On Wed, 2012-06-20 at 13:14 +1000, David Gibson wrote: > So, in fact the original comment is a bit out of date. With the > current version of this series, then a guest attempt to invalidate > will be delayed until the unmap occurs.
No, this code was dropped, including the tracking of the maps, following comments from Anthony and others. The API for providing a cancel callback is still there but nothing will call it unless the backend does its own tracking and decides to do so. As it is, the race exist but: - It will only hurt the guest - And only for a very buggy guest So the worst case is that it hurts something like kdump. I plan to re-introduce some of the mechanisms for cancellation eventually, but we agreed that it wasn't going to be a show stopper and that we could work on getting that sorted in a second phase. I'm looking at a more efficient way to deal with the tracking of the maps as well since some devices uses them often. Cheers, Ben.