On Wed, Sep 17, 2025 at 07:32:48PM +0900, Akihiko Odaki wrote:
> +static int del_memory_region(Object *child, void *opaque)
> +{
> + MemoryRegion *mr = (MemoryRegion *)object_dynamic_cast(child,
> TYPE_MEMORY_REGION);
> +
> + if (mr && mr->container) {
> + memory_region_del_subregion(mr->container, mr);
> + }
> +
> + return 0;
> +}
> +
> static void device_set_realized(Object *obj, bool value, Error **errp)
> {
> DeviceState *dev = DEVICE(obj);
> @@ -582,6 +593,7 @@ static void device_set_realized(Object *obj, bool value,
> Error **errp)
> if (dc->unrealize) {
> dc->unrealize(dev);
> }
> + object_child_foreach(OBJECT(dev), del_memory_region, NULL);
PS: I'll keep throwing some pure questions here, again, Paolo - it doesn't
need to block merging if you're confident with the general approach.
Said that, a few things I still want to mention after I read this series..
One thing I really feel hard to review such work is, you hardly describe
the problems the series is resolving.
For example, this patch proposed auto-detach MRs in unrealize() for qdev,
however there's nowhere describing "what will start to work, while it
doesn't", "how bad is the problem", etc.. All the rest patches are about
"what we can avoid do" after this patch.
Meanwhile, the cover letter is misleading. It is:
[PATCH 00/14] Fix memory region use-after-finalization
I believe it's simply wrong, because the whole series is not about
finalize() but unrealize(). For Device class, it also includes the exit()
which will be invoked in pci_qdev_unrealize(), but that is also part of the
unrealize() routine, not finalize().
The other question is, what if a MR has a owner that is not the device
itself? There's no place enforcing this, hence a qdev can logically have
some sub-objects (which may not really be qdev) that can be the owner of
the memory regions. Then the device emulation will found that some MRs are
auto-detached and some are not.
One example that I'm aware of is this:
https://lore.kernel.org/all/[email protected]/#t
TYPE_VIRTIO_SHARED_MEMORY_MAPPING is an object, not qdev here, which can be
the owner of the MR.
Thanks,
--
Peter Xu
