On 26/09/2015 08:54, David Gibson wrote: > On Fri, Sep 25, 2015 at 02:04:14PM +0200, Paolo Bonzini wrote: >> -----BEGIN PGP SIGNED MESSAGE----- Hash: SHA256 >> >> >> >> On 25/09/2015 13:33, David Gibson wrote: >>> 1) Is there a case where using the no-replay functions makes >>> sense? >>> >>> I'm not sure. I think vfio is the only user so far, so I >>> guess that's technically a no. I was reluctant to change the >>> interface and semantics just off the bat, though. >> >> Considering memory_region_listener does the reply, I think it's >> okay. > > Uh.. just to be clear, are you saying I should change this so > there's only the replaying interface?
Maybe... The only issue is the "granularity" argument, which is not in memory_region_register_iommu_notifier. That makes me wonder if the replay and registration make sense as separate operations. What about adding a new function memory_region_iommu_replay and separate the two phases? >> For solving the problem that Laurent mentioned, using int128 >> seems like the easiest solution... > > Maybe. It means I have to do all the address calculation in the > loop with an int128, then truncate it to do the actual call. That > seems harder to me than the overflow check I added, but I suppose > it's conceptually similar in some ways. Your overflow check is also okay, I wrote this before seeing the updated version. Paolo
