On Mon, Jun 29, 2015 at 08:36:01AM -0700, Arjan van de Ven wrote:
> On 6/29/2015 8:32 AM, Mark Brown wrote:
> >On Mon, Jun 29, 2015 at 07:35:20AM -0700, Arjan van de Ven wrote:

> >It's not that there's no heirachy of locks, it's that lockdep is unable
> >to understand what's going on since it's making simplifying assumptions
> >that just aren't true.  If I remember the problem correctly it's
> >grouping all locks allocated in the same place into one class which
> >doesn't work at all for scenarios where you've got a generic interface
> >providing services to many devices which may be stacked on top of each
> >other.

> but the stacking *IS* a lock hierarchy.

This is why I said "It's not that there is no heirachy of locks".  

> it just seems that the hierarchy is implied rather than explicit.

It's explicit for any given system, like I say it's just that lockdep's
simplifying assumptions don't cope.  As far as I can tell to do
something that robustly works without random magic thrown into
individual drivers with no clear logic we need to allocate a lock class
per regmap (or at least per regmap config that might be instantiated)
which is a problem as they need to be statically allocated.

> >>(I would be interested to know how you avoid ABBA deadlocks btw,
> >>can you have 2 devices, one with a hierarchy one way, and another
> >>with the hierarchy the other way?)

> >I'm not sure I fully understand what you mean here, sorry - do you mean
> >in terms of classes or individual devices?  The relationships between
> >devices are all device and system defined, individual regmaps should be
> >treated as separate classes.  From this perspective it's basically
> >eqivalent to asking how the mutex code avoids misuse of mutexes.

> well what I meant is inividual devices/ranges

> like device A is on devmap A but then ends up using devmap B underneath
> (e.g. the lock nesting case)

I'm not entirely sure what you mean by a "devmap" here - is that just a
regmap or do you mean something else?

> what prevents there from being a device B that is on devmap B but that
> uses devmap A underneath

Assuming you mean regmap nothing prevents that and we should be able to
detect if something messes up there.  It's a problem for the users, not
for regmap itself.