[RFC][PATCH RESEND] mm: vmalloc: remove ioremap align constraint

[Arnd Bergmann]

On Saturday 03 January 2015 18:59:46 Sergey Dyasly wrote:
> Hi Arnd,
> 
> First, some background information. We originally encountered high fragmentation
> issue in vmalloc area:
> 
> 	1. Total size of vmalloc area was 400 MB.
> 	2. 200 MB of vmalloc area was consumed by ioremaps of various sizes.
> 	3. Largest contiguous chunk of vmalloc area was 12 MB.
> 	4. ioremap of 10 MB failed due to 8 MB alignment requirement.

Interesting, can you describe how you end up with that many ioremap mappings?
200MB seems like a lot. Do you perhaps get a lot of duplicate entries for the
same hardware registers, or maybe a leak?

Can you send the output of /proc/vmallocinfo?
 
> It was decided to further increase the size of vmalloc area to resolve the above
> issue. And I don't like that solution because it decreases the amount of lowmem.

If all the mappings are in fact required, have you considered using
CONFIG_VMSPLIT_2G split to avoid the use of highmem?

> Now let's see how ioremap uses supersections. Judging from current implementation
> of __arm_ioremap_pfn_caller:
> 
> 	#if !defined(CONFIG_SMP) && !defined(CONFIG_ARM_LPAE)
> 		if (pfn >= 0x100000 && !((paddr | size | addr) & ~SUPERSECTION_MASK)) {
> 			remap_area_supersections();
> 		} else if (!((paddr | size | addr) & ~PMD_MASK)) {
> 			remap_area_sections();
> 		} else
> 	#endif
> 			err = ioremap_page_range();
> 
> supersections and sections mappings are used only in !SMP && !LPAE case.
> Otherwise, mapping is created using the usual 4K pages (and we are using SMP).
> The suggested patch removes alignment requirements for ioremap but it means that
> sections will not be used in !SMP case. So another solution is required.
> 
> __get_vm_area_node has align parameter, maybe it can be used to specify the
> required alignment of ioremap operation? Because I find current generic fls
> algorithm to be very restrictive in cases when it's not necessary to use such
> a big alignment.

I think using next-power-of-two alignment generally helps limit the effects of
fragmentation the same way that the buddy allocator works.

Since the section and supersection maps are only used with non-SMP non-LPAE
(why is that the case btw?), it would however make sense to use the default
(7 + PAGE_SHIFT) instead of the ARM-specific 24 here if one of them is set,
I don't see any downsides to that.

	Arnd