Re: + mm-memcontrol-fix-potential-oom_lock-recursion-deadlock.patch added to mm-unstable branch

[Tetsuo Handa]

On 2022/07/26 17:14, Michal Hocko wrote:
> As we have concluded there are two issues possible here which would be
> great to have reflected in the changelog.
> 
> On Mon 25-07-22 15:00:32, Andrew Morton wrote:
>> From: Tetsuo Handa <penguin-kernel@xxxxxxxxxxxxxxxxxxx>
>> Subject: mm: memcontrol: fix potential oom_lock recursion deadlock
>> Date: Fri, 22 Jul 2022 19:45:39 +0900
>>
>> syzbot is reporting GFP_KERNEL allocation with oom_lock held when
>> reporting memcg OOM [1].  Such allocation request might deadlock the
>> system, for __alloc_pages_may_oom() cannot invoke global OOM killer due to
>> oom_lock being already held by the caller.
> 
> I would phrase it like this:

This report is difficult to explain correctly.

> syzbot is reporting GFP_KERNEL allocation with oom_lock held when
> reporting memcg OOM [1].

Correct. But

>                          This is problematic because this creates a
> dependency between GFP_NOFS and GFP_KERNEL over oom_lock which could
> dead lock the system.

oom_lock is irrelevant when trying GFP_KERNEL allocation from GFP_NOFS
context. Therefore, something like:

----------
syzbot is reporting GFP_KERNEL allocation with oom_lock held when
reporting memcg OOM [1]. If this allocation triggers the global OOM
situation then the system can livelock because the GFP_KERNEL allocation
with oom_lock held cannot trigger the global OOM killer because
__alloc_pages_may_oom() fails to hold oom_lock.

Fix this problem by removing the allocation from memory_stat_format()
completely, and pass static buffer when calling from memcg OOM path.

Note that the caller holding filesystem lock was the trigger for syzbot
to report this locking dependency. Doing GFP_KERNEL allocation with
filesystem lock held can deadlock the system even without involving OOM
situation.
----------