On Thu, Oct 10, 2024 at 09:15:59AM +0200, Krzysztof Kozlowski wrote: > On 09/10/2024 16:04, Shiraz Hashim wrote: > > On Mon, Oct 07, 2024 at 06:25:01PM +0200, Dmitry Baryshkov wrote: > >> On Mon, 7 Oct 2024 at 17:35, Mukesh Ojha <quic_mojha@xxxxxxxxxxx> wrote: > >>> > >>> On Sun, Oct 06, 2024 at 10:38:01PM +0300, Dmitry Baryshkov wrote: > >>>> On Sat, Oct 05, 2024 at 02:53:54AM GMT, Mukesh Ojha wrote: > >>>>> From: Shiraz Hashim <quic_shashim@xxxxxxxxxxx> > >>>>> > >>>>> Qualcomm’s PAS implementation for remote processors only supports a > >>>>> single stage of IOMMU translation and is presently managed by the > >>>>> Qualcomm EL2 hypervisor (QHEE) if it is present. In the absence of QHEE, > >>>>> such as with a KVM hypervisor, IOMMU translations need to be set up by > >>>>> the KVM host. Remoteproc needs carveout memory region and its resource > >>>>> (device memory) permissions to be set before it comes up, and this > >>>>> information is presently available statically with QHEE. > >>>>> > >>>>> In the absence of QHEE, the boot firmware needs to overlay this > >>>>> information based on SoCs running with either QHEE or a KVM hypervisor > >>>>> (CPUs booted in EL2). > >>>>> > >>>>> The qcom,devmem property provides IOMMU devmem translation information > >>>>> intended for non-QHEE based systems. > >>>>> > >>>>> Signed-off-by: Shiraz Hashim <quic_shashim@xxxxxxxxxxx> > >>>>> Co-Developed-by: Mukesh Ojha <quic_mojha@xxxxxxxxxxx> > >>>>> Signed-off-by: Mukesh Ojha <quic_mojha@xxxxxxxxxxx> > >>>>> --- > >>>>> .../bindings/remoteproc/qcom,pas-common.yaml | 42 +++++++++++++++++++ > >>>>> .../bindings/remoteproc/qcom,sa8775p-pas.yaml | 20 +++++++++ > >>>>> 2 files changed, 62 insertions(+) > >>>>> > >>>>> diff --git a/Documentation/devicetree/bindings/remoteproc/qcom,pas-common.yaml b/Documentation/devicetree/bindings/remoteproc/qcom,pas-common.yaml > >>>>> index 63a82e7a8bf8..068e177ad934 100644 > >>>>> --- a/Documentation/devicetree/bindings/remoteproc/qcom,pas-common.yaml > >>>>> +++ b/Documentation/devicetree/bindings/remoteproc/qcom,pas-common.yaml > >>>>> @@ -52,6 +52,48 @@ properties: > >>>>> minItems: 1 > >>>>> maxItems: 3 > >>>>> > >>>>> + iommus: > >>>>> + maxItems: 1 > >>>>> + > >>>>> + qcom,devmem: > >>>>> + $ref: /schemas/types.yaml#/definitions/uint32-matrix > >>>>> + description: > >>>>> + Qualcomm’s PAS implementation for remote processors only supports a > >>>>> + single stage of IOMMU translation and is presently managed by the > >>>>> + Qualcomm EL2 hypervisor (QHEE) if it is present. In the absence of QHEE, > >>>>> + such as with a KVM hypervisor, IOMMU translations need to be set up by > >>>>> + the KVM host. Remoteproc might need some device resources and related > >>>>> + access permissions to be set before it comes up, and this information is > >>>>> + presently available statically with QHEE. > >>>>> + > >>>>> + In the absence of QHEE, the boot firmware needs to overlay this > >>>>> + information based on SoCs running with either QHEE or a KVM hypervisor > >>>>> + (CPUs booted in EL2). > >>>>> + > >>>>> + The qcom,devmem property provides IOMMU devmem translation information > >>>>> + intended for non-QHEE based systems. It is an array of u32 values > >>>>> + describing the device memory regions for which IOMMU translations need to > >>>>> + be set up before bringing up Remoteproc. This array consists of 4-tuples > >>>>> + defining the device address, physical address, size, and attribute flags > >>>>> + with which it has to be mapped. > >>>> > >>>> I'd expect that this kind of information is hardware-dependent. As such > >>>> it can go to the driver itself, rather than the device tree. The driver > >>>> can use compatible string to select the correct table. > >>>> > >>> > >>> IIUC, are you saying that to move this into driver file and override the > >>> compatible string via overlay ? > >> > >> Ideally we should live without compat overrides. On the other hand, > >> sc7180 and sc7280 provide an example of doing exactly that. > > > > I am not sure if there can arise a case where updated adsp firmware > > for particular board(s) may require additional access. > > > > Having it in device tree adds a convenience to deal with such > > variance. > > > > That's a downstream argument... Just look at the downstream DTS. > Everything, even software properties, can be added to DT, right? I was thinking this binding is similar to iommu-addresses approach, however that is under reserved-memory enumerating memory regions. regards Shiraz
