Re: [PATCH] atomic: Fix bugs in 'fetch_or()' and rename it to 'xchg_or()'

[Peter Zijlstra]

On Tue, Mar 15, 2016 at 10:32:45AM +0100, Ingo Molnar wrote:
> +#ifndef xchg_or
> +# define xchg_or(ptr, mask)						\
> +({									\
> +	typeof(ptr)  __ptr  = (ptr);					\
> +	typeof(mask) __mask = (mask);					\
> +									\
> +	typeof(*(__ptr)) __old, __val = *__ptr;				\
> +									\
>  	for (;;) {							\
> +		__old = cmpxchg(__ptr, __val, __val | __mask);		\
>  		if (__old == __val)					\
>  			break;						\
>  		__val = __old;						\
>  	}								\
> +									\
>  	__old;								\
>  })

As reported by you this explodes, and it obvious from the generated asm
why:

     48e1: 89 c2                 mov    %eax,%edx
     48e3: 41 89 d0              mov    %edx,%r8d
     48e6: 31 c9                 xor    %ecx,%ecx
     48e8: 89 d0                 mov    %edx,%eax
     48ea: 41 83 c8 08           or     $0x8,%r8d
     48ee: f0 44 0f b1 01        lock cmpxchg %r8d,(%rcx)
     48f3: 39 c2                 cmp    %eax,%edx
     48f5: 75 ea                 jne    48e1 <resched_curr+0x31>

That's an unconditional NULL deref.

What happens is that __ptr from xchg_or() aliasses with __ptr from
cmpxchg() and weird stuff happens.

If you do: s/__ptr/_ptr/ or similar on the xchg_or() code it all works
again.