On Tuesday 01 Jun 2021 at 13:02:00 (+0100), Marc Zyngier wrote: > On Thu, 27 May 2021 13:51:28 +0100, > Quentin Perret <qperret@xxxxxxxxxx> wrote: > > > > The hyp_page refcount helpers currently rely on the hyp_pool lock for > > serialization. However, this means the refcounts can't be changed from > > the buddy allocator core as it already holds the lock, which means pages > > have to go through odd transient states. > > > > For example, when a page is freed, its refcount is set to 0, and the > > lock is transiently released before the page can be attached to a free > > list in the buddy tree. This is currently harmless as the allocator > > checks the list node of each page to see if it is available for > > allocation or not, but it means the page refcount can't be trusted to > > represent the state of the page even if the pool lock is held. > > > > In order to fix this, remove the pool locking from the refcount helpers, > > and move all the logic to the buddy allocator. This will simplify the > > removal of the list node from struct hyp_page in a later patch. > > Is there any chance some documentation could be added so that we have > a record of what the locking boundaries are? Something along the line > of what we have in arch/arm64/kvm/vgic/vgic.c, for example. Sounds like a good idea, I'll go write something. Cheers, Quentin _______________________________________________ kvmarm mailing list kvmarm@xxxxxxxxxxxxxxxxxxxxx https://lists.cs.columbia.edu/mailman/listinfo/kvmarm
