Re: [PATCH v4 4/6] KVM: MMU: fast invalid all shadow pages

From: Gleb Natapov
Date: Tue May 07 2013 - 06:01:15 EST


On Tue, May 07, 2013 at 05:41:35PM +0800, Xiao Guangrong wrote:
> On 05/07/2013 04:58 PM, Gleb Natapov wrote:
> > On Tue, May 07, 2013 at 01:45:52AM +0800, Xiao Guangrong wrote:
> >> On 05/07/2013 01:24 AM, Gleb Natapov wrote:
> >>> On Mon, May 06, 2013 at 09:10:11PM +0800, Xiao Guangrong wrote:
> >>>> On 05/06/2013 08:36 PM, Gleb Natapov wrote:
> >>>>
> >>>>>>> Step 1) Fix kvm_mmu_zap_all's behaviour: introduce lockbreak via
> >>>>>>> spin_needbreak. Use generation numbers so that in case kvm_mmu_zap_all
> >>>>>>> releases mmu_lock and reacquires it again, only shadow pages
> >>>>>>> from the generation with which kvm_mmu_zap_all started are zapped (this
> >>>>>>> guarantees forward progress and eventual termination).
> >>>>>>>
> >>>>>>> kvm_mmu_zap_generation()
> >>>>>>> spin_lock(mmu_lock)
> >>>>>>> int generation = kvm->arch.mmu_generation;
> >>>>>>>
> >>>>>>> for_each_shadow_page(sp) {
> >>>>>>> if (sp->generation == kvm->arch.mmu_generation)
> >>>>>>> zap_page(sp)
> >>>>>>> if (spin_needbreak(mmu_lock)) {
> >>>>>>> kvm->arch.mmu_generation++;
> >>>>>>> cond_resched_lock(mmu_lock);
> >>>>>>> }
> >>>>>>> }
> >>>>>>>
> >>>>>>> kvm_mmu_zap_all()
> >>>>>>> spin_lock(mmu_lock)
> >>>>>>> for_each_shadow_page(sp) {
> >>>>>>> if (spin_needbreak(mmu_lock)) {
> >>>>>>> cond_resched_lock(mmu_lock);
> >>>>>>> }
> >>>>>>> }
> >>>>>>>
> >>>>>>> Use kvm_mmu_zap_generation for kvm_arch_flush_shadow_memslot.
> >>>>>>> Use kvm_mmu_zap_all for kvm_mmu_notifier_release,kvm_destroy_vm.
> >>>>>>>
> >>>>>>> This addresses the main problem: excessively long hold times
> >>>>>>> of kvm_mmu_zap_all with very large guests.
> >>>>>>>
> >>>>>>> Do you see any problem with this logic? This was what i was thinking
> >>>>>>> we agreed.
> >>>>>>
> >>>>>> No. I understand it and it can work.
> >>>>>>
> >>>>>> Actually, it is similar with Gleb's idea that "zapping stale shadow pages
> >>>>>> (and uses lock break technique)", after some discussion, we thought "only zap
> >>>>>> shadow pages that are reachable from the slot's rmap" is better, that is this
> >>>>>> patchset does.
> >>>>>> (https://lkml.org/lkml/2013/4/23/73)
> >>>>>>
> >>>>> But this is not what the patch is doing. Close, but not the same :)
> >>>>
> >>>> Okay. :)
> >>>>
> >>>>> Instead of zapping shadow pages reachable from slot's rmap the patch
> >>>>> does kvm_unmap_rmapp() which drop all spte without zapping shadow pages.
> >>>>> That is why you need special code to re-init lpage_info. What I proposed
> >>>>> was to call zap_page() on all shadow pages reachable from rmap. This
> >>>>> will take care of lpage_info counters. Does this make sense?
> >>>>
> >>>> Unfortunately, no! We still need to care lpage_info. lpage_info is used
> >>>> to count the number of guest page tables in the memslot.
> >>>>
> >>>> For example, there is a memslot:
> >>>> memslot[0].based_gfn = 0, memslot[0].npages = 100,
> >>>>
> >>>> and there is a shadow page:
> >>>> sp->role.direct =0, sp->role.level = 4, sp->gfn = 10.
> >>>>
> >>>> this sp is counted in the memslot[0] but it can not be found by walking
> >>>> memslot[0]->rmap since there is no last mapping in this shadow page.
> >>>>
> >>> Right, so what about walking mmu_page_hash for each gfn belonging to the
> >>> slot that is in process to be removed to find those?
> >>
> >> That will cost lots of time. The size of hashtable is 1 << 10. If the
> >> memslot has 4M memory, it will walk all the entries, the cost is the same
> >> as walking active_list (maybe litter more). And a memslot has 4M memory is
> >> the normal case i think.
> >>
> > Memslots will be much bigger with memory hotplug. Lock break should be
> > used while walking mmu_page_hash obviously, but still iterating over
> > entire memslot gfn space to find a few gfn that may be there is
> > suboptimal. We can keep a list of them in the memslot itself.
>
> It sounds good to me.
>
> BTW, this approach looks more complex and use more memory (new list_head
> added into every shadow page) used, why you dislike clearing lpage_info? ;)
>
Looks a little bit hackish, but now that I see we do not have easy way
to find all shadow pages counted in lpage_info I am not entirely against
it. If you convince Marcelo that clearing lpage_info like that is a good
idea I may reconsider. I think, regardless of tracking lpage_info,
having a way to find all shadow pages that reference a memslot is a good
thing though.

> >
> >> Another point is that lpage_info stops mmu to use large page. If we
> >> do not reset lpage_info, mmu is using 4K page until the invalid-sp is
> >> zapped.
> >>
> > I do not think this is a big issue. If lpage_info prevented the use of
> > large pages for some memory ranges before we zapped entire shadow pages
> > it was probably for a reason, so new shadow page will prevent large
> > pages from been created for the same memory ranges.
>
> Still worried, but I will try it if Marcelo does not have objects.
> Thanks a lot for your valuable suggestion, Gleb!
>
> Now, i am trying my best to catch Marcelo's idea of "zapping root
> pages", but......
>
Yes, I am missing what Marcelo means there too. We cannot free memslot
until we unmap its rmap one way or the other.

--
Gleb.
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/