Re: [PATCH v4 1/2] Documentation: Documentation for qcom, llcc

[Channa <ckadabi@codeaurora.org>]

On 2018-04-18 07:52, Rob Herring wrote:
> On Tue, Apr 17, 2018 at 5:12 PM,  <rishabhb@codeaurora.org> wrote:
>> On 2018-04-17 10:43, rishabhb@codeaurora.org wrote:
>>> 
>>> On 2018-04-16 07:59, Rob Herring wrote:
>>>> 
>>>> On Tue, Apr 10, 2018 at 01:08:12PM -0700, Rishabh Bhatnagar wrote:
>>>>> 
>>>>> Documentation for last level cache controller device tree bindings,
>>>>> client bindings usage examples.
>>>> 
>>>> 
>>>> "Documentation: Documentation ..."? That wastes a lot of the subject
>>>> line... The preferred prefix is "dt-bindings: ..."
>>>> 
>>>>> 
>>>>> Signed-off-by: Channagoud Kadabi <ckadabi@codeaurora.org>
>>>>> Signed-off-by: Rishabh Bhatnagar <rishabhb@codeaurora.org>
>>>>> ---
>>>>>  .../devicetree/bindings/arm/msm/qcom,llcc.txt      | 58
>>>>> ++++++++++++++++++++++
>>>>>  1 file changed, 58 insertions(+)
>>>>>  create mode 100644
>>>>> Documentation/devicetree/bindings/arm/msm/qcom,llcc.txt
>>>>> 
>>>>> diff --git 
>>>>> a/Documentation/devicetree/bindings/arm/msm/qcom,llcc.txt
>>>>> b/Documentation/devicetree/bindings/arm/msm/qcom,llcc.txt
>>>>> new file mode 100644
>>>>> index 0000000..497cf0f
>>>>> --- /dev/null
>>>>> +++ b/Documentation/devicetree/bindings/arm/msm/qcom,llcc.txt
>>>>> @@ -0,0 +1,58 @@
>>>>> +== Introduction==
>>>>> +
>>>>> +LLCC (Last Level Cache Controller) provides last level of cache 
>>>>> memory
>>>>> in SOC,
>>>>> +that can be shared by multiple clients. Clients here are different
>>>>> cores in the
>>>>> +SOC, the idea is to minimize the local caches at the clients and
>>>>> migrate to
>>>>> +common pool of memory
>>>>> +
>>>>> +Properties:
>>>>> +- compatible:
>>>>> +        Usage: required
>>>>> +        Value type: <string>
>>>>> +        Definition: must be "qcom,sdm845-llcc"
>>>>> +
>>>>> +- reg:
>>>>> +        Usage: required
>>>>> +        Value Type: <prop-encoded-array>
>>>>> +        Definition: must be addresses and sizes of the LLCC 
>>>>> registers
>>>> 
>>>> 
>>>> How many address ranges?
>>>> 
>>> It consists of just one address range. I'll edit the definition to 
>>> make
>>> it more clear.
>>>>> 
>>>>> +
>>>>> +- #cache-cells:
>>>> 
>>>> 
>>>> This is all written as it is a common binding, but it is not one.
>>>> 
>>>> You already have most of the configuration data for each client in 
>>>> the
>>>> driver, I think I'd just put the client connection there too. Is 
>>>> there
>>>> any variation of this for a given SoC?
>>>> 
>>> #cache-cells and max-slices won't change for a given SOC. So you want 
>>> me
>>> to hard-code in the driver itself?
>>> 
>> I can use of_parse_phandle_with_fixed_args function and fix the number 
>> of
>> args as 1 instead of keeping #cache-cells here in DT. Does that look 
>> fine?
> 
> No, I'm saying why even put cache-slices properties in DT to begin
> with? You could just define client id's within the kernel and clients
> can use those instead of getting the id from the DT.

The reason to add cache-slices here is to establish a connection between
client and system cache. For example if we have multiple instances of
system cache blocks and client wants to choose a system cache instance
based on the usecase then its easier to establish this connection using
device tree than hard coding in the driver.

> 
> I have a couple of hesitations with putting this into the DT. First, I
> think a cache is just one aspect of describing the interconnect
> between masters and memory (and there's been discussions on
> interconnect bindings too) and any binding needs to consider all of
> the aspects of the interconnect. Second, I'd expect this cache
> architecture will change SoC to SoC and the binding here is pretty
> closely tied to the current cache implementation (e.g. slices). If
> there were a bunch of SoCs with the same design and just different
> client IDs (like interrupt IDs), then I'd feel differently.

This is partially true, a bunch of SoCs would support this design but
clients IDs are not expected to change. So Ideally client drivers could
hard code these IDs.

However I have other concerns of moving the client Ids in the driver.
The way the APIs implemented today are as follows:
#1. Client calls into system cache driver to get cache slice handle
with the usecase Id as input.
#2. System cache driver gets the phandle of system cache instance from
the client device to obtain the private data.
#3. Based on the usecase Id perform look up in the private data to get
cache slice handle.
#4. Return the cache slice handle to client

If we don't have the connection between client & system cache then the
private data needs to declared as static global in the system cache 
driver,
that limits us to have just once instance of system cache block.


> 
> Rob

-- 
--
The Qualcomm Innovation Center, Inc. is a member of the Code Aurora 
Forum,
a Linux Foundation Collaborative Project