[dpdk-dev] virtio PMD is not working with master version

[Xie, Huawei]

On 2/26/2016 4:29 PM, David Marchand wrote:
> On Fri, Feb 26, 2016 at 3:23 AM, Yuanhan Liu
> <yuanhan.liu at linux.intel.com> wrote:
>> Mauricio, thanks for the testing and report.
>>
>> On Thu, Feb 25, 2016 at 02:30:18PM +0100, David Marchand wrote:
>>> >From the logs, I would say I broke uio_pci_generic since we are in
>>> "uio" case, but uio portio sysfs does not exist.
>>> virtio pmd fell back to ioports discovery before my change.
>> Maybe we can do same?
We shouldn't, :). I am now rebasing the patch to fix the issue that
virtio driver takes the virtio device blindly.
With the patch:
 if driver is VFIO/UIO, and errors happens, returns without falling back
to IO port.
 if no any kernel driver is managing the device, try IO port; otherwise
returns 1 to tell the layer we don't take over this device.

> I suppose, but see below.
>
>> ---
>> diff --git a/lib/librte_eal/linuxapp/eal/eal_pci.c b/lib/librte_eal/linuxapp/eal/eal_pci.c
>> index 4346973..579731c 100644
>> --- a/lib/librte_eal/linuxapp/eal/eal_pci.c
>> +++ b/lib/librte_eal/linuxapp/eal/eal_pci.c
>> @@ -685,12 +685,11 @@ int
>>  rte_eal_pci_ioport_map(struct rte_pci_device *dev, int bar,
>>                        struct rte_pci_ioport *p)
>>  {
>> -       int ret;
>> +       int ret = -1;
>>
>>         switch (dev->kdrv) {
>>  #ifdef VFIO_PRESENT
>>         case RTE_KDRV_VFIO:
>> -               ret = -1;
>>                 if (pci_vfio_is_enabled())
>>                         ret = pci_vfio_ioport_map(dev, bar, p);
>>                 break;
>> @@ -700,14 +699,14 @@ rte_eal_pci_ioport_map(struct rte_pci_device *dev, int bar,
>>                 ret = pci_uio_ioport_map(dev, bar, p);
>>                 break;
>>         default:
>> +               break;
>> +       }
>> +
>>  #if defined(RTE_ARCH_X86_64) || defined(RTE_ARCH_I686)
>> -               /* special case for x86 ... */
>> +       /* special case for x86 ... */
>> +       if (ret)
>>                 ret = pci_ioport_map(dev, bar, p);
>> -#else
>> -               ret = -1;
>>  #endif
>> -               break;
>> -       }
> What if we are supposed to do vfio here, but for some reason init failed ?
> Next thing, we will call ioport_read in vfio context, but init went
> through the ioports parsing => boom ?
>
> Another issue is that when device is bound to a kernel driver (let's
> say virtio-pci here), then init will succeed and pmd will kick in the
> device registers.
>
> This special case should really be narrowed down to "uio" and "none"
> driver cases.
>
>