Re: [PATCH 5/6] soudnwire: master: protect concurrecnt check for bus->md

[Pierre-Louis Bossart]

On 4/20/23 12:27, Krzysztof Kozlowski wrote:
> On 20/04/2023 18:42, Pierre-Louis Bossart wrote:
>> typos in commit title...
>>
>> On 4/20/23 05:16, Krzysztof Kozlowski wrote:
>>> The Soundwire master controllers might want to check for bus->md
>>
>> Apologies for being pedantic but 'manager' and 'controller' are
>> different concepts in SoundWire, see DisCo spec.
>> It's not a 1:1 mapping, a controller can rely on M managers
> 
> I wrote master, not manager. For the Qualcomm case one controller is one
> master, but in general I try to avoid the master/slave terminology.

The Soundwire 1.2.1 spec moved away from master/slave and uses
manager/peripheral. It's the same concepts, just different terms. At
some point we'll update the code, it's just been too busy in 2022/2023
to do this replacement. It doesn't hurt to use the new terms.

>>> initialization to avoid race between early interrupt and finish of
>>> sdw_bus_master_add()/sdw_master_device_add().  Such early interrupt can
>>> happen if Soundwire devices are not powered off during their probe.
>>>
>>> Add a store release barrier, so the Soundwire controllers can safely
>>> check it in concurrent (e.g. in interrupt) way.
>>
>> Can you elaborate on the race condition? I am not following what breaks,
>> and what entity generates the 'early interrupt'.
> 
> The condition is explained in next patch. If you think it's better, I
> can squash it with next.
> 
> If the condition is still not clear, drop a note in next patch, so I
> will elaborate there.

will do.

>> I am specifically concerned about adding this in common code without any
>> matching smp_load_acquire() - which is only added in the following patch
>> for the Qualcomm manager only, but not added for Intel/AMD managers. Is
>> this not a problem?
> 
> Shouldn't be. The barrier just won't be effective for these drivers, but
> that should not be a problem, because I also did not add to these
> checking bus->md in a concurrent path.
> 
> Basically the barrier here is necessary because I want to check bus->md
> in Qualcomm master interrupt handler.

I really don't have any understanding or background on what this does.

Is there actually a precedent for this? I mean, dealing with the
device/driver model is already complicated, if now we have to be careful
on when the device pointer is stored it adds a whole new element of
complexity or skillset required to understand the bus operation.

Re-looking at the code, the 'md' variable is allocated in
sdw_master_device_add(), initialized with all kinds of values, used by
device_register() so presumably when you store the value it's stored
somewhere consistent, no?