On Tue, Jun 14, 2022 at 12:19:24PM +0900, Sergey Senozhatsky wrote:
On (22/06/13 09:49), Minchan Kim wrote:
Many thanks for the tests.
Quite honestly I was hoping that the patch would not help :) Well, ok, we now know that it's mapping area lock and the lockdep part of its memory is zero-ed out. The question is - "why?" It really should not be zeroed out.
Ccing Mike and Sebastian who are author/expert of the culprit patch
Naresh found zsmalloc crashed on the testing [1] and confirmed that Sergey's patch[2] fixed the problem. However, I don't understand why we need reinit the local_lock on cpu_up handler[3].
Could you guys shed some light?
My guess is that it's either something very specific to Naresh's arch/config or a bug somewhere, which memset() per-CPU memory. Not sure how to track it down. KASAN maybe?
We certainly don't expect that
static DEFINE_PER_CPU(struct mapping_area, zs_map_area) = { .lock = INIT_LOCAL_LOCK(lock), };
would produce un-initialized dep_map. So I guess we start off with a valid per-CPU lock, but then it somehow gets zeroed-out.
Yes, I don't think we need to reinitialize the local_lock.
Naresh, we believe the patch Sergey provided for the test was just band aid to hide the problem.
Could you please try to bisect it?