On 31/10/2025 14:12, Tvrtko Ursulin wrote:
On 27/10/2025 10:26, Thomas Zimmermann wrote:
Hi
Am 27.10.25 um 10:46 schrieb Jocelyn Falempe:
On 24/10/2025 17:55, Tvrtko Ursulin wrote:
On 24/10/2025 16:18, Thomas Zimmermann wrote:
Hi
Am 24.10.25 um 15:33 schrieb Jocelyn Falempe:
On 24/10/2025 14:40, Thomas Zimmermann wrote:
Hi
Am 24.10.25 um 13:53 schrieb Tvrtko Ursulin:
On 24/10/2025 12:04, Jocelyn Falempe wrote:
On a lenovo se100 server, when using i915 GPU for rendering,
and the
ast driver for display, the graphic output is corrupted, and
almost
unusable.
Adding a clflush call in the vmap function fixes this issue
completely.
AST is importing i915 allocated buffer in this use case, or how
exactly is the relationship?
Wondering if some path is not calling dma_buf_begin/
end_cpu_access().
Yes, ast doesn't call begin/end_cpu_access in [1].
Jocelyn, if that fixes the issue, feel free to send me a patch
for review.
[1] https://elixir.bootlin.com/linux/v6.17.4/source/drivers/gpu/
drm/ ast/ ast_mode.c
I tried the following patch, but that doesn't fix the graphical
issue:
diff --git a/drivers/gpu/drm/ast/ast_mode.c b/drivers/gpu/drm/ast/
ast_mode.c
index b4e8edc7c767..e50f95a4c8a9 100644
--- a/drivers/gpu/drm/ast/ast_mode.c
+++ b/drivers/gpu/drm/ast/ast_mode.c
@@ -564,6 +564,7 @@ static void
ast_primary_plane_helper_atomic_update(struct drm_plane *plane,
struct drm_crtc_state *crtc_state =
drm_atomic_get_new_crtc_state(state, crtc);
struct drm_rect damage;
struct drm_atomic_helper_damage_iter iter;
+ int ret;
if (!old_fb || (fb->format != old_fb->format) ||
crtc_state- >mode_changed) {
struct ast_crtc_state *ast_crtc_state =
to_ast_crtc_state(crtc_state);
@@ -572,11 +573,16 @@ static void
ast_primary_plane_helper_atomic_update(struct drm_plane *plane,
ast_set_vbios_color_reg(ast, fb->format,
ast_crtc_state->vmode);
}
+ ret = drm_gem_fb_begin_cpu_access(fb, DMA_FROM_DEVICE);
+ pr_info("AST begin_cpu_access %d\n", ret);
Presumably, you end up in [1]. I cannot find the cflush there or in
[2]. Maybe you need to add this call somewhere in there, similar to
[3]. Just guessing.
Near [2] clflush can happen at [4] *if* the driver thinks it is
needed. Most GPUs are cache coherent so mostly it isn't. But if this
is a Meteorlake machine (when I google Lenovo se100 it makes me
think so?) then the userspace has some responsibility to manage
things since there it is only 1-way coherency. Or userspace could
have even told the driver to stay off in which case it then needs to
manage everything. From the top of my head I am not sure how exactly
this used to work, or how it is supposed to interact with exported
buffers.
If this is indeed on Meteorlake, maybe Joonas or Rodrigo remember
better how the special 1-way coherency is supposed to be managed there?
I've made an experiment, and if I add:
* a calls to drm_gem_fb_begin_cpu_access() in the ast driver.
* and in i915_gem_domain.c flush_write_domain():
case I915_GEM_DOMAIN_RENDER:
+ i915_gem_clflush_object(obj, I915_CLFLUSH_SYNC |
I915_CLFLUSH_FORCE);
Then that fixes the issue too.
So I think there are two things to fix:
* The missing call to drm_gem_fb_begin_cpu_access() in ast.
Yes. We definitely want to add these calls, as they are expected for
this case.
Browsing around a bit, I notice ast_primary_plane_helper_atomic_update()
calls to_drm_shadow_plane_state() to get the source of the memcpy.
Should there somewhere be calls to drm_gem_begin_shadow_fb_access() and
drm_gem_end_shadow_fb_access()? Or those should be set as vfuncs by
someone? Sorry I get lost easily in the DRM maze of helpers<->vfuncs<-
>helpers<->vfuncs..
I've already submitted [1] for the ast driver.
* The missing cache flush in i915 for the Arrowlake iGPU (but
probably not the way I've done it).
You call begin_cpu_access with DMA_FROM_DEVICE, but there's no support
for that flag in i915 AFAICT. Maybe this needs to be added somehow?
AFAIR the premise is GPU writes will not get stuck in the last level
cache but I might be remembering a reverse of what Meteorlake 1-way
coherency means. This area of the driver ended up a mess and was never
properly cleaned up. I even had a series to try and do it but it never
happened. We will need someone who actually remembers how Meteorlake works.
Logically clflush() shouldn't have any effect, because it flushes the
CPU cache, and then ast is reading from the CPU.
But from the tests, it also flushes the iGPU cache, and the iGPU cache
is not coherent with the CPU cache (at least for iGPU writes), or this
problem won't exist.
I've the setup ready, and this is easily reproducible, so you can send
me patches, and I can quickly check if that works or not.
[1]
https://lore.kernel.org/dri-devel/[email protected]/T/#t
Best regards,
--
Jocelyn
Regards,
Tvrtko
