Skip to contentSkip to conversation

Question about the sparse memory section size

DISCUSSION chenfeng 2 participants 3 messages Sep 22, 2015 – Sep 23, 2015 1 message not in the archive

lore
Share & tools
lore

chenfeng

author
#0 lore raw

Hi all,
The sparse memory section size, SECTION_SIZE_BITS, currently is 1GB for arm64 by default. However, it might generate wasted memmap memory space for those memory sections less than 1GB. e.g.
for 512MB memory section, still 14MB(sizeof(struct page) * PAGES_PER_SECTION) memmap needs to be reserved. The wasted memmap space could be eliminated by changing memory section size from 1GB to 512M, but still some questions to be answered,

1) why arm64 uses 1GB as default setting?
2) any risk to change section size from 1GB to 512MB? like, any impact to performance since memory section number is increased.

Any help will be appreciated.

Puck

#1 lore raw
Quoted 5 lines · chenfeng

Hi all,
The sparse memory section size, SECTION_SIZE_BITS, currently is 1GB for arm64 by default. However, it might generate wasted memmap memory space for those memory sections less than 1GB. e.g.

Quoted 8 lines

for 512MB memory section, still 14MB(sizeof(struct page) * PAGES_PER_SECTION) memmap needs to be reserved. The wasted memmap space could be eliminated by changing memory section size from 1GB to 512M, but still some questions to be answered,

1) why arm64 uses 1GB as default setting?
2) any risk to change section size from 1GB to 512MB? like, any impact to performance since memory section number is increased.

For arm64 we have SPARSEMEM_VMEMMAP enabled by default, which enables much of the wasted memmap backing memory to be reclaimed.

Take a look at arch/arm64/mm/init.c:free_unused_memmap().

This should reduce the amount of actual memory wasted on unused parts of memmap. The virtual space stays wasted as you describe, but that's plentiful on 64-bit arches.

You could try sticking some printks in there is you want to see how much of the memmap the code successfully frees.

Cheers
---Dave

#3 lore raw
In reply to a message not in the archive
Quoted 5 lines · Qijiwen

In fact the free_unused_memmap function is not invoked when CONFIG_SPARSEMEM_VMEMMAP is enabled. The reason is that the memmap region is mapped in unit of 2MB, not in unit of 4KB.

Below is the source code in arch/arm64/mm/init.c:
void __init mem_init(void)

Quoted 8 lines
{
	swiotlb_init(1);

	set_max_mapnr(pfn_to_page(max_pfn) - mem_map);
#ifndef CONFIG_SPARSEMEM_VMEMMAP
	free_unused_memmap();

#endif

Oops, you're right. But if we're not already doing so, it could still make sense to unmap useless backing memory from the vmemmap, or try harder to avoid mapping it in the first place.

However, whether this is is worth it depends on what fraction of memory is wasted on your platform. After all, every page struct is "wasted" memory. Which means 1-2% of RAM is always "wasted" for a kernel with 4KB page size anyway.

Cheers
---Dave

[...]

Quoted 38 lines

Best regards,

Jiwen Qi

-----????-----
???: Dave Martin [mailto:Dave.Martin at arm.com]
????: 2015?9?22? 19:17
???: Chenfeng (puck)
??: catalin.marinas at arm.com; will.deacon at arm.com; mark.rutland at arm.com; ard.biesheuvel at linaro.org; lauraa at codeaurora.org; linux-arm-kernel at lists.infradead.org; linux-kernel at vger.kernel.org; Dan zhao; Xuyiping; Suzhuangluan; Qijiwen; fujun (F); Panshilin (Peter) ??: Re: Question about the sparse memory section size

Quoted 5 lines · chenfeng

Hi all,
The sparse memory section size, SECTION_SIZE_BITS, currently is 1GB for arm64 by default. However, it might generate wasted memmap memory space for those memory sections less than 1GB. e.g.

Quoted 8 lines

for 512MB memory section, still 14MB(sizeof(struct page) * PAGES_PER_SECTION) memmap needs to be reserved. The wasted memmap space could be eliminated by changing memory section size from 1GB to 512M, but still some questions to be answered,

1) why arm64 uses 1GB as default setting?
2) any risk to change section size from 1GB to 512MB? like, any impact to performance since memory section number is increased.

For arm64 we have SPARSEMEM_VMEMMAP enabled by default, which enables much of the wasted memmap backing memory to be reclaimed.

Take a look at arch/arm64/mm/init.c:free_unused_memmap().

This should reduce the amount of actual memory wasted on unused parts of memmap. The virtual space stays wasted as you describe, but that's plentiful on 64-bit arches.

You could try sticking some printks in there is you want to see how much of the memmap the code successfully frees.

Cheers
---Dave