| Message ID | 20230730151606.2871391-1-usama.arif@bytedance.com (mailing list archive) |
|---|---|
| Headers | show |
| Series | mm/memblock: Skip prep and initialization of struct pages freed later by HVO | expand |
On 30/07/2023 16:16, Usama Arif wrote: > If the region is for gigantic hugepages and if HVO is enabled, then those > struct pages which will be freed later by HVO don't need to be prepared and > initialized. This can save significant time when a large number of hugepages > are allocated at boot time. > > For a 1G hugepage, this series avoid initialization and preparation of > 262144 - 64 = 262080 struct pages per hugepage. > > When tested on a 512G system (which can allocate max 500 1G hugepages), the > kexec-boot time with HVO and DEFERRED_STRUCT_PAGE_INIT enabled without this > patchseries to running init is 3.9 seconds. With this patch it is 1.2 seconds. > This represents an approximately 70% reduction in boot time and will > significantly reduce server downtime when using a large number of > gigantic pages. > > Thanks, > Usama > There were build errors reported by kernel-bot when CONFIG_HUGETLBFS/CONFIG_HUGETLB_PAGE_OPTIMIZE_VMEMMAP is disabled due to patches 5 and 6 which should be fixed by below diff. Will wait for review and include it in next revision as its a trivial diff diff --git a/mm/hugetlb_vmemmap.h b/mm/hugetlb_vmemmap.h index 3fff6f611c19..285b59b71203 100644 --- a/mm/hugetlb_vmemmap.h +++ b/mm/hugetlb_vmemmap.h @@ -38,6 +38,8 @@ static inline unsigned int hugetlb_vmemmap_optimizable_size(const struct hstate return 0; return size > 0 ? size : 0; } + +extern bool vmemmap_optimize_enabled; #else static inline int hugetlb_vmemmap_restore(const struct hstate *h, struct page *head) { @@ -58,6 +60,8 @@ static inline bool vmemmap_should_optimize(const struct hstate *h, const struct return false; } +static bool vmemmap_optimize_enabled = false; + #endif /* CONFIG_HUGETLB_PAGE_OPTIMIZE_VMEMMAP */ static inline bool hugetlb_vmemmap_optimizable(const struct hstate *h) @@ -65,6 +69,4 @@ static inline bool hugetlb_vmemmap_optimizable(const struct hstate *h) return hugetlb_vmemmap_optimizable_size(h) != 0; } -extern bool vmemmap_optimize_enabled; - #endif /* _LINUX_HUGETLB_VMEMMAP_H */ diff --git a/mm/internal.h b/mm/internal.h index 692bb1136a39..c3321afa36cb 100644 --- a/mm/internal.h +++ b/mm/internal.h @@ -1106,7 +1106,7 @@ struct vma_prepare { #ifdef CONFIG_HUGETLBFS void __init hugetlb_hstate_alloc_gigantic_pages(void); #else -static inline void __init hugetlb_hstate_alloc_gigantic_pages(void); +static inline void __init hugetlb_hstate_alloc_gigantic_pages(void) { } #endif /* CONFIG_HUGETLBFS */ > [v1->v2]: > - (Mike Rapoport) Code quality improvements (function names, arguments, > comments). > > [RFC->v1]: > - (Mike Rapoport) Change from passing hugepage_size in > memblock_alloc_try_nid_raw for skipping struct page initialization to > using MEMBLOCK_RSRV_NOINIT flag > > > > Usama Arif (6): > mm: hugetlb: Skip prep of tail pages when HVO is enabled > mm: hugetlb_vmemmap: Use nid of the head page to reallocate it > memblock: pass memblock_type to memblock_setclr_flag > memblock: introduce MEMBLOCK_RSRV_NOINIT flag > mm: move allocation of gigantic hstates to the start of mm_core_init > mm: hugetlb: Skip initialization of struct pages freed later by HVO > > include/linux/memblock.h | 9 +++++ > mm/hugetlb.c | 71 +++++++++++++++++++++++++--------------- > mm/hugetlb_vmemmap.c | 6 ++-- > mm/hugetlb_vmemmap.h | 18 +++++++--- > mm/internal.h | 9 +++++ > mm/memblock.c | 45 +++++++++++++++++-------- > mm/mm_init.c | 6 ++++ > 7 files changed, 118 insertions(+), 46 deletions(-) >