Skip to contentSkip to conversation

generic uaccess.h

PATCH Michal Simek 2 participants 5 messages Jul 24, 2009 – Jul 27, 2009 Generic include/asm header files

lore
Share & tools
lore
#0 lore raw

Hi Arnd,

I have just look at asm-generic uaccess.h and there is one thing which seems to me wrong.

For put_user macro - you use __copy_to_user but you have for 64bit case ifdef CONFIG_64BIT
but look at fs/eventfd: eventfd_read function. At least for this function(syscall) is necessary "return" 64bit
value on 32bit machines too.
IMHO that ifdef CONFIG_64BIT shouldn't be there.

What do you think?
If you agree with me, I'll generate proper patch with description.

Thanks,
Michal

include/asm-generic/uaccess.h +0 -0
index b218b85..bbbb983 100644
@@ -119,11 +119,9 @@ static inline __must_check long __copy_to_user(void

__user

                case 4:
                        *(u32 __force *)to = *(u32 *)from;
                        return 0;
-#ifdef CONFIG_64BIT
                case 8:
                        *(u64 __force *)to = *(u64 *)from;
                        return 0;
-#endif
                default:
                        break;
                }
Signature · 2 lines
Michal Simek, Ing. (M.Eng)
w: www.monstr.eu p: +42-0-721842854
maintainer
#1 lore raw
Quoted 10 lines · Michal Simek

I have just look at asm-generic uaccess.h and there is one thing which seems to me wrong.

For put_user macro - you use __copy_to_user but you have for 64bit case ifdef CONFIG_64BIT
but look at fs/eventfd: eventfd_read function. At least for this function(syscall) is necessary "return" 64bit
value on 32bit machines too.
IMHO that ifdef CONFIG_64BIT shouldn't be there.

Quoted 2 lines

What do you think?
If you agree with me, I'll generate proper patch with description.

The code was intentional, because 32 bit architectures normally don't acces u64 values efficiently. I would expect the memcpy to produce better object code in that case.
Did you see an actual bug in my version or are you only guessing that the assignment should work better than the memcpy?

What object code do you get with

int test(unsigned long long __user *out, unsigned long long in)

{
	return put_user(in, ptr);
}

in both cases?

	Arnd <><
#2 lore raw
Quoted 14 lines · Arnd Bergmann
Quoted 10 lines · Michal Simek

I have just look at asm-generic uaccess.h and there is one thing which seems to me wrong.

For put_user macro - you use __copy_to_user but you have for 64bit case ifdef CONFIG_64BIT
but look at fs/eventfd: eventfd_read function. At least for this function(syscall) is necessary "return" 64bit
value on 32bit machines too.
IMHO that ifdef CONFIG_64BIT shouldn't be there.

Quoted 2 lines

What do you think?
If you agree with me, I'll generate proper patch with description.

Quoted 6 lines

The code was intentional, because 32 bit architectures normally don't acces u64 values efficiently. I would expect the memcpy to produce better object code in that case.
Did you see an actual bug in my version or are you only guessing that the assignment should work better than the memcpy?

I have just compile noMMU version and first bootup failed. I miss that memcpy after it. Ooou.

I'll let you know what my results are.

Michal

Quoted 13 lines

What object code do you get with

int test(unsigned long long __user *out, unsigned long long in)

{
	return put_user(in, ptr);
}

in both cases?

	Arnd <><
Signature · 2 lines
Michal Simek, Ing. (M.Eng)
w: www.monstr.eu p: +42-0-721842854
#3 lore raw
In reply to Arnd Bergmann
Quoted 21 lines · Arnd Bergmann
Quoted 10 lines · Michal Simek

I have just look at asm-generic uaccess.h and there is one thing which seems to me wrong.

For put_user macro - you use __copy_to_user but you have for 64bit case ifdef CONFIG_64BIT
but look at fs/eventfd: eventfd_read function. At least for this function(syscall) is necessary "return" 64bit
value on 32bit machines too.
IMHO that ifdef CONFIG_64BIT shouldn't be there.

Quoted 2 lines

What do you think?
If you agree with me, I'll generate proper patch with description.

The code was intentional, because 32 bit architectures normally don't acces u64 values efficiently. I would expect the memcpy to produce better object code in that case.
Did you see an actual bug in my version or are you only guessing that the assignment should work better than the memcpy?

Quoted 8 lines

What object code do you get with

int test(unsigned long long __user *out, unsigned long long in)

{
	return put_user(in, ptr);
}

in both cases?

I found that I have problem on noMMU with replacing put_user macro

with

#define put_user(x, ptr) \

({								\
	might_sleep();						\
	access_ok(VERIFY_WRITE, ptr, sizeof(*ptr)) ?		\
		__put_user(x, ptr) :				\
		-EFAULT;					\
})

(The similar issue is for get_user) (__put_user function is origin - not asm-generic)

I am getting any short write error. It is caused with adding access_ok macro. It is weird.

RPC: Registered tcp transport module.
VFS: Mounted root (romfs filesystem) readonly on device 31:0. Freeing unused kernel memory: 100k freed
0: short write

Kernel panic - not syncing: Attempted to kill init!

I have more important work in front of me - I take a look at it later.

Michal

Quoted 2 lines
	Arnd <><
Signature · 2 lines
Michal Simek, Ing. (M.Eng)
w: www.monstr.eu p: +42-0-721842854
maintainer
#4 lore raw
Quoted 2 lines · Michal Simek

I am getting any short write error. It is caused with adding access_ok macro. It is weird.

Your implementation of access_ok() is minimal, but should be ok in theory. Note that even on nommu, it might make sense to check for kernel addresses here and implement get_fs/set_fs, e.g.

int access_ok(int type, unsigned long addr, size_t len)

{
	/* check wraparound */
	if (addr + len < addr)
		return 0;

	if (addr < memory_start || addr + len > memory_start)
		return 0;

	/* don't allow access to kernel memory */
	if (get_fs() == USER_DS) {
		if (is_kernel(addr) || is_kernel(addr + len))
			return 0;
	}

	return 1;
}
Quoted 1 line

I have more important work in front of me - I take a look at it later.

ok. When you get to debug this, I guess the easiest way is to stick some printk into your access_ok() function.

	Arnd <><