On 1/11/21 3:43 PM, David Hildenbrand wrote: > On 11.01.21 05:17, Anshuman Khandual wrote: >> >> >> On 1/8/21 9:00 PM, David Hildenbrand wrote: >>>> To summarize, the section size bits for each base page size config >>>> should always >>>> >>>> a. Avoid (MAX_ORDER - 1 + PAGE_SHIFT) > SECTION_SIZE_BITS >>> >>> Pageblocks must also always fall completely into a section. >>> >>>> >>>> b. Provide minimum possible section size for a given base page config to >>>> have increased agility during memory hotplug operations and reduced >>>> vmemmap wastage for sections with holes. >>> >>> OTOH, making the section size too small (e.g., 16MB) creates way to many >>> memory block devices in /sys/devices/system/memory/, and might consume >>> too many page->flags bits in the !vmemmap case. >>> >>> For bigger setups, we might, similar to x86-64 (e.g., >= 64 GiB), >>> determine the memory_block_size_bytes() at runtime (spanning multiple >>> sections then), once it becomes relevant. >>> >>>> >>>> c. Allow 4K base page configs to have PMD based vmemmap mappings >>> >>> Agreed. >>> >>>> >>>> Because CONFIG_FORCE_MAX_ZONEORDER is always defined on arm64 platform, >>>> the following would always avoid the condition (a) >>>> >>>> SECTION_SIZE_BITS (CONFIG_FORCE_MAX_ZONEORDER - 1 + PAGE_SHIFT) >>>> >>>> - 22 (11 - 1 + 12) for 4K pages >>>> - 24 (11 - 1 + 14) for 16K pages without THP >>>> - 25 (12 - 1 + 14) for 16K pages with THP >>>> - 26 (11 - 1 + 16) for 64K pages without THP >>>> - 29 (14 - 1 + 16) for 64K pages with THP >>>> >>>> Apart from overriding 4K base page size config to have 27 as section size >>>> bits, should not all other values be okay here ? But then wondering what >>>> benefit 128MB (27 bits) section size for 16K config would have ? OR the >>>> objective here is to match 16K page size config with default x86-64. >>> >>> We don't want to have sections that are too small. We don't want to have >>> sections that are too big :) >>> >>> Not sure if we really want to allow setting e.g., a section size of 4 >>> MB. That's just going to hurt. IMHO, something in the range of 64..256 >>> MB is quite a good choice, where possible. >>> >>>> >>>>> >>>>> (If we worry about the number of section bits in page->flags, we could >>>>> glue it to vmemmap support where that does not matter) >>>> >>>> Could you please elaborate ? Could smaller section size bits numbers like >>>> 22 or 24 create problems in page->flags without changing other parameters >>>> like NR_CPUS or NODES_SHIFT ? A quick test with 64K base page without THP >>> >>> Yes, in the !vmemmap case, we have to store the section_nr in there. >>> IIRC, it's less of an issue with section sizes like 128 MB. >>> >>>> i.e 26 bits in section size, fails to boot. >>> >>> 26 bits would mean 64 MB, no? Not sure if that's possible even without >>> THP (MAX_ORDER - 1, pageblock_order ...) on 64k pages. I'd assume 512 MB >>> is the lowest we can go. I'd assume this would crash :) >>> >>>> >>>> As you have suggested, probably constant defaults (128MB for 4K/16K, 512MB >>>> for 64K) might be better than depending on the CONFIG_FORCE_MAX_ZONEORDER, >>>> at least for now. >>> >>> That's also what I would prefer, keeping it simple. >> >> Okay sure, will send a RFC to begin with. >> > > Note that there is > > https://lkml.kernel.org/r/15cf9a2359197fee0168f820c5c904650d07939e.1610146597.git.sudaraja@xxxxxxxxxxxxxx > > (Sudarshan missed to cc linux-mm) > Right, some how missed that. Anyways, ended up spending some time testing the change for various configs.
