From: Gerald Schaefer Date: Tue, 29 Jul 2025 11:16:19 +0000 (+0200) Subject: s390/mm: Enable THP_SWAP and THP_MIGRATION X-Git-Tag: io_uring-6.17-20250815~27^2~3 X-Git-Url: https://git.kernel.dk/?a=commitdiff_plain;h=10dd5a0009898ba35eafeb6087e5c83b84742ff1;p=linux-block.git s390/mm: Enable THP_SWAP and THP_MIGRATION After hugetlbfs PTE_MARKER support for s390 introduced region-third and segment table swap entries, it is now possible to also enable THP_SWAP and THP_MIGRATION for s390. s390 has different layout for PTE and region / segment table entries (RSTE). This is also true for swap entries, and their swap type and offset encoding. For hugetlbfs PTE_MARKER support, s390 has internal __swp_type_rste() and __swp_offset_rste() helpers to correctly handle RSTE swap entries. But common swap code does not know about this difference, and only uses __swp_type(), __swp_offset() and __swp_entry() helpers for conversion between arch-dependent and arch-independent representation of swp_entry_t for all pagetable levels. On s390, those helpers only work for PTE swap entries. Therefore, implement __pmd_to_swp_entry() to build a fake PTE swap entry and return the arch-dependent representation of that. Correspondingly, implement __swp_entry_to_pmd() to convert that into a proper PMD swap entry again. With this, the arch-dependent swp_entry_t representation will always look like a PTE swap entry in common code. This is somewhat similar to fake PTEs in hugetlbfs code for s390, but only requires conversion of the swap type and offset, and not all the possible PTE bits. For PMD swap entry SOFT_DIRTY handling, use the same helpers as for normal PMDs. Similar to PTEs, the SOFT_DIRTY bit location is the same for swap and normal entries. Reviewed-by: Heiko Carstens Signed-off-by: Gerald Schaefer Signed-off-by: Alexander Gordeev --- diff --git a/arch/s390/Kconfig b/arch/s390/Kconfig index ac162e62da86..f41f604f0297 100644 --- a/arch/s390/Kconfig +++ b/arch/s390/Kconfig @@ -74,6 +74,7 @@ config S390 select ARCH_ENABLE_MEMORY_HOTPLUG if SPARSEMEM select ARCH_ENABLE_MEMORY_HOTREMOVE select ARCH_ENABLE_SPLIT_PMD_PTLOCK if PGTABLE_LEVELS > 2 + select ARCH_ENABLE_THP_MIGRATION if TRANSPARENT_HUGEPAGE select ARCH_HAS_CPU_FINALIZE_INIT select ARCH_HAS_CURRENT_STACK_POINTER select ARCH_HAS_DEBUG_VIRTUAL @@ -150,6 +151,7 @@ config S390 select ARCH_WANT_KERNEL_PMD_MKWRITE select ARCH_WANT_LD_ORPHAN_WARN select ARCH_WANT_OPTIMIZE_HUGETLB_VMEMMAP + select ARCH_WANTS_THP_SWAP select BUILDTIME_TABLE_SORT select CLONE_BACKWARDS2 select DCACHE_WORD_ACCESS if !KMSAN diff --git a/arch/s390/include/asm/pgtable.h b/arch/s390/include/asm/pgtable.h index 6d8bc27a366e..c1a7a92f0575 100644 --- a/arch/s390/include/asm/pgtable.h +++ b/arch/s390/include/asm/pgtable.h @@ -963,6 +963,12 @@ static inline pmd_t pmd_clear_soft_dirty(pmd_t pmd) return clear_pmd_bit(pmd, __pgprot(_SEGMENT_ENTRY_SOFT_DIRTY)); } +#ifdef CONFIG_ARCH_ENABLE_THP_MIGRATION +#define pmd_swp_soft_dirty(pmd) pmd_soft_dirty(pmd) +#define pmd_swp_mksoft_dirty(pmd) pmd_mksoft_dirty(pmd) +#define pmd_swp_clear_soft_dirty(pmd) pmd_clear_soft_dirty(pmd) +#endif + /* * query functions pte_write/pte_dirty/pte_young only work if * pte_present() is true. Undefined behaviour if not.. @@ -1979,6 +1985,45 @@ static inline unsigned long __swp_offset_rste(swp_entry_t entry) #define __rste_to_swp_entry(rste) ((swp_entry_t) { rste }) +/* + * s390 has different layout for PTE and region / segment table entries (RSTE). + * This is also true for swap entries, and their swap type and offset encoding. + * For hugetlbfs PTE_MARKER support, s390 has internal __swp_type_rste() and + * __swp_offset_rste() helpers to correctly handle RSTE swap entries. + * + * But common swap code does not know about this difference, and only uses + * __swp_type(), __swp_offset() and __swp_entry() helpers for conversion between + * arch-dependent and arch-independent representation of swp_entry_t for all + * pagetable levels. On s390, those helpers only work for PTE swap entries. + * + * Therefore, implement __pmd_to_swp_entry() to build a fake PTE swap entry + * and return the arch-dependent representation of that. Correspondingly, + * implement __swp_entry_to_pmd() to convert that into a proper PMD swap + * entry again. With this, the arch-dependent swp_entry_t representation will + * always look like a PTE swap entry in common code. + * + * This is somewhat similar to fake PTEs in hugetlbfs code for s390, but only + * requires conversion of the swap type and offset, and not all the possible + * PTE bits. + */ +static inline swp_entry_t __pmd_to_swp_entry(pmd_t pmd) +{ + swp_entry_t arch_entry; + pte_t pte; + + arch_entry = __rste_to_swp_entry(pmd_val(pmd)); + pte = mk_swap_pte(__swp_type_rste(arch_entry), __swp_offset_rste(arch_entry)); + return __pte_to_swp_entry(pte); +} + +static inline pmd_t __swp_entry_to_pmd(swp_entry_t arch_entry) +{ + pmd_t pmd; + + pmd = __pmd(mk_swap_rste(__swp_type(arch_entry), __swp_offset(arch_entry))); + return pmd; +} + extern int vmem_add_mapping(unsigned long start, unsigned long size); extern void vmem_remove_mapping(unsigned long start, unsigned long size); extern int __vmem_map_4k_page(unsigned long addr, unsigned long phys, pgprot_t prot, bool alloc);