Re: [PATCH net-next 0/3] nfp: support VF multi-queues configuration

[Leon Romanovsky <leon@kernel.org>]

On Thu, Oct 27, 2022 at 02:11:55AM +0000, Yinjun Zhang wrote:
> On Wed, 26 Oct 2022 15:22:21 +0100, Saeed Mahameed wrote:
> > On 25 Oct 11:39, Yinjun Zhang wrote:
> > >On Date: Tue, 25 Oct 2022 12:05:14 +0100, Saeed Mahameed wrote:
> > 
> > Usually you create the VFs unbound, configure them and then bind them.
> > otherwise a query will have to query any possible VF which for some vendors
> > can be thousands ! it's better to work on created but not yet deployed vfs
> 
> Usually creating and binding are not separated, that's why `sriov_drivers_autoprobe`
> is default true.

No, the situation is completely an opposite in a world which heavily uses SR-IOV.
Data centers which rely on SR-IOV to provide VMs to customers separate creation and
bind, as they two different stages in container/VM lifetime. Create is done when
physical server is booted and bind is done when user requested specific properties
of VM.

Various container orchestration frameworks do it for them.

> that's why `sriov_drivers_autoprobe` is default true.

It is not true either. The default value is chosen to keep kernel
backward compatible behavior. 

> unless some particular configuration requires it, like mlnx's msix
> practice. 

And it is not true either. I did MLNX implementation to be aligned with
PCI spec and in-kernel PCI subsystem implementation. Our device can change
MSI-X on-fly and from HW perspective unbind is not important.

Saeed is right "Usually you create the VFs unbound, configure them and
then bind them".

Thanks