Re: [PATCH] arm64: mm: support THP_SWAP on hardware with MTE

[Ryan Roberts]

On 08/12/2023 13:22, Matthew Wilcox wrote:
> On Fri, Dec 08, 2023 at 08:34:01PM +1300, Barry Song wrote:
>> arch_prepare_to_swap() should take folio rather than page as parameter
>> because we support THP swap-out as a whole. It saves tags for all
>> pages in a large folio.
>>
>> Meanwhile, arch_swap_restore() now moves to use page parameter rather
>> than folio because swap-in, refault and MTE tags are working at the
>> granularity of base pages rather than folio:
>> 1. a large folio in swapcache can be partially unmapped, thus, MTE
>> tags for the unmapped pages will be invalidated;
>> 2. users might use mprotect() to set MTEs on a part of a large folio.
> 
> I would argue that using mprotect() to set MTEs on part of a large folio
> should cause that folio to be split.  Could the user give us any
> stronger signal that this memory is being used for different purposes,
> and therefore should not be managed as a single entity?

I agree this probably makes sense here. But splitting is best effort as I
understand it? It can fail due to long-term GUP, right? In which case we still
have to handle the MTE on partial large folio case safely, even if not performantly.

As an aside, I don't think it's clear cut that we would always prefer to split
based on user space mprotect/madvise/etc calls. IIUC, there are garbage
collectors that temporarily mark pages RO then switch back to RW. I wouldn't
want to split here and lose the benefits of contpte forever. I'm handwaving
because I haven't looked into the exact mechanisms yet. But I think we need to
understand these users better before deciding on an "always split based on user
hints" policy.

> 
>> Thus, it won't be easy to manage MTE tags at the granularity of folios
>> since we do have some cases in which a part of pages in a large folios
>> have valid tags, while the other part of pages haven't. Furthermore,
>> trying to restore MTE tags for a whole folio can lead to many loops and
>> early exiting even if the large folio in swapcache are still entirely
>> mapped since do_swap_page() only sets PTE and frees swap for the base
>> page where PF is happening.
>>
>> But we still have a chance to restore tags for a whole large folio
>> once we support swap-in large folio. So this job is deferred till we
>> can do refault and swap-in as a large folio.
> 
> I strongly disagree with changing the interface to arch_swap_restore()
> from folio to page.