Re: Calling vmalloc_to_page() on ioremap memory?

[Alexander Potapenko <glider@google.com>]

On Mon, Jun 25, 2018 at 6:27 PM Mark Rutland <mark.rutland@arm.com> wrote:
>
> On Mon, Jun 25, 2018 at 06:24:57PM +0200, Alexander Potapenko wrote:
> > On Mon, Jun 25, 2018 at 6:00 PM Mark Rutland <mark.rutland@arm.com> wrote:
> > >
> > > On Mon, Jun 25, 2018 at 04:59:23PM +0200, Alexander Potapenko wrote:
> > > > Hi Ard, Mark, Andrew and others,
> > > >
> > > > AFAIU, commit 029c54b09599573015a5c18dbe59cbdf42742237 ("mm/vmalloc.c:
> > > > huge-vmap: fail gracefully on unexpected huge vmap mappings") was
> > > > supposed to make vmalloc_to_page() return NULL for pointers not
> > > > returned by vmalloc().
> > >
> > > It's a little more subtle than that -- avoiding an edge case where we
> > > unexpectedly hit huge mappings, rather than determining whether an
> > > address same from vmalloc().
> > Ok, but anyway, acpi_os_ioremap() creates a huge page mapping via
> > __ioremap_caller() (see
> > https://elixir.bootlin.com/linux/latest/source/arch/x86/mm/ioremap.c#L133)
> > Shouldn't these checks detect that as well?
>
> It should catch such mappings, yes.
>
> > > > For memory error detection purposes I'm trying to map the addresses
> > > > from the vmalloc range to valid struct pages, or at least make sure
> > > > there's no struct page for a given address.
> > > > Looking up the vmap_area_root rbtree isn't an option, as this must be
> > > > done from instrumented code, including interrupt handlers.
> > >
> > > I'm not sure how you can do this without looking at VMAs.
> > >
> > > In general, the vmalloc area can contain addresses which are not memory,
> > > and this cannot be detremined from the address alone.
> > I thought this was exactly what vmalloc_to_page() did, but apparently no.
> >
> > > You *might* be able to get away with pfn_valid(vmalloc_to_pfn(x)), but
> > > IIRC there's some disagreement on the precise meaning of pfn_valid(), so
> > > that might just tell you that the address happens to fall close to some
> > > valid memory.
> > This appears to work, at least for ACPI mappings. I'll check other cases though.
> > Thank you!
pfn_valid(vmalloc_to_pfn(x)) works for me, so I'll stick to this
solution for now. Thanks again!

But just to clarify, should vmalloc_to_page() return NULL for a huge
mapping returned by __ioremap_caller()?
Your answer and that of Ard seem to be contradictory.
Maybe it's a good idea to add the pfn_valid() check to
vmalloc_to_page() just to be sure?
> Great!
>
> Thanks,
> Mark.



-- 
Alexander Potapenko
Software Engineer

Google Germany GmbH
Erika-Mann-Straße, 33
80636 München

Geschäftsführer: Paul Manicle, Halimah DeLaine Prado
Registergericht und -nummer: Hamburg, HRB 86891
Sitz der Gesellschaft: Hamburg