Re: [PATCH v3] x86/VPMU: Clear last_vcpu when destroying VPMU

[Boris Ostrovsky <boris.ostrovsky@oracle.com>]

On 12/17/2014 12:28 PM, Jan Beulich wrote:
>>>> Boris Ostrovsky <boris.ostrovsky@oracle.com> 12/17/14 4:10 PM >>>
>> +    /* Need to clear last_vcpu in case it points to v */
>> +    (void)cmpxchg(last, v, NULL);
>   
> In a (later) reply to v2 I had indicated that it doesn't seem safe to so here but
> rely on an IPI in the other path altering that value. Can you explain why you
> think this is safe without changing the others to CPU-atomic accesses too?


So the local access to last_vcpu in vpmu_load() may be still raced on (I 
was thinking that access in vpmu_load() would on the same processor as 
vpmu_destroy(), which is obviously not the case).

So I will have to go back to IPI (I don't want to use cmpxchg in 
vpmu_load()).

-boris