Re: [thiscpuops upgrade 09/10] x86: this_cpu_cmpxchg and this_cpu_cmpxchg_double operations

[Mathieu Desnoyers <mathieu.desnoyers@efficios.com>]

* Christoph Lameter (cl@linux.com) wrote:
> Provide support as far as the hardware capabilities of the x86 cpus
> allow.
> 
> Signed-off-by: Christoph Lameter <cl@linux.com>
> 

[...]
> +#define percpu_cmpxchg16b(pcp, o1, o2, n1, n2)				\
> +({									\
> +	char __ret;							\
> +	typeof(o1) __o1 = o1;						\
> +	typeof(o1) __n1 = n1;						\
> +	typeof(o2) __o2 = o2;						\
> +	typeof(o2) __n2 = n2;						\
> +	asm("cmpxchg16b "__percpu_arg(1)"\n\tsetz %0\n\t"		\
> +		    : "=a"(__ret), "=m" (*pcp)				\
> +		    :  "b"(__n1), "c"(__n2), "a"(__o1), "d"(__o2));	\
> +	__ret;								\
> +})

Can we add a check to ensure that the target memory location is 16-byte aligned?
This is a documented limitation of cmpxchg16b.

Thanks,

Mathieu

> +
> +#define __this_cpu_cmpxchg_double_8(pcp, o1, o2, n1, n2) percpu_cmpxchg16b((pcp), o1, o2, n1, n2)
> +#define this_cpu_cmpxchg_double_8(pcp, o1, o2, n1, n2)	percpu_cmpxchg16b((pcp), o1, o2, n1, n2)
> +#define irqsafe_cmpxchg_double_8(pcp, o1, o2, n1, n2)	percpu_cmpxchg16b((pcp), o1, o2, n1, n2)
> +
>  #endif
>  
>  /* This is not atomic against other CPUs -- CPU preemption needs to be off */
> 

-- 
Mathieu Desnoyers
Operating System Efficiency R&D Consultant
EfficiOS Inc.
http://www.efficios.com