[PATCH 1/8] drm/amdgpu: grab an additional reference on the gang fence v2

[Christian König]

Am 14.03.25 um 05:09 schrieb SRINIVASAN SHANMUGAM:
>
> On 3/7/2025 7:18 PM, Christian König wrote:
>> We keep the gang submission fence around in adev, make sure that it
>> stays alive.
>>
>> v2: fix memory leak on retry
>>
>> Signed-off-by: Christian König <christian.koenig at amd.com>
>> ---
>>   drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 10 +++++++++-
>>   1 file changed, 9 insertions(+), 1 deletion(-)
>>
>> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
>> index 198d29faa754..337543ec615c 100644
>> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
>> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
>> @@ -6889,18 +6889,26 @@ struct dma_fence *amdgpu_device_switch_gang(struct amdgpu_device *adev,
>>   {
>>       struct dma_fence *old = NULL;
>>   +    dma_fence_get(gang);
>>       do {
>>           dma_fence_put(old);
>>           old = amdgpu_device_get_gang(adev);
>>           if (old == gang)
>>               break;
>>   -        if (!dma_fence_is_signaled(old))
>> +        if (!dma_fence_is_signaled(old)) {
>
> Here, should we need to check ?

The gang is initialized to a dummy fence on bootup. So even when there is never any gang submission the old value is never NULL.

Regards,
Christian.

>
>     // Check if old fence isn't signaled
>     if (old && !dma_fence_is_signaled(old)) {
>
>> +            dma_fence_put(gang);
>>               return old;
>> +        }
>>         } while (cmpxchg((struct dma_fence __force **)&adev->gang_submit,
>>                old, gang) != old);
>>   +    /*
>> +     * Drop it once for the exchanged reference in adev and once for the
>> +     * thread local reference acquired in amdgpu_device_get_gang().
>> +     */
>> +    dma_fence_put(old);
>
> if (old)
>     dma_fence_put(old); // Ensure to release old reference  only if it is valid?
>
>
>>       dma_fence_put(old);
>>       return NULL;
>>   }