Re: general protection fault in wb_workfn (2)

[Jens Axboe]

On 5/31/18 7:42 AM, Jan Kara wrote:
> On Thu 31-05-18 22:19:44, Tetsuo Handa wrote:
>> On 2018/05/31 20:42, Jan Kara wrote:
>>> On Thu 31-05-18 01:00:08, Tetsuo Handa wrote:
>>>> So, we have no idea what is happening...
>>>> Then, what about starting from temporary debug printk() patch shown below?
>>>>
>>>> >From 4f70f72ad3c9ae6ce1678024ef740aca4958e5b0 Mon Sep 17 00:00:00 2001
>>>> From: Tetsuo Handa <penguin-kernel@xxxxxxxxxxxxxxxxxxx>
>>>> Date: Wed, 30 May 2018 09:57:10 +0900
>>>> Subject: [PATCH] bdi: Add temporary config for debugging wb_workfn() versus
>>>>  bdi_unregister() race bug.
>>>>
>>>> syzbot is hitting NULL pointer dereference at wb_workfn() [1]. But due to
>>>> limitations that syzbot cannot find reproducer for this bug (frequency is
>>>> once or twice per a day) nor we can't capture vmcore in the environment
>>>> which syzbot is using, for now we need to rely on printk() debugging.
>>>>
>>>> [1] https://syzkaller.appspot.com/bug?id=e0818ccb7e46190b3f1038b0c794299208ed4206
>>>>
>>>> Signed-off-by: Tetsuo Handa <penguin-kernel@xxxxxxxxxxxxxxxxxxx>
>>>
>>> Hum a bit ugly solution but if others are fine with this, I can live with
>>> it for a while as well. Or would it be possible for syzkaller to just test
>>> some git tree where this patch is included? Then we would not even have to
>>> have the extra config option...
>>
>> If syzbot can reproduce this bug that way. While it is possible to add/remove
>> git trees syzbot tests, frequently adding/removing trees is bothering.
>>
>> syzbot can enable extra config option. Maybe the config name should be
>> something like CONFIG_DEBUG_FOR_SYZBOT rather than individual topic.
>>
>> I think that syzbot is using many VM instances. I don't know how many
>> instances will be needed for reproducing this bug within reasonable period.
>> More git trees syzbot tests, (I assume that) longer period will be needed
>> for reproducing this bug. The most reliable way is to use the shared part
>> of all trees (i.e. linux.git).
> 
> I understand this, I'd be just a bit reluctant to merge temporary debug
> patches like this to Linus' tree only to revert them later just because
> syzkaller... What do others think?

I guess I don't understand why having it in Linus's tree would make
a difference to syzkaller?

If there is a compelling reason why that absolutely has to be done,
I don't think it should be accompanied by a special Kconfig option
for it. It should just be on unconditionally, with the intent to
remove it before release.

-- 
Jens Axboe