On 27/06/2023 19:26, Yu Zhao wrote: > On Tue, Jun 27, 2023 at 3:41 AM Ryan Roberts <ryan.roberts@xxxxxxx> wrote: >> >> On 27/06/2023 09:29, Yu Zhao wrote: >>> On Tue, Jun 27, 2023 at 1:21 AM Ryan Roberts <ryan.roberts@xxxxxxx> wrote: >>>> >>>> On 27/06/2023 02:55, Yu Zhao wrote: >>>>> On Mon, Jun 26, 2023 at 11:14 AM Ryan Roberts <ryan.roberts@xxxxxxx> wrote: >>>>>> >>>>>> In preparation for extending vma_alloc_zeroed_movable_folio() to >>>>>> allocate a arbitrary order folio, expose clear_huge_page() >>>>>> unconditionally, so that it can be used to zero the allocated folio in >>>>>> the generic implementation of vma_alloc_zeroed_movable_folio(). >>>>>> >>>>>> Signed-off-by: Ryan Roberts <ryan.roberts@xxxxxxx> >>>>>> --- >>>>>> include/linux/mm.h | 3 ++- >>>>>> mm/memory.c | 2 +- >>>>>> 2 files changed, 3 insertions(+), 2 deletions(-) >>>>>> >>>>>> diff --git a/include/linux/mm.h b/include/linux/mm.h >>>>>> index 7f1741bd870a..7e3bf45e6491 100644 >>>>>> --- a/include/linux/mm.h >>>>>> +++ b/include/linux/mm.h >>>>>> @@ -3684,10 +3684,11 @@ enum mf_action_page_type { >>>>>> */ >>>>>> extern const struct attribute_group memory_failure_attr_group; >>>>>> >>>>>> -#if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS) >>>>>> extern void clear_huge_page(struct page *page, >>>>>> unsigned long addr_hint, >>>>>> unsigned int pages_per_huge_page); >>>>>> + >>>>>> +#if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS) >>>>> >>>>> We might not want to depend on THP eventually. Right now, we still >>>>> have to, unless splitting is optional, which seems to contradict >>>>> 06/10. (deferred_split_folio() is a nop without THP.) >>>> >>>> Yes, I agree - for large anon folios to work, we depend on THP. But I don't >>>> think that helps us here. >>>> >>>> In the next patch, I give vma_alloc_zeroed_movable_folio() an extra `order` >>>> parameter. So the generic/default version of the function now needs a way to >>>> clear a compound page. >>>> >>>> I guess I could do something like: >>>> >>>> static inline >>>> struct folio *vma_alloc_zeroed_movable_folio(struct vm_area_struct *vma, >>>> unsigned long vaddr, gfp_t gfp, int order) >>>> { >>>> struct folio *folio; >>>> >>>> folio = vma_alloc_folio(GFP_HIGHUSER_MOVABLE | gfp, >>>> order, vma, vaddr, false); >>>> if (folio) { >>>> #ifdef CONFIG_LARGE_FOLIO >>>> clear_huge_page(&folio->page, vaddr, 1U << order); >>>> #else >>>> BUG_ON(order != 0); >>>> clear_user_highpage(&folio->page, vaddr); >>>> #endif >>>> } >>>> >>>> return folio; >>>> } >>>> >>>> But that's pretty messy and there's no reason why other users might come along >>>> that pass order != 0 and will be surprised by the BUG_ON. >>> >>> #ifdef CONFIG_LARGE_ANON_FOLIO // depends on CONFIG_TRANSPARENT_HUGE_PAGE >>> struct folio *alloc_anon_folio(struct vm_area_struct *vma, unsigned >>> long vaddr, int order) >>> { >>> // how do_huge_pmd_anonymous_page() allocs and clears >>> vma_alloc_folio(..., *true*); >> >> This controls the mem allocation policy (see mempolicy.c::vma_alloc_folio()) not >> clearing. Clearing is done in __do_huge_pmd_anonymous_page(): >> >> clear_huge_page(page, vmf->address, HPAGE_PMD_NR); > > Sorry for rushing this previously. This is what I meant. The #ifdef > makes it safe to use clear_huge_page() without 01/10. I highlighted > the last parameter to vma_alloc_folio() only because it's different > from what you chose (not implying it clears the folio).> >>> } >>> #else >>> #define alloc_anon_folio(vma, addr, order) >>> vma_alloc_zeroed_movable_folio(vma, addr) >>> #endif >> >> Sorry I don't get this at all... If you are suggesting to bypass >> vma_alloc_zeroed_movable_folio() entirely for the LARGE_ANON_FOLIO case > > Correct. > >> I don't >> think that works because the arch code adds its own gfp flags there. For >> example, arm64 adds __GFP_ZEROTAGS for VM_MTE VMAs. > > I think it's the opposite: it should be safer to reuse the THP code because > 1. It's an existing case that has been working for PMD_ORDER folios > mapped by PTEs, and it's an arch-independent API which would be easier > to review. > 2. Use vma_alloc_zeroed_movable_folio() for large folios is a *new* > case. It's an arch-*dependent* API which I have no idea what VM_MTE > does (should do) to large folios and don't plan to answer that for > now. I've done some archaology on this now, and convinced myself that your suggestion is a good one - sorry for doubting it! If you are interested here are the details: Only arm64 and ia64 do something non-standard in vma_alloc_zeroed_movable_folio(). ia64 flushes the dcache for the folio - but given it does not support THP this is not a problem for the THP path. arm64 adds the __GFP_ZEROTAGS flag which means that the MTE tags will be zeroed at the same time as the page is zeroed. This is a perf optimization - if its not performed then it will be done at set_pte_at(), which is how this works for the THP path. So on that basis, I agree we can use your proposed alloc_anon_folio() approach. arm64 will lose the MTE optimization but that can be added back later if needed. So no need to unconditionally expose clear_huge_page() and no need to modify all the arch vma_alloc_zeroed_movable_folio() implementations. Thanks, Ryan > >> Perhaps we can do away with an arch-owned vma_alloc_zeroed_movable_folio() and >> replace it with a new arch_get_zeroed_movable_gfp_flags() then >> alloc_anon_folio() add in those flags? >> >> But I still think the cleanest, simplest change is just to unconditionally >> expose clear_huge_page() as I've done it. > > The fundamental choice there as I see it is to whether the first step > of large anon folios should lean toward the THP code base or the base > page code base (I'm a big fan of the answer "Neither -- we should > create something entirely new instead"). My POV is that the THP code > base would allow us to move faster, since it's proven to work for a > very similar case (PMD_ORDER folios mapped by PTEs).