Andrea Righi wrote: > cgroup_mm_owner_callbacks() can be called with task_lock() held in > mm_update_next_owner(), and all the .mm_owner_changed callbacks seem to > be *always* called with task_lock() held. > > Actually, memrlimit is using task_lock() via get_task_mm() in > memrlimit_cgroup_mm_owner_changed(), raising the following recursive locking > trace: [snip] Thanks for the BUG report() [snip] > static void memrlimit_cgroup_mm_owner_changed(struct cgroup_subsys *ss, > struct cgroup *old_cgrp, > @@ -246,7 +246,7 @@ static void memrlimit_cgroup_mm_owner_changed(struct cgroup_subsys *ss, > struct task_struct *p) > { > struct memrlimit_cgroup *memrcg, *old_memrcg; > - struct mm_struct *mm = get_task_mm(p); > + struct mm_struct *mm = get_task_mm_task_locked(p); > Since we hold task_lock(), we know that p->mm cannot change and we don't have to worry about incrementing mm_users. I think using just p->mm will work, we do have checks to make sure we don't pick a kernel thread. I vote for going down that road. -- Balbir _______________________________________________ Containers mailing list Containers@xxxxxxxxxxxxxxxxxxxxxxxxxx https://lists.linux-foundation.org/mailman/listinfo/containers