[PATCH] drm/amdgpu: Map all visible VRAM at startup

[Christian König]

Am 27.02.2018 um 16:22 schrieb Amber Lin:
> When using CPU to update page table, we need to kmap all the PDs/PTs after
> they are allocated and that requires a TLB shot down on each CPU, which is
> quite heavy.
>
> Instead, we map the whole visible VRAM to a kernel address at once. Pages
> can be obtained from the offset.
>
> Change-Id: I56574bd544dae273da50e8b5dd6894cd5d9454bd
> Signed-off-by: Amber Lin <Amber.Lin at amd.com>
> ---
>   drivers/gpu/drm/amd/amdgpu/amdgpu_gmc.h    | 1 +
>   drivers/gpu/drm/amd/amdgpu/amdgpu_object.c | 7 +++++++
>   drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c    | 5 +++++
>   3 files changed, 13 insertions(+)
>
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_gmc.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_gmc.h
> index 893c249..503672d 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_gmc.h
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_gmc.h
> @@ -77,6 +77,7 @@ struct amdgpu_gmc_funcs {
>   struct amdgpu_gmc {
>   	resource_size_t		aper_size;
>   	resource_size_t		aper_base;
> +	void __iomem		*aper_base_kaddr;

That isn't GMC related and belongs into structure struct amdgpu_mman in 
amdgpu_ttm.h.

>   	/* for some chips with <= 32MB we need to lie
>   	 * about vram size near mc fb location */
>   	u64			mc_vram_size;
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_object.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_object.c
> index 216799cc..56c7870 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_object.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_object.c
> @@ -829,6 +829,10 @@ int amdgpu_bo_init(struct amdgpu_device *adev)
>   	/* Add an MTRR for the VRAM */
>   	adev->gmc.vram_mtrr = arch_phys_wc_add(adev->gmc.aper_base,
>   					      adev->gmc.aper_size);
> +#ifdef CONFIG_64BIT
> +	adev->gmc.aper_base_kaddr = ioremap_wc(adev->gmc.aper_base,
> +					adev->gmc.visible_vram_size);
> +#endif

Put that into amdgpu_ttm_init in amdgpu_ttm.c.

>   	DRM_INFO("Detected VRAM RAM=%lluM, BAR=%lluM\n",
>   		 adev->gmc.mc_vram_size >> 20,
>   		 (unsigned long long)adev->gmc.aper_size >> 20);
> @@ -842,6 +846,9 @@ void amdgpu_bo_fini(struct amdgpu_device *adev)
>   	amdgpu_ttm_fini(adev);
>   	arch_phys_wc_del(adev->gmc.vram_mtrr);
>   	arch_io_free_memtype_wc(adev->gmc.aper_base, adev->gmc.aper_size);
> +	if (adev->gmc.aper_base_kaddr)
> +		iounmap(adev->gmc.aper_base_kaddr);
> +	adev->gmc.aper_base_kaddr = NULL;

And that into amdgpu_ttm_fini.

>   }
>   
>   int amdgpu_bo_fbdev_mmap(struct amdgpu_bo *bo,
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c
> index e38e6db..cde639d 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c
> @@ -621,6 +621,7 @@ static int amdgpu_ttm_io_mem_reserve(struct ttm_bo_device *bdev, struct ttm_mem_
>   {
>   	struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type];
>   	struct amdgpu_device *adev = amdgpu_ttm_adev(bdev);
> +	struct drm_mm_node *mm_node = mem->mm_node;
>   
>   	mem->bus.addr = NULL;
>   	mem->bus.offset = 0;
> @@ -640,6 +641,10 @@ static int amdgpu_ttm_io_mem_reserve(struct ttm_bo_device *bdev, struct ttm_mem_
>   		/* check if it's visible */
>   		if ((mem->bus.offset + mem->bus.size) > adev->gmc.visible_vram_size)
>   			return -EINVAL;
> +		if (adev->gmc.aper_base_kaddr &&
> +			(mm_node->size == mem->num_pages)) /* contiguous mem */
> +			mem->bus.addr = ((u8 *)adev->gmc.aper_base_kaddr +
> +						mem->bus.offset);

The coding style isn't correct here. The "(mm_node->size" should be on 
the same column as the "adev->gmc" above.

And please put the comment before the if, not on the same line after it.

Regards,
Christian.

>   		mem->bus.base = adev->gmc.aper_base;
>   		mem->bus.is_iomem = true;
>   		break;