On Wednesday 08 January 2014, Thierry Reding wrote:

On Wed, Jan 08, 2014 at 04:11:08PM +0100, Arnd Bergmann wrote:

quoted

On Wednesday 08 January 2014 15:55:27 Thierry Reding wrote:

quoted

It stands to reason that if they push back on the IOMMU variant of what
is essentially the same thing, they will push back on the IRQ variant as
well. One alternative I proposed was to, just as you suggested earlier,
move the code into platform_drv_probe() or rather a function called from
it. That proposal never got any replies, though.

	https://lkml.org/lkml/2013/12/14/39

I guess putting it into the platform_drv_probe function seems reasonable,
I would be more scared of the implications of a notifier based method.

I fully agree. Of course if we decide against moving things into the
core and in favour of a more generic API that drivers should use, then
this issue goes away silently at least for resources that the driver
needs to use explicitly (memory-mapped regions, interrupts, ...).

The issue remains for IOMMU which is meant to be used transparently
through the DMA API. Perhaps a good compromise would be to have some
sort of generic helper that can be called to initialize IOMMU support
for a particular device and support probe deferral on error. Something
like this perhaps:

	int iommu_attach(struct device *dev);
	int iommu_detach(struct device *dev);

I still don't like very much how that needs to be done in each driver
explicitly, but if we can't do it in the core, then the only other clean
way to handle it would be to treat it like any other sort of resource
and handle it explicitly. Perhaps handing out some sort of cookie would
be preferable to just an error code?

The more I think about the iommu case, the more I am convinced that it
does belong into the core, in whatever form we can find. As far as I
can tell from the little reliable information I have on the topic, I
would assume that we can keep it in the DT probing code, as there won't
be a need for multiple arbitrary IOMMUs with ACPI or with board files.

quoted

quoted

One downside of that approach is that, while it maps well to platform
devices or generic devices that have some sort of firmware interface
such as OF or ACPI, I don't see how it can be made to work with an I2C
client that's registered from board setup code for example. Well, I
suppose that problem could be solved by throwing another lookup table at
it, just like we do for clocks, regulators, PWMs and GPIOs.

Wouldn't you still be able to attach resources in the traditional
way for those, but use the same new interface to get at them?

I wouldn't know how. For instance platform devices store the IRQ number
within a struct resource of type IORESOURCE_IRQ, whereas I2C clients
store them in the struct i2c_client's .irq field.

Good point, I forgot about the special case for i2c_client->irq.
I looked now and noticed that very few i2c devices actually use this
field, but larger number uses platform_data, which has a similar
problem.

So without actually introspecting the struct device (possibly using the
.bus field for example) and upcasting you won't know how to get at the
resources. One possibility to remedy that would be to try and unify the
resources within struct device. But that doesn't feel right.

One other thing I had considered at one point was to extend the bus_type
structure and give it a way to obtain resources in a bus-specific way,
but that feel even more wrong.

Perhaps I'm missing something obvious, though, and this is actually much
more trivial to solve.

No trivial solution that I can see. I think we can deal with the case
where platform code uses platform_device->resources, and everything else
comes down to having multiple code branches in the driver, as we already
have to deal with platform_data and DT properties describing stuff that
doesn't fit in the resources.

	Arnd