Re: [PATCH v2 1/7] KVM: x86/mmu: Track if shadow MMU active

[Paolo Bonzini]

On 04/05/21 19:26, Ben Gardon wrote:
> On Mon, May 3, 2021 at 6:42 AM Paolo Bonzini <pbonzini@xxxxxxxxxx> wrote:
> > On 29/04/21 23:18, Ben Gardon wrote:
> > > +void activate_shadow_mmu(struct kvm *kvm)
> > > +{
> > > +     kvm->arch.shadow_mmu_active = true;
> > > +}
> > > +
> >
> > I think there's no lock protecting both the write and the read side.
> > Therefore this should be an smp_store_release, and all checks in
> > patch 2 should be an smp_load_acquire.
>
> That makes sense.
>
> > Also, the assignments to slot->arch.rmap in patch 4 (alloc_memslot_rmap)
> > should be an rcu_assign_pointer, while __gfn_to_rmap must be changed like so:
> >
> > +       struct kvm_rmap_head *head;
> > ...
> > -       return &slot->arch.rmap[level - PG_LEVEL_4K][idx];
> > +       head = srcu_dereference(slot->arch.rmap[level - PG_LEVEL_4K], &kvm->srcu,
> > +                                lockdep_is_held(&kvm->slots_arch_lock));
> > +       return &head[idx];
>
> I'm not sure I fully understand why this becomes necessary after patch
> 4. Isn't it already needed since the memslots are protected by RCU? Or
> is there already a higher level rcu dereference?
>
> __kvm_memslots already does an srcu dereference, so is there a path
> where we aren't getting the slots from that function where this is
> needed?

There are two point of views:

1) the easier one is just CONFIG_PROVE_RCU debugging: the rmaps need to 
be accessed under RCU because the memslots can disappear as soon as 
kvm->srcu is unlocked.

2) the harder one (though at this point I'm better at figuring out these 
ordering bugs than "traditional" mutex races) is what the happens before 
relation[1] looks like.  Consider what happens if the rmaps are 
allocated by *another thread* after the slots have been fetched.

thread 1		thread 2		thread 3
allocate memslots
rcu_assign_pointer
			slots = srcu_dereference
						allocate rmap
						rcu_assign_pointer
			head = slot->arch.rmap[]

Here, thread 3 is allocating the rmaps in the SRCU-protected 
kvm_memslots; those rmaps that didn't exist at the time thread 1 did the 
rcu_assign_pointer (which synchronizes with thread 2's srcu_dereference 
that retrieves slots), hence they were not covered by the release 
semantics of that rcu_assign_pointer and the "consume" semantics of the 
corresponding srcu_dereference.  Therefore, thread 2 needs another 
srcu_dereference when retrieving them.

Paolo

[1] https://lwn.net/Articles/844224/

> I wouldn't say that the rmaps are protected by RCU in any way that
> separate from the memslots.