cachepc-linux

Fork of AMDESE/linux with modifications for CachePC side-channel attack
git clone https://git.sinitax.com/sinitax/cachepc-linux
Log | Files | Refs | README | LICENSE | sfeed.txt

sysctl.c (59640B)


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