Re: [PATCH v2 9/9] KVM: X86: Optimize zapping rmap

[Peter Xu <peterx@redhat.com>]

On Wed, Jul 28, 2021 at 09:39:02PM +0000, Sean Christopherson wrote:
> On Fri, Jun 25, 2021, Peter Xu wrote:
> > Using rmap_get_first() and rmap_remove() for zapping a huge rmap list could be
> > slow.  The easy way is to travers the rmap list, collecting the a/d bits and
> > free the slots along the way.
> > 
> > Provide a pte_list_destroy() and do exactly that.
> > 
> > Signed-off-by: Peter Xu <peterx@redhat.com>
> > ---
> >  arch/x86/kvm/mmu/mmu.c | 45 +++++++++++++++++++++++++++++++-----------
> >  1 file changed, 33 insertions(+), 12 deletions(-)
> > 
> > diff --git a/arch/x86/kvm/mmu/mmu.c b/arch/x86/kvm/mmu/mmu.c
> > index ba0258bdebc4..45aac78dcabc 100644
> > --- a/arch/x86/kvm/mmu/mmu.c
> > +++ b/arch/x86/kvm/mmu/mmu.c
> > @@ -1014,6 +1014,38 @@ unsigned int pte_list_count(struct kvm_rmap_head *rmap_head)
> >  	return count;
> >  }
> >  
> > +/* Return true if rmap existed and callback called, false otherwise */
> > +static bool pte_list_destroy(struct kvm_rmap_head *rmap_head,
> > +			     int (*callback)(u64 *sptep))
> > +{
> > +	struct pte_list_desc *desc, *next;
> > +	int i;
> > +
> > +	if (!rmap_head->val)
> > +		return false;
> > +
> > +	if (!(rmap_head->val & 1)) {
> > +		if (callback)
> > +			callback((u64 *)rmap_head->val);
> > +		goto out;
> > +	}
> > +
> > +	desc = (struct pte_list_desc *)(rmap_head->val & ~1ul);
> > +
> > +	while (desc) {
> > +		if (callback)
> > +			for (i = 0; i < desc->spte_count; i++)
> > +				callback(desc->sptes[i]);
> > +		next = desc->more;
> > +		mmu_free_pte_list_desc(desc);
> > +		desc = next;
> 
> Alternatively, 
> 
> 	desc = (struct pte_list_desc *)(rmap_head->val & ~1ul);
> 	for ( ; desc; desc = next) {
> 		for (i = 0; i < desc->spte_count; i++)
> 			mmu_spte_clear_track_bits((u64 *)rmap_head->val);
> 		next = desc->more;
> 		mmu_free_pte_list_desc(desc);
> 	}
> 
> > +	}
> > +out:
> > +	/* rmap_head is meaningless now, remember to reset it */
> > +	rmap_head->val = 0;
> > +	return true;
> 
> Why implement this as a generic method with a callback?  gcc is suprisingly
> astute in optimizing callback(), but I don't see the point of adding a complex
> helper that has a single caller, and is extremely unlikely to gain new callers.
> Or is there another "zap everything" case I'm missing?

No other case; it's just that pte_list_*() helpers will be more self-contained.
If that'll be a performance concern, no objection to hard code it.

> 
> E.g. why not this?
> 
> static bool kvm_zap_rmapp(struct kvm *kvm, struct kvm_rmap_head *rmap_head,
> 			  const struct kvm_memory_slot *slot)
> {
> 	struct pte_list_desc *desc, *next;
> 	int i;
> 
> 	if (!rmap_head->val)
> 		return false;
> 
> 	if (!(rmap_head->val & 1)) {
> 		mmu_spte_clear_track_bits((u64 *)rmap_head->val);
> 		goto out;
> 	}
> 
> 	desc = (struct pte_list_desc *)(rmap_head->val & ~1ul);
> 	for ( ; desc; desc = next) {
> 		for (i = 0; i < desc->spte_count; i++)
> 			mmu_spte_clear_track_bits(desc->sptes[i]);
> 		next = desc->more;
> 		mmu_free_pte_list_desc(desc);
> 	}
> out:
> 	/* rmap_head is meaningless now, remember to reset it */
> 	rmap_head->val = 0;
> 	return true;
> }

Looks good, but so far I've no strong opinion on this.  I'll leave it for Paolo
to decide.

Thanks!

-- 
Peter Xu