Re: [PATCH v4 2/8] KVM: arm64: Introduce pkvm_alloc_private_va_range()

[Marc Zyngier]

Hi Kalesh,

On Wed, 02 Mar 2022 17:24:53 +0000,
Kalesh Singh <kaleshsingh@xxxxxxxxxx> wrote:
> 
> On Tue, Mar 1, 2022 at 11:46 PM Marc Zyngier <maz@xxxxxxxxxx> wrote:
> >
> > On Fri, 25 Feb 2022 03:34:47 +0000,
> > Kalesh Singh <kaleshsingh@xxxxxxxxxx> wrote:
> > >
> > > pkvm_hyp_alloc_private_va_range() can be used to reserve private VA ranges
> > > in the pKVM nVHE hypervisor (). Also update __pkvm_create_private_mapping()
> > > to allow specifying an alignment for the private VA mapping.
> > >
> > > These will be used to implement stack guard pages for pKVM nVHE hypervisor
> > > (in a subsequent patch in the series).
> > >
> > > Credits to Quentin Perret <qperret@xxxxxxxxxx> for the idea of moving
> > > private VA allocation out of __pkvm_create_private_mapping()
> > >
> > > Signed-off-by: Kalesh Singh <kaleshsingh@xxxxxxxxxx>
> > > ---
> > >
> > > Changes in v4:
> > >   - Handle null ptr in pkvm_alloc_private_va_range() and replace
> > >     IS_ERR_OR_NULL checks in callers with IS_ERR checks, per Fuad
> > >   - Fix kernel-doc comments format, per Fuad
> > >   - Format __pkvm_create_private_mapping() prototype args (< 80 col), per Fuad
> > >
> > > Changes in v3:
> > >   - Handle null ptr in IS_ERR_OR_NULL checks, per Mark
> > >
> > > Changes in v2:
> > >   - Allow specifying an alignment for the private VA allocations, per Marc
> >
> > I probably badly expressed my earlier concern.
> >
> > Yes, an alignment is necessary. But how often do we want an alignment
> > that isn't naturally aligned to the size of the allocation (i.e. the
> > power of 2 >= the size of the allocation)? This is what the rest of
> > the kernel does (get_order() and co), and I thing we should follow
> > this.
> 
> Hi Marc,
> 
> Thanks for clarifying. I think making the alignment implicitly based
> on the size here will create unnecessary holes where PAGE_SIZE
> alignment would be ok and potentially overflow the private VA space
> earlier. Is it not a concern?

I don't think we should worry too much about this. Even when building
the kernel with a very small VA space (commonly 39 bits), we still
have a quarter of that reserved for private EL2 mappings. That's
pretty big.

We will use a bit more of the memory that is set aside for EL2 page
tables, but this shouldn't be a problem either.

Thanks,

	M.

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