Re: [PATCH v4 10/13] memory: Change NotifyStateClear() definition to return the result

[David Hildenbrand <david@redhat.com>]

On 09.05.25 12:04, Chenyi Qiang wrote:
> 
> 
> On 5/9/2025 4:22 PM, Baolu Lu wrote:
>> On 4/27/2025 10:26 AM, Chenyi Qiang wrote:
>>> Hi David,
>>>
>>> Any thought on patch 10-12, which is to move the change attribute into a
>>> priority listener. A problem is how to handle the error handling of
>>> private_to_shared failure. Previously, we thought it would never be able
>>> to fail, but right now, it is possible in corner cases (e.g. -ENOMEM) in
>>> set_attribute_private(). At present, I simply raise an assert instead of
>>> adding any rollback work (see patch 11).
>>
>> Do the pages need to be pinned when converting them to a shared state
>> and unpinned when converting to a private state? Or is this handled
>> within the vfio_state_change_notify callbacks?
> 
> I think it is handled in vfio_state_change_notify(). Just like the
> device passthrough in legacy VM, the shared memory will be pinned during
> vfio dma-map and unpin during unmap.

We'll have to "unmap/unpin before shared->private" and "map/pin after 
private->shared" conversion.


vfio cannot fail unmap/unpin, but guest_memfd will be able to "easily" 
fail shared->private. But in that case (temporary references) we'll 
likely have to retry the conversion until it works.

guest_memfd cannot "easily" fail private->shared conversion, but vfio 
can fail map/pin, in which case we probably have to abort the conversion.

Error handling/recovery will be a bit more tricky than it is today.

-- 
Cheers,

David / dhildenb