Re: [RFC PATCH 09/17] KVM: arm64: Tear down unlinked page tables in parallel walk

[Oliver Upton <oupton@google.com>]

On Tue, May 03, 2022 at 02:17:25PM +0000, Quentin Perret wrote:
> On Friday 22 Apr 2022 at 20:41:47 (+0000), Oliver Upton wrote:
> > On Fri, Apr 22, 2022 at 04:00:45PM +0000, Quentin Perret wrote:
> > > On Thursday 21 Apr 2022 at 16:40:56 (+0000), Oliver Upton wrote:
> > > > The other option would be to not touch the subtree at all until the rcu
> > > > callback, as at that point software will not tweak the tables any more.
> > > > No need for atomics/spinning and can just do a boring traversal.
> > > 
> > > Right that is sort of what I had in mind. Note that I'm still trying to
> > > make my mind about the overall approach -- I can see how RCU protection
> > > provides a rather elegant solution to this problem, but this makes the
> > > whole thing inaccessible to e.g. pKVM where RCU is a non-starter.
> > 
> > Heh, figuring out how to do this for pKVM seemed hard hence my lazy
> > attempt :)
> > 
> > > A
> > > possible alternative that comes to mind would be to have all walkers
> > > take references on the pages as they walk down, and release them on
> > > their way back, but I'm still not sure how to make this race-safe. I'll
> > > have a think ...
> > 
> > Does pKVM ever collapse tables into blocks? That is the only reason any
> > of this mess ever gets roped in. If not I think it is possible to get
> > away with a rwlock with unmap on the write side and everything else on
> > the read side, right?
> > 
> > As far as regular KVM goes we get in this business when disabling dirty
> > logging on a memslot. Guest faults will lazily collapse the tables back
> > into blocks. An equally valid implementation would be just to unmap the
> > whole memslot and have the guest build out the tables again, which could
> > work with the aforementioned rwlock.
> 
> Apologies for the delay on this one, I was away for a while.
> 
> Yup, that all makes sense. FWIW the pKVM use-case I have in mind is
> slightly different. Specifically, in the pKVM world the hypervisor
> maintains a stage-2 for the host, that is all identity mapped. So we use
> nice big block mappings as much as we can. But when a protected guest
> starts, the hypervisor needs to break down the host stage-2 blocks to
> unmap the 4K guest pages from the host (which is where the protection
> comes from in pKVM). And when the guest is torn down, the host can
> reclaim its pages, hence putting us in a position to coallesce its
> stage-2 into nice big blocks again. Note that none of this coallescing
> is currently implemented even in our pKVM prototype, so it's a bit
> unfair to ask you to deal with this stuff now, but clearly it'd be cool
> if there was a way we could make these things coexist and even ideally
> share some code...

Oh, it certainly isn't unfair to make sure we've got good constructs
landing for everyone to use :-)

I'll need to chew on this a bit more to have a better answer. The reason
I hesitate to do the giant unmap for non-pKVM is that I believe we'd be
leaving some performance on the table for newer implementations of the
architecture. Having said that, avoiding a tlbi vmalls12e1is on every
collapsed table is highly desirable.

FEAT_BBM=2 semantics in the MMU is also on the todo list. In this case
we'd do a direct table->block transformation on the PTE and elide that
nasty tlbi.

Unless there's objections, I'll probably hobble this series along as-is
for the time being. My hope is that other table walkers can join in on
the parallel party later down the road.

Thanks for getting back to me.

--
Best,
Oliver