On 04.06.20 22:00, Dave Hansen wrote: > On 6/4/20 11:12 AM, Daniel Jordan wrote: >>> E.g., on powerpc that's 16MB so they have *a lot* of memory blocks. >>> That's why that's not papering over the problem. Increasing the memory >>> block size isn't always the answer. >> Ok. If you don't mind, what's the purpose of hotplugging at that granularity? >> I'm simply curious. > > FWIW, the 128MB on x86 came from the original sparsemem/hotplug > implementation. It was the size of the smallest DIMM that my server > system at the time would take. ppc64's huge page size was and is 16MB > and that's also the granularity with which hypervisors did hot-add way > back then. I'm not actually sure what they do now. > > My belief at the time was that the section size would grow over time as > DIMMs and hotplug units grew. I was young and naive. :) BTW, I recently studied your old hotplug papers and they are highly appreciated :) > > I actually can't think of anything that's *keeping* it at 128MB on x86 > though. We don't, for instance, require a whole section to be > pfn_valid(). Well, sub-section hotadd is only done for vmemmap and we only use it for !(memory block devices) stuff, a.k.a. ZONE_DEVICE. IIRC, sub-section hotadd works in granularity of 2M. AFAIK: - The lower limit for a section is MAX_ORDER - 1 / pageblock_order - The smaller the section, the more bits are wasted to store the section number in page->flags for page_to_pfn() (!vmemmap IIRC) - The smaller the section, the bigger the section array(s) - We want to make sure the section memmap always spans full pages (IIRC, not always the case e.g., arm64 with 256k page size. But arm64 is weird either way - 512MB (transparent) huge pages with 64k base pages ...) Changing the section size to get rid of sub-section memory hotadd does not seem to be easily possible. I assume we don't want to create memory block devices for something as small as current sub-section memory hotadd size (e.g., 2MB). So having significantly smaller sections might not make too much sense and your initial section size might have been a very good, initial pick :) -- Thanks, David / dhildenb
