[PATCH v4 1/4] drm/amdgpu/vcn: fix race condition issue for vcn start

[Christian König]

Am 11.03.20 um 15:15 schrieb James Zhu:
> Fix race condition issue when multiple vcn starts are called.
>
> v2: Removed checking the return value of cancel_delayed_work_sync()
> to prevent possible races here.
>
> v3: Add total_submission_cnt to avoid gate power unexpectedly.
>
> v4: Remove extra counter check, and reduce counter before idle
> work schedule
>
> Signed-off-by: James Zhu <James.Zhu at amd.com>
> ---
>   drivers/gpu/drm/amd/amdgpu/amdgpu_vcn.c | 22 +++++++++++++++-------
>   drivers/gpu/drm/amd/amdgpu/amdgpu_vcn.h |  2 ++
>   2 files changed, 17 insertions(+), 7 deletions(-)
>
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vcn.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_vcn.c
> index a41272f..2fa2891 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vcn.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vcn.c
> @@ -63,6 +63,8 @@ int amdgpu_vcn_sw_init(struct amdgpu_device *adev)
>   	int i, r;
>   
>   	INIT_DELAYED_WORK(&adev->vcn.idle_work, amdgpu_vcn_idle_work_handler);
> +	mutex_init(&adev->vcn.vcn_pg_lock);
> +	atomic_set(&adev->vcn.total_submission_cnt, 0);
>   
>   	switch (adev->asic_type) {
>   	case CHIP_RAVEN:
> @@ -210,6 +212,7 @@ int amdgpu_vcn_sw_fini(struct amdgpu_device *adev)
>   	}
>   
>   	release_firmware(adev->vcn.fw);
> +	mutex_destroy(&adev->vcn.vcn_pg_lock);
>   
>   	return 0;
>   }
> @@ -307,7 +310,8 @@ static void amdgpu_vcn_idle_work_handler(struct work_struct *work)
>   		fences += fence[j];
>   	}
>   
> -	if (fences == 0) {
> +	if (fences == 0 &&
> +		likely(atomic_read(&adev->vcn.total_submission_cnt) == 0)) {

The indentation here looks off, maybe just write that as !fences && 
!atomic_read(&adev->vcn.total_submission_cnt).

Apart from that looks good to me now,
Christian.

>   		amdgpu_gfx_off_ctrl(adev, true);
>   		amdgpu_device_ip_set_powergating_state(adev, AMD_IP_BLOCK_TYPE_VCN,
>   		       AMD_PG_STATE_GATE);
> @@ -319,13 +323,14 @@ static void amdgpu_vcn_idle_work_handler(struct work_struct *work)
>   void amdgpu_vcn_ring_begin_use(struct amdgpu_ring *ring)
>   {
>   	struct amdgpu_device *adev = ring->adev;
> -	bool set_clocks = !cancel_delayed_work_sync(&adev->vcn.idle_work);
>   
> -	if (set_clocks) {
> -		amdgpu_gfx_off_ctrl(adev, false);
> -		amdgpu_device_ip_set_powergating_state(adev, AMD_IP_BLOCK_TYPE_VCN,
> -		       AMD_PG_STATE_UNGATE);
> -	}
> +	atomic_inc(&adev->vcn.total_submission_cnt);
> +	cancel_delayed_work_sync(&adev->vcn.idle_work);
> +
> +	mutex_lock(&adev->vcn.vcn_pg_lock);
> +	amdgpu_gfx_off_ctrl(adev, false);
> +	amdgpu_device_ip_set_powergating_state(adev, AMD_IP_BLOCK_TYPE_VCN,
> +	       AMD_PG_STATE_UNGATE);
>   
>   	if (adev->pg_flags & AMD_PG_SUPPORT_VCN_DPG)	{
>   		struct dpg_pause_state new_state;
> @@ -345,10 +350,13 @@ void amdgpu_vcn_ring_begin_use(struct amdgpu_ring *ring)
>   
>   		adev->vcn.pause_dpg_mode(adev, ring->me, &new_state);
>   	}
> +	mutex_unlock(&adev->vcn.vcn_pg_lock);
>   }
>   
>   void amdgpu_vcn_ring_end_use(struct amdgpu_ring *ring)
>   {
> +	atomic_dec(&ring->adev->vcn.total_submission_cnt);
> +
>   	schedule_delayed_work(&ring->adev->vcn.idle_work, VCN_IDLE_TIMEOUT);
>   }
>   
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vcn.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_vcn.h
> index 6fe0573..111c4cc 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vcn.h
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vcn.h
> @@ -200,6 +200,8 @@ struct amdgpu_vcn {
>   	struct drm_gpu_scheduler *vcn_dec_sched[AMDGPU_MAX_VCN_INSTANCES];
>   	uint32_t		 num_vcn_enc_sched;
>   	uint32_t		 num_vcn_dec_sched;
> +	struct mutex		 vcn_pg_lock;
> +	atomic_t		 total_submission_cnt;
>   
>   	unsigned	harvest_config;
>   	int (*pause_dpg_mode)(struct amdgpu_device *adev,