On Wed, Jan 15, 2014 at 11:01:59AM +0100, Paolo Bonzini wrote:
Il 15/01/2014 02:48, Laszlo Ersek ha scritto:
When init_virtio_scsi() finds no SCSI targets connected to the HBA, it frees the virtio ring. Other code in SeaBIOS proceeds to overwrite the area. However, the ring is in use by qemu at that point -- not only did we report the (ACK|DRIVER|DRIVER_OK) status earlier, we even communicated over the ring.
Of course SeaBIOS doesn't "kick" the HBA ever again, hence qemu has no reason to look at the ring. However, when qemu uses KVM acceleration, and ioeventfd is enabled for the HBA, then a vmstate change to "running" (including stop->cont monitor commands and incoming migration) "forces" a kick (see qemu commit 25db9ebe). Qemu then tries to interpret whatever unrelated guest data is in the HBA's original ring area, as virtio protocol. Qemu exits upon seeing the garbage.
init_virtio_scsi() should reset the HBA before allowing the virtio ring memory to be reused. Device reset causes the hypervisor to drop its references.
This change is justified / underpinned by pure virtio-spec compliance as well.
Related RHBZ: https://bugzilla.redhat.com/show_bug.cgi?id=1013418
Signed-off-by: Laszlo Ersek email@example.com
Reviewed-by: Paolo Bonzini firstname.lastname@example.org
Thanks Laszlo - I'm sure that wasn't easy to find. I've pushed the patch.