[PATCH v3] drm/amdgpu: Don't flush/invalidate HDP for APUs and A+A

[Christian König]

Am 02.06.21 um 04:04 schrieb Eric Huang:
> Integrate two macros into two generic functions and add
> no_flush flag to determine if HDP flush is needed for
> all Asics.

Yes that starts looks like it should work, just a few comments below.

>
> Signed-off-by: Eric Huang <jinhuieric.huang at amd.com>
> ---
>   drivers/gpu/drm/amd/amdgpu/amdgpu.h        |  9 ++++---
>   drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 29 ++++++++++++++++++++++
>   drivers/gpu/drm/amd/amdgpu/amdgpu_hdp.h    |  1 +
>   drivers/gpu/drm/amd/amdgpu/amdgpu_ib.c     | 15 ++---------
>   4 files changed, 37 insertions(+), 17 deletions(-)
>
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu.h b/drivers/gpu/drm/amd/amdgpu/amdgpu.h
> index 1e3cd4ce9e91..224552d38240 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu.h
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu.h
> @@ -1297,10 +1297,6 @@ int emu_soc_asic_init(struct amdgpu_device *adev);
>   #define amdgpu_asic_read_bios_from_rom(adev, b, l) (adev)->asic_funcs->read_bios_from_rom((adev), (b), (l))
>   #define amdgpu_asic_read_register(adev, se, sh, offset, v)((adev)->asic_funcs->read_register((adev), (se), (sh), (offset), (v)))
>   #define amdgpu_asic_get_config_memsize(adev) (adev)->asic_funcs->get_config_memsize((adev))
> -#define amdgpu_asic_flush_hdp(adev, r) \
> -	((adev)->asic_funcs->flush_hdp ? (adev)->asic_funcs->flush_hdp((adev), (r)) : (adev)->hdp.funcs->flush_hdp((adev), (r)))
> -#define amdgpu_asic_invalidate_hdp(adev, r) \
> -	((adev)->asic_funcs->invalidate_hdp ? (adev)->asic_funcs->invalidate_hdp((adev), (r)) : (adev)->hdp.funcs->invalidate_hdp((adev), (r)))
>   #define amdgpu_asic_need_full_reset(adev) (adev)->asic_funcs->need_full_reset((adev))
>   #define amdgpu_asic_init_doorbell_index(adev) (adev)->asic_funcs->init_doorbell_index((adev))
>   #define amdgpu_asic_get_pcie_usage(adev, cnt0, cnt1) ((adev)->asic_funcs->get_pcie_usage((adev), (cnt0), (cnt1)))
> @@ -1314,6 +1310,11 @@ int emu_soc_asic_init(struct amdgpu_device *adev);
>   
>   #define amdgpu_inc_vram_lost(adev) atomic_inc(&((adev)->vram_lost_counter));
>   
> +void amdgpu_asic_flush_hdp(struct amdgpu_device *adev,
> +		struct amdgpu_ring *ring);
> +void amdgpu_asic_invalidate_hdp(struct amdgpu_device *adev,
> +		struct amdgpu_ring *ring);
> +
>   /* Common functions */
>   bool amdgpu_device_has_job_running(struct amdgpu_device *adev);
>   bool amdgpu_device_should_recover_gpu(struct amdgpu_device *adev);
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
> index 9c4d33f649f7..951feefb5e76 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
> @@ -3505,6 +3505,9 @@ int amdgpu_device_init(struct amdgpu_device *adev,
>   	if (r)
>   		goto failed_unmap;
>   
> +	if ((adev->flags & AMD_IS_APU) || adev->gmc.xgmi.connected_to_cpu)
> +		adev->hdp.no_flush = true;
> +

This is not correct. We can only skip HDP flush on X86_64, but not on 
32bit builds.

Additional to that we might want this somewhere in the GMC code and not 
here.

>   	/* doorbell bar mapping and doorbell index init*/
>   	amdgpu_device_doorbell_init(adev);
>   
> @@ -5548,3 +5551,29 @@ bool amdgpu_device_is_headless(struct amdgpu_device *adev)
>       /* Check if it is NV's VGA class while VCN is harvest, it is headless*/
>       return nv_is_headless_sku(adev->pdev);
>   }
> +
> +void amdgpu_asic_flush_hdp(struct amdgpu_device *adev,
> +		struct amdgpu_ring *ring)

Please name the functions amdgpu_device_, the amdgpu_asic_ names are for 
callbacks only.

> +{
> +	if (adev->hdp.no_flush)
> +		return;

Since we now have a function to check the different conditions I think 
we might want to check for X86_64 & APU and XGMI connection separately here.

But that is not a hard requirement, using the no_flush variable is fine 
with me if you think that this is better.

> +
> +	if (ring->funcs->emit_hdp_flush)
> +		amdgpu_ring_emit_hdp_flush(ring);
> +	else if (adev->asic_funcs->flush_hdp)
> +		adev->asic_funcs->flush_hdp(adev, ring);
> +	else
> +		adev->hdp.funcs->flush_hdp(adev, ring);
> +}
> +
> +void amdgpu_asic_invalidate_hdp(struct amdgpu_device *adev,
> +		struct amdgpu_ring *ring)
> +{
> +	if (adev->hdp.no_flush)
> +		return;
> +
> +	if (adev->asic_funcs->invalidate_hdp)
> +		adev->asic_funcs->invalidate_hdp(adev, ring);
> +	else
> +		adev->hdp.funcs->invalidate_hdp(adev, ring);
> +}
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_hdp.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_hdp.h
> index 7ec99d591584..1ca23f2f51d2 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_hdp.h
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_hdp.h
> @@ -44,6 +44,7 @@ struct amdgpu_hdp {
>   	struct ras_common_if			*ras_if;
>   	const struct amdgpu_hdp_funcs		*funcs;
>   	const struct amdgpu_hdp_ras_funcs	*ras_funcs;
> +	bool	no_flush;
>   };
>   
>   int amdgpu_hdp_ras_late_init(struct amdgpu_device *adev);
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ib.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_ib.c
> index aaa2574ce9bc..6db676c4c7e1 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ib.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ib.c
> @@ -222,15 +222,7 @@ int amdgpu_ib_schedule(struct amdgpu_ring *ring, unsigned num_ibs,
>   	if (job && ring->funcs->init_cond_exec)
>   		patch_offset = amdgpu_ring_init_cond_exec(ring);
>   
> -#ifdef CONFIG_X86_64
> -	if (!(adev->flags & AMD_IS_APU))
> -#endif
> -	{
> -		if (ring->funcs->emit_hdp_flush)
> -			amdgpu_ring_emit_hdp_flush(ring);
> -		else
> -			amdgpu_asic_flush_hdp(adev, ring);
> -	}
> +	amdgpu_asic_flush_hdp(adev, ring);

We also have two more HDP flush cases in the GART code and the CPU based 
VM code IIRC.

Thanks,
Christian.

>   
>   	if (need_ctx_switch)
>   		status |= AMDGPU_HAVE_CTX_SWITCH;
> @@ -267,10 +259,7 @@ int amdgpu_ib_schedule(struct amdgpu_ring *ring, unsigned num_ibs,
>   	if (job && ring->funcs->emit_frame_cntl)
>   		amdgpu_ring_emit_frame_cntl(ring, false, secure);
>   
> -#ifdef CONFIG_X86_64
> -	if (!(adev->flags & AMD_IS_APU))
> -#endif
> -		amdgpu_asic_invalidate_hdp(adev, ring);
> +	amdgpu_asic_invalidate_hdp(adev, ring);
>   
>   	if (ib->flags & AMDGPU_IB_FLAG_TC_WB_NOT_INVALIDATE)
>   		fence_flags |= AMDGPU_FENCE_FLAG_TC_WB_ONLY;