Re: [Ksummit-discuss] [TECH TOPIC] Addressing complex dependencies and semantics (v2)

[Jan Kara <jack@suse.cz>]

Hi Laurent!

On Thu 28-07-16 19:00:36, Laurent Pinchart wrote:
> On Thursday 28 Jul 2016 13:46:24 Jan Kara wrote:
> > On Thu 28-07-16 14:03:46, Laurent Pinchart wrote:
> > > On Thursday 28 Jul 2016 12:54:47 Hans Verkuil wrote:
> > >>> One problem with those other dependencies is that they can't always be
> > >>> expressed as a tree and may a graph instead. Worse, in some cases, the
> > >>> graph can be cyclic (I've recently been told about an external
> > >>> I2C-based PLL that takes an input clock and generates an output clock,
> > >>> with the input clock being produced by an on-SoC sound device and the
> > >>> output clock being used by the same sound device). Even when
> > >>> individual resource trees or graphs are not cyclic, combining them in
> > >>> a global dependency graph will often result in cycles. The challenge
> > >>> is to find a proper way to both express the dependency graph and break
> > >>> the cycles.
> > >> 
> > >> Do we need to capture 100% of all the weird and wonderful dependencies?
> > >> I think (speaking for the media subsystem) that the vast majority of the
> > >> dependencies are pretty simple trees without cycles. Being able to
> > >> capture that would be a huge help. The remaining more complex devices
> > >> could still fall back on deferred probe, I'd say.
> > > 
> > > When dealing with a single type of resources dependency graphs are very
> > > rarely cyclic. However, when merging multiple resource dependency graphs,
> > > cycles become quite common.
> > > 
> > > One example is the camera device found in OMAP3 chips. The SoC-side camera
> > > controller can output a clock to the outside world (and is thus a CCF
> > > clock provider). In most hardware designs that clock is used by the
> > > external camera sensors, which in turn outputs a video data stream (with a
> > > clock and other synchronization signals) fed to the camera controller. We
> > > thus get a dependency loop. The situation is even worse on some Atmel
> > > platforms where part of the camera controller logic is clocked by the
> > > pixel clock received from the sensor. In that case the camera controller
> > > can't be software reset without a sensor providing a pixel clock. We can
> > > blame the hardware designers, but at the end of the day we need to
> > > support such systems. Whether such cases need to be supported by generic
> > > code is a discussion we can have, and I'm not pushing in either
> > > direction, but I want to make sure we will always have *a* solution.
> > 
> > So I'm complete outsider to media drivers so maybe my question is stupid
> > but I cannot resist: So how is the device bringup supposed to work for
> > devices where you have these cycles in the dependency graph? Do you just
> > bring up the controller and then the sensor and it works because the
> > controller doesn't need the clock from the sensor for startup?
> 
> That's pretty much the idea, yes. We first probe the controller, and register 
> a notifier to react to sensors being probed. The controller registers at probe 
> time the clocks it provides, so the sensor can get hold of them when it gets 
> probed. The controller is then notified that the sensor is available and 
> proceeds normally.
> 
> We also have a suspend/resume ordering issue, where we have to stop the 
> controller first (the hardware doesn't like when it loses the incoming video 
> stream while running), but we can't stop the clock just yet before the sensor 
> driver has a chance to suspend the sensor. We thus use the PM prepare and 
> complete callbacks to stop/restart the controller, and the PM suspend and 
> resume callbacks to stop/restart the clock. This is a hack that assumes that 
> the sensor gets probed after the controller (which is guaranteed by the fact 
> that the sensor needs a clock provided by the controller and thus probe will 
> be deferred if the controller hasn't been probed yet) and that the sensor 
> driver will use the PM suspend/resume callbacks, not the prepare/complete 
> callbacks. Ugly, but for now it works.

Thanks for explanation. So from the explanation it seems to me that the
cycles in the directed graph of devices arise for two reasons:

1) Edges in the graph express different types of dependencies (i.e., boot
dependency graph is not necessarily the same as suspend dependency graph)
and thus cycles are created.

2) Some of the vertices in the graph currently express several separately
controllable entities which may possibly need to be brought up / suspended /
whatever at different times and this merging of vertices creates cycles.

So if we actually had several different graphs - one for each operation
(boot, suspend, I-don't-know-what-else-one-can-do-with-devices) -
expressing dependencies for that operation where vertices would correspond
to the smallest separately controllable entities for that operation, then
each of these graphs would be acyclic.

Now whether this would be a practical solution I don't know (one problem
may be the memory overhead, another may be number of different operations
one needs to track, and yet another whether one can easily determine
dependencies for each operation in advance). But it would be a theorically
clean one :)

								Honza 
-- 
Jan Kara <jack@suse.com>
SUSE Labs, CR