[PATCH] arm64: enable THP_SWAP for arm64

[Anshuman Khandual]

On 5/24/22 16:45, Barry Song wrote:
> On Tue, May 24, 2022 at 10:05 PM Barry Song <21cnbao at gmail.com> wrote:
>>
>> On Tue, May 24, 2022 at 8:12 PM Catalin Marinas <catalin.marinas at arm.com> wrote:
>>>
>>> On Tue, May 24, 2022 at 07:14:03PM +1200, Barry Song wrote:
>>>> From: Barry Song <v-songbaohua at oppo.com>
>>>>
>>>> THP_SWAP has been proved to improve the swap throughput significantly
>>>> on x86_64 according to commit bd4c82c22c367e ("mm, THP, swap: delay
>>>> splitting THP after swapped out").
>>>> As long as arm64 uses 4K page size, it is quite similar with x86_64
>>>> by having 2MB PMD THP. So we are going to get similar improvement.
>>>> For other page sizes such as 16KB and 64KB, PMD might be too large.
>>>> Negative side effects such as IO latency might be a problem. Thus,
>>>> we can only safely enable the counterpart of X86_64.
>>>>
>>>> Cc: "Huang, Ying" <ying.huang at intel.com>
>>>> Cc: Minchan Kim <minchan at kernel.org>
>>>> Cc: Johannes Weiner <hannes at cmpxchg.org>
>>>> Cc: Hugh Dickins <hughd at google.com>
>>>> Cc: Shaohua Li <shli at kernel.org>
>>>> Cc: Rik van Riel <riel at redhat.com>
>>>> Cc: Andrea Arcangeli <aarcange at redhat.com>
>>>> Signed-off-by: Barry Song <v-songbaohua at oppo.com>
>>>> ---
>>>>  arch/arm64/Kconfig | 1 +
>>>>  1 file changed, 1 insertion(+)
>>>>
>>>> diff --git a/arch/arm64/Kconfig b/arch/arm64/Kconfig
>>>> index d550f5acfaf3..8e3771c56fbf 100644
>>>> --- a/arch/arm64/Kconfig
>>>> +++ b/arch/arm64/Kconfig
>>>> @@ -98,6 +98,7 @@ config ARM64
>>>>       select ARCH_WANT_HUGE_PMD_SHARE if ARM64_4K_PAGES || (ARM64_16K_PAGES && !ARM64_VA_BITS_36)
>>>>       select ARCH_WANT_LD_ORPHAN_WARN
>>>>       select ARCH_WANTS_NO_INSTR
>>>> +     select ARCH_WANTS_THP_SWAP if ARM64_4K_PAGES
>>>
>>> I'm not opposed to this but I think it would break pages mapped with
>>> PROT_MTE. We have an assumption in mte_sync_tags() that compound pages
>>> are not swapped out (or in). With MTE, we store the tags in a slab
>>
>> I assume you mean mte_sync_tags() require that THP is not swapped as a whole,
>> as without THP_SWP, THP is still swapping after being splitted. MTE doesn't stop
>> THP from swapping through a couple of splitted pages, does it?
>>
>>> object (128-bytes per swapped page) and restore them when pages are
>>> swapped in. At some point we may teach the core swap code about such
>>> metadata but in the meantime that was the easiest way.
>>>
>>
>> If my previous assumption is true,  the easiest way to enable THP_SWP
>> for this moment
>> might be always letting mm fallback to the splitting way for MTE
>> hardware. For this
>> moment, I care about THP_SWP more as none of my hardware has MTE.
>>
>> diff --git a/arch/arm64/include/asm/pgtable.h b/arch/arm64/include/asm/pgtable.h
>> index 45c358538f13..d55a2a3e41a9 100644
>> --- a/arch/arm64/include/asm/pgtable.h
>> +++ b/arch/arm64/include/asm/pgtable.h
>> @@ -44,6 +44,8 @@
>>         __flush_tlb_range(vma, addr, end, PUD_SIZE, false, 1)
>>  #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
>>
>> +#define arch_thp_swp_supported !system_supports_mte
>> +
>>  /*
>>   * Outside of a few very special situations (e.g. hibernation), we always
>>   * use broadcast TLB invalidation instructions, therefore a spurious page
>> diff --git a/include/linux/huge_mm.h b/include/linux/huge_mm.h
>> index 2999190adc22..064b6b03df9e 100644
>> --- a/include/linux/huge_mm.h
>> +++ b/include/linux/huge_mm.h
>> @@ -447,4 +447,16 @@ static inline int split_folio_to_list(struct folio *folio,
>>         return split_huge_page_to_list(&folio->page, list);
>>  }
>>
>> +/*
>> + * archs that select ARCH_WANTS_THP_SWAP but don't support THP_SWP due to
>> + * limitations in the implementation like arm64 MTE can override this to
>> + * false
>> + */
>> +#ifndef arch_thp_swp_supported
>> +static inline bool arch_thp_swp_supported(void)
>> +{
>> +       return true;
>> +}
>> +#endif
>> +
>>  #endif /* _LINUX_HUGE_MM_H */
>> diff --git a/mm/swap_slots.c b/mm/swap_slots.c
>> index 2b5531840583..dde685836328 100644
>> --- a/mm/swap_slots.c
>> +++ b/mm/swap_slots.c
>> @@ -309,7 +309,7 @@ swp_entry_t get_swap_page(struct page *page)
>>         entry.val = 0;
>>
>>         if (PageTransHuge(page)) {
>> -               if (IS_ENABLED(CONFIG_THP_SWAP))
>> +               if (IS_ENABLED(CONFIG_THP_SWAP) && arch_thp_swp_supported())
>>                         get_swap_pages(1, &entry, HPAGE_PMD_NR);
>>                 goto out;
>>         }
>>
> 
> Am I actually able to go further to only split MTE tagged pages?
> 
> For mm core:
> 
> +/*
> + * archs that select ARCH_WANTS_THP_SWAP but don't support THP_SWP due to
> + * limitations in the implementation like arm64 MTE can override this to
> + * false
> + */
> +#ifndef arch_thp_swp_supported
> +static inline bool arch_thp_swp_supported(struct page *page)
> +{
> +       return true;
> +}
> +#endif
> +
> 
> For arm64:
> +#define arch_thp_swp_supported(page) !test_bit(PG_mte_tagged, &page->flags)

Although not entirely sure, but per page arch_thp_swp_supported() callback
seems bit risky. What if there scenarios or time windows when PG_mte_tagged
is cleared on an otherwise MTE tagged page ? I guess arch_thp_swp_supported()
just returning false on a system with MTE support, is a better option.

> 
> But I don't have MTE hardware to test. So to me, totally disabling THP_SWP
> is safer.
> 
> thoughts?
>>> --
>>> Catalin
>>
>> Thanks
>> Barry
>