On 31/03/21 18:20, Sean Christopherson wrote:
Every call
to .change_pte() is bookended by .invalidate_range_{start,end}(), i.e. the above
missing kvm->mmu_notifier_seq++ is benign because kvm->mmu_notifier_count is
guaranteed to be non-zero.
In fact ARM even relies on invalidate wrapping the change_pte handler.
/*
* The MMU notifiers will have unmapped a huge PMD before calling
* ->change_pte() (which in turn calls kvm_set_spte_hva()) and
* therefore we never need to clear out a huge PMD through this
* calling path and a memcache is not required.
*/
Assuming all of the above is correct, I'm very tempted to rip out .change_pte()
entirely.
There is still the performance benefit from immediately remapping the
page to the new destination without waiting for a fault. Yes it's
hypothetical but I would prefer to leave that change for later.
The fact that the count is nonzero means that you will not even have to
complicate kvm_mmu_notifier_change_pte to handle the removal of
mmu_notifier_seq; just add a patch before this one to WARN if it is
zero. (The rest of my review to patch 16 still holds).
Paolo
It's been dead weight for 8+ years and no one has complained about
KSM+KVM performance (I'd also be curious to know how much performance was gained
by shaving VM-Exits). As KVM is the only user of .change_pte(), dropping it in
KVM would mean the entire MMU notifier could also go away.
