[PATCH RFC v1 1/2] documentation/iommu: Add description of Hisilicon System MMU binding

[arnd@arndb.de (Arnd Bergmann)]

On Monday 16 June 2014 17:45:17 Will Deacon wrote:
> On Mon, Jun 16, 2014 at 05:42:10PM +0100, Arnd Bergmann wrote:
> > On Monday 16 June 2014 17:26:53 Will Deacon wrote:
> > > On Fri, Jun 06, 2014 at 12:07:11PM +0100, Dave Martin wrote:
> > > > On Fri, Jun 06, 2014 at 08:48:26AM +0200, Arnd Bergmann wrote:
> > > > > On Thursday 05 June 2014 21:37:09 Zhen Lei wrote:
> > > > > 
> > > > > > +- smmu-masters  : A list of phandles to device nodes representing bus
> > > > > > +                  masters for which the SMMU can provide a translation
> > > > > > +                  and their corresponding StreamIDs (see example below).
> > > > > > 
> > > > > 
> > > > > We're currently in the process of defining a generic binding for IOMMUs.
> > > > > 
> > > > > While the smmu-masters property was copied from an existing binding,
> > > > > I think we will end up migrating away from that towards a common way
> > > > > to express those things, and we shouldn't add another one doing this
> > > > > in a nonstandard way. Please have a look at the latest discussion
> > > > > about the iommu binding using #iommu-cells and a reference from the
> > > > > master to the iommu and see if you can migrate your code to use that.
> > > > 
> > > > Thanks for making this point -- I was going to do so yesterday but then
> > > > hesitated due to uncertainty about whether this should really be a new
> > > > driver.
> > > > 
> > > > Either way, it would be very valuable at least to attempt to describe
> > > > the Hisilicon SMMU implemenation using the new proposals, since that is
> > > > a good test of how reusable the proposed generic binding actually is.
> > > 
> > > If this ends up being an addition to the existing ARM SMMU driver, I'm
> > > really not keen on using the new DT bindings. We're already stuck with
> > > the old bindings for that driver -- supporting both old and new in the
> > > same code only buys us maintenance headaches and pointless divergence
> > > within the driver.
> > 
> > We have to migrate the driver to the new binding anyway, it may be
> > a bit painful, but there are not really any users yet so there
> > is a chance we can remove the nonstandard code at some point,
> > perhaps in a few years.
> 
> The only way I see this working is if we kill the existing binding and move
> exclusively to the new one. I'm actually ok with that (we have no in-tree
> users), but it needs to happen ASAP in my opinion, otherwise we increase the
> window where the old binding can be adopted.

I agree. I was hoping to get the generic binding ready for 3.16, but that
didn't happen. Maybe we can add a small patch to the binding to explain
that it will change in the future.

> Note that the next version of the ARM SMMU (v3) is considerably different to
> the current architecture, so a new driver (using the new bindings) will be
> required.
> 
> This actually opens a wider question: if we don't have an in-tree user for a
> device-tree binding, do we consider that binding to be unused?

Not in general, but often. We don't require dts files to be in the kernel,
so we have to apply a bit of common sense. If anyone knows of out-of-tree
users of the binding that are actually working with upstream kernels,
we need a migration path. Anything that also requires out-of-tree kernel
patches however is something we don't have to worry about.

	Arnd