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/mm.h>
  24 #include <linux/swap.h>
  25 #include <linux/slab.h>
  26 #include <linux/sysctl.h>
  27 #include <linux/bitmap.h>
  28 #include <linux/signal.h>
  29 #include <linux/panic.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/filter.h>
  36 #include <linux/fs.h>
  37 #include <linux/init.h>
  38 #include <linux/kernel.h>
  39 #include <linux/kobject.h>
  40 #include <linux/net.h>
  41 #include <linux/sysrq.h>
  42 #include <linux/highuid.h>
  43 #include <linux/writeback.h>
  44 #include <linux/ratelimit.h>
  45 #include <linux/compaction.h>
  46 #include <linux/hugetlb.h>
  47 #include <linux/initrd.h>
  48 #include <linux/key.h>
  49 #include <linux/times.h>
  50 #include <linux/limits.h>
  51 #include <linux/dcache.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/oom.h>
  60 #include <linux/kmod.h>
  61 #include <linux/capability.h>
  62 #include <linux/binfmts.h>
  63 #include <linux/sched/sysctl.h>
  64 #include <linux/mount.h>
  65 #include <linux/userfaultfd_k.h>
  66 #include <linux/pid.h>
  67
  68 #include "../lib/kstrtox.h"
  69
  70 #include <linux/uaccess.h>
  71 #include <asm/processor.h>
  72
  73 #ifdef CONFIG_X86
  74 #include <asm/nmi.h>
  75 #include <asm/stacktrace.h>
  76 #include <asm/io.h>
  77 #endif
  78 #ifdef CONFIG_SPARC
  79 #include <asm/setup.h>
  80 #endif
  81 #ifdef CONFIG_RT_MUTEXES
  82 #include <linux/rtmutex.h>
  83 #endif
  84
  85 /* shared constants to be used in various sysctls */
  86 const int sysctl_vals[] = { 0, 1, 2, 3, 4, 100, 200, 1000, 3000, INT_MAX, 65535, -1 };
  87 EXPORT_SYMBOL(sysctl_vals);
  88
  89 const unsigned long sysctl_long_vals[] = { 0, 1, LONG_MAX };
  90 EXPORT_SYMBOL_GPL(sysctl_long_vals);
  91
  92 #if defined(CONFIG_SYSCTL)
  93
  94 /* Constants used for minimum and maximum */
  95
  96 #ifdef CONFIG_PERF_EVENTS
  97 static const int six_hundred_forty_kb = 640 * 1024;
  98 #endif
  99
 100
 101 static const int ngroups_max = NGROUPS_MAX;
 102 static const int cap_last_cap = CAP_LAST_CAP;
 103
 104 #ifdef CONFIG_PROC_SYSCTL
 105
 106 /**
 107  * enum sysctl_writes_mode - supported sysctl write modes
 108  *
 109  * @SYSCTL_WRITES_LEGACY: each write syscall must fully contain the sysctl value
 110  *      to be written, and multiple writes on the same sysctl file descriptor
 111  *      will rewrite the sysctl value, regardless of file position. No warning
 112  *      is issued when the initial position is not 0.
 113  * @SYSCTL_WRITES_WARN: same as above but warn when the initial file position is
 114  *      not 0.
 115  * @SYSCTL_WRITES_STRICT: writes to numeric sysctl entries must always be at
 116  *      file position 0 and the value must be fully contained in the buffer
 117  *      sent to the write syscall. If dealing with strings respect the file
 118  *      position, but restrict this to the max length of the buffer, anything
 119  *      passed the max length will be ignored. Multiple writes will append
 120  *      to the buffer.
 121  *
 122  * These write modes control how current file position affects the behavior of
 123  * updating sysctl values through the proc interface on each write.
 124  */
 125 enum sysctl_writes_mode {
 126         SYSCTL_WRITES_LEGACY            = -1,
 127         SYSCTL_WRITES_WARN              = 0,
 128         SYSCTL_WRITES_STRICT            = 1,
 129 };
 130
 131 static enum sysctl_writes_mode sysctl_writes_strict = SYSCTL_WRITES_STRICT;
 132 #endif /* CONFIG_PROC_SYSCTL */
 133
 134 #if defined(HAVE_ARCH_PICK_MMAP_LAYOUT) || \
 135     defined(CONFIG_ARCH_WANT_DEFAULT_TOPDOWN_MMAP_LAYOUT)
 136 int sysctl_legacy_va_layout;
 137 #endif
 138
 139 #endif /* CONFIG_SYSCTL */
 140
 141 /*
 142  * /proc/sys support
 143  */
 144
 145 #ifdef CONFIG_PROC_SYSCTL
 146
 147 static int _proc_do_string(char *data, int maxlen, int write,
 148                 char *buffer, size_t *lenp, loff_t *ppos)
 149 {
 150         size_t len;
 151         char c, *p;
 152
 153         if (!data || !maxlen || !*lenp) {
 154                 *lenp = 0;
 155                 return 0;
 156         }
 157
 158         if (write) {
 159                 if (sysctl_writes_strict == SYSCTL_WRITES_STRICT) {
 160                         /* Only continue writes not past the end of buffer. */
 161                         len = strlen(data);
 162                         if (len > maxlen - 1)
 163                                 len = maxlen - 1;
 164
 165                         if (*ppos > len)
 166                                 return 0;
 167                         len = *ppos;
 168                 } else {
 169                         /* Start writing from beginning of buffer. */
 170                         len = 0;
 171                 }
 172
 173                 *ppos += *lenp;
 174                 p = buffer;
 175                 while ((p - buffer) < *lenp && len < maxlen - 1) {
 176                         c = *(p++);
 177                         if (c == 0 || c == '\n')
 178                                 break;
 179                         data[len++] = c;
 180                 }
 181                 data[len] = 0;
 182         } else {
 183                 len = strlen(data);
 184                 if (len > maxlen)
 185                         len = maxlen;
 186
 187                 if (*ppos > len) {
 188                         *lenp = 0;
 189                         return 0;
 190                 }
 191
 192                 data += *ppos;
 193                 len  -= *ppos;
 194
 195                 if (len > *lenp)
 196                         len = *lenp;
 197                 if (len)
 198                         memcpy(buffer, data, len);
 199                 if (len < *lenp) {
 200                         buffer[len] = '\n';
 201                         len++;
 202                 }
 203                 *lenp = len;
 204                 *ppos += len;
 205         }
 206         return 0;
 207 }
 208
 209 static void warn_sysctl_write(struct ctl_table *table)
 210 {
 211         pr_warn_once("%s wrote to %s when file position was not 0!\n"
 212                 "This will not be supported in the future. To silence this\n"
 213                 "warning, set kernel.sysctl_writes_strict = -1\n",
 214                 current->comm, table->procname);
 215 }
 216
 217 /**
 218  * proc_first_pos_non_zero_ignore - check if first position is allowed
 219  * @ppos: file position
 220  * @table: the sysctl table
 221  *
 222  * Returns true if the first position is non-zero and the sysctl_writes_strict
 223  * mode indicates this is not allowed for numeric input types. String proc
 224  * handlers can ignore the return value.
 225  */
 226 static bool proc_first_pos_non_zero_ignore(loff_t *ppos,
 227                                            struct ctl_table *table)
 228 {
 229         if (!*ppos)
 230                 return false;
 231
 232         switch (sysctl_writes_strict) {
 233         case SYSCTL_WRITES_STRICT:
 234                 return true;
 235         case SYSCTL_WRITES_WARN:
 236                 warn_sysctl_write(table);
 237                 return false;
 238         default:
 239                 return false;
 240         }
 241 }
 242
 243 /**
 244  * proc_dostring - read a string sysctl
 245  * @table: the sysctl table
 246  * @write: %TRUE if this is a write to the sysctl file
 247  * @buffer: the user buffer
 248  * @lenp: the size of the user buffer
 249  * @ppos: file position
 250  *
 251  * Reads/writes a string from/to the user buffer. If the kernel
 252  * buffer provided is not large enough to hold the string, the
 253  * string is truncated. The copied string is %NULL-terminated.
 254  * If the string is being read by the user process, it is copied
 255  * and a newline '\n' is added. It is truncated if the buffer is
 256  * not large enough.
 257  *
 258  * Returns 0 on success.
 259  */
 260 int proc_dostring(struct ctl_table *table, int write,
 261                   void *buffer, size_t *lenp, loff_t *ppos)
 262 {
 263         if (write)
 264                 proc_first_pos_non_zero_ignore(ppos, table);
 265
 266         return _proc_do_string(table->data, table->maxlen, write, buffer, lenp,
 267                         ppos);
 268 }
 269
 270 static void proc_skip_spaces(char **buf, size_t *size)
 271 {
 272         while (*size) {
 273                 if (!isspace(**buf))
 274                         break;
 275                 (*size)--;
 276                 (*buf)++;
 277         }
 278 }
 279
 280 static void proc_skip_char(char **buf, size_t *size, const char v)
 281 {
 282         while (*size) {
 283                 if (**buf != v)
 284                         break;
 285                 (*size)--;
 286                 (*buf)++;
 287         }
 288 }
 289
 290 /**
 291  * strtoul_lenient - parse an ASCII formatted integer from a buffer and only
 292  *                   fail on overflow
 293  *
 294  * @cp: kernel buffer containing the string to parse
 295  * @endp: pointer to store the trailing characters
 296  * @base: the base to use
 297  * @res: where the parsed integer will be stored
 298  *
 299  * In case of success 0 is returned and @res will contain the parsed integer,
 300  * @endp will hold any trailing characters.
 301  * This function will fail the parse on overflow. If there wasn't an overflow
 302  * the function will defer the decision what characters count as invalid to the
 303  * caller.
 304  */
 305 static int strtoul_lenient(const char *cp, char **endp, unsigned int base,
 306                            unsigned long *res)
 307 {
 308         unsigned long long result;
 309         unsigned int rv;
 310
 311         cp = _parse_integer_fixup_radix(cp, &base);
 312         rv = _parse_integer(cp, base, &result);
 313         if ((rv & KSTRTOX_OVERFLOW) || (result != (unsigned long)result))
 314                 return -ERANGE;
 315
 316         cp += rv;
 317
 318         if (endp)
 319                 *endp = (char *)cp;
 320
 321         *res = (unsigned long)result;
 322         return 0;
 323 }
 324
 325 #define TMPBUFLEN 22
 326 /**
 327  * proc_get_long - reads an ASCII formatted integer from a user buffer
 328  *
 329  * @buf: a kernel buffer
 330  * @size: size of the kernel buffer
 331  * @val: this is where the number will be stored
 332  * @neg: set to %TRUE if number is negative
 333  * @perm_tr: a vector which contains the allowed trailers
 334  * @perm_tr_len: size of the perm_tr vector
 335  * @tr: pointer to store the trailer character
 336  *
 337  * In case of success %0 is returned and @buf and @size are updated with
 338  * the amount of bytes read. If @tr is non-NULL and a trailing
 339  * character exists (size is non-zero after returning from this
 340  * function), @tr is updated with the trailing character.
 341  */
 342 static int proc_get_long(char **buf, size_t *size,
 343                           unsigned long *val, bool *neg,
 344                           const char *perm_tr, unsigned perm_tr_len, char *tr)
 345 {
 346         char *p, tmp[TMPBUFLEN];
 347         ssize_t len = *size;
 348
 349         if (len <= 0)
 350                 return -EINVAL;
 351
 352         if (len > TMPBUFLEN - 1)
 353                 len = TMPBUFLEN - 1;
 354
 355         memcpy(tmp, *buf, len);
 356
 357         tmp[len] = 0;
 358         p = tmp;
 359         if (*p == '-' && *size > 1) {
 360                 *neg = true;
 361                 p++;
 362         } else
 363                 *neg = false;
 364         if (!isdigit(*p))
 365                 return -EINVAL;
 366
 367         if (strtoul_lenient(p, &p, 0, val))
 368                 return -EINVAL;
 369
 370         len = p - tmp;
 371
 372         /* We don't know if the next char is whitespace thus we may accept
 373          * invalid integers (e.g. 1234...a) or two integers instead of one
 374          * (e.g. 123...1). So lets not allow such large numbers. */
 375         if (len == TMPBUFLEN - 1)
 376                 return -EINVAL;
 377
 378         if (len < *size && perm_tr_len && !memchr(perm_tr, *p, perm_tr_len))
 379                 return -EINVAL;
 380
 381         if (tr && (len < *size))
 382                 *tr = *p;
 383
 384         *buf += len;
 385         *size -= len;
 386
 387         return 0;
 388 }
 389
 390 /**
 391  * proc_put_long - converts an integer to a decimal ASCII formatted string
 392  *
 393  * @buf: the user buffer
 394  * @size: the size of the user buffer
 395  * @val: the integer to be converted
 396  * @neg: sign of the number, %TRUE for negative
 397  *
 398  * In case of success @buf and @size are updated with the amount of bytes
 399  * written.
 400  */
 401 static void proc_put_long(void **buf, size_t *size, unsigned long val, bool neg)
 402 {
 403         int len;
 404         char tmp[TMPBUFLEN], *p = tmp;
 405
 406         sprintf(p, "%s%lu", neg ? "-" : "", val);
 407         len = strlen(tmp);
 408         if (len > *size)
 409                 len = *size;
 410         memcpy(*buf, tmp, len);
 411         *size -= len;
 412         *buf += len;
 413 }
 414 #undef TMPBUFLEN
 415
 416 static void proc_put_char(void **buf, size_t *size, char c)
 417 {
 418         if (*size) {
 419                 char **buffer = (char **)buf;
 420                 **buffer = c;
 421
 422                 (*size)--;
 423                 (*buffer)++;
 424                 *buf = *buffer;
 425         }
 426 }
 427
 428 static int do_proc_dobool_conv(bool *negp, unsigned long *lvalp,
 429                                 int *valp,
 430                                 int write, void *data)
 431 {
 432         if (write) {
 433                 *(bool *)valp = *lvalp;
 434         } else {
 435                 int val = *(bool *)valp;
 436
 437                 *lvalp = (unsigned long)val;
 438                 *negp = false;
 439         }
 440         return 0;
 441 }
 442
 443 static int do_proc_dointvec_conv(bool *negp, unsigned long *lvalp,
 444                                  int *valp,
 445                                  int write, void *data)
 446 {
 447         if (write) {
 448                 if (*negp) {
 449                         if (*lvalp > (unsigned long) INT_MAX + 1)
 450                                 return -EINVAL;
 451                         WRITE_ONCE(*valp, -*lvalp);
 452                 } else {
 453                         if (*lvalp > (unsigned long) INT_MAX)
 454                                 return -EINVAL;
 455                         WRITE_ONCE(*valp, *lvalp);
 456                 }
 457         } else {
 458                 int val = READ_ONCE(*valp);
 459                 if (val < 0) {
 460                         *negp = true;
 461                         *lvalp = -(unsigned long)val;
 462                 } else {
 463                         *negp = false;
 464                         *lvalp = (unsigned long)val;
 465                 }
 466         }
 467         return 0;
 468 }
 469
 470 static int do_proc_douintvec_conv(unsigned long *lvalp,
 471                                   unsigned int *valp,
 472                                   int write, void *data)
 473 {
 474         if (write) {
 475                 if (*lvalp > UINT_MAX)
 476                         return -EINVAL;
 477                 WRITE_ONCE(*valp, *lvalp);
 478         } else {
 479                 unsigned int val = READ_ONCE(*valp);
 480                 *lvalp = (unsigned long)val;
 481         }
 482         return 0;
 483 }
 484
 485 static const char proc_wspace_sep[] = { ' ', '\t', '\n' };
 486
 487 static int __do_proc_dointvec(void *tbl_data, struct ctl_table *table,
 488                   int write, void *buffer,
 489                   size_t *lenp, loff_t *ppos,
 490                   int (*conv)(bool *negp, unsigned long *lvalp, int *valp,
 491                               int write, void *data),
 492                   void *data)
 493 {
 494         int *i, vleft, first = 1, err = 0;
 495         size_t left;
 496         char *p;
 497
 498         if (!tbl_data || !table->maxlen || !*lenp || (*ppos && !write)) {
 499                 *lenp = 0;
 500                 return 0;
 501         }
 502
 503         i = (int *) tbl_data;
 504         vleft = table->maxlen / sizeof(*i);
 505         left = *lenp;
 506
 507         if (!conv)
 508                 conv = do_proc_dointvec_conv;
 509
 510         if (write) {
 511                 if (proc_first_pos_non_zero_ignore(ppos, table))
 512                         goto out;
 513
 514                 if (left > PAGE_SIZE - 1)
 515                         left = PAGE_SIZE - 1;
 516                 p = buffer;
 517         }
 518
 519         for (; left && vleft--; i++, first=0) {
 520                 unsigned long lval;
 521                 bool neg;
 522
 523                 if (write) {
 524                         proc_skip_spaces(&p, &left);
 525
 526                         if (!left)
 527                                 break;
 528                         err = proc_get_long(&p, &left, &lval, &neg,
 529                                              proc_wspace_sep,
 530                                              sizeof(proc_wspace_sep), NULL);
 531                         if (err)
 532                                 break;
 533                         if (conv(&neg, &lval, i, 1, data)) {
 534                                 err = -EINVAL;
 535                                 break;
 536                         }
 537                 } else {
 538                         if (conv(&neg, &lval, i, 0, data)) {
 539                                 err = -EINVAL;
 540                                 break;
 541                         }
 542                         if (!first)
 543                                 proc_put_char(&buffer, &left, '\t');
 544                         proc_put_long(&buffer, &left, lval, neg);
 545                 }
 546         }
 547
 548         if (!write && !first && left && !err)
 549                 proc_put_char(&buffer, &left, '\n');
 550         if (write && !err && left)
 551                 proc_skip_spaces(&p, &left);
 552         if (write && first)
 553                 return err ? : -EINVAL;
 554         *lenp -= left;
 555 out:
 556         *ppos += *lenp;
 557         return err;
 558 }
 559
 560 static int do_proc_dointvec(struct ctl_table *table, int write,
 561                   void *buffer, size_t *lenp, loff_t *ppos,
 562                   int (*conv)(bool *negp, unsigned long *lvalp, int *valp,
 563                               int write, void *data),
 564                   void *data)
 565 {
 566         return __do_proc_dointvec(table->data, table, write,
 567                         buffer, lenp, ppos, conv, data);
 568 }
 569
 570 static int do_proc_douintvec_w(unsigned int *tbl_data,
 571                                struct ctl_table *table,
 572                                void *buffer,
 573                                size_t *lenp, loff_t *ppos,
 574                                int (*conv)(unsigned long *lvalp,
 575                                            unsigned int *valp,
 576                                            int write, void *data),
 577                                void *data)
 578 {
 579         unsigned long lval;
 580         int err = 0;
 581         size_t left;
 582         bool neg;
 583         char *p = buffer;
 584
 585         left = *lenp;
 586
 587         if (proc_first_pos_non_zero_ignore(ppos, table))
 588                 goto bail_early;
 589
 590         if (left > PAGE_SIZE - 1)
 591                 left = PAGE_SIZE - 1;
 592
 593         proc_skip_spaces(&p, &left);
 594         if (!left) {
 595                 err = -EINVAL;
 596                 goto out_free;
 597         }
 598
 599         err = proc_get_long(&p, &left, &lval, &neg,
 600                              proc_wspace_sep,
 601                              sizeof(proc_wspace_sep), NULL);
 602         if (err || neg) {
 603                 err = -EINVAL;
 604                 goto out_free;
 605         }
 606
 607         if (conv(&lval, tbl_data, 1, data)) {
 608                 err = -EINVAL;
 609                 goto out_free;
 610         }
 611
 612         if (!err && left)
 613                 proc_skip_spaces(&p, &left);
 614
 615 out_free:
 616         if (err)
 617                 return -EINVAL;
 618
 619         return 0;
 620
 621         /* This is in keeping with old __do_proc_dointvec() */
 622 bail_early:
 623         *ppos += *lenp;
 624         return err;
 625 }
 626
 627 static int do_proc_douintvec_r(unsigned int *tbl_data, void *buffer,
 628                                size_t *lenp, loff_t *ppos,
 629                                int (*conv)(unsigned long *lvalp,
 630                                            unsigned int *valp,
 631                                            int write, void *data),
 632                                void *data)
 633 {
 634         unsigned long lval;
 635         int err = 0;
 636         size_t left;
 637
 638         left = *lenp;
 639
 640         if (conv(&lval, tbl_data, 0, data)) {
 641                 err = -EINVAL;
 642                 goto out;
 643         }
 644
 645         proc_put_long(&buffer, &left, lval, false);
 646         if (!left)
 647                 goto out;
 648
 649         proc_put_char(&buffer, &left, '\n');
 650
 651 out:
 652         *lenp -= left;
 653         *ppos += *lenp;
 654
 655         return err;
 656 }
 657
 658 static int __do_proc_douintvec(void *tbl_data, struct ctl_table *table,
 659                                int write, void *buffer,
 660                                size_t *lenp, loff_t *ppos,
 661                                int (*conv)(unsigned long *lvalp,
 662                                            unsigned int *valp,
 663                                            int write, void *data),
 664                                void *data)
 665 {
 666         unsigned int *i, vleft;
 667
 668         if (!tbl_data || !table->maxlen || !*lenp || (*ppos && !write)) {
 669                 *lenp = 0;
 670                 return 0;
 671         }
 672
 673         i = (unsigned int *) tbl_data;
 674         vleft = table->maxlen / sizeof(*i);
 675
 676         /*
 677          * Arrays are not supported, keep this simple. *Do not* add
 678          * support for them.
 679          */
 680         if (vleft != 1) {
 681                 *lenp = 0;
 682                 return -EINVAL;
 683         }
 684
 685         if (!conv)
 686                 conv = do_proc_douintvec_conv;
 687
 688         if (write)
 689                 return do_proc_douintvec_w(i, table, buffer, lenp, ppos,
 690                                            conv, data);
 691         return do_proc_douintvec_r(i, buffer, lenp, ppos, conv, data);
 692 }
 693
 694 int do_proc_douintvec(struct ctl_table *table, int write,
 695                       void *buffer, size_t *lenp, loff_t *ppos,
 696                       int (*conv)(unsigned long *lvalp,
 697                                   unsigned int *valp,
 698                                   int write, void *data),
 699                       void *data)
 700 {
 701         return __do_proc_douintvec(table->data, table, write,
 702                                    buffer, lenp, ppos, conv, data);
 703 }
 704
 705 /**
 706  * proc_dobool - read/write a bool
 707  * @table: the sysctl table
 708  * @write: %TRUE if this is a write to the sysctl file
 709  * @buffer: the user buffer
 710  * @lenp: the size of the user buffer
 711  * @ppos: file position
 712  *
 713  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
 714  * values from/to the user buffer, treated as an ASCII string.
 715  *
 716  * Returns 0 on success.
 717  */
 718 int proc_dobool(struct ctl_table *table, int write, void *buffer,
 719                 size_t *lenp, loff_t *ppos)
 720 {
 721         return do_proc_dointvec(table, write, buffer, lenp, ppos,
 722                                 do_proc_dobool_conv, NULL);
 723 }
 724
 725 /**
 726  * proc_dointvec - read a vector of integers
 727  * @table: the sysctl table
 728  * @write: %TRUE if this is a write to the sysctl file
 729  * @buffer: the user buffer
 730  * @lenp: the size of the user buffer
 731  * @ppos: file position
 732  *
 733  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
 734  * values from/to the user buffer, treated as an ASCII string.
 735  *
 736  * Returns 0 on success.
 737  */
 738 int proc_dointvec(struct ctl_table *table, int write, void *buffer,
 739                   size_t *lenp, loff_t *ppos)
 740 {
 741         return do_proc_dointvec(table, write, buffer, lenp, ppos, NULL, NULL);
 742 }
 743
 744 #ifdef CONFIG_COMPACTION
 745 static int proc_dointvec_minmax_warn_RT_change(struct ctl_table *table,
 746                 int write, void *buffer, size_t *lenp, loff_t *ppos)
 747 {
 748         int ret, old;
 749
 750         if (!IS_ENABLED(CONFIG_PREEMPT_RT) || !write)
 751                 return proc_dointvec_minmax(table, write, buffer, lenp, ppos);
 752
 753         old = *(int *)table->data;
 754         ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos);
 755         if (ret)
 756                 return ret;
 757         if (old != *(int *)table->data)
 758                 pr_warn_once("sysctl attribute %s changed by %s[%d]\n",
 759                              table->procname, current->comm,
 760                              task_pid_nr(current));
 761         return ret;
 762 }
 763 #endif
 764
 765 /**
 766  * proc_douintvec - read a vector of unsigned integers
 767  * @table: the sysctl table
 768  * @write: %TRUE if this is a write to the sysctl file
 769  * @buffer: the user buffer
 770  * @lenp: the size of the user buffer
 771  * @ppos: file position
 772  *
 773  * Reads/writes up to table->maxlen/sizeof(unsigned int) unsigned integer
 774  * values from/to the user buffer, treated as an ASCII string.
 775  *
 776  * Returns 0 on success.
 777  */
 778 int proc_douintvec(struct ctl_table *table, int write, void *buffer,
 779                 size_t *lenp, loff_t *ppos)
 780 {
 781         return do_proc_douintvec(table, write, buffer, lenp, ppos,
 782                                  do_proc_douintvec_conv, NULL);
 783 }
 784
 785 /*
 786  * Taint values can only be increased
 787  * This means we can safely use a temporary.
 788  */
 789 static int proc_taint(struct ctl_table *table, int write,
 790                                void *buffer, size_t *lenp, loff_t *ppos)
 791 {
 792         struct ctl_table t;
 793         unsigned long tmptaint = get_taint();
 794         int err;
 795
 796         if (write && !capable(CAP_SYS_ADMIN))
 797                 return -EPERM;
 798
 799         t = *table;
 800         t.data = &tmptaint;
 801         err = proc_doulongvec_minmax(&t, write, buffer, lenp, ppos);
 802         if (err < 0)
 803                 return err;
 804
 805         if (write) {
 806                 int i;
 807
 808                 /*
 809                  * If we are relying on panic_on_taint not producing
 810                  * false positives due to userspace input, bail out
 811                  * before setting the requested taint flags.
 812                  */
 813                 if (panic_on_taint_nousertaint && (tmptaint & panic_on_taint))
 814                         return -EINVAL;
 815
 816                 /*
 817                  * Poor man's atomic or. Not worth adding a primitive
 818                  * to everyone's atomic.h for this
 819                  */
 820                 for (i = 0; i < TAINT_FLAGS_COUNT; i++)
 821                         if ((1UL << i) & tmptaint)
 822                                 add_taint(i, LOCKDEP_STILL_OK);
 823         }
 824
 825         return err;
 826 }
 827
 828 /**
 829  * struct do_proc_dointvec_minmax_conv_param - proc_dointvec_minmax() range checking structure
 830  * @min: pointer to minimum allowable value
 831  * @max: pointer to maximum allowable value
 832  *
 833  * The do_proc_dointvec_minmax_conv_param structure provides the
 834  * minimum and maximum values for doing range checking for those sysctl
 835  * parameters that use the proc_dointvec_minmax() handler.
 836  */
 837 struct do_proc_dointvec_minmax_conv_param {
 838         int *min;
 839         int *max;
 840 };
 841
 842 static int do_proc_dointvec_minmax_conv(bool *negp, unsigned long *lvalp,
 843                                         int *valp,
 844                                         int write, void *data)
 845 {
 846         int tmp, ret;
 847         struct do_proc_dointvec_minmax_conv_param *param = data;
 848         /*
 849          * If writing, first do so via a temporary local int so we can
 850          * bounds-check it before touching *valp.
 851          */
 852         int *ip = write ? &tmp : valp;
 853
 854         ret = do_proc_dointvec_conv(negp, lvalp, ip, write, data);
 855         if (ret)
 856                 return ret;
 857
 858         if (write) {
 859                 if ((param->min && *param->min > tmp) ||
 860                     (param->max && *param->max < tmp))
 861                         return -EINVAL;
 862                 WRITE_ONCE(*valp, tmp);
 863         }
 864
 865         return 0;
 866 }
 867
 868 /**
 869  * proc_dointvec_minmax - read a vector of integers with min/max values
 870  * @table: the sysctl table
 871  * @write: %TRUE if this is a write to the sysctl file
 872  * @buffer: the user buffer
 873  * @lenp: the size of the user buffer
 874  * @ppos: file position
 875  *
 876  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
 877  * values from/to the user buffer, treated as an ASCII string.
 878  *
 879  * This routine will ensure the values are within the range specified by
 880  * table->extra1 (min) and table->extra2 (max).
 881  *
 882  * Returns 0 on success or -EINVAL on write when the range check fails.
 883  */
 884 int proc_dointvec_minmax(struct ctl_table *table, int write,
 885                   void *buffer, size_t *lenp, loff_t *ppos)
 886 {
 887         struct do_proc_dointvec_minmax_conv_param param = {
 888                 .min = (int *) table->extra1,
 889                 .max = (int *) table->extra2,
 890         };
 891         return do_proc_dointvec(table, write, buffer, lenp, ppos,
 892                                 do_proc_dointvec_minmax_conv, &param);
 893 }
 894
 895 /**
 896  * struct do_proc_douintvec_minmax_conv_param - proc_douintvec_minmax() range checking structure
 897  * @min: pointer to minimum allowable value
 898  * @max: pointer to maximum allowable value
 899  *
 900  * The do_proc_douintvec_minmax_conv_param structure provides the
 901  * minimum and maximum values for doing range checking for those sysctl
 902  * parameters that use the proc_douintvec_minmax() handler.
 903  */
 904 struct do_proc_douintvec_minmax_conv_param {
 905         unsigned int *min;
 906         unsigned int *max;
 907 };
 908
 909 static int do_proc_douintvec_minmax_conv(unsigned long *lvalp,
 910                                          unsigned int *valp,
 911                                          int write, void *data)
 912 {
 913         int ret;
 914         unsigned int tmp;
 915         struct do_proc_douintvec_minmax_conv_param *param = data;
 916         /* write via temporary local uint for bounds-checking */
 917         unsigned int *up = write ? &tmp : valp;
 918
 919         ret = do_proc_douintvec_conv(lvalp, up, write, data);
 920         if (ret)
 921                 return ret;
 922
 923         if (write) {
 924                 if ((param->min && *param->min > tmp) ||
 925                     (param->max && *param->max < tmp))
 926                         return -ERANGE;
 927
 928                 WRITE_ONCE(*valp, tmp);
 929         }
 930
 931         return 0;
 932 }
 933
 934 /**
 935  * proc_douintvec_minmax - read a vector of unsigned ints with min/max values
 936  * @table: the sysctl table
 937  * @write: %TRUE if this is a write to the sysctl file
 938  * @buffer: the user buffer
 939  * @lenp: the size of the user buffer
 940  * @ppos: file position
 941  *
 942  * Reads/writes up to table->maxlen/sizeof(unsigned int) unsigned integer
 943  * values from/to the user buffer, treated as an ASCII string. Negative
 944  * strings are not allowed.
 945  *
 946  * This routine will ensure the values are within the range specified by
 947  * table->extra1 (min) and table->extra2 (max). There is a final sanity
 948  * check for UINT_MAX to avoid having to support wrap around uses from
 949  * userspace.
 950  *
 951  * Returns 0 on success or -ERANGE on write when the range check fails.
 952  */
 953 int proc_douintvec_minmax(struct ctl_table *table, int write,
 954                           void *buffer, size_t *lenp, loff_t *ppos)
 955 {
 956         struct do_proc_douintvec_minmax_conv_param param = {
 957                 .min = (unsigned int *) table->extra1,
 958                 .max = (unsigned int *) table->extra2,
 959         };
 960         return do_proc_douintvec(table, write, buffer, lenp, ppos,
 961                                  do_proc_douintvec_minmax_conv, &param);
 962 }
 963
 964 /**
 965  * proc_dou8vec_minmax - read a vector of unsigned chars with min/max values
 966  * @table: the sysctl table
 967  * @write: %TRUE if this is a write to the sysctl file
 968  * @buffer: the user buffer
 969  * @lenp: the size of the user buffer
 970  * @ppos: file position
 971  *
 972  * Reads/writes up to table->maxlen/sizeof(u8) unsigned chars
 973  * values from/to the user buffer, treated as an ASCII string. Negative
 974  * strings are not allowed.
 975  *
 976  * This routine will ensure the values are within the range specified by
 977  * table->extra1 (min) and table->extra2 (max).
 978  *
 979  * Returns 0 on success or an error on write when the range check fails.
 980  */
 981 int proc_dou8vec_minmax(struct ctl_table *table, int write,
 982                         void *buffer, size_t *lenp, loff_t *ppos)
 983 {
 984         struct ctl_table tmp;
 985         unsigned int min = 0, max = 255U, val;
 986         u8 *data = table->data;
 987         struct do_proc_douintvec_minmax_conv_param param = {
 988                 .min = &min,
 989                 .max = &max,
 990         };
 991         int res;
 992
 993         /* Do not support arrays yet. */
 994         if (table->maxlen != sizeof(u8))
 995                 return -EINVAL;
 996
 997         if (table->extra1) {
 998                 min = *(unsigned int *) table->extra1;
 999                 if (min > 255U)
1000                         return -EINVAL;
1001         }
1002         if (table->extra2) {
1003                 max = *(unsigned int *) table->extra2;
1004                 if (max > 255U)
1005                         return -EINVAL;
1006         }
1007
1008         tmp = *table;
1009
1010         tmp.maxlen = sizeof(val);
1011         tmp.data = &val;
1012         val = READ_ONCE(*data);
1013         res = do_proc_douintvec(&tmp, write, buffer, lenp, ppos,
1014                                 do_proc_douintvec_minmax_conv, &param);
1015         if (res)
1016                 return res;
1017         if (write)
1018                 WRITE_ONCE(*data, val);
1019         return 0;
1020 }
1021 EXPORT_SYMBOL_GPL(proc_dou8vec_minmax);
1022
1023 #ifdef CONFIG_MAGIC_SYSRQ
1024 static int sysrq_sysctl_handler(struct ctl_table *table, int write,
1025                                 void *buffer, size_t *lenp, loff_t *ppos)
1026 {
1027         int tmp, ret;
1028
1029         tmp = sysrq_mask();
1030
1031         ret = __do_proc_dointvec(&tmp, table, write, buffer,
1032                                lenp, ppos, NULL, NULL);
1033         if (ret || !write)
1034                 return ret;
1035
1036         if (write)
1037                 sysrq_toggle_support(tmp);
1038
1039         return 0;
1040 }
1041 #endif
1042
1043 static int __do_proc_doulongvec_minmax(void *data, struct ctl_table *table,
1044                 int write, void *buffer, size_t *lenp, loff_t *ppos,
1045                 unsigned long convmul, unsigned long convdiv)
1046 {
1047         unsigned long *i, *min, *max;
1048         int vleft, first = 1, err = 0;
1049         size_t left;
1050         char *p;
1051
1052         if (!data || !table->maxlen || !*lenp || (*ppos && !write)) {
1053                 *lenp = 0;
1054                 return 0;
1055         }
1056
1057         i = data;
1058         min = table->extra1;
1059         max = table->extra2;
1060         vleft = table->maxlen / sizeof(unsigned long);
1061         left = *lenp;
1062
1063         if (write) {
1064                 if (proc_first_pos_non_zero_ignore(ppos, table))
1065                         goto out;
1066
1067                 if (left > PAGE_SIZE - 1)
1068                         left = PAGE_SIZE - 1;
1069                 p = buffer;
1070         }
1071
1072         for (; left && vleft--; i++, first = 0) {
1073                 unsigned long val;
1074
1075                 if (write) {
1076                         bool neg;
1077
1078                         proc_skip_spaces(&p, &left);
1079                         if (!left)
1080                                 break;
1081
1082                         err = proc_get_long(&p, &left, &val, &neg,
1083                                              proc_wspace_sep,
1084                                              sizeof(proc_wspace_sep), NULL);
1085                         if (err || neg) {
1086                                 err = -EINVAL;
1087                                 break;
1088                         }
1089
1090                         val = convmul * val / convdiv;
1091                         if ((min && val < *min) || (max && val > *max)) {
1092                                 err = -EINVAL;
1093                                 break;
1094                         }
1095                         WRITE_ONCE(*i, val);
1096                 } else {
1097                         val = convdiv * READ_ONCE(*i) / convmul;
1098                         if (!first)
1099                                 proc_put_char(&buffer, &left, '\t');
1100                         proc_put_long(&buffer, &left, val, false);
1101                 }
1102         }
1103
1104         if (!write && !first && left && !err)
1105                 proc_put_char(&buffer, &left, '\n');
1106         if (write && !err)
1107                 proc_skip_spaces(&p, &left);
1108         if (write && first)
1109                 return err ? : -EINVAL;
1110         *lenp -= left;
1111 out:
1112         *ppos += *lenp;
1113         return err;
1114 }
1115
1116 static int do_proc_doulongvec_minmax(struct ctl_table *table, int write,
1117                 void *buffer, size_t *lenp, loff_t *ppos, unsigned long convmul,
1118                 unsigned long convdiv)
1119 {
1120         return __do_proc_doulongvec_minmax(table->data, table, write,
1121                         buffer, lenp, ppos, convmul, convdiv);
1122 }
1123
1124 /**
1125  * proc_doulongvec_minmax - read a vector of long integers with min/max values
1126  * @table: the sysctl table
1127  * @write: %TRUE if this is a write to the sysctl file
1128  * @buffer: the user buffer
1129  * @lenp: the size of the user buffer
1130  * @ppos: file position
1131  *
1132  * Reads/writes up to table->maxlen/sizeof(unsigned long) unsigned long
1133  * values from/to the user buffer, treated as an ASCII string.
1134  *
1135  * This routine will ensure the values are within the range specified by
1136  * table->extra1 (min) and table->extra2 (max).
1137  *
1138  * Returns 0 on success.
1139  */
1140 int proc_doulongvec_minmax(struct ctl_table *table, int write,
1141                            void *buffer, size_t *lenp, loff_t *ppos)
1142 {
1143     return do_proc_doulongvec_minmax(table, write, buffer, lenp, ppos, 1l, 1l);
1144 }
1145
1146 /**
1147  * proc_doulongvec_ms_jiffies_minmax - read a vector of millisecond values with min/max values
1148  * @table: the sysctl table
1149  * @write: %TRUE if this is a write to the sysctl file
1150  * @buffer: the user buffer
1151  * @lenp: the size of the user buffer
1152  * @ppos: file position
1153  *
1154  * Reads/writes up to table->maxlen/sizeof(unsigned long) unsigned long
1155  * values from/to the user buffer, treated as an ASCII string. The values
1156  * are treated as milliseconds, and converted to jiffies when they are stored.
1157  *
1158  * This routine will ensure the values are within the range specified by
1159  * table->extra1 (min) and table->extra2 (max).
1160  *
1161  * Returns 0 on success.
1162  */
1163 int proc_doulongvec_ms_jiffies_minmax(struct ctl_table *table, int write,
1164                                       void *buffer, size_t *lenp, loff_t *ppos)
1165 {
1166     return do_proc_doulongvec_minmax(table, write, buffer,
1167                                      lenp, ppos, HZ, 1000l);
1168 }
1169
1170
1171 static int do_proc_dointvec_jiffies_conv(bool *negp, unsigned long *lvalp,
1172                                          int *valp,
1173                                          int write, void *data)
1174 {
1175         if (write) {
1176                 if (*lvalp > INT_MAX / HZ)
1177                         return 1;
1178                 if (*negp)
1179                         WRITE_ONCE(*valp, -*lvalp * HZ);
1180                 else
1181                         WRITE_ONCE(*valp, *lvalp * HZ);
1182         } else {
1183                 int val = READ_ONCE(*valp);
1184                 unsigned long lval;
1185                 if (val < 0) {
1186                         *negp = true;
1187                         lval = -(unsigned long)val;
1188                 } else {
1189                         *negp = false;
1190                         lval = (unsigned long)val;
1191                 }
1192                 *lvalp = lval / HZ;
1193         }
1194         return 0;
1195 }
1196
1197 static int do_proc_dointvec_userhz_jiffies_conv(bool *negp, unsigned long *lvalp,
1198                                                 int *valp,
1199                                                 int write, void *data)
1200 {
1201         if (write) {
1202                 if (USER_HZ < HZ && *lvalp > (LONG_MAX / HZ) * USER_HZ)
1203                         return 1;
1204                 *valp = clock_t_to_jiffies(*negp ? -*lvalp : *lvalp);
1205         } else {
1206                 int val = *valp;
1207                 unsigned long lval;
1208                 if (val < 0) {
1209                         *negp = true;
1210                         lval = -(unsigned long)val;
1211                 } else {
1212                         *negp = false;
1213                         lval = (unsigned long)val;
1214                 }
1215                 *lvalp = jiffies_to_clock_t(lval);
1216         }
1217         return 0;
1218 }
1219
1220 static int do_proc_dointvec_ms_jiffies_conv(bool *negp, unsigned long *lvalp,
1221                                             int *valp,
1222                                             int write, void *data)
1223 {
1224         if (write) {
1225                 unsigned long jif = msecs_to_jiffies(*negp ? -*lvalp : *lvalp);
1226
1227                 if (jif > INT_MAX)
1228                         return 1;
1229                 WRITE_ONCE(*valp, (int)jif);
1230         } else {
1231                 int val = READ_ONCE(*valp);
1232                 unsigned long lval;
1233                 if (val < 0) {
1234                         *negp = true;
1235                         lval = -(unsigned long)val;
1236                 } else {
1237                         *negp = false;
1238                         lval = (unsigned long)val;
1239                 }
1240                 *lvalp = jiffies_to_msecs(lval);
1241         }
1242         return 0;
1243 }
1244
1245 static int do_proc_dointvec_ms_jiffies_minmax_conv(bool *negp, unsigned long *lvalp,
1246                                                 int *valp, int write, void *data)
1247 {
1248         int tmp, ret;
1249         struct do_proc_dointvec_minmax_conv_param *param = data;
1250         /*
1251          * If writing, first do so via a temporary local int so we can
1252          * bounds-check it before touching *valp.
1253          */
1254         int *ip = write ? &tmp : valp;
1255
1256         ret = do_proc_dointvec_ms_jiffies_conv(negp, lvalp, ip, write, data);
1257         if (ret)
1258                 return ret;
1259
1260         if (write) {
1261                 if ((param->min && *param->min > tmp) ||
1262                                 (param->max && *param->max < tmp))
1263                         return -EINVAL;
1264                 *valp = tmp;
1265         }
1266         return 0;
1267 }
1268
1269 /**
1270  * proc_dointvec_jiffies - read a vector of integers as seconds
1271  * @table: the sysctl table
1272  * @write: %TRUE if this is a write to the sysctl file
1273  * @buffer: the user buffer
1274  * @lenp: the size of the user buffer
1275  * @ppos: file position
1276  *
1277  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
1278  * values from/to the user buffer, treated as an ASCII string.
1279  * The values read are assumed to be in seconds, and are converted into
1280  * jiffies.
1281  *
1282  * Returns 0 on success.
1283  */
1284 int proc_dointvec_jiffies(struct ctl_table *table, int write,
1285                           void *buffer, size_t *lenp, loff_t *ppos)
1286 {
1287     return do_proc_dointvec(table,write,buffer,lenp,ppos,
1288                             do_proc_dointvec_jiffies_conv,NULL);
1289 }
1290
1291 int proc_dointvec_ms_jiffies_minmax(struct ctl_table *table, int write,
1292                           void *buffer, size_t *lenp, loff_t *ppos)
1293 {
1294         struct do_proc_dointvec_minmax_conv_param param = {
1295                 .min = (int *) table->extra1,
1296                 .max = (int *) table->extra2,
1297         };
1298         return do_proc_dointvec(table, write, buffer, lenp, ppos,
1299                         do_proc_dointvec_ms_jiffies_minmax_conv, &param);
1300 }
1301
1302 /**
1303  * proc_dointvec_userhz_jiffies - read a vector of integers as 1/USER_HZ seconds
1304  * @table: the sysctl table
1305  * @write: %TRUE if this is a write to the sysctl file
1306  * @buffer: the user buffer
1307  * @lenp: the size of the user buffer
1308  * @ppos: pointer to the file position
1309  *
1310  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
1311  * values from/to the user buffer, treated as an ASCII string.
1312  * The values read are assumed to be in 1/USER_HZ seconds, and
1313  * are converted into jiffies.
1314  *
1315  * Returns 0 on success.
1316  */
1317 int proc_dointvec_userhz_jiffies(struct ctl_table *table, int write,
1318                                  void *buffer, size_t *lenp, loff_t *ppos)
1319 {
1320         return do_proc_dointvec(table, write, buffer, lenp, ppos,
1321                                 do_proc_dointvec_userhz_jiffies_conv, NULL);
1322 }
1323
1324 /**
1325  * proc_dointvec_ms_jiffies - read a vector of integers as 1 milliseconds
1326  * @table: the sysctl table
1327  * @write: %TRUE if this is a write to the sysctl file
1328  * @buffer: the user buffer
1329  * @lenp: the size of the user buffer
1330  * @ppos: file position
1331  * @ppos: the current position in the file
1332  *
1333  * Reads/writes up to table->maxlen/sizeof(unsigned int) integer
1334  * values from/to the user buffer, treated as an ASCII string.
1335  * The values read are assumed to be in 1/1000 seconds, and
1336  * are converted into jiffies.
1337  *
1338  * Returns 0 on success.
1339  */
1340 int proc_dointvec_ms_jiffies(struct ctl_table *table, int write, void *buffer,
1341                 size_t *lenp, loff_t *ppos)
1342 {
1343         return do_proc_dointvec(table, write, buffer, lenp, ppos,
1344                                 do_proc_dointvec_ms_jiffies_conv, NULL);
1345 }
1346
1347 static int proc_do_cad_pid(struct ctl_table *table, int write, void *buffer,
1348                 size_t *lenp, loff_t *ppos)
1349 {
1350         struct pid *new_pid;
1351         pid_t tmp;
1352         int r;
1353
1354         tmp = pid_vnr(cad_pid);
1355
1356         r = __do_proc_dointvec(&tmp, table, write, buffer,
1357                                lenp, ppos, NULL, NULL);
1358         if (r || !write)
1359                 return r;
1360
1361         new_pid = find_get_pid(tmp);
1362         if (!new_pid)
1363                 return -ESRCH;
1364
1365         put_pid(xchg(&cad_pid, new_pid));
1366         return 0;
1367 }
1368
1369 /**
1370  * proc_do_large_bitmap - read/write from/to a large bitmap
1371  * @table: the sysctl table
1372  * @write: %TRUE if this is a write to the sysctl file
1373  * @buffer: the user buffer
1374  * @lenp: the size of the user buffer
1375  * @ppos: file position
1376  *
1377  * The bitmap is stored at table->data and the bitmap length (in bits)
1378  * in table->maxlen.
1379  *
1380  * We use a range comma separated format (e.g. 1,3-4,10-10) so that
1381  * large bitmaps may be represented in a compact manner. Writing into
1382  * the file will clear the bitmap then update it with the given input.
1383  *
1384  * Returns 0 on success.
1385  */
1386 int proc_do_large_bitmap(struct ctl_table *table, int write,
1387                          void *buffer, size_t *lenp, loff_t *ppos)
1388 {
1389         int err = 0;
1390         size_t left = *lenp;
1391         unsigned long bitmap_len = table->maxlen;
1392         unsigned long *bitmap = *(unsigned long **) table->data;
1393         unsigned long *tmp_bitmap = NULL;
1394         char tr_a[] = { '-', ',', '\n' }, tr_b[] = { ',', '\n', 0 }, c;
1395
1396         if (!bitmap || !bitmap_len || !left || (*ppos && !write)) {
1397                 *lenp = 0;
1398                 return 0;
1399         }
1400
1401         if (write) {
1402                 char *p = buffer;
1403                 size_t skipped = 0;
1404
1405                 if (left > PAGE_SIZE - 1) {
1406                         left = PAGE_SIZE - 1;
1407                         /* How much of the buffer we'll skip this pass */
1408                         skipped = *lenp - left;
1409                 }
1410
1411                 tmp_bitmap = bitmap_zalloc(bitmap_len, GFP_KERNEL);
1412                 if (!tmp_bitmap)
1413                         return -ENOMEM;
1414                 proc_skip_char(&p, &left, '\n');
1415                 while (!err && left) {
1416                         unsigned long val_a, val_b;
1417                         bool neg;
1418                         size_t saved_left;
1419
1420                         /* In case we stop parsing mid-number, we can reset */
1421                         saved_left = left;
1422                         err = proc_get_long(&p, &left, &val_a, &neg, tr_a,
1423                                              sizeof(tr_a), &c);
1424                         /*
1425                          * If we consumed the entirety of a truncated buffer or
1426                          * only one char is left (may be a "-"), then stop here,
1427                          * reset, & come back for more.
1428                          */
1429                         if ((left <= 1) && skipped) {
1430                                 left = saved_left;
1431                                 break;
1432                         }
1433
1434                         if (err)
1435                                 break;
1436                         if (val_a >= bitmap_len || neg) {
1437                                 err = -EINVAL;
1438                                 break;
1439                         }
1440
1441                         val_b = val_a;
1442                         if (left) {
1443                                 p++;
1444                                 left--;
1445                         }
1446
1447                         if (c == '-') {
1448                                 err = proc_get_long(&p, &left, &val_b,
1449                                                      &neg, tr_b, sizeof(tr_b),
1450                                                      &c);
1451                                 /*
1452                                  * If we consumed all of a truncated buffer or
1453                                  * then stop here, reset, & come back for more.
1454                                  */
1455                                 if (!left && skipped) {
1456                                         left = saved_left;
1457                                         break;
1458                                 }
1459
1460                                 if (err)
1461                                         break;
1462                                 if (val_b >= bitmap_len || neg ||
1463                                     val_a > val_b) {
1464                                         err = -EINVAL;
1465                                         break;
1466                                 }
1467                                 if (left) {
1468                                         p++;
1469                                         left--;
1470                                 }
1471                         }
1472
1473                         bitmap_set(tmp_bitmap, val_a, val_b - val_a + 1);
1474                         proc_skip_char(&p, &left, '\n');
1475                 }
1476                 left += skipped;
1477         } else {
1478                 unsigned long bit_a, bit_b = 0;
1479                 bool first = 1;
1480
1481                 while (left) {
1482                         bit_a = find_next_bit(bitmap, bitmap_len, bit_b);
1483                         if (bit_a >= bitmap_len)
1484                                 break;
1485                         bit_b = find_next_zero_bit(bitmap, bitmap_len,
1486                                                    bit_a + 1) - 1;
1487
1488                         if (!first)
1489                                 proc_put_char(&buffer, &left, ',');
1490                         proc_put_long(&buffer, &left, bit_a, false);
1491                         if (bit_a != bit_b) {
1492                                 proc_put_char(&buffer, &left, '-');
1493                                 proc_put_long(&buffer, &left, bit_b, false);
1494                         }
1495
1496                         first = 0; bit_b++;
1497                 }
1498                 proc_put_char(&buffer, &left, '\n');
1499         }
1500
1501         if (!err) {
1502                 if (write) {
1503                         if (*ppos)
1504                                 bitmap_or(bitmap, bitmap, tmp_bitmap, bitmap_len);
1505                         else
1506                                 bitmap_copy(bitmap, tmp_bitmap, bitmap_len);
1507                 }
1508                 *lenp -= left;
1509                 *ppos += *lenp;
1510         }
1511
1512         bitmap_free(tmp_bitmap);
1513         return err;
1514 }
1515
1516 #else /* CONFIG_PROC_SYSCTL */
1517
1518 int proc_dostring(struct ctl_table *table, int write,
1519                   void *buffer, size_t *lenp, loff_t *ppos)
1520 {
1521         return -ENOSYS;
1522 }
1523
1524 int proc_dobool(struct ctl_table *table, int write,
1525                 void *buffer, size_t *lenp, loff_t *ppos)
1526 {
1527         return -ENOSYS;
1528 }
1529
1530 int proc_dointvec(struct ctl_table *table, int write,
1531                   void *buffer, size_t *lenp, loff_t *ppos)
1532 {
1533         return -ENOSYS;
1534 }
1535
1536 int proc_douintvec(struct ctl_table *table, int write,
1537                   void *buffer, size_t *lenp, loff_t *ppos)
1538 {
1539         return -ENOSYS;
1540 }
1541
1542 int proc_dointvec_minmax(struct ctl_table *table, int write,
1543                     void *buffer, size_t *lenp, loff_t *ppos)
1544 {
1545         return -ENOSYS;
1546 }
1547
1548 int proc_douintvec_minmax(struct ctl_table *table, int write,
1549                           void *buffer, size_t *lenp, loff_t *ppos)
1550 {
1551         return -ENOSYS;
1552 }
1553
1554 int proc_dou8vec_minmax(struct ctl_table *table, int write,
1555                         void *buffer, size_t *lenp, loff_t *ppos)
1556 {
1557         return -ENOSYS;
1558 }
1559
1560 int proc_dointvec_jiffies(struct ctl_table *table, int write,
1561                     void *buffer, size_t *lenp, loff_t *ppos)
1562 {
1563         return -ENOSYS;
1564 }
1565
1566 int proc_dointvec_ms_jiffies_minmax(struct ctl_table *table, int write,
1567                                     void *buffer, size_t *lenp, loff_t *ppos)
1568 {
1569         return -ENOSYS;
1570 }
1571
1572 int proc_dointvec_userhz_jiffies(struct ctl_table *table, int write,
1573                     void *buffer, size_t *lenp, loff_t *ppos)
1574 {
1575         return -ENOSYS;
1576 }
1577
1578 int proc_dointvec_ms_jiffies(struct ctl_table *table, int write,
1579                              void *buffer, size_t *lenp, loff_t *ppos)
1580 {
1581         return -ENOSYS;
1582 }
1583
1584 int proc_doulongvec_minmax(struct ctl_table *table, int write,
1585                     void *buffer, size_t *lenp, loff_t *ppos)
1586 {
1587         return -ENOSYS;
1588 }
1589
1590 int proc_doulongvec_ms_jiffies_minmax(struct ctl_table *table, int write,
1591                                       void *buffer, size_t *lenp, loff_t *ppos)
1592 {
1593         return -ENOSYS;
1594 }
1595
1596 int proc_do_large_bitmap(struct ctl_table *table, int write,
1597                          void *buffer, size_t *lenp, loff_t *ppos)
1598 {
1599         return -ENOSYS;
1600 }
1601
1602 #endif /* CONFIG_PROC_SYSCTL */
1603
1604 #if defined(CONFIG_SYSCTL)
1605 int proc_do_static_key(struct ctl_table *table, int write,
1606                        void *buffer, size_t *lenp, loff_t *ppos)
1607 {
1608         struct static_key *key = (struct static_key *)table->data;
1609         static DEFINE_MUTEX(static_key_mutex);
1610         int val, ret;
1611         struct ctl_table tmp = {
1612                 .data   = &val,
1613                 .maxlen = sizeof(val),
1614                 .mode   = table->mode,
1615                 .extra1 = SYSCTL_ZERO,
1616                 .extra2 = SYSCTL_ONE,
1617         };
1618
1619         if (write && !capable(CAP_SYS_ADMIN))
1620                 return -EPERM;
1621
1622         mutex_lock(&static_key_mutex);
1623         val = static_key_enabled(key);
1624         ret = proc_dointvec_minmax(&tmp, write, buffer, lenp, ppos);
1625         if (write && !ret) {
1626                 if (val)
1627                         static_key_enable(key);
1628                 else
1629                         static_key_disable(key);
1630         }
1631         mutex_unlock(&static_key_mutex);
1632         return ret;
1633 }
1634
1635 static struct ctl_table kern_table[] = {
1636 #ifdef CONFIG_NUMA_BALANCING
1637         {
1638                 .procname       = "numa_balancing",
1639                 .data           = NULL, /* filled in by handler */
1640                 .maxlen         = sizeof(unsigned int),
1641                 .mode           = 0644,
1642                 .proc_handler   = sysctl_numa_balancing,
1643                 .extra1         = SYSCTL_ZERO,
1644                 .extra2         = SYSCTL_FOUR,
1645         },
1646         {
1647                 .procname       = "numa_balancing_promote_rate_limit_MBps",
1648                 .data           = &sysctl_numa_balancing_promote_rate_limit,
1649                 .maxlen         = sizeof(unsigned int),
1650                 .mode           = 0644,
1651                 .proc_handler   = proc_dointvec_minmax,
1652                 .extra1         = SYSCTL_ZERO,
1653         },
1654 #endif /* CONFIG_NUMA_BALANCING */
1655         {
1656                 .procname       = "panic",
1657                 .data           = &panic_timeout,
1658                 .maxlen         = sizeof(int),
1659                 .mode           = 0644,
1660                 .proc_handler   = proc_dointvec,
1661         },
1662 #ifdef CONFIG_PROC_SYSCTL
1663         {
1664                 .procname       = "tainted",
1665                 .maxlen         = sizeof(long),
1666                 .mode           = 0644,
1667                 .proc_handler   = proc_taint,
1668         },
1669         {
1670                 .procname       = "sysctl_writes_strict",
1671                 .data           = &sysctl_writes_strict,
1672                 .maxlen         = sizeof(int),
1673                 .mode           = 0644,
1674                 .proc_handler   = proc_dointvec_minmax,
1675                 .extra1         = SYSCTL_NEG_ONE,
1676                 .extra2         = SYSCTL_ONE,
1677         },
1678 #endif
1679         {
1680                 .procname       = "print-fatal-signals",
1681                 .data           = &print_fatal_signals,
1682                 .maxlen         = sizeof(int),
1683                 .mode           = 0644,
1684                 .proc_handler   = proc_dointvec,
1685         },
1686 #ifdef CONFIG_SPARC
1687         {
1688                 .procname       = "reboot-cmd",
1689                 .data           = reboot_command,
1690                 .maxlen         = 256,
1691                 .mode           = 0644,
1692                 .proc_handler   = proc_dostring,
1693         },
1694         {
1695                 .procname       = "stop-a",
1696                 .data           = &stop_a_enabled,
1697                 .maxlen         = sizeof (int),
1698                 .mode           = 0644,
1699                 .proc_handler   = proc_dointvec,
1700         },
1701         {
1702                 .procname       = "scons-poweroff",
1703                 .data           = &scons_pwroff,
1704                 .maxlen         = sizeof (int),
1705                 .mode           = 0644,
1706                 .proc_handler   = proc_dointvec,
1707         },
1708 #endif
1709 #ifdef CONFIG_SPARC64
1710         {
1711                 .procname       = "tsb-ratio",
1712                 .data           = &sysctl_tsb_ratio,
1713                 .maxlen         = sizeof (int),
1714                 .mode           = 0644,
1715                 .proc_handler   = proc_dointvec,
1716         },
1717 #endif
1718 #ifdef CONFIG_PARISC
1719         {
1720                 .procname       = "soft-power",
1721                 .data           = &pwrsw_enabled,
1722                 .maxlen         = sizeof (int),
1723                 .mode           = 0644,
1724                 .proc_handler   = proc_dointvec,
1725         },
1726 #endif
1727 #ifdef CONFIG_SYSCTL_ARCH_UNALIGN_ALLOW
1728         {
1729                 .procname       = "unaligned-trap",
1730                 .data           = &unaligned_enabled,
1731                 .maxlen         = sizeof (int),
1732                 .mode           = 0644,
1733                 .proc_handler   = proc_dointvec,
1734         },
1735 #endif
1736 #ifdef CONFIG_STACK_TRACER
1737         {
1738                 .procname       = "stack_tracer_enabled",
1739                 .data           = &stack_tracer_enabled,
1740                 .maxlen         = sizeof(int),
1741                 .mode           = 0644,
1742                 .proc_handler   = stack_trace_sysctl,
1743         },
1744 #endif
1745 #ifdef CONFIG_TRACING
1746         {
1747                 .procname       = "ftrace_dump_on_oops",
1748                 .data           = &ftrace_dump_on_oops,
1749                 .maxlen         = sizeof(int),
1750                 .mode           = 0644,
1751                 .proc_handler   = proc_dointvec,
1752         },
1753         {
1754                 .procname       = "traceoff_on_warning",
1755                 .data           = &__disable_trace_on_warning,
1756                 .maxlen         = sizeof(__disable_trace_on_warning),
1757                 .mode           = 0644,
1758                 .proc_handler   = proc_dointvec,
1759         },
1760         {
1761                 .procname       = "tracepoint_printk",
1762                 .data           = &tracepoint_printk,
1763                 .maxlen         = sizeof(tracepoint_printk),
1764                 .mode           = 0644,
1765                 .proc_handler   = tracepoint_printk_sysctl,
1766         },
1767 #endif
1768 #ifdef CONFIG_MODULES
1769         {
1770                 .procname       = "modprobe",
1771                 .data           = &modprobe_path,
1772                 .maxlen         = KMOD_PATH_LEN,
1773                 .mode           = 0644,
1774                 .proc_handler   = proc_dostring,
1775         },
1776         {
1777                 .procname       = "modules_disabled",
1778                 .data           = &modules_disabled,
1779                 .maxlen         = sizeof(int),
1780                 .mode           = 0644,
1781                 /* only handle a transition from default "0" to "1" */
1782                 .proc_handler   = proc_dointvec_minmax,
1783                 .extra1         = SYSCTL_ONE,
1784                 .extra2         = SYSCTL_ONE,
1785         },
1786 #endif
1787 #ifdef CONFIG_UEVENT_HELPER
1788         {
1789                 .procname       = "hotplug",
1790                 .data           = &uevent_helper,
1791                 .maxlen         = UEVENT_HELPER_PATH_LEN,
1792                 .mode           = 0644,
1793                 .proc_handler   = proc_dostring,
1794         },
1795 #endif
1796 #ifdef CONFIG_MAGIC_SYSRQ
1797         {
1798                 .procname       = "sysrq",
1799                 .data           = NULL,
1800                 .maxlen         = sizeof (int),
1801                 .mode           = 0644,
1802                 .proc_handler   = sysrq_sysctl_handler,
1803         },
1804 #endif
1805 #ifdef CONFIG_PROC_SYSCTL
1806         {
1807                 .procname       = "cad_pid",
1808                 .data           = NULL,
1809                 .maxlen         = sizeof (int),
1810                 .mode           = 0600,
1811                 .proc_handler   = proc_do_cad_pid,
1812         },
1813 #endif
1814         {
1815                 .procname       = "threads-max",
1816                 .data           = NULL,
1817                 .maxlen         = sizeof(int),
1818                 .mode           = 0644,
1819                 .proc_handler   = sysctl_max_threads,
1820         },
1821         {
1822                 .procname       = "usermodehelper",
1823                 .mode           = 0555,
1824                 .child          = usermodehelper_table,
1825         },
1826         {
1827                 .procname       = "overflowuid",
1828                 .data           = &overflowuid,
1829                 .maxlen         = sizeof(int),
1830                 .mode           = 0644,
1831                 .proc_handler   = proc_dointvec_minmax,
1832                 .extra1         = SYSCTL_ZERO,
1833                 .extra2         = SYSCTL_MAXOLDUID,
1834         },
1835         {
1836                 .procname       = "overflowgid",
1837                 .data           = &overflowgid,
1838                 .maxlen         = sizeof(int),
1839                 .mode           = 0644,
1840                 .proc_handler   = proc_dointvec_minmax,
1841                 .extra1         = SYSCTL_ZERO,
1842                 .extra2         = SYSCTL_MAXOLDUID,
1843         },
1844 #ifdef CONFIG_S390
1845         {
1846                 .procname       = "userprocess_debug",
1847                 .data           = &show_unhandled_signals,
1848                 .maxlen         = sizeof(int),
1849                 .mode           = 0644,
1850                 .proc_handler   = proc_dointvec,
1851         },
1852 #endif
1853         {
1854                 .procname       = "pid_max",
1855                 .data           = &pid_max,
1856                 .maxlen         = sizeof (int),
1857                 .mode           = 0644,
1858                 .proc_handler   = proc_dointvec_minmax,
1859                 .extra1         = &pid_max_min,
1860                 .extra2         = &pid_max_max,
1861         },
1862         {
1863                 .procname       = "panic_on_oops",
1864                 .data           = &panic_on_oops,
1865                 .maxlen         = sizeof(int),
1866                 .mode           = 0644,
1867                 .proc_handler   = proc_dointvec,
1868         },
1869         {
1870                 .procname       = "panic_print",
1871                 .data           = &panic_print,
1872                 .maxlen         = sizeof(unsigned long),
1873                 .mode           = 0644,
1874                 .proc_handler   = proc_doulongvec_minmax,
1875         },
1876         {
1877                 .procname       = "ngroups_max",
1878                 .data           = (void *)&ngroups_max,
1879                 .maxlen         = sizeof (int),
1880                 .mode           = 0444,
1881                 .proc_handler   = proc_dointvec,
1882         },
1883         {
1884                 .procname       = "cap_last_cap",
1885                 .data           = (void *)&cap_last_cap,
1886                 .maxlen         = sizeof(int),
1887                 .mode           = 0444,
1888                 .proc_handler   = proc_dointvec,
1889         },
1890 #if defined(CONFIG_X86_LOCAL_APIC) && defined(CONFIG_X86)
1891         {
1892                 .procname       = "unknown_nmi_panic",
1893                 .data           = &unknown_nmi_panic,
1894                 .maxlen         = sizeof (int),
1895                 .mode           = 0644,
1896                 .proc_handler   = proc_dointvec,
1897         },
1898 #endif
1899
1900 #if (defined(CONFIG_X86_32) || defined(CONFIG_PARISC)) && \
1901         defined(CONFIG_DEBUG_STACKOVERFLOW)
1902         {
1903                 .procname       = "panic_on_stackoverflow",
1904                 .data           = &sysctl_panic_on_stackoverflow,
1905                 .maxlen         = sizeof(int),
1906                 .mode           = 0644,
1907                 .proc_handler   = proc_dointvec,
1908         },
1909 #endif
1910 #if defined(CONFIG_X86)
1911         {
1912                 .procname       = "panic_on_unrecovered_nmi",
1913                 .data           = &panic_on_unrecovered_nmi,
1914                 .maxlen         = sizeof(int),
1915                 .mode           = 0644,
1916                 .proc_handler   = proc_dointvec,
1917         },
1918         {
1919                 .procname       = "panic_on_io_nmi",
1920                 .data           = &panic_on_io_nmi,
1921                 .maxlen         = sizeof(int),
1922                 .mode           = 0644,
1923                 .proc_handler   = proc_dointvec,
1924         },
1925         {
1926                 .procname       = "bootloader_type",
1927                 .data           = &bootloader_type,
1928                 .maxlen         = sizeof (int),
1929                 .mode           = 0444,
1930                 .proc_handler   = proc_dointvec,
1931         },
1932         {
1933                 .procname       = "bootloader_version",
1934                 .data           = &bootloader_version,
1935                 .maxlen         = sizeof (int),
1936                 .mode           = 0444,
1937                 .proc_handler   = proc_dointvec,
1938         },
1939         {
1940                 .procname       = "io_delay_type",
1941                 .data           = &io_delay_type,
1942                 .maxlen         = sizeof(int),
1943                 .mode           = 0644,
1944                 .proc_handler   = proc_dointvec,
1945         },
1946 #endif
1947 #if defined(CONFIG_MMU)
1948         {
1949                 .procname       = "randomize_va_space",
1950                 .data           = &randomize_va_space,
1951                 .maxlen         = sizeof(int),
1952                 .mode           = 0644,
1953                 .proc_handler   = proc_dointvec,
1954         },
1955 #endif
1956 #if defined(CONFIG_S390) && defined(CONFIG_SMP)
1957         {
1958                 .procname       = "spin_retry",
1959                 .data           = &spin_retry,
1960                 .maxlen         = sizeof (int),
1961                 .mode           = 0644,
1962                 .proc_handler   = proc_dointvec,
1963         },
1964 #endif
1965 #if     defined(CONFIG_ACPI_SLEEP) && defined(CONFIG_X86)
1966         {
1967                 .procname       = "acpi_video_flags",
1968                 .data           = &acpi_realmode_flags,
1969                 .maxlen         = sizeof (unsigned long),
1970                 .mode           = 0644,
1971                 .proc_handler   = proc_doulongvec_minmax,
1972         },
1973 #endif
1974 #ifdef CONFIG_SYSCTL_ARCH_UNALIGN_NO_WARN
1975         {
1976                 .procname       = "ignore-unaligned-usertrap",
1977                 .data           = &no_unaligned_warning,
1978                 .maxlen         = sizeof (int),
1979                 .mode           = 0644,
1980                 .proc_handler   = proc_dointvec,
1981         },
1982 #endif
1983 #ifdef CONFIG_IA64
1984         {
1985                 .procname       = "unaligned-dump-stack",
1986                 .data           = &unaligned_dump_stack,
1987                 .maxlen         = sizeof (int),
1988                 .mode           = 0644,
1989                 .proc_handler   = proc_dointvec,
1990         },
1991 #endif
1992 #ifdef CONFIG_RT_MUTEXES
1993         {
1994                 .procname       = "max_lock_depth",
1995                 .data           = &max_lock_depth,
1996                 .maxlen         = sizeof(int),
1997                 .mode           = 0644,
1998                 .proc_handler   = proc_dointvec,
1999         },
2000 #endif
2001 #ifdef CONFIG_KEYS
2002         {
2003                 .procname       = "keys",
2004                 .mode           = 0555,
2005                 .child          = key_sysctls,
2006         },
2007 #endif
2008 #ifdef CONFIG_PERF_EVENTS
2009         /*
2010          * User-space scripts rely on the existence of this file
2011          * as a feature check for perf_events being enabled.
2012          *
2013          * So it's an ABI, do not remove!
2014          */
2015         {
2016                 .procname       = "perf_event_paranoid",
2017                 .data           = &sysctl_perf_event_paranoid,
2018                 .maxlen         = sizeof(sysctl_perf_event_paranoid),
2019                 .mode           = 0644,
2020                 .proc_handler   = proc_dointvec,
2021         },
2022         {
2023                 .procname       = "perf_event_mlock_kb",
2024                 .data           = &sysctl_perf_event_mlock,
2025                 .maxlen         = sizeof(sysctl_perf_event_mlock),
2026                 .mode           = 0644,
2027                 .proc_handler   = proc_dointvec,
2028         },
2029         {
2030                 .procname       = "perf_event_max_sample_rate",
2031                 .data           = &sysctl_perf_event_sample_rate,
2032                 .maxlen         = sizeof(sysctl_perf_event_sample_rate),
2033                 .mode           = 0644,
2034                 .proc_handler   = perf_proc_update_handler,
2035                 .extra1         = SYSCTL_ONE,
2036         },
2037         {
2038                 .procname       = "perf_cpu_time_max_percent",
2039                 .data           = &sysctl_perf_cpu_time_max_percent,
2040                 .maxlen         = sizeof(sysctl_perf_cpu_time_max_percent),
2041                 .mode           = 0644,
2042                 .proc_handler   = perf_cpu_time_max_percent_handler,
2043                 .extra1         = SYSCTL_ZERO,
2044                 .extra2         = SYSCTL_ONE_HUNDRED,
2045         },
2046         {
2047                 .procname       = "perf_event_max_stack",
2048                 .data           = &sysctl_perf_event_max_stack,
2049                 .maxlen         = sizeof(sysctl_perf_event_max_stack),
2050                 .mode           = 0644,
2051                 .proc_handler   = perf_event_max_stack_handler,
2052                 .extra1         = SYSCTL_ZERO,
2053                 .extra2         = (void *)&six_hundred_forty_kb,
2054         },
2055         {
2056                 .procname       = "perf_event_max_contexts_per_stack",
2057                 .data           = &sysctl_perf_event_max_contexts_per_stack,
2058                 .maxlen         = sizeof(sysctl_perf_event_max_contexts_per_stack),
2059                 .mode           = 0644,
2060                 .proc_handler   = perf_event_max_stack_handler,
2061                 .extra1         = SYSCTL_ZERO,
2062                 .extra2         = SYSCTL_ONE_THOUSAND,
2063         },
2064 #endif
2065         {
2066                 .procname       = "panic_on_warn",
2067                 .data           = &panic_on_warn,
2068                 .maxlen         = sizeof(int),
2069                 .mode           = 0644,
2070                 .proc_handler   = proc_dointvec_minmax,
2071                 .extra1         = SYSCTL_ZERO,
2072                 .extra2         = SYSCTL_ONE,
2073         },
2074 #ifdef CONFIG_TREE_RCU
2075         {
2076                 .procname       = "panic_on_rcu_stall",
2077                 .data           = &sysctl_panic_on_rcu_stall,
2078                 .maxlen         = sizeof(sysctl_panic_on_rcu_stall),
2079                 .mode           = 0644,
2080                 .proc_handler   = proc_dointvec_minmax,
2081                 .extra1         = SYSCTL_ZERO,
2082                 .extra2         = SYSCTL_ONE,
2083         },
2084         {
2085                 .procname       = "max_rcu_stall_to_panic",
2086                 .data           = &sysctl_max_rcu_stall_to_panic,
2087                 .maxlen         = sizeof(sysctl_max_rcu_stall_to_panic),
2088                 .mode           = 0644,
2089                 .proc_handler   = proc_dointvec_minmax,
2090                 .extra1         = SYSCTL_ONE,
2091                 .extra2         = SYSCTL_INT_MAX,
2092         },
2093 #endif
2094         { }
2095 };
2096
2097 static struct ctl_table vm_table[] = {
2098         {
2099                 .procname       = "overcommit_memory",
2100                 .data           = &sysctl_overcommit_memory,
2101                 .maxlen         = sizeof(sysctl_overcommit_memory),
2102                 .mode           = 0644,
2103                 .proc_handler   = overcommit_policy_handler,
2104                 .extra1         = SYSCTL_ZERO,
2105                 .extra2         = SYSCTL_TWO,
2106         },
2107         {
2108                 .procname       = "overcommit_ratio",
2109                 .data           = &sysctl_overcommit_ratio,
2110                 .maxlen         = sizeof(sysctl_overcommit_ratio),
2111                 .mode           = 0644,
2112                 .proc_handler   = overcommit_ratio_handler,
2113         },
2114         {
2115                 .procname       = "overcommit_kbytes",
2116                 .data           = &sysctl_overcommit_kbytes,
2117                 .maxlen         = sizeof(sysctl_overcommit_kbytes),
2118                 .mode           = 0644,
2119                 .proc_handler   = overcommit_kbytes_handler,
2120         },
2121         {
2122                 .procname       = "page-cluster",
2123                 .data           = &page_cluster,
2124                 .maxlen         = sizeof(int),
2125                 .mode           = 0644,
2126                 .proc_handler   = proc_dointvec_minmax,
2127                 .extra1         = SYSCTL_ZERO,
2128         },
2129         {
2130                 .procname       = "dirtytime_expire_seconds",
2131                 .data           = &dirtytime_expire_interval,
2132                 .maxlen         = sizeof(dirtytime_expire_interval),
2133                 .mode           = 0644,
2134                 .proc_handler   = dirtytime_interval_handler,
2135                 .extra1         = SYSCTL_ZERO,
2136         },
2137         {
2138                 .procname       = "swappiness",
2139                 .data           = &vm_swappiness,
2140                 .maxlen         = sizeof(vm_swappiness),
2141                 .mode           = 0644,
2142                 .proc_handler   = proc_dointvec_minmax,
2143                 .extra1         = SYSCTL_ZERO,
2144                 .extra2         = SYSCTL_TWO_HUNDRED,
2145         },
2146 #ifdef CONFIG_NUMA
2147         {
2148                 .procname       = "numa_stat",
2149                 .data           = &sysctl_vm_numa_stat,
2150                 .maxlen         = sizeof(int),
2151                 .mode           = 0644,
2152                 .proc_handler   = sysctl_vm_numa_stat_handler,
2153                 .extra1         = SYSCTL_ZERO,
2154                 .extra2         = SYSCTL_ONE,
2155         },
2156 #endif
2157 #ifdef CONFIG_HUGETLB_PAGE
2158         {
2159                 .procname       = "nr_hugepages",
2160                 .data           = NULL,
2161                 .maxlen         = sizeof(unsigned long),
2162                 .mode           = 0644,
2163                 .proc_handler   = hugetlb_sysctl_handler,
2164         },
2165 #ifdef CONFIG_NUMA
2166         {
2167                 .procname       = "nr_hugepages_mempolicy",
2168                 .data           = NULL,
2169                 .maxlen         = sizeof(unsigned long),
2170                 .mode           = 0644,
2171                 .proc_handler   = &hugetlb_mempolicy_sysctl_handler,
2172         },
2173 #endif
2174          {
2175                 .procname       = "hugetlb_shm_group",
2176                 .data           = &sysctl_hugetlb_shm_group,
2177                 .maxlen         = sizeof(gid_t),
2178                 .mode           = 0644,
2179                 .proc_handler   = proc_dointvec,
2180          },
2181         {
2182                 .procname       = "nr_overcommit_hugepages",
2183                 .data           = NULL,
2184                 .maxlen         = sizeof(unsigned long),
2185                 .mode           = 0644,
2186                 .proc_handler   = hugetlb_overcommit_handler,
2187         },
2188 #endif
2189         {
2190                 .procname       = "lowmem_reserve_ratio",
2191                 .data           = &sysctl_lowmem_reserve_ratio,
2192                 .maxlen         = sizeof(sysctl_lowmem_reserve_ratio),
2193                 .mode           = 0644,
2194                 .proc_handler   = lowmem_reserve_ratio_sysctl_handler,
2195         },
2196         {
2197                 .procname       = "drop_caches",
2198                 .data           = &sysctl_drop_caches,
2199                 .maxlen         = sizeof(int),
2200                 .mode           = 0200,
2201                 .proc_handler   = drop_caches_sysctl_handler,
2202                 .extra1         = SYSCTL_ONE,
2203                 .extra2         = SYSCTL_FOUR,
2204         },
2205 #ifdef CONFIG_COMPACTION
2206         {
2207                 .procname       = "compact_memory",
2208                 .data           = NULL,
2209                 .maxlen         = sizeof(int),
2210                 .mode           = 0200,
2211                 .proc_handler   = sysctl_compaction_handler,
2212         },
2213         {
2214                 .procname       = "compaction_proactiveness",
2215                 .data           = &sysctl_compaction_proactiveness,
2216                 .maxlen         = sizeof(sysctl_compaction_proactiveness),
2217                 .mode           = 0644,
2218                 .proc_handler   = compaction_proactiveness_sysctl_handler,
2219                 .extra1         = SYSCTL_ZERO,
2220                 .extra2         = SYSCTL_ONE_HUNDRED,
2221         },
2222         {
2223                 .procname       = "extfrag_threshold",
2224                 .data           = &sysctl_extfrag_threshold,
2225                 .maxlen         = sizeof(int),
2226                 .mode           = 0644,
2227                 .proc_handler   = proc_dointvec_minmax,
2228                 .extra1         = SYSCTL_ZERO,
2229                 .extra2         = SYSCTL_ONE_THOUSAND,
2230         },
2231         {
2232                 .procname       = "compact_unevictable_allowed",
2233                 .data           = &sysctl_compact_unevictable_allowed,
2234                 .maxlen         = sizeof(int),
2235                 .mode           = 0644,
2236                 .proc_handler   = proc_dointvec_minmax_warn_RT_change,
2237                 .extra1         = SYSCTL_ZERO,
2238                 .extra2         = SYSCTL_ONE,
2239         },
2240
2241 #endif /* CONFIG_COMPACTION */
2242         {
2243                 .procname       = "min_free_kbytes",
2244                 .data           = &min_free_kbytes,
2245                 .maxlen         = sizeof(min_free_kbytes),
2246                 .mode           = 0644,
2247                 .proc_handler   = min_free_kbytes_sysctl_handler,
2248                 .extra1         = SYSCTL_ZERO,
2249         },
2250         {
2251                 .procname       = "watermark_boost_factor",
2252                 .data           = &watermark_boost_factor,
2253                 .maxlen         = sizeof(watermark_boost_factor),
2254                 .mode           = 0644,
2255                 .proc_handler   = proc_dointvec_minmax,
2256                 .extra1         = SYSCTL_ZERO,
2257         },
2258         {
2259                 .procname       = "watermark_scale_factor",
2260                 .data           = &watermark_scale_factor,
2261                 .maxlen         = sizeof(watermark_scale_factor),
2262                 .mode           = 0644,
2263                 .proc_handler   = watermark_scale_factor_sysctl_handler,
2264                 .extra1         = SYSCTL_ONE,
2265                 .extra2         = SYSCTL_THREE_THOUSAND,
2266         },
2267         {
2268                 .procname       = "percpu_pagelist_high_fraction",
2269                 .data           = &percpu_pagelist_high_fraction,
2270                 .maxlen         = sizeof(percpu_pagelist_high_fraction),
2271                 .mode           = 0644,
2272                 .proc_handler   = percpu_pagelist_high_fraction_sysctl_handler,
2273                 .extra1         = SYSCTL_ZERO,
2274         },
2275         {
2276                 .procname       = "page_lock_unfairness",
2277                 .data           = &sysctl_page_lock_unfairness,
2278                 .maxlen         = sizeof(sysctl_page_lock_unfairness),
2279                 .mode           = 0644,
2280                 .proc_handler   = proc_dointvec_minmax,
2281                 .extra1         = SYSCTL_ZERO,
2282         },
2283 #ifdef CONFIG_MMU
2284         {
2285                 .procname       = "max_map_count",
2286                 .data           = &sysctl_max_map_count,
2287                 .maxlen         = sizeof(sysctl_max_map_count),
2288                 .mode           = 0644,
2289                 .proc_handler   = proc_dointvec_minmax,
2290                 .extra1         = SYSCTL_ZERO,
2291         },
2292 #else
2293         {
2294                 .procname       = "nr_trim_pages",
2295                 .data           = &sysctl_nr_trim_pages,
2296                 .maxlen         = sizeof(sysctl_nr_trim_pages),
2297                 .mode           = 0644,
2298                 .proc_handler   = proc_dointvec_minmax,
2299                 .extra1         = SYSCTL_ZERO,
2300         },
2301 #endif
2302         {
2303                 .procname       = "vfs_cache_pressure",
2304                 .data           = &sysctl_vfs_cache_pressure,
2305                 .maxlen         = sizeof(sysctl_vfs_cache_pressure),
2306                 .mode           = 0644,
2307                 .proc_handler   = proc_dointvec_minmax,
2308                 .extra1         = SYSCTL_ZERO,
2309         },
2310 #if defined(HAVE_ARCH_PICK_MMAP_LAYOUT) || \
2311     defined(CONFIG_ARCH_WANT_DEFAULT_TOPDOWN_MMAP_LAYOUT)
2312         {
2313                 .procname       = "legacy_va_layout",
2314                 .data           = &sysctl_legacy_va_layout,
2315                 .maxlen         = sizeof(sysctl_legacy_va_layout),
2316                 .mode           = 0644,
2317                 .proc_handler   = proc_dointvec_minmax,
2318                 .extra1         = SYSCTL_ZERO,
2319         },
2320 #endif
2321 #ifdef CONFIG_NUMA
2322         {
2323                 .procname       = "zone_reclaim_mode",
2324                 .data           = &node_reclaim_mode,
2325                 .maxlen         = sizeof(node_reclaim_mode),
2326                 .mode           = 0644,
2327                 .proc_handler   = proc_dointvec_minmax,
2328                 .extra1         = SYSCTL_ZERO,
2329         },
2330         {
2331                 .procname       = "min_unmapped_ratio",
2332                 .data           = &sysctl_min_unmapped_ratio,
2333                 .maxlen         = sizeof(sysctl_min_unmapped_ratio),
2334                 .mode           = 0644,
2335                 .proc_handler   = sysctl_min_unmapped_ratio_sysctl_handler,
2336                 .extra1         = SYSCTL_ZERO,
2337                 .extra2         = SYSCTL_ONE_HUNDRED,
2338         },
2339         {
2340                 .procname       = "min_slab_ratio",
2341                 .data           = &sysctl_min_slab_ratio,
2342                 .maxlen         = sizeof(sysctl_min_slab_ratio),
2343                 .mode           = 0644,
2344                 .proc_handler   = sysctl_min_slab_ratio_sysctl_handler,
2345                 .extra1         = SYSCTL_ZERO,
2346                 .extra2         = SYSCTL_ONE_HUNDRED,
2347         },
2348 #endif
2349 #ifdef CONFIG_SMP
2350         {
2351                 .procname       = "stat_interval",
2352                 .data           = &sysctl_stat_interval,
2353                 .maxlen         = sizeof(sysctl_stat_interval),
2354                 .mode           = 0644,
2355                 .proc_handler   = proc_dointvec_jiffies,
2356         },
2357         {
2358                 .procname       = "stat_refresh",
2359                 .data           = NULL,
2360                 .maxlen         = 0,
2361                 .mode           = 0600,
2362                 .proc_handler   = vmstat_refresh,
2363         },
2364 #endif
2365 #ifdef CONFIG_MMU
2366         {
2367                 .procname       = "mmap_min_addr",
2368                 .data           = &dac_mmap_min_addr,
2369                 .maxlen         = sizeof(unsigned long),
2370                 .mode           = 0644,
2371                 .proc_handler   = mmap_min_addr_handler,
2372         },
2373 #endif
2374 #ifdef CONFIG_NUMA
2375         {
2376                 .procname       = "numa_zonelist_order",
2377                 .data           = &numa_zonelist_order,
2378                 .maxlen         = NUMA_ZONELIST_ORDER_LEN,
2379                 .mode           = 0644,
2380                 .proc_handler   = numa_zonelist_order_handler,
2381         },
2382 #endif
2383 #if (defined(CONFIG_X86_32) && !defined(CONFIG_UML))|| \
2384    (defined(CONFIG_SUPERH) && defined(CONFIG_VSYSCALL))
2385         {
2386                 .procname       = "vdso_enabled",
2387 #ifdef CONFIG_X86_32
2388                 .data           = &vdso32_enabled,
2389                 .maxlen         = sizeof(vdso32_enabled),
2390 #else
2391                 .data           = &vdso_enabled,
2392                 .maxlen         = sizeof(vdso_enabled),
2393 #endif
2394                 .mode           = 0644,
2395                 .proc_handler   = proc_dointvec,
2396                 .extra1         = SYSCTL_ZERO,
2397         },
2398 #endif
2399 #ifdef CONFIG_MEMORY_FAILURE
2400         {
2401                 .procname       = "memory_failure_early_kill",
2402                 .data           = &sysctl_memory_failure_early_kill,
2403                 .maxlen         = sizeof(sysctl_memory_failure_early_kill),
2404                 .mode           = 0644,
2405                 .proc_handler   = proc_dointvec_minmax,
2406                 .extra1         = SYSCTL_ZERO,
2407                 .extra2         = SYSCTL_ONE,
2408         },
2409         {
2410                 .procname       = "memory_failure_recovery",
2411                 .data           = &sysctl_memory_failure_recovery,
2412                 .maxlen         = sizeof(sysctl_memory_failure_recovery),
2413                 .mode           = 0644,
2414                 .proc_handler   = proc_dointvec_minmax,
2415                 .extra1         = SYSCTL_ZERO,
2416                 .extra2         = SYSCTL_ONE,
2417         },
2418 #endif
2419         {
2420                 .procname       = "user_reserve_kbytes",
2421                 .data           = &sysctl_user_reserve_kbytes,
2422                 .maxlen         = sizeof(sysctl_user_reserve_kbytes),
2423                 .mode           = 0644,
2424                 .proc_handler   = proc_doulongvec_minmax,
2425         },
2426         {
2427                 .procname       = "admin_reserve_kbytes",
2428                 .data           = &sysctl_admin_reserve_kbytes,
2429                 .maxlen         = sizeof(sysctl_admin_reserve_kbytes),
2430                 .mode           = 0644,
2431                 .proc_handler   = proc_doulongvec_minmax,
2432         },
2433 #ifdef CONFIG_HAVE_ARCH_MMAP_RND_BITS
2434         {
2435                 .procname       = "mmap_rnd_bits",
2436                 .data           = &mmap_rnd_bits,
2437                 .maxlen         = sizeof(mmap_rnd_bits),
2438                 .mode           = 0600,
2439                 .proc_handler   = proc_dointvec_minmax,
2440                 .extra1         = (void *)&mmap_rnd_bits_min,
2441                 .extra2         = (void *)&mmap_rnd_bits_max,
2442         },
2443 #endif
2444 #ifdef CONFIG_HAVE_ARCH_MMAP_RND_COMPAT_BITS
2445         {
2446                 .procname       = "mmap_rnd_compat_bits",
2447                 .data           = &mmap_rnd_compat_bits,
2448                 .maxlen         = sizeof(mmap_rnd_compat_bits),
2449                 .mode           = 0600,
2450                 .proc_handler   = proc_dointvec_minmax,
2451                 .extra1         = (void *)&mmap_rnd_compat_bits_min,
2452                 .extra2         = (void *)&mmap_rnd_compat_bits_max,
2453         },
2454 #endif
2455 #ifdef CONFIG_USERFAULTFD
2456         {
2457                 .procname       = "unprivileged_userfaultfd",
2458                 .data           = &sysctl_unprivileged_userfaultfd,
2459                 .maxlen         = sizeof(sysctl_unprivileged_userfaultfd),
2460                 .mode           = 0644,
2461                 .proc_handler   = proc_dointvec_minmax,
2462                 .extra1         = SYSCTL_ZERO,
2463                 .extra2         = SYSCTL_ONE,
2464         },
2465 #endif
2466         { }
2467 };
2468
2469 static struct ctl_table debug_table[] = {
2470 #ifdef CONFIG_SYSCTL_EXCEPTION_TRACE
2471         {
2472                 .procname       = "exception-trace",
2473                 .data           = &show_unhandled_signals,
2474                 .maxlen         = sizeof(int),
2475                 .mode           = 0644,
2476                 .proc_handler   = proc_dointvec
2477         },
2478 #endif
2479         { }
2480 };
2481
2482 static struct ctl_table dev_table[] = {
2483         { }
2484 };
2485
2486 DECLARE_SYSCTL_BASE(kernel, kern_table);
2487 DECLARE_SYSCTL_BASE(vm, vm_table);
2488 DECLARE_SYSCTL_BASE(debug, debug_table);
2489 DECLARE_SYSCTL_BASE(dev, dev_table);
2490
2491 int __init sysctl_init_bases(void)
2492 {
2493         register_sysctl_base(kernel);
2494         register_sysctl_base(vm);
2495         register_sysctl_base(debug);
2496         register_sysctl_base(dev);
2497
2498         return 0;
2499 }
2500 #endif /* CONFIG_SYSCTL */
2501 /*
2502  * No sense putting this after each symbol definition, twice,
2503  * exception granted :-)
2504  */
2505 EXPORT_SYMBOL(proc_dobool);
2506 EXPORT_SYMBOL(proc_dointvec);
2507 EXPORT_SYMBOL(proc_douintvec);
2508 EXPORT_SYMBOL(proc_dointvec_jiffies);
2509 EXPORT_SYMBOL(proc_dointvec_minmax);
2510 EXPORT_SYMBOL_GPL(proc_douintvec_minmax);
2511 EXPORT_SYMBOL(proc_dointvec_userhz_jiffies);
2512 EXPORT_SYMBOL(proc_dointvec_ms_jiffies);
2513 EXPORT_SYMBOL(proc_dostring);
2514 EXPORT_SYMBOL(proc_doulongvec_minmax);
2515 EXPORT_SYMBOL(proc_doulongvec_ms_jiffies_minmax);
2516 EXPORT_SYMBOL(proc_do_large_bitmap);