Re: [PATCH v2 0/4] gpio: fix an incorrect lockdep warning

From: Peter Rosin
Date: Mon Sep 19 2016 - 11:37:20 EST

On 2016-09-18 21:45, Bartosz Golaszewski wrote:
> 2016-09-18 21:43 GMT+02:00 Bartosz Golaszewski <bgolaszewski@xxxxxxxxxxxx>:
>> 2016-09-18 10:52 GMT+02:00 Peter Rosin <peda@xxxxxxxxxx>:
>>> On 2016-09-16 19:58, Wolfram Sang wrote:
>>>> Same here. And if it prevents us from false positive lockdep reports, I
>>>> am all for fixing it.
>>> Except it doesn't, when I think some more about it...
>>> If you have two gpio-expanders on the same depth but on different i2c
>>> branches you still end up with a splat if one is used to control a mux
>>> to reach the other.
>>> The only way to solve it for good, that I see, is to have every instance
>>> of the gpio-expander mutex in its own class. That might lead to many
>>> lockdep classes but then again, how many gpio expanders could there be
>>> in a system? A dozen or two seems extreme, so maybe that is the correct
>>> approach anyway?
>> Wouldn't it be enough to have a separate class for every base (as in:
>> not having any parent adapters) i2c adapter?
> Of course not - since we could have two branches deeper on the tree
> with the same problem.
> Nevermind my last e-mail.

Right, but you have a point, we can avoid some lockdep classes if
we work at it.

We could recognize that no gpio-expander can be involved with
muxing for a gpio-expander sitting on a root i2c adapter. That
much is trivially obvious.

That means that we could treat all gpio-expanders sitting
directly on a root i2c-adapter as we always have, just like
v1 (and v2, implicitly) do, but instead of putting all other
gpio-exanders in one lockdep class (as in v1) or in one lockdep
class per depth (as in v2), we put all other gpio-mutexes in
a lockdep class of their own.

Then we do not get any extra lockdep classes unless there are
i2c muxes, which is better.

But only i2c-mux-gpio can cause the recursion.

I think the theoretical minimal number of lockdep classes is to
create a new lockdep class for every i2c-mux-gpio instance.
gpio-expanders needing a mutex could walk up the adapter tree
looking for a i2c-mux-gpio and getting the lockdep class from
there. If the root i2c adapter is reached, use some default
lockdep class.

That would consolidate muxes from different gpio-expander
drivers into the one lockdep class. I don't know if that is
a big no-no? I guess that can be handled too with more code,
but it is starting to get messy...

Or, do what the i2c-mux code is doing and use an rt_mutex instead
of an ordinary mutex. That way you are very sure to not get any
lockdep splat ... at all. Ok, sorry, that was not a serious
suggestion, but it would be a tad bit simpler to implement...