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