[PATCH v2] drm/amdgpu: Map all visible VRAM at startup

[PATCH v2] drm/amdgpu: Map all visible VRAM at startup

[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.

v2: move the mapping base from gmc to amdgpu_mman structure, and the
    implementation in amdgpu_ttm_* functions

Change-Id: I56574bd544dae273da50e8b5dd6894cd5d9454bd
Signed-off-by: Amber Lin <Amber.Lin@amd.com>
---
 drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c | 17 +++++++++++++++++
 drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.h |  1 +
 2 files changed, 18 insertions(+)

diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c
index e38e6db..f126a5a 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 +641,15 @@ 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;
+		/* Only physically contiguous buffers apply. In a contiguous
+		 * buffer, size of the first mm_node would match the number of
+		 * pages in ttm_mem_reg.
+		 */
+		if (adev->mman.aper_base_kaddr &&
+		    (mm_node->size == mem->num_pages))
+			mem->bus.addr = (u8 *)adev->mman.aper_base_kaddr +
+					mem->bus.offset;
+
 		mem->bus.base = adev->gmc.aper_base;
 		mem->bus.is_iomem = true;
 		break;
@@ -1402,6 +1412,10 @@ int amdgpu_ttm_init(struct amdgpu_device *adev)
 
 	/* Change the size here instead of the init above so only lpfn is affected */
 	amdgpu_ttm_set_active_vram_size(adev, adev->gmc.visible_vram_size);
+#ifdef CONFIG_64BIT
+	adev->mman.aper_base_kaddr = ioremap_wc(adev->gmc.aper_base,
+						adev->gmc.visible_vram_size);
+#endif
 
 	/*
 	 *The reserved vram for firmware must be pinned to the specified
@@ -1494,6 +1508,9 @@ void amdgpu_ttm_fini(struct amdgpu_device *adev)
 	amdgpu_ttm_debugfs_fini(adev);
 	amdgpu_bo_free_kernel(&adev->stolen_vga_memory, NULL, NULL);
 	amdgpu_ttm_fw_reserve_vram_fini(adev);
+	if (adev->mman.aper_base_kaddr)
+		iounmap(adev->mman.aper_base_kaddr);
+	adev->mman.aper_base_kaddr = NULL;
 
 	ttm_bo_clean_mm(&adev->mman.bdev, TTM_PL_VRAM);
 	ttm_bo_clean_mm(&adev->mman.bdev, TTM_PL_TT);
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.h
index 1e275c7..d314910 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.h
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.h
@@ -44,6 +44,7 @@ struct amdgpu_mman {
 	struct ttm_bo_device		bdev;
 	bool				mem_global_referenced;
 	bool				initialized;
+	void __iomem			*aper_base_kaddr;
 
 #if defined(CONFIG_DEBUG_FS)
 	struct dentry			*debugfs_entries[8];
-- 
2.7.4

_______________________________________________
amd-gfx mailing list
amd-gfx@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

Re: [PATCH v2] drm/amdgpu: Map all visible VRAM at startup

[Christian König]

Am 27.02.2018 um 19:36 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.
>
> v2: move the mapping base from gmc to amdgpu_mman structure, and the
>      implementation in amdgpu_ttm_* functions
>
> Change-Id: I56574bd544dae273da50e8b5dd6894cd5d9454bd
> Signed-off-by: Amber Lin <Amber.Lin@amd.com>

Reviewed-by: Christian König <christian.koenig@amd.com>

> ---
>   drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c | 17 +++++++++++++++++
>   drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.h |  1 +
>   2 files changed, 18 insertions(+)
>
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c
> index e38e6db..f126a5a 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 +641,15 @@ 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;
> +		/* Only physically contiguous buffers apply. In a contiguous
> +		 * buffer, size of the first mm_node would match the number of
> +		 * pages in ttm_mem_reg.
> +		 */
> +		if (adev->mman.aper_base_kaddr &&
> +		    (mm_node->size == mem->num_pages))
> +			mem->bus.addr = (u8 *)adev->mman.aper_base_kaddr +
> +					mem->bus.offset;
> +
>   		mem->bus.base = adev->gmc.aper_base;
>   		mem->bus.is_iomem = true;
>   		break;
> @@ -1402,6 +1412,10 @@ int amdgpu_ttm_init(struct amdgpu_device *adev)
>   
>   	/* Change the size here instead of the init above so only lpfn is affected */
>   	amdgpu_ttm_set_active_vram_size(adev, adev->gmc.visible_vram_size);
> +#ifdef CONFIG_64BIT
> +	adev->mman.aper_base_kaddr = ioremap_wc(adev->gmc.aper_base,
> +						adev->gmc.visible_vram_size);
> +#endif
>   
>   	/*
>   	 *The reserved vram for firmware must be pinned to the specified
> @@ -1494,6 +1508,9 @@ void amdgpu_ttm_fini(struct amdgpu_device *adev)
>   	amdgpu_ttm_debugfs_fini(adev);
>   	amdgpu_bo_free_kernel(&adev->stolen_vga_memory, NULL, NULL);
>   	amdgpu_ttm_fw_reserve_vram_fini(adev);
> +	if (adev->mman.aper_base_kaddr)
> +		iounmap(adev->mman.aper_base_kaddr);
> +	adev->mman.aper_base_kaddr = NULL;
>   
>   	ttm_bo_clean_mm(&adev->mman.bdev, TTM_PL_VRAM);
>   	ttm_bo_clean_mm(&adev->mman.bdev, TTM_PL_TT);
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.h
> index 1e275c7..d314910 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.h
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.h
> @@ -44,6 +44,7 @@ struct amdgpu_mman {
>   	struct ttm_bo_device		bdev;
>   	bool				mem_global_referenced;
>   	bool				initialized;
> +	void __iomem			*aper_base_kaddr;
>   
>   #if defined(CONFIG_DEBUG_FS)
>   	struct dentry			*debugfs_entries[8];

_______________________________________________
amd-gfx mailing list
amd-gfx@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/amd-gfx