Re: [PATCH] drm/amdgpu: Cancel delayed work when GFXOFF is disabled

["Lazar, Lijo" <lijo.lazar@amd.com>]

On 8/17/2021 4:36 PM, Michel Dänzer wrote:
> On 2021-08-17 12:37 p.m., Lazar, Lijo wrote:
>>
>>
>> On 8/17/2021 3:29 PM, Michel Dänzer wrote:
>>> On 2021-08-17 11:37 a.m., Lazar, Lijo wrote:
>>>>
>>>>
>>>> On 8/17/2021 2:56 PM, Michel Dänzer wrote:
>>>>> On 2021-08-17 11:12 a.m., Lazar, Lijo wrote:
>>>>>>
>>>>>>
>>>>>> On 8/17/2021 1:53 PM, Michel Dänzer wrote:
>>>>>>> From: Michel Dänzer <mdaenzer@redhat.com>
>>>>>>>
>>>>>>> schedule_delayed_work does not push back the work if it was already
>>>>>>> scheduled before, so amdgpu_device_delay_enable_gfx_off ran ~100 ms
>>>>>>> after the first time GFXOFF was disabled and re-enabled, even if GFXOFF
>>>>>>> was disabled and re-enabled again during those 100 ms.
>>>>>>>
>>>>>>> This resulted in frame drops / stutter with the upcoming mutter 41
>>>>>>> release on Navi 14, due to constantly enabling GFXOFF in the HW and
>>>>>>> disabling it again (for getting the GPU clock counter).
>>>>>>>
>>>>>>> To fix this, call cancel_delayed_work_sync when the disable count
>>>>>>> transitions from 0 to 1, and only schedule the delayed work on the
>>>>>>> reverse transition, not if the disable count was already 0. This makes
>>>>>>> sure the delayed work doesn't run at unexpected times, and allows it to
>>>>>>> be lock-free.
>>>>>>>
>>>>>>> v2:
>>>>>>> * Use cancel_delayed_work_sync & mutex_trylock instead of
>>>>>>>       mod_delayed_work.
>>>>>>> v3:
>>>>>>> * Make amdgpu_device_delay_enable_gfx_off lock-free (Christian König)
>>>>>>> v4:
>>>>>>> * Fix race condition between amdgpu_gfx_off_ctrl incrementing
>>>>>>>       adev->gfx.gfx_off_req_count and amdgpu_device_delay_enable_gfx_off
>>>>>>>       checking for it to be 0 (Evan Quan)
>>>>>>>
>>>>>>> Cc: stable@vger.kernel.org
>>>>>>> Reviewed-by: Lijo Lazar <lijo.lazar@amd.com> # v3
>>>>>>> Acked-by: Christian König <christian.koenig@amd.com> # v3
>>>>>>> Signed-off-by: Michel Dänzer <mdaenzer@redhat.com>
>>>>>>> ---
>>>>>>>
>>>>>>> Alex, probably best to wait a bit longer before picking this up. :)
>>>>>>>
>>>>>>>      drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 11 +++----
>>>>>>>      drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c    | 36 +++++++++++++++-------
>>>>>>>      2 files changed, 30 insertions(+), 17 deletions(-)
>>>>>>>
>>>>>>> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
>>>>>>> index f3fd5ec710b6..f944ed858f3e 100644
>>>>>>> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
>>>>>>> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
>>>>>>> @@ -2777,12 +2777,11 @@ static void amdgpu_device_delay_enable_gfx_off(struct work_struct *work)
>>>>>>>          struct amdgpu_device *adev =
>>>>>>>              container_of(work, struct amdgpu_device, gfx.gfx_off_delay_work.work);
>>>>>>>      -    mutex_lock(&adev->gfx.gfx_off_mutex);
>>>>>>> -    if (!adev->gfx.gfx_off_state && !adev->gfx.gfx_off_req_count) {
>>>>>>> -        if (!amdgpu_dpm_set_powergating_by_smu(adev, AMD_IP_BLOCK_TYPE_GFX, true))
>>>>>>> -            adev->gfx.gfx_off_state = true;
>>>>>>> -    }
>>>>>>> -    mutex_unlock(&adev->gfx.gfx_off_mutex);
>>>>>>> +    WARN_ON_ONCE(adev->gfx.gfx_off_state);
>>>>>>> +    WARN_ON_ONCE(adev->gfx.gfx_off_req_count);
>>>>>>> +
>>>>>>> +    if (!amdgpu_dpm_set_powergating_by_smu(adev, AMD_IP_BLOCK_TYPE_GFX, true))
>>>>>>> +        adev->gfx.gfx_off_state = true;
>>>>>>>      }
>>>>>>>        /**
>>>>>>> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c
>>>>>>> index a0be0772c8b3..b4ced45301be 100644
>>>>>>> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c
>>>>>>> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c
>>>>>>> @@ -563,24 +563,38 @@ void amdgpu_gfx_off_ctrl(struct amdgpu_device *adev, bool enable)
>>>>>>>            mutex_lock(&adev->gfx.gfx_off_mutex);
>>>>>>>      -    if (!enable)
>>>>>>> -        adev->gfx.gfx_off_req_count++;
>>>>>>> -    else if (adev->gfx.gfx_off_req_count > 0)
>>>>>>> +    if (enable) {
>>>>>>> +        /* If the count is already 0, it means there's an imbalance bug somewhere.
>>>>>>> +         * Note that the bug may be in a different caller than the one which triggers the
>>>>>>> +         * WARN_ON_ONCE.
>>>>>>> +         */
>>>>>>> +        if (WARN_ON_ONCE(adev->gfx.gfx_off_req_count == 0))
>>>>>>> +            goto unlock;
>>>>>>> +
>>>>>>>              adev->gfx.gfx_off_req_count--;
>>>>>>>      -    if (enable && !adev->gfx.gfx_off_state && !adev->gfx.gfx_off_req_count) {
>>>>>>> -        schedule_delayed_work(&adev->gfx.gfx_off_delay_work, GFX_OFF_DELAY_ENABLE);
>>>>>>> -    } else if (!enable && adev->gfx.gfx_off_state) {
>>>>>>> -        if (!amdgpu_dpm_set_powergating_by_smu(adev, AMD_IP_BLOCK_TYPE_GFX, false)) {
>>>>>>> -            adev->gfx.gfx_off_state = false;
>>>>>>> +        if (adev->gfx.gfx_off_req_count == 0 && !adev->gfx.gfx_off_state)
>>>>>>> +            schedule_delayed_work(&adev->gfx.gfx_off_delay_work, GFX_OFF_DELAY_ENABLE);
>>>>>>> +    } else {
>>>>>>> +        if (adev->gfx.gfx_off_req_count == 0) {
>>>>>>> +            cancel_delayed_work_sync(&adev->gfx.gfx_off_delay_work);
>>>>>>> +
>>>>>>> +            if (adev->gfx.gfx_off_state &&
>>>>>>
>>>>>> More of a question which I didn't check last time - Is this expected to be true when the disable call comes in first?
>>>>>
>>>>> My assumption is that cancel_delayed_work_sync guarantees amdgpu_device_delay_enable_gfx_off's assignment is visible here.
>>>>>
>>>>
>>>> To clarify - when nothing is scheduled. If enable() is called when the count is 0, it goes to unlock. Now the expectation is someone to call Disable first.
>>>
>>> Yes, the very first amdgpu_gfx_off_ctrl call must pass enable=false, or it's a bug, which
>>>
>>>           if (WARN_ON_ONCE(adev->gfx.gfx_off_req_count == 0))
>>>
>>> will catch.
>>>
>>>
>>>> Let's say  Disable() is called first, then the variable will be false, right?
>>>
>>> Ohh, I see what you mean. The first time amdgpu_gfx_off_ctrl is called with enable=false, adev->gfx.gfx_off_state == false (what it was initialized to), so it doesn't actually disable GFXOFF in HW.
>>
>> Exactly.
> 
> Turns out that's not the end of that rabbit (side-)hole yet. :)
> 
> amdgpu_device_init initializes adev->gfx.gfx_off_req_count = 1. amdgpu_gfx_off_ctrl is then called with enable=true from amdgpu_device_init → amdgpu_device_ip_late_init → amdgpu_device_set_pg_state. This schedules amdgpu_device_delay_enable_gfx_off, which runs ~100ms later, enables GFXOFF in the HW and sets adev->gfx.gfx_off_state = true.
> 

What if a disable comes at < 100ms? Quite unlikely, neverthless in that 
case pending work will get cancelled and the variable won't be set until 
the work gets a chance to fully run. The assumption that GFXOFF disable 
succeeded in a subsequent amdgpu_gfx_off_ctrl  enable = false won't be 
correct as PMFW will by default enable GFXOFF when there is no activity.

Otherwise, keep an assumption that amdgpu_device_delay_enable_gfx_off 
gets a chance to run before any disable call comes - maybe that's the 
case in most cases.

> So it looks fine as is actually, if a bit convoluted. 

> (I wonder if GFXOFF shouldn't rather be enabled synchronously during initialization though)

Yes, that is logical. But amdgpu_device_ip_late_init is called also 
during amdgpu_device_resume. amdgpu_device_resume is used in pm_ops or 
runtime pm. In those cases it makes sense to delay it as there could be 
an immediate usage of GFX.

Thanks,
Lijo

> 
>