On 06.12.18 10:21, Wei Yang wrote: > On Thu, Dec 06, 2018 at 10:00:05AM +0100, David Hildenbrand wrote: >> On 05.12.18 23:31, Wei Yang wrote: >>> On Wed, Dec 05, 2018 at 03:37:33PM +0000, Mel Gorman wrote: >>>> On Wed, Dec 05, 2018 at 12:08:20PM +0000, Wei Yang wrote: >>>>> On Wed, Dec 05, 2018 at 11:15:13AM +0000, Mel Gorman wrote: >>>>>> On Wed, Dec 05, 2018 at 05:19:04PM +0800, Wei Yang wrote: >>>>>>> When SPARSEMEM is used, there is an indication that pageblock is not >>>>>>> allowed to exceed one mem_section. Current code doesn't have this >>>>>>> constrain explicitly. >>>>>>> >>>>>>> This patch adds this to make sure it won't. >>>>>>> >>>>>>> Signed-off-by: Wei Yang <richard.weiyang@xxxxxxxxx> >>>>>> >>>>>> Is this even possible? This would imply that the section size is smaller >>>>>> than max order which would be quite a crazy selection for a sparesemem >>>>>> section size. A lot of assumptions on the validity of PFNs within a >>>>>> max-order boundary would be broken with such a section size. I'd be >>>>>> surprised if such a setup could even boot, let alone run. >>>>> >>>>> pageblock_order has two definitions. >>>>> >>>>> #define pageblock_order HUGETLB_PAGE_ORDER >>>>> >>>>> #define pageblock_order (MAX_ORDER-1) >>>>> >>>>> If CONFIG_HUGETLB_PAGE is not enabled, pageblock_order is related to >>>>> MAX_ORDER, which ensures it is smaller than section size. >>>>> >>>>> If CONFIG_HUGETLB_PAGE is enabled, pageblock_order is not related to >>>>> MAX_ORDER. I don't see HUGETLB_PAGE_ORDER is ensured to be less than >>>>> section size. Maybe I missed it? >>>>> >>>> >>>> HUGETLB_PAGE_ORDER is less than MAX_ORDER on the basis that normal huge >>>> pages (not gigantic) pages are served from the buddy allocator which is >>>> limited by MAX_ORDER. >>>> >>> >>> Maybe I am lost here, I got one possible definition on x86. >>> >>> #define pageblock_order HUGETLB_PAGE_ORDER >>> #define HUGETLB_PAGE_ORDER (HPAGE_SHIFT - PAGE_SHIFT) >>> #define HPAGE_SHIFT PMD_SHIFT >>> #define PMD_SHIFT PUD_SHIFT >> >> PMD_SHIFT is usually 21 >> >> arch/x86/include/asm/pgtable-3level_types.h:#define PMD_SHIFT 21 >> arch/x86/include/asm/pgtable_64_types.h:#define PMD_SHIFT 21 >> >> Unless CONFIG_PGTABLE_LEVELS <= 2 >> >> Then include/asm-generic/pgtable-nopmd.h will be used in >> arch/x86/include/asm/pgtable_types.h >> #define PMD_SHIFT PUD_SHIFT >> >> In that case, also include/asm-generic/pgtable-nopmd.h is uses >> #define PUD_SHIFT P4D_SHIFT >> >> ... include/asm-generic/pgtable-nop4d.h >> #define P4D_SHIFT PGDIR_SHIFT >> >> >> And that would be >> arch/x86/include/asm/pgtable-2level_types.h:#define PGDIR_SHIFT 22 >> >> If I am not wrong. >> >> So we would have pageblock_order = (22 - 12) = 10 >> > > Thank, David :-) > > I think current configuration is correct, while all these digits are > written by programmer. > > My concern and suggestion is to add a compiler check to enforce this. So > that we would avoid this situation if someone miss this constrain. Just > as the check on MAX_ORDER and SECION_SIZE. I am not completely against this, I rather wonder if it is needed because I assume other things will break horribly in case this is violated. And at that would only be helpful for somebody developing for a new architecture/flavor. As I am a friend of documenting things that are not obvious, I would rather suggest to add a comment to the #define pageblock_order HUGETLB_PAGE_ORDER line, stating what we just learned. /* * HUGETLB_PAGE_ORDER will always be smaller than MAX_ORDER, so that * huge (not gigantic) pages can be served from the buddy allocator. */ -- Thanks, David / dhildenb
