Re: [REGRESSION v5.13-rc1] NULL dereference in do_shrink_slab()

[NOMURA JUNICHI(野村 淳一)]

On 2021/05/13 1:31, Yang Shi wrote:
> On Wed, May 12, 2021 at 5:36 AM Shakeel Butt <shakeelb@xxxxxxxxxx> wrote:
>>
>> +Tejun Heo
>>
>> On Wed, May 12, 2021 at 3:48 AM NOMURA JUNICHI(野村　淳一)
>> <junichi.nomura@xxxxxxx> wrote:
>>> With the commit 476b30a0949a, if a memcg-aware shrinker is registered before
>>> cgroup_init(), shrinker->nr_deferred is NULL.  However xchg_nr_deferred()
>>> tries to use it as memcg is turned off via "cgroup_disable=memory".
>>>
>>> Any thoughts?
> 
> Thanks for the report.
> 
>>
>> Is there a way to find the call chain of "memcg-aware shrinker is
>> registered before cgroup_init()"?
> 
> Other than adding some printk in prealloc_memcg_shrinker() then
> checking out the output of dmesg I didn't think of a better way. Not
> sure if we have something like early trace.

This is the first registration of memcg-aware shrinker:

[    1.933693] Call Trace:
[    1.934694]  sget_fc+0x20d/0x2f0
[    1.935693]  ? compare_single+0x10/0x10
[    1.936693]  ? shmem_create+0x30/0x30
[    1.937693]  vfs_get_super+0x3e/0x100
[    1.938693]  get_tree_nodev+0x16/0x20
[    1.939693]  shmem_get_tree+0x15/0x20
[    1.940693]  vfs_get_tree+0x2a/0xc0
[    1.941693]  fc_mount+0x12/0x40
[    1.942693]  vfs_kern_mount.part.43+0x61/0xa0
[    1.943693]  kern_mount+0x24/0x40
[    1.944693]  shmem_init+0x5c/0xc8
[    1.945693]  mnt_init+0x12f/0x24a
[    1.946693]  ? __percpu_counter_init+0x8f/0xb0
[    1.947693]  vfs_caches_init+0xce/0xda
[    1.948693]  start_kernel+0x479/0x4e3
[    1.949693]  x86_64_start_reservations+0x24/0x26
[    1.950693]  x86_64_start_kernel+0x8a/0x8d
[    1.951693]  secondary_startup_64_no_verify+0xc2/0xcb

That is done after command line parsing but before cgroup_init.

>> Irrespective I think we can revert a3e72739b7a7e ("cgroup: fix too
>> early usage of static_branch_disable()") as 6041186a3258 ("init:
>> initialize jump labels before command line option parsing") has moved
>> the initialization of jump labels before command line parsing.
> 
> Seems make sense to me. If some memcg aware shrinker is registered
> before cgroup_init(), the mem_cgroup_disabled() check in
> prealloc_memcg_shrinker() would return false negative. And I don't
> think any shrinker could be registered before parsing boot
> commandline.

Thank you.  Shakeel 's patch works for me:

   [PATCH] cgroup: disable controllers at parse time
   https://lore.kernel.org/linux-mm/20210512201946.2949351-1-shakeelb@xxxxxxxxxx/

-- 
Jun'ichi Nomura, NEC Corporation / NEC Solution Innovators, Ltd.