[Intel-gfx] How to convert drivers/gpu/drm/i915/ to use local workqueue?

[Tvrtko Ursulin]

On 30/06/2022 09:06, Tetsuo Handa wrote:
> On 2022/06/30 16:46, Tvrtko Ursulin wrote:
>>
>> Hi,
>>
>> On 10/06/2022 15:57, Tetsuo Handa wrote:
>>> Hello.
>>>
>>> Like commit c4f135d643823a86 ("workqueue: Wrap flush_workqueue() using
>>> a macro") explains, we are removing flush_scheduled_work() calls. And now
>>>
>>>     drivers/gpu/drm/i915/display/intel_display.c
>>>     drivers/gpu/drm/i915/gt/selftest_execlists.c
>>>
>>> are the last flush_scheduled_work() callers which have no patch proposed.
>>> I want to make a patch like
>>> https://lkml.kernel.org/r/e9b95132-89cd-5cfc-1a09-966393c5ecb0@I-love.SAKURA.ne.jp
>>> but I couldn't understand how to interpret drivers/gpu/drm/i915/ part.
>>
>> Could you provide some more context please? I did not immediately understand
>> whether the goal is remove flush_schedule_work helper with no arguments, or
>> actually stop drivers using the system work queues.
> 
> The goal is to remove flush_schedule_work().
> 
> Any kernel module is expected to stop using system workqueues if that module
> needs to call flush_scheduled_work() or flush_workqueue(system_*_wq).
> Continuing use of system workqueues is OK as long as that module does not
> need to call flush_scheduled_work() nor flush_workqueue(system_*_wq).

Could you give more context on reasoning here please? What is the 
difference between using the system_wq and flushing it from a random 
context? Or concern is about flushing from specific contexts?

> All in-tree kernel modules stopped calling flush_workqueue(system_*_wq) in 5.19.
> 
> Many of in-tree kernel modules already have patches for stop calling
> flush_scheduled_work(). But gpu/drm/i915 is one of in-tree kernel modules
> which do not have patches for stop calling flush_scheduled_work().
> 
> I want help from those who are familiar with this module.

On the i915 specifics, the caller in 
drivers/gpu/drm/i915/gt/selftest_execlists.c I am pretty sure can be 
removed. It is synchronized with the error capture side of things which 
is not required for the test to work.

I can send a patch for that or you can, as you prefer?

The drivers/gpu/drm/i915/display/intel_display.c one will have to be 
looked at by Jani or someone familiar with display code paths.

Regards,

Tvrtko