Re: [PATCH v2] KVM: arm64: Correctly handle page aging notifiers for unaligned memslot

[Marc Zyngier <maz@kernel.org>]

Hi Oliver.,

On Tue, 11 Jul 2023 19:04:27 +0100,
Oliver Upton <oliver.upton@linux.dev> wrote:
> 
> Hey Marc,
> 
> On Tue, Jul 11, 2023 at 11:10:01AM +0100, Marc Zyngier wrote:
> > On Wed, 28 Jun 2023 00:54:05 +0100, Oliver Upton <oliver.upton@linux.dev> wrote:
> > > +static int stage2_age_walker(const struct kvm_pgtable_visit_ctx *ctx,
> > > +			     enum kvm_pgtable_walk_flags visit)
> > >  {
> > > -	kvm_pte_t pte = 0;
> > > -	stage2_update_leaf_attrs(pgt, addr, 1, 0, KVM_PTE_LEAF_ATTR_LO_S2_AF,
> > > -				 &pte, NULL, 0);
> > > +	kvm_pte_t new = ctx->old & ~KVM_PTE_LEAF_ATTR_LO_S2_AF;
> > > +	struct stage2_age_data *data = ctx->arg;
> > > +
> > > +	if (!kvm_pte_valid(ctx->old) || new == ctx->old)
> > > +		return 0;
> > > +
> > > +	data->young = true;
> > > +
> > > +	if (data->mkold && !stage2_try_set_pte(ctx, new))
> > > +		return -EAGAIN;
> > > +
> > >  	/*
> > >  	 * "But where's the TLBI?!", you scream.
> > >  	 * "Over in the core code", I sigh.
> > >  	 *
> > >  	 * See the '->clear_flush_young()' callback on the KVM mmu notifier.
> > >  	 */
> > > -	return pte;
> > > +	return 0;
> > >  }
> > >  
> > > -bool kvm_pgtable_stage2_is_young(struct kvm_pgtable *pgt, u64 addr)
> > > +bool kvm_pgtable_stage2_test_clear_young(struct kvm_pgtable *pgt, u64 addr,
> > > +					 u64 size, bool mkold)
> > >  {
> > > -	kvm_pte_t pte = 0;
> > > -	stage2_update_leaf_attrs(pgt, addr, 1, 0, 0, &pte, NULL, 0);
> > > -	return pte & KVM_PTE_LEAF_ATTR_LO_S2_AF;
> > > +	struct stage2_age_data data = {
> > > +		.mkold		= mkold,
> > > +	};
> > > +	struct kvm_pgtable_walker walker = {
> > > +		.cb		= stage2_age_walker,
> > > +		.arg		= &data,
> > > +		.flags		= KVM_PGTABLE_WALK_LEAF,
> > > +	};
> > > +
> > > +	WARN_ON(kvm_pgtable_walk(pgt, addr, size, &walker));
> > 
> > Do we really want a WARN_ON() here? From what I can tell, it can be
> > (trivially?) triggered by the previous function returning -EAGAIN if
> > the pte update fails in the case of a shared walk.
> 
> I threw the -EAGAIN in there just due to reflexes, we're holding the MMU
> write lock at this point so stage2_try_set_pte() will always succeed. A
> tad fragile, but wanted to make it trivial to change the locking around
> stage2_age_walker() in the future.

Right, the notifier takes the write lock, ensuring that there is never
a concurrent access. Maybe a small comment above the return would
help, as '-EAGAIN' is easily interpreted as "we missed the boat, but
we'll take the next one, no big deal".

> 
> The reason I wanted to have a WARN here is because we're unable to
> return an error on the MMU notifier and might need some breadcrumbs to
> debug any underlying issues in the table walker. I'd really like to keep
> it in some form.
> 
> I can either replace stage2_try_set_pte() with a direct WRITE_ONCE()
> (eliminating the error path) or leave it as-is. Which do you prefer?

With the above nit addressed:

Reviewed-by: Marc Zyngier <maz@kernel.org>

	M.

-- 
Without deviation from the norm, progress is not possible.