On Thu, Nov 03, 2016 at 12:02:28PM +1100, Paul Mackerras wrote: > On Wed, Nov 02, 2016 at 01:44:03PM +1100, Alexey Kardashevskiy wrote: > > On 31/10/16 15:23, David Gibson wrote: > [...] > > > > > > Um.. yeah.. that's not really ok. Prohibiting overlapping > > > registrations on the same container is reasonable enough. Having a > > > container not be able to register memory because some completely > > > different container has registered something overlapping is getting > > > very ugly. > > > > I am lost here. Does this mean the patches cannot go upstream? > > > > Also how would I implement overlapping if we are not teaching KVM about > > VFIO containers? The mm list has a counter of how many times each memory > > region was mapped via TCE (and this prevents unregistration), and if we > > want overlapping regions - a "mapped" counter of which one would I update > > in real mode (where I only have a user address and a LIOBN)? > > The patches fix a real bug, where we run out of memory to run VMs. > > The patches don't change the interface, and don't introduce the > constraint that is being discussed here (that the regions being > registered may not overlap unless they are identical to a previously > registered region). That constraint is already present in the > upstream code.
Ah, good point. I hadn't thought that through and realized that
limitation was already through.
> They do change the behaviour when you use a container fd from a
> different process from the one which opened the fd, but that is not
> something that worked in any meaningful way before anyway.
>
> So David, do you still see any reason why the patches should not be
> accepted?
Given the (still hideously ugly) limitation is not new, then no, not
any more. I'll send an R-b.
--
David Gibson | I'll have my music baroque, and my code
david AT gibson.dropbear.id.au | minimalist, thank you. NOT _the_ _other_
| _way_ _around_!
http://www.ozlabs.org/~dgibson
signature.asc
Description: PGP signature
