On 12/14/21 23:36, Davidlohr Bueso wrote:
On Tue, 14 Dec 2021, Waiman Long wrote:
@@ -2189,7 +2194,7 @@ static void drain_local_stock(struct
work_struct *dummy)
* drain_stock races is that we always operate on local CPU stock
* here with IRQ disabled
*/
- local_irq_save(flags);
+ local_lock_irqsave(&memcg_stock.lock, flags);
stock = this_cpu_ptr(&memcg_stock);
drain_obj_stock(&stock->irq_obj);
So here there is still the problem that you can end up taking sleeping
locks
with irqs disabled via obj_cgroup_put() >> obj_cgroup_release() - ie: the
percpu_ref_switch_lock and css_set_lock. It had occurred to me to promote
the former to a raw spinlock, but doubt we can get away with the latter.
Yes, it is tricky to avoid all these. One possibility to to delay
obj_cgroup_release() through RCU. However, this is outside the scope of
this patch.
Cheers,
Longman
