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 2021/1/21 16:29, Maxime Coquelin wrote:
>
> On 1/20/21 3:54 PM, 谢华伟(此时此刻） wrote:
>> On 2021/1/13 0:58, Maxime Coquelin wrote:
>>> On 1/12/21 10:37 AM, Maxime Coquelin wrote:
>>>> bus/pci: ...
>>>>
>>>> On 10/22/20 5:51 PM, 谢华伟(此时此刻） wrote:
>>>>> From: "huawei.xhw" <huawei.xhw@alibaba-inc.com>
>>>>>
>>>>> VFIO should use the same way to map/read/write PORT IO as UIO, for
>>>>> virtio PMD.
>>>> Please provide more details in the commit message on why the way VFIO
>>>> works today is wrong (The cover letter is lost once applied).
>> ok
>>>>> Signed-off-by: huawei.xhw <huawei.xhw@alibaba-inc.com>
>>>> Same comment about name format as on previous patches.
>>>>
>>>>> ---
>>>>>    drivers/bus/pci/linux/pci.c     | 8 ++++----
>>>>>    drivers/bus/pci/linux/pci_uio.c | 4 +++-
>>>>>    2 files changed, 7 insertions(+), 5 deletions(-)
>>>>>
>>>>> diff --git a/drivers/bus/pci/linux/pci.c b/drivers/bus/pci/linux/pci.c
>>>>> index 0dc99e9..2ed9f2b 100644
>>>>> --- a/drivers/bus/pci/linux/pci.c
>>>>> +++ b/drivers/bus/pci/linux/pci.c
>>>>> @@ -687,7 +687,7 @@ int rte_pci_write_config(const struct
>>>>> rte_pci_device *device,
>>>>>    #ifdef VFIO_PRESENT
>>>>>        case RTE_PCI_KDRV_VFIO:
>>>>>            if (pci_vfio_is_enabled())
>>>>> -            ret = pci_vfio_ioport_map(dev, bar, p);
>>>>> +            ret = pci_uio_ioport_map(dev, bar, p);
>>>> Doesn't it create a regression with regards to needed capabilities?
>>>> My understanding is that before this patch we don't need to call iopl(),
>>>> whereas once applied it is required, correct?
>>> I did some testing today, and think it is not a regression with para-
>>> virtualized Virtio devices.
>>>
>>> Indeed, I thought it would be a regression with Legacy devices when
>>> IOMMU is enabled and the program is run as non-root (IOMMU enabled
>>> just to suport IOVA as VA mode). But it turns out para-virtualized
>>> Virtio legacy device and vIOMMU enabled is not a supported configuration
>>> by QEMU.
>>>
>>> Note that when noiommu mode is enabled, the app needs cap_sys_rawio, so
>>> same as iopl(). No regression in this case too.
>>>
>>> That said, with real (non para-virtualized) Virtio device using PIO like
>>> yours, doesn't your patch introduce a restriction for your device that
>>> it will require cap_sys_rawio whereas it would not be needed?
>> I don't catch the regression issue.
>>
>> With real virtio device(hardware implemented), if it is using MMIO, no
>> cap_sys_rawio is required.
>>
>> If it is using PIO, iopl is required always.
> My understanding of the Kernel VFIO driver is that cap_sys_rawio is only
> necessary in noiommu mode, i.e. when VFIO is loaded with
> enable_unsafe_noiommu parameter set. The doc for this parameters seems
> to validate my understanding of the code:
> "
> MODULE_PARM_DESC(enable_unsafe_noiommu_mode, "Enable UNSAFE, no-IOMMU
> mode.  This mode provides no device isolation, no DMA translation, no
> host kernel protection, cannot be used for device assignment to virtual
> machines, requires RAWIO permissions, and will taint the kernel.  If you
> do not know what this is for, step away. (default: false)");
> "
>
> I think that using inb/outb in the case of VFIO with IOMMU enabled won't
> work without cap_sys_rawio, and using it in the case of VFIO with IOMMU
> disabled just bypasses VFIO and so is not correct.

Get your concern.

PIO bar:

     HW virtio on HW machine: any vendor implements hardware virtio 
using PIO bar? I think this isn't right. And i dout if vfio doesn't 
check rawio perssion in the syscall in this case.

     Para virtio:  you have no choice to enable unsafe no-iommu mode.  
You must have RAWIO permission.

so with PIO bar, the regression doesn't exist in real world.

Btw, our virtio device is basically MMIO bar, either in hardware machine 
or in pass-throughed virtual machine.


Do you mean we apply or abandon patch 3? I am both OK. The first 
priority to me is to enable MMIO bar support.


> In my opinion, what we should do is to add something like this in the
> DPDK documentation:
>
>   - MMIO BAR: VFIO with IOMMU enabled recommended. Equivalent performance
> as with IGB UIO or VFIO with NOIOMMU. VFIO with IOMMU is recommended for
> security reasons.
>   - PIO BAR: VFIO with IOMMU enabled is recommended for security reasons,
> providing proper isolation and not requiring cap_sys_rawio. However, use
> of IOMMU is not always possible in some cases (e.g. para-virtualized
> Virtio-net legacy device). Also, performance of using VFIO for PIO BARs
> accesses has an impact on performance as it uses pread/pwrite syscalls,
> whereas UIO drivers use inb/outb. If security is not a concern or IOMMU
> is not available, one might consider using UIO driver in this case for
> performance reasons.
>
> What do you think?
>>> Thanks,
>>> Maxime
>>>
>>>> Regards,
>>>> Maxime
>>>>