On Sun, Mar 10, 2024 at 9:46 PM Oliver Upton <oliver.upton@xxxxxxxxx> wrote: > > > > > > 2. What is your best guess as to when KVM userfault patches will be available, > > > even if only in RFC form? > > > > We're aiming for the end of April for RFC with KVM/ARM support. > > Just to make sure everyone is read in on what this entails -- is this > the implementation that only worries about vCPUs touching non-present > memory, leaving the question of other UAPIs that consume guest memory > (e.g. GIC/ITS table save/restore) up for further discussion? Yes. The initial version will only support returning to userspace on invalid vCPU accesses with KVM_EXIT_MEMORY_FAULT. Non-vCPU accesses to invalid pages (e.g. GIC/ITS table save/restore) will trigger an error return from __gfn_to_hva_many() (which will cause the corresponding ioctl to fail). It will be userspace's responsibility to clear the invalid attribute before invoking those ioctls. For x86 we may need an blocking kernel-to-userspace notification mechanism for code paths in the emulator, but we'd like to investigate and discuss if there are any other cleaner alternatives before going too far down that route.
