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

[Robin Murphy]

Hi Will,

On 05/12/14 12:10, Will Deacon wrote:
[...]
>>> Do you expect drivers to modify that *priv pointer after the ops
>>> structure is registered? I'd be very surprised if that was the use
>>> case. It's fine for the driver to register a non-const version, but
>>> once it is registered, the infrastructure can treat it as const from
>>> then on.
>>
>> Possibly not - certainly my current port of the ARM SMMU which makes use
>> of *priv is only ever reading it - although we did also wave around
>> reasons for mutable ops like dynamically changing the pgsize_bitmap and
>> possibly even swizzling individual ops for runtime reconfiguration. On
>> consideration though, I'd agree that things like that are mad enough to
>> stay well within individual drivers if they did ever happen, and
>> certainly shouldn't apply to this bit of the infrastructure at any rate.
>
> I certainly need to update the pgsize_bitmap at runtime because I don't
> know the supported page sizes until I've both (a) probed the hardware
> and (b) allocated page tables for a domain. We've already discussed
> moving the pgsize_bitmap out of the ops, but moving it somewhere where
> it remains const doesn't really help.

We can safely cast the call to get_ops in the SMMU driver though, since 
we'll know that we put a mutable per-instance ops in there in the first 
place. At least that way drivers that aren't taking advantage and just 
pass their static const ops around shouldn't provoke warnings. I 
deliberately didn't touch anything beyond get_ops as that would be too 
disruptive.

> Can I just take the patch that Grant acked, in the interest of getting
> something merged? As you say, there's plenty of planned changes in this
> area anyway. I plan to send Olof a pull request this afternoon.

Grant, Thierry? Personally I'm not fussed either way - the sooner 
something goes in, the sooner I can carry on working at replacing it :D

Robin.