On Wed, 20 Mar 2013, Paolo Bonzini wrote: > Il 20/03/2013 09:33, Alex Bligh ha scritto: > > Stefano, > > > > --On 19 March 2013 15:13:29 +0000 Stefano Stabellini > > <stefano.stabell...@eu.citrix.com> wrote: > > > >> Therefore I think that the current change is not safe, but it is pretty > >> easy to make it safe. > >> You just need to move the call to blk_open from blk_init to blk_connect. > >> Actually you can move most of blk_init to blk_connect, you just need to > >> keep the 4 xenstore_write_be_int at the end of the function. > > > > The final of these 4 writes in blk_init is: > > > > xenstore_write_be_int(&blkdev->xendev, "sectors", > > blkdev->file_size / blkdev->file_blk); > > > > and blkdev->filesize comes from bdrv_getlength(blkdev->bs), and that > > requires (I think) bdrv_open to have been called. Can these > > 4 writes move safely to blk_connect? Or can we write the sectors > > count as (say) 0 here and fix it later in blk_connect? The remainder > > look ok. > > I think so. blkfront reads "sectors" when QEMU moves to > XenbusStateConnected, in blkfront_connect. > > blk_connect is called from xen_be_try_initialise, which moves to > XenbusStateConnected on success. So, QEMU's blk_connect will always be > called before blkfront's blkfront_connect.
Alex, do you have any updates on this patch?