On s390x, we actually need a pte_mkyoung() instead of a
mark_page_accessed() when doing a FOLL_TOUCH to clear the HW invalid bit
in the pte and allow subsequent accesses via the MMU to succeed without
triggering a pagefault.
Otherwise, buffered I/O will loop forever because it will keep stumlbing
over the set HW invalid bit, requiring a page fault, which is disabled.
Reported-by: Andreas Gruenbacher <agruenba@xxxxxxxxxx>
Signed-off-by: David Hildenbrand <david@xxxxxxxxxx>
---
mm/gup.c | 23 ++++++++++++++++++++---
1 file changed, 20 insertions(+), 3 deletions(-)
diff --git a/mm/gup.c b/mm/gup.c
index a9d4d724aef7..d6c65474ed72 100644
--- a/mm/gup.c
+++ b/mm/gup.c
@@ -592,10 +592,27 @@ static struct page *follow_page_pte(struct vm_area_struct *vma,
set_page_dirty(page);
/*
* pte_mkyoung() would be more correct here, but atomic care
- * is needed to avoid losing the dirty bit: it is easier to use
- * mark_page_accessed().
+ * is needed for architectures that have a hw dirty bit, to
+ * avoid losing the dirty bit: it is easier to use
+ * mark_page_accessed() for these architectures.
+ *
+ * s390x doesn't have a hw reference/dirty bit and sets the
+ * hw invalid bit in pte_mkold(), to catch further references.
+ * We have to update the pte via pte_mkyoung() here to clear the
+ * invalid bit and mark the page young; otherwise, callers that
+ * rely on not requiring a MMU fault once GUP(FOLL_TOUCH)
+ * succeeded will loop forever because the page won't be
+ * actually accessible via the MMU.
*/
- mark_page_accessed(page);
+ if (IS_ENABLED(CONFIG_S390)) {
+ pte = pte_mkyoung(pte);
+ if (!pte_same(pte, *ptep)) {
+ set_pte_at(vma->vm_mm, address, ptep, pte);
+ update_mmu_cache(vma, address, ptep);
+ }
+ } else {
+ mark_page_accessed(page);
+ }
}
if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
/* Do not mlock pte-mapped THP */
--
2.35.1
We should probably generalize this, using an ARCH config that says that we
don't have HW dirty bits and can do a pte_mkyoung() here without losing
any concurrent updates to the pte via the hw.
Further, I wonder if we might have to do a pte_mkdirty() in case of FOLL_WRITE
for these architectures as well, instead of going via the set_page_dirty().
Could be that that might be required as well here, haven't looked into the
details.
The follow_trans_huge_pmd()->touch_pmd() case should be fine I guess,
and it does both, the pmd_mkyoung and the pmd_mkdirty.
--
Thanks,
David / dhildenb
