Re: [PATCH v3 03/10] arm64, kfence: enable KFENCE for ARM64

From: Mark Rutland
Date: Tue Sep 29 2020 - 09:54:17 EST


On Tue, Sep 22, 2020 at 11:56:26AM +0200, Marco Elver wrote:
> On Mon, 21 Sep 2020 at 19:44, Will Deacon <will@xxxxxxxxxx> wrote:
> [...]
> > > > > > For ARM64, we would like to solicit feedback on what the best option is
> > > > > > to obtain a constant address for __kfence_pool. One option is to declare
> > > > > > a memory range in the memory layout to be dedicated to KFENCE (like is
> > > > > > done for KASAN), however, it is unclear if this is the best available
> > > > > > option. We would like to avoid touching the memory layout.
> > > > >
> > > > > Sorry for the delay on this.
> > > >
> > > > NP, thanks for looking!
> > > >
> > > > > Given that the pool is relatively small (i.e. when compared with our virtual
> > > > > address space), dedicating an area of virtual space sounds like it makes
> > > > > the most sense here. How early do you need it to be available?
> > > >
> > > > Yes, having a dedicated address sounds good.
> > > > We're inserting kfence_init() into start_kernel() after timekeeping_init().
> > > > So way after mm_init(), if that matters.
> > >
> > > The question is though, how big should that dedicated area be?
> > > Right now KFENCE_NUM_OBJECTS can be up to 16383 (which makes the pool
> > > size 64MB), but this number actually comes from the limitation on
> > > static objects, so we might want to increase that number on arm64.
> >
> > What happens on x86 and why would we do something different?
>
> On x86 we just do `char __kfence_pool[KFENCE_POOL_SIZE] ...;` to
> statically allocate the pool. On arm64 this doesn't seem to work
> because static memory doesn't have struct pages?

Are you using virt_to_page() directly on that statically-allocated
__kfence_pool? If so you'll need to use lm_alias() if so, as is done in
mm/kasan/init.c.

Anything statically allocated is part of the kernel image address range
rather than the linear/direct map, and doesn't have a valid virt addr,
but its linear map alias does.

If you enable CONFIG_DEBUG_VIRTUAL you should get warnings if missing
lm_alias() calls.

Thanks,
Mark.