On Tue, 2020-12-01 at 16:40 -0800, Ankur Arora wrote:
> > How come we get to pin the page and directly dereference it every time,
> > while kvm_setup_pvclock_page() has to use kvm_write_guest_cached()
> > instead?
> 
> So looking at my WIP trees from the time, this is something that
> we went back and forth on as well with using just a pinned page or a
> persistent kvm_vcpu_map().

I have a new plan.

Screw pinning it and mapping it into kernel space just because we need
to do atomic operations on it. 

Futexes can do atomic operations on userspace; so can we. We just add
atomic_cmpxchg_user() and use that.

We can add kvm_cmpxchg_guest_offset_cached() and use that from places
like record_steal_time().

That works nicely for all of the Xen support needs too, I think — since
test-and-set and test-and-clear bit operations all want to be built on
cmpxchg.

The one thing I can think off offhand that it doesn't address easily is
the testing of vcpu_info->evtchn_upcall_pending in
kvm_xen_cpu_has_interrupt(), which really does want to be fast so I'm
not sure I want to be using copy_from_user() for that.


But I think I can open-code that to do the cache checking that
kvm_read_guest_offset_cached() does, and use __get_user() directly in
the fast path, falling back to actually calling
kvm_read_guest_offset_cached() when it needs to. That will suffice
until/unless we grow more use cases which would make it worth providing
that as a kvm_get_guest...() set of functions for generic use.