[PATCH v2] drm/amdgpu: always use legacy tlb flush on cyan_skilfish

[Christian König]

Is a single legacy flush sufficient to emulate an heavyweight flush as well?

On previous generations we needed to issue at least two legacy flushes 
for this.

And please don't push before getting an rb from Felix as well.

Regards,
Christian.


Am 14.09.23 um 11:23 schrieb Lang Yu:
> cyan_skilfish has problems with other flush types.
>
> v2: fix incorrect ternary conditional operator usage.(Yifan)
>
> Signed-off-by: Lang Yu <Lang.Yu at amd.com>
> Cc: <stable at vger.kernel.org> # v5.15+
> ---
>   drivers/gpu/drm/amd/amdgpu/gmc_v10_0.c | 7 ++++++-
>   1 file changed, 6 insertions(+), 1 deletion(-)
>
> diff --git a/drivers/gpu/drm/amd/amdgpu/gmc_v10_0.c b/drivers/gpu/drm/amd/amdgpu/gmc_v10_0.c
> index d3da13f4c80e..c6d11047169a 100644
> --- a/drivers/gpu/drm/amd/amdgpu/gmc_v10_0.c
> +++ b/drivers/gpu/drm/amd/amdgpu/gmc_v10_0.c
> @@ -236,7 +236,8 @@ static void gmc_v10_0_flush_vm_hub(struct amdgpu_device *adev, uint32_t vmid,
>   {
>   	bool use_semaphore = gmc_v10_0_use_invalidate_semaphore(adev, vmhub);
>   	struct amdgpu_vmhub *hub = &adev->vmhub[vmhub];
> -	u32 inv_req = hub->vmhub_funcs->get_invalidate_req(vmid, flush_type);
> +	u32 inv_req = hub->vmhub_funcs->get_invalidate_req(vmid,
> +		      (adev->asic_type != CHIP_CYAN_SKILLFISH) ? flush_type : 0);
>   	u32 tmp;
>   	/* Use register 17 for GART */
>   	const unsigned int eng = 17;
> @@ -331,6 +332,8 @@ static void gmc_v10_0_flush_gpu_tlb(struct amdgpu_device *adev, uint32_t vmid,
>   
>   	int r;
>   
> +	flush_type = (adev->asic_type != CHIP_CYAN_SKILLFISH) ? flush_type : 0;
> +
>   	/* flush hdp cache */
>   	adev->hdp.funcs->flush_hdp(adev, NULL);
>   
> @@ -426,6 +429,8 @@ static int gmc_v10_0_flush_gpu_tlb_pasid(struct amdgpu_device *adev,
>   	struct amdgpu_ring *ring = &adev->gfx.kiq[0].ring;
>   	struct amdgpu_kiq *kiq = &adev->gfx.kiq[0];
>   
> +	flush_type = (adev->asic_type != CHIP_CYAN_SKILLFISH) ? flush_type : 0;
> +
>   	if (amdgpu_emu_mode == 0 && ring->sched.ready) {
>   		spin_lock(&adev->gfx.kiq[0].ring_lock);
>   		/* 2 dwords flush + 8 dwords fence */