Simplify (hopefully it's simpler) zapping collapsible SPTEs by first
simplifying retrieving the host mapping level. KVM currently requires
memory be backed by a refcounted struct page in order to be mapped as
a huge page. This requires KVM to acquire the pfn that corresponds to
the gfn/hva before checking whether or not the gfn/hva can be mapped
huge.
Dropping that requirement allow the "zap collapsible" path to detect
that a shadow page can be zapped without having to first bottom out on
leaf entries. This could theoretically be a minor performance win,
e.g. then KVM doesn't need to walk all not-present leaf SPTEs to find
out that a shadow page has no children. In basic testing I didn't see
any meaningful difference (the actual zapping dominates).
There are also potential use cases for allow any mappings to be huge,
e.g. GPU buffers (IIUC). Dropping the struct page requirement makes
KVM play nice with those.
This is most definitely post-5.20 material.
Sean Christopherson (4):
KVM: x86/mmu: Don't require refcounted "struct page" to create huge
SPTEs
KVM: x86/mmu: Document the "rules" for using host_pfn_mapping_level()
KVM: x86/mmu: Don't bottom out on leafs when zapping collapsible SPTEs
KVM: selftests: Add an option to run vCPUs while disabling dirty
logging
arch/x86/kvm/mmu/mmu.c | 65 ++++++++++++-------
arch/x86/kvm/mmu/mmu_internal.h | 2 +-
arch/x86/kvm/mmu/tdp_iter.c | 9 ---
arch/x86/kvm/mmu/tdp_iter.h | 1 -
arch/x86/kvm/mmu/tdp_mmu.c | 61 ++++++++---------
.../selftests/kvm/dirty_log_perf_test.c | 30 ++++++++-
6 files changed, 94 insertions(+), 74 deletions(-)
base-commit: 8031d87aa9953ddeb047a5356ebd0b240c30f233
--
2.37.0.170.g444d1eabd0-goog
