arm64: virt_to_page() does not return right page for a kernel image address

[Mark Rutland]

On Wed, Jan 04, 2017 at 05:28:32PM +0530, Pratyush Anand wrote:
> Hi Mark,

Hi Pratyush,

> On Wednesday 04 January 2017 04:41 PM, Mark Rutland wrote:
> >Other than the (new) crypto test and the (not yet upstream) kdump
> >patches, does any code rely on virt_to_page() working for a kernel image
> >address? If so, and if we cannot fix those in the short term, we may
> >want to temporarily revert commit 9f2875912dac35d9 until those are fixed
> >up.
> >
> >Regardless, I think that the kdump code should not rely on
> >virt_to_page() for a kernel image (or kmap) result.
> 
> Its not the kdump code which is relying on virt_to_page() for a
> kernel image.

Sorry, I had misread the above. I understand now.

> Its only crypto test which does that. In the first kernel (none
> kdump case) also crypto test gets a wrong kmap_atomic() address,
> however luckily there exists an entry for that address in page table
> and so a valid corresponding physical location. Therefore, it is
> just that we do not see the crash. I see still the crypto test
> failure print in the 1st kernel.
> However, in crash kernel there was no valid physical address and so
> the kernel crashed.

So it seems we need to fix the crypto test.

Looking at crypto/testmgr.c, I can't spot the kmap_atomic() or the
page_address()/virt_to_page(). I guess that's hidden behind helpers,
which might also be used elsewhere?

Could you elaborate on where exactly the problem is in the crypto test?

Thanks,
Mark.