Re: [PATCH for 4.6 07/13] xen: Introduce a generic way to describe device

[Julien Grall <julien.grall@linaro.org>]

Hi Jan,

On 17/12/2014 17:17, Jan Beulich wrote:
>>>> Julien Grall <julien.grall@linaro.org> 12/17/14 2:04 PM >>>
>> On 17/12/14 10:46, Jan Beulich wrote:
>>>>>> On 17.12.14 at 11:30, <julien.grall@linaro.org> wrote:
>>>> Having a generic way to describe device will really help ARM code (see
>>>> IOMMU).
>>>>
>>>> If we don't have a such thing, we may need to duplicate quite a lots of
>>>> code. Which will make hard to maintain.
>>>
>>> Not really, if e.g. "device" was simply an alias of "pci_dev" on x86.
>>
>> How many pci_dev instance you could have on a platform? 1000? Though it
>> might be a high value but that mean we use 2k more of RAM.
>
> Sure the total amount isn't big. But these days everyone thinks that way, and
> data size gets grown without much consideration. And you shouldn't just think
> about RAM cache utilization.

I will go ahead with the aliasing.

>> It doesn't seem to bad for the benefit to have a clear code.
>
> Aliasing device and pci_dev for x86 would yield similar clarity afaict.

To be sure, by aliasing you mean creating a typedef?

For x86:
typedef struct pci_dev device_t;

And for ARM:
typedef struct device device_t;

>
>>>>>> +#define pci_to_dev(pcidev)  (&(pcidev)->dev)
>>>>>> +
>>>>>> +static inline struct pci_dev *dev_to_pci(struct device *dev)
>>>>>> +{
>>>>>> +    ASSERT(dev->type == DEV_PCI);
>>>>>> +
>>>>>> +    return container_of(dev, struct pci_dev, dev);
>>>>>> +}
>>>>>
>>>>> While the former is const-correct, I dislike the inability of passing
>>>>> pointers to const into helper functions like the latter. I can't think
>>>>> of a good solution other than introducing a second const variant
>>>>> of it, but I suppose we should try to find alternatives before
>>>>> adding such a construct that moves us in a direction opposite to
>>>>> getting our code more const-correct.
>>>>
>>>> Oh right. I didn't though about that case. I will turn this inline
>>>> function into a macro.
>>>
>>> I'm afraid that won't help, as you still need to specify a type as
>>> 2nd argument to container_of(), and that type can't be both
>>> const and non-const at the same time, i.e. you can't easily
>>> inherit the const-ness of the passed in pointer.
>>
>> I agree that we will drop the const-ness. But is it really an issue?
>>
>> We won't have many place where we don't want to modify the pci_dev.
>
> Did you check (including places where const could be added)? But at least
> you didn't have to drop and const-s, so I'm not heavily objecting the change
> you propose.

The only usage will be in the IOMMU code where most of the time we 
require a non const version.

At least on the SMMU driver, we have to store data per-device. This is 
to know what is the SMMU master for this device.

-- 
Julien Grall