| Message ID | 20240125164256.4147-32-alexandru.elisei@arm.com (mailing list archive) |
|---|---|
| State | Handled Elsewhere |
| Headers | show |
| Series | Add support for arm64 MTE dynamic tag storage reuse | expand |
On 1/25/24 22:12, Alexandru Elisei wrote: > copy_user_highpage() will do memory allocation if there are saved tags for > the destination page, and the page is missing tag storage. > > After commit a349d72fd9ef ("mm/pgtable: add rcu_read_lock() and > rcu_read_unlock()s"), collapse_huge_page() calls > __collapse_huge_page_copy() -> .. -> copy_user_highpage() with the RCU lock > held, which means that copy_user_highpage() can only allocate memory using > GFP_ATOMIC or equivalent. > > Get around this by refusing to collapse pages into a transparent huge page > if the VMA is MTE-enabled. Makes sense when copy_user_highpage() will allocate memory for tag storage. > > Signed-off-by: Alexandru Elisei <alexandru.elisei@arm.com> > --- > > Changes since rfc v2: > > * New patch. I think an agreement on whether copy*_user_highpage() should be > always allowed to sleep, or should not be allowed, would be useful. This is a good question ! Even after preventing the collapse of MTE VMA here, there still might be more paths where a sleeping (i.e memory allocating) copy*_user_highpage() becomes problematic ? > > arch/arm64/include/asm/pgtable.h | 3 +++ > arch/arm64/kernel/mte_tag_storage.c | 5 +++++ > include/linux/khugepaged.h | 5 +++++ > mm/khugepaged.c | 4 ++++ > 4 files changed, 17 insertions(+) > > diff --git a/arch/arm64/include/asm/pgtable.h b/arch/arm64/include/asm/pgtable.h > index 87ae59436162..d0473538c926 100644 > --- a/arch/arm64/include/asm/pgtable.h > +++ b/arch/arm64/include/asm/pgtable.h > @@ -1120,6 +1120,9 @@ static inline bool arch_alloc_cma(gfp_t gfp_mask) > return true; > } > > +bool arch_hugepage_vma_revalidate(struct vm_area_struct *vma, unsigned long address); > +#define arch_hugepage_vma_revalidate arch_hugepage_vma_revalidate > + > #endif /* CONFIG_ARM64_MTE_TAG_STORAGE */ > #endif /* CONFIG_ARM64_MTE */ > > diff --git a/arch/arm64/kernel/mte_tag_storage.c b/arch/arm64/kernel/mte_tag_storage.c > index ac7b9c9c585c..a99959b70573 100644 > --- a/arch/arm64/kernel/mte_tag_storage.c > +++ b/arch/arm64/kernel/mte_tag_storage.c > @@ -636,3 +636,8 @@ void arch_alloc_page(struct page *page, int order, gfp_t gfp) > if (tag_storage_enabled() && alloc_requires_tag_storage(gfp)) > reserve_tag_storage(page, order, gfp); > } > + > +bool arch_hugepage_vma_revalidate(struct vm_area_struct *vma, unsigned long address) > +{ > + return !(vma->vm_flags & VM_MTE); > +} > diff --git a/include/linux/khugepaged.h b/include/linux/khugepaged.h > index f68865e19b0b..461e4322dff2 100644 > --- a/include/linux/khugepaged.h > +++ b/include/linux/khugepaged.h > @@ -38,6 +38,11 @@ static inline void khugepaged_exit(struct mm_struct *mm) > if (test_bit(MMF_VM_HUGEPAGE, &mm->flags)) > __khugepaged_exit(mm); > } > + > +#ifndef arch_hugepage_vma_revalidate > +#define arch_hugepage_vma_revalidate(vma, address) 1 Please replace s/1/true as arch_hugepage_vma_revalidate() returns bool ? > +#endif Right, above construct is much better than __HAVE_ARCH_XXXX based one. > + > #else /* CONFIG_TRANSPARENT_HUGEPAGE */ > static inline void khugepaged_fork(struct mm_struct *mm, struct mm_struct *oldmm) > { > diff --git a/mm/khugepaged.c b/mm/khugepaged.c > index 2b219acb528e..cb9a9ddb4d86 100644 > --- a/mm/khugepaged.c > +++ b/mm/khugepaged.c > @@ -935,6 +935,10 @@ static int hugepage_vma_revalidate(struct mm_struct *mm, unsigned long address, > */ > if (expect_anon && (!(*vmap)->anon_vma || !vma_is_anonymous(*vmap))) > return SCAN_PAGE_ANON; > + > + if (!arch_hugepage_vma_revalidate(vma, address)) > + return SCAN_VMA_CHECK; > + > return SCAN_SUCCEED; > } > Otherwise this LGTM.
On Thu, Feb 01, 2024 at 01:42:08PM +0530, Anshuman Khandual wrote: > > > On 1/25/24 22:12, Alexandru Elisei wrote: > > copy_user_highpage() will do memory allocation if there are saved tags for > > the destination page, and the page is missing tag storage. > > > > After commit a349d72fd9ef ("mm/pgtable: add rcu_read_lock() and > > rcu_read_unlock()s"), collapse_huge_page() calls > > __collapse_huge_page_copy() -> .. -> copy_user_highpage() with the RCU lock > > held, which means that copy_user_highpage() can only allocate memory using > > GFP_ATOMIC or equivalent. > > > > Get around this by refusing to collapse pages into a transparent huge page > > if the VMA is MTE-enabled. > > Makes sense when copy_user_highpage() will allocate memory for tag storage. > > > > > Signed-off-by: Alexandru Elisei <alexandru.elisei@arm.com> > > --- > > > > Changes since rfc v2: > > > > * New patch. I think an agreement on whether copy*_user_highpage() should be > > always allowed to sleep, or should not be allowed, would be useful. > > This is a good question ! Even after preventing the collapse of MTE VMA here, > there still might be more paths where a sleeping (i.e memory allocating) > copy*_user_highpage() becomes problematic ? Exactly! > > > > > arch/arm64/include/asm/pgtable.h | 3 +++ > > arch/arm64/kernel/mte_tag_storage.c | 5 +++++ > > include/linux/khugepaged.h | 5 +++++ > > mm/khugepaged.c | 4 ++++ > > 4 files changed, 17 insertions(+) > > > > diff --git a/arch/arm64/include/asm/pgtable.h b/arch/arm64/include/asm/pgtable.h > > index 87ae59436162..d0473538c926 100644 > > --- a/arch/arm64/include/asm/pgtable.h > > +++ b/arch/arm64/include/asm/pgtable.h > > @@ -1120,6 +1120,9 @@ static inline bool arch_alloc_cma(gfp_t gfp_mask) > > return true; > > } > > > > +bool arch_hugepage_vma_revalidate(struct vm_area_struct *vma, unsigned long address); > > +#define arch_hugepage_vma_revalidate arch_hugepage_vma_revalidate > > + > > #endif /* CONFIG_ARM64_MTE_TAG_STORAGE */ > > #endif /* CONFIG_ARM64_MTE */ > > > > diff --git a/arch/arm64/kernel/mte_tag_storage.c b/arch/arm64/kernel/mte_tag_storage.c > > index ac7b9c9c585c..a99959b70573 100644 > > --- a/arch/arm64/kernel/mte_tag_storage.c > > +++ b/arch/arm64/kernel/mte_tag_storage.c > > @@ -636,3 +636,8 @@ void arch_alloc_page(struct page *page, int order, gfp_t gfp) > > if (tag_storage_enabled() && alloc_requires_tag_storage(gfp)) > > reserve_tag_storage(page, order, gfp); > > } > > + > > +bool arch_hugepage_vma_revalidate(struct vm_area_struct *vma, unsigned long address) > > +{ > > + return !(vma->vm_flags & VM_MTE); > > +} > > diff --git a/include/linux/khugepaged.h b/include/linux/khugepaged.h > > index f68865e19b0b..461e4322dff2 100644 > > --- a/include/linux/khugepaged.h > > +++ b/include/linux/khugepaged.h > > @@ -38,6 +38,11 @@ static inline void khugepaged_exit(struct mm_struct *mm) > > if (test_bit(MMF_VM_HUGEPAGE, &mm->flags)) > > __khugepaged_exit(mm); > > } > > + > > +#ifndef arch_hugepage_vma_revalidate > > +#define arch_hugepage_vma_revalidate(vma, address) 1 > > Please replace s/1/true as arch_hugepage_vma_revalidate() returns bool ? Yeah, that's strange, I don't know why I used 1 there. Will change it to true, thanks for spotting it. > > > +#endif > > Right, above construct is much better than __HAVE_ARCH_XXXX based one. Thanks! Alex > > > + > > #else /* CONFIG_TRANSPARENT_HUGEPAGE */ > > static inline void khugepaged_fork(struct mm_struct *mm, struct mm_struct *oldmm) > > { > > diff --git a/mm/khugepaged.c b/mm/khugepaged.c > > index 2b219acb528e..cb9a9ddb4d86 100644 > > --- a/mm/khugepaged.c > > +++ b/mm/khugepaged.c > > @@ -935,6 +935,10 @@ static int hugepage_vma_revalidate(struct mm_struct *mm, unsigned long address, > > */ > > if (expect_anon && (!(*vmap)->anon_vma || !vma_is_anonymous(*vmap))) > > return SCAN_PAGE_ANON; > > + > > + if (!arch_hugepage_vma_revalidate(vma, address)) > > + return SCAN_VMA_CHECK; > > + > > return SCAN_SUCCEED; > > } > > > > Otherwise this LGTM.
diff --git a/arch/arm64/include/asm/pgtable.h b/arch/arm64/include/asm/pgtable.h index 87ae59436162..d0473538c926 100644 --- a/arch/arm64/include/asm/pgtable.h +++ b/arch/arm64/include/asm/pgtable.h @@ -1120,6 +1120,9 @@ static inline bool arch_alloc_cma(gfp_t gfp_mask) return true; } +bool arch_hugepage_vma_revalidate(struct vm_area_struct *vma, unsigned long address); +#define arch_hugepage_vma_revalidate arch_hugepage_vma_revalidate + #endif /* CONFIG_ARM64_MTE_TAG_STORAGE */ #endif /* CONFIG_ARM64_MTE */ diff --git a/arch/arm64/kernel/mte_tag_storage.c b/arch/arm64/kernel/mte_tag_storage.c index ac7b9c9c585c..a99959b70573 100644 --- a/arch/arm64/kernel/mte_tag_storage.c +++ b/arch/arm64/kernel/mte_tag_storage.c @@ -636,3 +636,8 @@ void arch_alloc_page(struct page *page, int order, gfp_t gfp) if (tag_storage_enabled() && alloc_requires_tag_storage(gfp)) reserve_tag_storage(page, order, gfp); } + +bool arch_hugepage_vma_revalidate(struct vm_area_struct *vma, unsigned long address) +{ + return !(vma->vm_flags & VM_MTE); +} diff --git a/include/linux/khugepaged.h b/include/linux/khugepaged.h index f68865e19b0b..461e4322dff2 100644 --- a/include/linux/khugepaged.h +++ b/include/linux/khugepaged.h @@ -38,6 +38,11 @@ static inline void khugepaged_exit(struct mm_struct *mm) if (test_bit(MMF_VM_HUGEPAGE, &mm->flags)) __khugepaged_exit(mm); } + +#ifndef arch_hugepage_vma_revalidate +#define arch_hugepage_vma_revalidate(vma, address) 1 +#endif + #else /* CONFIG_TRANSPARENT_HUGEPAGE */ static inline void khugepaged_fork(struct mm_struct *mm, struct mm_struct *oldmm) { diff --git a/mm/khugepaged.c b/mm/khugepaged.c index 2b219acb528e..cb9a9ddb4d86 100644 --- a/mm/khugepaged.c +++ b/mm/khugepaged.c @@ -935,6 +935,10 @@ static int hugepage_vma_revalidate(struct mm_struct *mm, unsigned long address, */ if (expect_anon && (!(*vmap)->anon_vma || !vma_is_anonymous(*vmap))) return SCAN_PAGE_ANON; + + if (!arch_hugepage_vma_revalidate(vma, address)) + return SCAN_VMA_CHECK; + return SCAN_SUCCEED; }
copy_user_highpage() will do memory allocation if there are saved tags for the destination page, and the page is missing tag storage. After commit a349d72fd9ef ("mm/pgtable: add rcu_read_lock() and rcu_read_unlock()s"), collapse_huge_page() calls __collapse_huge_page_copy() -> .. -> copy_user_highpage() with the RCU lock held, which means that copy_user_highpage() can only allocate memory using GFP_ATOMIC or equivalent. Get around this by refusing to collapse pages into a transparent huge page if the VMA is MTE-enabled. Signed-off-by: Alexandru Elisei <alexandru.elisei@arm.com> --- Changes since rfc v2: * New patch. I think an agreement on whether copy*_user_highpage() should be always allowed to sleep, or should not be allowed, would be useful. arch/arm64/include/asm/pgtable.h | 3 +++ arch/arm64/kernel/mte_tag_storage.c | 5 +++++ include/linux/khugepaged.h | 5 +++++ mm/khugepaged.c | 4 ++++ 4 files changed, 17 insertions(+)