Re: [RFC PATCH 1/7] iommu: provide early initialisation hook for IOMMU drivers

[Arnd Bergmann <arnd@arndb.de>]

On Tuesday 02 September 2014 14:47:54 Varun Sethi wrote:
> > +void __init of_iommu_init(void)
> > +{
> > +     struct device_node *np;
> > +     const struct of_device_id *match, *matches = &__iommu_of_table;
> > +
> > +     for_each_matching_node_and_match(np, matches, &match) {
> > +             const int (*init_fn)(struct device_node *) = match->data;
> 
> Is the init function also responsible for setting iommu_ops (per bus)?
> We need to take in to consideration that iommu API (iommu.c)
> initialization happens during  "arch_init".

I would hope that as part as one of the next steps, we can skip
the step of setting the iommu_ops per bus_type. It's really
not that meaningful when only some of the devices on one bus
are able to use an iommu, and some may need other ops than others.

> When we setup bus iommu
> ops add_device would be called for devices which have already been
> probed.

As soon as you move to the generic way of probing the IOMMU for
DT, you should not need an add_device callback any more.

Each iommu driver should do one one or the other, but not both.

> Also, as I can see from the code we have of_platform_populate
> which gets called right after of_iommu_init, which would in turn also
> lead to add_device invocation (after add_device_master_ids). 
> I believe this would happen before iommu API initialization which
> would cause issues. 

of_iommu_init should be the point where the iommu API gets
initialized.

	Arnd