KVM: pfncache: Implement range-based invalidation check for hva_to_pfn_retry()

KVM: pfncache: Implement range-based invalidation check for hva_to_pfn_retry()

The existing retry loop in hva_to_pfn_retry() is extremely pessimistic.
If there are any concurrent invalidations running, it's effectively just
a complex busy wait loop because its local mmu_notifier_retry_cache()
function will always return true.

Since multiple invalidations can be running in parallel, this can result
in a situation where hva_to_pfn_retry() just backs off and keep retrying
for ever, not making any progress.

Solve this by being a bit more selective about when to retry. In
addition to the ->needs_invalidation flag added in a previous commit,
check before calling hva_to_pfn() if there is any pending invalidation
(i.e. between invalidate_range_start() and invalidate_range_end()) which
affects the uHVA of the GPC being updated. This catches the case where
hva_to_pfn() would return a soon-to-be-stale mapping of a range for which
the invalidate_range_start() hook had already been called before the uHVA
was even set in the GPC and the ->needs_invalidation flag set.

An invalidation which *starts* after the lock is dropped in the loop is
fine because it will clear the ->needs_revalidation flag and also trigger
a retry.

This is slightly more complex than it needs to be; moving the
invalidation to occur in the invalidate_range_end() hook will make life
simpler, in a subsequent commit.

Signed-off-by: David Woodhouse <dwmw@amazon.co.uk>