On Jul 6, 2026, at 18:07, Michał Cłapiński [off-list ref] wrote:

On 6/12/26 5:58 AM, Muchun Song wrote:

quoted

Commit 622026e87c40 ("mm/hugetlb: remove fake head pages") switched
HVO to reuse per-zone shared tail pages from zone->vmemmap_tails[].
Those shared tail pages were initialized in hugetlb_vmemmap_init(), but
bootmem HugeTLB folios are prepared earlier from gather_bootmem_prealloc().
With hugetlb_free_vmemmap=on, prep_and_add_bootmem_folios() can access
pageblock flags on bootmem HugeTLB pages whose mirrored tail struct pages
already point to the shared tail page. On CONFIG_DEBUG_VM kernels,
get_pfnblock_bitmap_bitidx() then dereferences the still-uninitialized
shared tail page and can panic during boot.
Initialize zone->vmemmap_tails[] from gather_bootmem_prealloc(), before
bootmem HugeTLB folios are processed, and drop the later initialization
from hugetlb_vmemmap_init().
This bug only affects CONFIG_DEBUG_VM kernels, where the relevant
assertion is evaluated.
Fixes: 622026e87c40 ("mm/hugetlb: remove fake head pages")
Signed-off-by: Muchun Song <redacted>
Acked-by: Oscar Salvador <osalvador@suse.de>
---
 mm/hugetlb.c         | 25 +++++++++++++++++++++++++
 mm/hugetlb_vmemmap.c | 17 -----------------
 mm/sparse-vmemmap.c  |  2 +-
 3 files changed, 26 insertions(+), 18 deletions(-)

diff --git a/mm/hugetlb.c b/mm/hugetlb.c
index 571212b80835..cd55524c7e30 100644
--- a/mm/hugetlb.c
+++ b/mm/hugetlb.c

@@ -3365,6 +3365,31 @@ static void __init gather_bootmem_prealloc(void)
  .max_threads = num_node_state(N_MEMORY),
  .numa_aware = true,
  };
+#ifdef CONFIG_HUGETLB_PAGE_OPTIMIZE_VMEMMAP
+ struct zone *zone;
+
+ for_each_zone(zone) {
+ for (int i = 0; i < NR_VMEMMAP_TAILS; i++) {
+ struct page *tail, *p;
+ unsigned int order;
+
+ tail = zone->vmemmap_tails[i];
+ if (!tail)
+ continue;
+
+ order = i + VMEMMAP_TAIL_MIN_ORDER;
+ p = page_to_virt(tail);
+ /*
+ * prep_and_add_bootmem_folios() can access pageblock
+ * flags on bootmem HugeTLB pages, so initialize the
+ * shared tail struct pages here before bootmem folios
+ * start using them.
+ */
+ for (int j = 0; j < PAGE_SIZE / sizeof(struct page); j++)
+ init_compound_tail(p + j, NULL, order, zone);
+ }
+ }
+#endif

I think it would be better if this was located in mm/hugetlb_vmemmap.c as a separate function.

This is actually just temporary. I already have another patch [1] that
completely removes this code later on. So, to avoid having to update
the entire patchset unnecessarily, I figured it's best to just leave
it as is for now, as it's not a major issue..

[1] https://lore.kernel.org/all/20260702093821.2740183-10-songmuchun@bytedance.com/ (local)

quoted

    padata_do_multithreaded(&job);
 }

diff --git a/mm/hugetlb_vmemmap.c b/mm/hugetlb_vmemmap.c
index 133b46dfb09f..c713c0d2593a 100644
--- a/mm/hugetlb_vmemmap.c
+++ b/mm/hugetlb_vmemmap.c

@@ -870,27 +870,10 @@ static const struct ctl_table hugetlb_vmemmap_sysctls[] = {
 static int __init hugetlb_vmemmap_init(void)
 {
  const struct hstate *h;
- struct zone *zone;
    /* HUGETLB_VMEMMAP_RESERVE_SIZE should cover all used struct pages */
  BUILD_BUG_ON(__NR_USED_SUBPAGE > HUGETLB_VMEMMAP_RESERVE_PAGES);
 - for_each_zone(zone) {
- for (int i = 0; i < NR_VMEMMAP_TAILS; i++) {
- struct page *tail, *p;
- unsigned int order;
-
- tail = zone->vmemmap_tails[i];
- if (!tail)
- continue;
-
- order = i + VMEMMAP_TAIL_MIN_ORDER;
- p = page_to_virt(tail);
- for (int j = 0; j < PAGE_SIZE / sizeof(struct page); j++)
- init_compound_tail(p + j, NULL, order, zone);
- }
- }
-
  for_each_hstate(h) {
  if (hugetlb_vmemmap_optimizable(h)) {
  register_sysctl_init("vm", hugetlb_vmemmap_sysctls);

diff --git a/mm/sparse-vmemmap.c b/mm/sparse-vmemmap.c
index 99e2be39671b..bb23fb3077a3 100644
--- a/mm/sparse-vmemmap.c
+++ b/mm/sparse-vmemmap.c

@@ -342,7 +342,7 @@ static __meminit struct page *vmemmap_get_tail(unsigned int order, struct zone *
  *
  * Any initialization done here will be overwritten by memmap_init().
  *
- * hugetlb_vmemmap_init() will take care of initialization after
+ * gather_bootmem_prealloc() will take care of initialization after
  * memmap_init().
  */
 

It gets the job done.

Tested-by: Michal Clapinski <redacted>
Reviewed-by: Michal Clapinski <redacted>

Really thanks for your test.

Muchun,
Thanks.