RFI: I2c muxes I2C_MUX_LOCKED and interrupt support

RFI: I2c muxes I2C_MUX_LOCKED and interrupt support

[Phil Reid]

G'day Peter,

I'm looking into modifying the i2c-mux-pca954x driver to add support for
the pca_9543 interrupt mux function.

So the first thing I need to add is a reg read function.
However based on the changes to the i2c mux code in the 4.6 series the
locking work around shouldn't be needed now if the mux is allocated with
I2C_MUX_LOCKED. Currently this driver is not doing this.
Also the same with the similar i2c-mux-pca9541 driver which does implement read.

So my question is should I change the driver to use I2C_MUX_LOCKED
or implement the read operation the same as the i2c-mux-pca9541?



-- 
Regards
Phil Reid

Re: RFI: I2c muxes I2C_MUX_LOCKED and interrupt support

[Peter Rosin]

Hi Phil,

On 2016-07-21 05:20, Phil Reid wrote:
> G'day Peter,
> 
> I'm looking into modifying the i2c-mux-pca954x driver to add support for
> the pca_9543 interrupt mux function.
> 
> So the first thing I need to add is a reg read function.
> However based on the changes to the i2c mux code in the 4.6 series the
> locking work around shouldn't be needed now if the mux is allocated with
> I2C_MUX_LOCKED. Currently this driver is not doing this.
> Also the same with the similar i2c-mux-pca9541 driver which does implement read.
> 
> So my question is should I change the driver to use I2C_MUX_LOCKED
> or implement the read operation the same as the i2c-mux-pca9541?

Good question. I didn't dare changing the pca9541/pca954x drivers to
be mux locked. Maybe I am too conservative?

The issue is that if you have a multi-level hierarchy of muxes, the rules
are more relaxed for mux locked muxed compared to adapter locked muxes.

I.e.
             mux3
            /
        mux1
       /    \
   root      mux4
       \
        mux2

accesses to devices on e.g. mux3 and mux2 may interleave if all muxes are
mux-locked, that will never happen for adapter-locked muxes.

Building complex hierarchies feels more likely with pca954x that with the
other muxing options. But I don't know that, and maybe none exist at all?

Anyway, the safe option is to do it like in pca9541...

Cheers,
Peter

Re: RFI: I2c muxes I2C_MUX_LOCKED and interrupt support

[Phil Reid]

On 21/07/2016 14:52, Peter Rosin wrote:
> Hi Phil,
>
> On 2016-07-21 05:20, Phil Reid wrote:
>> G'day Peter,
>>
>> I'm looking into modifying the i2c-mux-pca954x driver to add support for
>> the pca_9543 interrupt mux function.
>>
>> So the first thing I need to add is a reg read function.
>> However based on the changes to the i2c mux code in the 4.6 series the
>> locking work around shouldn't be needed now if the mux is allocated with
>> I2C_MUX_LOCKED. Currently this driver is not doing this.
>> Also the same with the similar i2c-mux-pca9541 driver which does implement read.
>>
>> So my question is should I change the driver to use I2C_MUX_LOCKED
>> or implement the read operation the same as the i2c-mux-pca9541?
>
> Good question. I didn't dare changing the pca9541/pca954x drivers to
> be mux locked. Maybe I am too conservative?
>
> The issue is that if you have a multi-level hierarchy of muxes, the rules
> are more relaxed for mux locked muxed compared to adapter locked muxes.
>
> I.e.
>              mux3
>             /
>         mux1
>        /    \
>    root      mux4
>        \
>         mux2
>
> accesses to devices on e.g. mux3 and mux2 may interleave if all muxes are
> mux-locked, that will never happen for adapter-locked muxes.
>
> Building complex hierarchies feels more likely with pca954x that with the
> other muxing options. But I don't know that, and maybe none exist at all?
>
> Anyway, the safe option is to do it like in pca9541...
>
G'day Peter

Thanks for the explanation.

However I've thought about this a bit more as I've started implementation.
The irq status reading probably doesn't need to got thru the lock work around
as they won't be getting called in the mux select / release functions.

Data read will occur on a threaded interrupt request. Which would be a similar
context to the drivers resume function which directly calls i2c_smbus_write_byte.

Is my thinking right here?


-- 
Regards
Phil Reid

Re: RFI: I2c muxes I2C_MUX_LOCKED and interrupt support

[Peter Rosin]

On 2016-07-21 11:59, Phil Reid wrote:
> On 21/07/2016 14:52, Peter Rosin wrote:
>> Hi Phil,
>>
>> On 2016-07-21 05:20, Phil Reid wrote:
>>> G'day Peter,
>>>
>>> I'm looking into modifying the i2c-mux-pca954x driver to add support for
>>> the pca_9543 interrupt mux function.
>>>
>>> So the first thing I need to add is a reg read function.
>>> However based on the changes to the i2c mux code in the 4.6 series the
>>> locking work around shouldn't be needed now if the mux is allocated with
>>> I2C_MUX_LOCKED. Currently this driver is not doing this.
>>> Also the same with the similar i2c-mux-pca9541 driver which does implement read.
>>>
>>> So my question is should I change the driver to use I2C_MUX_LOCKED
>>> or implement the read operation the same as the i2c-mux-pca9541?
>>
>> Good question. I didn't dare changing the pca9541/pca954x drivers to
>> be mux locked. Maybe I am too conservative?
>>
>> The issue is that if you have a multi-level hierarchy of muxes, the rules
>> are more relaxed for mux locked muxed compared to adapter locked muxes.
>>
>> I.e.
>>              mux3
>>             /
>>         mux1
>>        /    \
>>    root      mux4
>>        \
>>         mux2
>>
>> accesses to devices on e.g. mux3 and mux2 may interleave if all muxes are
>> mux-locked, that will never happen for adapter-locked muxes.
>>
>> Building complex hierarchies feels more likely with pca954x that with the
>> other muxing options. But I don't know that, and maybe none exist at all?
>>
>> Anyway, the safe option is to do it like in pca9541...
>>
> G'day Peter
> 
> Thanks for the explanation.
> 
> However I've thought about this a bit more as I've started implementation.
> The irq status reading probably doesn't need to got thru the lock work around
> as they won't be getting called in the mux select / release functions.
> 
> Data read will occur on a threaded interrupt request. Which would be a similar
> context to the drivers resume function which directly calls i2c_smbus_write_byte.
> 
> Is my thinking right here?

Yes, I think so. Accesses to registers in the device itself is independent from
muxing, hopefully...

Cheers,
Peter