Re: [PATCH] ARM: memblock limit must be pmd-aligned

[Mark Rutland <mark.rutland@arm.com>]

On Mon, Jun 26, 2017 at 05:50:03PM -0700, Doug Berger wrote:
> On 06/26/2017 04:43 PM, Laura Abbott wrote:
> > On 06/26/2017 10:23 AM, Doug Berger wrote:
> >> There is a path through the adjust_lowmem_bounds() routine where if all
> >> memory regions start and end on pmd-aligned addresses the memblock_limit
> >> will be set to arm_lowmem_limit.
> >>
> >> However, since arm_lowmem_limit can be affected by the vmalloc early
> >> parameter, the value of arm_lowmem_limit may not be pmd-aligned. This
> >> commit corrects this oversight such that memblock_limit is always rounded
> >> down to pmd-alignment.
> >>
> >> The pmd containing arm_lowmem_limit is cleared by prepare_page_table()
> >> and without this commit it is possible for early_alloc() to allocate
> >> unmapped memory in that range when mapping the lowmem.
> >>
> > 
> > Do you have an example system or configuration where you see this
> > crash?
> I have observed this crash occur on systems like the bcm7445 when a
> customer uses the vmalloc boot parameter to specify an odd number of
> Megabytes of VMALLOC space (e.g. vmalloc=751m).  This seems to be a
> popular way for them to set the low memory boundary.
> 
> As long as vmalloc is a multiple of the pmd (e.g. 2MB) there isn't a
> problem, so documenting this constraint is another possible solution.
> However, educating the user is more difficult in this case than working
> around a questionable value to allow the boot to succeed.

It sounds like this leads to the same issue as we tried to fix in
commit:

  965278dcb8ab0b1f ("ARM: 8356/1: mm: handle non-pmd-aligned end of RAM")

... where with !LPAE page tables, we don't map the last section (as we
can't map the whole PMD containig it), but arm_lowmem_limit doesn't
account for this, and we try to access memroy from the unmapped section,
blowing up.

We're just failing to account for this where we don't have an inital
memblock_limit.

> 
> -Doug
> > 
> > Thanks,
> > Laura
> > 
> >> Signed-off-by: Doug Berger <opendmb@gmail.com>
> >> ---
> >>  arch/arm/mm/mmu.c | 2 +-
> >>  1 file changed, 1 insertion(+), 1 deletion(-)
> >>
> >> diff --git a/arch/arm/mm/mmu.c b/arch/arm/mm/mmu.c
> >> index 31af3cb59a60..2ae4f9c9d757 100644
> >> --- a/arch/arm/mm/mmu.c
> >> +++ b/arch/arm/mm/mmu.c
> >> @@ -1226,7 +1226,7 @@ void __init adjust_lowmem_bounds(void)
> >>  	if (memblock_limit)
> >>  		memblock_limit = round_down(memblock_limit, PMD_SIZE);
> >>  	if (!memblock_limit)
> >> -		memblock_limit = arm_lowmem_limit;
> >> +		memblock_limit = round_down(arm_lowmem_limit, PMD_SIZE);
> >>  

Given we're always going to do the rounding, how about we move that out
of the existing conditional, i.e. get rid of the first if, and have:

	if (!memblock_limit)
		memblock_limit = arm_lowmem_limit;

	/*
	 * Round the memblock limit down to a pmd size.  This
	 * helps to ensure that we will allocate memory from the
	 * last full pmd, which should be mapped.
	 */
	memblock_limit = round_down(memblock_limit, PMD_SIZE);

Thanks,
Mark.