Re: [PATCH v5 1/2] mm/madvise: optimize lazyfreeing with mTHP in madvise_free

[Ryan Roberts <ryan.roberts@arm.com>]

On 11/04/2024 12:20, David Hildenbrand wrote:
> On 11.04.24 13:11, Ryan Roberts wrote:
>> On 08/04/2024 05:24, Lance Yang wrote:
>>> This patch optimizes lazyfreeing with PTE-mapped mTHP[1]
>>> (Inspired by David Hildenbrand[2]). We aim to avoid unnecessary folio
>>> splitting if the large folio is fully mapped within the target range.
>>>
>>> If a large folio is locked or shared, or if we fail to split it, we just
>>> leave it in place and advance to the next PTE in the range. But note that
>>> the behavior is changed; previously, any failure of this sort would cause
>>> the entire operation to give up. As large folios become more common,
>>> sticking to the old way could result in wasted opportunities.
>>>
>>> On an Intel I5 CPU, lazyfreeing a 1GiB VMA backed by PTE-mapped folios of
>>> the same size results in the following runtimes for madvise(MADV_FREE) in
>>> seconds (shorter is better):
>>>
>>> Folio Size |   Old    |   New    | Change
>>> ------------------------------------------
>>>        4KiB | 0.590251 | 0.590259 |    0%
>>>       16KiB | 2.990447 | 0.185655 |  -94%
>>>       32KiB | 2.547831 | 0.104870 |  -95%
>>>       64KiB | 2.457796 | 0.052812 |  -97%
>>>      128KiB | 2.281034 | 0.032777 |  -99%
>>>      256KiB | 2.230387 | 0.017496 |  -99%
>>>      512KiB | 2.189106 | 0.010781 |  -99%
>>>     1024KiB | 2.183949 | 0.007753 |  -99%
>>>     2048KiB | 0.002799 | 0.002804 |    0%
>>>
>>> [1] https://lkml.kernel.org/r/20231207161211.2374093-5-ryan.roberts@arm.com
>>> [2] https://lore.kernel.org/linux-mm/20240214204435.167852-1-david@redhat.com
>>>
>>> Signed-off-by: Lance Yang <ioworker0@gmail.com>
>>> ---
>>>   include/linux/pgtable.h |  34 +++++++++
>>>   mm/internal.h           |  12 +++-
>>>   mm/madvise.c            | 149 ++++++++++++++++++++++------------------
>>>   mm/memory.c             |   4 +-
>>>   4 files changed, 129 insertions(+), 70 deletions(-)
>>>
>>> diff --git a/include/linux/pgtable.h b/include/linux/pgtable.h
>>> index 0f4b2faa1d71..4dd442787420 100644
>>> --- a/include/linux/pgtable.h
>>> +++ b/include/linux/pgtable.h
>>> @@ -489,6 +489,40 @@ static inline pte_t ptep_get_and_clear(struct mm_struct
>>> *mm,
>>>   }
>>>   #endif
>>>   +#ifndef mkold_clean_ptes
>>> +/**
>>> + * mkold_clean_ptes - Mark PTEs that map consecutive pages of the same folio
>>> + *        as old and clean.
>>> + * @mm: Address space the pages are mapped into.
>>> + * @addr: Address the first page is mapped at.
>>> + * @ptep: Page table pointer for the first entry.
>>> + * @nr: Number of entries to mark old and clean.
>>> + *
>>> + * May be overridden by the architecture; otherwise, implemented by
>>> + * get_and_clear/modify/set for each pte in the range.
>>> + *
>>> + * Note that PTE bits in the PTE range besides the PFN can differ. For example,
>>> + * some PTEs might be write-protected.
>>> + *
>>> + * Context: The caller holds the page table lock.  The PTEs map consecutive
>>> + * pages that belong to the same folio.  The PTEs are all in the same PMD.
>>> + */
>>> +static inline void mkold_clean_ptes(struct mm_struct *mm, unsigned long addr,
>>> +                    pte_t *ptep, unsigned int nr)
>>
>> Just thinking out loud, I wonder if it would be cleaner to convert mkold_ptes()
>> (which I added as part of swap-out) to something like:
>>
>> clear_young_dirty_ptes(struct mm_struct *mm, unsigned long addr,
>>                pte_t *ptep, unsigned int nr,
>>                bool clear_young, bool clear_dirty);
>>
>> Then we can use the same function for both use cases and also have the ability
>> to only clear dirty in future if we ever need it. The other advantage is that we
>> only need to plumb a single function down the arm64 arch code. As it currently
>> stands, those 2 functions would be duplicating most of their code.
> 
> Yes. Maybe better use proper __bitwise flags, the compiler should be smart
> enough to optimize either way.

Agreed. I was also thinking perhaps it makes sense to start using output bitwise
flags for folio_pte_batch() since this patch set takes us up to 3 optional bool
pointers for different things. Might be cleaner to have input flags to tell it
what we care about and output flags to highlight those things. I guess the
compiler should be able to optimize in the same way.