On Thu, Jul 23, 2020 at 12:29:26PM +0100, Catalin Marinas wrote: > On Wed, Jul 22, 2020 at 01:40:34PM +0000, liwei (CM) wrote: > > Catalin Marinas wrote: > > > On Wed, Jul 22, 2020 at 08:41:17AM +0000, liwei (CM) wrote: > > > > Mike Rapoport wrote: > > > > > On Tue, Jul 21, 2020 at 03:32:03PM +0800, Wei Li wrote: > > > > > > For the memory hole, sparse memory model that define > > > > > > SPARSEMEM_VMEMMAP do not free the reserved memory for the page > > > > > > map, this patch do it. > > > > > > > > > > Are there numbers showing how much memory is actually freed? > > > > > > > > > > The freeing of empty memmap would become rather complex with these > > > > > changes, do the memory savings justify it? > > > > > > > > In the sparse memory model, the size of a section is 1 GB > > > > (SECTION_SIZE_BITS 30) by default. > > > > > > Can we reduce SECTION_SIZE_BITS instead? Say 26? > > > > Yes, you are right, reduce SECTION_SIZE_BITS to 26 can save almost the > > same memory as the patch. > > > > 1) However, it is not clear whether changing the section size has any > > other impact. > > Well, we should analyse this. > > > 2) Just like the flat memory model and the sparse memory model that > > does not define VMEMMAP, both of them have their own ways to free > > unused memmap. I think we've given a similar way for sparse memory > > define VMEMMAP. > > I think we did it for flatmem initially (on arm32) and added support for > sparsemem later on, so free_unused_memmap() had to cope with sparse > sections. On arm64 we introduced vmemmap support and didn't bother with > the freeing at all because of the added complexity of the vmemmap page > tables. > > I wonder whether we should just disallow flatmem and non-vmemmap > sparsemem on arm64. Is there any value in keeping them around?
FLATMEM is useful for UMA systems with a single memory bank, so probably it's worth keeping it for low end machines. Non-vmemmap sparsemem is essentially disable in arch/arm64/Kconfig, so for NUMA configurations SPARSEMEM_VMEMMAP is the only choice. > > 3) This explicit free unused memmap method does reduce unnecessary > > memory waste for users who do not notice the section size > > modification. > > But if we changed SECTION_SIZE_BITS in the mainline kernel, then we > wouldn't need additional code to free the unused memmap. Moreover if we reduce SECTION_SIZE_BITS, we can drop free_unused_memmap() and since the arm64 memory map for sparse will not differ from other arches we can drop custom pfn_valid() as well. > -- > Catalin -- Sincerely yours, Mike.