Re: [PATCH v2] KVM: x86: Preserve TDP MMU roots until they are explicitly invalidated

[David Matlack]

On Fri, Apr 21, 2023 at 6:56 PM Sean Christopherson <seanjc@xxxxxxxxxx> wrote:
>
> On Fri, Apr 21, 2023, David Matlack wrote:
> > On Fri, Apr 21, 2023 at 2:49 PM Sean Christopherson <seanjc@xxxxxxxxxx> wrote:
> > >
> > >  void kvm_tdp_mmu_invalidate_all_roots(struct kvm *kvm)
> > >  {
> > >         struct kvm_mmu_page *root;
> > >
> > > -       lockdep_assert_held_write(&kvm->mmu_lock);
> > > -       list_for_each_entry(root, &kvm->arch.tdp_mmu_roots, link) {
> > > -               if (!root->role.invalid &&
> > > -                   !WARN_ON_ONCE(!kvm_tdp_mmu_get_root(root))) {
> > > +       /*
> > > +        * Note!  mmu_lock isn't held when destroying the VM!  There can't be
> > > +        * other references to @kvm, i.e. nothing else can invalidate roots,
> > > +        * but walking the list of roots does need to be guarded against roots
> > > +        * being deleted by the asynchronous zap worker.
> > > +        */
> > > +       rcu_read_lock();
> > > +
> > > +       list_for_each_entry_rcu(root, &kvm->arch.tdp_mmu_roots, link) {
> >
> > I see that roots are removed from the list with list_del_rcu(), so I
> > agree this should be safe.
> >
> > KVM could, alternatively, acquire the mmu_lock in
> > kvm_mmu_uninit_tdp_mmu(), which would let us keep the lockdep
> > assertion and drop the rcu_read_lock() + comment. That might be worth
> > it in case someone accidentally adds a call to
> > kvm_tdp_mmu_invalidate_all_roots() without mmu_lock outside of VM
> > teardown. kvm_mmu_uninit_tdp_mmu() is not a particularly performance
> > sensitive path and adding the mmu_lock wouldn't add much overhead
> > anyway (it would block for at most a few milliseconds waiting for the
> > async work to reschedule).
>
> Heh, I actually started to ping you off-list to specifically discuss this option,
> but then decided that not waiting those few milliseconds might be worthwhile for
> some use cases.  I also couldn't quite convince myself that it would only be a few
> milliseconds, e.g. if the worker is zapping a fully populated 5-level root, there
> are no other tasks scheduled on _its_ CPU, and CONFIG_PREEMPTION=n (which neuters
> rwlock_needbreak()).

Good point. At some point we're going to have to fix that.

>
> The other reason I opted for not taking mmu_lock is that, with the persistent roots
> approach, I don't think it's actually strictly necessary for kvm_mmu_zap_all_fast()
> to invaliate roots while holding mmu_lock for write.  Holding slots_lock ensures
> that only a single task can be doing invalidations, versus the old model where
> putting the last reference to a root could happen just about anywhere.  And
> allocating a new root and zapping from mmu_noitifiers requires holding mmu_lock for
> write, so I _think_ we could getaway with holding mmu_lock for read.  Maybe.
>
> It's largely a moot point since kvm_mmu_zap_all_fast() needs to hold mmu_lock for
> write anyways to play nice with the shadow MMU, i.e. I don't expect us to ever
> want to pursue a change in this area.  But at the same time I was struggling to
> write a comment explaining why the VM destruction path "had" to take mmu_lock.

Yeah, probably because it really isn't necessary :). It'd be nice to keep
around the lockdep assertion though for the other (and future)
callers. The cleanest options I can think of are:

1. Pass in a bool "vm_teardown" kvm_tdp_mmu_invalidate_all_roots() and
use that to gate the lockdep assertion.
2. Take the mmu_lock for read in kvm_mmu_uninit_tdp_mmu() and pass
down bool shared to kvm_tdp_mmu_invalidate_all_roots().

Both would satisfy your concern of not blocking teardown on the async
worker and my concern of keeping the lockdep check. I think I prefer
(1) since, as you point out, taking the mmu_lock at all is
unnecessary.