[PATCH v1 0/2] CMA-memory-to-offload-FNN-PLUS-module-to-eNPU

[PATCH v1 0/2] CMA-memory-to-offload-FNN-PLUS-module-to-eNPU

[Karthik S]

This series contains patches for allocating CMA memory to offload
FNN PLUS module to eNPU.

Karthik S (2):
  arm64: dts: qcom: qcs6490-rb3gen2: add audio CMA memory region for
    AudioReach
  arm64: dts: qcom: kodiak-el2: add LPASS AML IOMMU SIDs for eNPU audio
    offload

 arch/arm64/boot/dts/qcom/kodiak-el2.dtso         | 5 ++++-
 arch/arm64/boot/dts/qcom/qcs6490-audioreach.dtsi | 1 +
 arch/arm64/boot/dts/qcom/qcs6490-rb3gen2.dts     | 8 ++++++++
 3 files changed, 13 insertions(+), 1 deletion(-)

This patch depends on: 
 arm64: dts: qcom: kodiak: EL2 overlay for qcm6490-idp / qcs6490-rb3gen2
 
 Link: https://lore.kernel.org/all/20260803121042.3313824-1-mukesh.ojha@oss.qualcomm.com/
 
 The prerequisite patch introduces kodiak el2 overlay.

-- 
2.34.1

[PATCH v1 1/2] arm64: dts: qcom: qcs6490-rb3gen2: add audio CMA memory region for AudioReach

[Karthik S]

Add a dynamic CMA pool for the AudioReach audio stack on qcs6490-rb3gen2.
Use alloc-ranges to let the kernel pick a free address rather than a fixed
reg, which avoided EFI-reserved memory at 0x9cc00000. Add 4 MiB alignment
to satisfy CMA_MIN_ALIGNMENT_BYTES. Wire memory-region into q6apmdai so
the q6apm-audio-mem-cma child driver can attach the pool.

Based on the equivalent implementation for lemans (SA8295P).

Signed-off-by: Karthik S <karthik.s@oss.qualcomm.com>
---
 arch/arm64/boot/dts/qcom/qcs6490-audioreach.dtsi | 1 +
 arch/arm64/boot/dts/qcom/qcs6490-rb3gen2.dts     | 8 ++++++++
 2 files changed, 9 insertions(+)

diff --git a/arch/arm64/boot/dts/qcom/qcs6490-audioreach.dtsi b/arch/arm64/boot/dts/qcom/qcs6490-audioreach.dtsi
index c1867711298b..06f65ae8afbc 100644
--- a/arch/arm64/boot/dts/qcom/qcs6490-audioreach.dtsi
+++ b/arch/arm64/boot/dts/qcom/qcs6490-audioreach.dtsi
@@ -97,6 +97,7 @@ q6apm: service@1 {
 			q6apmdai: dais {
 				compatible = "qcom,q6apm-dais";
 				iommus = <&apps_smmu 0x1801 0x0>;
+				memory-region = <&audio_cma_mem>;
 			};
 
 			q6apmbedai: bedais {
diff --git a/arch/arm64/boot/dts/qcom/qcs6490-rb3gen2.dts b/arch/arm64/boot/dts/qcom/qcs6490-rb3gen2.dts
index a13315bf0fb0..d39246f34172 100644
--- a/arch/arm64/boot/dts/qcom/qcs6490-rb3gen2.dts
+++ b/arch/arm64/boot/dts/qcom/qcs6490-rb3gen2.dts
@@ -174,6 +174,14 @@ rmtfs_mem: memory@f8500000 {
 			qcom,client-id = <1>;
 			qcom,vmid = <QCOM_SCM_VMID_MSS_MSA>, <QCOM_SCM_VMID_NAV>;
 		};
+
+		audio_cma_mem: qcom,audio-ml {
+			compatible = "shared-dma-pool";
+			alloc-ranges = <0x0 0x00000000 0x0 0xffffffff>;
+			reusable;
+			alignment = <0x0 0x400000>;
+			size = <0x0 0x1000000>;
+		};
 	};
 
 	gpio-keys {
-- 
2.34.1

[PATCH v1 2/2] arm64: dts: qcom: kodiak-el2: add LPASS AML IOMMU SIDs for eNPU audio offload

[Karthik S]

Add the three LPASS AML stream IDs to remoteproc_adsp in the
kodiak EL2 overlay so that the q6apm-audio-mem-cma child driver
can obtain an IOMMU domain via iommu_get_domain_for_dev(dev).

Without these SIDs the CMA device has no IOMMU domain (has_iommu=0)
and IOCTL_MAP_HYP_ASSIGN falls through to the SCM path, which fails
on RB3Gen2 because Gunyah owns the SMMU.

SID assignments:
  0x1860 / mask 0x9 — ML Activation + Weights + Write streams
  0x1862 / mask 0x1 — ML Config + Bias streams
  0x1864 / mask 0x0 — ML Normalized Data stream

Signed-off-by: Karthik S <karthik.s@oss.qualcomm.com>
---
 arch/arm64/boot/dts/qcom/kodiak-el2.dtso | 5 ++++-
 1 file changed, 4 insertions(+), 1 deletion(-)

diff --git a/arch/arm64/boot/dts/qcom/kodiak-el2.dtso b/arch/arm64/boot/dts/qcom/kodiak-el2.dtso
index 8daf7f02fd75..d050857cab31 100644
--- a/arch/arm64/boot/dts/qcom/kodiak-el2.dtso
+++ b/arch/arm64/boot/dts/qcom/kodiak-el2.dtso
@@ -13,7 +13,10 @@ &gpu_zap_shader {
 };
 
 &remoteproc_adsp {
-	iommus = <&apps_smmu 0x1800 0x0>;
+	iommus = <&apps_smmu 0x1800 0x0>,
+		 <&apps_smmu 0x1860 0x9>,
+		 <&apps_smmu 0x1862 0x1>,
+		 <&apps_smmu 0x1864 0x0>;
 };
 
 &remoteproc_cdsp {
-- 
2.34.1

Re: [PATCH v1 1/2] arm64: dts: qcom: qcs6490-rb3gen2: add audio CMA memory region for AudioReach

[Krzysztof Kozlowski]

On 25/08/2026 15:33, Karthik S wrote:
> Add a dynamic CMA pool for the AudioReach audio stack on qcs6490-rb3gen2.
> Use alloc-ranges to let the kernel pick a free address rather than a fixed
> reg, which avoided EFI-reserved memory at 0x9cc00000. Add 4 MiB alignment
> to satisfy CMA_MIN_ALIGNMENT_BYTES. Wire memory-region into q6apmdai so
> the q6apm-audio-mem-cma child driver can attach the pool.
> 
> Based on the equivalent implementation for lemans (SA8295P).
> 
> Signed-off-by: Karthik S <karthik.s@oss.qualcomm.com>
> ---
>  arch/arm64/boot/dts/qcom/qcs6490-audioreach.dtsi | 1 +
>  arch/arm64/boot/dts/qcom/qcs6490-rb3gen2.dts     | 8 ++++++++
>  2 files changed, 9 insertions(+)
> 
> diff --git a/arch/arm64/boot/dts/qcom/qcs6490-audioreach.dtsi b/arch/arm64/boot/dts/qcom/qcs6490-audioreach.dtsi
> index c1867711298b..06f65ae8afbc 100644
> --- a/arch/arm64/boot/dts/qcom/qcs6490-audioreach.dtsi
> +++ b/arch/arm64/boot/dts/qcom/qcs6490-audioreach.dtsi
> @@ -97,6 +97,7 @@ q6apm: service@1 {
>  			q6apmdai: dais {
>  				compatible = "qcom,q6apm-dais";
>  				iommus = <&apps_smmu 0x1801 0x0>;
> +				memory-region = <&audio_cma_mem>;
>  			};
>  
>  			q6apmbedai: bedais {
> diff --git a/arch/arm64/boot/dts/qcom/qcs6490-rb3gen2.dts b/arch/arm64/boot/dts/qcom/qcs6490-rb3gen2.dts
> index a13315bf0fb0..d39246f34172 100644
> --- a/arch/arm64/boot/dts/qcom/qcs6490-rb3gen2.dts
> +++ b/arch/arm64/boot/dts/qcom/qcs6490-rb3gen2.dts
> @@ -174,6 +174,14 @@ rmtfs_mem: memory@f8500000 {
>  			qcom,client-id = <1>;
>  			qcom,vmid = <QCOM_SCM_VMID_MSS_MSA>, <QCOM_SCM_VMID_NAV>;
>  		};
> +
> +		audio_cma_mem: qcom,audio-ml {

Don't send downstream code. There are no such patterns in upstream for
years.

Best regards,
Krzysztof

Re: [PATCH v1 1/2] arm64: dts: qcom: qcs6490-rb3gen2: add audio CMA memory region for AudioReach

[Krzysztof Kozlowski]

On 25/08/2026 15:34, Krzysztof Kozlowski wrote:
> On 25/08/2026 15:33, Karthik S wrote:
>> Add a dynamic CMA pool for the AudioReach audio stack on qcs6490-rb3gen2.
>> Use alloc-ranges to let the kernel pick a free address rather than a fixed
>> reg, which avoided EFI-reserved memory at 0x9cc00000. Add 4 MiB alignment
>> to satisfy CMA_MIN_ALIGNMENT_BYTES. Wire memory-region into q6apmdai so
>> the q6apm-audio-mem-cma child driver can attach the pool.
>>
>> Based on the equivalent implementation for lemans (SA8295P).

And the "equivalent implementation for lemans" received comments which
you missed and re-implemented the same issues.

Best regards,
Krzysztof

Re: [PATCH v1 1/2] arm64: dts: qcom: qcs6490-rb3gen2: add audio CMA memory region for AudioReach

[Konrad Dybcio]

On 8/25/26 3:33 PM, Karthik S wrote:
> Add a dynamic CMA pool for the AudioReach audio stack on qcs6490-rb3gen2.
> Use alloc-ranges to let the kernel pick a free address rather than a fixed
> reg, which avoided EFI-reserved memory at 0x9cc00000. Add 4 MiB alignment
> to satisfy CMA_MIN_ALIGNMENT_BYTES. Wire memory-region into q6apmdai so
> the q6apm-audio-mem-cma child driver can attach the pool.
> 
> Based on the equivalent implementation for lemans (SA8295P).
> 
> Signed-off-by: Karthik S <karthik.s@oss.qualcomm.com>
> ---

Please don't send patches that you haven't compile-tested

Konrad