On 02/08/2024 11:39, Christian König wrote:
Am 02.08.24 um 09:17 schrieb Lu Yao:
Add support for the drm_panic module, which displays a pretty user
friendly message on the screen when a Linux kernel panic occurs.
Signed-off-by: Lu Yao <yaolu@xxxxxxxxxx>
---
The patch can work properly on the TTY, but after start X, drawn
image is messy, it looks like the data isn't linearly arranged.
However at this time 'fb->modifier' is 'DRM_FORMAT_MOD_LINEAR'.
Another difference I found is:
For TTY, the amdgpu_bo is created with flag
'AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED|AMDGPU_GEM_CREATE_CPU_GTT_USWC|
AMDGPU_GEM_CREATE_VRAM_CLEARED|AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS'.
For X, the amdgpu_bo is created with flag
'AMDGPU_GEM_CREATE_NO_CPU_ACCESS|AMDGPU_GEM_CREATE_CPU_GTT_USWC'
I try to use same flag for X, it looks like no difference.
Can someone provide some insight into this problem or where I am going
wrong. Thanks a lot.
Test environment: X86 arch + v6.6 kernel + R7340.
---
drivers/gpu/drm/amd/amdgpu/dce_v6_0.c | 32 +++++++++++++++++++++++++++
1 file changed, 32 insertions(+)
diff --git a/drivers/gpu/drm/amd/amdgpu/dce_v6_0.c
b/drivers/gpu/drm/amd/amdgpu/dce_v6_0.c
index 05c0df97f01d..12c3801c264a 100644
--- a/drivers/gpu/drm/amd/amdgpu/dce_v6_0.c
+++ b/drivers/gpu/drm/amd/amdgpu/dce_v6_0.c
@@ -28,6 +28,8 @@
#include <drm/drm_modeset_helper.h>
#include <drm/drm_modeset_helper_vtables.h>
#include <drm/drm_vblank.h>
+#include <drm/drm_panic.h>
+#include "../../drm_internal.h"
Well that this file is named "internal" and not in a common include
directory is a strong indicator that you should absolutely *not* include
it in a driver.
#include "amdgpu.h"
#include "amdgpu_pm.h"
@@ -2600,6 +2602,35 @@ static const struct drm_crtc_helper_funcs
dce_v6_0_crtc_helper_funcs = {
.get_scanout_position = amdgpu_crtc_get_scanout_position,
};
+static int dce_v6_0_drm_primary_plane_get_scanout_buffer(struct
drm_plane *plane,
+ struct drm_scanout_buffer *sb)
+{
+ struct drm_framebuffer *fb;
+ struct drm_gem_object *obj;
+ struct amdgpu_bo *abo;
+ int ret = 0;
+
+ if (!plane->fb || plane->fb->modifier != DRM_FORMAT_MOD_LINEAR)
+ return -ENODEV;
+
+ fb = plane->fb;
+ sb->width = fb->width;
+ sb->height = fb->height;
+ sb->format = fb->format;
+ sb->pitch[0] = fb->pitches[0];
+
+ obj = fb->obj[0];
+ abo = gem_to_amdgpu_bo(obj);
+ if (!abo || abo->flags & AMDGPU_GEM_CREATE_NO_CPU_ACCESS)
+ return -EINVAL;
+
+ return drm_gem_vmap(obj, &sb->map[0]);
Yeah that will almost always not work. Most display buffers are tilled
and not CPU accessible.
For the CPU accessible issue, Christian mentioned there was a debug
interface on AMD GPU that can be used, to work around this:
https://lore.kernel.org/dri-devel/0baabe1f-8924-2c9a-5cd4-59084a37dbb2@xxxxxxxxx/
and
https://lore.kernel.org/dri-devel/d233c376-ed07-2127-6084-8292d313dac7@xxxxxxx/
And you will need to use the scanout_buffer->set_pixel() callback to
write the pixels one by one, similar to what I've tried for nouveau with
https://patchwork.freedesktop.org/series/133963/
For the tiling format, the problem is that it is internal to the GPU,
and currently the driver don't know which tiling format is being used.
It might be possible to disable tiling and compression, but it requires
some internal DC knowledge:
https://lore.kernel.org/dri-devel/f76a3297-7d63-8615-45c5-47f02b64a1d5@xxxxxxx/
Best regards,
--
Jocelyn
Regards,
Christian.
+}
+
+static const struct drm_plane_helper_funcs
dce_v6_0_drm_primary_plane_helper_funcs = {
+ .get_scanout_buffer = dce_v6_0_drm_primary_plane_get_scanout_buffer
+};
+
static int dce_v6_0_crtc_init(struct amdgpu_device *adev, int index)
{
struct amdgpu_crtc *amdgpu_crtc;
@@ -2627,6 +2658,7 @@ static int dce_v6_0_crtc_init(struct
amdgpu_device *adev, int index)
amdgpu_crtc->encoder = NULL;
amdgpu_crtc->connector = NULL;
drm_crtc_helper_add(&amdgpu_crtc->base,
&dce_v6_0_crtc_helper_funcs);
+ drm_plane_helper_add(amdgpu_crtc->base.primary,
&dce_v6_0_drm_primary_plane_helper_funcs);
return 0;
}
