proc_handler   = proc_dointvec_minmax,
1555                 .extra1         = &zero,
1556                 .extra2         = &one,
1557         },
1558         {
1559                 .procname       = "memory_failure_recovery",
1560                 .data           = &sysctl_memory_failure_recovery,
1561                 .maxlen         = sizeof(sysctl_memory_failure_recovery),
1562                 .mode           = 0644,
1563                 .proc_handler   = proc_dointvec_minmax,
1564                 .extra1         = &zero,
1565                 .extra2         = &one,
1566         },
1567 #endif
1568         {
1569                 .procname       = "user_reserve_kbytes",
1570                 .data           = &sysctl_user_reserve_kbytes,
1571                 .maxlen         = sizeof(sysctl_user_reserve_kbytes),
1572                 .mode           = 0644,
1573                 .proc_handler   = proc_doulongvec_minmax,
1574         },
1575         {
1576                 .procname       = "admin_reserve_kbytes",
1577                 .data           = &sysctl_admin_reserve_kbytes,
1578                 .maxlen         = sizeof(sysctl_admin_reserve_kbytes),
1579                 .mode           = 0644,
1580                 .proc_handler   = proc_doulongvec_minmax,
1581         },
1582 #ifdef CONFIG_HAVE_ARCH_MMAP_RND_BITS
1583         {
1584                 .procname       = "mmap_rnd_bits",
1585                 .data           = &mmap_rnd_bits,
1586                 .maxlen         = sizeof(mmap_rnd_bits),
1587                 .mode           = 0600,
1588                 .proc_handler   = proc_dointvec_minmax,
1589                 .extra1         = (void *)&mmap_rnd_bits_min,
1590                 .extra2         = (void *)&mmap_rnd_bits_max,
1591         },
1592 #endif
1593 #ifdef CONFIG_HAVE_ARCH_MMAP_RND_COMPAT_BITS
1594         {
1595                 .procname       = "mmap_rnd_compat_bits",
1596                 .data           = &mmap_rnd_compat_bits,
1597                 .maxlen         = sizeof(mmap_rnd_compat_bits),
1598                 .mode           = 0600,
1599                 .proc_handler   = proc_dointvec_minmax,
1600                 .extra1         = (void *)&mmap_rnd_compat_bits_min,
1601                 .extra2         = (void *)&mmap_rnd_compat_bits_max,
1602         },
1603 #endif
1604         { }
1605 };
1606
1607 static struct ctl_table fs_table[] = {
1608         {
1609                 .procname       = "inode-nr",
1610                 .data           = &inodes_stat,
1611                 .maxlen         = 2*sizeof(long),
1612                 .mode           = 0444,
1613                 .proc_handler   = proc_nr_inodes,
1614         },
1615         {
1616                 .procname       = "inode-state",
1617                 .data           = &inodes_stat,
1618                 .maxlen         = 7*sizeof(long),
1619                 .mode           = 0444,
1620                 .proc_handler   = proc_nr_inodes,
1621         },
1622         {
1623                 .procname       = "file-nr",
1624                 .data           = &files_stat,
1625                 .maxlen         = sizeof(files_stat),
1626                 .mode           = 0444,
1627                 .proc_handler   = proc_nr_files,
1628         },
1629         {
1630                 .procname       = "file-max",
1631                 .data           = &files_stat.max_files,
1632                 .maxlen         = sizeof(files_stat.max_files),
1633                 .mode           = 0644,
1634                 .proc_handler   = proc_doulongvec_minmax,
1635         },
1636         {
1637                 .procname       = "nr_open",
1638                 .data           = &sysctl_nr_open,
1639                 .maxlen         = sizeof(int),
1640                 .mode           = 0644,
1641                 .proc_handler   = proc_dointvec_minmax,
1642                 .extra1         = &sysctl_nr_open_min,
1643                 .extra2         = &sysctl_nr_open_max,
1644         },
1645         {
1646                 .procname       = "dentry-state",
1647                 .data           = &dentry_stat,
1648                 .maxlen         = 6*sizeof(long),
1649                 .mode           = 0444,
1650                 .proc_handler   = proc_nr_dentry,
1651         },
1652         {
1653                 .procname       = "overflowuid",
1654                 .data           = &fs_overflowuid,
1655                 .maxlen         = sizeof(int),
1656                 .mode           = 0644,
1657                 .proc_handler   = proc_dointvec_minmax,
1658                 .extra1         = &minolduid,
1659                 .extra2         = &maxolduid,
1660         },
1661         {
1662                 .procname       = "overflowgid",
1663                 .data           = &fs_overflowgid,
1664                 .maxlen         = sizeof(int),
1665                 .mode           = 0644,
1666                 .proc_handler   = proc_dointvec_minmax,
1667                 .extra1         = &minolduid,
1668                 .extra2         = &maxolduid,
1669         },
1670 #ifdef CONFIG_FILE_LOCKING
1671         {
1672                 .procname       = "leases-enable",
1673                 .data           = &leases_enable,
1674                 .maxlen         = sizeof(int),
1675                 .mode           = 0644,
1676                 .proc_handler   = proc_dointvec,
1677         },
1678 #endif
1679 #ifdef CONFIG_DNOTIFY
1680         {
1681                 .procname       = "dir-notify-enable",
1682                 .data           = &dir_notify_enable,
1683                 .maxlen         = sizeof(int),
1684                 .mode           = 0644,
1685                 .proc_handler   = proc_dointvec,
1686         },
1687 #endif
1688 #ifdef CONFIG_MMU
1689 #ifdef CONFIG_FILE_LOCKING
1690         {
1691                 .procname       = "lease-break-time",
1692                 .data           = &lease_break_time,
1693                 .maxlen         = sizeof(int),
1694                 .mode           = 0644,
1695                 .proc_handler   = proc_dointvec,
1696         },
1697 #endif
1698 #ifdef CONFIG_AIO
1699         {
1700                 .procname       = "aio-nr",
1701                 .data           = &aio_nr,
1702                 .maxlen         = sizeof(aio_nr),
1703                 .mode           = 0444,
1704                 .proc_handler   = proc_doulongvec_minmax,
1705         },
1706         {
1707                 .procname       = "aio-max-nr",
1708                 .data           = &aio_max_nr,
1709                 .maxlen         = sizeof(aio_max_nr),
1710                 .mode           = 0644,
1711                 .proc_handler   = proc_doulongvec_minmax,
1712         },
1713 #if IS_ENABLED(CONFIG_AIO_THREAD)
1714         {
1715                 .procname       = "aio-auto-threads",
1716                 .data           = &aio_auto_threads,
1717                 .maxlen         = sizeof(aio_auto_threads),
1718                 .mode           = 0644,
1719                 .proc_handler   = proc_doulongvec_minmax,
1720         },
1721 #endif
1722 #endif /* CONFIG_AIO */
1723 #ifdef CONFIG_INOTIFY_USER
1724         {
1725                 .procname       = "inotify",
1726                 .mode           = 0555,
1727                 .child          = inotify_table,
1728         },
1729 #endif
1730 #ifdef CONFIG_EPOLL
1731         {
1732                 .procname       = "epoll",
1733                 .mode           = 0555,
1734                 .child          = epoll_table,
1735         },
1736 #endif
1737 #endif
1738         {
1739                 .procname       = "protected_symlinks",
1740                 .data           = &sysctl_protected_symlinks,
1741                 .maxlen         = sizeof(int),
1742                 .mode           = 0600,
1743                 .proc_handler   = proc_dointvec_minmax,
1744                 .extra1         = &zero,
1745                 .extra2         = &one,
1746         },
1747         {
1748                 .procname       = "protected_hardlinks",
1749                 .data           = &sysctl_protected_hardlinks,
1750                 .maxlen         = sizeof(int),
1751                 .mode           = 0600,
1752                 .proc_handler   = proc_dointvec_minmax,
1753                 .extra1         = &zero,
1754                 .extra2         = &one,
1755         },
1756         {
1757                 .procname       = "suid_dumpable",
1758                 .data           = &suid_dumpable,
1759                 .maxlen         = sizeof(int),
1760                 .mode           = 0644,
1761                 .proc_handler   = proc_dointvec_minmax_coredump,
1762                 .extra1         = &zero,
1763                 .extra2         = &two,
1764         },
1765 #if defined(CONFIG_BINFMT_MISC) || defined(CONFIG_BINFMT_MISC_MODULE)
1766         {
1767                 .procname       = "binfmt_misc",
1768                 .mode           = 0555,
1769                 .child          = sysctl_mount_point,
1770         },
1771 #endif
1772         {
1773                 .procname       = "pipe-max-size",
1774                 .data           = &pipe_max_size,
1775                 .maxlen         = sizeof(int),
1776                 .mode           = 0644,
1777                 .proc_handler   = &pipe_proc_fn,
1778                 .extra1         = &pipe_min_size,
1779         },
1780         {
1781                 .procname       = "pipe-user-pages-hard",
1782                 .data           = &pipe_user_pages_hard,
1783                 .maxlen         = sizeof(pipe_user_pages_hard),
1784                 .mode           = 0644,
1785                 .proc_handler   = proc_doulongvec_minmax,
1786         },
1787         {
1788                 .procname       = "pipe-user-pages-soft",
1789                 .data           = &pipe_user_pages_soft,
1790                 .maxlen         = sizeof(pipe_user_pages_soft),
1791                 .mode           = 0644,
1792                 .proc_handler   = proc_doulongvec_minmax,
1793         },
1794         { }
1795 };
1796
1797 static struct ctl_table debug_table[] = {
1798 #ifdef CONFIG_SYSCTL_EXCEPTION_TRACE
1799         {
1800                 .procname       = "exception-trace",
1801                 .data           = &show_unhandled_signals,
1802                 .maxlen         = sizeof(int),
1803                 .mode           = 0644,
1804                 .proc_handler   = proc_dointvec
1805         },
1806 #endif
1807 #if defined(CONFIG_OPTPROBES)
1808         {
1809                 .procname       = "kprobes-optimization",
1810                 .data           = &sysctl_kprobes_optimization,
1811                 .maxlen         = sizeof(int),
1812                 .mode           = 0644,
1813                 .proc_handler   = proc_kprobes_optimization_handler,
1814                 .extra1         = &zero,
1815                 .extra2         = &one,
1816         },
1817 #endif
1818         { }
1819 };
1820
1821 static struct ctl_table dev_table[] = {
1822         { }
1823 };
1824
1825 int __init sysctl_init(void)
1826 {
1827         struct ctl_table_header *hdr;
1828
1829         hdr = register_sysctl_table(sysctl_base_table);
1830         kmemleak_not_leak(hdr);
1831         return 0;
1832 }
1833
1834 #endif /* CONFIG_SYSCTL */
1835
1836 /*
1837  * /proc/sys support
1838  */
1839
1840 #ifdef CONFIG_PROC_SYSCTL
1841
1842 static int _proc_do_string(char *data, int maxlen, int write,
1843                            char __user *buffer,
1844                            size_t *lenp, loff_t *ppos)
1845 {
1846         size_t len;
1847         char __user *p;
1848         char c;
1849
1850         if (!data || !maxlen || !*lenp) {
1851                 *lenp = 0;
1852                 return 0;
1853         }
1854
1855         if (write) {
1856                 if (sysctl_writes_strict == SYSCTL_WRITES_STRICT) {
1857                         /* Only continue writes not past the end of buffer. */
1858                         len = strlen(data);
1859                         if (len > maxlen - 1)
1860                                 len = maxlen - 1;
1861
1862                         if (*ppos > len)
1863                                 return 0;
1864                         len = *ppos;
1865                 } else {
1866                         /* Start writing from beginning of buffer. */
1867                         len = 0;
1868                 }
1869
1870                 *ppos += *lenp;
1871                 p = buffer;
1872                 while ((p - buffer) < *lenp && len < maxlen - 1) {
1873                         if (get_user(c, p++))
1874                                 return -EFAULT;
1875                         if (c == 0 || c == '\n')
1876                                 break;
1877                         data[len++] = c;
1878                 }
1879                 data[len] = 0;
1880         } else {
1881                 len = strlen(data);
1882                 if (len > maxlen)
1883                         len = maxlen;
1884
1885                 if (*ppos > len) {
1886                         *lenp = 0;
1887                         return 0;
1888                 }
1889
1890                 data += *ppos;
1891                 len  -= *ppos;
1892
1893                 if (len > *lenp)
1894                         len = *lenp;
1895                 if (len)
1896                         if (copy_to_user(buffer, data, len))
1897                                 return -EFAULT;
1898                 if (len < *lenp) {
1899                         if (put_user('\n', buffer + len))
1900                                 return -EFAULT;
1901                         len++;
1902                 }
1903                 *lenp = len;
1904                 *ppos += len;
1905         }
1906         return 0;
1907 }
1908
1909 static void warn_sysctl_write(struct ctl_table *table)
1910 {
1911         pr_warn_once("%s wrote to %s when file position was not 0!\n"
1912                 "This will not be supported in the future. To silence this\n"
1913                 "warning, set kernel.sysctl_writes_strict = -1\n",
1914                 current->comm, table->procname);
1915 }
1916
1917 /**
1918  * proc_dostring - read a string sysctl
1919  * @table: the sysctl table
1920  * @write: %TRUE if this is a write to the sysctl file
1921  * @buffer: the user buffer
1922  * @lenp: the size of the user buffer
1923  * @ppos: file position
1924  *
1925  * Reads/writes a string from/to the user buffer. If the kernel
1926  * buffer provided is not large enough to hold the string, the
1927  * string is truncated. The copied string is %NULL-terminated.
1928  * If the string is being read by the user process, it is copied
1929  * and a newline '\n' is added. It is truncated if the buffer is
1930  * not large enough.
1931  *
1932  * Returns 0 on success.
1933  */
1934 int proc_dostring(struct ctl_table *table, int write,
1935                   void __user *buffer, size_t *lenp, loff_t *ppos)
1936 {
1937         if (write && *ppos && sysctl_writes_strict == SYSCTL_WRITES_WARN)
1938                 warn_sysctl_write(table);
1939
1940         return _proc_do_string((char *)(table->data), table->maxlen, write,
1941                                (char __user *)buffer, lenp, ppos);
1942 }
1943
1944 static size_t proc_skip_spaces(char **buf)
1945 {
1946         size_t ret;
1947         char *tmp = skip_spaces(*buf);
1948         ret = tmp - *buf;
1949         *buf = tmp;
1950         return ret;
1951 }
1952
1953 static void proc_skip_char(char **buf, size_t *size, const char v)
1954 {
1955         while (*size) {
1956                 if (**buf != v)
1957                         break;
1958                 (*size)--;
1959                 (*buf)++;
1960         }
1961 }
1962
1963 #define TMPBUFLEN 22
1964 /**
1965  * proc_get_long - reads an ASCII formatted integer from a user buffer
1966  *
1967  * @buf: a kernel buffer
1968  * @size: size of the kernel buffer
1969  * @val: this is where the number will be stored
1970  * @neg: set to %TRUE if number is negative
1971  * @perm_tr: a vector which contains the allowed trailers
1972  * @perm_tr_len: size of the perm_tr vector
1973  * @tr: pointer to store the trailer character
1974  *
1975  * In case of success %0 is returned and @buf and @size are updated with
1976  * the amount of bytes read. If @tr is non-NULL and a trailing
1977  * character exists (size is non-zero after returning from this
1978  * function), @tr is updated with the trailing character.
1979  */
1980 static int proc_get_long(char **buf, size_t *size,
1981                           unsigned long *val, bool *neg,
1982                           const char *perm_tr, unsigned perm_tr_len, char *tr)
1983 {
1984         int len;
1985         char *p, tmp[TMPBUFLEN];
1986
1987         if (!*size)
1988                 return -EINVAL;
1989
1990         len = *size;
1991         if (len > TMPBUFLEN - 1)
1992                 len = TMPBUFLEN - 1;
1993
1994         memcpy(tmp, *buf, len);
1995
1996         tmp[len] = 0;
1997         p = tmp;
1998         if (*p == '-' && *size > 1) {
1999                 *neg = true;
2000                 p++;
2001         } else
2002                 *neg = false;
2003         if (!isdigit(*p))
2004                 return -EINVAL;
2005
2006         *val = simple_strtoul(p, &p, 0);
2007
2008         len = p - tmp;
2009
2010         /* We don't know if the next char is whitespace thus we may accept
2011          * invalid integers (e.g. 1234...a) or two integers instead of one
2012          * (e.g. 123...1). So lets not allow such large numbers. */
2013         if (len == TMPBUFLEN - 1)
2014                 return -EINVAL;
2015
2016         if (len < *size && perm_tr_len && !memchr(perm_tr, *p, perm_tr_len))
2017                 return -EINVAL;
2018
2019         if (tr && (len < *size))
2020                 *tr = *p;
2021
2022         *buf += len;
2023         *size -= len;
2024
2025         return 0;
2026 }
2027
2028 /**
2029  * proc_put_long - converts an integer to a decimal ASCII formatted string
2030  *
2031  * @buf: the user buffer
2032  * @size: the size of the user buffer
2033  * @val: the integer to be converted
2034  * @neg: sign of the number, %TRUE for negative
2035  *
2036  * In case of success %0 is returned and @buf and @size are updated with
2037  * the amount of bytes written.
2038  */
2039 static int proc_put_long(void __user **buf, size_t *size, unsigned long val,
2040                           bool neg)
2041 {
2042         int len;
2043         char tmp[TMPBUFLEN], *p = tmp;
2044
2045         sprintf(p, "%s%lu", neg ? "-" : "", val);
2046         len = strlen(tmp);
2047         if (len > *size)
2048                 len = *size;
2049         if (copy_to_user(*buf, tmp, len))
2050                 return -EFAULT;
2051         *size -= len;
2052         *buf += len;
2053         return 0;
2054 }
2055 #undef TMPBUFLEN
2056
2057 static int proc_put_char(void __user **buf, size_t *size, char c)
2058 {
2059         if (*size) {
2060                 char __user **buffer = (char __user **)buf;
2061                 if (put_user(c, *buffer))
2062                         return -EFAULT;
2063                 (*size)--, (*buffer)++;
2064                 *buf = *buffer;
2065         }
2066         return 0;
2067 }
2068
2069 static int do_proc_dointvec_conv(bool *negp, unsigned long *lvalp,
2070                                  int *valp,
2071                                  int write, void *data)
2072 {
2073         if (write) {
2074                 if (*negp) {
2075                         if (*lvalp > (unsigned long) INT_MAX + 1)
2076                                 return -EINVAL;
2077                         *valp = -*lvalp;
2078                 } else {
2079                         if (*lvalp > (unsigned long) INT_MAX)
2080                                 return -EINVAL;
2081                         *valp = *lvalp;
2082                 }
2083         } else {
2084                 int val = *valp;
2085                 if (val < 0) {
2086                         *negp = true;
2087                         *lvalp = -(unsigned long)val;
2088                 } else {
2089                         *negp = false;
2090                         *lvalp = (unsigned long)val;
2091                 }
2092         }
2093         return 0;
2094 }
2095
2096 static const char proc_wspace_sep[] = { ' ', '\t', '\n' };
2097
2098 static int __do_proc_dointvec(void *tbl_data, struct ctl_table *table,
2099                   int write, void __user *buffer,
2100                   size_t *lenp, loff_t *ppos,
2101                   int (*conv)(bool *negp, unsigned long *lvalp, int *valp,
2102                               int write, void *data),
2103                   void *data)
2104 {
2105         int *i, vleft, first = 1, err = 0;
2106         size_t left;
2107         char *kbuf = NULL, *p;
2108
2109         if (!tbl_data || !table->maxlen || !*lenp || (*ppos && !write)) {
2110                 *lenp = 0;
2111                 return 0;
2112         }
2113
2114         i = (int *) tbl_data;
2115         vleft = table->maxlen / sizeof(*i);
2116         left = *lenp;
2117
2118         if (!conv)
2119                 conv = do_proc_dointvec_conv;
2120
2121         if (write) {
2122                 if (*ppos) {
2123                         switch (sysctl_writes_strict) {
2124                         case SYSCTL_WRITES_STRICT:
2125                                 goto out;
2126                         case SYSCTL_WRITES_WARN:
2127                                 warn_sysctl_write(table);
2128                                 break;
2129                         default:
2130                                 break;
2131                         }
2132                 }
2133
2134                 if (left > PAGE_SIZE - 1)
2135                         left = PAGE_SIZE - 1;
2136                 p = kbuf = memdup_user_nul(buffer, left);
2137                 if (IS_ERR(kbuf))
2138                         return PTR_ERR(kbuf);
2139         }
2140
2141         for (; left && vleft--; i++, first=0) {
2142                 unsigned long lval;
2143                 bool neg;
2144
2145                 if (write) {
2146                         left -= proc_skip_spaces(&p);
2147
2148                         if (!left)
2149                                 break;
2150                         err = proc_get_long(&p, &left, &lval, &neg,
2151                                              proc_wspace_sep,
2152                                              sizeof(proc_wspace_sep), NULL);
2153                         if (err)
2154                                 break;
2155                         if (conv(&neg, &lval, i, 1, data)) {
2156                                 err = -EINVAL;
2157                                 break;
2158                         }
2159                 } else {
2160                         if (conv(&neg, &lval, i, 0, data)) {
2161                                 err = -EINVAL;
2162                                 break;
2163                         }
2164                         if (!first)
2165                                 err = proc_put_char(&buffer, &left, '\t');
2166                         if (err)
2167                                 break;
2168                         err = proc_put_long(&buffer, &left, lval, neg);
2169                         if (err)
2170                                 break;
2171                 }
2172         }
2173
2174         if (!write && !first && left && !err)
2175                 err = proc_put_char(&buffer, &left, '\n');
2176         if (write && !err && left)
2177                 left -= proc_skip_spaces(&p);
2178         if (write) {
2179                 kfree(kbuf);
2180                 if (first)
2181                         return err ? : -EINVAL;
2182         }
2183         *lenp -= left;
2184 out:
2185         *ppos += *lenp;
2186         return err;
2187 }
2188
2189 static int do_proc_dointvec(struct ctl_table *table, int write,
2190                   void __user *buffer, size_t *lenp, loff_t *ppos,
2191                   int (*conv)(bool *negp, unsigned long *lvalp, int *valp,
2192                               int write, void *data),
2193                   void *data)
2194 {
2195         return __do_proc_dointvec(table->data, table, write,
2196                         buffer, lenp, ppos, conv, data);
2197 }
2198
2199 /**
2200  * proc_dointvec - read a vector of integers
2201  * @table: the sysctl table
2202  * @write: %TRUE if this is a write to the sysctl file
2203  * @buffer: the user buffer
2204  * @lenp: the size of the user buffer
2205  * @ppos: file position
2206  *
2207  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2208  * values from/to the user buffer, treated as an ASCII string.
2209  *
2210  * Returns 0 on success.
2211  */
2212 int proc_dointvec(struct ctl_table *table, int write,
2213                      void __user *buffer, size_t *lenp, loff_t *ppos)
2214 {
2215     return do_proc_dointvec(table,write,buffer,lenp,ppos,
2216                             NULL,NULL);
2217 }
2218
2219 /*
2220  * Taint values can only be increased
2221  * This means we can safely use a temporary.
2222  */
2223 static int proc_taint(struct ctl_table *table, int write,
2224                                void __user *buffer, size_t *lenp, loff_t *ppos)
2225 {
2226         struct ctl_table t;
2227         unsigned long tmptaint = get_taint();
2228         int err;
2229
2230         if (write && !capable(CAP_SYS_ADMIN))
2231                 return -EPERM;
2232
2233         t = *table;
2234         t.data = &tmptaint;
2235         err = proc_doulongvec_minmax(&t, write, buffer, lenp, ppos);
2236         if (err < 0)
2237                 return err;
2238
2239         if (write) {
2240                 /*
2241                  * Poor man's atomic or. Not worth adding a primitive
2242                  * to everyone's atomic.h for this
2243                  */
2244                 int i;
2245                 for (i = 0; i < BITS_PER_LONG && tmptaint >> i; i++) {
2246                         if ((tmptaint >> i) & 1)
2247                                 add_taint(i, LOCKDEP_STILL_OK);
2248                 }
2249         }
2250
2251         return err;
2252 }
2253
2254 #ifdef CONFIG_PRINTK
2255 static int proc_dointvec_minmax_sysadmin(struct ctl_table *table, int write,
2256                                 void __user *buffer, size_t *lenp, loff_t *ppos)
2257 {
2258         if (write && !capable(CAP_SYS_ADMIN))
2259                 return -EPERM;
2260
2261         return proc_dointvec_minmax(table, write, buffer, lenp, ppos);
2262 }
2263 #endif
2264
2265 struct do_proc_dointvec_minmax_conv_param {
2266         int *min;
2267         int *max;
2268 };
2269
2270 static int do_proc_dointvec_minmax_conv(bool *negp, unsigned long *lvalp,
2271                                         int *valp,
2272                                         int write, void *data)
2273 {
2274         struct do_proc_dointvec_minmax_conv_param *param = data;
2275         if (write) {
2276                 int val = *negp ? -*lvalp : *lvalp;
2277                 if ((param->min && *param->min > val) ||
2278                     (param->max && *param->max < val))
2279                         return -EINVAL;
2280                 *valp = val;
2281         } else {
2282                 int val = *valp;
2283                 if (val < 0) {
2284                         *negp = true;
2285                         *lvalp = -(unsigned long)val;
2286                 } else {
2287                         *negp = false;
2288                         *lvalp = (unsigned long)val;
2289                 }
2290         }
2291         return 0;
2292 }
2293
2294 /**
2295  * proc_dointvec_minmax - read a vector of integers with min/max values
2296  * @table: the sysctl table
2297  * @write: %TRUE if this is a write to the sysctl file
2298  * @buffer: the user buffer
2299  * @lenp: the size of the user buffer
2300  * @ppos: file position
2301  *
2302  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2303  * values from/to the user buffer, treated as an ASCII string.
2304  *
2305  * This routine will ensure the values are within the range specified by
2306  * table->extra1 (min) and table->extra2 (max).
2307  *
2308  * Returns 0 on success.
2309  */
2310 int proc_dointvec_minmax(struct ctl_table *table, int write,
2311                   void __user *buffer, size_t *lenp, loff_t *ppos)
2312 {
2313         struct do_proc_dointvec_minmax_conv_param param = {
2314                 .min = (int *) table->extra1,
2315                 .max = (int *) table->extra2,
2316         };
2317         return do_proc_dointvec(table, write, buffer, lenp, ppos,
2318                                 do_proc_dointvec_minmax_conv, &param);
2319 }
2320
2321 static void validate_coredump_safety(void)
2322 {
2323 #ifdef CONFIG_COREDUMP
2324         if (suid_dumpable == SUID_DUMP_ROOT &&
2325             core_pattern[0] != '/' && core_pattern[0] != '|') {
2326                 printk(KERN_WARNING "Unsafe core_pattern used with "\
2327                         "suid_dumpable=2. Pipe handler or fully qualified "\
2328                         "core dump path required.\n");
2329         }
2330 #endif
2331 }
2332
2333 static int proc_dointvec_minmax_coredump(struct ctl_table *table, int write,
2334                 void __user *buffer, size_t *lenp, loff_t *ppos)
2335 {
2336         int error = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
2337         if (!error)
2338                 validate_coredump_safety();
2339         return error;
2340 }
2341
2342 #ifdef CONFIG_COREDUMP
2343 static int proc_dostring_coredump(struct ctl_table *table, int write,
2344                   void __user *buffer, size_t *lenp, loff_t *ppos)
2345 {
2346         int error = proc_dostring(table, write, buffer, lenp, ppos);
2347         if (!error)
2348                 validate_coredump_safety();
2349         return error;
2350 }
2351 #endif
2352
2353 static int __do_proc_doulongvec_minmax(void *data, struct ctl_table *table, int write,
2354                                      void __user *buffer,
2355                                      size_t *lenp, loff_t *ppos,
2356                                      unsigned long convmul,
2357                                      unsigned long convdiv)
2358 {
2359         unsigned long *i, *min, *max;
2360         int vleft, first = 1, err = 0;
2361         size_t left;
2362         char *kbuf = NULL, *p;
2363
2364         if (!data || !table->maxlen || !*lenp || (*ppos && !write)) {
2365                 *lenp = 0;
2366                 return 0;
2367         }
2368
2369         i = (unsigned long *) data;
2370         min = (unsigned long *) table->extra1;
2371         max = (unsigned long *) table->extra2;
2372         vleft = table->maxlen / sizeof(unsigned long);
2373         left = *lenp;
2374
2375         if (write) {
2376                 if (*ppos) {
2377                         switch (sysctl_writes_strict) {
2378                         case SYSCTL_WRITES_STRICT:
2379                                 goto out;
2380                         case SYSCTL_WRITES_WARN:
2381                                 warn_sysctl_write(table);
2382                                 break;
2383                         default:
2384                                 break;
2385                         }
2386                 }
2387
2388                 if (left > PAGE_SIZE - 1)
2389                         left = PAGE_SIZE - 1;
2390                 p = kbuf = memdup_user_nul(buffer, left);
2391                 if (IS_ERR(kbuf))
2392                         return PTR_ERR(kbuf);
2393         }
2394
2395         for (; left && vleft--; i++, first = 0) {
2396                 unsigned long val;
2397
2398                 if (write) {
2399                         bool neg;
2400
2401                         left -= proc_skip_spaces(&p);
2402
2403                         err = proc_get_long(&p, &left, &val, &neg,
2404                                              proc_wspace_sep,
2405                                              sizeof(proc_wspace_sep), NULL);
2406                         if (err)
2407                                 break;
2408                         if (neg)
2409                                 continue;
2410                         if ((min && val < *min) || (max && val > *max))
2411                                 continue;
2412                         *i = val;
2413                 } else {
2414                         val = convdiv * (*i) / convmul;
2415                         if (!first) {
2416                                 err = proc_put_char(&buffer, &left, '\t');
2417                                 if (err)
2418                                         break;
2419                         }
2420                         err = proc_put_long(&buffer, &left, val, false);
2421                         if (err)
2422                                 break;
2423                 }
2424         }
2425
2426         if (!write && !first && left && !err)
2427                 err = proc_put_char(&buffer, &left, '\n');
2428         if (write && !err)
2429                 left -= proc_skip_spaces(&p);
2430         if (write) {
2431                 kfree(kbuf);
2432                 if (first)
2433                         return err ? : -EINVAL;
2434         }
2435         *lenp -= left;
2436 out:
2437         *ppos += *lenp;
2438         return err;
2439 }
2440
2441 static int do_proc_doulongvec_minmax(struct ctl_table *table, int write,
2442                                      void __user *buffer,
2443                                      size_t *lenp, loff_t *ppos,
2444                                      unsigned long convmul,
2445                                      unsigned long convdiv)
2446 {
2447         return __do_proc_doulongvec_minmax(table->data, table, write,
2448                         buffer, lenp, ppos, convmul, convdiv);
2449 }
2450
2451 /**
2452  * proc_doulongvec_minmax - read a vector of long integers with min/max values
2453  * @table: the sysctl table
2454  * @write: %TRUE if this is a write to the sysctl file
2455  * @buffer: the user buffer
2456  * @lenp: the size of the user buffer
2457  * @ppos: file position
2458  *
2459  * Reads/writes up to table->maxlen/sizeof(unsigned long) unsigned long
2460  * values from/to the user buffer, treated as an ASCII string.
2461  *
2462  * This routine will ensure the values are within the range specified by
2463  * table->extra1 (min) and table->extra2 (max).
2464  *
2465  * Returns 0 on success.
2466  */
2467 int proc_doulongvec_minmax(struct ctl_table *table, int write,
2468                            void __user *buffer, size_t *lenp, loff_t *ppos)
2469 {
2470     return do_proc_doulongvec_minmax(table, write, buffer, lenp, ppos, 1l, 1l);
2471 }
2472
2473 /**
2474  * proc_doulongvec_ms_jiffies_minmax - read a vector of millisecond values with min/max values
2475  * @table: the sysctl table
2476  * @write: %TRUE if this is a write to the sysctl file
2477  * @buffer: the user buffer
2478  * @lenp: the size of the user buffer
2479  * @ppos: file position
2480  *
2481  * Reads/writes up to table->maxlen/sizeof(unsigned long) unsigned long
2482  * values from/to the user buffer, treated as an ASCII string. The values
2483  * are treated as milliseconds, and converted to jiffies when they are stored.
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_ms_jiffies_minmax(struct ctl_table *table, int write,
2491                                       void __user *buffer,
2492                                       size_t *lenp, loff_t *ppos)
2493 {
2494     return do_proc_doulongvec_minmax(table, write, buffer,
2495                                      lenp, ppos, HZ, 1000l);
2496 }
2497
2498
2499 static int do_proc_dointvec_jiffies_conv(bool *negp, unsigned long *lvalp,
2500                                          int *valp,
2501                                          int write, void *data)
2502 {
2503         if (write) {
2504                 if (*lvalp > LONG_MAX / HZ)
2505                         return 1;
2506                 *valp = *negp ? -(*lvalp*HZ) : (*lvalp*HZ);
2507         } else {
2508                 int val = *valp;
2509                 unsigned long lval;
2510                 if (val < 0) {
2511                         *negp = true;
2512                         lval = -(unsigned long)val;
2513                 } else {
2514                         *negp = false;
2515                         lval = (unsigned long)val;
2516                 }
2517                 *lvalp = lval / HZ;
2518         }
2519         return 0;
2520 }
2521
2522 static int do_proc_dointvec_userhz_jiffies_conv(bool *negp, unsigned long *lvalp,
2523                                                 int *valp,
2524                                                 int write, void *data)
2525 {
2526         if (write) {
2527                 if (USER_HZ < HZ && *lvalp > (LONG_MAX / HZ) * USER_HZ)
2528                         return 1;
2529                 *valp = clock_t_to_jiffies(*negp ? -*lvalp : *lvalp);
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 = jiffies_to_clock_t(lval);
2541         }
2542         return 0;
2543 }
2544
2545 static int do_proc_dointvec_ms_jiffies_conv(bool *negp, unsigned long *lvalp,
2546                                             int *valp,
2547                                             int write, void *data)
2548 {
2549         if (write) {
2550                 unsigned long jif = msecs_to_jiffies(*negp ? -*lvalp : *lvalp);
2551
2552                 if (jif > INT_MAX)
2553                         return 1;
2554                 *valp = (int)jif;
2555         } else {
2556                 int val = *valp;
2557                 unsigned long lval;
2558                 if (val < 0) {
2559                         *negp = true;
2560                         lval = -(unsigned long)val;
2561                 } else {
2562                         *negp = false;
2563                         lval = (unsigned long)val;
2564                 }
2565                 *lvalp = jiffies_to_msecs(lval);
2566         }
2567         return 0;
2568 }
2569
2570 /**
2571  * proc_dointvec_jiffies - read a vector of integers as seconds
2572  * @table: the sysctl table
2573  * @write: %TRUE if this is a write to the sysctl file
2574  * @buffer: the user buffer
2575  * @lenp: the size of the user buffer
2576  * @ppos: file position
2577  *
2578  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2579  * values from/to the user buffer, treated as an ASCII string.
2580  * The values read are assumed to be in seconds, and are converted into
2581  * jiffies.
2582  *
2583  * Returns 0 on success.
2584  */
2585 int proc_dointvec_jiffies(struct ctl_table *table, int write,
2586                           void __user *buffer, size_t *lenp, loff_t *ppos)
2587 {
2588     return do_proc_dointvec(table,write,buffer,lenp,ppos,
2589                             do_proc_dointvec_jiffies_conv,NULL);
2590 }
2591
2592 /**
2593  * proc_dointvec_userhz_jiffies - read a vector of integers as 1/USER_HZ seconds
2594  * @table: the sysctl table
2595  * @write: %TRUE if this is a write to the sysctl file
2596  * @buffer: the user buffer
2597  * @lenp: the size of the user buffer
2598  * @ppos: pointer to the file position
2599  *
2600  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
2601  * values from/to the user buffer, treated as an ASCII string.
2602  * The values read are assumed to be in 1/USER_HZ seconds, and
2603  * are converted into jiffies.
2604  *
2605  * Returns 0 on success.
2606  */
2607 int proc_dointvec_userhz_jiffies(struct ctl_table *table, int write,
2608                                  void __user *buffer, size_t *lenp, loff_t *ppos)
2609 {
2610     return do_proc_dointvec(table,write,buffer,lenp,ppos,
2611                             do_proc_dointvec_userhz_jiffies_conv,NULL);
2612 }
2613
2614 /**
2615  * proc_dointvec_ms_jiffies - read a vector of integers as 1 milliseconds
2616  * @table: the sysctl table
2617  * @write: %TRUE if this is a write to the sysctl file
2618  * @buffer: the user buffer
2619  * @lenp: the size of the user buffer
2620  * @ppos: file position
2621  * @ppos: the current position in the file
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 1/1000 seconds, and
2626  * are converted into jiffies.
2627  *
2628  * Returns 0 on success.
2629  */
2630 int proc_dointvec_ms_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_ms_jiffies_conv, NULL);
2635 }
2636
2637 static int proc_do_cad_pid(struct ctl_table *table, int write,
2638                            void __user *buffer, size_t *lenp, loff_t *ppos)
2639 {
2640         struct pid *new_pid;
2641         pid_t tmp;
2642         int r;
2643
2644         tmp = pid_vnr(cad_pid);
2645
2646         r = __do_proc_dointvec(&tmp, table, write, buffer,
2647                                lenp, ppos, NULL, NULL);
2648         if (r || !write)
2649                 return r;
2650
2651         new_pid = find_get_pid(tmp);
2652         if (!new_pid)
2653                 return -ESRCH;
2654
2655         put_pid(xchg(&cad_pid, new_pid));
2656         return 0;
2657 }
2658
2659 /**
2660  * proc_do_large_bitmap - read/write from/to a large bitmap
2661  * @table: the sysctl table
2662  * @write: %TRUE if this is a write to the sysctl file
2663  * @buffer: the user buffer
2664  * @lenp: the size of the user buffer
2665  * @ppos: file position
2666  *
2667  * The bitmap is stored at table->data and the bitmap length (in bits)
2668  * in table->maxlen.
2669  *
2670  * We use a range comma separated format (e.g. 1,3-4,10-10) so that
2671  * large bitmaps may be represented in a compact manner. Writing into
2672  * the file will clear the bitmap then update it with the given input.
2673  *
2674  * Returns 0 on success.
2675  */
2676 int proc_do_large_bitmap(struct ctl_table *table, int write,
2677                          void __user *buffer, size_t *lenp, loff_t *ppos)
2678 {
2679         int err = 0;
2680         bool first = 1;
2681         size_t left = *lenp;
2682         unsigned long bitmap_len = table->maxlen;
2683         unsigned long *bitmap = *(unsigned long **) table->data;
2684         unsigned long *tmp_bitmap = NULL;
2685         char tr_a[] = { '-', ',', '\n' }, tr_b[] = { ',', '\n', 0 }, c;
2686
2687         if (!bitmap || !bitmap_len || !left || (*ppos && !write)) {
2688                 *lenp = 0;
2689                 return 0;
2690         }
2691
2692         if (write) {
2693                 char *kbuf, *p;
2694
2695                 if (left > PAGE_SIZE - 1)
2696                         left = PAGE_SIZE - 1;
2697
2698                 p = kbuf = memdup_user_nul(buffer, left);
2699                 if (IS_ERR(kbuf))
2700                         return PTR_ERR(kbuf);
2701
2702                 tmp_bitmap = kzalloc(BITS_TO_LONGS(bitmap_len) * sizeof(unsigned long),
2703                                      GFP_KERNEL);
2704                 if (!tmp_bitmap) {
2705                         kfree(kbuf);
2706                         return -ENOMEM;
2707                 }
2708                 proc_skip_char(&p, &left, '\n');
2709                 while (!err && left) {
2710                         unsigned long val_a, val_b;
2711                         bool neg;
2712
2713                         err = proc_get_long(&p, &left, &val_a, &neg, tr_a,
2714                                              sizeof(tr_a), &c);
2715                         if (err)
2716                                 break;
2717                         if (val_a >= bitmap_len || neg) {
2718                                 err = -EINVAL;
2719                                 break;
2720                         }
2721
2722                         val_b = val_a;
2723                         if (left) {
2724                                 p++;
2725                                 left--;
2726                         }
2727
2728                         if (c == '-') {
2729                                 err = proc_get_long(&p, &left, &val_b,
2730                                                      &neg, tr_b, sizeof(tr_b),
2731                                                      &c);
2732                                 if (err)
2733                                         break;
2734                                 if (val_b >= bitmap_len || neg ||
2735                                     val_a > val_b) {
2736                                         err = -EINVAL;
2737                                         break;
2738                                 }
2739                                 if (left) {
2740                                         p++;
2741                                         left--;
2742                                 }
2743                         }
2744
2745                         bitmap_set(tmp_bitmap, val_a, val_b - val_a + 1);
2746                         first = 0;
2747                         proc_skip_char(&p, &left, '\n');
2748                 }
2749                 kfree(kbuf);
2750         } else {
2751                 unsigned long bit_a, bit_b = 0;
2752
2753                 while (left) {
2754                         bit_a = find_next_bit(bitmap, bitmap_len, bit_b);
2755                         if (bit_a >= bitmap_len)
2756                                 break;
2757                         bit_b = find_next_zero_bit(bitmap, bitmap_len,
2758                                                    bit_a + 1) - 1;
2759
2760                         if (!first) {
2761                                 err = proc_put_char(&buffer, &left, ',');
2762                                 if (err)
2763                                         break;
2764                         }
2765                         err = proc_put_long(&buffer, &left, bit_a, false);
2766                         if (err)
2767                                 break;
2768                         if (bit_a != bit_b) {
2769                                 err = proc_put_char(&buffer, &left, '-');
2770                                 if (err)
2771                                         break;
2772                                 err = proc_put_long(&buffer, &left, bit_b, false);
2773                                 if (err)
2774                                         break;
2775                         }
2776
2777                         first = 0; bit_b++;
2778                 }
2779                 if (!err)
2780                         err = proc_put_char(&buffer, &left, '\n');
2781         }
2782
2783         if (!err) {
2784                 if (write) {
2785                         if (*ppos)
2786                                 bitmap_or(bitmap, bitmap, tmp_bitmap, bitmap_len);
2787                         else
2788                                 bitmap_copy(bitmap, tmp_bitmap, bitmap_len);
2789                 }
2790                 kfree(tmp_bitmap);
2791                 *lenp -= left;
2792                 *ppos += *lenp;
2793                 return 0;
2794         } else {
2795                 kfree(tmp_bitmap);
2796                 return err;
2797         }
2798 }
2799
2800 #else /* CONFIG_PROC_SYSCTL */
2801
2802 int proc_dostring(struct ctl_table *table, int write,
2803                   void __user *buffer, size_t *lenp, loff_t *ppos)
2804 {
2805         return -ENOSYS;
2806 }
2807
2808 int proc_dointvec(struct ctl_table *table, int write,
2809                   void __user *buffer, size_t *lenp, loff_t *ppos)
2810 {
2811         return -ENOSYS;
2812 }
2813
2814 int proc_dointvec_minmax(struct ctl_table *table, int write,
2815                     void __user *buffer, size_t *lenp, loff_t *ppos)
2816 {
2817         return -ENOSYS;
2818 }
2819
2820 int proc_dointvec_jiffies(struct ctl_table *table, int write,
2821                     void __user *buffer, size_t *lenp, loff_t *ppos)
2822 {
2823         return -ENOSYS;
2824 }
2825
2826 int proc_dointvec_userhz_jiffies(struct ctl_table *table, int write,
2827                     void __user *buffer, size_t *lenp, loff_t *ppos)
2828 {
2829         return -ENOSYS;
2830 }
2831
2832 int proc_dointvec_ms_jiffies(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_doulongvec_minmax(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_doulongvec_ms_jiffies_minmax(struct ctl_table *table, int write,
2845                                       void __user *buffer,
2846                                       size_t *lenp, loff_t *ppos)
2847 {
2848     return -ENOSYS;
2849 }
2850
2851
2852 #endif /* CONFIG_PROC_SYSCTL */
2853
2854 /*
2855  * No sense putting this after each symbol definition, twice,
2856  * exception granted :-)
2857  */
2858 EXPORT_SYMBOL(proc_dointvec);
2859 EXPORT_SYMBOL(proc_dointvec_jiffies);
2860 EXPORT_SYMBOL(proc_dointvec_minmax);
2861 EXPORT_SYMBOL(proc_dointvec_userhz_jiffies);
2862 EXPORT_SYMBOL(proc_dointvec_ms_jiffies);
2863 EXPORT_SYMBOL(proc_dostring);
2864 EXPORT_SYMBOL(proc_doulongvec_minmax);
2865 EXPORT_SYMBOL(proc_doulongvec_ms_jiffies_minmax);