[GIT PULL] Allwinner drivers changes for 4.2

[Arnd Bergmann]

On Thursday 28 May 2015 21:08:13 Maxime Ripard wrote:
> Hi Arnd,
> 
> On Thu, May 28, 2015 at 07:16:52PM +0200, Arnd Bergmann wrote:
> > On Thursday 21 May 2015 14:20:19 Maxime Ripard wrote:
> > > > Preface: I only did the reserved sections so I could claim parts of my 
> > > > Rockchip sram for the smp code that needed to reside at a specific place in it, 
> > > > so I guess I don't necessarily feel qualified to judge one against the other 
> > > > :-), but anyway
> > > > 
> > > > 
> > > > The commit message for the driver contains
> > > > 
> > > > "We could also imagine changing this at runtime for example to change the
> > > > mapping of these SRAMs to use them for suspend/resume or runtime memory rate
> > > > change, if that ever happens."
> > > > 
> > > > which sounds to me a lot like the generic use case for the current sram driver 
> > > > - for example in conjuction with the PIE stuff if it ever resurfaces.
> > > > 
> > > > 
> > > > But from my short glance I also don't see how this quite custom mapping thing 
> > > > (device vs. cpu) would be able to fit into the generic description.
> > > 
> > > So, what's the conclusion on this?
> > > 
> > > This driver has been properly sent, without any kind of review from
> > > you. I then sent a pull request with it for 4.1, which has only been
> > > silently ignored.
> > > 
> > > And now, it seems like this is going to be the same for 4.2. I'd
> > > *really* like to have some kind of a discussion here, and not let it
> > > fall into oblivion. It fixes some real issues we have.
> > 
> > I've looked at the driver some more now, and tried to come up with
> > a binding that I think would fit better with the existing mmio-sram
> > one. Do you think you could get it to work like this?
> > 
> > sram at 00000000 {
> > 	// we bind to the first one here
> > 	compatible = "allwinner,sun4i-a10-sram-controller", "mmio-sram";
> > 	// first the SRAM, second the controller regs
> > 	reg = <0x10000000 0x11000>, <0x01c00000 0x30>;
> 
> Even if we consider the various contiguous SRAMs as a single one with
> sections, that would cause some issues, because we have some other
> SRAMs too, that are mapped at different addresses than those and still
> controlled by the SRAM controller (C and D).
> 
> Depending on the SoC, and the current state of the driver, that would
> mean a various number of reg cells...

Yes, good point.

> > 	ranges = <0 0x10000000 0x11000>;
> 
> With ranges being wrong for most of them.
> 
> > 	#address-cells = <1>;
> > 	#size-cells = <1>;
> > 
> > 	otg-sram: otg-sram at 10000 {
> > 		compatible = "allwinner,sun4i-a10-sram";
> > 		reg = <0x10000 0x1000>;
> > 	};
> 
> And that would reserve a section of the SRAM, even if the device is
> not used at all, which would mean that we would lose the benefit of
> the mmio-sram genpool stuff.

What it essentially means it that we would not list the section
in the DT if it's not used. Removing it from DT would be a bit
impractical, but if we use 'status="disabled"' (and ensure
that actually works), it should be ok.

> > };
> > 
> > &usb-otg {
> > 	// allow otg driver to find the sram by phandle and
> > 	// not need a table in the driver
> > 	sram = <&sram-D>;
> 
> That would be the best solution to reference the link between a device
> and its SRAM though.
> 
> What about some kind of a "hybrid" solution: have all the SRAM
> declared as separate node to avoid the ranges and reg issues described
> above, and use an allwinner,sram or whatever to reference the link,
> without requiring the string used by the client you were finding odd?

yes, sounds reasonable.
 
> And then, the SRAM controller driver would simply parse this property
> using the struct device when the client driver claim the SRAM?

Ok. I'm a bit torn between using "allwinner,sram" and making it
a standard "sram" property that could be used for a generic subsystem
if other platforms need the same thing in the future.

> > One important advantage here would be that we later have the option
> > of turning it into a subsystem with multiple sram controller drivers
> > and have sram clients just ask for a node by referencing a phandle.
> 
> I'd be all in favor of a subsystem, but most likely, when such a
> subsystem will be introduced, we will not have considered all possible
> cases, and would end up with different bindings anyway...

We can to some degree accomodate earlier bindings that are "mostly"
compatible by special-casing them. E.g. a generic subsystem could
look for both "srams" and "allwinner,sram" if the first is not found.

A related question is whether we want to pass arguments here, and how
to link the controller to the sram nodes.

	Arnd