[PATCH v2 0/2] mm: optimize HugeTLB bootmem hugepage init

[PATCH v2 0/2] mm: optimize HugeTLB bootmem hugepage init

[Li Zhe]

This series speeds up the fallback initialization path for HugeTLB tail
struct pages during boot-time hugepage allocation, when HugeTLB vmemmap
optimization (HVO) is unavailable or fails.

Patch 1 factors the existing ZONE_DEVICE template-copy operation into
memmap_init_page_from_template().

Patch 2 reuses that helper in hugetlb_folio_init_tail_vmemmap():
initialize the first tail page normally, copy it into a local template,
then initialize the remaining tail pages by refreshing only PFN-dependent
fields before copying the template.

In the earlier discussion, Mike asked whether the same kind of
template-based initialization also helps the ordinary struct page
initialization path:

https://lore.kernel.org/all/akfB6Qe-jKDEMUxz@kernel.org/

I tested that direction as well, but did not see a measurable benefit
for ordinary struct page initialization. The measurable improvement I
could find is in the HugeTLB bootmem hugepage tail path, where gigantic
pages can require initializing a large number of tail struct pages when
HVO does not cover the allocation.

Tested on an arm64 QEMU/KVM guest with 4K base pages and 64 boot-time 1G
hugepages. HVO is unavailable on arm64 in v7.4-rc1, so the fallback tail
initialization path is exercised.

Results averaged over five boots, measured with a temporary
jiffies-based printk:

  base (v7.4-rc1):  318.4 ms
  with this patch:  202.4 ms

This is a 36.43% reduction in the measured fallback tail initialization
time. The temporary measurement instrumentation is not part of this
series.

Li Zhe (2):
  mm/mm_init: factor out memmap_init_page_from_template()
  mm/hugetlb: optimize bootmem hugepage tail vmemmap init

 mm/hugetlb.c | 20 +++++++++++++++-----
 mm/mm_init.c | 34 ++++++++++++++++++++--------------
 mm/mm_init.h |  2 ++
 3 files changed, 37 insertions(+), 19 deletions(-)

---
v1: https://lore.kernel.org/all/20261105120615.16817-1-lizhe.67@bytedance.com/#t

Changes in v2:
- Add a comment for the template-before-copy ordering, as suggested by
  Muchun Song.
- Drop the empty-range guard in hugetlb_folio_init_tail_vmemmap()
  as suggested by Balbir Singh and Mike Rapoport; current callers are
  expected to pass a non-empty range.

-- 
2.20.1

[PATCH v2 1/2] mm/mm_init: factor out memmap_init_page_from_template()

[Li Zhe]

ZONE_DEVICE memmap initialization already uses a reusable struct page
template when initializing compound tail pages.  The template mostly
contains invariant state, but a few fields still depend on the target PFN
and must be refreshed before copying the template into each destination
struct page.

Factor that operation into memmap_init_page_from_template() so other
boot-time memmap initialization paths can reuse the same helper.

No functional change intended.

Signed-off-by: Li Zhe <lizhe.67@bytedance.com>
---
 mm/mm_init.c | 34 ++++++++++++++++++++--------------
 mm/mm_init.h |  2 ++
 2 files changed, 22 insertions(+), 14 deletions(-)

diff --git a/mm/mm_init.c b/mm/mm_init.c
index d99af80ec752..905222772548 100644
--- a/mm/mm_init.c
+++ b/mm/mm_init.c
@@ -625,6 +625,22 @@ void __meminit __init_single_page(struct page *page, unsigned long pfn,
 	vmemmap_optimization_verify_zone(page);
 }
 
+void __meminit memmap_init_page_from_template(struct page *page,
+		unsigned long pfn, unsigned long 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().
+	 */
+	set_page_section_from_pfn(template, pfn);
+#ifdef WANT_PAGE_VIRTUAL
+	if (!is_highmem_idx(zone))
+		set_page_address(template, __va(pfn << PAGE_SHIFT));
+#endif
+	memcpy_nontemporal(page, template, sizeof(*page));
+}
+
 #ifdef CONFIG_NUMA
 /*
  * During memory init memblocks map pfns to nids. The search is expensive and
@@ -1073,17 +1089,6 @@ static void __ref __init_zone_device_page(struct page *page, unsigned long pfn,
 	}
 }
 
-static void zone_device_page_init_from_template(struct page *page,
-		unsigned long pfn, struct page *template)
-{
-	set_page_section_from_pfn(template, pfn);
-#ifdef WANT_PAGE_VIRTUAL
-	if (!is_highmem_idx(ZONE_DEVICE))
-		set_page_address(template, __va(pfn << PAGE_SHIFT));
-#endif
-	memcpy_nontemporal(page, template, sizeof(*page));
-}
-
 /*
  * With compound page geometry and when struct pages are stored in ram most
  * tail pages are reused. Consequently, the amount of unique struct pages to
@@ -1136,8 +1141,8 @@ static void __ref memmap_init_compound(struct page *head,
 
 	/* Initialize the remaining tail pages from template. */
 	for (pfn = head_pfn + 2; pfn < end_pfn; pfn++)
-		zone_device_page_init_from_template(pfn_to_page(pfn), pfn,
-						    &template);
+		memmap_init_page_from_template(pfn_to_page(pfn), pfn,
+					       ZONE_DEVICE, &template);
 	prep_compound_head(head, order);
 }
 
@@ -1192,7 +1197,8 @@ void __ref memmap_init_zone_device(struct zone *zone,
 	for (; pfn < end_pfn; pfn += pfns_per_compound) {
 		page = pfn_to_page(pfn);
 
-		zone_device_page_init_from_template(page, pfn, &template);
+		memmap_init_page_from_template(page, pfn, ZONE_DEVICE,
+					       &template);
 
 		if (IS_ALIGNED(pfn, PAGES_PER_SECTION))
 			cond_resched();
diff --git a/mm/mm_init.h b/mm/mm_init.h
index 0034e37dbabd..22c465122b47 100644
--- a/mm/mm_init.h
+++ b/mm/mm_init.h
@@ -111,5 +111,7 @@ void memblock_free_all(void);
 
 void __meminit __init_single_page(struct page *page, unsigned long pfn,
 				unsigned long zone, int nid);
+void __meminit memmap_init_page_from_template(struct page *page,
+		unsigned long pfn, unsigned long zone, struct page *template);
 
 #endif /* __MM_MM_INIT_H */
-- 
2.20.1

[PATCH v2 2/2] mm/hugetlb: optimize bootmem hugepage tail vmemmap init

[Li Zhe]

Boot-time HugeTLB allocation has to initialize tail struct pages itself
when HugeTLB vmemmap optimization (HVO) is unavailable or fails.  For
gigantic pages, this can cover a large number of tail pages.  Most state
installed by __init_single_page() and prep_compound_tail() is identical
for every tail page in the same huge page; only PFN-dependent fields need
refreshing.

Initialize the first tail page normally, copy it into a local template,
then initialize the remaining tails from that template with
memmap_init_page_from_template().  This avoids repeatedly doing the full
__init_single_page() and prep_compound_tail() sequence for state that is
identical across tail pages in the same huge page.

On an arm64 QEMU/KVM guest using 4K pages and 64 boot-time 1G hugepages
(HVO unavailable on arm64 in v7.4-rc1), the measured fallback tail
initialization time improved by 36.43%, measured with a temporary
jiffies-based printk modeled after the existing timing in
memmap_init_zone_device():

  base (v7.4-rc1):  318.4 ms
  with this patch:  202.4 ms

Signed-off-by: Li Zhe <lizhe.67@bytedance.com>
Acked-by: Balbir Singh <balbirs@nvidia.com>
---
 mm/hugetlb.c | 20 +++++++++++++++-----
 1 file changed, 15 insertions(+), 5 deletions(-)

diff --git a/mm/hugetlb.c b/mm/hugetlb.c
index 630c96cd1e88..3f6731161d77 100644
--- a/mm/hugetlb.c
+++ b/mm/hugetlb.c
@@ -3223,6 +3223,7 @@ static void __init hugetlb_folio_init_tail_vmemmap(struct folio *folio,
 	enum zone_type zone = folio_zonenum(folio);
 	int nid = folio_nid(folio);
 	struct page *page = folio_page(folio, start_page_number);
+	struct page template;
 	unsigned long head_pfn = folio_pfn(folio);
 	unsigned long pfn, end_pfn = head_pfn + end_page_number;
 	unsigned int order = huge_page_order(h);
@@ -3231,11 +3232,20 @@ static void __init hugetlb_folio_init_tail_vmemmap(struct folio *folio,
 	 * 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);
-	}
+	pfn = head_pfn + start_page_number;
+
+	/*
+	 * Seed a template from the first real tail page and refresh only
+	 * PFN-dependent fields before copying it.
+	 */
+	__init_single_page(page, pfn, zone, nid);
+	prep_compound_tail(page, &folio->page, order);
+	set_page_count(page, 0);
+	memcpy(&template, page, sizeof(template));
+
+	for (pfn++; pfn < end_pfn; pfn++)
+		memmap_init_page_from_template(pfn_to_page(pfn), pfn, zone,
+					       &template);
 }
 
 static void __init hugetlb_folio_init_vmemmap(struct folio *folio,
-- 
2.20.1

Re: [PATCH v2 1/2] mm/mm_init: factor out memmap_init_page_from_template()

[Muchun Song]

> On Nov 10, 2026, at 19:48, Li Zhe <lizhe.67@bytedance.com> wrote:
> 
> ZONE_DEVICE memmap initialization already uses a reusable struct page
> template when initializing compound tail pages.  The template mostly
> contains invariant state, but a few fields still depend on the target PFN
> and must be refreshed before copying the template into each destination
> struct page.
> 
> Factor that operation into memmap_init_page_from_template() so other
> boot-time memmap initialization paths can reuse the same helper.
> 
> No functional change intended.
> 
> Signed-off-by: Li Zhe <lizhe.67@bytedance.com>

Reviewed-by: Muchun Song (ByteDance) <muchun.song@linux.dev>

Thanks.

Re: [PATCH v2 0/2] mm: optimize HugeTLB bootmem hugepage init

[Mike Rapoport]

Hi,

On Tue, Nov 10, 2026 at 07:48:57PM +0800, Li Zhe wrote:
> This series speeds up the fallback initialization path for HugeTLB tail
> struct pages during boot-time hugepage allocation, when HugeTLB vmemmap
> optimization (HVO) is unavailable or fails.
> 
> Patch 1 factors the existing ZONE_DEVICE template-copy operation into
> memmap_init_page_from_template().
> 
> Patch 2 reuses that helper in hugetlb_folio_init_tail_vmemmap():
> initialize the first tail page normally, copy it into a local template,
> then initialize the remaining tail pages by refreshing only PFN-dependent
> fields before copying the template.
> 
> In the earlier discussion, Mike asked whether the same kind of
> template-based initialization also helps the ordinary struct page
> initialization path:
> 
> https://lore.kernel.org/all/akfB6Qe-jKDEMUxz@kernel.org/
> 
> I tested that direction as well, but did not see a measurable benefit
> for ordinary struct page initialization. The measurable improvement I
> could find is in the HugeTLB bootmem hugepage tail path, where gigantic
> pages can require initializing a large number of tail struct pages when
> HVO does not cover the allocation.
> 
> Tested on an arm64 QEMU/KVM guest with 4K base pages and 64 boot-time 1G
> hugepages. HVO is unavailable on arm64 in v7.4-rc1, so the fallback tail
> initialization path is exercised.
> 
> Results averaged over five boots, measured with a temporary
> jiffies-based printk:
> 
>   base (v7.4-rc1):  318.4 ms
>   with this patch:  202.4 ms

Are there numbers for x86 without HVO as well?
We need to be sure this does not regress that use case.
 
> This is a 36.43% reduction in the measured fallback tail initialization
> time. The temporary measurement instrumentation is not part of this
> series.
> 
> Li Zhe (2):
>   mm/mm_init: factor out memmap_init_page_from_template()
>   mm/hugetlb: optimize bootmem hugepage tail vmemmap init
> 
>  mm/hugetlb.c | 20 +++++++++++++++-----
>  mm/mm_init.c | 34 ++++++++++++++++++++--------------
>  mm/mm_init.h |  2 ++
>  3 files changed, 37 insertions(+), 19 deletions(-)
> 
> ---
> v1: https://lore.kernel.org/all/20261105120615.16817-1-lizhe.67@bytedance.com/#t
> 
> Changes in v2:
> - Add a comment for the template-before-copy ordering, as suggested by
>   Muchun Song.
> - Drop the empty-range guard in hugetlb_folio_init_tail_vmemmap()
>   as suggested by Balbir Singh and Mike Rapoport; current callers are
>   expected to pass a non-empty range.
> 
> -- 
> 2.20.1

-- 
Sincerely yours,
Mike.