[PATCH v3 07/11] mm/hugetlb: use common compound page initialization helper
From: Muchun Song
Date: Fri Nov 20 2026 - 05:10:11 EST
Boot-time gigantic hugetlb folios currently use a separate path to
initialize their compound struct pages.
Use memmap_init_compound_page() for these folios so that HugeTLB and
ZONE_DEVICE share the common compound page initialization path. This
removes the HugeTLB-specific implementation and improves initialization
performance.
ZONE_DEVICE initializes struct pages by walking one contiguous memory
range. Boot-time HugeTLB instead initializes multiple disjoint ranges,
one for each contiguous hugepage. Measurements showed that memcpy()
performed better than memcpy_nontemporal() for the latter access pattern.
Therefore, use memcpy() for normal memory while retaining
memcpy_nontemporal() for ZONE_DEVICE.
The initialization time for 64 boot-time 1G hugepages with 4K base
pages was measured as follows:
arch without this patch with this patch reduction
x86-64 195.60 ms 170.40 ms 12.88%
arm64 374.40 ms 237.60 ms 36.54%
Co-developed-by: Li Zhe <lizhe.67@xxxxxxxxxxxxx>
Signed-off-by: Li Zhe <lizhe.67@xxxxxxxxxxxxx>
Signed-off-by: Muchun Song <songmuchun@xxxxxxxxxxxxx>
---
v3:
- Integrate the template-copy optimization into the common helper switch
(suggested by Mike Rapoport)
- Use memcpy() for normal memory and keep memcpy_nontemporal() for
ZONE_DEVICE
- Update performance results for the integrated implementation (measured
by Li Zhe)
- Add Co-developed-by for Li Zhe
---
mm/hugetlb.c | 52 ++--------------------------------------------------
mm/mm_init.c | 10 ++++++----
2 files changed, 8 insertions(+), 54 deletions(-)
diff --git a/mm/hugetlb.c b/mm/hugetlb.c
index 2a94f8f60870..6a613d9c9332 100644
--- a/mm/hugetlb.c
+++ b/mm/hugetlb.c
@@ -3319,50 +3319,6 @@ static bool __init alloc_bootmem_huge_page(struct hstate *h, int nid)
return true;
}
-/* Initialize [start_page:end_page_number] tail struct pages of a hugepage */
-static void __init hugetlb_folio_init_tail_vmemmap(struct folio *folio,
- struct hstate *h,
- unsigned long start_page_number,
- unsigned long end_page_number)
-{
- enum zone_type zone = folio_zonenum(folio);
- int nid = folio_nid(folio);
- struct page *page = folio_page(folio, start_page_number);
- unsigned long head_pfn = folio_pfn(folio);
- unsigned long pfn, end_pfn = head_pfn + end_page_number;
- unsigned int order = huge_page_order(h);
-
- /*
- * As we marked all tail pages with memblock_reserved_mark_noinit(),
- * we must initialize them ourselves here.
- */
- for (pfn = head_pfn + start_page_number; pfn < end_pfn; page++, pfn++) {
- __init_single_page(page, pfn, zone, nid);
- prep_compound_tail(page, &folio->page, order);
- set_page_count(page, 0);
- }
-}
-
-static void __init hugetlb_folio_init_vmemmap(struct folio *folio,
- unsigned long pfn,
- struct hstate *h,
- enum zone_type zone, int nid,
- unsigned long nr_pages)
-{
- struct page *head = &folio->page;
-
- /*
- * This is an open-coded prep_compound_page() whereby we avoid
- * walking pages twice by initializing/preparing+freezing them in the
- * same go.
- */
- __init_single_page(head, pfn, zone, nid);
- __folio_set_head(folio);
- set_page_count(head, 0);
- hugetlb_folio_init_tail_vmemmap(folio, h, 1, nr_pages);
- prep_compound_head(head, huge_page_order(h));
-}
-
/*
* memblock-allocated pageblocks might not have the migrate type set
* if marked with the 'noinit' flag. Set it to the default (MIGRATE_MOVABLE)
@@ -3420,8 +3376,7 @@ static void __init gather_bootmem_prealloc_node(unsigned long nid)
struct folio *folio = (void *)page;
const unsigned long pfn = PHYS_PFN(__pa(m));
const unsigned long nr_pages = pages_per_huge_page(m->hstate);
- const enum zone_type zone = zone_idx(pfn_to_zone(pfn, nid));
- unsigned long nr_struct_pages = nr_pages;
+ const struct zone *zone = pfn_to_zone(pfn, nid);
h = m->hstate;
/*
@@ -3433,10 +3388,7 @@ static void __init gather_bootmem_prealloc_node(unsigned long nid)
prev_h = h;
VM_BUG_ON(!hstate_is_gigantic(h));
-
- if (vmemmap_optimizable_order(pfn_to_section_compound_order(pfn)))
- nr_struct_pages = VMEMMAP_OPTIMIZATION_NR_STRUCT_PAGES;
- hugetlb_folio_init_vmemmap(folio, pfn, h, zone, nid, nr_struct_pages);
+ memmap_init_compound_page(page, pfn, zone, huge_page_order(h));
init_new_hugetlb_folio(folio);
if (vmemmap_optimizable_order(pfn_to_section_compound_order(pfn)))
diff --git a/mm/mm_init.c b/mm/mm_init.c
index 3e645302377e..8bf350d5a695 100644
--- a/mm/mm_init.c
+++ b/mm/mm_init.c
@@ -722,16 +722,18 @@ static void __meminit memmap_init_page_from_template(struct page *page,
unsigned long pfn, const struct zone *zone, struct page *template)
{
/*
- * Refresh PFN-dependent fields in the template before the non-temporal
- * copy, so the destination page does not need cached stores after
- * memcpy_nontemporal().
+ * Refresh PFN-dependent fields in the template before copying it, so
+ * the destination page does not need follow-up stores.
*/
set_page_section_from_pfn(template, pfn);
#ifdef WANT_PAGE_VIRTUAL
if (!is_highmem_idx(zone_idx(zone)))
set_page_address(template, __va(pfn << PAGE_SHIFT));
#endif
- memcpy_nontemporal(page, template, sizeof(*page));
+ if (zone_is_zone_device(zone))
+ memcpy_nontemporal(page, template, sizeof(*page));
+ else
+ memcpy(page, template, sizeof(*page));
vmemmap_optimization_verify_zone(page);
}
--
2.54.0