[dpdk-dev] [PATCH v2 08/16] Add support for mapping devices through VFIO.

[Burakov, Anatoly]

Hi Thomas,

> How did you test this feature?
> Did you see some performance differences with igb_uio?

The same way everything else is tested - bind a NIC to the driver and see if it works :-)

As for performance differences, potentially it can be degraded a bit because of mandatory IOMMU involvement, but I did not see any performance impact during my tests.

> For history reason, it's better to explain in another patch that eal_hpet has
> been renamed eal_timer and there is no such need anymore in this file.

Agreed.

> 
> We should discuss a way to request igb_uio or VFIO binding of a device.

Why? The device can either be bound to VFIO or igb_uio. So unless you want binding code in DPDK EAL (to avoid which the pci_unbind/igb_uio_bind/dpdk_bind script was created in the first place), I see no point in that. The dpdk_bind script already does that (you bind either to igb_uio or to vfio-pci).

> This whole socket communication deserves a separated patch with protocol
> description.

Agreed, I'll break it up and provide a more detailed explanation.

> By the way, I'm not a big fan of the suffix "_socket" which can be misleading.
> But I have no other good naming idea.

Would _mp_socket do?
 
> So we have another thread to manage.
> I don't see where it is spawned?

In rte_eal_pci_init().

> You are defining some variables in a .h file. I think it is a problem.

Well, they need to be shared between several .c files.
 
> Here are some other relevant errors from checkpatch.pl:

Thanks, I'll fix those.

Best regards,
Anatoly Burakov
DPDK SW Engineer