On 12/10/21 08:01, Sebastian Andrzej Siewior wrote:

On 2021-12-09 21:52:28 [-0500], Waiman Long wrote:
…

quoted

--- a/mm/memcontrol.c
+++ b/mm/memcontrol.c

…

quoted

@@ -2210,7 +2211,7 @@ static void refill_stock(struct mem_cgroup *memcg, unsigned int nr_pages)
  	struct memcg_stock_pcp *stock;
  	unsigned long flags;
  
-	local_irq_save(flags);
+	local_lock_irqsave(&memcg_stock.lock, flags);

Why is this one using the lock? It isn't accessing irq_obj, right?

Well, the lock isn't just for irq_obj. It protects the whole memcg_stock 
structure which include irq_obj. Sometimes, data in irq_obj (or 
task_obj) will get transfer to nr_pages and vice versa. So it is easier 
to use one single lock for the whole thing.

quoted

  	stock = this_cpu_ptr(&memcg_stock);
  	if (stock->cached != memcg) { /* reset if necessary */

@@ -2779,29 +2780,28 @@ static struct mem_cgroup *get_mem_cgroup_from_objcg(struct obj_cgroup *objcg)
   * which is cheap in non-preempt kernel. The interrupt context object stock
   * can only be accessed after disabling interrupt. User context code can
   * access interrupt object stock, but not vice versa.
+ *
+ * This task and interrupt context optimization is disabled for PREEMPT_RT
+ * as there is no performance gain in this case.
   */
  static inline struct obj_stock *get_obj_stock(unsigned long *pflags)
  {
-	struct memcg_stock_pcp *stock;
-
-	if (likely(in_task())) {
+	if (likely(in_task()) && !IS_ENABLED(CONFIG_PREEMPT_RT)) {
  		*pflags = 0UL;
  		preempt_disable();
-		stock = this_cpu_ptr(&memcg_stock);
-		return &stock->task_obj;
+		return this_cpu_ptr(&memcg_stock.task_obj);
  	}

We usually add the local_lock_t to the object it protects, struct
obj_stock it this case.
That would give you two different locks (instead of one) so you wouldn't
have to use preempt_disable() to avoid lockdep's complains. Also it
would warn you if you happen to use that obj_stock in !in_task() which
is isn't possible now.
The only downside would be that drain_local_stock() needs to acquire two
locks.

As said above, having separate locks will complicate the interaction 
between irq_obj and the broader memcg_stock fields. Besides throughput 
is a less important matrix for PREEMPT_RT, so I am not trying to 
optimize throughput performance for PREEMPT_RT here.

Cheers,
Longman