Re: [Qemu-block] [Qemu-devel] [PATCH v2 0/6] Acquire the AioContext during _realize()

[Kevin Wolf]

Am 18.01.2019 um 10:56 hat Stefan Hajnoczi geschrieben:
> On Thu, Jan 17, 2019 at 02:24:10PM +0100, Alberto Garcia wrote:
> > On Wed 16 Jan 2019 02:54:44 PM CET, Stefan Hajnoczi wrote:
> > > This patch series makes virtio-blk and virtio-scsi more robust,
> > > although I haven't checked what happens if the drive is attached to a
> > > different IOThread than the device - will the switchover work?
> > 
> > You mean if you use x-blockdev-set-iothread with a drive and then attach
> > it to a device using a different thread? That seems to work (with my
> > patch).
> 
> Does it work with a running VM that will do I/O?  If yes, then that's
> great and we should merge the virtio-blk/scsi patches to make them
> robust for this case.  I'm a little surprised it works though :-)
> because the code wasn't designed or tested for this.

There are two ways to trigger the crash even without
x-blockdev-set-iothread:

* device_del, then device_add for a device with iothread (virtio-scsi;
  may or may not exist with virtio-blk)
  https://bugzilla.redhat.com/show_bug.cgi?id=1656276

* Simply attach two devices with iothread to the the same node
  https://bugzilla.redhat.com/show_bug.cgi?id=1662508

Maybe this series is actually papering over the real problems, though.
Unplug should move the BDS back to the main AioContext, and attaching
two users with different AioContexts must fail. If we don't take care
there, the locking will be wrong.

Note that even though there are more problems, we still need this series
so we can allow attaching two devices to the same node while using the
same iothread.

Kevin

signature.asc
Description: PGP signature