Re: [RFC V1 10/16] arm64/mm: Route all pgtable writes via ptdesc_set()
From: Ryan Roberts <ryan.roberts@arm.com>
Date: 2026-02-26 13:19:17
Also in:
linux-mm, lkml
On 26/02/2026 12:54, Anshuman Khandual wrote:
On 26/02/26 6:07 PM, Usama Arif wrote:quoted
On Tue, 24 Feb 2026 10:41:47 +0530 Anshuman Khandual [off-list ref] wrote:quoted
Currently ptdesc_set() is defined as WRITE_ONCE() but this will change for D128 pgtable builds, for which WRITE_ONCE() is not sufficient for single copy atomicity. In future this infrastructure can be used for D128 to maintain single copy atomicity semantics with inline asm blocks. Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: Will Deacon <will@kernel.org> Cc: Ryan Roberts <ryan.roberts@arm.com> Cc: Mark Rutland <mark.rutland@arm.com> Cc: linux-arm-kernel@lists.infradead.org Cc: linux-kernel@vger.kernel.org Signed-off-by: Anshuman Khandual <redacted> --- arch/arm64/include/asm/pgtable.h | 11 ++++++----- arch/arm64/mm/mmu.c | 4 ++-- mm/debug_vm_pgtable.c | 4 ++-- 3 files changed, 10 insertions(+), 9 deletions(-)diff --git a/arch/arm64/include/asm/pgtable.h b/arch/arm64/include/asm/pgtable.h index 804ef49aea88..42124d2f323d 100644 --- a/arch/arm64/include/asm/pgtable.h +++ b/arch/arm64/include/asm/pgtable.h@@ -85,6 +85,7 @@ static inline void arch_leave_lazy_mmu_mode(void) } #define ptdesc_get(x) READ_ONCE(x) +#define ptdesc_set(x, val) WRITE_ONCE(x, val) #define pmdp_get pmdp_get static inline pmd_t pmdp_get(pmd_t *pmdp)@@ -389,7 +390,7 @@ static inline pte_t pte_clear_uffd_wp(pte_t pte) static inline void __set_pte_nosync(pte_t *ptep, pte_t pte) { - WRITE_ONCE(*ptep, pte); + ptdesc_set(*ptep, pte); } static inline void __set_pte_complete(pte_t pte)@@ -856,7 +857,7 @@ static inline void set_pmd(pmd_t *pmdp, pmd_t pmd) } #endif /* __PAGETABLE_PMD_FOLDED */ - WRITE_ONCE(*pmdp, pmd); + ptdesc_set(*pmdp, pmd); if (pmd_valid(pmd)) queue_pte_barriers();@@ -917,7 +918,7 @@ static inline void set_pud(pud_t *pudp, pud_t pud) return; } - WRITE_ONCE(*pudp, pud); + ptdesc_set(*pudp, pud); if (pud_valid(pud)) queue_pte_barriers();@@ -999,7 +1000,7 @@ static inline void set_p4d(p4d_t *p4dp, p4d_t p4d) return; } - WRITE_ONCE(*p4dp, p4d); + ptdesc_set(*p4dp, p4d); queue_pte_barriers(); }@@ -1120,7 +1121,7 @@ static inline void set_pgd(pgd_t *pgdp, pgd_t pgd) return; } - WRITE_ONCE(*pgdp, pgd); + ptdesc_set(*pgdp, pgd); queue_pte_barriers(); }diff --git a/arch/arm64/mm/mmu.c b/arch/arm64/mm/mmu.c index bcf32d1a92de..ffd307c546f5 100644 --- a/arch/arm64/mm/mmu.c +++ b/arch/arm64/mm/mmu.c@@ -83,7 +83,7 @@ void noinstr set_swapper_pgd(pgd_t *pgdp, pgd_t pgd) * writable in the kernel mapping. */ if (rodata_is_rw) { - WRITE_ONCE(*pgdp, pgd); + ptdesc_set(*pgdp, pgd); dsb(ishst); isb(); return;@@ -91,7 +91,7 @@ void noinstr set_swapper_pgd(pgd_t *pgdp, pgd_t pgd) spin_lock(&swapper_pgdir_lock); fixmap_pgdp = pgd_set_fixmap(__pa_symbol(pgdp)); - WRITE_ONCE(*fixmap_pgdp, pgd); + ptdesc_set(*fixmap_pgdp, pgd); /* * We need dsb(ishst) here to ensure the page-table-walker sees * our new entry before set_p?d() returns. The fixmap'sdiff --git a/mm/debug_vm_pgtable.c b/mm/debug_vm_pgtable.c index 83cf07269f13..faf6a19a89a1 100644 --- a/mm/debug_vm_pgtable.c +++ b/mm/debug_vm_pgtable.c@@ -445,7 +445,7 @@ static void __init pmd_huge_tests(struct pgtable_debug_args *args) * X86 defined pmd_set_huge() verifies that the given * PMD is not a populated non-leaf entry. */ - WRITE_ONCE(*args->pmdp, __pmd(0)); + ptdesc_set(*args->pmdp, __pmd(0));The ptdesc_set() and ptdesc_get() macros are defined in arch/arm64/include/asm/pgtable.h and are arm64-specific. This change is in mm/debug_vm_pgtable.c which is generic code compiled for all architectures. Other architectures do not define ptdesc_set(), so this will would cause a build failure on other architectures..Agreed. Probably will add a local fallback for ptdesc_set() which defaults to WRITE_ONCE() for platforms not overriding. OR does it require making ptdesc_get/set() much more generic construct ? Probably missed this problem while building on some other platforms as DEBUG_VM_PGTABLE might not have been enabled in their defconfig.
I don't think there is any need for WRITE_ONCE() in this file? (and I feel like this has come up before in some other context). Could you just change it to a plain C assignment? Or even better, pmd_clear()?
quoted
quoted
WARN_ON(!pmd_set_huge(args->pmdp, __pfn_to_phys(args->fixed_pmd_pfn), args->page_prot)); WARN_ON(!pmd_clear_huge(args->pmdp)); pmd = pmdp_get(args->pmdp);@@ -465,7 +465,7 @@ static void __init pud_huge_tests(struct pgtable_debug_args *args) * X86 defined pud_set_huge() verifies that the given * PUD is not a populated non-leaf entry. */ - WRITE_ONCE(*args->pudp, __pud(0)); + ptdesc_set(*args->pudp, __pud(0)); WARN_ON(!pud_set_huge(args->pudp, __pfn_to_phys(args->fixed_pud_pfn), args->page_prot)); WARN_ON(!pud_clear_huge(args->pudp)); pud = pudp_get(args->pudp);-- 2.43.0