kernel/sysctl.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * sysctl.c: General linux system control interface
   4  *
   5  * Begun 24 March 1995, Stephen Tweedie
   6  * Added /proc support, Dec 1995
   7  * Added bdflush entry and intvec min/max checking, 2/23/96, Tom Dyas.
   8  * Added hooks for /proc/sys/net (minor, minor patch), 96/4/1, Mike Shaver.
   9  * Added kernel/java-{interpreter,appletviewer}, 96/5/10, Mike Shaver.
  10  * Dynamic registration fixes, Stephen Tweedie.
  11  * Added kswapd-interval, ctrl-alt-del, printk stuff, 1/8/97, Chris Horn.
  12  * Made sysctl support optional via CONFIG_SYSCTL, 1/10/97, Chris
  13  *  Horn.
  14  * Added proc_doulongvec_ms_jiffies_minmax, 09/08/99, Carlos H. Bauer.
  15  * Added proc_doulongvec_minmax, 09/08/99, Carlos H. Bauer.
  16  * Changed linked lists to use list.h instead of lists.h, 02/24/00, Bill
  17  *  Wendling.
  18  * The list_for_each() macro wasn't appropriate for the sysctl loop.
  19  *  Removed it and replaced it with older style, 03/23/00, Bill Wendling
  20  */
  21
  22 #include <linux/module.h>
  23 #include <linux/aio.h>
  24 #include <linux/mm.h>
  25 #include <linux/swap.h>
  26 #include <linux/slab.h>
  27 #include <linux/sysctl.h>
  28 #include <linux/bitmap.h>
  29 #include <linux/signal.h>
  30 #include <linux/printk.h>
  31 #include <linux/proc_fs.h>
  32 #include <linux/security.h>
  33 #include <linux/ctype.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/sched/coredump.h>
  67 #include <linux/kexec.h>
  68 #include <linux/bpf.h>
  69 #include <linux/mount.h>
  70 #include <linux/userfaultfd_k.h>
  71 #include <linux/coredump.h>
  72 #include <linux/latencytop.h>
  73 #include <linux/pid.h>
  74 #include <linux/delayacct.h>
  75
  76 #include "../lib/kstrtox.h"
  77
  78 #include <linux/uaccess.h>
  79 #include <asm/processor.h>
  80
  81 #ifdef CONFIG_X86
  82 #include <asm/nmi.h>
  83 #include <asm/stacktrace.h>
  84 #include <asm/io.h>
  85 #endif
  86 #ifdef CONFIG_SPARC
  87 #include <asm/setup.h>
  88 #endif
  89 #ifdef CONFIG_BSD_PROCESS_ACCT
  90 #include <linux/acct.h>
  91 #endif
  92 #ifdef CONFIG_RT_MUTEXES
  93 #include <linux/rtmutex.h>
  94 #endif
  95 #if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_LOCK_STAT)
  96 #include <linux/lockdep.h>
  97 #endif
  98 #ifdef CONFIG_CHR_DEV_SG
  99 #include <scsi/sg.h>
 100 #endif
 101 #ifdef CONFIG_STACKLEAK_RUNTIME_DISABLE
 102 #include <linux/stackleak.h>
 103 #endif
 104 #ifdef CONFIG_LOCKUP_DETECTOR
 105 #include <linux/nmi.h>
 106 #endif
 107
 108 #if defined(CONFIG_SYSCTL)
 109
 110 /* Constants used for minimum and  maximum */
 111 #ifdef CONFIG_LOCKUP_DETECTOR
 112 static int sixty = 60;
 113 #endif
 114
 115 static int __maybe_unused neg_one = -1;
 116 static int __maybe_unused two = 2;
 117 static int __maybe_unused four = 4;
 118 static unsigned long zero_ul;
 119 static unsigned long one_ul = 1;
 120 static unsigned long long_max = LONG_MAX;
 121 static int one_hundred = 100;
 122 static int two_hundred = 200;
 123 static int one_thousand = 1000;
 124 #ifdef CONFIG_PRINTK
 125 static int ten_thousand = 10000;
 126 #endif
 127 #ifdef CONFIG_PERF_EVENTS
 128 static int six_hundred_forty_kb = 640 * 1024;
 129 #endif
 130
 131 /* this is needed for the proc_doulongvec_minmax of vm_dirty_bytes */
 132 static unsigned long dirty_bytes_min = 2 * PAGE_SIZE;
 133
 134 /* this is needed for the proc_dointvec_minmax for [fs_]overflow UID and GID */
 135 static int maxolduid = 65535;
 136 static int minolduid;
 137
 138 static int ngroups_max = NGROUPS_MAX;
 139 static const int cap_last_cap = CAP_LAST_CAP;
 140
 141 /*
 142  * This is needed for proc_doulongvec_minmax of sysctl_hung_task_timeout_secs
 143  * and hung_task_check_interval_secs
 144  */
 145 #ifdef CONFIG_DETECT_HUNG_TASK
 146 static unsigned long hung_task_timeout_max = (LONG_MAX/HZ);
 147 #endif
 148
 149 #ifdef CONFIG_INOTIFY_USER
 150 #include <linux/inotify.h>
 151 #endif
 152 #ifdef CONFIG_FANOTIFY
 153 #include <linux/fanotify.h>
 154 #endif
 155
 156 #ifdef CONFIG_PROC_SYSCTL
 157
 158 /**
 159  * enum sysctl_writes_mode - supported sysctl write modes
 160  *
 161  * @SYSCTL_WRITES_LEGACY: each write syscall must fully contain the sysctl value
 162  *      to be written, and multiple writes on the same sysctl file descriptor
 163  *      will rewrite the sysctl value, regardless of file position. No warning
 164  *      is issued when the initial position is not 0.
 165  * @SYSCTL_WRITES_WARN: same as above but warn when the initial file position is
 166  *      not 0.
 167  * @SYSCTL_WRITES_STRICT: writes to numeric sysctl entries must always be at
 168  *      file position 0 and the value must be fully contained in the buffer
 169  *      sent to the write syscall. If dealing with strings respect the file
 170  *      position, but restrict this to the max length of the buffer, anything
 171  *      passed the max length will be ignored. Multiple writes will append
 172  *      to the buffer.
 173  *
 174  * These write modes control how current file position affects the behavior of
 175  * updating sysctl values through the proc interface on each write.
 176  */
 177 enum sysctl_writes_mode {
 178         SYSCTL_WRITES_LEGACY            = -1,
 179         SYSCTL_WRITES_WARN              = 0,
 180         SYSCTL_WRITES_STRICT            = 1,
 181 };
 182
 183 static enum sysctl_writes_mode sysctl_writes_strict = SYSCTL_WRITES_STRICT;
 184 #endif /* CONFIG_PROC_SYSCTL */
 185
 186 #if defined(HAVE_ARCH_PICK_MMAP_LAYOUT) || \
 187     defined(CONFIG_ARCH_WANT_DEFAULT_TOPDOWN_MMAP_LAYOUT)
 188 int sysctl_legacy_va_layout;
 189 #endif
 190
 191 #ifdef CONFIG_COMPACTION
 192 static int min_extfrag_threshold;
 193 static int max_extfrag_threshold = 1000;
 194 #endif
 195
 196 #endif /* CONFIG_SYSCTL */
 197
 198 #if defined(CONFIG_BPF_SYSCALL) && defined(CONFIG_SYSCTL)
 199 static int bpf_stats_handler(struct ctl_table *table, int write,
 200                              void *buffer, size_t *lenp, loff_t *ppos)
 201 {
 202         struct static_key *key = (struct static_key *)table->data;
 203         static int saved_val;
 204         int val, ret;
 205         struct ctl_table tmp = {
 206                 .data   = &val,
 207                 .maxlen = sizeof(val),
 208                 .mode   = table->mode,
 209                 .extra1 = SYSCTL_ZERO,
 210                 .extra2 = SYSCTL_ONE,
 211         };
 212
 213         if (write && !capable(CAP_SYS_ADMIN))
 214                 return -EPERM;
 215
 216         mutex_lock(&bpf_stats_enabled_mutex);
 217         val = saved_val;
 218         ret = proc_dointvec_minmax(&tmp, write, buffer, lenp, ppos);
 219         if (write && !ret && val != saved_val) {
 220                 if (val)
 221                         static_key_slow_inc(key);
 222                 else
 223                         static_key_slow_dec(key);
 224                 saved_val = val;
 225         }
 226         mutex_unlock(&bpf_stats_enabled_mutex);
 227         return ret;
 228 }
 229
 230 static int bpf_unpriv_handler(struct ctl_table *table, int write,
 231                               void *buffer, size_t *lenp, loff_t *ppos)
 232 {
 233         int ret, unpriv_enable = *(int *)table->data;
 234         bool locked_state = unpriv_enable == 1;
 235         struct ctl_table tmp = *table;
 236
 237         if (write && !capable(CAP_SYS_ADMIN))
 238                 return -EPERM;
 239
 240         tmp.data = &unpriv_enable;
 241         ret = proc_dointvec_minmax(&tmp, write, buffer, lenp, ppos);
 242         if (write && !ret) {
 243                 if (locked_state && unpriv_enable != 1)
 244                         return -EPERM;
 245                 *(int *)table->data = unpriv_enable;
 246         }
 247         return ret;
 248 }
 249 #endif /* CONFIG_BPF_SYSCALL && CONFIG_SYSCTL */
 250
 251 /*
 252  * /proc/sys support
 253  */
 254
 255 #ifdef CONFIG_PROC_SYSCTL
 256
 257 static int _proc_do_string(char *data, int maxlen, int write,
 258                 char *buffer, size_t *lenp, loff_t *ppos)
 259 {
 260         size_t len;
 261         char c, *p;
 262
 263         if (!data || !maxlen || !*lenp) {
 264                 *lenp = 0;
 265                 return 0;
 266         }
 267
 268         if (write) {
 269                 if (sysctl_writes_strict == SYSCTL_WRITES_STRICT) {
 270                         /* Only continue writes not past the end of buffer. */
 271                         len = strlen(data);
 272                         if (len > maxlen - 1)
 273                                 len = maxlen - 1;
 274
 275                         if (*ppos > len)
 276                                 return 0;
 277                         len = *ppos;
 278                 } else {
 279                         /* Start writing from beginning of buffer. */
 280                         len = 0;
 281                 }
 282
 283                 *ppos += *lenp;
 284                 p = buffer;
 285                 while ((p - buffer) < *lenp && len < maxlen - 1) {
 286                         c = *(p++);
 287                         if (c == 0 || c == '\n')
 288                                 break;
 289                         data[len++] = c;
 290                 }
 291                 data[len] = 0;
 292         } else {
 293                 len = strlen(data);
 294                 if (len > maxlen)
 295                         len = maxlen;
 296
 297                 if (*ppos > len) {
 298                         *lenp = 0;
 299                         return 0;
 300                 }
 301
 302                 data += *ppos;
 303                 len  -= *ppos;
 304
 305                 if (len > *lenp)
 306                         len = *lenp;
 307                 if (len)
 308                         memcpy(buffer, data, len);
 309                 if (len < *lenp) {
 310                         buffer[len] = '\n';
 311                         len++;
 312                 }
 313                 *lenp = len;
 314                 *ppos += len;
 315         }
 316         return 0;
 317 }
 318
 319 static void warn_sysctl_write(struct ctl_table *table)
 320 {
 321         pr_warn_once("%s wrote to %s when file position was not 0!\n"
 322                 "This will not be supported in the future. To silence this\n"
 323                 "warning, set kernel.sysctl_writes_strict = -1\n",
 324                 current->comm, table->procname);
 325 }
 326
 327 /**
 328  * proc_first_pos_non_zero_ignore - check if first position is allowed
 329  * @ppos: file position
 330  * @table: the sysctl table
 331  *
 332  * Returns true if the first position is non-zero and the sysctl_writes_strict
 333  * mode indicates this is not allowed for numeric input types. String proc
 334  * handlers can ignore the return value.
 335  */
 336 static bool proc_first_pos_non_zero_ignore(loff_t *ppos,
 337                                            struct ctl_table *table)
 338 {
 339         if (!*ppos)
 340                 return false;
 341
 342         switch (sysctl_writes_strict) {
 343         case SYSCTL_WRITES_STRICT:
 344                 return true;
 345         case SYSCTL_WRITES_WARN:
 346                 warn_sysctl_write(table);
 347                 return false;
 348         default:
 349                 return false;
 350         }
 351 }
 352
 353 /**
 354  * proc_dostring - read a string sysctl
 355  * @table: the sysctl table
 356  * @write: %TRUE if this is a write to the sysctl file
 357  * @buffer: the user buffer
 358  * @lenp: the size of the user buffer
 359  * @ppos: file position
 360  *
 361  * Reads/writes a string from/to the user buffer. If the kernel
 362  * buffer provided is not large enough to hold the string, the
 363  * string is truncated. The copied string is %NULL-terminated.
 364  * If the string is being read by the user process, it is copied
 365  * and a newline '\n' is added. It is truncated if the buffer is
 366  * not large enough.
 367  *
 368  * Returns 0 on success.
 369  */
 370 int proc_dostring(struct ctl_table *table, int write,
 371                   void *buffer, size_t *lenp, loff_t *ppos)
 372 {
 373         if (write)
 374                 proc_first_pos_non_zero_ignore(ppos, table);
 375
 376         return _proc_do_string(table->data, table->maxlen, write, buffer, lenp,
 377                         ppos);
 378 }
 379
 380 static size_t proc_skip_spaces(char **buf)
 381 {
 382         size_t ret;
 383         char *tmp = skip_spaces(*buf);
 384         ret = tmp - *buf;
 385         *buf = tmp;
 386         return ret;
 387 }
 388
 389 static void proc_skip_char(char **buf, size_t *size, const char v)
 390 {
 391         while (*size) {
 392                 if (**buf != v)
 393                         break;
 394                 (*size)--;
 395                 (*buf)++;
 396         }
 397 }
 398
 399 /**
 400  * strtoul_lenient - parse an ASCII formatted integer from a buffer and only
 401  *                   fail on overflow
 402  *
 403  * @cp: kernel buffer containing the string to parse
 404  * @endp: pointer to store the trailing characters
 405  * @base: the base to use
 406  * @res: where the parsed integer will be stored
 407  *
 408  * In case of success 0 is returned and @res will contain the parsed integer,
 409  * @endp will hold any trailing characters.
 410  * This function will fail the parse on overflow. If there wasn't an overflow
 411  * the function will defer the decision what characters count as invalid to the
 412  * caller.
 413  */
 414 static int strtoul_lenient(const char *cp, char **endp, unsigned int base,
 415                            unsigned long *res)
 416 {
 417         unsigned long long result;
 418         unsigned int rv;
 419
 420         cp = _parse_integer_fixup_radix(cp, &base);
 421         rv = _parse_integer(cp, base, &result);
 422         if ((rv & KSTRTOX_OVERFLOW) || (result != (unsigned long)result))
 423                 return -ERANGE;
 424
 425         cp += rv;
 426
 427         if (endp)
 428                 *endp = (char *)cp;
 429
 430         *res = (unsigned long)result;
 431         return 0;
 432 }
 433
 434 #define TMPBUFLEN 22
 435 /**
 436  * proc_get_long - reads an ASCII formatted integer from a user buffer
 437  *
 438  * @buf: a kernel buffer
 439  * @size: size of the kernel buffer
 440  * @val: this is where the number will be stored
 441  * @neg: set to %TRUE if number is negative
 442  * @perm_tr: a vector which contains the allowed trailers
 443  * @perm_tr_len: size of the perm_tr vector
 444  * @tr: pointer to store the trailer character
 445  *
 446  * In case of success %0 is returned and @buf and @size are updated with
 447  * the amount of bytes read. If @tr is non-NULL and a trailing
 448  * character exists (size is non-zero after returning from this
 449  * function), @tr is updated with the trailing character.
 450  */
 451 static int proc_get_long(char **buf, size_t *size,
 452                           unsigned long *val, bool *neg,
 453                           const char *perm_tr, unsigned perm_tr_len, char *tr)
 454 {
 455         int len;
 456         char *p, tmp[TMPBUFLEN];
 457
 458         if (!*size)
 459                 return -EINVAL;
 460
 461         len = *size;
 462         if (len > TMPBUFLEN - 1)
 463                 len = TMPBUFLEN - 1;
 464
 465         memcpy(tmp, *buf, len);
 466
 467         tmp[len] = 0;
 468         p = tmp;
 469         if (*p == '-' && *size > 1) {
 470                 *neg = true;
 471                 p++;
 472         } else
 473                 *neg = false;
 474         if (!isdigit(*p))
 475                 return -EINVAL;
 476
 477         if (strtoul_lenient(p, &p, 0, val))
 478                 return -EINVAL;
 479
 480         len = p - tmp;
 481
 482         /* We don't know if the next char is whitespace thus we may accept
 483          * invalid integers (e.g. 1234...a) or two integers instead of one
 484          * (e.g. 123...1). So lets not allow such large numbers. */
 485         if (len == TMPBUFLEN - 1)
 486                 return -EINVAL;
 487
 488         if (len < *size && perm_tr_len && !memchr(perm_tr, *p, perm_tr_len))
 489                 return -EINVAL;
 490
 491         if (tr && (len < *size))
 492                 *tr = *p;
 493
 494         *buf += len;
 495         *size -= len;
 496
 497         return 0;
 498 }
 499
 500 /**
 501  * proc_put_long - converts an integer to a decimal ASCII formatted string
 502  *
 503  * @buf: the user buffer
 504  * @size: the size of the user buffer
 505  * @val: the integer to be converted
 506  * @neg: sign of the number, %TRUE for negative
 507  *
 508  * In case of success @buf and @size are updated with the amount of bytes
 509  * written.
 510  */
 511 static void proc_put_long(void **buf, size_t *size, unsigned long val, bool neg)
 512 {
 513         int len;
 514         char tmp[TMPBUFLEN], *p = tmp;
 515
 516         sprintf(p, "%s%lu", neg ? "-" : "", val);
 517         len = strlen(tmp);
 518         if (len > *size)
 519                 len = *size;
 520         memcpy(*buf, tmp, len);
 521         *size -= len;
 522         *buf += len;
 523 }
 524 #undef TMPBUFLEN
 525
 526 static void proc_put_char(void **buf, size_t *size, char c)
 527 {
 528         if (*size) {
 529                 char **buffer = (char **)buf;
 530                 **buffer = c;
 531
 532                 (*size)--;
 533                 (*buffer)++;
 534                 *buf = *buffer;
 535         }
 536 }
 537
 538 static int do_proc_dointvec_conv(bool *negp, unsigned long *lvalp,
 539                                  int *valp,
 540                                  int write, void *data)
 541 {
 542         if (write) {
 543                 if (*negp) {
 544                         if (*lvalp > (unsigned long) INT_MAX + 1)
 545                                 return -EINVAL;
 546                         *valp = -*lvalp;
 547                 } else {
 548                         if (*lvalp > (unsigned long) INT_MAX)
 549                                 return -EINVAL;
 550                         *valp = *lvalp;
 551                 }
 552         } else {
 553                 int val = *valp;
 554                 if (val < 0) {
 555                         *negp = true;
 556                         *lvalp = -(unsigned long)val;
 557                 } else {
 558                         *negp = false;
 559                         *lvalp = (unsigned long)val;
 560                 }
 561         }
 562         return 0;
 563 }
 564
 565 static int do_proc_douintvec_conv(unsigned long *lvalp,
 566                                   unsigned int *valp,
 567                                   int write, void *data)
 568 {
 569         if (write) {
 570                 if (*lvalp > UINT_MAX)
 571                         return -EINVAL;
 572                 *valp = *lvalp;
 573         } else {
 574                 unsigned int val = *valp;
 575                 *lvalp = (unsigned long)val;
 576         }
 577         return 0;
 578 }
 579
 580 static const char proc_wspace_sep[] = { ' ', '\t', '\n' };
 581
 582 static int __do_proc_dointvec(void *tbl_data, struct ctl_table *table,
 583                   int write, void *buffer,
 584                   size_t *lenp, loff_t *ppos,
 585                   int (*conv)(bool *negp, unsigned long *lvalp, int *valp,
 586                               int write, void *data),
 587                   void *data)
 588 {
 589         int *i, vleft, first = 1, err = 0;
 590         size_t left;
 591         char *p;
 592
 593         if (!tbl_data || !table->maxlen || !*lenp || (*ppos && !write)) {
 594                 *lenp = 0;
 595                 return 0;
 596         }
 597
 598         i = (int *) tbl_data;
 599         vleft = table->maxlen / sizeof(*i);
 600         left = *lenp;
 601
 602         if (!conv)
 603                 conv = do_proc_dointvec_conv;
 604
 605         if (write) {
 606                 if (proc_first_pos_non_zero_ignore(ppos, table))
 607                         goto out;
 608
 609                 if (left > PAGE_SIZE - 1)
 610                         left = PAGE_SIZE - 1;
 611                 p = buffer;
 612         }
 613
 614         for (; left && vleft--; i++, first=0) {
 615                 unsigned long lval;
 616                 bool neg;
 617
 618                 if (write) {
 619                         left -= proc_skip_spaces(&p);
 620
 621                         if (!left)
 622                                 break;
 623                         err = proc_get_long(&p, &left, &lval, &neg,
 624                                              proc_wspace_sep,
 625                                              sizeof(proc_wspace_sep), NULL);
 626                         if (err)
 627                                 break;
 628                         if (conv(&neg, &lval, i, 1, data)) {
 629                                 err = -EINVAL;
 630                                 break;
 631                         }
 632                 } else {
 633                         if (conv(&neg, &lval, i, 0, data)) {
 634                                 err = -EINVAL;
 635                                 break;
 636                         }
 637                         if (!first)
 638                                 proc_put_char(&buffer, &left, '\t');
 639                         proc_put_long(&buffer, &left, lval, neg);
 640                 }
 641         }
 642
 643         if (!write && !first && left && !err)
 644                 proc_put_char(&buffer, &left, '\n');
 645         if (write && !err && left)
 646                 left -= proc_skip_spaces(&p);
 647         if (write && first)
 648                 return err ? : -EINVAL;
 649         *lenp -= left;
 650 out:
 651         *ppos += *lenp;
 652         return err;
 653 }
 654
 655 static int do_proc_dointvec(struct ctl_table *table, int write,
 656                   void *buffer, size_t *lenp, loff_t *ppos,
 657                   int (*conv)(bool *negp, unsigned long *lvalp, int *valp,
 658                               int write, void *data),
 659                   void *data)
 660 {
 661         return __do_proc_dointvec(table->data, table, write,
 662                         buffer, lenp, ppos, conv, data);
 663 }
 664
 665 static int do_proc_douintvec_w(unsigned int *tbl_data,
 666                                struct ctl_table *table,
 667                                void *buffer,
 668                                size_t *lenp, loff_t *ppos,
 669                                int (*conv)(unsigned long *lvalp,
 670                                            unsigned int *valp,
 671                                            int write, void *data),
 672                                void *data)
 673 {
 674         unsigned long lval;
 675         int err = 0;
 676         size_t left;
 677         bool neg;
 678         char *p = buffer;
 679
 680         left = *lenp;
 681
 682         if (proc_first_pos_non_zero_ignore(ppos, table))
 683                 goto bail_early;
 684
 685         if (left > PAGE_SIZE - 1)
 686                 left = PAGE_SIZE - 1;
 687
 688         left -= proc_skip_spaces(&p);
 689         if (!left) {
 690                 err = -EINVAL;
 691                 goto out_free;
 692         }
 693
 694         err = proc_get_long(&p, &left, &lval, &neg,
 695                              proc_wspace_sep,
 696                              sizeof(proc_wspace_sep), NULL);
 697         if (err || neg) {
 698                 err = -EINVAL;
 699                 goto out_free;
 700         }
 701
 702         if (conv(&lval, tbl_data, 1, data)) {
 703                 err = -EINVAL;
 704                 goto out_free;
 705         }
 706
 707         if (!err && left)
 708                 left -= proc_skip_spaces(&p);
 709
 710 out_free:
 711         if (err)
 712                 return -EINVAL;
 713
 714         return 0;
 715
 716         /* This is in keeping with old __do_proc_dointvec() */
 717 bail_early:
 718         *ppos += *lenp;
 719         return err;
 720 }
 721
 722 static int do_proc_douintvec_r(unsigned int *tbl_data, void *buffer,
 723                                size_t *lenp, loff_t *ppos,
 724                                int (*conv)(unsigned long *lvalp,
 725                                            unsigned int *valp,
 726                                            int write, void *data),
 727                                void *data)
 728 {
 729         unsigned long lval;
 730         int err = 0;
 731         size_t left;
 732
 733         left = *lenp;
 734
 735         if (conv(&lval, tbl_data, 0, data)) {
 736                 err = -EINVAL;
 737                 goto out;
 738         }
 739
 740         proc_put_long(&buffer, &left, lval, false);
 741         if (!left)
 742                 goto out;
 743
 744         proc_put_char(&buffer, &left, '\n');
 745
 746 out:
 747         *lenp -= left;
 748         *ppos += *lenp;
 749
 750         return err;
 751 }
 752
 753 static int __do_proc_douintvec(void *tbl_data, struct ctl_table *table,
 754                                int write, void *buffer,
 755                                size_t *lenp, loff_t *ppos,
 756                                int (*conv)(unsigned long *lvalp,
 757                                            unsigned int *valp,
 758                                            int write, void *data),
 759                                void *data)
 760 {
 761         unsigned int *i, vleft;
 762
 763         if (!tbl_data || !table->maxlen || !*lenp || (*ppos && !write)) {
 764                 *lenp = 0;
 765                 return 0;
 766         }
 767
 768         i = (unsigned int *) tbl_data;
 769         vleft = table->maxlen / sizeof(*i);
 770
 771         /*
 772          * Arrays are not supported, keep this simple. *Do not* add
 773          * support for them.
 774          */
 775         if (vleft != 1) {
 776                 *lenp = 0;
 777                 return -EINVAL;
 778         }
 779
 780         if (!conv)
 781                 conv = do_proc_douintvec_conv;
 782
 783         if (write)
 784                 return do_proc_douintvec_w(i, table, buffer, lenp, ppos,
 785                                            conv, data);
 786         return do_proc_douintvec_r(i, buffer, lenp, ppos, conv, data);
 787 }
 788
 789 static int do_proc_douintvec(struct ctl_table *table, int write,
 790                              void *buffer, size_t *lenp, loff_t *ppos,
 791                              int (*conv)(unsigned long *lvalp,
 792                                          unsigned int *valp,
 793                                          int write, void *data),
 794                              void *data)
 795 {
 796         return __do_proc_douintvec(table->data, table, write,
 797                                    buffer, lenp, ppos, conv, data);
 798 }
 799
 800 /**
 801  * proc_dointvec - read a vector of integers
 802  * @table: the sysctl table
 803  * @write: %TRUE if this is a write to the sysctl file
 804  * @buffer: the user buffer
 805  * @lenp: the size of the user buffer
 806  * @ppos: file position
 807  *
 808  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
 809  * values from/to the user buffer, treated as an ASCII string.
 810  *
 811  * Returns 0 on success.
 812  */
 813 int proc_dointvec(struct ctl_table *table, int write, void *buffer,
 814                   size_t *lenp, loff_t *ppos)
 815 {
 816         return do_proc_dointvec(table, write, buffer, lenp, ppos, NULL, NULL);
 817 }
 818
 819 #ifdef CONFIG_COMPACTION
 820 static int proc_dointvec_minmax_warn_RT_change(struct ctl_table *table,
 821                 int write, void *buffer, size_t *lenp, loff_t *ppos)
 822 {
 823         int ret, old;
 824
 825         if (!IS_ENABLED(CONFIG_PREEMPT_RT) || !write)
 826                 return proc_dointvec_minmax(table, write, buffer, lenp, ppos);
 827
 828         old = *(int *)table->data;
 829         ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
 830         if (ret)
 831                 return ret;
 832         if (old != *(int *)table->data)
 833                 pr_warn_once("sysctl attribute %s changed by %s[%d]\n",
 834                              table->procname, current->comm,
 835                              task_pid_nr(current));
 836         return ret;
 837 }
 838 #endif
 839
 840 /**
 841  * proc_douintvec - read a vector of unsigned integers
 842  * @table: the sysctl table
 843  * @write: %TRUE if this is a write to the sysctl file
 844  * @buffer: the user buffer
 845  * @lenp: the size of the user buffer
 846  * @ppos: file position
 847  *
 848  * Reads/writes up to table->maxlen/sizeof(unsigned int) unsigned integer
 849  * values from/to the user buffer, treated as an ASCII string.
 850  *
 851  * Returns 0 on success.
 852  */
 853 int proc_douintvec(struct ctl_table *table, int write, void *buffer,
 854                 size_t *lenp, loff_t *ppos)
 855 {
 856         return do_proc_douintvec(table, write, buffer, lenp, ppos,
 857                                  do_proc_douintvec_conv, NULL);
 858 }
 859
 860 /*
 861  * Taint values can only be increased
 862  * This means we can safely use a temporary.
 863  */
 864 static int proc_taint(struct ctl_table *table, int write,
 865                                void *buffer, size_t *lenp, loff_t *ppos)
 866 {
 867         struct ctl_table t;
 868         unsigned long tmptaint = get_taint();
 869         int err;
 870
 871         if (write && !capable(CAP_SYS_ADMIN))
 872                 return -EPERM;
 873
 874         t = *table;
 875         t.data = &tmptaint;
 876         err = proc_doulongvec_minmax(&t, write, buffer, lenp, ppos);
 877         if (err < 0)
 878                 return err;
 879
 880         if (write) {
 881                 int i;
 882
 883                 /*
 884                  * If we are relying on panic_on_taint not producing
 885                  * false positives due to userspace input, bail out
 886                  * before setting the requested taint flags.
 887                  */
 888                 if (panic_on_taint_nousertaint && (tmptaint & panic_on_taint))
 889                         return -EINVAL;
 890
 891                 /*
 892                  * Poor man's atomic or. Not worth adding a primitive
 893                  * to everyone's atomic.h for this
 894                  */
 895                 for (i = 0; i < TAINT_FLAGS_COUNT; i++)
 896                         if ((1UL << i) & tmptaint)
 897                                 add_taint(i, LOCKDEP_STILL_OK);
 898         }
 899
 900         return err;
 901 }
 902
 903 #ifdef CONFIG_PRINTK
 904 static int proc_dointvec_minmax_sysadmin(struct ctl_table *table, int write,
 905                                 void *buffer, size_t *lenp, loff_t *ppos)
 906 {
 907         if (write && !capable(CAP_SYS_ADMIN))
 908                 return -EPERM;
 909
 910         return proc_dointvec_minmax(table, write, buffer, lenp, ppos);
 911 }
 912 #endif
 913
 914 /**
 915  * struct do_proc_dointvec_minmax_conv_param - proc_dointvec_minmax() range checking structure
 916  * @min: pointer to minimum allowable value
 917  * @max: pointer to maximum allowable value
 918  *
 919  * The do_proc_dointvec_minmax_conv_param structure provides the
 920  * minimum and maximum values for doing range checking for those sysctl
 921  * parameters that use the proc_dointvec_minmax() handler.
 922  */
 923 struct do_proc_dointvec_minmax_conv_param {
 924         int *min;
 925         int *max;
 926 };
 927
 928 static int do_proc_dointvec_minmax_conv(bool *negp, unsigned long *lvalp,
 929                                         int *valp,
 930                                         int write, void *data)
 931 {
 932         int tmp, ret;
 933         struct do_proc_dointvec_minmax_conv_param *param = data;
 934         /*
 935          * If writing, first do so via a temporary local int so we can
 936          * bounds-check it before touching *valp.
 937          */
 938         int *ip = write ? &tmp : valp;
 939
 940         ret = do_proc_dointvec_conv(negp, lvalp, ip, write, data);
 941         if (ret)
 942                 return ret;
 943
 944         if (write) {
 945                 if ((param->min && *param->min > tmp) ||
 946                     (param->max && *param->max < tmp))
 947                         return -EINVAL;
 948                 *valp = tmp;
 949         }
 950
 951         return 0;
 952 }
 953
 954 /**
 955  * proc_dointvec_minmax - read a vector of integers with min/max values
 956  * @table: the sysctl table
 957  * @write: %TRUE if this is a write to the sysctl file
 958  * @buffer: the user buffer
 959  * @lenp: the size of the user buffer
 960  * @ppos: file position
 961  *
 962  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
 963  * values from/to the user buffer, treated as an ASCII string.
 964  *
 965  * This routine will ensure the values are within the range specified by
 966  * table->extra1 (min) and table->extra2 (max).
 967  *
 968  * Returns 0 on success or -EINVAL on write when the range check fails.
 969  */
 970 int proc_dointvec_minmax(struct ctl_table *table, int write,
 971                   void *buffer, size_t *lenp, loff_t *ppos)
 972 {
 973         struct do_proc_dointvec_minmax_conv_param param = {
 974                 .min = (int *) table->extra1,
 975                 .max = (int *) table->extra2,
 976         };
 977         return do_proc_dointvec(table, write, buffer, lenp, ppos,
 978                                 do_proc_dointvec_minmax_conv, &param);
 979 }
 980
 981 /**
 982  * struct do_proc_douintvec_minmax_conv_param - proc_douintvec_minmax() range checking structure
 983  * @min: pointer to minimum allowable value
 984  * @max: pointer to maximum allowable value
 985  *
 986  * The do_proc_douintvec_minmax_conv_param structure provides the
 987  * minimum and maximum values for doing range checking for those sysctl
 988  * parameters that use the proc_douintvec_minmax() handler.
 989  */
 990 struct do_proc_douintvec_minmax_conv_param {
 991         unsigned int *min;
 992         unsigned int *max;
 993 };
 994
 995 static int do_proc_douintvec_minmax_conv(unsigned long *lvalp,
 996                                          unsigned int *valp,
 997                                          int write, void *data)
 998 {
 999         int ret;
1000         unsigned int tmp;
1001         struct do_proc_douintvec_minmax_conv_param *param = data;
1002         /* write via temporary local uint for bounds-checking */
1003         unsigned int *up = write ? &tmp : valp;
1004
1005         ret = do_proc_douintvec_conv(lvalp, up, write, data);
1006         if (ret)
1007                 return ret;
1008
1009         if (write) {
1010                 if ((param->min && *param->min > tmp) ||
1011                     (param->max && *param->max < tmp))
1012                         return -ERANGE;
1013
1014                 *valp = tmp;
1015         }
1016
1017         return 0;
1018 }
1019
1020 /**
1021  * proc_douintvec_minmax - read a vector of unsigned ints with min/max values
1022  * @table: the sysctl table
1023  * @write: %TRUE if this is a write to the sysctl file
1024  * @buffer: the user buffer
1025  * @lenp: the size of the user buffer
1026  * @ppos: file position
1027  *
1028  * Reads/writes up to table->maxlen/sizeof(unsigned int) unsigned integer
1029  * values from/to the user buffer, treated as an ASCII string. Negative
1030  * strings are not allowed.
1031  *
1032  * This routine will ensure the values are within the range specified by
1033  * table->extra1 (min) and table->extra2 (max). There is a final sanity
1034  * check for UINT_MAX to avoid having to support wrap around uses from
1035  * userspace.
1036  *
1037  * Returns 0 on success or -ERANGE on write when the range check fails.
1038  */
1039 int proc_douintvec_minmax(struct ctl_table *table, int write,
1040                           void *buffer, size_t *lenp, loff_t *ppos)
1041 {
1042         struct do_proc_douintvec_minmax_conv_param param = {
1043                 .min = (unsigned int *) table->extra1,
1044                 .max = (unsigned int *) table->extra2,
1045         };
1046         return do_proc_douintvec(table, write, buffer, lenp, ppos,
1047                                  do_proc_douintvec_minmax_conv, &param);
1048 }
1049
1050 /**
1051  * proc_dou8vec_minmax - read a vector of unsigned chars with min/max values
1052  * @table: the sysctl table
1053  * @write: %TRUE if this is a write to the sysctl file
1054  * @buffer: the user buffer
1055  * @lenp: the size of the user buffer
1056  * @ppos: file position
1057  *
1058  * Reads/writes up to table->maxlen/sizeof(u8) unsigned chars
1059  * values from/to the user buffer, treated as an ASCII string. Negative
1060  * strings are not allowed.
1061  *
1062  * This routine will ensure the values are within the range specified by
1063  * table->extra1 (min) and table->extra2 (max).
1064  *
1065  * Returns 0 on success or an error on write when the range check fails.
1066  */
1067 int proc_dou8vec_minmax(struct ctl_table *table, int write,
1068                         void *buffer, size_t *lenp, loff_t *ppos)
1069 {
1070         struct ctl_table tmp;
1071         unsigned int min = 0, max = 255U, val;
1072         u8 *data = table->data;
1073         struct do_proc_douintvec_minmax_conv_param param = {
1074                 .min = &min,
1075                 .max = &max,
1076         };
1077         int res;
1078
1079         /* Do not support arrays yet. */
1080         if (table->maxlen != sizeof(u8))
1081                 return -EINVAL;
1082
1083         if (table->extra1) {
1084                 min = *(unsigned int *) table->extra1;
1085                 if (min > 255U)
1086                         return -EINVAL;
1087         }
1088         if (table->extra2) {
1089                 max = *(unsigned int *) table->extra2;
1090                 if (max > 255U)
1091                         return -EINVAL;
1092         }
1093
1094         tmp = *table;
1095
1096         tmp.maxlen = sizeof(val);
1097         tmp.data = &val;
1098         val = *data;
1099         res = do_proc_douintvec(&tmp, write, buffer, lenp, ppos,
1100                                 do_proc_douintvec_minmax_conv, &param);
1101         if (res)
1102                 return res;
1103         if (write)
1104                 *data = val;
1105         return 0;
1106 }
1107 EXPORT_SYMBOL_GPL(proc_dou8vec_minmax);
1108
1109 static int do_proc_dopipe_max_size_conv(unsigned long *lvalp,
1110                                         unsigned int *valp,
1111                                         int write, void *data)
1112 {
1113         if (write) {
1114                 unsigned int val;
1115
1116                 val = round_pipe_size(*lvalp);
1117                 if (val == 0)
1118                         return -EINVAL;
1119
1120                 *valp = val;
1121         } else {
1122                 unsigned int val = *valp;
1123                 *lvalp = (unsigned long) val;
1124         }
1125
1126         return 0;
1127 }
1128
1129 static int proc_dopipe_max_size(struct ctl_table *table, int write,
1130                                 void *buffer, size_t *lenp, loff_t *ppos)
1131 {
1132         return do_proc_douintvec(table, write, buffer, lenp, ppos,
1133                                  do_proc_dopipe_max_size_conv, NULL);
1134 }
1135
1136 static void validate_coredump_safety(void)
1137 {
1138 #ifdef CONFIG_COREDUMP
1139         if (suid_dumpable == SUID_DUMP_ROOT &&
1140             core_pattern[0] != '/' && core_pattern[0] != '|') {
1141                 printk(KERN_WARNING
1142 "Unsafe core_pattern used with fs.suid_dumpable=2.\n"
1143 "Pipe handler or fully qualified core dump path required.\n"
1144 "Set kernel.core_pattern before fs.suid_dumpable.\n"
1145                 );
1146         }
1147 #endif
1148 }
1149
1150 static int proc_dointvec_minmax_coredump(struct ctl_table *table, int write,
1151                 void *buffer, size_t *lenp, loff_t *ppos)
1152 {
1153         int error = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
1154         if (!error)
1155                 validate_coredump_safety();
1156         return error;
1157 }
1158
1159 #ifdef CONFIG_COREDUMP
1160 static int proc_dostring_coredump(struct ctl_table *table, int write,
1161                   void *buffer, size_t *lenp, loff_t *ppos)
1162 {
1163         int error = proc_dostring(table, write, buffer, lenp, ppos);
1164         if (!error)
1165                 validate_coredump_safety();
1166         return error;
1167 }
1168 #endif
1169
1170 #ifdef CONFIG_MAGIC_SYSRQ
1171 static int sysrq_sysctl_handler(struct ctl_table *table, int write,
1172                                 void *buffer, size_t *lenp, loff_t *ppos)
1173 {
1174         int tmp, ret;
1175
1176         tmp = sysrq_mask();
1177
1178         ret = __do_proc_dointvec(&tmp, table, write, buffer,
1179                                lenp, ppos, NULL, NULL);
1180         if (ret || !write)
1181                 return ret;
1182
1183         if (write)
1184                 sysrq_toggle_support(tmp);
1185
1186         return 0;
1187 }
1188 #endif
1189
1190 static int __do_proc_doulongvec_minmax(void *data, struct ctl_table *table,
1191                 int write, void *buffer, size_t *lenp, loff_t *ppos,
1192                 unsigned long convmul, unsigned long convdiv)
1193 {
1194         unsigned long *i, *min, *max;
1195         int vleft, first = 1, err = 0;
1196         size_t left;
1197         char *p;
1198
1199         if (!data || !table->maxlen || !*lenp || (*ppos && !write)) {
1200                 *lenp = 0;
1201                 return 0;
1202         }
1203
1204         i = (unsigned long *) data;
1205         min = (unsigned long *) table->extra1;
1206         max = (unsigned long *) table->extra2;
1207         vleft = table->maxlen / sizeof(unsigned long);
1208         left = *lenp;
1209
1210         if (write) {
1211                 if (proc_first_pos_non_zero_ignore(ppos, table))
1212                         goto out;
1213
1214                 if (left > PAGE_SIZE - 1)
1215                         left = PAGE_SIZE - 1;
1216                 p = buffer;
1217         }
1218
1219         for (; left && vleft--; i++, first = 0) {
1220                 unsigned long val;
1221
1222                 if (write) {
1223                         bool neg;
1224
1225                         left -= proc_skip_spaces(&p);
1226                         if (!left)
1227                                 break;
1228
1229                         err = proc_get_long(&p, &left, &val, &neg,
1230                                              proc_wspace_sep,
1231                                              sizeof(proc_wspace_sep), NULL);
1232                         if (err)
1233                                 break;
1234                         if (neg)
1235                                 continue;
1236                         val = convmul * val / convdiv;
1237                         if ((min && val < *min) || (max && val > *max)) {
1238                                 err = -EINVAL;
1239                                 break;
1240                         }
1241                         *i = val;
1242                 } else {
1243                         val = convdiv * (*i) / convmul;
1244                         if (!first)
1245                                 proc_put_char(&buffer, &left, '\t');
1246                         proc_put_long(&buffer, &left, val, false);
1247                 }
1248         }
1249
1250         if (!write && !first && left && !err)
1251                 proc_put_char(&buffer, &left, '\n');
1252         if (write && !err)
1253                 left -= proc_skip_spaces(&p);
1254         if (write && first)
1255                 return err ? : -EINVAL;
1256         *lenp -= left;
1257 out:
1258         *ppos += *lenp;
1259         return err;
1260 }
1261
1262 static int do_proc_doulongvec_minmax(struct ctl_table *table, int write,
1263                 void *buffer, size_t *lenp, loff_t *ppos, unsigned long convmul,
1264                 unsigned long convdiv)
1265 {
1266         return __do_proc_doulongvec_minmax(table->data, table, write,
1267                         buffer, lenp, ppos, convmul, convdiv);
1268 }
1269
1270 /**
1271  * proc_doulongvec_minmax - read a vector of long integers with min/max values
1272  * @table: the sysctl table
1273  * @write: %TRUE if this is a write to the sysctl file
1274  * @buffer: the user buffer
1275  * @lenp: the size of the user buffer
1276  * @ppos: file position
1277  *
1278  * Reads/writes up to table->maxlen/sizeof(unsigned long) unsigned long
1279  * values from/to the user buffer, treated as an ASCII string.
1280  *
1281  * This routine will ensure the values are within the range specified by
1282  * table->extra1 (min) and table->extra2 (max).
1283  *
1284  * Returns 0 on success.
1285  */
1286 int proc_doulongvec_minmax(struct ctl_table *table, int write,
1287                            void *buffer, size_t *lenp, loff_t *ppos)
1288 {
1289     return do_proc_doulongvec_minmax(table, write, buffer, lenp, ppos, 1l, 1l);
1290 }
1291
1292 /**
1293  * proc_doulongvec_ms_jiffies_minmax - read a vector of millisecond values with min/max values
1294  * @table: the sysctl table
1295  * @write: %TRUE if this is a write to the sysctl file
1296  * @buffer: the user buffer
1297  * @lenp: the size of the user buffer
1298  * @ppos: file position
1299  *
1300  * Reads/writes up to table->maxlen/sizeof(unsigned long) unsigned long
1301  * values from/to the user buffer, treated as an ASCII string. The values
1302  * are treated as milliseconds, and converted to jiffies when they are stored.
1303  *
1304  * This routine will ensure the values are within the range specified by
1305  * table->extra1 (min) and table->extra2 (max).
1306  *
1307  * Returns 0 on success.
1308  */
1309 int proc_doulongvec_ms_jiffies_minmax(struct ctl_table *table, int write,
1310                                       void *buffer, size_t *lenp, loff_t *ppos)
1311 {
1312     return do_proc_doulongvec_minmax(table, write, buffer,
1313                                      lenp, ppos, HZ, 1000l);
1314 }
1315
1316
1317 static int do_proc_dointvec_jiffies_conv(bool *negp, unsigned long *lvalp,
1318                                          int *valp,
1319                                          int write, void *data)
1320 {
1321         if (write) {
1322                 if (*lvalp > INT_MAX / HZ)
1323                         return 1;
1324                 *valp = *negp ? -(*lvalp*HZ) : (*lvalp*HZ);
1325         } else {
1326                 int val = *valp;
1327                 unsigned long lval;
1328                 if (val < 0) {
1329                         *negp = true;
1330                         lval = -(unsigned long)val;
1331                 } else {
1332                         *negp = false;
1333                         lval = (unsigned long)val;
1334                 }
1335                 *lvalp = lval / HZ;
1336         }
1337         return 0;
1338 }
1339
1340 static int do_proc_dointvec_userhz_jiffies_conv(bool *negp, unsigned long *lvalp,
1341                                                 int *valp,
1342                                                 int write, void *data)
1343 {
1344         if (write) {
1345                 if (USER_HZ < HZ && *lvalp > (LONG_MAX / HZ) * USER_HZ)
1346                         return 1;
1347                 *valp = clock_t_to_jiffies(*negp ? -*lvalp : *lvalp);
1348         } else {
1349                 int val = *valp;
1350                 unsigned long lval;
1351                 if (val < 0) {
1352                         *negp = true;
1353                         lval = -(unsigned long)val;
1354                 } else {
1355                         *negp = false;
1356                         lval = (unsigned long)val;
1357                 }
1358                 *lvalp = jiffies_to_clock_t(lval);
1359         }
1360         return 0;
1361 }
1362
1363 static int do_proc_dointvec_ms_jiffies_conv(bool *negp, unsigned long *lvalp,
1364                                             int *valp,
1365                                             int write, void *data)
1366 {
1367         if (write) {
1368                 unsigned long jif = msecs_to_jiffies(*negp ? -*lvalp : *lvalp);
1369
1370                 if (jif > INT_MAX)
1371                         return 1;
1372                 *valp = (int)jif;
1373         } else {
1374                 int val = *valp;
1375                 unsigned long lval;
1376                 if (val < 0) {
1377                         *negp = true;
1378                         lval = -(unsigned long)val;
1379                 } else {
1380                         *negp = false;
1381                         lval = (unsigned long)val;
1382                 }
1383                 *lvalp = jiffies_to_msecs(lval);
1384         }
1385         return 0;
1386 }
1387
1388 /**
1389  * proc_dointvec_jiffies - read a vector of integers as seconds
1390  * @table: the sysctl table
1391  * @write: %TRUE if this is a write to the sysctl file
1392  * @buffer: the user buffer
1393  * @lenp: the size of the user buffer
1394  * @ppos: file position
1395  *
1396  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
1397  * values from/to the user buffer, treated as an ASCII string.
1398  * The values read are assumed to be in seconds, and are converted into
1399  * jiffies.
1400  *
1401  * Returns 0 on success.
1402  */
1403 int proc_dointvec_jiffies(struct ctl_table *table, int write,
1404                           void *buffer, size_t *lenp, loff_t *ppos)
1405 {
1406     return do_proc_dointvec(table,write,buffer,lenp,ppos,
1407                             do_proc_dointvec_jiffies_conv,NULL);
1408 }
1409
1410 /**
1411  * proc_dointvec_userhz_jiffies - read a vector of integers as 1/USER_HZ seconds
1412  * @table: the sysctl table
1413  * @write: %TRUE if this is a write to the sysctl file
1414  * @buffer: the user buffer
1415  * @lenp: the size of the user buffer
1416  * @ppos: pointer to the file position
1417  *
1418  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
1419  * values from/to the user buffer, treated as an ASCII string.
1420  * The values read are assumed to be in 1/USER_HZ seconds, and
1421  * are converted into jiffies.
1422  *
1423  * Returns 0 on success.
1424  */
1425 int proc_dointvec_userhz_jiffies(struct ctl_table *table, int write,
1426                                  void *buffer, size_t *lenp, loff_t *ppos)
1427 {
1428     return do_proc_dointvec(table,write,buffer,lenp,ppos,
1429                             do_proc_dointvec_userhz_jiffies_conv,NULL);
1430 }
1431
1432 /**
1433  * proc_dointvec_ms_jiffies - read a vector of integers as 1 milliseconds
1434  * @table: the sysctl table
1435  * @write: %TRUE if this is a write to the sysctl file
1436  * @buffer: the user buffer
1437  * @lenp: the size of the user buffer
1438  * @ppos: file position
1439  * @ppos: the current position in the file
1440  *
1441  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
1442  * values from/to the user buffer, treated as an ASCII string.
1443  * The values read are assumed to be in 1/1000 seconds, and
1444  * are converted into jiffies.
1445  *
1446  * Returns 0 on success.
1447  */
1448 int proc_dointvec_ms_jiffies(struct ctl_table *table, int write, void *buffer,
1449                 size_t *lenp, loff_t *ppos)
1450 {
1451         return do_proc_dointvec(table, write, buffer, lenp, ppos,
1452                                 do_proc_dointvec_ms_jiffies_conv, NULL);
1453 }
1454
1455 static int proc_do_cad_pid(struct ctl_table *table, int write, void *buffer,
1456                 size_t *lenp, loff_t *ppos)
1457 {
1458         struct pid *new_pid;
1459         pid_t tmp;
1460         int r;
1461
1462         tmp = pid_vnr(cad_pid);
1463
1464         r = __do_proc_dointvec(&tmp, table, write, buffer,
1465                                lenp, ppos, NULL, NULL);
1466         if (r || !write)
1467                 return r;
1468
1469         new_pid = find_get_pid(tmp);
1470         if (!new_pid)
1471                 return -ESRCH;
1472
1473         put_pid(xchg(&cad_pid, new_pid));
1474         return 0;
1475 }
1476
1477 /**
1478  * proc_do_large_bitmap - read/write from/to a large bitmap
1479  * @table: the sysctl table
1480  * @write: %TRUE if this is a write to the sysctl file
1481  * @buffer: the user buffer
1482  * @lenp: the size of the user buffer
1483  * @ppos: file position
1484  *
1485  * The bitmap is stored at table->data and the bitmap length (in bits)
1486  * in table->maxlen.
1487  *
1488  * We use a range comma separated format (e.g. 1,3-4,10-10) so that
1489  * large bitmaps may be represented in a compact manner. Writing into
1490  * the file will clear the bitmap then update it with the given input.
1491  *
1492  * Returns 0 on success.
1493  */
1494 int proc_do_large_bitmap(struct ctl_table *table, int write,
1495                          void *buffer, size_t *lenp, loff_t *ppos)
1496 {
1497         int err = 0;
1498         bool first = 1;
1499         size_t left = *lenp;
1500         unsigned long bitmap_len = table->maxlen;
1501         unsigned long *bitmap = *(unsigned long **) table->data;
1502         unsigned long *tmp_bitmap = NULL;
1503         char tr_a[] = { '-', ',', '\n' }, tr_b[] = { ',', '\n', 0 }, c;
1504
1505         if (!bitmap || !bitmap_len || !left || (*ppos && !write)) {
1506                 *lenp = 0;
1507                 return 0;
1508         }
1509
1510         if (write) {
1511                 char *p = buffer;
1512                 size_t skipped = 0;
1513
1514                 if (left > PAGE_SIZE - 1) {
1515                         left = PAGE_SIZE - 1;
1516                         /* How much of the buffer we'll skip this pass */
1517                         skipped = *lenp - left;
1518                 }
1519
1520                 tmp_bitmap = bitmap_zalloc(bitmap_len, GFP_KERNEL);
1521                 if (!tmp_bitmap)
1522                         return -ENOMEM;
1523                 proc_skip_char(&p, &left, '\n');
1524                 while (!err && left) {
1525                         unsigned long val_a, val_b;
1526                         bool neg;
1527                         size_t saved_left;
1528
1529                         /* In case we stop parsing mid-number, we can reset */
1530                         saved_left = left;
1531                         err = proc_get_long(&p, &left, &val_a, &neg, tr_a,
1532                                              sizeof(tr_a), &c);
1533                         /*
1534                          * If we consumed the entirety of a truncated buffer or
1535                          * only one char is left (may be a "-"), then stop here,
1536                          * reset, & come back for more.
1537                          */
1538                         if ((left <= 1) && skipped) {
1539                                 left = saved_left;
1540                                 break;
1541                         }
1542
1543                         if (err)
1544                                 break;
1545                         if (val_a >= bitmap_len || neg) {
1546                                 err = -EINVAL;
1547                                 break;
1548                         }
1549
1550                         val_b = val_a;
1551                         if (left) {
1552                                 p++;
1553                                 left--;
1554                         }
1555
1556                         if (c == '-') {
1557                                 err = proc_get_long(&p, &left, &val_b,
1558                                                      &neg, tr_b, sizeof(tr_b),
1559                                                      &c);
1560                                 /*
1561                                  * If we consumed all of a truncated buffer or
1562                                  * then stop here, reset, & come back for more.
1563                                  */
1564                                 if (!left && skipped) {
1565                                         left = saved_left;
1566                                         break;
1567                                 }
1568
1569                                 if (err)
1570                                         break;
1571                                 if (val_b >= bitmap_len || neg ||
1572                                     val_a > val_b) {
1573                                         err = -EINVAL;
1574                                         break;
1575                                 }
1576                                 if (left) {
1577                                         p++;
1578                                         left--;
1579                                 }
1580                         }
1581
1582                         bitmap_set(tmp_bitmap, val_a, val_b - val_a + 1);
1583                         first = 0;
1584                         proc_skip_char(&p, &left, '\n');
1585                 }
1586                 left += skipped;
1587         } else {
1588                 unsigned long bit_a, bit_b = 0;
1589
1590                 while (left) {
1591                         bit_a = find_next_bit(bitmap, bitmap_len, bit_b);
1592                         if (bit_a >= bitmap_len)
1593                                 break;
1594                         bit_b = find_next_zero_bit(bitmap, bitmap_len,
1595                                                    bit_a + 1) - 1;
1596
1597                         if (!first)
1598                                 proc_put_char(&buffer, &left, ',');
1599                         proc_put_long(&buffer, &left, bit_a, false);
1600                         if (bit_a != bit_b) {
1601                                 proc_put_char(&buffer, &left, '-');
1602                                 proc_put_long(&buffer, &left, bit_b, false);
1603                         }
1604
1605                         first = 0; bit_b++;
1606                 }
1607                 proc_put_char(&buffer, &left, '\n');
1608         }
1609
1610         if (!err) {
1611                 if (write) {
1612                         if (*ppos)
1613                                 bitmap_or(bitmap, bitmap, tmp_bitmap, bitmap_len);
1614                         else
1615                                 bitmap_copy(bitmap, tmp_bitmap, bitmap_len);
1616                 }
1617                 *lenp -= left;
1618                 *ppos += *lenp;
1619         }
1620
1621         bitmap_free(tmp_bitmap);
1622         return err;
1623 }
1624
1625 #else /* CONFIG_PROC_SYSCTL */
1626
1627 int proc_dostring(struct ctl_table *table, int write,
1628                   void *buffer, size_t *lenp, loff_t *ppos)
1629 {
1630         return -ENOSYS;
1631 }
1632
1633 int proc_dointvec(struct ctl_table *table, int write,
1634                   void *buffer, size_t *lenp, loff_t *ppos)
1635 {
1636         return -ENOSYS;
1637 }
1638
1639 int proc_douintvec(struct ctl_table *table, int write,
1640                   void *buffer, size_t *lenp, loff_t *ppos)
1641 {
1642         return -ENOSYS;
1643 }
1644
1645 int proc_dointvec_minmax(struct ctl_table *table, int write,
1646                     void *buffer, size_t *lenp, loff_t *ppos)
1647 {
1648         return -ENOSYS;
1649 }
1650
1651 int proc_douintvec_minmax(struct ctl_table *table, int write,
1652                           void *buffer, size_t *lenp, loff_t *ppos)
1653 {
1654         return -ENOSYS;
1655 }
1656
1657 int proc_dou8vec_minmax(struct ctl_table *table, int write,
1658                         void *buffer, size_t *lenp, loff_t *ppos)
1659 {
1660         return -ENOSYS;
1661 }
1662
1663 int proc_dointvec_jiffies(struct ctl_table *table, int write,
1664                     void *buffer, size_t *lenp, loff_t *ppos)
1665 {
1666         return -ENOSYS;
1667 }
1668
1669 int proc_dointvec_userhz_jiffies(struct ctl_table *table, int write,
1670                     void *buffer, size_t *lenp, loff_t *ppos)
1671 {
1672         return -ENOSYS;
1673 }
1674
1675 int proc_dointvec_ms_jiffies(struct ctl_table *table, int write,
1676                              void *buffer, size_t *lenp, loff_t *ppos)
1677 {
1678         return -ENOSYS;
1679 }
1680
1681 int proc_doulongvec_minmax(struct ctl_table *table, int write,
1682                     void *buffer, size_t *lenp, loff_t *ppos)
1683 {
1684         return -ENOSYS;
1685 }
1686
1687 int proc_doulongvec_ms_jiffies_minmax(struct ctl_table *table, int write,
1688                                       void *buffer, size_t *lenp, loff_t *ppos)
1689 {
1690         return -ENOSYS;
1691 }
1692
1693 int proc_do_large_bitmap(struct ctl_table *table, int write,
1694                          void *buffer, size_t *lenp, loff_t *ppos)
1695 {
1696         return -ENOSYS;
1697 }
1698
1699 #endif /* CONFIG_PROC_SYSCTL */
1700
1701 #if defined(CONFIG_SYSCTL)
1702 int proc_do_static_key(struct ctl_table *table, int write,
1703                        void *buffer, size_t *lenp, loff_t *ppos)
1704 {
1705         struct static_key *key = (struct static_key *)table->data;
1706         static DEFINE_MUTEX(static_key_mutex);
1707         int val, ret;
1708         struct ctl_table tmp = {
1709                 .data   = &val,
1710                 .maxlen = sizeof(val),
1711                 .mode   = table->mode,
1712                 .extra1 = SYSCTL_ZERO,
1713                 .extra2 = SYSCTL_ONE,
1714         };
1715
1716         if (write && !capable(CAP_SYS_ADMIN))
1717                 return -EPERM;
1718
1719         mutex_lock(&static_key_mutex);
1720         val = static_key_enabled(key);
1721         ret = proc_dointvec_minmax(&tmp, write, buffer, lenp, ppos);
1722         if (write && !ret) {
1723                 if (val)
1724                         static_key_enable(key);
1725                 else
1726                         static_key_disable(key);
1727         }
1728         mutex_unlock(&static_key_mutex);
1729         return ret;
1730 }
1731
1732 static struct ctl_table kern_table[] = {
1733         {
1734                 .procname       = "sched_child_runs_first",
1735                 .data           = &sysctl_sched_child_runs_first,
1736                 .maxlen         = sizeof(unsigned int),
1737                 .mode           = 0644,
1738                 .proc_handler   = proc_dointvec,
1739         },
1740 #ifdef CONFIG_SCHEDSTATS
1741         {
1742                 .procname       = "sched_schedstats",
1743                 .data           = NULL,
1744                 .maxlen         = sizeof(unsigned int),
1745                 .mode           = 0644,
1746                 .proc_handler   = sysctl_schedstats,
1747                 .extra1         = SYSCTL_ZERO,
1748                 .extra2         = SYSCTL_ONE,
1749         },
1750 #endif /* CONFIG_SCHEDSTATS */
1751 #ifdef CONFIG_TASK_DELAY_ACCT
1752         {
1753                 .procname       = "task_delayacct",
1754                 .data           = NULL,
1755                 .maxlen         = sizeof(unsigned int),
1756                 .mode           = 0644,
1757                 .proc_handler   = sysctl_delayacct,
1758                 .extra1         = SYSCTL_ZERO,
1759                 .extra2         = SYSCTL_ONE,
1760         },
1761 #endif /* CONFIG_TASK_DELAY_ACCT */
1762 #ifdef CONFIG_NUMA_BALANCING
1763         {
1764                 .procname       = "numa_balancing",
1765                 .data           = NULL, /* filled in by handler */
1766                 .maxlen         = sizeof(unsigned int),
1767                 .mode           = 0644,
1768                 .proc_handler   = sysctl_numa_balancing,
1769                 .extra1         = SYSCTL_ZERO,
1770                 .extra2         = SYSCTL_ONE,
1771         },
1772 #endif /* CONFIG_NUMA_BALANCING */
1773         {
1774                 .procname       = "sched_rt_period_us",
1775                 .data           = &sysctl_sched_rt_period,
1776                 .maxlen         = sizeof(unsigned int),
1777                 .mode           = 0644,
1778                 .proc_handler   = sched_rt_handler,
1779         },
1780         {
1781                 .procname       = "sched_rt_runtime_us",
1782                 .data           = &sysctl_sched_rt_runtime,
1783                 .maxlen         = sizeof(int),
1784                 .mode           = 0644,
1785                 .proc_handler   = sched_rt_handler,
1786         },
1787         {
1788                 .procname       = "sched_deadline_period_max_us",
1789                 .data           = &sysctl_sched_dl_period_max,
1790                 .maxlen         = sizeof(unsigned int),
1791                 .mode           = 0644,
1792                 .proc_handler   = proc_dointvec,
1793         },
1794         {
1795                 .procname       = "sched_deadline_period_min_us",
1796                 .data           = &sysctl_sched_dl_period_min,
1797                 .maxlen         = sizeof(unsigned int),
1798                 .mode           = 0644,
1799                 .proc_handler   = proc_dointvec,
1800         },
1801         {
1802                 .procname       = "sched_rr_timeslice_ms",
1803                 .data           = &sysctl_sched_rr_timeslice,
1804                 .maxlen         = sizeof(int),
1805                 .mode           = 0644,
1806                 .proc_handler   = sched_rr_handler,
1807         },
1808 #ifdef CONFIG_UCLAMP_TASK
1809         {
1810                 .procname       = "sched_util_clamp_min",
1811                 .data           = &sysctl_sched_uclamp_util_min,
1812                 .maxlen         = sizeof(unsigned int),
1813                 .mode           = 0644,
1814                 .proc_handler   = sysctl_sched_uclamp_handler,
1815         },
1816         {
1817                 .procname       = "sched_util_clamp_max",
1818                 .data           = &sysctl_sched_uclamp_util_max,
1819                 .maxlen         = sizeof(unsigned int),
1820                 .mode           = 0644,
1821                 .proc_handler   = sysctl_sched_uclamp_handler,
1822         },
1823         {
1824                 .procname       = "sched_util_clamp_min_rt_default",
1825                 .data           = &sysctl_sched_uclamp_util_min_rt_default,
1826                 .maxlen         = sizeof(unsigned int),
1827                 .mode           = 0644,
1828                 .proc_handler   = sysctl_sched_uclamp_handler,
1829         },
1830 #endif
1831 #ifdef CONFIG_SCHED_AUTOGROUP
1832         {
1833                 .procname       = "sched_autogroup_enabled",
1834                 .data           = &sysctl_sched_autogroup_enabled,
1835                 .maxlen         = sizeof(unsigned int),
1836                 .mode           = 0644,
1837                 .proc_handler   = proc_dointvec_minmax,
1838                 .extra1         = SYSCTL_ZERO,
1839                 .extra2         = SYSCTL_ONE,
1840         },
1841 #endif
1842 #ifdef CONFIG_CFS_BANDWIDTH
1843         {
1844                 .procname       = "sched_cfs_bandwidth_slice_us",
1845                 .data           = &sysctl_sched_cfs_bandwidth_slice,
1846                 .maxlen         = sizeof(unsigned int),
1847                 .mode           = 0644,
1848                 .proc_handler   = proc_dointvec_minmax,
1849                 .extra1         = SYSCTL_ONE,
1850         },
1851 #endif
1852 #if defined(CONFIG_ENERGY_MODEL) && defined(CONFIG_CPU_FREQ_GOV_SCHEDUTIL)
1853         {
1854                 .procname       = "sched_energy_aware",
1855                 .data           = &sysctl_sched_energy_aware,
1856                 .maxlen         = sizeof(unsigned int),
1857                 .mode           = 0644,
1858                 .proc_handler   = sched_energy_aware_handler,
1859                 .extra1         = SYSCTL_ZERO,
1860                 .extra2         = SYSCTL_ONE,
1861         },
1862 #endif
1863 #ifdef CONFIG_PROVE_LOCKING
1864         {
1865                 .procname       = "prove_locking",
1866                 .data           = &prove_locking,
1867                 .maxlen         = sizeof(int),
1868                 .mode           = 0644,
1869                 .proc_handler   = proc_dointvec,
1870         },
1871 #endif
1872 #ifdef CONFIG_LOCK_STAT
1873         {
1874                 .procname       = "lock_stat",
1875                 .data           = &lock_stat,
1876                 .maxlen         = sizeof(int),
1877                 .mode           = 0644,
1878                 .proc_handler   = proc_dointvec,
1879         },
1880 #endif
1881         {
1882                 .procname       = "panic",
1883                 .data           = &panic_timeout,
1884                 .maxlen         = sizeof(int),
1885                 .mode           = 0644,
1886                 .proc_handler   = proc_dointvec,
1887         },
1888 #ifdef CONFIG_COREDUMP
1889         {
1890                 .procname       = "core_uses_pid",
1891                 .data           = &core_uses_pid,
1892                 .maxlen         = sizeof(int),
1893                 .mode           = 0644,
1894                 .proc_handler   = proc_dointvec,
1895         },
1896         {
1897                 .procname       = "core_pattern",
1898                 .data           = core_pattern,
1899                 .maxlen         = CORENAME_MAX_SIZE,
1900                 .mode           = 0644,
1901                 .proc_handler   = proc_dostring_coredump,
1902         },
1903         {
1904                 .procname       = "core_pipe_limit",
1905                 .data           = &core_pipe_limit,
1906                 .maxlen         = sizeof(unsigned int),
1907                 .mode           = 0644,
1908                 .proc_handler   = proc_dointvec,
1909         },
1910 #endif
1911 #ifdef CONFIG_PROC_SYSCTL
1912         {
1913                 .procname       = "tainted",
1914                 .maxlen         = sizeof(long),
1915                 .mode           = 0644,
1916                 .proc_handler   = proc_taint,
1917         },
1918         {
1919                 .procname       = "sysctl_writes_strict",
1920                 .data           = &sysctl_writes_strict,
1921                 .maxlen         = sizeof(int),
1922                 .mode           = 0644,
1923                 .proc_handler   = proc_dointvec_minmax,
1924                 .extra1         = &neg_one,
1925                 .extra2         = SYSCTL_ONE,
1926         },
1927 #endif
1928 #ifdef CONFIG_LATENCYTOP
1929         {
1930                 .procname       = "latencytop",
1931                 .data           = &latencytop_enabled,
1932                 .maxlen         = sizeof(int),
1933                 .mode           = 0644,
1934                 .proc_handler   = sysctl_latencytop,
1935         },
1936 #endif
1937 #ifdef CONFIG_BLK_DEV_INITRD
1938         {
1939                 .procname       = "real-root-dev",
1940                 .data           = &real_root_dev,
1941                 .maxlen         = sizeof(int),
1942                 .mode           = 0644,
1943                 .proc_handler   = proc_dointvec,
1944         },
1945 #endif
1946         {
1947                 .procname       = "print-fatal-signals",
1948                 .data           = &print_fatal_signals,
1949                 .maxlen         = sizeof(int),
1950                 .mode           = 0644,
1951                 .proc_handler   = proc_dointvec,
1952         },
1953 #ifdef CONFIG_SPARC
1954         {
1955                 .procname       = "reboot-cmd",
1956                 .data           = reboot_command,
1957                 .maxlen         = 256,
1958                 .mode           = 0644,
1959                 .proc_handler   = proc_dostring,
1960         },
1961         {
1962                 .procname       = "stop-a",
1963                 .data           = &stop_a_enabled,
1964                 .maxlen         = sizeof (int),
1965                 .mode           = 0644,
1966                 .proc_handler   = proc_dointvec,
1967         },
1968         {
1969                 .procname       = "scons-poweroff",
1970                 .data           = &scons_pwroff,
1971                 .maxlen         = sizeof (int),
1972                 .mode           = 0644,
1973                 .proc_handler   = proc_dointvec,
1974         },
1975 #endif
1976 #ifdef CONFIG_SPARC64
1977         {
1978                 .procname       = "tsb-ratio",
1979                 .data           = &sysctl_tsb_ratio,
1980                 .maxlen         = sizeof (int),
1981                 .mode           = 0644,
1982                 .proc_handler   = proc_dointvec,
1983         },
1984 #endif
1985 #ifdef CONFIG_PARISC
1986         {
1987                 .procname       = "soft-power",
1988                 .data           = &pwrsw_enabled,
1989                 .maxlen         = sizeof (int),
1990                 .mode           = 0644,
1991                 .proc_handler   = proc_dointvec,
1992         },
1993 #endif
1994 #ifdef CONFIG_SYSCTL_ARCH_UNALIGN_ALLOW
1995         {
1996                 .procname       = "unaligned-trap",
1997                 .data           = &unaligned_enabled,
1998                 .maxlen         = sizeof (int),
1999                 .mode           = 0644,
2000                 .proc_handler   = proc_dointvec,
2001         },
2002 #endif
2003         {
2004                 .procname       = "ctrl-alt-del",
2005                 .data           = &C_A_D,
2006                 .maxlen         = sizeof(int),
2007                 .mode           = 0644,
2008                 .proc_handler   = proc_dointvec,
2009         },
2010 #ifdef CONFIG_FUNCTION_TRACER
2011         {
2012                 .procname       = "ftrace_enabled",
2013                 .data           = &ftrace_enabled,
2014                 .maxlen         = sizeof(int),
2015                 .mode           = 0644,
2016                 .proc_handler   = ftrace_enable_sysctl,
2017         },
2018 #endif
2019 #ifdef CONFIG_STACK_TRACER
2020         {
2021                 .procname       = "stack_tracer_enabled",
2022                 .data           = &stack_tracer_enabled,
2023                 .maxlen         = sizeof(int),
2024                 .mode           = 0644,
2025                 .proc_handler   = stack_trace_sysctl,
2026         },
2027 #endif
2028 #ifdef CONFIG_TRACING
2029         {
2030                 .procname       = "ftrace_dump_on_oops",
2031                 .data           = &ftrace_dump_on_oops,
2032                 .maxlen         = sizeof(int),
2033                 .mode           = 0644,
2034                 .proc_handler   = proc_dointvec,
2035         },
2036         {
2037                 .procname       = "traceoff_on_warning",
2038                 .data           = &__disable_trace_on_warning,
2039                 .maxlen         = sizeof(__disable_trace_on_warning),
2040                 .mode           = 0644,
2041                 .proc_handler   = proc_dointvec,
2042         },
2043         {
2044                 .procname       = "tracepoint_printk",
2045                 .data           = &tracepoint_printk,
2046                 .maxlen         = sizeof(tracepoint_printk),
2047                 .mode           = 0644,
2048                 .proc_handler   = tracepoint_printk_sysctl,
2049         },
2050 #endif
2051 #ifdef CONFIG_KEXEC_CORE
2052         {
2053                 .procname       = "kexec_load_disabled",
2054                 .data           = &kexec_load_disabled,
2055                 .maxlen         = sizeof(int),
2056                 .mode           = 0644,
2057                 /* only handle a transition from default "0" to "1" */
2058                 .proc_handler   = proc_dointvec_minmax,
2059                 .extra1         = SYSCTL_ONE,
2060                 .extra2         = SYSCTL_ONE,
2061         },
2062 #endif
2063 #ifdef CONFIG_MODULES
2064         {
2065                 .procname       = "modprobe",
2066                 .data           = &modprobe_path,
2067                 .maxlen         = KMOD_PATH_LEN,
2068                 .mode           = 0644,
2069                 .proc_handler   = proc_dostring,
2070         },
2071         {
2072                 .procname       = "modules_disabled",
2073                 .data           = &modules_disabled,
2074                 .maxlen         = sizeof(int),
2075                 .mode           = 0644,
2076                 /* only handle a transition from default "0" to "1" */
2077                 .proc_handler   = proc_dointvec_minmax,
2078                 .extra1         = SYSCTL_ONE,
2079                 .extra2         = SYSCTL_ONE,
2080         },
2081 #endif
2082 #ifdef CONFIG_UEVENT_HELPER
2083         {
2084                 .procname       = "hotplug",
2085                 .data           = &uevent_helper,
2086                 .maxlen         = UEVENT_HELPER_PATH_LEN,
2087                 .mode           = 0644,
2088                 .proc_handler   = proc_dostring,
2089         },
2090 #endif
2091 #ifdef CONFIG_CHR_DEV_SG
2092         {
2093                 .procname       = "sg-big-buff",
2094                 .data           = &sg_big_buff,
2095                 .maxlen         = sizeof (int),
2096                 .mode           = 0444,
2097                 .proc_handler   = proc_dointvec,
2098         },
2099 #endif
2100 #ifdef CONFIG_BSD_PROCESS_ACCT
2101         {
2102                 .procname       = "acct",
2103                 .data           = &acct_parm,
2104                 .maxlen         = 3*sizeof(int),
2105                 .mode           = 0644,
2106                 .proc_handler   = proc_dointvec,
2107         },
2108 #endif
2109 #ifdef CONFIG_MAGIC_SYSRQ
2110         {
2111                 .procname       = "sysrq",
2112                 .data           = NULL,
2113                 .maxlen         = sizeof (int),
2114                 .mode           = 0644,
2115                 .proc_handler   = sysrq_sysctl_handler,
2116         },
2117 #endif
2118 #ifdef CONFIG_PROC_SYSCTL
2119         {
2120                 .procname       = "cad_pid",
2121                 .data           = NULL,
2122                 .maxlen         = sizeof (int),
2123                 .mode           = 0600,
2124                 .proc_handler   = proc_do_cad_pid,
2125         },
2126 #endif
2127         {
2128                 .procname       = "threads-max",
2129                 .data           = NULL,
2130                 .maxlen         = sizeof(int),
2131                 .mode           = 0644,
2132                 .proc_handler   = sysctl_max_threads,
2133         },
2134         {
2135                 .procname       = "random",
2136                 .mode           = 0555,
2137                 .child          = random_table,
2138         },
2139         {
2140                 .procname       = "usermodehelper",
2141                 .mode           = 0555,
2142                 .child          = usermodehelper_table,
2143         },
2144 #ifdef CONFIG_FW_LOADER_USER_HELPER
2145         {
2146                 .procname       = "firmware_config",
2147                 .mode           = 0555,
2148                 .child          = firmware_config_table,
2149         },
2150 #endif
2151         {
2152                 .procname       = "overflowuid",
2153                 .data           = &overflowuid,
2154                 .maxlen         = sizeof(int),
2155                 .mode           = 0644,
2156                 .proc_handler   = proc_dointvec_minmax,
2157                 .extra1         = &minolduid,
2158                 .extra2         = &maxolduid,
2159         },
2160         {
2161                 .procname       = "overflowgid",
2162                 .data           = &overflowgid,
2163                 .maxlen         = sizeof(int),
2164                 .mode           = 0644,
2165                 .proc_handler   = proc_dointvec_minmax,
2166                 .extra1         = &minolduid,
2167                 .extra2         = &maxolduid,
2168         },
2169 #ifdef CONFIG_S390
2170         {
2171                 .procname       = "userprocess_debug",
2172                 .data           = &show_unhandled_signals,
2173                 .maxlen         = sizeof(int),
2174                 .mode           = 0644,
2175                 .proc_handler   = proc_dointvec,
2176         },
2177 #endif
2178 #ifdef CONFIG_SMP
2179         {
2180                 .procname       = "oops_all_cpu_backtrace",
2181                 .data           = &sysctl_oops_all_cpu_backtrace,
2182                 .maxlen         = sizeof(int),
2183                 .mode           = 0644,
2184                 .proc_handler   = proc_dointvec_minmax,
2185                 .extra1         = SYSCTL_ZERO,
2186                 .extra2         = SYSCTL_ONE,
2187         },
2188 #endif /* CONFIG_SMP */
2189         {
2190                 .procname       = "pid_max",
2191                 .data           = &pid_max,
2192                 .maxlen         = sizeof (int),
2193                 .mode           = 0644,
2194                 .proc_handler   = proc_dointvec_minmax,
2195                 .extra1         = &pid_max_min,
2196                 .extra2         = &pid_max_max,
2197         },
2198         {
2199                 .procname       = "panic_on_oops",
2200                 .data           = &panic_on_oops,
2201                 .maxlen         = sizeof(int),
2202                 .mode           = 0644,
2203                 .proc_handler   = proc_dointvec,
2204         },
2205         {
2206                 .procname       = "panic_print",
2207                 .data           = &panic_print,
2208                 .maxlen         = sizeof(unsigned long),
2209                 .mode           = 0644,
2210                 .proc_handler   = proc_doulongvec_minmax,
2211         },
2212 #if defined CONFIG_PRINTK
2213         {
2214                 .procname       = "printk",
2215                 .data           = &console_loglevel,
2216                 .maxlen         = 4*sizeof(int),
2217                 .mode           = 0644,
2218                 .proc_handler   = proc_dointvec,
2219         },
2220         {
2221                 .procname       = "printk_ratelimit",
2222                 .data           = &printk_ratelimit_state.interval,
2223                 .maxlen         = sizeof(int),
2224                 .mode           = 0644,
2225                 .proc_handler   = proc_dointvec_jiffies,
2226         },
2227         {
2228                 .procname       = "printk_ratelimit_burst",
2229                 .data           = &printk_ratelimit_state.burst,
2230                 .maxlen         = sizeof(int),
2231                 .mode           = 0644,
2232                 .proc_handler   = proc_dointvec,
2233         },
2234         {
2235                 .procname       = "printk_delay",
2236                 .data           = &printk_delay_msec,
2237                 .maxlen         = sizeof(int),
2238                 .mode           = 0644,
2239                 .proc_handler   = proc_dointvec_minmax,
2240                 .extra1         = SYSCTL_ZERO,
2241                 .extra2         = &ten_thousand,
2242         },
2243         {
2244                 .procname       = "printk_devkmsg",
2245                 .data           = devkmsg_log_str,
2246                 .maxlen         = DEVKMSG_STR_MAX_SIZE,
2247                 .mode           = 0644,
2248                 .proc_handler   = devkmsg_sysctl_set_loglvl,
2249         },
2250         {
2251                 .procname       = "dmesg_restrict",
2252                 .data           = &dmesg_restrict,
2253                 .maxlen         = sizeof(int),
2254                 .mode           = 0644,
2255                 .proc_handler   = proc_dointvec_minmax_sysadmin,
2256                 .extra1         = SYSCTL_ZERO,
2257                 .extra2         = SYSCTL_ONE,
2258         },
2259         {
2260                 .procname       = "kptr_restrict",
2261                 .data           = &kptr_restrict,
2262                 .maxlen         = sizeof(int),
2263                 .mode           = 0644,
2264                 .proc_handler   = proc_dointvec_minmax_sysadmin,
2265                 .extra1         = SYSCTL_ZERO,
2266                 .extra2         = &two,
2267         },
2268 #endif
2269         {
2270                 .procname       = "ngroups_max",
2271                 .data           = &ngroups_max,
2272                 .maxlen         = sizeof (int),
2273                 .mode           = 0444,
2274                 .proc_handler   = proc_dointvec,
2275         },
2276         {
2277                 .procname       = "cap_last_cap",
2278                 .data           = (void *)&cap_last_cap,
2279                 .maxlen         = sizeof(int),
2280                 .mode           = 0444,
2281                 .proc_handler   = proc_dointvec,
2282         },
2283 #if defined(CONFIG_LOCKUP_DETECTOR)
2284         {
2285                 .procname       = "watchdog",
2286                 .data           = &watchdog_user_enabled,
2287                 .maxlen         = sizeof(int),
2288                 .mode           = 0644,
2289                 .proc_handler   = proc_watchdog,
2290                 .extra1         = SYSCTL_ZERO,
2291                 .extra2         = SYSCTL_ONE,
2292         },
2293         {
2294                 .procname       = "watchdog_thresh",
2295                 .data           = &watchdog_thresh,
2296                 .maxlen         = sizeof(int),
2297                 .mode           = 0644,
2298                 .proc_handler   = proc_watchdog_thresh,
2299                 .extra1         = SYSCTL_ZERO,
2300                 .extra2         = &sixty,
2301         },
2302         {
2303                 .procname       = "nmi_watchdog",
2304                 .data           = &nmi_watchdog_user_enabled,
2305                 .maxlen         = sizeof(int),
2306                 .mode           = NMI_WATCHDOG_SYSCTL_PERM,
2307                 .proc_handler   = proc_nmi_watchdog,
2308                 .extra1         = SYSCTL_ZERO,
2309                 .extra2         = SYSCTL_ONE,
2310         },
2311         {
2312                 .procname       = "watchdog_cpumask",
2313                 .data           = &watchdog_cpumask_bits,
2314                 .maxlen         = NR_CPUS,
2315                 .mode           = 0644,
2316                 .proc_handler   = proc_watchdog_cpumask,
2317         },
2318 #ifdef CONFIG_SOFTLOCKUP_DETECTOR
2319         {
2320                 .procname       = "soft_watchdog",
2321                 .data           = &soft_watchdog_user_enabled,
2322                 .maxlen         = sizeof(int),
2323                 .mode           = 0644,
2324                 .proc_handler   = proc_soft_watchdog,
2325                 .extra1         = SYSCTL_ZERO,
2326                 .extra2         = SYSCTL_ONE,
2327         },
2328         {
2329                 .procname       = "softlockup_panic",
2330                 .data           = &softlockup_panic,
2331                 .maxlen         = sizeof(int),
2332                 .mode           = 0644,
2333                 .proc_handler   = proc_dointvec_minmax,
2334                 .extra1         = SYSCTL_ZERO,
2335                 .extra2         = SYSCTL_ONE,
2336         },
2337 #ifdef CONFIG_SMP
2338         {
2339                 .procname       = "softlockup_all_cpu_backtrace",
2340                 .data           = &sysctl_softlockup_all_cpu_backtrace,
2341                 .maxlen         = sizeof(int),
2342                 .mode           = 0644,
2343                 .proc_handler   = proc_dointvec_minmax,
2344                 .extra1         = SYSCTL_ZERO,
2345                 .extra2         = SYSCTL_ONE,
2346         },
2347 #endif /* CONFIG_SMP */
2348 #endif
2349 #ifdef CONFIG_HARDLOCKUP_DETECTOR
2350         {
2351                 .procname       = "hardlockup_panic",
2352                 .data           = &hardlockup_panic,
2353                 .maxlen         = sizeof(int),
2354                 .mode           = 0644,
2355                 .proc_handler   = proc_dointvec_minmax,
2356                 .extra1         = SYSCTL_ZERO,
2357                 .extra2         = SYSCTL_ONE,
2358         },
2359 #ifdef CONFIG_SMP
2360         {
2361                 .procname       = "hardlockup_all_cpu_backtrace",
2362                 .data           = &sysctl_hardlockup_all_cpu_backtrace,
2363                 .maxlen         = sizeof(int),
2364                 .mode           = 0644,
2365                 .proc_handler   = proc_dointvec_minmax,
2366                 .extra1         = SYSCTL_ZERO,
2367                 .extra2         = SYSCTL_ONE,
2368         },
2369 #endif /* CONFIG_SMP */
2370 #endif
2371 #endif
2372
2373 #if defined(CONFIG_X86_LOCAL_APIC) && defined(CONFIG_X86)
2374         {
2375                 .procname       = "unknown_nmi_panic",
2376                 .data           = &unknown_nmi_panic,
2377                 .maxlen         = sizeof (int),
2378                 .mode           = 0644,
2379                 .proc_handler   = proc_dointvec,
2380         },
2381 #endif
2382
2383 #if (defined(CONFIG_X86_32) || defined(CONFIG_PARISC)) && \
2384         defined(CONFIG_DEBUG_STACKOVERFLOW)
2385         {
2386                 .procname       = "panic_on_stackoverflow",
2387                 .data           = &sysctl_panic_on_stackoverflow,
2388                 .maxlen         = sizeof(int),
2389                 .mode           = 0644,
2390                 .proc_handler   = proc_dointvec,
2391         },
2392 #endif
2393 #if defined(CONFIG_X86)
2394         {
2395                 .procname       = "panic_on_unrecovered_nmi",
2396                 .data           = &panic_on_unrecovered_nmi,
2397                 .maxlen         = sizeof(int),
2398                 .mode           = 0644,
2399                 .proc_handler   = proc_dointvec,
2400         },
2401         {
2402                 .procname       = "panic_on_io_nmi",
2403                 .data           = &panic_on_io_nmi,
2404                 .maxlen         = sizeof(int),
2405                 .mode           = 0644,
2406                 .proc_handler   = proc_dointvec,
2407         },
2408         {
2409                 .procname       = "bootloader_type",
2410                 .data           = &bootloader_type,
2411                 .maxlen         = sizeof (int),
2412                 .mode           = 0444,
2413                 .proc_handler   = proc_dointvec,
2414         },
2415         {
2416                 .procname       = "bootloader_version",
2417                 .data           = &bootloader_version,
2418                 .maxlen         = sizeof (int),
2419                 .mode           = 0444,
2420                 .proc_handler   = proc_dointvec,
2421         },
2422         {
2423                 .procname       = "io_delay_type",
2424                 .data           = &io_delay_type,
2425                 .maxlen         = sizeof(int),
2426                 .mode           = 0644,
2427                 .proc_handler   = proc_dointvec,
2428         },
2429 #endif
2430 #if defined(CONFIG_MMU)
2431         {
2432                 .procname       = "randomize_va_space",
2433                 .data           = &randomize_va_space,
2434                 .maxlen         = sizeof(int),
2435                 .mode           = 0644,
2436                 .proc_handler   = proc_dointvec,
2437         },
2438 #endif
2439 #if defined(CONFIG_S390) && defined(CONFIG_SMP)
2440         {
2441                 .procname       = "spin_retry",
2442                 .data           = &spin_retry,
2443                 .maxlen         = sizeof (int),
2444                 .mode           = 0644,
2445                 .proc_handler   = proc_dointvec,
2446         },
2447 #endif
2448 #if     defined(CONFIG_ACPI_SLEEP) && defined(CONFIG_X86)
2449         {
2450                 .procname       = "acpi_video_flags",
2451                 .data           = &acpi_realmode_flags,
2452                 .maxlen         = sizeof (unsigned long),
2453                 .mode           = 0644,
2454                 .proc_handler   = proc_doulongvec_minmax,
2455         },
2456 #endif
2457 #ifdef CONFIG_SYSCTL_ARCH_UNALIGN_NO_WARN
2458         {
2459                 .procname       = "ignore-unaligned-usertrap",
2460                 .data           = &no_unaligned_warning,
2461                 .maxlen         = sizeof (int),
2462                 .mode           = 0644,
2463                 .proc_handler   = proc_dointvec,
2464         },
2465 #endif
2466 #ifdef CONFIG_IA64
2467         {
2468                 .procname       = "unaligned-dump-stack",
2469                 .data           = &unaligned_dump_stack,
2470                 .maxlen         = sizeof (int),
2471                 .mode           = 0644,
2472                 .proc_handler   = proc_dointvec,
2473         },
2474 #endif
2475 #ifdef CONFIG_DETECT_HUNG_TASK
2476 #ifdef CONFIG_SMP
2477         {
2478                 .procname       = "hung_task_all_cpu_backtrace",
2479                 .data           = &sysctl_hung_task_all_cpu_backtrace,
2480                 .maxlen         = sizeof(int),
2481                 .mode           = 0644,
2482                 .proc_handler   = proc_dointvec_minmax,
2483                 .extra1         = SYSCTL_ZERO,
2484                 .extra2         = SYSCTL_ONE,
2485         },
2486 #endif /* CONFIG_SMP */
2487         {
2488                 .procname       = "hung_task_panic",
2489                 .data           = &sysctl_hung_task_panic,
2490                 .maxlen         = sizeof(int),
2491                 .mode           = 0644,
2492                 .proc_handler   = proc_dointvec_minmax,
2493                 .extra1         = SYSCTL_ZERO,
2494                 .extra2         = SYSCTL_ONE,
2495         },
2496         {
2497                 .procname       = "hung_task_check_count",
2498                 .data           = &sysctl_hung_task_check_count,
2499                 .maxlen         = sizeof(int),
2500                 .mode           = 0644,
2501                 .proc_handler   = proc_dointvec_minmax,
2502                 .extra1         = SYSCTL_ZERO,
2503         },
2504         {
2505                 .procname       = "hung_task_timeout_secs",
2506                 .data           = &sysctl_hung_task_timeout_secs,
2507                 .maxlen         = sizeof(unsigned long),
2508                 .mode           = 0644,
2509                 .proc_handler   = proc_dohung_task_timeout_secs,
2510                 .extra2         = &hung_task_timeout_max,
2511         },
2512         {
2513                 .procname       = "hung_task_check_interval_secs",
2514                 .data           = &sysctl_hung_task_check_interval_secs,
2515                 .maxlen         = sizeof(unsigned long),
2516                 .mode           = 0644,
2517                 .proc_handler   = proc_dohung_task_timeout_secs,
2518                 .extra2         = &hung_task_timeout_max,
2519         },
2520         {
2521                 .procname       = "hung_task_warnings",
2522                 .data           = &sysctl_hung_task_warnings,
2523                 .maxlen         = sizeof(int),
2524                 .mode           = 0644,
2525                 .proc_handler   = proc_dointvec_minmax,
2526                 .extra1         = &neg_one,
2527         },
2528 #endif
2529 #ifdef CONFIG_RT_MUTEXES
2530         {
2531                 .procname       = "max_lock_depth",
2532                 .data           = &max_lock_depth,
2533                 .maxlen         = sizeof(int),
2534                 .mode           = 0644,
2535                 .proc_handler   = proc_dointvec,
2536         },
2537 #endif
2538         {
2539                 .procname       = "poweroff_cmd",
2540                 .data           = &poweroff_cmd,
2541                 .maxlen         = POWEROFF_CMD_PATH_LEN,
2542                 .mode           = 0644,
2543                 .proc_handler   = proc_dostring,
2544         },
2545 #ifdef CONFIG_KEYS
2546         {
2547                 .procname       = "keys",
2548                 .mode           = 0555,
2549                 .child          = key_sysctls,
2550         },
2551 #endif
2552 #ifdef CONFIG_PERF_EVENTS
2553         /*
2554          * User-space scripts rely on the existence of this file
2555          * as a feature check for perf_events being enabled.
2556          *
2557          * So it's an ABI, do not remove!
2558          */
2559         {
2560                 .procname       = "perf_event_paranoid",
2561                 .data           = &sysctl_perf_event_paranoid,
2562                 .maxlen         = sizeof(sysctl_perf_event_paranoid),
2563                 .mode           = 0644,
2564                 .proc_handler   = proc_dointvec,
2565         },
2566         {
2567                 .procname       = "perf_event_mlock_kb",
2568                 .data           = &sysctl_perf_event_mlock,
2569                 .maxlen         = sizeof(sysctl_perf_event_mlock),
2570                 .mode           = 0644,
2571                 .proc_handler   = proc_dointvec,
2572         },
2573         {
2574                 .procname       = "perf_event_max_sample_rate",
2575                 .data           = &sysctl_perf_event_sample_rate,
2576                 .maxlen         = sizeof(sysctl_perf_event_sample_rate),
2577                 .mode           = 0644,
2578                 .proc_handler   = perf_proc_update_handler,
2579                 .extra1         = SYSCTL_ONE,
2580         },
2581         {
2582                 .procname       = "perf_cpu_time_max_percent",
2583                 .data           = &sysctl_perf_cpu_time_max_percent,
2584                 .maxlen         = sizeof(sysctl_perf_cpu_time_max_percent),
2585                 .mode           = 0644,
2586                 .proc_handler   = perf_cpu_time_max_percent_handler,
2587                 .extra1         = SYSCTL_ZERO,
2588                 .extra2         = &one_hundred,
2589         },
2590         {
2591                 .procname       = "perf_event_max_stack",
2592                 .data           = &sysctl_perf_event_max_stack,
2593                 .maxlen         = sizeof(sysctl_perf_event_max_stack),
2594                 .mode           = 0644,
2595                 .proc_handler   = perf_event_max_stack_handler,
2596                 .extra1         = SYSCTL_ZERO,
2597                 .extra2         = &six_hundred_forty_kb,
2598         },
2599         {
2600                 .procname       = "perf_event_max_contexts_per_stack",
2601                 .data           = &sysctl_perf_event_max_contexts_per_stack,
2602                 .maxlen         = sizeof(sysctl_perf_event_max_contexts_per_stack),
2603                 .mode           = 0644,
2604                 .proc_handler   = perf_event_max_stack_handler,
2605                 .extra1         = SYSCTL_ZERO,
2606                 .extra2         = &one_thousand,
2607         },
2608 #endif
2609         {
2610                 .procname       = "panic_on_warn",
2611                 .data           = &panic_on_warn,
2612                 .maxlen         = sizeof(int),
2613                 .mode           = 0644,
2614                 .proc_handler   = proc_dointvec_minmax,
2615                 .extra1         = SYSCTL_ZERO,
2616                 .extra2         = SYSCTL_ONE,
2617         },
2618 #if defined(CONFIG_SMP) && defined(CONFIG_NO_HZ_COMMON)
2619         {
2620                 .procname       = "timer_migration",
2621                 .data           = &sysctl_timer_migration,
2622                 .maxlen         = sizeof(unsigned int),
2623                 .mode           = 0644,
2624                 .proc_handler   = timer_migration_handler,
2625                 .extra1         = SYSCTL_ZERO,
2626                 .extra2         = SYSCTL_ONE,
2627         },
2628 #endif
2629 #ifdef CONFIG_BPF_SYSCALL
2630         {
2631                 .procname       = "unprivileged_bpf_disabled",
2632                 .data           = &sysctl_unprivileged_bpf_disabled,
2633                 .maxlen         = sizeof(sysctl_unprivileged_bpf_disabled),
2634                 .mode           = 0644,
2635                 .proc_handler   = bpf_unpriv_handler,
2636                 .extra1         = SYSCTL_ZERO,
2637                 .extra2         = &two,
2638         },
2639         {
2640                 .procname       = "bpf_stats_enabled",
2641                 .data           = &bpf_stats_enabled_key.key,
2642                 .maxlen         = sizeof(bpf_stats_enabled_key),
2643                 .mode           = 0644,
2644                 .proc_handler   = bpf_stats_handler,
2645         },
2646 #endif
2647 #if defined(CONFIG_TREE_RCU)
2648         {
2649                 .procname       = "panic_on_rcu_stall",
2650                 .data           = &sysctl_panic_on_rcu_stall,
2651                 .maxlen         = sizeof(sysctl_panic_on_rcu_stall),
2652                 .mode           = 0644,
2653                 .proc_handler   = proc_dointvec_minmax,
2654                 .extra1         = SYSCTL_ZERO,
2655                 .extra2         = SYSCTL_ONE,
2656         },
2657 #endif
2658 #if defined(CONFIG_TREE_RCU)
2659         {
2660                 .procname       = "max_rcu_stall_to_panic",
2661                 .data           = &sysctl_max_rcu_stall_to_panic,
2662                 .maxlen         = sizeof(sysctl_max_rcu_stall_to_panic),
2663                 .mode           = 0644,
2664                 .proc_handler   = proc_dointvec_minmax,
2665                 .extra1         = SYSCTL_ONE,
2666                 .extra2         = SYSCTL_INT_MAX,
2667         },
2668 #endif
2669 #ifdef CONFIG_STACKLEAK_RUNTIME_DISABLE
2670         {
2671                 .procname       = "stack_erasing",
2672                 .data           = NULL,
2673                 .maxlen         = sizeof(int),
2674                 .mode           = 0600,
2675                 .proc_handler   = stack_erasing_sysctl,
2676                 .extra1         = SYSCTL_ZERO,
2677                 .extra2         = SYSCTL_ONE,
2678         },
2679 #endif
2680         { }
2681 };
2682
2683 static struct ctl_table vm_table[] = {
2684         {
2685                 .procname       = "overcommit_memory",
2686                 .data           = &sysctl_overcommit_memory,
2687                 .maxlen         = sizeof(sysctl_overcommit_memory),
2688                 .mode           = 0644,
2689                 .proc_handler   = overcommit_policy_handler,
2690                 .extra1         = SYSCTL_ZERO,
2691                 .extra2         = &two,
2692         },
2693         {
2694                 .procname       = "panic_on_oom",
2695                 .data           = &sysctl_panic_on_oom,
2696                 .maxlen         = sizeof(sysctl_panic_on_oom),
2697                 .mode           = 0644,
2698                 .proc_handler   = proc_dointvec_minmax,
2699                 .extra1         = SYSCTL_ZERO,
2700                 .extra2         = &two,
2701         },
2702         {
2703                 .procname       = "oom_kill_allocating_task",
2704                 .data           = &sysctl_oom_kill_allocating_task,
2705                 .maxlen         = sizeof(sysctl_oom_kill_allocating_task),
2706                 .mode           = 0644,
2707                 .proc_handler   = proc_dointvec,
2708         },
2709         {
2710                 .procname       = "oom_dump_tasks",
2711                 .data           = &sysctl_oom_dump_tasks,
2712                 .maxlen         = sizeof(sysctl_oom_dump_tasks),
2713                 .mode           = 0644,
2714                 .proc_handler   = proc_dointvec,
2715         },
2716         {
2717                 .procname       = "overcommit_ratio",
2718                 .data           = &sysctl_overcommit_ratio,
2719                 .maxlen         = sizeof(sysctl_overcommit_ratio),
2720                 .mode           = 0644,
2721                 .proc_handler   = overcommit_ratio_handler,
2722         },
2723         {
2724                 .procname       = "overcommit_kbytes",
2725                 .data           = &sysctl_overcommit_kbytes,
2726                 .maxlen         = sizeof(sysctl_overcommit_kbytes),
2727                 .mode           = 0644,
2728                 .proc_handler   = overcommit_kbytes_handler,
2729         },
2730         {
2731                 .procname       = "page-cluster",
2732                 .data           = &page_cluster,
2733                 .maxlen         = sizeof(int),
2734                 .mode           = 0644,
2735                 .proc_handler   = proc_dointvec_minmax,
2736                 .extra1         = SYSCTL_ZERO,
2737         },
2738         {
2739                 .procname       = "dirty_background_ratio",
2740                 .data           = &dirty_background_ratio,
2741                 .maxlen         = sizeof(dirty_background_ratio),
2742                 .mode           = 0644,
2743                 .proc_handler   = dirty_background_ratio_handler,
2744                 .extra1         = SYSCTL_ZERO,
2745                 .extra2         = &one_hundred,
2746         },
2747         {
2748                 .procname       = "dirty_background_bytes",
2749                 .data           = &dirty_background_bytes,
2750                 .maxlen         = sizeof(dirty_background_bytes),
2751                 .mode           = 0644,
2752                 .proc_handler   = dirty_background_bytes_handler,
2753                 .extra1         = &one_ul,
2754         },
2755         {
2756                 .procname       = "dirty_ratio",
2757                 .data           = &vm_dirty_ratio,
2758                 .maxlen         = sizeof(vm_dirty_ratio),
2759                 .mode           = 0644,
2760                 .proc_handler   = dirty_ratio_handler,
2761                 .extra1         = SYSCTL_ZERO,
2762                 .extra2         = &one_hundred,
2763         },
2764         {
2765                 .procname       = "dirty_bytes",
2766                 .data           = &vm_dirty_bytes,
2767                 .maxlen         = sizeof(vm_dirty_bytes),
2768                 .mode           = 0644,
2769                 .proc_handler   = dirty_bytes_handler,
2770                 .extra1         = &dirty_bytes_min,
2771         },
2772         {
2773                 .procname       = "dirty_writeback_centisecs",
2774                 .data           = &dirty_writeback_interval,
2775                 .maxlen         = sizeof(dirty_writeback_interval),
2776                 .mode           = 0644,
2777                 .proc_handler   = dirty_writeback_centisecs_handler,
2778         },
2779         {
2780                 .procname       = "dirty_expire_centisecs",
2781                 .data           = &dirty_expire_interval,
2782                 .maxlen         = sizeof(dirty_expire_interval),
2783                 .mode           = 0644,
2784                 .proc_handler   = proc_dointvec_minmax,
2785                 .extra1         = SYSCTL_ZERO,
2786         },
2787         {
2788                 .procname       = "dirtytime_expire_seconds",
2789                 .data           = &dirtytime_expire_interval,
2790                 .maxlen         = sizeof(dirtytime_expire_interval),
2791                 .mode           = 0644,
2792                 .proc_handler   = dirtytime_interval_handler,
2793                 .extra1         = SYSCTL_ZERO,
2794         },
2795         {
2796                 .procname       = "swappiness",
2797                 .data           = &vm_swappiness,
2798                 .maxlen         = sizeof(vm_swappiness),
2799                 .mode           = 0644,
2800                 .proc_handler   = proc_dointvec_minmax,
2801                 .extra1         = SYSCTL_ZERO,
2802                 .extra2         = &two_hundred,
2803         },
2804 #ifdef CONFIG_HUGETLB_PAGE
2805         {
2806                 .procname       = "nr_hugepages",
2807                 .data           = NULL,
2808                 .maxlen         = sizeof(unsigned long),
2809                 .mode           = 0644,
2810                 .proc_handler   = hugetlb_sysctl_handler,
2811         },
2812 #ifdef CONFIG_NUMA
2813         {
2814                 .procname       = "nr_hugepages_mempolicy",
2815                 .data           = NULL,
2816                 .maxlen         = sizeof(unsigned long),
2817                 .mode           = 0644,
2818                 .proc_handler   = &hugetlb_mempolicy_sysctl_handler,
2819         },
2820         {
2821                 .procname               = "numa_stat",
2822                 .data                   = &sysctl_vm_numa_stat,
2823                 .maxlen                 = sizeof(int),
2824                 .mode                   = 0644,
2825                 .proc_handler   = sysctl_vm_numa_stat_handler,
2826                 .extra1                 = SYSCTL_ZERO,
2827                 .extra2                 = SYSCTL_ONE,
2828         },
2829 #endif
2830          {
2831                 .procname       = "hugetlb_shm_group",
2832                 .data           = &sysctl_hugetlb_shm_group,
2833                 .maxlen         = sizeof(gid_t),
2834                 .mode           = 0644,
2835                 .proc_handler   = proc_dointvec,
2836          },
2837         {
2838                 .procname       = "nr_overcommit_hugepages",
2839                 .data           = NULL,
2840                 .maxlen         = sizeof(unsigned long),
2841                 .mode           = 0644,
2842                 .proc_handler   = hugetlb_overcommit_handler,
2843         },
2844 #endif
2845         {
2846                 .procname       = "lowmem_reserve_ratio",
2847                 .data           = &sysctl_lowmem_reserve_ratio,
2848                 .maxlen         = sizeof(sysctl_lowmem_reserve_ratio),
2849                 .mode           = 0644,
2850                 .proc_handler   = lowmem_reserve_ratio_sysctl_handler,
2851         },
2852         {
2853                 .procname       = "drop_caches",
2854                 .data           = &sysctl_drop_caches,
2855                 .maxlen         = sizeof(int),
2856                 .mode           = 0200,
2857                 .proc_handler   = drop_caches_sysctl_handler,
2858                 .extra1         = SYSCTL_ONE,
2859                 .extra2         = &four,
2860         },
2861 #ifdef CONFIG_COMPACTION
2862         {
2863                 .procname       = "compact_memory",
2864                 .data           = NULL,
2865                 .maxlen         = sizeof(int),
2866                 .mode           = 0200,
2867                 .proc_handler   = sysctl_compaction_handler,
2868         },
2869         {
2870                 .procname       = "compaction_proactiveness",
2871                 .data           = &sysctl_compaction_proactiveness,
2872                 .maxlen         = sizeof(sysctl_compaction_proactiveness),
2873                 .mode           = 0644,
2874                 .proc_handler   = proc_dointvec_minmax,
2875                 .extra1         = SYSCTL_ZERO,
2876                 .extra2         = &one_hundred,
2877         },
2878         {
2879                 .procname       = "extfrag_threshold",
2880                 .data           = &sysctl_extfrag_threshold,
2881                 .maxlen         = sizeof(int),
2882                 .mode           = 0644,
2883                 .proc_handler   = proc_dointvec_minmax,
2884                 .extra1         = &min_extfrag_threshold,
2885                 .extra2         = &max_extfrag_threshold,
2886         },
2887         {
2888                 .procname       = "compact_unevictable_allowed",
2889                 .data           = &sysctl_compact_unevictable_allowed,
2890                 .maxlen         = sizeof(int),
2891                 .mode           = 0644,
2892                 .proc_handler   = proc_dointvec_minmax_warn_RT_change,
2893                 .extra1         = SYSCTL_ZERO,
2894                 .extra2         = SYSCTL_ONE,
2895         },
2896
2897 #endif /* CONFIG_COMPACTION */
2898         {
2899                 .procname       = "min_free_kbytes",
2900                 .data           = &min_free_kbytes,
2901                 .maxlen         = sizeof(min_free_kbytes),
2902                 .mode           = 0644,
2903                 .proc_handler   = min_free_kbytes_sysctl_handler,
2904                 .extra1         = SYSCTL_ZERO,
2905         },
2906         {
2907                 .procname       = "watermark_boost_factor",
2908                 .data           = &watermark_boost_factor,
2909                 .maxlen         = sizeof(watermark_boost_factor),
2910                 .mode           = 0644,
2911                 .proc_handler   = proc_dointvec_minmax,
2912                 .extra1         = SYSCTL_ZERO,
2913         },
2914         {
2915                 .procname       = "watermark_scale_factor",
2916                 .data           = &watermark_scale_factor,
2917                 .maxlen         = sizeof(watermark_scale_factor),
2918                 .mode           = 0644,
2919                 .proc_handler   = watermark_scale_factor_sysctl_handler,
2920                 .extra1         = SYSCTL_ONE,
2921                 .extra2         = &one_thousand,
2922         },
2923         {
2924                 .procname       = "percpu_pagelist_high_fraction",
2925                 .data           = &percpu_pagelist_high_fraction,
2926                 .maxlen         = sizeof(percpu_pagelist_high_fraction),
2927                 .mode           = 0644,
2928                 .proc_handler   = percpu_pagelist_high_fraction_sysctl_handler,
2929                 .extra1         = SYSCTL_ZERO,
2930         },
2931         {
2932                 .procname       = "page_lock_unfairness",
2933                 .data           = &sysctl_page_lock_unfairness,
2934                 .maxlen         = sizeof(sysctl_page_lock_unfairness),
2935                 .mode           = 0644,
2936                 .proc_handler   = proc_dointvec_minmax,
2937                 .extra1         = SYSCTL_ZERO,
2938         },
2939 #ifdef CONFIG_MMU
2940         {
2941                 .procname       = "max_map_count",
2942                 .data           = &sysctl_max_map_count,
2943                 .maxlen         = sizeof(sysctl_max_map_count),
2944                 .mode           = 0644,
2945                 .proc_handler   = proc_dointvec_minmax,
2946                 .extra1         = SYSCTL_ZERO,
2947         },
2948 #else
2949         {
2950                 .procname       = "nr_trim_pages",
2951                 .data           = &sysctl_nr_trim_pages,
2952                 .maxlen         = sizeof(sysctl_nr_trim_pages),
2953                 .mode           = 0644,
2954                 .proc_handler   = proc_dointvec_minmax,
2955                 .extra1         = SYSCTL_ZERO,
2956         },
2957 #endif
2958         {
2959                 .procname       = "laptop_mode",
2960                 .data           = &laptop_mode,
2961                 .maxlen         = sizeof(laptop_mode),
2962                 .mode           = 0644,
2963                 .proc_handler   = proc_dointvec_jiffies,
2964         },
2965         {
2966                 .procname       = "block_dump",
2967                 .data           = &block_dump,
2968                 .maxlen         = sizeof(block_dump),
2969                 .mode           = 0644,
2970                 .proc_handler   = proc_dointvec_minmax,
2971                 .extra1         = SYSCTL_ZERO,
2972         },
2973         {
2974                 .procname       = "vfs_cache_pressure",
2975                 .data           = &sysctl_vfs_cache_pressure,
2976                 .maxlen         = sizeof(sysctl_vfs_cache_pressure),
2977                 .mode           = 0644,
2978                 .proc_handler   = proc_dointvec_minmax,
2979                 .extra1         = SYSCTL_ZERO,
2980         },
2981 #if defined(HAVE_ARCH_PICK_MMAP_LAYOUT) || \
2982     defined(CONFIG_ARCH_WANT_DEFAULT_TOPDOWN_MMAP_LAYOUT)
2983         {
2984                 .procname       = "legacy_va_layout",
2985                 .data           = &sysctl_legacy_va_layout,
2986                 .maxlen         = sizeof(sysctl_legacy_va_layout),
2987                 .mode           = 0644,
2988                 .proc_handler   = proc_dointvec_minmax,
2989                 .extra1         = SYSCTL_ZERO,
2990         },
2991 #endif
2992 #ifdef CONFIG_NUMA
2993         {
2994                 .procname       = "zone_reclaim_mode",
2995                 .data           = &node_reclaim_mode,
2996                 .maxlen         = sizeof(node_reclaim_mode),
2997                 .mode           = 0644,
2998                 .proc_handler   = proc_dointvec_minmax,
2999                 .extra1         = SYSCTL_ZERO,
3000         },
3001         {
3002                 .procname       = "min_unmapped_ratio",
3003                 .data           = &sysctl_min_unmapped_ratio,
3004                 .maxlen         = sizeof(sysctl_min_unmapped_ratio),
3005                 .mode           = 0644,
3006                 .proc_handler   = sysctl_min_unmapped_ratio_sysctl_handler,
3007                 .extra1         = SYSCTL_ZERO,
3008                 .extra2         = &one_hundred,
3009         },
3010         {
3011                 .procname       = "min_slab_ratio",
3012                 .data           = &sysctl_min_slab_ratio,
3013                 .maxlen         = sizeof(sysctl_min_slab_ratio),
3014                 .mode           = 0644,
3015                 .proc_handler   = sysctl_min_slab_ratio_sysctl_handler,
3016                 .extra1         = SYSCTL_ZERO,
3017                 .extra2         = &one_hundred,
3018         },
3019 #endif
3020 #ifdef CONFIG_SMP
3021         {
3022                 .procname       = "stat_interval",
3023                 .data           = &sysctl_stat_interval,
3024                 .maxlen         = sizeof(sysctl_stat_interval),
3025                 .mode           = 0644,
3026                 .proc_handler   = proc_dointvec_jiffies,
3027         },
3028         {
3029                 .procname       = "stat_refresh",
3030                 .data           = NULL,
3031                 .maxlen         = 0,
3032                 .mode           = 0600,
3033                 .proc_handler   = vmstat_refresh,
3034         },
3035 #endif
3036 #ifdef CONFIG_MMU
3037         {
3038                 .procname       = "mmap_min_addr",
3039                 .data           = &dac_mmap_min_addr,
3040                 .maxlen         = sizeof(unsigned long),
3041                 .mode           = 0644,
3042                 .proc_handler   = mmap_min_addr_handler,
3043         },
3044 #endif
3045 #ifdef CONFIG_NUMA
3046         {
3047                 .procname       = "numa_zonelist_order",
3048                 .data           = &numa_zonelist_order,
3049                 .maxlen         = NUMA_ZONELIST_ORDER_LEN,
3050                 .mode           = 0644,
3051                 .proc_handler   = numa_zonelist_order_handler,
3052         },
3053 #endif
3054 #if (defined(CONFIG_X86_32) && !defined(CONFIG_UML))|| \
3055    (defined(CONFIG_SUPERH) && defined(CONFIG_VSYSCALL))
3056         {
3057                 .procname       = "vdso_enabled",
3058 #ifdef CONFIG_X86_32
3059                 .data           = &vdso32_enabled,
3060                 .maxlen         = sizeof(vdso32_enabled),
3061 #else
3062                 .data           = &vdso_enabled,
3063                 .maxlen         = sizeof(vdso_enabled),
3064 #endif
3065                 .mode           = 0644,
3066                 .proc_handler   = proc_dointvec,
3067                 .extra1         = SYSCTL_ZERO,
3068         },
3069 #endif
3070 #ifdef CONFIG_HIGHMEM
3071         {
3072                 .procname       = "highmem_is_dirtyable",
3073                 .data           = &vm_highmem_is_dirtyable,
3074                 .maxlen         = sizeof(vm_highmem_is_dirtyable),
3075                 .mode           = 0644,
3076                 .proc_handler   = proc_dointvec_minmax,
3077                 .extra1         = SYSCTL_ZERO,
3078                 .extra2         = SYSCTL_ONE,
3079         },
3080 #endif
3081 #ifdef CONFIG_MEMORY_FAILURE
3082         {
3083                 .procname       = "memory_failure_early_kill",
3084                 .data           = &sysctl_memory_failure_early_kill,
3085                 .maxlen         = sizeof(sysctl_memory_failure_early_kill),
3086                 .mode           = 0644,
3087                 .proc_handler   = proc_dointvec_minmax,
3088                 .extra1         = SYSCTL_ZERO,
3089                 .extra2         = SYSCTL_ONE,
3090         },
3091         {
3092                 .procname       = "memory_failure_recovery",
3093                 .data           = &sysctl_memory_failure_recovery,
3094                 .maxlen         = sizeof(sysctl_memory_failure_recovery),
3095                 .mode           = 0644,
3096                 .proc_handler   = proc_dointvec_minmax,
3097                 .extra1         = SYSCTL_ZERO,
3098                 .extra2         = SYSCTL_ONE,
3099         },
3100 #endif
3101         {
3102                 .procname       = "user_reserve_kbytes",
3103                 .data           = &sysctl_user_reserve_kbytes,
3104                 .maxlen         = sizeof(sysctl_user_reserve_kbytes),
3105                 .mode           = 0644,
3106                 .proc_handler   = proc_doulongvec_minmax,
3107         },
3108         {
3109                 .procname       = "admin_reserve_kbytes",
3110                 .data           = &sysctl_admin_reserve_kbytes,
3111                 .maxlen         = sizeof(sysctl_admin_reserve_kbytes),
3112                 .mode           = 0644,
3113                 .proc_handler   = proc_doulongvec_minmax,
3114         },
3115 #ifdef CONFIG_HAVE_ARCH_MMAP_RND_BITS
3116         {
3117                 .procname       = "mmap_rnd_bits",
3118                 .data           = &mmap_rnd_bits,
3119                 .maxlen         = sizeof(mmap_rnd_bits),
3120                 .mode           = 0600,
3121                 .proc_handler   = proc_dointvec_minmax,
3122                 .extra1         = (void *)&mmap_rnd_bits_min,
3123                 .extra2         = (void *)&mmap_rnd_bits_max,
3124         },
3125 #endif
3126 #ifdef CONFIG_HAVE_ARCH_MMAP_RND_COMPAT_BITS
3127         {
3128                 .procname       = "mmap_rnd_compat_bits",
3129                 .data           = &mmap_rnd_compat_bits,
3130                 .maxlen         = sizeof(mmap_rnd_compat_bits),
3131                 .mode           = 0600,
3132                 .proc_handler   = proc_dointvec_minmax,
3133                 .extra1         = (void *)&mmap_rnd_compat_bits_min,
3134                 .extra2         = (void *)&mmap_rnd_compat_bits_max,
3135         },
3136 #endif
3137 #ifdef CONFIG_USERFAULTFD
3138         {
3139                 .procname       = "unprivileged_userfaultfd",
3140                 .data           = &sysctl_unprivileged_userfaultfd,
3141                 .maxlen         = sizeof(sysctl_unprivileged_userfaultfd),
3142                 .mode           = 0644,
3143                 .proc_handler   = proc_dointvec_minmax,
3144                 .extra1         = SYSCTL_ZERO,
3145                 .extra2         = SYSCTL_ONE,
3146         },
3147 #endif
3148         { }
3149 };
3150
3151 static struct ctl_table fs_table[] = {
3152         {
3153                 .procname       = "inode-nr",
3154                 .data           = &inodes_stat,
3155                 .maxlen         = 2*sizeof(long),
3156                 .mode           = 0444,
3157                 .proc_handler   = proc_nr_inodes,
3158         },
3159         {
3160                 .procname       = "inode-state",
3161                 .data           = &inodes_stat,
3162                 .maxlen         = 7*sizeof(long),
3163                 .mode           = 0444,
3164                 .proc_handler   = proc_nr_inodes,
3165         },
3166         {
3167                 .procname       = "file-nr",
3168                 .data           = &files_stat,
3169                 .maxlen         = sizeof(files_stat),
3170                 .mode           = 0444,
3171                 .proc_handler   = proc_nr_files,
3172         },
3173         {
3174                 .procname       = "file-max",
3175                 .data           = &files_stat.max_files,
3176                 .maxlen         = sizeof(files_stat.max_files),
3177                 .mode           = 0644,
3178                 .proc_handler   = proc_doulongvec_minmax,
3179                 .extra1         = &zero_ul,
3180                 .extra2         = &long_max,
3181         },
3182         {
3183                 .procname       = "nr_open",
3184                 .data           = &sysctl_nr_open,
3185                 .maxlen         = sizeof(unsigned int),
3186                 .mode           = 0644,
3187                 .proc_handler   = proc_dointvec_minmax,
3188                 .extra1         = &sysctl_nr_open_min,
3189                 .extra2         = &sysctl_nr_open_max,
3190         },
3191         {
3192                 .procname       = "dentry-state",
3193                 .data           = &dentry_stat,
3194                 .maxlen         = 6*sizeof(long),
3195                 .mode           = 0444,
3196                 .proc_handler   = proc_nr_dentry,
3197         },
3198         {
3199                 .procname       = "overflowuid",
3200                 .data           = &fs_overflowuid,
3201                 .maxlen         = sizeof(int),
3202                 .mode           = 0644,
3203                 .proc_handler   = proc_dointvec_minmax,
3204                 .extra1         = &minolduid,
3205                 .extra2         = &maxolduid,
3206         },
3207         {
3208                 .procname       = "overflowgid",
3209                 .data           = &fs_overflowgid,
3210                 .maxlen         = sizeof(int),
3211                 .mode           = 0644,
3212                 .proc_handler   = proc_dointvec_minmax,
3213                 .extra1         = &minolduid,
3214                 .extra2         = &maxolduid,
3215         },
3216 #ifdef CONFIG_FILE_LOCKING
3217         {
3218                 .procname       = "leases-enable",
3219                 .data           = &leases_enable,
3220                 .maxlen         = sizeof(int),
3221                 .mode           = 0644,
3222                 .proc_handler   = proc_dointvec,
3223         },
3224 #endif
3225 #ifdef CONFIG_DNOTIFY
3226         {
3227                 .procname       = "dir-notify-enable",
3228                 .data           = &dir_notify_enable,
3229                 .maxlen         = sizeof(int),
3230                 .mode           = 0644,
3231                 .proc_handler   = proc_dointvec,
3232         },
3233 #endif
3234 #ifdef CONFIG_MMU
3235 #ifdef CONFIG_FILE_LOCKING
3236         {
3237                 .procname       = "lease-break-time",
3238                 .data           = &lease_break_time,
3239                 .maxlen         = sizeof(int),
3240                 .mode           = 0644,
3241                 .proc_handler   = proc_dointvec,
3242         },
3243 #endif
3244 #ifdef CONFIG_AIO
3245         {
3246                 .procname       = "aio-nr",
3247                 .data           = &aio_nr,
3248                 .maxlen         = sizeof(aio_nr),
3249                 .mode           = 0444,
3250                 .proc_handler   = proc_doulongvec_minmax,
3251         },
3252         {
3253                 .procname       = "aio-max-nr",
3254                 .data           = &aio_max_nr,
3255                 .maxlen         = sizeof(aio_max_nr),
3256                 .mode           = 0644,
3257                 .proc_handler   = proc_doulongvec_minmax,
3258         },
3259 #endif /* CONFIG_AIO */
3260 #ifdef CONFIG_INOTIFY_USER
3261         {
3262                 .procname       = "inotify",
3263                 .mode           = 0555,
3264                 .child          = inotify_table,
3265         },
3266 #endif
3267 #ifdef CONFIG_FANOTIFY
3268         {
3269                 .procname       = "fanotify",
3270                 .mode           = 0555,
3271                 .child          = fanotify_table,
3272         },
3273 #endif
3274 #ifdef CONFIG_EPOLL
3275         {
3276                 .procname       = "epoll",
3277                 .mode           = 0555,
3278                 .child          = epoll_table,
3279         },
3280 #endif
3281 #endif
3282         {
3283                 .procname       = "protected_symlinks",
3284                 .data           = &sysctl_protected_symlinks,
3285                 .maxlen         = sizeof(int),
3286                 .mode           = 0600,
3287                 .proc_handler   = proc_dointvec_minmax,
3288                 .extra1         = SYSCTL_ZERO,
3289                 .extra2         = SYSCTL_ONE,
3290         },
3291         {
3292                 .procname       = "protected_hardlinks",
3293                 .data           = &sysctl_protected_hardlinks,
3294                 .maxlen         = sizeof(int),
3295                 .mode           = 0600,
3296                 .proc_handler   = proc_dointvec_minmax,
3297                 .extra1         = SYSCTL_ZERO,
3298                 .extra2         = SYSCTL_ONE,
3299         },
3300         {
3301                 .procname       = "protected_fifos",
3302                 .data           = &sysctl_protected_fifos,
3303                 .maxlen         = sizeof(int),
3304                 .mode           = 0600,
3305                 .proc_handler   = proc_dointvec_minmax,
3306                 .extra1         = SYSCTL_ZERO,
3307                 .extra2         = &two,
3308         },
3309         {
3310                 .procname       = "protected_regular",
3311                 .data           = &sysctl_protected_regular,
3312                 .maxlen         = sizeof(int),
3313                 .mode           = 0600,
3314                 .proc_handler   = proc_dointvec_minmax,
3315                 .extra1         = SYSCTL_ZERO,
3316                 .extra2         = &two,
3317         },
3318         {
3319                 .procname       = "suid_dumpable",
3320                 .data           = &suid_dumpable,
3321                 .maxlen         = sizeof(int),
3322                 .mode           = 0644,
3323                 .proc_handler   = proc_dointvec_minmax_coredump,
3324                 .extra1         = SYSCTL_ZERO,
3325                 .extra2         = &two,
3326         },
3327 #if defined(CONFIG_BINFMT_MISC) || defined(CONFIG_BINFMT_MISC_MODULE)
3328         {
3329                 .procname       = "binfmt_misc",
3330                 .mode           = 0555,
3331                 .child          = sysctl_mount_point,
3332         },
3333 #endif
3334         {
3335                 .procname       = "pipe-max-size",
3336                 .data           = &pipe_max_size,
3337                 .maxlen         = sizeof(pipe_max_size),
3338                 .mode           = 0644,
3339                 .proc_handler   = proc_dopipe_max_size,
3340         },
3341         {
3342                 .procname       = "pipe-user-pages-hard",
3343                 .data           = &pipe_user_pages_hard,
3344                 .maxlen         = sizeof(pipe_user_pages_hard),
3345                 .mode           = 0644,
3346                 .proc_handler   = proc_doulongvec_minmax,
3347         },
3348         {
3349                 .procname       = "pipe-user-pages-soft",
3350                 .data           = &pipe_user_pages_soft,
3351                 .maxlen         = sizeof(pipe_user_pages_soft),
3352                 .mode           = 0644,
3353                 .proc_handler   = proc_doulongvec_minmax,
3354         },
3355         {
3356                 .procname       = "mount-max",
3357                 .data           = &sysctl_mount_max,
3358                 .maxlen         = sizeof(unsigned int),
3359                 .mode           = 0644,
3360                 .proc_handler   = proc_dointvec_minmax,
3361                 .extra1         = SYSCTL_ONE,
3362         },
3363         { }
3364 };
3365
3366 static struct ctl_table debug_table[] = {
3367 #ifdef CONFIG_SYSCTL_EXCEPTION_TRACE
3368         {
3369                 .procname       = "exception-trace",
3370                 .data           = &show_unhandled_signals,
3371                 .maxlen         = sizeof(int),
3372                 .mode           = 0644,
3373                 .proc_handler   = proc_dointvec
3374         },
3375 #endif
3376 #if defined(CONFIG_OPTPROBES)
3377         {
3378                 .procname       = "kprobes-optimization",
3379                 .data           = &sysctl_kprobes_optimization,
3380                 .maxlen         = sizeof(int),
3381                 .mode           = 0644,
3382                 .proc_handler   = proc_kprobes_optimization_handler,
3383                 .extra1         = SYSCTL_ZERO,
3384                 .extra2         = SYSCTL_ONE,
3385         },
3386 #endif
3387         { }
3388 };
3389
3390 static struct ctl_table dev_table[] = {
3391         { }
3392 };
3393
3394 static struct ctl_table sysctl_base_table[] = {
3395         {
3396                 .procname       = "kernel",
3397                 .mode           = 0555,
3398                 .child          = kern_table,
3399         },
3400         {
3401                 .procname       = "vm",
3402                 .mode           = 0555,
3403                 .child          = vm_table,
3404         },
3405         {
3406                 .procname       = "fs",
3407                 .mode           = 0555,
3408                 .child          = fs_table,
3409         },
3410         {
3411                 .procname       = "debug",
3412                 .mode           = 0555,
3413                 .child          = debug_table,
3414         },
3415         {
3416                 .procname       = "dev",
3417                 .mode           = 0555,
3418                 .child          = dev_table,
3419         },
3420         { }
3421 };
3422
3423 int __init sysctl_init(void)
3424 {
3425         struct ctl_table_header *hdr;
3426
3427         hdr = register_sysctl_table(sysctl_base_table);
3428         kmemleak_not_leak(hdr);
3429         return 0;
3430 }
3431 #endif /* CONFIG_SYSCTL */
3432 /*
3433  * No sense putting this after each symbol definition, twice,
3434  * exception granted :-)
3435  */
3436 EXPORT_SYMBOL(proc_dointvec);
3437 EXPORT_SYMBOL(proc_douintvec);
3438 EXPORT_SYMBOL(proc_dointvec_jiffies);
3439 EXPORT_SYMBOL(proc_dointvec_minmax);
3440 EXPORT_SYMBOL_GPL(proc_douintvec_minmax);
3441 EXPORT_SYMBOL(proc_dointvec_userhz_jiffies);
3442 EXPORT_SYMBOL(proc_dointvec_ms_jiffies);
3443 EXPORT_SYMBOL(proc_dostring);
3444 EXPORT_SYMBOL(proc_doulongvec_minmax);
3445 EXPORT_SYMBOL(proc_doulongvec_ms_jiffies_minmax);
3446 EXPORT_SYMBOL(proc_do_large_bitmap);