Re: [PATCH v2 1/3] arm64: dts: qcom: sdm630/660: Add CDSP-related nodes

[Nickolay Goppen <setotau@mainlining.org>]

26.11.2025 17:00, Nickolay Goppen пишет:
>
> 21.11.2025 15:09, Dmitry Baryshkov пишет:
>> On Fri, Nov 21, 2025 at 01:41:21PM +0530, Ekansh Gupta wrote:
>>>
>>> On 11/20/2025 5:17 PM, Konrad Dybcio wrote:
>>>> On 11/20/25 11:54 AM, Ekansh Gupta wrote:
>>>>> On 11/20/2025 1:27 PM, Nickolay Goppen wrote:
>>>>>> 20.11.2025 07:55, Ekansh Gupta пишет:
>>>>>>> On 11/20/2025 1:58 AM, Srinivas Kandagatla wrote:
>>>>>>>> On 11/12/25 1:52 PM, Konrad Dybcio wrote:
>>>>>>>>> On 11/10/25 6:41 PM, Srinivas Kandagatla wrote:
>>>>>>>>>> On 11/3/25 12:52 PM, Konrad Dybcio wrote:
>>>>>>>>>>> On 10/31/25 12:30 PM, Nickolay Goppen wrote:
>>>>>>>>>>>> 24.10.2025 16:58, Nickolay Goppen пишет:
>>>>>>>>>>>>> 24.10.2025 11:28, Konrad Dybcio пишет:
>>>>>>>>>>>>>> On 10/23/25 9:51 PM, Nickolay Goppen wrote:
>>>>>>>>>>>>>>> In order to enable CDSP support for SDM660 SoC:
>>>>>>>>>>>>>>>     * add shared memory p2p nodes for CDSP
>>>>>>>>>>>>>>>     * add CDSP-specific smmu node
>>>>>>>>>>>>>>>     * add CDSP peripheral image loader node
>>>>>>>>>>>>>>>
>>>>>>>>>>>>>>> Memory region for CDSP in SDM660 occupies the same spot as
>>>>>>>>>>>>>>> TZ buffer mem defined in sdm630.dtsi (which does not 
>>>>>>>>>>>>>>> have CDSP).
>>>>>>>>>>>>>>> In sdm660.dtsi replace buffer_mem inherited from SDM630 
>>>>>>>>>>>>>>> with
>>>>>>>>>>>>>>> cdsp_region, which is also larger in size.
>>>>>>>>>>>>>>>
>>>>>>>>>>>>>>> SDM636 also doesn't have CDSP, so remove inherited from 
>>>>>>>>>>>>>>> sdm660.dtsi
>>>>>>>>>>>>>>> related nodes and add buffer_mem back.
>>>>>>>>>>>>>>>
>>>>>>>>>>>>>>> Signed-off-by: Nickolay Goppen <setotau@mainlining.org>
>>>>>>>>>>>>>>> ---
>>>>>>>>>>>>>> [...]
>>>>>>>>>>>>>>
>>>>>>>>>>>>>>> + label = "turing";
>>>>>>>>>>>>>> "cdsp"
>>>>>>>>>>>>> Ok, I'll change this in the next revision.
>>>>>>>>>>>>>>> + mboxes = <&apcs_glb 29>;
>>>>>>>>>>>>>>> +            qcom,remote-pid = <5>;
>>>>>>>>>>>>>>> +
>>>>>>>>>>>>>>> +            fastrpc {
>>>>>>>>>>>>>>> +                compatible = "qcom,fastrpc";
>>>>>>>>>>>>>>> +                qcom,glink-channels = 
>>>>>>>>>>>>>>> "fastrpcglink-apps-dsp";
>>>>>>>>>>>>>>> +                label = "cdsp";
>>>>>>>>>>>>>>> + qcom,non-secure-domain;
>>>>>>>>>>>>>> This shouldn't matter, both a secure and a non-secure 
>>>>>>>>>>>>>> device is
>>>>>>>>>>>>>> created for CDSP
>>>>>>>>>>>>> I've added this property, because it is used in other 
>>>>>>>>>>>>> SoC's, such as SDM845 and SM6115 for both ADSP and CDSP
>>>>>>>>>>>> Is this property not neccessary anymore?
>>>>>>>>>>> +Srini?
>>>>>>>>>> That is true, we do not require this for CDSP, as CDSP allows 
>>>>>>>>>> both
>>>>>>>>>> unsigned and signed loading, we create both secured and 
>>>>>>>>>> non-secure node
>>>>>>>>>> by default. May be we can provide that clarity in yaml 
>>>>>>>>>> bindings so that
>>>>>>>>>> it gets caught during dtb checks.
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>> However in ADSP case, we only support singed modules, due to 
>>>>>>>>>> historical
>>>>>>>>>> reasons how this driver evolved over years, we have this flag 
>>>>>>>>>> to allow
>>>>>>>>>> compatiblity for such users.
>>>>>>>>> Does that mean that we can only load signed modules on the 
>>>>>>>>> ADSP, but
>>>>>>>>> the driver behavior was previously such that unsigned modules 
>>>>>>>>> were
>>>>>>>>> allowed (which was presumably fine on devboards, but not on fused
>>>>>>>>> devices)?
>>>>>>>> Yes, its true that we allowed full access to adsp device nodes 
>>>>>>>> when we
>>>>>>>> first started upstreaming fastrpc driver.
>>>>>>>>
>>>>>>>> irrespective of the board only signed modules are supported on 
>>>>>>>> the ADSP.
>>>>>>>> I think there was one version of SoC i think 8016 or some older 
>>>>>>>> one
>>>>>>>> which had adsp with hvx which can load unsigned modules for 
>>>>>>>> compute
>>>>>>>> usecase only.
>>>>>>>>
>>>>>>>> I have added @Ekansh for more clarity.
>>>>>>>>
>>>>>>>> --srini
>>>>>>> For all the available platforms, ADSP supports only signed 
>>>>>>> modules. Unsigned
>>>>>>> modules(as well as signed) are supported by CDSP and GDSP 
>>>>>>> subsystems.
>>>>>>>
>>>>>>> qcom,non-secure-domain property marks the corresponding DSP as 
>>>>>>> non-secure DSP.
>>>>>>> The implications of adding this property would be the following:
>>>>>>> on ADSP, SDSP, MDSP:
>>>>>>> - Only non-secure device node(/dev/fastrpc-Xdsp) is created.
>>>>>>> - Non-secure device node can be used for signed DSP PD offload.
>>>>>>>
>>>>>>> on CDSP, GDSP:
>>>>>>> - Both secure(/dev/fastrpc-Xdsp-secure) and 
>>>>>>> non-secure(/dev/fastrpc-Xdsp) devices
>>>>>>>     are created, regardless of this property.
>>>>>>> - Both the nodes can be used for signed and unsigned DSP PD 
>>>>>>> offload.
>>>>>>>
>>>>>>> Note: If the property is not added for CDSP/GDSP, only secure 
>>>>>>> device node can
>>>>>>> be used for signed PD offload, if non-secure device is used, the 
>>>>>>> request gets
>>>>>>> rejected[1].
>>>>>>>
>>>>>>> [1] 
>>>>>>> https://git.kernel.org/pub/scm/linux/kernel/git/next/linux-next.git/tree/drivers/misc/fastrpc.c#n1245
>>>>>>>
>>>>>>> //Ekansh
>>>>>> Does this mean that the qcom,non-secure-domain property should be 
>>>>>> dropped from both nodes?
>>>>> I checked again and found that unsigned module support for CDSP is
>>>>> not available on this platform. Given this, the safest approach would
>>>>> be to add the property for both ADSP and CDSP, ensuring that all
>>>>> created device nodes can be used for signed PD offload. I can provide
>>>> The property allows *unsigned* PD offload though
>>> I don't think I can directly relate this property to unsigned PD 
>>> offload. This is just
>>> defining what type of device node will be created and whether the 
>>> channel is secure
>>> or not. There is a possibility of making unsigned PD request(on 
>>> CDSP/GDSP) irrespective
>>> of whether this property is added or not. If DSP does not support 
>>> unsigned offload, it
>>> should return failures for such requests.
>> Which part of the hardware and/or firmware interface does it define? If
>> it simply declared Linux behaviour, it is incorrect and probably should
>> be dropped.
>>
> When I've removed the qcom,non-secure-domain property from cdsp and 
> tried to run hexagonrpcd via this command:
>
> sudo -u fastrpc hexagonrpcd -f /dev/fastrpc-cdsp  -R 
> /usr/share/qcom/sdm660/Xiaomi/clover/ -d cdsp -c 
> /usr/share/qcom/sdm660/Xiaomi/clover/dsp/cdsp/fastrpc_shell_3
>
> It raised the following error:
>
> qcom,fastrpc 
> 1a300000.remoteproc:glink-edge.fastrpcglink-apps-dsp.-1.-1: Error: 
> Untrusted application trying to offload to signed PD
>
I've tried to add "hlos2_vote_turing_adsp_smmu_clk "as "iface" clock and 
"hlos2_vote_turing_adsp_gdsc" as a second power-domain and cdsp_smmu 
gave the following error:

hlos1_vote_turing_adsp_smmu_clk status stuck at 'off'

>
>
>>>>> a more definitive recommendation once I know the specific use cases
>>>>> you plan to run.
>>>> Why would the usecase affect this?
>>> I'm saying this as per past discussions where some application was 
>>> relying on non-secure
>>> device node on some old platform(on postmarketOS)[1] and having this 
>>> property in place.
>>> So if similar usecase is being enabled here, the property might be 
>>> required[1].
>> DT files are not usecase-based.
>>
>>> [1] https://lkml.org/lkml/2024/8/15/117
>
-- 
Best regards,
Nickolay