Kevin Wolf <kw...@redhat.com> writes:
> Am 23.09.2011 10:55, schrieb Markus Armbruster:
[...]
>> Apropos enforcing. Currently, -drive accepts any werror and rerror
>> action with if={ide,virtio,scsi,none}. We rely on device models not
>> implementing an action to check and fail during initialization.
>> scsi-generic and the floppy devices do. All the other qdevified devices
>> don't, and that's broken.
>>
>> Are werror and rerror really host state?
>
> If you want to convince me of the opposite, show me the real device that
> implements them.
It's an honest question.
> In fact, for the implementation of rerror/werror the very same thing is
> true as you say about I/O status updates. Other than being easy to
> implement, there is really no reason to do it in the devices. The block
> layer could just as well stop the VM and queue the requests for
> resubmission when the VM is restarted. Of course, you can't keep
> migration compatibility when doing such a fundamental change to the
> design, but if you take the host state thing serious, this is what you
> should do.
Good answer. I'd love to have it spelled out in the code.
