This RFC is just the beginning. The same problem exists for virtio-net and
other devices. I am looking for feedback before I convert all of virtio.
The virtio transport/device split is broken as follows:
1. The virtio-blk device is never finalized because the transport devices
(virtio-blk-pci and friends) leak the refcount.
2. If we fix the refcount leak then we double-free the 'serial' string property
upon hot unplug since its char* is copied into the virtio-blk device which
has an identical 'serial' qdev property.
This series solves both of these problems as follows:
1. Introduce a qdev child alias property that lets the transport device forward
qdev property accesses into the virtio device (the child).
2. Use qdev child alias properties in transport devices, instead of keeping a
duplicate copy of the VirtIOBlkConf struct.
3. Fix the virtio-blk device refcount leak. It's now safe to do this since the
double-free has been resolved.
Note about the new qdev child alias property type: I haven't made the alias
fully transparent yet. Perhaps we should hide the alias completely?
$ qemu-system-x86_64 -device virtio-blk-pci,\?
...
virtio-blk-pci.logical_block_size=ChildAlias <--- should be uint64 or similar
Stefan Hajnoczi (5):
qdev: add child alias properties
virtio: add child alias properties for virtio-blk
virtio: use child aliases for virtio-blk transport properties
virtio-blk: drop virtio_blk_set_conf()
virtio: fix virtio-blk child refcount in transports
hw/block/virtio-blk.c | 6 ------
hw/core/qdev-properties.c | 28 ++++++++++++++++++++++++++++
hw/s390x/s390-virtio-bus.c | 4 ++--
hw/s390x/s390-virtio-bus.h | 1 -
hw/s390x/virtio-ccw.c | 6 +++---
hw/s390x/virtio-ccw.h | 1 -
hw/virtio/virtio-pci.c | 6 +++---
hw/virtio/virtio-pci.h | 1 -
include/hw/block/block.h | 14 ++++++++++++++
include/hw/qdev-properties.h | 28 ++++++++++++++++++++++++++++
include/hw/virtio/virtio-blk.h | 17 ++++++++++++++++-
11 files changed, 94 insertions(+), 18 deletions(-)