Re: [Qemu-devel] [PATCH v6 1/3] IOMMU: add option to enable VTD_CAP_CM to vIOMMU capility exposoed to guest

[Jason Wang]

On 2016年11月21日 21:35, Michael S. Tsirkin wrote:
> On Fri, Nov 11, 2016 at 12:15:48PM +0800, Jason Wang wrote:
> > >
> > >
> > >On 2016年11月11日 11:39, Michael S. Tsirkin wrote:
> > > > >On Fri, Nov 11, 2016 at 10:32:42AM +0800, Jason Wang wrote:
> > > > > > >
> > > > > > >On 2016年11月10日 06:00, Michael S. Tsirkin wrote:
> > > > > > > > >On Wed, Nov 09, 2016 at 03:28:02PM +0800, Jason Wang wrote:
> > > > > > > > > > > > >
> > > > > > > > > > > > >On 2016年11月08日 19:04, Aviv B.D wrote:
> > > > > > > > > > > > > > > > >From: "Aviv Ben-David"<address@hidden>
> > > > > > > > > > > > > > > > >
> > > > > > > > > > > > > > > > >This capability asks the guest to invalidate cache before each 
> > > > > > > > > map operation.
> > > > > > > > > > > > > > > > >We can use this invalidation to trap map operations in the 
> > > > > > > > > hypervisor.
> > > > > > > > > > > > >Hi:
> > > > > > > > > > > > >
> > > > > > > > > > > > >Like I've asked twice in the past, I want to know why don't you cache
> > > > > > > > > > > > >translation faults as what spec required (especially this is a guest 
> > > > > > > visible
> > > > > > > > > > > > >behavior)?
> > > > > > > > > > > > >
> > > > > > > > > > > > >Btw, please cc me on posting future versions.
> > > > > > > > > > > > >
> > > > > > > > > > > > >Thanks
> > > > > > > > >Caching isn't guest visible.
> > > > > > >Seems not, if one fault mapping were cached by IOTLB. Guest can notice this
> > > > > > >behavior.
> > > > >Sorry, I don't get what you are saying.
> > > > >
> > > > > > > > >Spec just says you*can*  cache,
> > > > > > > > >not that you must.
> > > > > > > > >
> > > > > > >Yes, but what did in this patch is "don't". What I suggest is just a "can",
> > > > > > >since anyway the IOTLB entries were limited and could be replaced by other.
> > > > > > >
> > > > > > >Thanks
> > > > >Have trouble understanding this. Can you given an example of
> > > > >a guest visible difference?
> > >
> > >I guess this may do the detection:
> > >
> > >1) map iova A to be non-present.
> > >2) invalidate iova A
> > >3) map iova A to addr B
> > >4) access iova A
> > >
> > >A correct implemented CM may meet fault in step 4, but with this patch, we
> > >never.
> I think that IOTLB is free to invalidate entries at any point,
> so the fault is not guaranteed on bare metal.

Yes, that's the interesting point. The fault is not guaranteed but 
conditional. And we have similar issue for IEC.

So in conclusion (since I can't find a hardware IOMMU that have CM set):

1) If we don't cache fault conditions, we are still in question that 
whether it was spec compatible.
2) If we do cache fault conditions, we are 100% sure it was spec 
compatible.

Consider 2) is not complicated, we'd better do it I believe?

Thanks