Re: [dpdk-dev] [PATCH v5 3/3] PCI: don't use vfio ioctl call to access PIO resource

["谢华伟(此时此刻）" <huawei.xhw@alibaba-inc.com>]

On 2021/1/28 0:45, Ferruh Yigit wrote:
> On 1/27/2021 2:43 PM, 谢华伟(此时此刻） wrote:
>>
>> On 2021/1/27 18:32, Ferruh Yigit wrote:
>>> I was waiting for clarification if this can be solved in virtio, 
>>> which seems clarified and decided to go with this patch, I am OK to 
>>> proceed with patch 1 & 2.
>>>
>>> But first patch changes how PIO address get, it changes the Linux 
>>> interface used to get the PIO.
>>> And as far as I can see second patch requires this new interface to 
>>> be able to access the MEM resources.
>>>
>>> I have a concern that this interface change may cause issues with 
>>> various distros, kernel versions etc.. And prefer it goes through a 
>>> full -rc1 validation cycle.
>>>
>>> Huawei, I am aware the patch is around for a while but to play safe, 
>>> I suggest considering it for early next release, so it can be tested 
>>> enough, instead of getting if for -rc2/3 in this release.
>>>
>>> Thanks,
>>> ferruh
>>>
>> Hi Ferruh and Maxime:
>>
>> igb_uio kernel driver gets resource through pci_resource_start, i.e, 
>> (dev)->resource[(bar)].start
>>
>> uio_pci_generic and the generic way in my patch 1 gets resource 
>> through the same interface:
>>
>>          pci dev driver exports to userspace the bar resource 
>> attributes through pci_dev->resource (check resource_show in kernel's 
>> drivers/pci/pci-sysfs.c)
>>
>> Other arch than x86 uses the same interface in their pci_uio_ioport_map.
>>
>> So patch 1 is the most generic way and shouldn't break things. 
>> /proc/ioports should be fully dropped.
>>
>> Using /proc/ioport is partly my fault at the very beginning. It 
>> causes so much mess.
>>
>> Could you please confirm this?
>>
>
> Hi Huawei,
>
> I confirm that interface is already in use, 
> 'pci_parse_sysfs_resource()' does similar parsing.
>
> Most probably it is safe as you and Maxime said, and I am not trying 
> to be difficult but extra conscious here.
>
> Will it cause too much trouble to consider the patch early next 
> release? This gives more time and testing after the patch merged.
>
> Thanks,
> ferruh

Hi Ferruh:

If early next release, what is about the schedule, early next February?


In summary, patch 1 is simple and straightforward. It just don't use 
/proc/ioports and don't use resource attribute created by igb_uio and 
use the standard resource attribute under /sys/pci/.

As i explained, the resource attribute created by igb_uio is exactly the 
same thing as the standard resource attribute under /sys/pci/.

Patch 1 fixes messy things. Patch 2 fixes wrong assumptions (BAR is IO bar).


Customers have been pushing us for quite a long time. Besides, at least 
in China,  virtio in VM with MMIO bar is a de-facto implementation. It 
brings quite much trouble not only to cloud users, but also to us self 
when we run DPDK.


>
>> Thanks huawei
>>
>>>> Could you please post v6?
>>>>
>>>> Thanks,
>>>> Maxime
>>>>
>>>>>>
>>>>>> It is not too late for this release, as the change will not be that
>>>>>> intrusive. But if you prepare such patch, please base it on top 
>>>>>> of my
>>>>>> virtio rework series; To make it easier to you, I added it to the 
>>>>>> dpdk-
>>>>>> next-virtio tree:
>>>>>> https://git.dpdk.org/next/dpdk-next-virtio/log/?h=virtio_pmd_rework_v2 
>>>>>>
>>>>>>
>>>>>> Thanks,
>>>>>> Maxime
>>>>>>
>>>>>