Re: [Qemu-devel] [PATCH v3 3/3] IOMMU: Integrate between VFIO and vIOMMU to support device assignment

[Peter Xu <peterx@redhat.com>]

On Thu, Jun 02, 2016 at 07:34:17AM -0600, Alex Williamson wrote:
> On Thu, 02 Jun 2016 13:09:27 +0000
> "Aviv B.D." <bd.aviv@gmail.com> wrote:
> 
> > Hi,
> > 
> > In case of hot plug vfio device there should not be any active mapping
> > to this device prior the device addition.
> 
> Counter example - a device is hot added to a guest booted with iommu=pt.

I got the same question with Aviv...

For hot-plug devices, even if it is using iommu=pt, shouldn't it still
follow the steps that first init vfio device, then configure device
context entry? Let me list the steps for device addition in case I got
any mistake:

1. user add new VFIO device A

2. vfio_listener_region_add() called for device A on the IOMMU mr,
   here we should create the iommu notifier. However since the context
   entry still does not exist, memory_region_iommu_replay() will got
   all invalid IOTLB (IOMMU_NONE entries)

3. guest kernel found the device, enabled the device, filled in
   context entry for device A with "pass-through" (so the SLPTPTR is
   invalid)

4. guest sent context invalidation to QEMU vIOMMU since we have CM=1
   set for guest vIOMMU

5. QEMU vIOMMU handle the invalidation, trigger VFIO notify to do
   correct VFIO mapping for device A

Though here step 5 should still be missing (IIUC Aviv's patch 3 still
not handled context invalidation). Just want to know whether we can
avoid the replay operation for Intel vIOMMUs (for Intel only, because
Intel has context invalidation and cache mode support, not sure about
other platform)?

Thanks,

-- peterx