Re: [PATCH 4/5] xen: use masking operation instead of test_bit for VPF bits

[George Dunlap <george.dunlap@citrix.com>]

On 05/10/15 14:36, Jan Beulich wrote:
>>>> On 05.10.15 at 15:18, <george.dunlap@citrix.com> wrote:
>> On 02/10/15 05:40, Juergen Gross wrote:
>>> Use a bit mask for testing of a set bit instead of test_bit in case no
>>> atomic operation is needed, as this will lead to smaller and more
>>> effective code.
>>>
>>> Signed-off-by: Juergen Gross <jgross@suse.com>
>>
>> I'm a bit confused here -- exactly when is an atomic operation needed or
>> not needed?  Isn't it the case that we always need to do an atomic read
>> if we ever do an atomic write without a lock held?
> 
> First of all - what is an atomic read from CPU perspective other than
> just a read? Since we talk about individual bits here, we don't care
> about the granularity the compiler may convert the read to, and even
> if the compiler chose to do multiple reads the result would still be
> correct, as only one of those reads can possibly have read the bit
> in question.
> 
> And then, the old mechanism was in no way "atomic", all it added was
> a kind of compiler barrier (due to the cast to volatile). Yet in none of
> the cases changed I was able to spot a need for such a barrier.

OK, so the key thing about test_bit isn't that it's atomic, so much that
it's an implicit memory barrier.  So as long as you're not doing a
lockless careful-ordering sort of thing, then a simple memory read
should be fine.  Is that correct?

In that case, it's likely that the patch is correct (though I'll take a
closer look just to be sure).

 -George