On 2016-09-20 12:07, Bartosz Golaszewski wrote:
> 2016-09-20 10:48 GMT+02:00 Peter Rosin <p...@axentia.se>:
>> One pretty simple problematic case is:
>>   .---.          .----.
>>   |   |          |    |-- i2c2
>>   |   |-- i2c0 --|mux0|          .----.
>>   | l |          |    |-- i2c3 --|gpio|
>>   | i |          '----'          '----'
>>   | n |             .--------------'
>>   | u |          .----.          .----.
>>   | x |          |    |-- i2c4 --|dev0|
>>   |   |-- i2c1 --|mux1|          '----'
>>   |   |          |    |-- i2c5
>>   '---'          '----'
>> Accesses to dev0 will:
>> 1. lock i2c1:mux_lock (depth 0)
>> 2. switch mux1 to i2c4 using gpio
>>  a lock i2c0:mux_lock (depth 0)
>>  b switch mux0 to i2c3 using whatever
>>  c access gpio
>>  d unlock i2c0:mux_lock
>> 3. access dev0
>> 4. unlock i2c1:mux_lock
>> 2a will cause a lockdep splat if i2c0:mux_lock is in the same
>> lockdep class & subclass as i2c1:mux_lock. So, lockdep needs
>> separate lockdep classes depending on the i2c root adapter
>> (subclasses are needed to handle deeper trees, so they are off
>> limits). Great fun. How do I go about creating a new lockdep
>> class for every i2c root adapter instance?
> I feel like it's just wrong to set an arbitrary limit on the number of
> i2c branches - and this is what the result of this approach would be.

What arbitrary limit would that be? The number of lockdep classes
can't be *that* limited? Or?

I mean one lockdep class per root adapter and one subclass within that
class per mux level doesn't sound too bad. How many root adapters do
we need to design for?

> One solution that comes to mind is to have a separate, global set of
> lock classes solely for gpio expanders. I think you mentioned earlier
> that it's the only thing that can cause this kind of lockdep false
> positives.

No, that's not true, so if I said that, it was simply wrong. E.g. I
think the new mlxcpld i2c mux can mux another i2c adapter and will
cause the same problem without involving any gpio expander.

[OT: Wolfram, are you still following this? Are you picking up the
 mlxcpld mux driver, or did my ack fly by without you noticing?]

Oh, and the there is a pinctrl-based i2c mux that also suffers as
there are pinctrl chips that are i2c clients.

And even though I don't really know the pinctrl demux, it will
probably also cause interesting nesting if an i2c based pinctrl
is used as demuxer. If that's relevant, probably isn't...

>            We could potentially have a limited set of lock classes and
> every expander that would need one would request it using some kind of
> API ensuring that every instance gets a separate class. But this
> sounds like a big hack too I'm afraid... And regmap would need to be
> aware of that as well.
> Anyways, we're past rc7 already and 4.9 will be the next LTS kernel.
> We have real hardware here that runs on mainline linux and is
> suffering from this issue. Are there any objections against merging
> this series now and continuing the work on improving the solution for
> 4.10?

No, I have no objection. It's not wrong per se, it's just not complete.


Reply via email to