The ARM __kuser_cmpxchg routine is meant to implement an atomic cmpxchg
in user space.  It however can produce spurious false negative if a
processor exception occurs in the middle of the operation.  Normally
this is not a problem since cmpxchg is typically called in a loop until
it succeeds to implement an atomic increment for example.

Some use cases which don't involve a loop require that the operation be
100% reliable though.  This patch changes the implementation so to
reattempt the operation after an exception has occurred in the critical
section rather than abort it.

Here's a simple program to test the fix (don't use CONFIG_NO_HZ in your
kernel as this depends on a sufficiently high interrupt rate):

	#include <stdio.h>

	typedef int (__kernel_cmpxchg_t)(int oldval, int newval, int *ptr);
	#define __kernel_cmpxchg (*(__kernel_cmpxchg_t *)0xffff0fc0)

	int main()
	{
		int i, x = 0;
		for (i = 0; i < 100000000; i++) {
			int v = x;
			if (__kernel_cmpxchg(v, v+1, &x))
				printf("failed at %d: %d vs %d\n", i, v, x);
		}
		printf("done with %d vs %d\n", i, x);
		return 0;
	}

Signed-off-by: Nicolas Pitre <nico@marvell.com>
Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>