pmdp_collapse_flush() should be given the start address at which the huge
page is mapped, haddr: it was given addr, which at that point has been
used as a local variable, incremented to the end address of the extent.
Found by source inspection while chasing a hugepage locking bug, which
I then could not explain by this. At first I thought this was very bad;
then saw that all of the page translations that were not flushed would
actually still point to the right pages afterwards, so harmless; then
realized that I know nothing of how different architectures and models
cache intermediate paging structures, so maybe it matters after all -
particularly since the page table concerned is immediately freed.
Much easier to fix than to think about.
Fixes: 27e1f8273113 ("khugepaged: enable collapse pmd for pte-mapped THP")
Signed-off-by: Hugh Dickins <[email protected]>
Cc: [email protected] # v5.4+
---
mm/khugepaged.c | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
--- 5.8-rc7/mm/khugepaged.c 2020-07-26 16:58:02.189038680 -0700
+++ linux/mm/khugepaged.c 2020-08-02 10:48:59.890925896 -0700
@@ -1502,7 +1502,7 @@ void collapse_pte_mapped_thp(struct mm_s
/* step 4: collapse pmd */
ptl = pmd_lock(vma->vm_mm, pmd);
- _pmd = pmdp_collapse_flush(vma, addr, pmd);
+ _pmd = pmdp_collapse_flush(vma, haddr, pmd);
spin_unlock(ptl);
mm_dec_nr_ptes(mm);
pte_free(mm, pmd_pgtable(_pmd));
