v5:
 * Plumbing vm change state handlers into DeviceClass/BusClass is a rather
   large bug fix.  Instead I've combined the previous priorities approach with
   the observation from Kevin and Paolo that we really want to order by qdev
   tree depth.

   The new qdev_add_vm_change_state_handler() API lets DeviceStates register
   callbacks that execute in qdev tree depth order.  This solves the
   virtio-scsi bug since the virtio-scsi device's callback must complete before
   its child scsi-disk's callback runs.

   Is this a good compromise for everyone?

Stefan Hajnoczi (3):
  vl: add qemu_add_vm_change_state_handler_prio()
  qdev: add qdev_add_vm_change_state_handler()
  virtio-scsi: restart DMA after iothread

 hw/core/Makefile.objs             |  1 +
 include/hw/qdev-core.h            |  5 +++
 include/sysemu/sysemu.h           |  2 +
 hw/core/vm-change-state-handler.c | 61 +++++++++++++++++++++++++++++++
 hw/scsi/scsi-bus.c                |  4 +-
 hw/virtio/virtio.c                |  4 +-
 vl.c                              | 59 ++++++++++++++++++++++++------
 7 files changed, 120 insertions(+), 16 deletions(-)
 create mode 100644 hw/core/vm-change-state-handler.c

-- 
2.21.0

On 20/06/19 19:37, Stefan Hajnoczi wrote:
> v5:
>  * Plumbing vm change state handlers into DeviceClass/BusClass is a rather
>    large bug fix.  Instead I've combined the previous priorities approach with
>    the observation from Kevin and Paolo that we really want to order by qdev
>    tree depth.
> 
>    The new qdev_add_vm_change_state_handler() API lets DeviceStates register
>    callbacks that execute in qdev tree depth order.  This solves the
>    virtio-scsi bug since the virtio-scsi device's callback must complete before
>    its child scsi-disk's callback runs.
> 
>    Is this a good compromise for everyone?

Yes!  Perhaps a bit of a hack, but it works and both the API and the
implementation are very sane.  Converting other devices to use
qdev_add_vm_change_state_handler() is left as an exercise for the
reviewer, I guess? :)

Paolo

Am 20.06.2019 um 19:37 hat Stefan Hajnoczi geschrieben:
> v5:
>  * Plumbing vm change state handlers into DeviceClass/BusClass is a rather
>    large bug fix.  Instead I've combined the previous priorities approach with
>    the observation from Kevin and Paolo that we really want to order by qdev
>    tree depth.
> 
>    The new qdev_add_vm_change_state_handler() API lets DeviceStates register
>    callbacks that execute in qdev tree depth order.  This solves the
>    virtio-scsi bug since the virtio-scsi device's callback must complete before
>    its child scsi-disk's callback runs.
> 
>    Is this a good compromise for everyone?

I'd still call it a hack, but I can also see that doing the real thing
would be a lot of work that might not be worth the effort.

Thanks, applied to the block branch.

Kevin