Re: [PATCH v7 1/5] dt-bindings: interconnect: Add Qualcomm IPQ9574 support

[Krzysztof Kozlowski <krzk@kernel.org>]

On 09/04/2024 13:03, Varadarajan Narayanan wrote:
> On Tue, Apr 09, 2024 at 11:45:51AM +0200, Krzysztof Kozlowski wrote:
>> On 09/04/2024 09:41, Varadarajan Narayanan wrote:
>>> On Thu, Apr 04, 2024 at 02:25:06PM +0530, Varadarajan Narayanan wrote:
>>>> On Wed, Apr 03, 2024 at 04:59:40PM +0200, Krzysztof Kozlowski wrote:
>>>>> On 03/04/2024 12:42, Varadarajan Narayanan wrote:
>>>>>> Add interconnect-cells to clock provider so that it can be
>>>>>> used as icc provider.
>>>>>>
>>>>>> Add master/slave ids for Qualcomm IPQ9574 Network-On-Chip
>>>>>> interfaces. This will be used by the gcc-ipq9574 driver
>>>>>> that will for providing interconnect services using the
>>>>>> icc-clk framework.
>>>>>>
>>>>>> Signed-off-by: Varadarajan Narayanan <quic_varada@quicinc.com>
>>>>>> ---
>>>>>> v7:
>>>>>> Fix macro names to be consistent with other bindings
>>>>>> v6:
>>>>>> Removed Reviewed-by: Krzysztof Kozlowski
>>>>>> Redefine the bindings such that driver and DT can share them
>>>>>>
>>>>>> v3:
>>>>>> Squash Documentation/ and include/ changes into same patch
>>>>>>
>>>>>> qcom,ipq9574.h
>>>>>> 	Move 'first id' to clock driver
>>>>>>
>>>>>> ---
>>>>>>  .../bindings/clock/qcom,ipq9574-gcc.yaml      |  3 +
>>>>>>  .../dt-bindings/interconnect/qcom,ipq9574.h   | 87 +++++++++++++++++++
>>>>>>  2 files changed, 90 insertions(+)
>>>>>>  create mode 100644 include/dt-bindings/interconnect/qcom,ipq9574.h
>>>>>>
>>>>>> diff --git a/Documentation/devicetree/bindings/clock/qcom,ipq9574-gcc.yaml b/Documentation/devicetree/bindings/clock/qcom,ipq9574-gcc.yaml
>>>>>> index 944a0ea79cd6..824781cbdf34 100644
>>>>>> --- a/Documentation/devicetree/bindings/clock/qcom,ipq9574-gcc.yaml
>>>>>> +++ b/Documentation/devicetree/bindings/clock/qcom,ipq9574-gcc.yaml
>>>>>> @@ -33,6 +33,9 @@ properties:
>>>>>>        - description: PCIE30 PHY3 pipe clock source
>>>>>>        - description: USB3 PHY pipe clock source
>>>>>>
>>>>>> +  '#interconnect-cells':
>>>>>> +    const: 1
>>>>>> +
>>>>>>  required:
>>>>>>    - compatible
>>>>>>    - clocks
>>>>>> diff --git a/include/dt-bindings/interconnect/qcom,ipq9574.h b/include/dt-bindings/interconnect/qcom,ipq9574.h
>>>>>> new file mode 100644
>>>>>> index 000000000000..0b076b0cf880
>>>>>> --- /dev/null
>>>>>> +++ b/include/dt-bindings/interconnect/qcom,ipq9574.h
>>>>>> @@ -0,0 +1,87 @@
>>>>>> +/* SPDX-License-Identifier: (GPL-2.0-only OR BSD-2-Clause) */
>>>>>> +#ifndef INTERCONNECT_QCOM_IPQ9574_H
>>>>>> +#define INTERCONNECT_QCOM_IPQ9574_H
>>>>>> +
>>>>>> +#define ICC_ANOC_PCIE0		0
>>>>>> +#define ICC_SNOC_PCIE0		1
>>>>>> +#define ICC_ANOC_PCIE1		2
>>>>>> +#define ICC_SNOC_PCIE1		3
>>>>>> +#define ICC_ANOC_PCIE2		4
>>>>>> +#define ICC_SNOC_PCIE2		5
>>>>>> +#define ICC_ANOC_PCIE3		6
>>>>>> +#define ICC_SNOC_PCIE3		7
>>>>>> +#define ICC_SNOC_USB		8
>>>>>> +#define ICC_ANOC_USB_AXI	9
>>>>>> +#define ICC_NSSNOC_NSSCC	10
>>>>>> +#define ICC_NSSNOC_SNOC_0	11
>>>>>> +#define ICC_NSSNOC_SNOC_1	12
>>>>>> +#define ICC_NSSNOC_PCNOC_1	13
>>>>>> +#define ICC_NSSNOC_QOSGEN_REF	14
>>>>>> +#define ICC_NSSNOC_TIMEOUT_REF	15
>>>>>> +#define ICC_NSSNOC_XO_DCD	16
>>>>>> +#define ICC_NSSNOC_ATB		17
>>>>>> +#define ICC_MEM_NOC_NSSNOC	18
>>>>>> +#define ICC_NSSNOC_MEMNOC	19
>>>>>> +#define ICC_NSSNOC_MEM_NOC_1	20
>>>>>> +
>>>>>> +#define ICC_NSSNOC_PPE		0
>>>>>> +#define ICC_NSSNOC_PPE_CFG	1
>>>>>> +#define ICC_NSSNOC_NSS_CSR	2
>>>>>> +#define ICC_NSSNOC_IMEM_QSB	3
>>>>>> +#define ICC_NSSNOC_IMEM_AHB	4
>>>>>> +
>>>>>> +#define MASTER_ANOC_PCIE0		(ICC_ANOC_PCIE0 * 2)
>>>>>> +#define SLAVE_ANOC_PCIE0		((ICC_ANOC_PCIE0 * 2) + 1)
>>>>>
>>>>> Which existing Qualcomm platform has such code?
>>>>
>>>> Existing Qualcomm platforms don't use icc-clk. They use icc-rpm
>>>> or icc-rpmh. clk-cbf-msm8996.c is the only driver that uses icc-clk.
>>>>
>>>> The icc_clk_register automatically creates master & slave nodes
>>>> for each clk entry provided as input with the node-ids 'n' and
>>>> 'n+1'. Since clk-cbf-msm8996.c has only one entry, it could just
>>>> define MASTER_CBF_M4M and SLAVE_CBF_M4M with 0 and 1 and avoid these
>>>> calculations.
>>>>
>>>> However, ipq9574 gives an array of clock entries as input to
>>>> icc_clk_register. To tie the order/sequence of these clock
>>>> entries correctly with the node-ids, this calculation is needed.
>>>>
>>>>> This is the third time I am asking for consistent headers. Open
>>>>> existing, recently added headers and look how it is done there. Why?
>>>>> Because I am against such calculations and see no reason for them.
>>>>
>>>> Apologies. Regret that I have to trouble you.
>>>>
>>>> In this ipq9574 case, have to reconcile between the following
>>>> feedbacks.
>>>>
>>>> 1. https://lore.kernel.org/linux-arm-msm/fe40b307-26d0-4b2a-869b-5d093415b9d1@linaro.org/
>>>>    We could probably use indexed identifiers here to avoid confusion:
>>>>    [ICC_BINDING_NAME] = CLK_BINDING_NAME
>>>>
>>>> 2. https://lore.kernel.org/linux-arm-msm/95f4e99a60cc97770fc3cee850b62faf.sboyd@kernel.org/
>>>>    Are these supposed to be in a dt-binding header?
>>>>
>>>> 3. https://lore.kernel.org/linux-arm-msm/031d0a35-b192-4161-beef-97b89d5d1da6@linaro.org/
>>>>    Do you use them as well in the DTS?
>>>>
>>>> Having the defines (with the calculations) seemed to to comply
>>>> with the above three feedbacks.
>>>>
>>>> Please let me know if this can be handled in a different way that
>>>> would be consistent with other Qualcomm platforms.
>>>
>>> Krzysztof,
>>>
>>> Is this ok? Can I post a new version addressing other review comments?
>>
>> I don't understand and you did not answered before, why you have to do
>> it differently than all other Qualcomm interconnect providers. Maybe the
>> code here needs it, maybe not, but I don't see any argument proving this.
> 
> Other Qualcomm interconnect providers use the icc-rpm.
> 
> 	1. The SoC specific interconnect providers have control
> 	   over the master/slave id-numbers and is hard coded.
> 
> 	2. These id-numbers are used by the RPM firmware.
> 
> IPQ9574 uses icc-clk.
> 
> 	1. The ipq9574 specific interconnect provider doesn't
> 	   have control over the master/slave id-numbers. The
> 	   icc-clk framework auto generates it in the order of
> 	   the clock entries given as input.

Okay, so what happens if icc-clk way of generating them changes a bit?
It can change, why not, driver implementation is not an ABI.

> 
> 	2. These auto-generated id-numbers have to be correctly
> 	   tied to the DT nodes. Else, the relevant clocks may
> 	   not get enabled.

Sorry, I don't get, how auto generated ID number is tied to DT node.
What DT node?

> 
> Since ICC-CLK creates two ids per clock entry (one MASTER_xxx and
> one SLAVE_xxx), using those MASTER/SLAVE_xxx macros as indices in
> the below array would create holes.
> 
> 	static int icc_ipq9574_hws[] = {
> 		[MASTER_ANOC_PCIE0] = GCC_ANOC_PCIE0_1LANE_M_CLK,
> 		[MASTER_SNOC_PCIE0] = GCC_SNOC_PCIE0_1LANE_S_CLK,
> 		[MASTER_ANOC_PCIE1] = GCC_ANOC_PCIE1_1LANE_M_CLK,
> 		[MASTER_SNOC_PCIE1] = GCC_SNOC_PCIE1_1LANE_S_CLK,
> 		. . .
> 	};
> 
> Other Qualcomm drivers don't have this issue and they can
> directly use the MASTER/SLAVE_xxx macros.

I understand, thanks, yet your last patch keeps adding fake IDs, means
IDs which are not part of ABI.

> 
> As the MASTER_xxx macros cannot be used, have to define a new set
> of macros that can be used for indices in the above array. This
> is the reason for the ICC_BINDING_NAME macros.

Then maybe fix the driver, instead of adding something which is not an
ABI to bindings and completely skipping the actual ABI.

Best regards,
Krzysztof