By the time we lock a page in collapse_pte_mapped_thp(), the page
mapped by the address pushed onto the slot's .pte_mapped_thp[] array
might have changed arbitrarily since we last looked at it. We
revalidate that the page is still the head of a compound page, but we
don't revalidate if the compound page is of order HPAGE_PMD_ORDER before
applying rmap and page table updates.
Since the kernel now supports large folios of arbitrary order, and since
replacing page's pte mappings by a pmd mapping only makes sense for
compound pages of order HPAGE_PMD_ORDER, revalidate that the compound
order is indeed of order HPAGE_PMD_ORDER before proceeding.
Suggested-by: Yang Shi <shy828301@xxxxxxxxx>
Signed-off-by: Zach O'Keefe <zokeefe@xxxxxxxxxx>
---
Andrew, could you please take this into mm-unstable as a prerequisite
patch for the "mm: add file/shmem support to MADV_COLLAPSE" series?
Thank you.
mm/khugepaged.c | 5 +++++
1 file changed, 5 insertions(+)
diff --git a/mm/khugepaged.c b/mm/khugepaged.c
index 122cb72435e3..30f35d646f11 100644
--- a/mm/khugepaged.c
+++ b/mm/khugepaged.c
@@ -1448,6 +1448,11 @@ int collapse_pte_mapped_thp(struct mm_struct *mm, unsigned long addr,
goto drop_hpage;
}
+ if (compound_order(hpage) != HPAGE_PMD_ORDER) {
+ result = SCAN_PAGE_COMPOUND;
+ goto drop_hpage;
+ }
+
result = find_pmd_or_thp_or_none(mm, haddr, &pmd);
switch (result) {
case SCAN_SUCCEED:
--
2.37.3.998.g577e59143f-goog
