Re: [linuxsh-dev] bad alignment fixup

[Philipp R. <pr...@pa...>]

On Sat, Nov 11, 2000 at 10:59:04AM +0900, NIIBE Yutaka wrote:
> Jesper Skov wrote:
>  > The below patch is based on an old patch from Philipp Rumpf. It allows
>  > the kernel to make unaligned access if it should happen to think that
>  > it a good thing to do.
> 
> Thanks, installed.

I think the patch leaves a security hole: if you put_user to an unaligned
invalid address, the kernel will oops.

The fix is something like (optimized for code size not speed)

static inline void unaligned_memcpy(void *dst, void *src, int count,
				    struct pt_regs *regs)
{
	int failed;

	while(count--) {
		asm volatile("
1:	mov.b	@(%2),%0
	bra	4f
2:	mov.b	%0,@(%3)
3:	mov	#1,%1
4:
	.section	__ex_table,\"a\"
	.long	1b,2b
	.previous" : "=r" (tmp), "=r" (failed) : "r" (dst++), "r" (src++),
		"0" (0), "1" (0));

		if (failed) {
			die_if_no_fixup("invalid unaligned access", regs, ???);
			return;
		}
	}
}

(my apologies for the ugly code).

An alternative fix, and one which I think I like better, is to try the
search_exception_table before doing any unaligned handling - unaligned
userspace pointers aren't the kernel's job to deal with.