[devel] [PATCH 28/35] arm64: device tree: baikal: mark GPU as dma-coherent

[asheplyakov на yandex.ru]

From: Alexey Sheplyakov <asheplyakov на altlinux.org>

With memattr 0x888d88 (set by arm_mali_lpae_alloc_pgtable) GPU
(Mali T628 r1p0) experiences a lot of DATA_INVALID faults,
unhandled page faults, and other errors. Also the screen goes
black almost immediately.

On the other hand with memattr 0x484d48 (as set by mali_kbase)
the GPU appears to work just fine.

Robin Murphy <robin.murphy на arm.com> explains:

> using the outer-cacheable attribute is deliberate because it is necessary
> for I/O-coherent GPUs to work properly (and should be irrelevant for
> non-coherent integrations)

> I'd note that panfrost has been working OK - to the extent that Mesa
> supports its older ISA - on the T624 (single core group) in Arm's
> Juno SoC for over a year now since commit 268af50f38b1.

> If you have to force outer non-cacheable to avoid getting translation
> faults and other errors that look like the GPU is inexplicably seeing
> the wrong data, I'd check whether you have the same thing where your
> integration is actually I/O-coherent and you're missing the "dma-coherent"
> property in your DT.

Indeed setting "gpu-coherent" property (and adjusting jobs affinity
for dual core group GPU) makes panfrost work just fine on Baikal-M.

X-DONTUPSTREAM
X-feature-Baikal-M
---
 arch/arm64/boot/dts/baikal/bm1000.dtsi | 1 +
 1 file changed, 1 insertion(+)

diff --git a/arch/arm64/boot/dts/baikal/bm1000.dtsi b/arch/arm64/boot/dts/baikal/bm1000.dtsi
index bc69835c4d4f..cc4e7d199c27 100644
--- a/arch/arm64/boot/dts/baikal/bm1000.dtsi
+++ b/arch/arm64/boot/dts/baikal/bm1000.dtsi
@@ -692,6 +692,7 @@ gpu: gpu на 2a200000 {
 			interrupt-names = "job", "mmu", "gpu";
 			clocks = <&cmu_mali>;
 			clock-names = "gpuclk";
+			dma-coherent;
 			operating-points-v2 = <&gpu_opp_table>;
 		};
 
-- 
2.32.0