Re: [PATCH 11/20] drm/amdgpu: fix KFD eviction fence enable_signaling path

["Christian König" <christian.koenig@amd.com>]

On 11/6/25 18:07, Kuehling, Felix wrote:
> 
> On 2025-11-06 11:46, Christian König wrote:
>> On 11/6/25 17:37, Kuehling, Felix wrote:
>>> On 2025-11-06 08:43, Christian König wrote:
>>>> On 11/4/25 17:28, Philipp Stanner wrote:
>>>>> On Fri, 2025-10-31 at 14:16 +0100, Christian König wrote:
>>>>>> Calling dma_fence_is_signaled() here is illegal!
>>>>> The series was sent as a v2. But is this still an RFC?
>>>> I think when Matthew came up with the XE patches we pretty much agreed that this is the way to go.
>>>>
>>>>> If not, more detailed commit messages are a desirable thing.
>>>> Good point, how about:
>>>>
>>>> The enable_signaling callback is called with the same irqsave spinlock held than dma_fence_is_signaled() tries to grab. That will 100% reliable deadlock if that happens.
>>> I guess we could use dma_fence_is_signaled_locked instead. That said, it only tries to take the lock (in dma_fence_signal) if fence->ops->signal is set, which isn't the case for these fences. That's why this has never caused a problem up till now.
>> But when fence->ops->signal isn't set then why are we calling this?
> 
> There is no need to enable-signaling (and trigger a preemption), if the eviction fence has already signaled.

But when the evicted fence has already been signaled then enable_signaling is not called in the first place:

        if (test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &fence->flags))
                return false;

        if (!was_set && fence->ops->enable_signaling) {
                trace_dma_fence_enable_signal(fence);

                if (!fence->ops->enable_signaling(fence)) {
...

So the extra check is actually completely superfluous as far as I can see.

Regards,
Christian.

> 
> Regards,
>   Felix
> 
> 
>>
>> Regards,
>> Christian.
>>
>>> Regards,
>>>    Felix
>>>
>>>
>>>> Thanks,
>>>> Christian.
>>>>
>>>>> P.
>>>>>
>>>>>> Signed-off-by: Christian König <christian.koenig@amd.com>
>>>>>> ---
>>>>>>    drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_fence.c | 6 ------
>>>>>>    1 file changed, 6 deletions(-)
>>>>>>
>>>>>> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_fence.c
>>>>>> b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_fence.c
>>>>>> index 1ef758ac5076..09c919f72b6c 100644
>>>>>> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_fence.c
>>>>>> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_fence.c
>>>>>> @@ -120,12 +120,6 @@ static bool amdkfd_fence_enable_signaling(struct
>>>>>> dma_fence *f)
>>>>>>    {
>>>>>>        struct amdgpu_amdkfd_fence *fence =
>>>>>> to_amdgpu_amdkfd_fence(f);
>>>>>>    -    if (!fence)
>>>>>> -        return false;
>>>>>> -
>>>>>> -    if (dma_fence_is_signaled(f))
>>>>>> -        return true;
>>>>>> -
>>>>>>        if (!fence->svm_bo) {
>>>>>>            if
>>>>>> (!kgd2kfd_schedule_evict_and_restore_process(fence->mm, f))
>>>>>>                return true;