]> git.ipfire.org Git - thirdparty/linux.git/commitdiff
drm/amdgpu/gmc9: rework stolen vga memory handling
authorAlex Deucher <alexander.deucher@amd.com>
Thu, 30 Aug 2018 14:31:56 +0000 (09:31 -0500)
committerAlex Deucher <alexander.deucher@amd.com>
Sun, 2 Sep 2018 15:16:51 +0000 (10:16 -0500)
No functional change, just rework it in order to adjust the
behavior on a per asic level.  The problem is that on vega10,
something corrupts the lower 8 MB of vram on the second
resume from S3.  This does not seem to affect Raven, other
gmc9 based asics need testing.

Reviewed-by: Junwei Zhang <Jerry.Zhang@amd.com>
Acked-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c

index 04d50893a6f2b514f898694ad728baa963c75acd..46cff7d8b3759963469a93f4335b54d73d1a6be1 100644 (file)
@@ -692,6 +692,28 @@ static int gmc_v9_0_ecc_available(struct amdgpu_device *adev)
        return lost_sheep == 0;
 }
 
+static bool gmc_v9_0_keep_stolen_memory(struct amdgpu_device *adev)
+{
+
+       /*
+        * TODO:
+        * Currently there is a bug where some memory client outside
+        * of the driver writes to first 8M of VRAM on S3 resume,
+        * this overrides GART which by default gets placed in first 8M and
+        * causes VM_FAULTS once GTT is accessed.
+        * Keep the stolen memory reservation until the while this is not solved.
+        * Also check code in gmc_v9_0_get_vbios_fb_size and gmc_v9_0_late_init
+        */
+       switch (adev->asic_type) {
+       case CHIP_RAVEN:
+       case CHIP_VEGA10:
+       case CHIP_VEGA12:
+       case CHIP_VEGA20:
+       default:
+               return true;
+       }
+}
+
 static int gmc_v9_0_late_init(void *handle)
 {
        struct amdgpu_device *adev = (struct amdgpu_device *)handle;
@@ -708,10 +730,8 @@ static int gmc_v9_0_late_init(void *handle)
        unsigned i;
        int r;
 
-       /*
-        * TODO - Uncomment once GART corruption issue is fixed.
-        */
-       /* amdgpu_bo_late_init(adev); */
+       if (!gmc_v9_0_keep_stolen_memory(adev))
+               amdgpu_bo_late_init(adev);
 
        for(i = 0; i < adev->num_rings; ++i) {
                struct amdgpu_ring *ring = adev->rings[i];
@@ -848,18 +868,16 @@ static int gmc_v9_0_gart_init(struct amdgpu_device *adev)
 
 static unsigned gmc_v9_0_get_vbios_fb_size(struct amdgpu_device *adev)
 {
-#if 0
        u32 d1vga_control = RREG32_SOC15(DCE, 0, mmD1VGA_CONTROL);
-#endif
        unsigned size;
 
        /*
         * TODO Remove once GART corruption is resolved
         * Check related code in gmc_v9_0_sw_fini
         * */
-       size = 9 * 1024 * 1024;
+       if (gmc_v9_0_keep_stolen_memory(adev))
+               return 9 * 1024 * 1024;
 
-#if 0
        if (REG_GET_FIELD(d1vga_control, D1VGA_CONTROL, D1VGA_MODE_ENABLE)) {
                size = 9 * 1024 * 1024; /* reserve 8MB for vga emulator and 1 MB for FB */
        } else {
@@ -876,6 +894,7 @@ static unsigned gmc_v9_0_get_vbios_fb_size(struct amdgpu_device *adev)
                        break;
                case CHIP_VEGA10:
                case CHIP_VEGA12:
+               case CHIP_VEGA20:
                default:
                        viewport = RREG32_SOC15(DCE, 0, mmSCL0_VIEWPORT_SIZE);
                        size = (REG_GET_FIELD(viewport, SCL0_VIEWPORT_SIZE, VIEWPORT_HEIGHT) *
@@ -888,7 +907,6 @@ static unsigned gmc_v9_0_get_vbios_fb_size(struct amdgpu_device *adev)
        if ((adev->gmc.real_vram_size - size) < (8 * 1024 * 1024))
                return 0;
 
-#endif
        return size;
 }
 
@@ -1000,16 +1018,8 @@ static int gmc_v9_0_sw_fini(void *handle)
        amdgpu_gem_force_release(adev);
        amdgpu_vm_manager_fini(adev);
 
-       /*
-       * TODO:
-       * Currently there is a bug where some memory client outside
-       * of the driver writes to first 8M of VRAM on S3 resume,
-       * this overrides GART which by default gets placed in first 8M and
-       * causes VM_FAULTS once GTT is accessed.
-       * Keep the stolen memory reservation until the while this is not solved.
-       * Also check code in gmc_v9_0_get_vbios_fb_size and gmc_v9_0_late_init
-       */
-       amdgpu_bo_free_kernel(&adev->stolen_vga_memory, NULL, NULL);
+       if (gmc_v9_0_keep_stolen_memory(adev))
+               amdgpu_bo_free_kernel(&adev->stolen_vga_memory, NULL, NULL);
 
        amdgpu_gart_table_vram_free(adev);
        amdgpu_bo_fini(adev);