[PATCH v3 04/15] iommu: Move bus setup to IOMMU device registration

[Baolu Lu]

On 2022/7/6 22:37, Robin Murphy wrote:
> On 2022-07-06 03:35, Baolu Lu wrote:
>> On 2022/7/6 01:08, Robin Murphy wrote:
>>> @@ -202,12 +210,32 @@ int iommu_device_register(struct iommu_device 
>>> *iommu,
>>>       spin_lock(&iommu_device_lock);
>>>       list_add_tail(&iommu->list, &iommu_device_list);
>>>       spin_unlock(&iommu_device_lock);
>>> +
>>> +    for (int i = 0; i < ARRAY_SIZE(iommu_buses); i++) {
>>> +        struct bus_type *bus = iommu_buses[i];
>>> +        int err;
>>> +
>>> +        if (bus->iommu_ops && bus->iommu_ops != ops) {
>>> +            err = -EBUSY;
>>> +        } else {
>>> +            bus->iommu_ops = ops;
>>> +            err = bus_iommu_probe(bus);
>>> +        }
>>> +        if (err) {
>>> +            iommu_device_unregister(iommu);
>>> +            return err;
>>> +        }
>>> +    }
>>> +
>>>       return 0;
>>>   }
>>>   EXPORT_SYMBOL_GPL(iommu_device_register);
>>
>> With bus_set_iommu() retired, my understanding is that now we embrace
>> the first-come-first-serve policy for bus->iommu_ops setting. This will
>> lead to problem in different iommu_ops for different bus case. Did I
>> overlook anything?
> 
> This is just formalising the de-facto situation that we don't actually 
> have any combination of drivers that could load on the same system 
> without already attempting to claim at least one bus in common. It's 
> also only temporary until the bus ops are removed completely and we 
> fully support multiple drivers coexisting, which only actually takes a 
> handful more patches - I've realised I could even bring that change 
> *ahead* of the big job of converting iommu_domain_alloc() (I'm not 
> convinced that the tree-wide flag-day patch for that I currently have in 
> the dev branch is really viable, nor that I've actually got the correct 
> device at some of the callsites), although whether it's worth the 
> potentially-surprising behaviour that might result I'm less sure.
> 
> If we already had systems where in-tree drivers successfully coexisted 
> on different buses then I'd have split this up and done something a bit 
> more involved to keep a vestigial bus_set_iommu() around until the final 
> bus ops removal, but since we don't, it seemed neatest to do all the 
> related work in one go.

Fair enough. I've never seen a mixed system as far. It's fine for us to
retire bus_set_iommu() for now and then formally support mixed IOMMU
drivers later.

Best regards,
baolu