On 2025/01/16 1:14, Peter Xu wrote:
On Thu, Jan 16, 2025 at 12:52:56AM +0900, Akihiko Odaki wrote:
Functionally, the ordering of container/subregion finalization matters if
some device tries to a container during finalization. In such a case,
|
^ something is missing here, feel free to complete this.
Oops, I meant: functionally, the ordering of container/subregion
finalization matters if some device tries to use a container during
finalization.
removing subregions from the container at random timing can result in an
unexpected behavior. There is little chance to have such a scenario but we
should stay the safe side if possible.
It sounds like a future feature, and I'm not sure we'll get there, so I
don't worry that much. Keeping refcount core idea simple is still very
attractive to me. I still prefer we have complete MR refcounting iff when
necessary. It's also possible it'll never happen to QEMU.
It's not just about the future but also about compatibility with the
current device implementations. I will not be surprised even if the
random ordering of subregion finalization breaks one of dozens of
devices we already have. We should pay attention the details as we are
touching the core infrastructure.
