答复: [PATCH 2/2] PCI: Layerscape: Add Layerscape PCIe driver

[B31939@freescale.com (Lian Minghuan-B31939)]

Hi Arnd,

On 2014?09?09? 11:58, Arnd Bergmann wrote:
> On Tuesday 09 September 2014 19:16:01 Lian Minghuan-B31939 wrote:
>> On 2014?09?09? 10:50, Arnd Bergmann wrote:
>>> On Tuesday 09 September 2014 18:46:59 Lian Minghuan-B31939 wrote:
>>>> On 2014?09?09? 09:56, Arnd Bergmann wrote:
>>>>> On Tuesday 09 September 2014 17:25:57 Lian Minghuan-B31939 wrote:
>>>>>> [Minghuan] I discussed with my colleague. They worry about performance
>>>>>> degradation if using regmap API,
>>>>>> because there are some fast device use scfg. We tend to use a simple way
>>>>>> to map andread/write scfg directly.
>>>>> I see. In this case, I would probably create a separate msi controller
>>>>> driver that owns the "fsl,ls1021a-scfg" device, and is referenced
>>>>> through the "msi-parent" property in the pcie controller.
>>>>>
>>>>> You can use of_pci_find_msi_chip_by_node() to get the msi_chip
>>>>> instance and then connect that to your pci host. This will also
>>>>> take care of the case where you may want to use the main GICv3
>>>>> on a future SoC.
>>>> [Minghuan] There is something wrong with LS1021A MSI hardware that it
>>>> only supports one interrupt not 32 interrupts.  Now, I do not want to
>>>> create a separate msi controller driver just for incorrect hardware.
>>>> I may provide complete MSI driver for the new hardware when it is ready.
>>> Would you just leave out MSI support for the LS1021A PCIe variant?
>>> I guess that's fine because all device drivers should also support
>>> legacy interrupts and there is no performance gain in MSI in this
>>> case.
>> [Minghuan] I have added MSI support for LS1021A PCIe just reserved 31
>> interrupts as used.
> I don't understand your logic then. If LS1021A has an incorrect MSI
> implementation, and you may want to reuse the PCIe driver with a
> future chip that either includes a correct MSI implementation, or
> with one that uses the GICv3 instead, isn't that even more reason
> to split out the MSI support into a separate driver?
>
> That way you can at least separate the normal code path from the
> broken one and don't need any special run-time or compile-conditionals
> beyond calling of_pci_find_msi_chip_by_node().
[Minghuan] The new MSI hardware implementation is not ready, I do not 
know how to describe its dts node
and how to implement its driver. When it is ready, I will implement the 
driver and add a workaround
for the current incorrect MSI. The current MSI driver is based on 
pci-designware.
> 	Arnd