Re: [PATCH for-7.2] block-backend: avoid bdrv_unregister_buf() NULL pointer deref

[Stefan Hajnoczi]

On Tue, 22 Nov 2022 at 03:22, Kevin Wolf <kwolf@redhat.com> wrote:
>
> Am 21.11.2022 um 22:19 hat Stefan Hajnoczi geschrieben:
> > bdrv_*() APIs expect a valid BlockDriverState. Calling them with bs=NULL
> > leads to undefined behavior.
> >
> > Jonathan Cameron reported this following NULL pointer dereference when a
> > VM with a virtio-blk device and a memory-backend-file object is
> > terminated:
> > 1. qemu_cleanup() closes all drives, setting blk->root to NULL
> > 2. qemu_cleanup() calls user_creatable_cleanup(), which results in a RAM
> >    block notifier callback because the memory-backend-file is destroyed.
> > 3. blk_unregister_buf() is called by virtio-blk's BlockRamRegistrar
> >    notifier callback and undefined behavior occurs.
> >
> > Fixes: baf422684d73 ("virtio-blk: use BDRV_REQ_REGISTERED_BUF optimization 
> > hint")
> > Co-authored-by: Jonathan Cameron <Jonathan.Cameron@huawei.com>
> > Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
>
> Reviewed-by: Kevin Wolf <kwolf@redhat.com>
>
> This raises some questions, though. What happens if the graph isn't
> static between creation and deletion of the device? Do we need to do
> something with registered buffers when a node is attached to or detached
> from an existing device?

I think you are right. Graph changes need to be handled. Right now they aren't.

Stefan