[PATCH] arm64: mm: always enable CONFIG_HOLES_IN_ZONE
mhocko at kernel.org
Mon Sep 10 02:25:11 PDT 2018
On Fri 07-09-18 18:47:24, James Morse wrote:
> Hi Michal,
> On 03/09/18 20:47, Michal Hocko wrote:
> > On Thu 30-08-18 16:05:32, James Morse wrote:
> >> Commit 6d526ee26ccd ("arm64: mm: enable CONFIG_HOLES_IN_ZONE for NUMA")
> >> only enabled HOLES_IN_ZONE for NUMA systems because the NUMA code was
> >> choking on the missing zone for nomap pages. This problem doesn't just
> >> apply to NUMA systems.
> >> If the architecture doesn't set HAVE_ARCH_PFN_VALID, pfn_valid() will
> >> return true if the pfn is part of a valid sparsemem section.
> >> When working with multiple pages, the mm code uses pfn_valid_within()
> >> to test each page it uses within the sparsemem section is valid. On
> >> most systems memory comes in MAX_ORDER_NR_PAGES chunks which all
> >> have valid/initialised struct pages. In this case pfn_valid_within()
> >> is optimised out.
> >> Systems where this isn't true (e.g. due to nomap) should set
> >> HOLES_IN_ZONE and provide HAVE_ARCH_PFN_VALID so that mm tests each
> >> page as it works with it.
> >> Currently non-NUMA arm64 systems can't enable HOLES_IN_ZONE, leading to
> >> VM_BUG_ON()
> >> Remove the NUMA dependency.
> >> Reported-by: Mikulas Patocka <mpatocka at redhat.com>
> >> Link: https://www.spinics.net/lists/arm-kernel/msg671851.html
> >> Fixes: 6d526ee26ccd ("arm64: mm: enable CONFIG_HOLES_IN_ZONE for NUMA")
> >> CC: Ard Biesheuvel <ard.biesheuvel at linaro.org>
> >> Signed-off-by: James Morse <james.morse at arm.com>
> > OK. I guess you are also going to post a patch to drop
> > ARCH_HAS_HOLES_MEMORYMODEL, right?
> After all this I'm suspicious about arm64's support for FLATMEM given we always
> set HAVE_ARCH_PFN_VALID.
> > Anyway
> > Acked-by: Michal Hocko <mhocko at suse.com>
> > I wish we could simplify the pfn validation code a bit. I find
> > pfn_valid_within quite confusing and I would bet it is not used
> > consistently.
> > This will require a non trivial audit. I am wondering
> > whether we really need to make the code more complicated rather than
> > simply establish a contract that we always have a pageblock worth of
> > struct pages always available. Even when there is no physical memory
> > backing it. Such a page can be reserved and never used by the page
> > allocator. pfn walkers should back off for reserved pages already.
> Is PG_Reserved really where this stops?
> Going through the mail archive it looks like whenever this crops up on arm64 the
> issues are with nomap pages needing a 'correct' node or zone, where-as we would
> prefer it if linux knew nothing about them.
Well, I will not pretend I have a clear view on early mem init code. I
have seen so many surprises lately that I just gave up. I can clearly
see why you want nomap pages to have no backing struct pages. It just
makes sense but I strongly suspect that pfn_valid_within is not the
right approach. If for no other reason it is basically unmaintainable
interface. All/Most pfn walkers should use it but I do not see this
being the case. I strongly suspect that initializing sub section memmaps
is quite wastefull on its own (especially with VMEMAP) because you are
losing the large kernel pagetables for those memmaps. So having a full
section worth of initialized memory and then reserving holes should
result in a better maintainable code because pfn_valid_within shouldn't
be really needed. But I might easily miss something subtle here.
Especially arm specific.
More information about the linux-arm-kernel