Re: net: SOFTIRQ-safe -> SOFTIRQ-unsafe lock order detected in skb_array_produce
From: Dmitry Vyukov
Date: Sat Feb 18 2017 - 12:37:36 EST
On Sat, Feb 18, 2017 at 6:28 PM, Dmitry Vyukov <dvyukov@xxxxxxxxxx> wrote:
> On Fri, Feb 10, 2017 at 6:17 AM, Jason Wang <jasowang@xxxxxxxxxx> wrote:
>>
>>
>> On 2017å02æ10æ 02:10, Michael S. Tsirkin wrote:
>>>
>>> On Thu, Feb 09, 2017 at 05:02:31AM -0500, Jason Wang wrote:
>>>>
>>>> ----- Original Message -----
>>>>>
>>>>> Hello,
>>>>>
>>>>> I've got the following report while running syzkaller fuzzer on mmotm
>>>>> (git://git.kernel.org/pub/scm/linux/kernel/git/mhocko/mm.git)
>>>>> remotes/mmotm/auto-latest ee4ba7533626ba7bf2f8b992266467ac9fdc045e:
>>>>>
>>>> [...]
>>>>
>>>>> other info that might help us debug this:
>>>>>
>>>>> Possible interrupt unsafe locking scenario:
>>>>>
>>>>> CPU0 CPU1
>>>>> ---- ----
>>>>> lock(&(&r->consumer_lock)->rlock);
>>>>> local_irq_disable();
>>>>> lock(&(&r->producer_lock)->rlock);
>>>>> lock(&(&r->consumer_lock)->rlock);
>>>>> <Interrupt>
>>>>> lock(&(&r->producer_lock)->rlock);
>>>>>
>>>> Thanks a lot for the testing.
>>>>
>>>> Looks like we could address this by using skb_array_consume_bh() instead.
>>>>
>>>> Could you pls verify if the following patch works?
>>>
>>> I think we should use _bh for the produce call as well,
>>> since resizing takes the producer lock.
>>
>> Looks not since irq was disabled during resizing?
>
>
> Hello,
>
> Is there a fix for this that we can pick up?
> This killed 10'000 VMs on our testing infra over the last day. Still
> happening on linux-next.
Ah, sorry, I see the patch above with skb_array_consume_bh. It's just
that it's not in linux-next. Will manually apply it now then.
Should we also do something with produce_skb?