On 13.02.2025 4:22 PM, Stephan Gerhold wrote: > On Thu, Feb 13, 2025 at 03:00:00PM +0100, Konrad Dybcio wrote: >> On 13.02.2025 10:13 AM, Stephan Gerhold wrote: >>> On Wed, Feb 12, 2025 at 10:01:59PM +0100, Konrad Dybcio wrote: >>>> On 12.02.2025 6:03 PM, Stephan Gerhold wrote: >>>>> num-channels and qcom,num-ees are required when there are no clocks >>>>> specified in the device tree, because we have no reliable way to read them >>>>> from the hardware registers if we cannot ensure the BAM hardware is up when >>>>> the device is being probed. >>>>> >>>>> This has often been forgotten when adding new SoC device trees, so make >>>>> this clear by describing this requirement in the schema. >>>>> >>>>> Signed-off-by: Stephan Gerhold <stephan.gerhold@xxxxxxxxxx> >>>>> --- >>>>> Documentation/devicetree/bindings/dma/qcom,bam-dma.yaml | 4 ++++ >>>>> 1 file changed, 4 insertions(+) >>>>> >>>>> diff --git a/Documentation/devicetree/bindings/dma/qcom,bam-dma.yaml b/Documentation/devicetree/bindings/dma/qcom,bam-dma.yaml >>>>> index 3ad0d9b1fbc5e4f83dd316d1ad79773c288748ba..5f7e7763615578717651014cfd52745ea2132115 100644 >>>>> --- a/Documentation/devicetree/bindings/dma/qcom,bam-dma.yaml >>>>> +++ b/Documentation/devicetree/bindings/dma/qcom,bam-dma.yaml >>>>> @@ -90,8 +90,12 @@ required: >>>>> anyOf: >>>>> - required: >>>>> - qcom,powered-remotely >>>>> + - num-channels >>>>> + - qcom,num-ees >>>>> - required: >>>>> - qcom,controlled-remotely >>>>> + - num-channels >>>>> + - qcom,num-ees >>>> >>>> I think I'd rather see these deprecated and add the clock everywhere.. >>>> Do we know which one we need to add on newer platforms? Or maybe it's >>>> been transformed into an icc path? >>> >>> This isn't feasible, there are too many different setups. Also often the >>> BAM power management is tightly integrated into the consumer interface. >>> To give a short excerpt (I'm sure there are even more obscure uses): >>> >>> - BLSP BAM (UART, I2C, SPI on older SoCs): >>> 1. Enable GCC_BLSP_AHB_CLK >>> -> This is what the bam_dma driver supports currently. >>> >>> - Crypto BAM: Either >>> OR 1. Vote for single RPM clock >>> OR 1. Enable 3 separate clocks (CE, CE_AHB, CE_AXI) >>> OR 1. Vote dummy bandwidth for interconnect >>> >>> - BAM DMUX (WWAN on older SoCs): >>> 1. Start modem firmware >>> 2. Wait for BAM DMUX service to be up >>> 3. Vote for power up via the BAM-DMUX-specific SMEM state >>> 4. Hope the firmware agrees and brings up the BAM >>> >>> - SLIMbus BAM (audio on some SoCs): >>> 1. Start ADSP firmware >>> 2. Wait for QMI SLIMBUS service to be up via QRTR >>> 3. Vote for power up via SLIMbus-specific QMI messages >>> 4. Hope the firmware agrees and brings up the BAM >>> >>> Especially for the last two, we can't implement support for those >>> consumer-specific interfaces in the BAM driver. Implementing support for >>> the 3 variants of the Crypto BAM would be possible, but it's honestly >>> the least interesting use case of all these. It's not really clear why >>> we are bothing with the crypto engine on newer SoCs at all, see e.g. [1]. >>> >>> [1]: https://lore.kernel.org/linux-arm-msm/20250118080604.GA721573@sol.localdomain/ >>> >>>> Reading back things from this piece of HW only to add it to DT to avoid >>>> reading it later is a really messy solution. >>> >>> In retrospect, it could have been cleaner to avoid describing the BAM as >>> device node independent of the consumer. We wouldn't have this problem >>> if the BAM driver would only probe when the consumer is already ready. >>> >>> But I think specifying num-channels in the device tree is the cleanest >>> way out of this mess. I have a second patch series ready that drops >>> qcom,num-ees and validates the num-channels once it's safe reading from >>> the BAM registers. That way, you just need one boot test to ensure the >>> device tree description is really correct. >> >> Thanks for the detailed explanation! >> >> Do you think it could maybe make sense to expose a clock/power-domain >> from the modem/adsp rproc and feed it to the DMUX / SLIM instances when >> an appropriate ping arrives? This way we'd also defer probing the drivers >> until the device is actually accessible. >> > > Maybe, but that would result in a cyclic dependency between the DMA > provider and consumer. E.g. > > bam_dmux_dma: dma-controller@ { > #dma-cells = <1>; > power-domains = <&bam_dmux>; > }; > > remoteproc@ { > /* ... */ > > bam_dmux: bam-dmux { > dmas = <&bam_dmux_dma 4>, <&bam_dmux_dma 5>; > dma-names = "tx", "rx"; > }; > }; > > fw_devlink will likely get confused by that. > > At the end my thought process here is the following: > > 1. BAM-DMA is a legacy block at this point, it doesn't look like there > are any new use cases being added on new SoCs > 2. We need to preserve compatibility with the old bindings anyway > 3. I trimmed it down to having to specify just "num-channels" > 4. Everything else is read from the hardware registers, and > num-channels gets validated when the first DMA channel is requested > > I think it's the best we can do here at this point. Alright, let's go this route then. Konrad
