linux/arch/csky/lib/usercopy.c

259 lines
7.7 KiB
C
Raw Normal View History

// SPDX-License-Identifier: GPL-2.0
// Copyright (C) 2018 Hangzhou C-SKY Microsystems co.,ltd.
#include <linux/uaccess.h>
#include <linux/types.h>
unsigned long raw_copy_from_user(void *to, const void *from,
unsigned long n)
{
csky: Fixup raw_copy_from_user() If raw_copy_from_user(to, from, N) returns K, callers expect the first N - K bytes starting at to to have been replaced with the contents of corresponding area starting at from and the last K bytes of destination *left* *unmodified*. What arch/sky/lib/usercopy.c is doing is broken - it can lead to e.g. data corruption on write(2). raw_copy_to_user() is inaccurate about return value, which is a bug, but consequences are less drastic than for raw_copy_from_user(). And just what are those access_ok() doing in there? I mean, look into linux/uaccess.h; that's where we do that check (as well as zero tail on failure in the callers that need zeroing). AFAICS, all of that shouldn't be hard to fix; something like a patch below might make a useful starting point. I would suggest moving these macros into usercopy.c (they are never used anywhere else) and possibly expanding them there; if you leave them alive, please at least rename __copy_user_zeroing(). Again, it must not zero anything on failed read. Said that, I'm not sure we won't be better off simply turning usercopy.c into usercopy.S - all that is left there is a couple of functions, each consisting only of inline asm. Guo Ren reply: Yes, raw_copy_from_user is wrong, it's no need zeroing code. unsigned long _copy_from_user(void *to, const void __user *from, unsigned long n) { unsigned long res = n; might_fault(); if (likely(access_ok(from, n))) { kasan_check_write(to, n); res = raw_copy_from_user(to, from, n); } if (unlikely(res)) memset(to + (n - res), 0, res); return res; } EXPORT_SYMBOL(_copy_from_user); You are right and access_ok() should be removed. but, how about: do { ... "2: stw %3, (%1, 0) \n" \ + " subi %0, 4 \n" \ "9: stw %4, (%1, 4) \n" \ + " subi %0, 4 \n" \ "10: stw %5, (%1, 8) \n" \ + " subi %0, 4 \n" \ "11: stw %6, (%1, 12) \n" \ + " subi %0, 4 \n" \ " addi %2, 16 \n" \ " addi %1, 16 \n" \ Don't expand __ex_table AI Viro reply: Hey, I've no idea about the instruction scheduling on csky - if that doesn't slow the things down, all the better. It's just that copy_to_user() and friends are on fairly hot codepaths, and in quite a few situations they will dominate the speed of e.g. read(2). So I tried to keep the fast path unchanged. Up to the architecture maintainers, obviously. Which would be you... As for the fixups size increase (__ex_table size is unchanged)... You have each of those macros expanded exactly once. So the size is not a serious argument, IMO - useless complexity would be, if it is, in fact, useless; the size... not really, especially since those extra subi will at least offset it. Again, up to you - asm optimizations of (essentially) memcpy()-style loops are tricky and can depend upon the fairly subtle details of architecture. So even on something I know reasonably well I would resort to direct experiments if I can't pass the buck to architecture maintainers. It *is* worth optimizing - this is where read() from a file that is already in page cache spends most of the time, etc. Guo Ren reply: Thx, after fixup some typo “sub %0, 4”, apply the patch. TODO: - user copy/from codes are still need optimizing. Signed-off-by: Al Viro <viro@zeniv.linux.org.uk> Signed-off-by: Guo Ren <guoren@linux.alibaba.com>
2020-04-07 09:40:11 +08:00
___copy_from_user(to, from, n);
return n;
}
EXPORT_SYMBOL(raw_copy_from_user);
unsigned long raw_copy_to_user(void *to, const void *from,
unsigned long n)
{
csky: Fixup raw_copy_from_user() If raw_copy_from_user(to, from, N) returns K, callers expect the first N - K bytes starting at to to have been replaced with the contents of corresponding area starting at from and the last K bytes of destination *left* *unmodified*. What arch/sky/lib/usercopy.c is doing is broken - it can lead to e.g. data corruption on write(2). raw_copy_to_user() is inaccurate about return value, which is a bug, but consequences are less drastic than for raw_copy_from_user(). And just what are those access_ok() doing in there? I mean, look into linux/uaccess.h; that's where we do that check (as well as zero tail on failure in the callers that need zeroing). AFAICS, all of that shouldn't be hard to fix; something like a patch below might make a useful starting point. I would suggest moving these macros into usercopy.c (they are never used anywhere else) and possibly expanding them there; if you leave them alive, please at least rename __copy_user_zeroing(). Again, it must not zero anything on failed read. Said that, I'm not sure we won't be better off simply turning usercopy.c into usercopy.S - all that is left there is a couple of functions, each consisting only of inline asm. Guo Ren reply: Yes, raw_copy_from_user is wrong, it's no need zeroing code. unsigned long _copy_from_user(void *to, const void __user *from, unsigned long n) { unsigned long res = n; might_fault(); if (likely(access_ok(from, n))) { kasan_check_write(to, n); res = raw_copy_from_user(to, from, n); } if (unlikely(res)) memset(to + (n - res), 0, res); return res; } EXPORT_SYMBOL(_copy_from_user); You are right and access_ok() should be removed. but, how about: do { ... "2: stw %3, (%1, 0) \n" \ + " subi %0, 4 \n" \ "9: stw %4, (%1, 4) \n" \ + " subi %0, 4 \n" \ "10: stw %5, (%1, 8) \n" \ + " subi %0, 4 \n" \ "11: stw %6, (%1, 12) \n" \ + " subi %0, 4 \n" \ " addi %2, 16 \n" \ " addi %1, 16 \n" \ Don't expand __ex_table AI Viro reply: Hey, I've no idea about the instruction scheduling on csky - if that doesn't slow the things down, all the better. It's just that copy_to_user() and friends are on fairly hot codepaths, and in quite a few situations they will dominate the speed of e.g. read(2). So I tried to keep the fast path unchanged. Up to the architecture maintainers, obviously. Which would be you... As for the fixups size increase (__ex_table size is unchanged)... You have each of those macros expanded exactly once. So the size is not a serious argument, IMO - useless complexity would be, if it is, in fact, useless; the size... not really, especially since those extra subi will at least offset it. Again, up to you - asm optimizations of (essentially) memcpy()-style loops are tricky and can depend upon the fairly subtle details of architecture. So even on something I know reasonably well I would resort to direct experiments if I can't pass the buck to architecture maintainers. It *is* worth optimizing - this is where read() from a file that is already in page cache spends most of the time, etc. Guo Ren reply: Thx, after fixup some typo “sub %0, 4”, apply the patch. TODO: - user copy/from codes are still need optimizing. Signed-off-by: Al Viro <viro@zeniv.linux.org.uk> Signed-off-by: Guo Ren <guoren@linux.alibaba.com>
2020-04-07 09:40:11 +08:00
___copy_to_user(to, from, n);
return n;
}
EXPORT_SYMBOL(raw_copy_to_user);
/*
* copy a null terminated string from userspace.
*/
#define __do_strncpy_from_user(dst, src, count, res) \
do { \
int tmp; \
long faultres; \
asm volatile( \
" cmpnei %3, 0 \n" \
" bf 4f \n" \
"1: cmpnei %1, 0 \n" \
" bf 5f \n" \
"2: ldb %4, (%3, 0) \n" \
" stb %4, (%2, 0) \n" \
" cmpnei %4, 0 \n" \
" bf 3f \n" \
" addi %3, 1 \n" \
" addi %2, 1 \n" \
" subi %1, 1 \n" \
" br 1b \n" \
"3: subu %0, %1 \n" \
" br 5f \n" \
"4: mov %0, %5 \n" \
" br 5f \n" \
".section __ex_table, \"a\" \n" \
".align 2 \n" \
".long 2b, 4b \n" \
".previous \n" \
"5: \n" \
: "=r"(res), "=r"(count), "=r"(dst), \
"=r"(src), "=r"(tmp), "=r"(faultres) \
: "5"(-EFAULT), "0"(count), "1"(count), \
"2"(dst), "3"(src) \
: "memory", "cc"); \
} while (0)
/*
* __strncpy_from_user: - Copy a NUL terminated string from userspace,
* with less checking.
* @dst: Destination address, in kernel space. This buffer must be at
* least @count bytes long.
* @src: Source address, in user space.
* @count: Maximum number of bytes to copy, including the trailing NUL.
*
* Copies a NUL-terminated string from userspace to kernel space.
* Caller must check the specified block with access_ok() before calling
* this function.
*
* On success, returns the length of the string (not including the trailing
* NUL).
*
* If access to userspace fails, returns -EFAULT (some data may have been
* copied).
*
* If @count is smaller than the length of the string, copies @count bytes
* and returns @count.
*/
long __strncpy_from_user(char *dst, const char *src, long count)
{
long res;
__do_strncpy_from_user(dst, src, count, res);
return res;
}
EXPORT_SYMBOL(__strncpy_from_user);
/*
* strncpy_from_user: - Copy a NUL terminated string from userspace.
* @dst: Destination address, in kernel space. This buffer must be at
* least @count bytes long.
* @src: Source address, in user space.
* @count: Maximum number of bytes to copy, including the trailing NUL.
*
* Copies a NUL-terminated string from userspace to kernel space.
*
* On success, returns the length of the string (not including the trailing
* NUL).
*
* If access to userspace fails, returns -EFAULT (some data may have been
* copied).
*
* If @count is smaller than the length of the string, copies @count bytes
* and returns @count.
*/
long strncpy_from_user(char *dst, const char *src, long count)
{
long res = -EFAULT;
Remove 'type' argument from access_ok() function Nobody has actually used the type (VERIFY_READ vs VERIFY_WRITE) argument of the user address range verification function since we got rid of the old racy i386-only code to walk page tables by hand. It existed because the original 80386 would not honor the write protect bit when in kernel mode, so you had to do COW by hand before doing any user access. But we haven't supported that in a long time, and these days the 'type' argument is a purely historical artifact. A discussion about extending 'user_access_begin()' to do the range checking resulted this patch, because there is no way we're going to move the old VERIFY_xyz interface to that model. And it's best done at the end of the merge window when I've done most of my merges, so let's just get this done once and for all. This patch was mostly done with a sed-script, with manual fix-ups for the cases that weren't of the trivial 'access_ok(VERIFY_xyz' form. There were a couple of notable cases: - csky still had the old "verify_area()" name as an alias. - the iter_iov code had magical hardcoded knowledge of the actual values of VERIFY_{READ,WRITE} (not that they mattered, since nothing really used it) - microblaze used the type argument for a debug printout but other than those oddities this should be a total no-op patch. I tried to fix up all architectures, did fairly extensive grepping for access_ok() uses, and the changes are trivial, but I may have missed something. Any missed conversion should be trivially fixable, though. Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-01-04 10:57:57 +08:00
if (access_ok(src, 1))
__do_strncpy_from_user(dst, src, count, res);
return res;
}
EXPORT_SYMBOL(strncpy_from_user);
/*
* strlen_user: - Get the size of a string in user space.
* @str: The string to measure.
* @n: The maximum valid length
*
* Get the size of a NUL-terminated string in user space.
*
* Returns the size of the string INCLUDING the terminating NUL.
* On exception, returns 0.
* If the string is too long, returns a value greater than @n.
*/
long strnlen_user(const char *s, long n)
{
unsigned long res, tmp;
if (s == NULL)
return 0;
asm volatile(
" cmpnei %1, 0 \n"
" bf 3f \n"
"1: cmpnei %0, 0 \n"
" bf 3f \n"
"2: ldb %3, (%1, 0) \n"
" cmpnei %3, 0 \n"
" bf 3f \n"
" subi %0, 1 \n"
" addi %1, 1 \n"
" br 1b \n"
"3: subu %2, %0 \n"
" addi %2, 1 \n"
" br 5f \n"
"4: movi %0, 0 \n"
" br 5f \n"
".section __ex_table, \"a\" \n"
".align 2 \n"
".long 2b, 4b \n"
".previous \n"
"5: \n"
: "=r"(n), "=r"(s), "=r"(res), "=r"(tmp)
: "0"(n), "1"(s), "2"(n)
: "memory", "cc");
return res;
}
EXPORT_SYMBOL(strnlen_user);
#define __do_clear_user(addr, size) \
do { \
int __d0, zvalue, tmp; \
\
asm volatile( \
"0: cmpnei %1, 0 \n" \
" bf 7f \n" \
" mov %3, %1 \n" \
" andi %3, 3 \n" \
" cmpnei %3, 0 \n" \
" bf 1f \n" \
" br 5f \n" \
"1: cmplti %0, 32 \n" /* 4W */ \
" bt 3f \n" \
"8: stw %2, (%1, 0) \n" \
"10: stw %2, (%1, 4) \n" \
"11: stw %2, (%1, 8) \n" \
"12: stw %2, (%1, 12) \n" \
"13: stw %2, (%1, 16) \n" \
"14: stw %2, (%1, 20) \n" \
"15: stw %2, (%1, 24) \n" \
"16: stw %2, (%1, 28) \n" \
" addi %1, 32 \n" \
" subi %0, 32 \n" \
" br 1b \n" \
"3: cmplti %0, 4 \n" /* 1W */ \
" bt 5f \n" \
"4: stw %2, (%1, 0) \n" \
" addi %1, 4 \n" \
" subi %0, 4 \n" \
" br 3b \n" \
"5: cmpnei %0, 0 \n" /* 1B */ \
"9: bf 7f \n" \
"6: stb %2, (%1, 0) \n" \
" addi %1, 1 \n" \
" subi %0, 1 \n" \
" br 5b \n" \
".section __ex_table,\"a\" \n" \
".align 2 \n" \
".long 8b, 9b \n" \
".long 10b, 9b \n" \
".long 11b, 9b \n" \
".long 12b, 9b \n" \
".long 13b, 9b \n" \
".long 14b, 9b \n" \
".long 15b, 9b \n" \
".long 16b, 9b \n" \
".long 4b, 9b \n" \
".long 6b, 9b \n" \
".previous \n" \
"7: \n" \
: "=r"(size), "=r" (__d0), \
"=r"(zvalue), "=r"(tmp) \
: "0"(size), "1"(addr), "2"(0) \
: "memory", "cc"); \
} while (0)
/*
* clear_user: - Zero a block of memory in user space.
* @to: Destination address, in user space.
* @n: Number of bytes to zero.
*
* Zero a block of memory in user space.
*
* Returns number of bytes that could not be cleared.
* On success, this will be zero.
*/
unsigned long
clear_user(void __user *to, unsigned long n)
{
Remove 'type' argument from access_ok() function Nobody has actually used the type (VERIFY_READ vs VERIFY_WRITE) argument of the user address range verification function since we got rid of the old racy i386-only code to walk page tables by hand. It existed because the original 80386 would not honor the write protect bit when in kernel mode, so you had to do COW by hand before doing any user access. But we haven't supported that in a long time, and these days the 'type' argument is a purely historical artifact. A discussion about extending 'user_access_begin()' to do the range checking resulted this patch, because there is no way we're going to move the old VERIFY_xyz interface to that model. And it's best done at the end of the merge window when I've done most of my merges, so let's just get this done once and for all. This patch was mostly done with a sed-script, with manual fix-ups for the cases that weren't of the trivial 'access_ok(VERIFY_xyz' form. There were a couple of notable cases: - csky still had the old "verify_area()" name as an alias. - the iter_iov code had magical hardcoded knowledge of the actual values of VERIFY_{READ,WRITE} (not that they mattered, since nothing really used it) - microblaze used the type argument for a debug printout but other than those oddities this should be a total no-op patch. I tried to fix up all architectures, did fairly extensive grepping for access_ok() uses, and the changes are trivial, but I may have missed something. Any missed conversion should be trivially fixable, though. Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2019-01-04 10:57:57 +08:00
if (access_ok(to, n))
__do_clear_user(to, n);
return n;
}
EXPORT_SYMBOL(clear_user);
/*
* __clear_user: - Zero a block of memory in user space, with less checking.
* @to: Destination address, in user space.
* @n: Number of bytes to zero.
*
* Zero a block of memory in user space. Caller must check
* the specified block with access_ok() before calling this function.
*
* Returns number of bytes that could not be cleared.
* On success, this will be zero.
*/
unsigned long
__clear_user(void __user *to, unsigned long n)
{
__do_clear_user(to, n);
return n;
}
EXPORT_SYMBOL(__clear_user);