On Fri, Apr 21, 2017 at 10:51:00AM +0200, Juergen Gross wrote: > Today disconnecting xen-blkback is broken in case there are still > I/Os in flight: xen_blkif_disconnect() will bail out early without > releasing all resources in the hope it will be called again when > the last request has terminated. This, however, won't happen as > xen_blkif_free() won't be called on termination of the last running > request: xen_blkif_put() won't decrement the blkif refcnt to 0 as > xen_blkif_disconnect() didn't finish before thus some xen_blkif_put() > calls in xen_blkif_disconnect() didn't happen. > > To solve this deadlock xen_blkif_disconnect() and > xen_blkif_alloc_rings() shouldn't use xen_blkif_put() and > xen_blkif_get() but use some other way to do their accounting of > resources. > > This at once fixes another error in xen_blkif_disconnect(): when it > returned early with -EBUSY for another ring than 0 it would call > xen_blkif_put() again for already handled rings on a subsequent call. > This will lead to inconsistencies in the refcnt handling. > > Cc: stable@xxxxxxxxxxxxxxx > Signed-off-by: Juergen Gross <jgross@xxxxxxxx> > Tested-by: Steven Haigh <netwiz@xxxxxxxxx> Acked-by: Roger Pau Monné <roger.pau@xxxxxxxxxx> TBH, it seems like all this accounting in blkback could be improved quite a lot. Right now this is just a complete caos. Thanks, Roger.