[PATCH] drm/amdgpu: move GTT allocation from gmc_sw_init to gmc_hw_init

[Chen, Guchun]

[Public]

This will create sdma_access_bo only for ASIC with gmc v10? Original creation occurs in amdgpu_ttm_init, it's not limited to ASICs with gmc v10.

Regards,
Guchun

-----Original Message-----
From: amd-gfx <amd-gfx-bounces at lists.freedesktop.org> On Behalf Of Aaron Liu
Sent: Thursday, January 27, 2022 3:04 PM
To: amd-gfx at lists.freedesktop.org
Cc: Ji, Ruili <Ruili.Ji at amd.com>; Kim, Jonathan <Jonathan.Kim at amd.com>; Kuehling, Felix <Felix.Kuehling at amd.com>; Liu, Aaron <Aaron.Liu at amd.com>; Huang, Ray <Ray.Huang at amd.com>; Deucher, Alexander <Alexander.Deucher at amd.com>
Subject: [PATCH] drm/amdgpu: move GTT allocation from gmc_sw_init to gmc_hw_init

The below patch causes system hang for harvested ASICs.
d015e9861e55 drm/amdgpu: improve debug VRAM access performance using sdma

The root cause is that GTT buffer should be allocated after GC SA harvest programming completed.

For harvested AISC, the GC SA harvest process(see utcl2_harvest) is programmed in gmc_v10_0_hw_init function. This is a hardware programming.
Therefore should be located in hw init. Hence need to move GTT allocation from gmc_v10_0_sw_init to gmc_v10_0_hw_init.

Signed-off-by: Aaron Liu <aaron.liu at amd.com>
---
 drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c | 8 --------  drivers/gpu/drm/amd/amdgpu/gmc_v10_0.c  | 9 +++++++++
 2 files changed, 9 insertions(+), 8 deletions(-)

diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c
index 3d8a20956b74..7ce0478b2908 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c
@@ -1823,12 +1823,6 @@ int amdgpu_ttm_init(struct amdgpu_device *adev)
 		return r;
 	}
 
-	if (amdgpu_bo_create_kernel(adev, PAGE_SIZE, PAGE_SIZE,
-				AMDGPU_GEM_DOMAIN_GTT,
-				&adev->mman.sdma_access_bo, NULL,
-				&adev->mman.sdma_access_ptr))
-		DRM_WARN("Debug VRAM access will use slowpath MM access\n");
-
 	return 0;
 }
 
@@ -1850,8 +1844,6 @@ void amdgpu_ttm_fini(struct amdgpu_device *adev)
 	if (adev->mman.stolen_reserved_size)
 		amdgpu_bo_free_kernel(&adev->mman.stolen_reserved_memory,
 				      NULL, NULL);
-	amdgpu_bo_free_kernel(&adev->mman.sdma_access_bo, NULL,
-					&adev->mman.sdma_access_ptr);
 	amdgpu_ttm_fw_reserve_vram_fini(adev);
 
 	if (drm_dev_enter(adev_to_drm(adev), &idx)) { diff --git a/drivers/gpu/drm/amd/amdgpu/gmc_v10_0.c b/drivers/gpu/drm/amd/amdgpu/gmc_v10_0.c
index 73ab0eebe4e2..c560bdc2a93c 100644
--- a/drivers/gpu/drm/amd/amdgpu/gmc_v10_0.c
+++ b/drivers/gpu/drm/amd/amdgpu/gmc_v10_0.c
@@ -1062,6 +1062,12 @@ static int gmc_v10_0_hw_init(void *handle)
 	if (adev->umc.funcs && adev->umc.funcs->init_registers)
 		adev->umc.funcs->init_registers(adev);
 
+	if (amdgpu_bo_create_kernel(adev, PAGE_SIZE, PAGE_SIZE,
+				AMDGPU_GEM_DOMAIN_GTT,
+				&adev->mman.sdma_access_bo, NULL,
+				&adev->mman.sdma_access_ptr))
+		DRM_WARN("Debug VRAM access will use slowpath MM access\n");
+
 	return 0;
 }
 
@@ -1082,6 +1088,9 @@ static int gmc_v10_0_hw_fini(void *handle)  {
 	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
 
+	amdgpu_bo_free_kernel(&adev->mman.sdma_access_bo, NULL,
+					&adev->mman.sdma_access_ptr);
+
 	gmc_v10_0_gart_disable(adev);
 
 	if (amdgpu_sriov_vf(adev)) {
--
2.25.1