On Tue, May 23, 2023, Kautuk Consul wrote:
On 2022-07-06 16:20:10, Chao Peng wrote:
diff --git a/include/linux/kvm_host.h b/include/linux/kvm_host.h index e9153b54e2a4..c262ebb168a7 100644 --- a/include/linux/kvm_host.h +++ b/include/linux/kvm_host.h @@ -765,10 +765,10 @@ struct kvm { #if defined(CONFIG_MMU_NOTIFIER) && defined(KVM_ARCH_WANT_MMU_NOTIFIER) struct mmu_notifier mmu_notifier;
- unsigned long mmu_notifier_seq;
- long mmu_notifier_count;
- gfn_t mmu_notifier_range_start;
- gfn_t mmu_notifier_range_end;
- unsigned long mmu_updating_seq;
- long mmu_updating_count;
Can we convert mmu_updating_seq and mmu_updating_count to atomic_t ?
Heh, can we? Yes. Should we? No.
I see that not all accesses to these are under the kvm->mmu_lock spinlock.
Ya, working as intended. Ignoring gfn_to_pfn_cache for the moment, all accesses to mmu_invalidate_in_progress (was mmu_notifier_count / mmu_updating_count above) are done under mmu_lock. And for for mmu_notifier_seq (mmu_updating_seq above), all writes and some reads are done under mmu_lock. The only reads that are done outside of mmu_lock are the initial snapshots of the sequence number.
gfn_to_pfn_cache uses a different locking scheme, the comments in mmu_notifier_retry_cache() do a good job explaining the ordering.
This will also remove the need for putting separate smp_wmb() and smp_rmb() memory barriers while accessing these structure members.
No, the memory barriers aren't there to provide any kind of atomicity. The barriers exist to ensure that stores and loads to/from the sequence and invalidate in-progress counts are ordered relative to the invalidation (stores to counts) and creation (loads) of SPTEs. Making the counts atomic changes nothing because atomic operations don't guarantee the necessary ordering.
E.g. when handling a page fault, KVM snapshots the sequence outside of mmu_lock _before_ touching any state that is involved in resolving the host pfn, e.g. primary MMU state (VMAs, host page tables, etc.). After the page fault task acquires mmu_lock, KVM checks that there are no in-progress invalidations and that the sequence count is the same. This ensures that if there is a concurrent page fault and invalidation event, the page fault task will either acquire mmu_lock and create SPTEs _before_ the invalidation is processed, or the page fault task will observe either an elevated mmu_invalidate_in_progress or a different sequence count, and thus retry the page fault, if the page fault task acquires mmu_lock after the invalidation event.