Re: [PATCH 1/3] xen/vt-d: need barriers to workaround CLFLUSH

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

On 05/06/2015 03:12 AM, Jan Beulich wrote:
>>>> On 05.05.15 at 18:11, <boris.ostrovsky@oracle.com> wrote:
>> On 05/05/2015 11:58 AM, Jan Beulich wrote:
>>>>>> On 05.05.15 at 17:46, <boris.ostrovsky@oracle.com> wrote:
>>>> On 05/04/2015 05:14 AM, Andrew Cooper wrote:
>>>>> On 04/05/2015 09:52, Jan Beulich wrote:
>>>>>>>>> On 04.05.15 at 04:16, <tiejun.chen@intel.com> wrote:
>>>>>>> --- a/xen/drivers/passthrough/vtd/x86/vtd.c
>>>>>>> +++ b/xen/drivers/passthrough/vtd/x86/vtd.c
>>>>>>> @@ -56,7 +56,9 @@ unsigned int get_cache_line_size(void)
>>>>>>>     
>>>>>>>     void cacheline_flush(char * addr)
>>>>>>>     {
>>>>>>> +    mb();
>>>>>>>         clflush(addr);
>>>>>>> +    mb();
>>>>>>>     }
>>>>>> I think the purpose of the flush is to force write back, not to evict
>>>>>> the cache line, and if so wmb() would appear to be sufficient. As
>>>>>> the SDM says that's not the case, a comment explaining why wmb()
>>>>>> is not sufficient would seem necessary. Plus in the description I
>>>>>> think "serializing" needs to be changed to "fencing", as serialization
>>>>>> is not what we really care about here. If you and the maintainers
>>>>>> agree, I could certainly fix up both aspects while committing.
>>>>> On the subject of writebacks, we should get around to alternating-up the
>>>>> use of clflushopt and clwb, either of which would be better than a
>>>>> clflush in this case (avoiding the need for the leading mfence).
>>>>>
>>>>> However, the ISA extension document does not indicate which processors
>>>>> will have support for these new instructions.
>>>> https://software.intel.com/sites/default/files/managed/0d/53/319433-022.pdf
>>>>
>>>> We really should add support for this. On shutting down a very large
>>>> guest (hundreds of GB) we observed *minutes* spent in flushing IOMMU.
>>>> This was due to serializing nature of CLFLUSH.
>>> But flushing the IOMMU isn't being done via CPU instructions, but
>>> rather via commands sent to the IOMMU. I.e. I'm somewhat
>>> confused by your reply.
>> I didn't mean flushing IOMMU itself, sorry. I meant
>> __iommu_flush_cache() (or whatever it's equivalent we had in the
>> product, which was 4.1-based).
> In that case I wonder how much of that flushing is really necessary
> during IOMMU teardown. VT-d maintainers?

I should mention that we saw this on 4.1, where iommu teardown is done 
in the domain destruction code path. After we backported code that moves 
this into a tasklet things got much better.

Of course we still are doing flushing, even if in the background, and 
therefore system resources are still being consumed (and memory is not 
available until flushing is done). So if this is unnecessary then there 
are good reasons not to do it.

-boris