On Fri, Mar 08, 2019 at 03:10:08PM +0800, zhong jiang wrote: > I can reproduce the issue in arm64 qemu machine. The issue will leave after applying the > patch. > > Tested-by: zhong jiang <zhongjiang@xxxxxxxxxx> Thanks a lot for the quick testing! > Meanwhile, I just has a little doubt whether it is necessary to use RCU to free the task struct or not. > I think that mm->owner alway be NULL after failing to create to process. Because we call mm_clear_owner. I wish it was enough, but the problem is that the other CPU may be in the middle of get_mem_cgroup_from_mm() while this runs, and it would dereference mm->owner while it is been freed without the call_rcu affter we clear mm->owner. What prevents this race is the rcu_read_lock() in get_mem_cgroup_from_mm() and the corresponding call_rcu to free the task struct in the fork failure path (again only if CONFIG_MEMCG=y is defined). Considering you can reproduce this tiny race on arm64 qemu (perhaps tcg JIT timing variantions helps?), you might also in theory be able to still reproduce the race condition if you remove the call_rcu from delayed_free_task and you replace it with free_task.
