[PATCH v2] drm/amdgpu: Ensure the DMA engine is deactivated during set ups

[Christian König]

Am 27.04.22 um 14:31 schrieb ricetons at gmail.com:
> From: Haohui Mai <haohui at alt-chain.io>
>
> The patch fully deactivates the DMA engine before setting up the ring
> buffer to avoid potential data races and crashes.
>
> Signed-off-by: Haohui Mai <haohui at alt-chain.io>
> ---
>   drivers/gpu/drm/amd/amdgpu/sdma_v5_2.c | 94 ++++++++++++++------------
>   1 file changed, 50 insertions(+), 44 deletions(-)
>
> diff --git a/drivers/gpu/drm/amd/amdgpu/sdma_v5_2.c b/drivers/gpu/drm/amd/amdgpu/sdma_v5_2.c
> index 013d2dec81d0..cb3987dd3bda 100644
> --- a/drivers/gpu/drm/amd/amdgpu/sdma_v5_2.c
> +++ b/drivers/gpu/drm/amd/amdgpu/sdma_v5_2.c
> @@ -459,7 +459,6 @@ static void sdma_v5_2_ring_emit_fence(struct amdgpu_ring *ring, u64 addr, u64 se
>   	}
>   }
>   
> -
>   /**
>    * sdma_v5_2_gfx_stop - stop the gfx async dma engines
>    *
> @@ -505,17 +504,19 @@ static void sdma_v5_2_rlc_stop(struct amdgpu_device *adev)
>   }
>   
>   /**
> - * sdma_v5_2_ctx_switch_enable - stop the async dma engines context switch
> + * sdma_v5_2_ctx_switch_enable_for_instance - start the async dma engines
> + * context switch for an instance
>    *
>    * @adev: amdgpu_device pointer
> - * @enable: enable/disable the DMA MEs context switch.
> + * @i: the index of the SDMA instance
>    *
> - * Halt or unhalt the async dma engines context switch.
> + * Unhalt the async dma engines context switch.
>    */
> -static void sdma_v5_2_ctx_switch_enable(struct amdgpu_device *adev, bool enable)
> +static void sdma_v5_2_ctx_switch_enable_for_instance(struct amdgpu_device *adev, int i)

Better name the parameter instance.

Need to give that a full review when I have time.

Thanks for taking a look into this,
Christian.

>   {
>   	u32 f32_cntl, phase_quantum = 0;
> -	int i;
> +
> +	WARN_ON(i >= adev->sdma.num_instances);
>   
>   	if (amdgpu_sdma_phase_quantum) {
>   		unsigned value = amdgpu_sdma_phase_quantum;
> @@ -539,26 +540,44 @@ static void sdma_v5_2_ctx_switch_enable(struct amdgpu_device *adev, bool enable)
>   		phase_quantum =
>   			value << SDMA0_PHASE0_QUANTUM__VALUE__SHIFT |
>   			unit  << SDMA0_PHASE0_QUANTUM__UNIT__SHIFT;
> -	}
>   
> -	for (i = 0; i < adev->sdma.num_instances; i++) {
> -		if (enable && amdgpu_sdma_phase_quantum) {
> -			WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_PHASE0_QUANTUM),
> -			       phase_quantum);
> -			WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_PHASE1_QUANTUM),
> -			       phase_quantum);
> -			WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_PHASE2_QUANTUM),
> -			       phase_quantum);
> -		}
> +		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_PHASE0_QUANTUM),
> +		       phase_quantum);
> +		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_PHASE1_QUANTUM),
> +		       phase_quantum);
> +		WREG32_SOC15_IP(GC, sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_PHASE2_QUANTUM),
> +		       phase_quantum);
> +	}
>   
> -		if (!amdgpu_sriov_vf(adev)) {
> -			f32_cntl = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CNTL));
> -			f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_CNTL,
> -					AUTO_CTXSW_ENABLE, enable ? 1 : 0);
> -			WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CNTL), f32_cntl);
> -		}
> +	if (!amdgpu_sriov_vf(adev)) {
> +		f32_cntl = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CNTL));
> +		f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_CNTL,
> +				AUTO_CTXSW_ENABLE, 1);
> +		WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CNTL), f32_cntl);
>   	}
> +}
>   
> +/**
> + * sdma_v5_2_ctx_switch_disable_all - stop the async dma engines context switch
> + *
> + * @adev: amdgpu_device pointer
> + *
> + * Halt the async dma engines context switch.
> + */
> +static void sdma_v5_2_ctx_switch_disable_all(struct amdgpu_device *adev)
> +{
> +	u32 f32_cntl;
> +	int i;
> +
> +	if (amdgpu_sriov_vf(adev))
> +		return;
> +
> +	for (i = 0; i < adev->sdma.num_instances; i++) {
> +		f32_cntl = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CNTL));
> +		f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_CNTL,
> +				AUTO_CTXSW_ENABLE, 0);
> +		WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_CNTL), f32_cntl);
> +	}
>   }
>   
>   /**
> @@ -571,21 +590,10 @@ static void sdma_v5_2_ctx_switch_enable(struct amdgpu_device *adev, bool enable)
>    */
>   static void sdma_v5_2_enable(struct amdgpu_device *adev, bool enable)
>   {
> -	u32 f32_cntl;
> -	int i;
> -
>   	if (!enable) {
>   		sdma_v5_2_gfx_stop(adev);
>   		sdma_v5_2_rlc_stop(adev);
>   	}
> -
> -	if (!amdgpu_sriov_vf(adev)) {
> -		for (i = 0; i < adev->sdma.num_instances; i++) {
> -			f32_cntl = RREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_F32_CNTL));
> -			f32_cntl = REG_SET_FIELD(f32_cntl, SDMA0_F32_CNTL, HALT, enable ? 0 : 1);
> -			WREG32(sdma_v5_2_get_reg_offset(adev, i, mmSDMA0_F32_CNTL), f32_cntl);
> -		}
> -	}
>   }
>   
>   /**
> @@ -594,6 +602,9 @@ static void sdma_v5_2_enable(struct amdgpu_device *adev, bool enable)
>    * @adev: amdgpu_device pointer
>    *
>    * Set up the gfx DMA ring buffers and enable them.
> + * It assumes that the dma engine is stopped for each instance.
> + * The function enables the engine and preemptions sequentially for each instance.
> + *
>    * Returns 0 for success, error for failure.
>    */
>   static int sdma_v5_2_gfx_resume(struct amdgpu_device *adev)
> @@ -737,10 +748,7 @@ static int sdma_v5_2_gfx_resume(struct amdgpu_device *adev)
>   
>   		ring->sched.ready = true;
>   
> -		if (amdgpu_sriov_vf(adev)) { /* bare-metal sequence doesn't need below to lines */
> -			sdma_v5_2_ctx_switch_enable(adev, true);
> -			sdma_v5_2_enable(adev, true);
> -		}
> +		sdma_v5_2_ctx_switch_enable_for_instance(adev, i);
>   
>   		r = amdgpu_ring_test_ring(ring);
>   		if (r) {
> @@ -856,7 +864,7 @@ static int sdma_v5_2_start(struct amdgpu_device *adev)
>   	int r = 0;
>   
>   	if (amdgpu_sriov_vf(adev)) {
> -		sdma_v5_2_ctx_switch_enable(adev, false);
> +		sdma_v5_2_ctx_switch_disable_all(adev);
>   		sdma_v5_2_enable(adev, false);
>   
>   		/* set RB registers */
> @@ -881,12 +889,10 @@ static int sdma_v5_2_start(struct amdgpu_device *adev)
>   		amdgpu_gfx_off_ctrl(adev, false);
>   
>   	sdma_v5_2_soft_reset(adev);
> -	/* unhalt the MEs */
> -	sdma_v5_2_enable(adev, true);
> -	/* enable sdma ring preemption */
> -	sdma_v5_2_ctx_switch_enable(adev, true);
>   
> -	/* start the gfx rings and rlc compute queues */
> +	/* Soft reset supposes to disable the dma engine and preemption.
> +	 * Now start the gfx rings and rlc compute queues.
> +	 */
>   	r = sdma_v5_2_gfx_resume(adev);
>   	if (adev->in_s0ix)
>   		amdgpu_gfx_off_ctrl(adev, true);
> @@ -1340,7 +1346,7 @@ static int sdma_v5_2_hw_fini(void *handle)
>   	if (amdgpu_sriov_vf(adev))
>   		return 0;
>   
> -	sdma_v5_2_ctx_switch_enable(adev, false);
> +	sdma_v5_2_ctx_switch_disable_all(adev);
>   	sdma_v5_2_enable(adev, false);
>   
>   	return 0;