Message-ID: <20210312093205.GB32016@willie-the-truck>
Date:   Fri, 12 Mar 2021 09:32:06 +0000
From:   Will Deacon <will@...nel.org>
To:     Quentin Perret <qperret@...gle.com>
Cc:     catalin.marinas@....com, maz@...nel.org, james.morse@....com,
        julien.thierry.kdev@...il.com, suzuki.poulose@....com,
        android-kvm@...gle.com, linux-kernel@...r.kernel.org,
        kernel-team@...roid.com, kvmarm@...ts.cs.columbia.edu,
        linux-arm-kernel@...ts.infradead.org, tabba@...gle.com,
        mark.rutland@....com, dbrazdil@...gle.com, mate.toth-pal@....com,
        seanjc@...gle.com, robh+dt@...nel.org, ardb@...nel.org
Subject: Re: [PATCH v4 28/34] KVM: arm64: Use page-table to track page
 ownership

On Fri, Mar 12, 2021 at 06:23:00AM +0000, Quentin Perret wrote:
> On Thursday 11 Mar 2021 at 18:38:36 (+0000), Will Deacon wrote:
> > On Wed, Mar 10, 2021 at 05:57:45PM +0000, Quentin Perret wrote:
> > > As the host stage 2 will be identity mapped, all the .hyp memory regions
> > > and/or memory pages donated to protected guestis will have to marked
> > > invalid in the host stage 2 page-table. At the same time, the hypervisor
> > > will need a way to track the ownership of each physical page to ensure
> > > memory sharing or donation between entities (host, guests, hypervisor) is
> > > legal.
> > > 
> > > In order to enable this tracking at EL2, let's use the host stage 2
> > > page-table itself. The idea is to use the top bits of invalid mappings
> > > to store the unique identifier of the page owner. The page-table owner
> > > (the host) gets identifier 0 such that, at boot time, it owns the entire
> > > IPA space as the pgd starts zeroed.
> > > 
> > > Provide kvm_pgtable_stage2_set_owner() which allows to modify the
> > > ownership of pages in the host stage 2. It re-uses most of the map()
> > > logic, but ends up creating invalid mappings instead. This impacts
> > > how we do refcount as we now need to count invalid mappings when they
> > > are used for ownership tracking.
> > > 
> > > Signed-off-by: Quentin Perret <qperret@...gle.com>
> > > ---
> > >  arch/arm64/include/asm/kvm_pgtable.h | 21 +++++++
> > >  arch/arm64/kvm/hyp/pgtable.c         | 92 ++++++++++++++++++++++++----
> > >  2 files changed, 101 insertions(+), 12 deletions(-)
> > > 
> > > diff --git a/arch/arm64/include/asm/kvm_pgtable.h b/arch/arm64/include/asm/kvm_pgtable.h
> > > index 4ae19247837b..b09af4612656 100644
> > > --- a/arch/arm64/include/asm/kvm_pgtable.h
> > > +++ b/arch/arm64/include/asm/kvm_pgtable.h
> > > @@ -238,6 +238,27 @@ int kvm_pgtable_stage2_map(struct kvm_pgtable *pgt, u64 addr, u64 size,
> > >  			   u64 phys, enum kvm_pgtable_prot prot,
> > >  			   void *mc);
> > >  
> > > +/**
> > > + * kvm_pgtable_stage2_set_owner() - Annotate invalid mappings with metadata
> > > + *				    encoding the ownership of a page in the
> > > + *				    IPA space.
> > > + * @pgt:	Page-table structure initialised by kvm_pgtable_stage2_init().
> > > + * @addr:	Intermediate physical address at which to place the annotation.
> > 
> > This confused me a bit, as the annotation is stored in the page-table, not
> > at the memory identified by @addr. How about:
> > 
> >   "Base intermediate physical address to annotate"
> > 
> > > + * @size:	Size of the IPA range to annotate.
> > 
> >   "Size of the annotated range"
> > 
> > > + * @mc:		Cache of pre-allocated and zeroed memory from which to allocate
> > > + *		page-table pages.
> > > + * @owner_id:	Unique identifier for the owner of the page.
> > > + *
> > > + * The page-table owner has identifier 0.
> > 
> > Perhaps, "By default, all page-tables are owned by identifier 0"
> 
> Ack all of the above.
> 
> > > + *
> > > + * Return: 0 on success, negative error code on failure.
> > > + */
> > > +int kvm_pgtable_stage2_set_owner(struct kvm_pgtable *pgt, u64 addr, u64 size,
> > > +				 void *mc, u32 owner_id);
> > 
> > Is there a need for the owner_id to be 32-bit rather than e.g. 16-bit? Just
> > strikes me that it might be difficult to recover these bits in future if we
> > give them out freely now.
> 
> I figured we might want to use identifiers that are stable for the
> lifetime of protected VMs. I wasn't sure using e.g. VMIDs would be a
> better choice here as re-using them will cause a lot of pain for the
> host stage 2 pgtable maintenance.

I'm not saying to use the VMID directly, just that allocating half of the
pte feels a bit OTT given that the state of things after this patch series
is that we're using exactly 1 bit.

> > > @@ -517,28 +543,36 @@ static int stage2_map_walker_try_leaf(u64 addr, u64 end, u32 level,
> > >  	if (!kvm_block_mapping_supported(addr, end, phys, level))
> > >  		return -E2BIG;
> > >  
> > > -	new = kvm_init_valid_leaf_pte(phys, data->attr, level);
> > > -	if (kvm_pte_valid(old)) {
> > > +	if (kvm_pte_valid(data->attr))
> > 
> > This feels like a bit of a hack to me: the 'attr' field in stage2_map_data
> > is intended to correspond directly to the lower/upper attributes of the
> > descriptor as per the architecture, so tagging the valid bit in there is
> > pretty grotty. However, I can see the significant advantage in being able
> > to re-use the stage2_map_walker functionality, so about instead of nobbling
> > attr, you set phys to something invalid instead, e.g.:
> > 
> > 	#define KVM_PHYS_SET_OWNER	(-1ULL)
> 
> That'll confuse kvm_block_mapping_supported() and friends I think, at
> least in their current form. If you _really_ don't like this, maybe we
> could have an extra 'flags' field in stage2_map_data?

I was pondering this last night and I thought of two ways to do it:

1. Add a 'bool valid' and then stick the owner and the phys in a union.
   (yes, you'll need to update the block mapping checks to look at the
    valid flag)

2. Go with my latter suggestion:

> > Is there ever a reason to use kvm_pgtable_stage2_set_owner() to set an
> > owner of 0, or should you just use the map/unmap APIs for that? If so,
> > then maybe the key is simply if owner_id is non-zero, then an invalid
> > entry is installed?
> 
> I couldn't find a good reason to restrict it, as that wouldn't change
> the implementation much anyway. Also, if we added the right CMOs, we
> could probably remove the unmap walker and re-express it in terms of
> set_owner(0) ... But I suppose that is for later :-)

The idea being that if owner is 0, then we install a mapping for phys, but
if owner is !0 then we set the invalid mapping.

(1) is probably the less hacky option... what do you reckon?

Will

Powered by blists - more mailing lists