Re: [PATCH 1/4] mm: page_alloc: reduce unnecessary binary search in memblock_next_valid_pfn()

[Eugeniu Rosca]

On Wed, Mar 21, 2018 at 08:28:18PM +0800, Jia He wrote:
> 
> 
> On 3/21/2018 6:14 PM, Daniel Vacek Wrote:
> >On Wed, Mar 21, 2018 at 9:09 AM, Jia He <hejianet@xxxxxxxxx> wrote:
> >>Commit b92df1de5d28 ("mm: page_alloc: skip over regions of invalid pfns
> >>where possible") optimized the loop in memmap_init_zone(). But there is
> >>still some room for improvement. E.g. if pfn and pfn+1 are in the same
> >>memblock region, we can simply pfn++ instead of doing the binary search
> >>in memblock_next_valid_pfn.
> >There is a revert-mm-page_alloc-skip-over-regions-of-invalid-pfns-where-possible.patch
> >in -mm reverting b92df1de5d289c0b as it is fundamentally wrong by
> >design causing system panics on some machines with rare but still
> >valid mappings. Basically it skips valid pfns which are outside of
> >usable memory ranges (outside of memblock memory regions).
> Thanks for the infomation.
> quote from you patch description:
> >But given some specific memory mapping on x86_64 (or more generally
> theoretically anywhere but on arm with CONFIG_HAVE_ARCH_PFN_VALID) > the
> implementation also skips valid pfns which is plain wrong and causes >
> 'kernel BUG at mm/page_alloc.c:1389!'
> 
> Do you think memblock_next_valid_pfn can remain to be not reverted on arm64
> with CONFIG_HAVE_ARCH_PFN_VALID? Arm64 can benifit from this optimization.

I confirm that the boot time of Rcar-H3 arm64 platform greatly
benefits from v4.11-rc1 commit b92df1de5d28 ("mm: page_alloc: skip over
regions of invalid pfns where possible"). The startup improvement is
roughly ~140ms, which will be lost if the mentioned commit is reverted.

For more details on my measurements, please see linux-next commit
283f1645e236 ("mm: page_alloc: skip over regions of invalid pfns on
UMA").

Whichever way you decide to go forward (reimplement/fix b92df1de5d28
or create an <arch>_next_valid_pfn), I am willing to participate in
testing your proposals on RCAR SoCs. TIA.

Thanks,
Eugeniu.