[PATCH 0/2] Hibernate fixes for 'Fix memmap to be initialized for the entire section'

[Will Deacon]

On Wed, Dec 07, 2016 at 10:06:38AM +0100, Robert Richter wrote:
> On 06.12.16 17:38:11, Will Deacon wrote:
> > On Mon, Dec 05, 2016 at 03:42:14PM +0000, Ard Biesheuvel wrote:
> > > On 2 December 2016 at 14:49, James Morse <james.morse at arm.com> wrote:
> > > > Patch "arm64: mm: Fix memmap to be initialized for the entire section"
> > > > changes pfn_valid() in a way that breaks hibernate. These patches fix
> > > > hibernate, and provided struct page's are allocated for nomap pages,
> > > > can be applied before [0].
> > > >
> > > > Hibernate core code belives 'valid' to mean "I can access this". It
> > > > uses pfn_valid() to test the page if the page is 'valid'.
> > > >
> > > > pfn_valid() needs to be changed so that all struct pages in a numa
> > > > node have the same node-id. Currently 'nomap' pages are skipped, and
> > > > retain their pre-numa node-ids, which leads to a later BUG_ON().
> > > >
> > > > These patches make hibernate's savable_page() take its escape route
> > > > via 'if (PageReserved(page) && pfn_is_nosave(pfn))'.
> > > >
> > > 
> > > This makes me feel slightly uneasy. Robert makes a convincing point,
> > > but I wonder if we can expect more fallout from the ambiguity of
> > > pfn_valid(). Now we are not only forced to assign non-existing (as far
> > > as the OS is concerned) pages to the correct NUMA node, we also need
> > > to set certain page flags.
> > 
> > Yes, I really don't know how to proceed here. Playing whack-a-mole with
> > pfn_valid() users doesn't sounds like an improvement on the current
> > situation to me.
> > 
> > Robert -- if we leave pfn_valid() as it is, would a point-hack to
> > memmap_init_zone help, or do you anticipate other problems?
> 
> I would suggest to fix the hibernation code as I commented on before
> to use pfn_is_nosave() that defaults to pfn_valid() but uses memblock_
> is_nomap() for arm64. Let's just fix it and see if no other issues
> arise. I am trying to send a patch for this until tomorrow.

I'd rather not use mainline as a guinea pig like this, since I'd be very
surprised if other places don't break given the scope for different
interpretations of pfn_valid.

> I am also going to see how early_pfn_valid() could be redirected to
> use memblock_is_nomap() on arm64. That would "quick fix" the problem,
> though I rather prefer to go further with the current solution.

I don't like either of them, but early_pfn_valid is easier to revert so
let's go with that.

Will