summaryrefslogtreecommitdiff
diff options
context:
space:
mode:
-rw-r--r--mm/migrate.c19
1 files changed, 14 insertions, 5 deletions
diff --git a/mm/migrate.c b/mm/migrate.c
index a73504ff5ab..4e0eccca5e2 100644
--- a/mm/migrate.c
+++ b/mm/migrate.c
@@ -153,11 +153,6 @@ static void remove_migration_pte(struct vm_area_struct *vma,
return;
}
- if (mem_cgroup_charge(new, mm, GFP_KERNEL)) {
- pte_unmap(ptep);
- return;
- }
-
ptl = pte_lockptr(mm, pmd);
spin_lock(ptl);
pte = *ptep;
@@ -169,6 +164,20 @@ static void remove_migration_pte(struct vm_area_struct *vma,
if (!is_migration_entry(entry) || migration_entry_to_page(entry) != old)
goto out;
+ /*
+ * Yes, ignore the return value from a GFP_ATOMIC mem_cgroup_charge.
+ * Failure is not an option here: we're now expected to remove every
+ * migration pte, and will cause crashes otherwise. Normally this
+ * is not an issue: mem_cgroup_prepare_migration bumped up the old
+ * page_cgroup count for safety, that's now attached to the new page,
+ * so this charge should just be another incrementation of the count,
+ * to keep in balance with rmap.c's mem_cgroup_uncharging. But if
+ * there's been a force_empty, those reference counts may no longer
+ * be reliable, and this charge can actually fail: oh well, we don't
+ * make the situation any worse by proceeding as if it had succeeded.
+ */
+ mem_cgroup_charge(new, mm, GFP_ATOMIC);
+
get_page(new);
pte = pte_mkold(mk_pte(new, vma->vm_page_prot));
if (is_write_migration_entry(entry))