Re: [Intel-gfx] [4.2-rc4] acpi|drm|i915: circular locking dependency: acpi_video_get_backlight_type

[Hans de Goede <hdegoede@redhat.com>]

Hi,

On 13-08-15 16:33, Hans de Goede wrote:
> Hi,
>
> On 12-08-15 21:26, Ville Syrjälä wrote:
>> On Mon, Aug 10, 2015 at 08:29:00PM +0200, Sedat Dilek wrote:
>>> On Sat, Aug 1, 2015 at 2:23 PM, Sedat Dilek <sedat.dilek@gmail.com> wrote:
>>>> On Mon, Jul 27, 2015 at 12:33 AM, Sedat Dilek <sedat.dilek@gmail.com> wrote:
>>>>> Hi,
>>>>>
>>>>> this my first build of a 4.2-rcN Linux-kernel and I see this...
>>>>>
>>>>
>>>> Just FYI:
>>>>
>>>> I am *not* seeing this with drm-intel-nightly from below url.
>>>>
>>>> Also, I plan to test Linux v4.2-rc5.
>>>>
>>>
>>> [ CC Linus ]
>>>
>>> Knock Knock Knock.
>>>
>>> This issue still remains here (with CONFIG_DRM_I915=m)...
>>>
>>> [   18.269792] ======================================================
>>> [   18.269798] [ INFO: possible circular locking dependency detected ]
>>> [   18.269805] 4.2.0-rc6-1-iniza-small #1 Not tainted
>>> [   18.269810] -------------------------------------------------------
>>> [   18.269816] modprobe/727 is trying to acquire lock:
>>> [   18.269822]  (init_mutex){+.+.+.}, at: [<ffffffffa0090f2d>]
>>> acpi_video_get_backlight_type+0x17/0x164 [video]
>>> [   18.269840]
>>> [   18.269840] but task is already holding lock:
>>> [   18.269848]  (&(&backlight_notifier)->rwsem){++++..}, at:
>>> [<ffffffff810a6519>] __blocking_notifier_call_chain+0x39/0x70
>>> [   18.269864]
>>> [   18.269864] which lock already depends on the new lock.
>>> [   18.269864]
>>> [   18.269875]
>>> [   18.269875] the existing dependency chain (in reverse order) is:
>>> [   18.269884]
>>> ...
>>>
>>> Full dmesg log and kernel-config attached.
>>>
>>> Shall I add Rusty and modules/modprobe folks?
>>
>> Just got back from vacation and was greeted by this same lockdep splat.
>>
>> On a hunch I reverted
>>
>> commit 93a291dfaf9c328ca5a9cea1733af1a128efe890
>> Author: Hans de Goede <hdegoede@redhat.com>
>> Date:   Tue Jun 16 16:27:52 2015 +0200
>>
>>      ACPI / video: Move backlight notifier to video_detect.c
>>
>> and the problem seems to be gone. Hans, any thoughts?
>
> Looking into this atm, lockdep clearly is right.
>
> Sorry about this I have put a lot of thinking into avoiding
> these kind of issues with this patch-set, but I did not realize
> there was another lock "hiding" inside the notifier-chain.
>
> Further analysis shows that the lock inside the notifier-chain
> causes similar problems vs register_count_mutex from
> drivers/acpi/acpi_video.c. I'm working on a fix for this
> atm.

Heh, look at what I just found (I'm shifting my work focus
in the direction of nouveau) :

https://bugzilla.redhat.com/show_bug.cgi?id=1152876

So it looks like we have had the root cause of this issue
for a long time already, maybe my recent backlight selection
logic cleanup / rework has made it easier to trigger the
lockdep warning for this though.

Anyhow assuming people are ok with the fix I submitted
yesterday we've a fix for this now.

Regards,

Hans