[Qemu-devel] [PATCH RFC 0/3] virtio-scsi unplug of active device

[Eric Farman]

In working with hot-plug/unplug of virtio-scsi devices on s390,
we have occasionally noticed some erratic behavior when an unplug
occurs while I/O is in flight.  Ideally a device is not being used
when it is removed from a guest configuration, but not guarantee
can be made that this will be the case.  And while this scenario
is meant for I/O that occurs during normal use of a device, it
includes the pathological case of an unplug that occurs while the
asynchronous Inquiry loop (initiated by a hotplug) is still ongoing.

Symptoms vary depending on when the unplug is recognized.  Sometimes
a hang occurs, because a reference is not properly released and thus
never reaches zero.  Sometimes a reference is released too early,
allowing the count to go negative and trip an assertion (or more
unpredictable results, if storage is released but still used).

Of course there are many times when things work perfectly, though
that seems to be when the I/O was able to complete in time.  These
patches simply straighten out the completion of I/Os during an
unplug, such that it results in predictable behavior whenever the
device is not idle.

Eric Farman (3):
  scsi/virtio-scsi: Properly flush I/Os inflight during an unplug
  scsi/virtio-scsi: Cleanup of I/Os that never started
  scsi/virtio-scsi: Prevent assertion on missed events

 hw/scsi/virtio-scsi.c |    8 +++++++-
 1 file changed, 7 insertions(+), 1 deletion(-)

-- 
1.7.9.5