On 2025/10/03 0:03, Paolo Bonzini wrote:
On 9/17/25 12:32, Akihiko Odaki wrote:
Based-on: <[email protected]>
("[PATCH v3 0/7] Do not unparent in instance_finalize()")
This patch series was spun off from "[PATCH v2 00/15] Fix memory region
leaks and use-after-finalization":
https://lore.kernel.org/qemu-devel/20250915-use-v2-0-
[email protected]/
When developing the next version of "[PATCH 00/16] memory: Stop
piggybacking on memory region owners*", I faced multiple memory region
leaks and use-after-finalization. This series extracts their fixes so
that the number of Cc: won't explode.
Patch "qdev: Automatically delete memory subregions" and the succeeding
patches are for refactoring, but patch "vfio-user: Do not delete the
subregion" does fix use-after-finalization.
* https://lore.kernel.org/qemu-devel/20250901-mr-v1-0-
[email protected]/
Signed-off-by: Akihiko Odaki <[email protected]>
This makes sense, but I think it is not bisectable, because of this in
memory_region_del_subregion():
assert(subregion->container == mr);
subregion->container = NULL;
You would need to add a temporary
if (subregion->container == NULL) {
return;
}
and undo it at the end of the series. Do you agree? With this change I
can apply it.
It is unnecessary because patch "qdev: Automatically delete memory
subregions" satisfies the following:
1. the device-specific code can assume that subregions they added are
present until it finishes unrealization. The unrealize() callback can
also assume the subregions are present and delete them. qdev satisfies
this by deleting subregions only after calling the unrealize().
2. qdev should delete the remaining subregions before it finishes
unrealization to ensure that the devices are hidden from the guest. qdev
satisfies this by checking if memory regions have containers before
deleting.
Regards,
Akihiko Odaki
