memcg,thp: fix res_counter:96 regression

Occasionally, testing memcg's move_charge_at_immigrate on rc7 shows
a flurry of hundreds of warnings at kernel/res_counter.c:96, where
res_counter_uncharge_locked() does WARN_ON(counter->usage < val).

The first trace of each flurry implicates __mem_cgroup_cancel_charge()
of mc.precharge, and an audit of mc.precharge handling points to
mem_cgroup_move_charge_pte_range()'s THP handling in commit 12724850e806
("memcg: avoid THP split in task migration").

Checking !mc.precharge is good everywhere else, when a single page is to
be charged; but here the "mc.precharge -= HPAGE_PMD_NR" likely to
follow, is liable to result in underflow (a lot can change since the
precharge was estimated).

Simply check against HPAGE_PMD_NR: there's probably a better
alternative, trying precharge for more, splitting if unsuccessful; but
this one-liner is safer for now - no kernel/res_counter.c:96 warnings
seen in 26 hours.

Signed-off-by: Hugh Dickins <hughd@google.com>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

diff --git a/mm/memcontrol.c b/mm/memcontrol.c
index b659260..7685d4a 100644 (file)
--- a/mm/memcontrol.c
+++ b/mm/memcontrol.c
@@ -5481,7 +5481,7 @@ static int mem_cgroup_move_charge_pte_range(pmd_t *pmd,
         *    part of thp split is not executed yet.
         */
        if (pmd_trans_huge_lock(pmd, vma) == 1) {
-               if (!mc.precharge) {
+               if (mc.precharge < HPAGE_PMD_NR) {
                        spin_unlock(&vma->vm_mm->page_table_lock);
                        return 0;
                }