[PATCH v6 1/8] iommu: provide early initialisation hook for IOMMU drivers

[Grant Likely]

On Thu, Dec 4, 2014 at 11:19 AM, Arnd Bergmann <arnd at arndb.de> wrote:
> On Thursday 04 December 2014 10:21:27 Will Deacon wrote:
>> On Thu, Dec 04, 2014 at 10:10:17AM +0000, Arnd Bergmann wrote:
>> > On Thursday 04 December 2014 09:49:53 Will Deacon wrote:
>> > > On Wed, Dec 03, 2014 at 07:57:50PM +0000, Arnd Bergmann wrote:
>> > > > On Tuesday 02 December 2014 14:16:57 Grant Likely wrote:
>> > > > > On Mon, Dec 1, 2014 at 11:54 PM, Rob Herring <robherring2 at gmail.com> wrote:
>> > > > > > On Mon, Dec 1, 2014 at 10:57 AM, Will Deacon <will.deacon at arm.com> wrote:
>> > > > > >>
>> > > > > >> +static inline void of_iommu_set_ops(struct device_node *np,
>> > > > > >> +                                   const struct iommu_ops *ops)
>> > > > > >> +{
>> > > > > >> +       np->data = (struct iommu_ops *)ops;
>> > > > > >> +}
>> > > > > >> +
>> > > > > >> +static inline struct iommu_ops *of_iommu_get_ops(struct device_node *np)
>> > > > > >> +{
>> > > > > >> +       return np->data;
>> > > > > >> +}
>> > > > > >
>> > > > > > This may collide with other users. While use of it is rare, PPC uses
>> > > > > > it in its PCI code. The OF_DYNAMIC code frees it but never actually
>> > > > > > sets it. There may be some coming usage with the DT overlay code or
>> > > > > > that's just a bug. Pantelis or Grant can comment. If not, I think we
>> > > > > > really should try to get rid of this pointer rather than expand it's
>> > > > > > usage.
>> > > > > >
>> > > > > > I didn't see a user of this. I'm guessing that is coming in a SMMU patch?
>> > > > >
>> > > > > Good catch. This is not good. The data pointer should be avoided since
>> > > > > there are no controls over its use. Until a better solution can be
>> > > > > implemented, probably the safest thing to do is add a struct iommu_ops
>> > > > > pointer to struct device_node. However, assuming that only a small
>> > > > > portion of nodes will actually have iommu_ops set, I'd rather see a
>> > > > > separate registry that matches device_nodes to iommu_ops.
>> > > >
>> > > > Fair enough. Will, can you take a copy of drivers/dma/of-dma.c and
>> > > > adapt it as needed? It should be exactly what we need to start
>> > > > out and can be extended and generalized later.
>> > >
>> > > Sure, I'll add this to my list of stuff to do for 3.20.
>> >
>> > Does that mean the we don't get any of the patches for 3.19 despite the
>> > Acks?
>>
>> Hmm, I don't know how useful they are without the get/set ops and I don't
>> think I can get those ready for 3.19 given where we currently are.
>>
>> Grant's suggestion of adding an iommu_ops pointer to device_node would work
>> as a temporary hack, but anything more advanced is going to need proper
>> review.
>
> Right. I guess it doesn't hurt much if we put the new pointer inside
> #ifdef CONFIG_OF_IOMMU, then at least there is no significant size
> increase in most DT based platforms.

Yes, I can live with that hack on the proviso that it will be removed by v3.20

Oh, and please put an ugly /* */ comment block in the #ifdef
CONFIG_OF_IOMMU section that makes it really clear that it is an ugly
hack and will be removed in the next release. I don't want anyone
getting ideas that adding pointers to struct device_node is a good
idea.

g.