Message-ID: <f2b11a1f-8085-45e8-a3f5-b07477b9ef91@oss.qualcomm.com>
Date: Tue, 3 Feb 2026 12:39:23 +0100
From: Konrad Dybcio <konrad.dybcio@....qualcomm.com>
To: Dmitry Torokhov <dmitry.torokhov@...il.com>
Cc: Luca Weiss <luca.weiss@...rphone.com>,
        Griffin Kroah-Hartman <griffin.kroah@...rphone.com>,
        Rob Herring <robh@...nel.org>,
        Krzysztof Kozlowski <krzk+dt@...nel.org>,
        Conor Dooley <conor+dt@...nel.org>,
        Bjorn Andersson <andersson@...nel.org>,
        Konrad Dybcio <konradybcio@...nel.org>, linux-input@...r.kernel.org,
        devicetree@...r.kernel.org, linux-kernel@...r.kernel.org,
        linux-arm-msm@...r.kernel.org
Subject: Re: [PATCH v2 2/3] Input: aw86938 - add driver for Awinic AW86938

On 2/3/26 10:49 AM, Dmitry Torokhov wrote:
> On Mon, Feb 02, 2026 at 04:11:51PM +0100, Konrad Dybcio wrote:
>> On 2/2/26 12:04 PM, Dmitry Torokhov wrote:
>>> On Mon, Feb 02, 2026 at 11:19:36AM +0100, Konrad Dybcio wrote:
>>>> On 2/2/26 11:14 AM, Luca Weiss wrote:
>>>>> Hi Konrad,
>>>>>
>>>>> On Mon Feb 2, 2026 at 11:12 AM CET, Konrad Dybcio wrote:
>>>>>> On 2/1/26 2:49 AM, Dmitry Torokhov wrote:
>>>>>>> Hi Griffin,
>>>>>>>
>>>>>>> On Wed, Jan 28, 2026 at 04:51:14PM +0100, Griffin Kroah-Hartman wrote:
>>>>>>>> @@ -717,9 +746,19 @@ static int aw86927_detect(struct aw86927_data *haptics)
>>>>>>>>  
>>>>>>>>  	chip_id = be16_to_cpu(read_buf);
>>>>>>>>  
>>>>>>>> -	if (chip_id != AW86927_CHIPID) {
>>>>>>>> -		dev_err(haptics->dev, "Unexpected CHIPID value 0x%x\n", chip_id);
>>>>>>>> -		return -ENODEV;
>>>>>>>> +	switch (haptics->model) {
>>>>>>>> +	case AW86927:
>>>>>>>> +		if (chip_id != AW86927_CHIPID) {
>>>>>>>> +			dev_err(haptics->dev, "Unexpected CHIPID value 0x%x\n", chip_id);
>>>>>>>> +			return -ENODEV;
>>>>>>>> +		}
>>>>>>>
>>>>>>> If we are able to query chip ID why do we need to have separate
>>>>>>> compatibles? I would define chip data structure with differences between
>>>>>>> variants and assign and use it instead of having separate compatible.
>>>>>>
>>>>>> dt-bindings guidelines explicitly call for this, a chipid comparison
>>>>>> then works as a safety net
>>>>>
>>>>> Are you saying, that
>>>>>
>>>>> 1. we should enforce dt-bindings == CHIP_ID (what's currently done)
>>>>
>>>> This
>>>
>>> No. If there is a compatible chip with different ID (for whatever reason
>>> - maybe there is additional functionality that either board does not
>>> need or the driver does not implement) we absolutely should not refuse
>>> to bind the driver.
>>>
>>> Hint: this thing is called _compatible_ for a reason.
>>
>> Right, the reason you have in mind is fulfilled by fallback compatibles
>>
>> (i.e. "vendor,actualchipname", "vendor,similarchipname" where the driver
>> only considers the latter becuase the software interface hasn't changed)
> 
> And having chip_id checks will break this...

Depends on how they're implemented and how different the chips are.

If the software interface is exactly 1:1, with the only difference in the
ID register, something like

if (chipid == SIMILARCHIPNAME_ID || chipid == ACTUALCHIPNAME_ID)
	...

would be fitting.

However, more often than not, like in this case, there's actual differences
that need to be taken into account, meaning we already need to act upon
the "actual" compatible

> 
>>
>>>
>>>>
>>>>>
>>>>> or
>>>>>
>>>>> 2. we should have both compatibles with no handling based on compatible,
>>>>>    but only use CHIP_ID at runtime to change behavior
>>>>
>>>> This is spaghetti
>>>
>>> I really do not understand the aversion of DT maintainers to generic
>>> compatibles. We see this in I2C HID where we keep adding compatibles
>>> for what could be described via device properties.
>>
>> This is because it's the only way to allow for retroactive changes that
>> do not require changing firmware. That's why ACPI carries new identifiers
>> for even very slightly different devices too. Once the firmware containing
>> (ACPI tables / DTB) is put on a production device, it is generally not
>> going to ever change.
> 
> They are actually solving slightly different problem. In ACPI world they
> allocate a new ID to represent a peripheral in a given design, down to
> it's firmware behavior. It encodes much more than chip ID that DT
> maintainers want to key off of.

DT sort of does this to. In the Qualcomm world, how you get to interact
with the platform changes dramatically depending on the firmware flavor
it's flashed with (which I'd hope to see go away one day..) - if you have
a Chrome firmware, you're basically free/required to configure everything
from Linux. With Android firmware, much of that heavy lifting must be done
by the hypervisor or the secure world. With Windows firmware, you get the
Android experience + UEFI services. And at the tail end of the scale,
there's the automotive firmware where you don't even get to toggle clocks,
but instead all peripherals' power and performance states are exposed
through dozens of SCMI servers..

Somehow we can try and be smart, deducing the behavior based on the
properties present in DT, but often times, a separate compatible for
"this SoC except with this firmware" needs to exist, as the OS-accessible
software interface is simply different, as if this wasn't the same SoC
anymore.

> 
>>
>> CHIP_ID registers are a good tool to validate that the author of the
>> firmware table is doing the right thing, but solely relying on them
>> encourages creating a "vendor,haptic" compatible, which I'm sure you'll
>> agree is totally meaningless.
> 
> Is it? If a piece of hardware speaks i2c-hid protocol why do I need to
> know the exact chip that is being used? Depending on the chassis and the
> size of the sensing element and the version of the firmware that is
> loaded into it the behavior and timings of the same chip may be very
> different.

I agree this argument gets overused at times

>>
>> That's especially if the naming scheme makes no sense and you can't
>> even factor out a common wildcard-name (which also happens to be the case
>> quite often)
>>
>> Plus a compatible is used to restrict/modify the set of allowed/required
>> properties, so having an "actual" compatible is required for schema
>> validation to work
> 
> Yes, in cases where there is not a common set of properties having
> different compatibles makes sense. But in cases when the device is
> supposed to have vendor-agnostic behavior insisting on myriad
> compatibles makes little sense.

Some people may be thrown off by the golden rule of implementing standards,
which is to break or bend them immediately, claiming full compatibility.

But for cases like hid-over-i2c, I symphatize with the "no one needs
to know if it's a synaptics a00001 or a synaptics a00002" sentiment

Konrad

Powered by blists - more mailing lists