Messages in this thread | | | Date | Thu, 11 Apr 2024 15:55:08 +0200 | Subject | Re: [PATCH v5 1/2] mm/madvise: optimize lazyfreeing with mTHP in madvise_free | From | David Hildenbrand <> |
| |
On 11.04.24 15:51, Ryan Roberts wrote: > On 11/04/2024 13:23, Lance Yang wrote: >> On Thu, Apr 11, 2024 at 7:27 PM Ryan Roberts <ryan.roberts@arm.com> wrote: >>> >>> 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) >>>>> >> >> Thanks for the suggestions, Ryan, David! >> >>>>> 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: >> >> Yeah, this is definitely cleaner than before. >> >>>>> >>>>> 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. >> >> Agreed. It's indeed a good idea to use a single function for both use cases. >> >>>> >>>> Yes. Maybe better use proper __bitwise flags, the compiler should be smart >>>> enough to optimize either way. >> >> Nice. I'll use the __bitwise flags as the input. >> >>> >>> 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. >>> >> >> Should I start using output bitwise flags for folio_pte_batch() in >> this patch set? > > I don't think its crucial (yet). I'd leave it as you have done it for now, > unless David shouts.
Let's do that separately, and investigate if the compiler actually is smart enough ... :)
-- Cheers,
David / dhildenb
| |