Re: [dpdk-dev] [PATCH v4] pci: support both PIO and MMIO BAR for legacy virtio on x86

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

On 2020/10/22 1:24, Ferruh Yigit wrote:
> On 10/21/2020 1:32 PM, 谢华伟(此时此刻） wrote:
>>
>> On 2020/10/21 19:49, Ferruh Yigit wrote:
>>> On 10/13/2020 9:41 AM, 谢华伟(此时此刻） wrote:
>>>> From: "huawei.xhw" <huawei.xhw@alibaba-inc.com>
>>>>
>>>> Legacy virtio-pci only supports PIO BAR resource. As we need to 
>>>> create lots of
>>>> virtio devices and PIO resource on x86 is very limited, we expose 
>>>> MMIO BAR.
>>>>
>>>> Kernel supports both PIO  and MMIO BAR for legacy virtio-pci 
>>>> device. We handles
>>>> different type of BAR in the similar way.
>>>>
>>>> In previous implementation, with igb_uio we get PIO address from 
>>>> igb_uio
>>>> sysfs entry; with uio_pci_generic, we get PIO address from
>>>> /proc/ioports.
>>>> For PIO/MMIO RW, there is different path for different drivers and 
>>>> arch.
>>>> For VFIO, PIO/MMIO RW is through syscall, which has big performance
>>>> issue.
>>>> On X86, it assumes only PIO is supported.
>>>>
>>>> All of the above is too much twisted.
>>>> This patch unifies the way to get both PIO and MMIO address for 
>>>> different driver
>>>> and arch, all from standard resource attr under pci sysfs.
>>>>
>>>
>>> As mentined above this patch does multiple things.
>>>
>>> The main target is, as far as I understand, you have a legacy virtio 
>>> device which supports "memory-mapped I/O" and "port-mapped I/O", but 
>>> virtio logic forces legacy devices to use the PIO but you want to be 
>>> able to use the MMIO with this device.
>> yes.
>>>
>>> The solution below is adding MMIO support in the PIO funciton, and 
>>> distinguish MMIO or PIO based on their address check.
>> Yes, kernel does this in the similar way.
>>>
>>>
>>> Instead of this, can't this be resolved in the virtio side, like if 
>>> the legacy device supports MMIO (detect this somehow) use the MMIO 
>>> istead of hacking PIO mapping to support MMIO?
>>
>> Get your concern.
>>
>> 1>
>>
>> If we move, I think we should move all those PCI codes into virtio 
>> side, not just the mmio part.
>>
>> Without my patch, those PCI codes are virtio-pci device specific, not 
>> generic.
>>
>> With this patch, those pci ioport map/rw code could also be used for 
>> other devices if they support both PIO and MMIO.
>>
>
> I was not suggesting moving any code into virtio, but within 
> 'vtpci_init()' what happens when "hw->modern = 1;" is set?
> And if this is set for your device, will it work without change?

Yes, this will only affect legacy_device, which uses legacy_ops to 
access port io.

If is is modern_device, port access will go through modern_ops.

We only change the implementation in legacy_ops.


>
>> Every option is ok. Hope i make myself clear.
>>
>> 2>  I don't think this is hacking. for rte_pci_ioport_map/read/write, 
>> if ioport could be both PIO and MMIO, then everything is reasonable.
>>
>> Take how kernel does port map for example:
>>
>>      vp_dev->ioaddr = pci_iomap(pci_dev, 0, 0);
>>
>> Here io doesn't mean PIO only. It could also be MMIO. Kernel then 
>> uses ioread/write to access PIO/MMIO port.
>>
>> Actually we are pretty much the same in the interface.
>>
>> I think this patch extends rather then hacks the ioport interface to 
>> support MMIO.
>>
>>>
>>> I have other concerns, specially mergin VFIO mapping too, but lets 
>>> clarify above first.
>>
>> vfio doesn't affect other driver but only virtio.
>>
>
> Why it doesn't affect other drivers, can't there be other driver using 
> PIO?

Currently only virtio-pci uses PIO, and only virtio PMD uses these port 
map/read/write functions.

I don't foresee in future any new device uses PIO.

/huawei

>
>> igb_uio, uio_pci_generic and vfio-pci all uses the same way to map/rw 
>> ioport.
>>
>
> For vfio, code changes 'pci_vfio_ioport_read()' to the direct address 
> read, first I don't know if this is always safe, and my question why 
> there is a syscall introduced at first place if you can read from 
> address directly?

Original vfio way works, but we don't need that syscall. Under whatever 
driver, we could use the simple way as in this patch.

/huawei

>
> Is your device works as expected when vfio-pci kernel module used? 
> Since it is not suffering from PIO limitation, right?

Certainly i tested vfio module. Firstly, i didn't intend to fix vfio 
performance issue, but i heard that igb_uio will be removed.

/huawei

>
>
> And I wonder if the patch can be done as three patches to simply it, as:
> 1) Combine 'RTE_PCI_KDRV_IGB_UIO' & 'RTE_PCI_KDRV_UIO_GENERIC' (remove 
> pci_ioport_map)
> 2) Update 'pci_uio_ioport_map()' to add memory map support (and update 
> read/write functions according)
> 3) Combine vfio & uio
>
Got it. It makes sense to split, but i think this patch is already 
simple enough.

Let me check.

/huawei

>>>
>>> Thanks,
>>> ferruh
>>>
>>>
>>>
>>>> We distinguish PIO and MMIO by their address like how kernel does. 
>>>> It is ugly but works.
>>>>
>>>> Signed-off-by: huawei.xhw <huawei.xhw@alibaba-inc.com>
>
> <...>