[PATCH v3 0/7] Rework device probing

[Bruce Richardson]

On Mon, Apr 13, 2026 at 03:34:05PM +0200, Morten Brørup wrote:
> > From: David Marchand [mailto:david.marchand at redhat.com]
> > Sent: Saturday, 11 April 2026 12.11
> > 
> > On Fri, 10 Apr 2026 at 16:35, Morten Brørup <mb at smartsharesystems.com>
> > wrote:
> > >
> > > Is there a workaround until this patch is added, if I don't want
> > rte_eal_init() to probe any devices, but I want my application to probe
> > devices later by calling rte_bus_scan() followed by rte_bus_probe()?
> > >
> > > It's not something we really need, so if you don't know it off the
> > top of your head, don't spend time trying to figure it out.
> > 
> > Well, starting from the well known -a pci:0000:00:00.0 trick, I wrote
> > an incomplete workaround implemented for OVS.
> > https://patchwork.ozlabs.org/project/openvswitch/patch/20260226172429.2
> > 12034-3-david.marchand at redhat.com/
> > 
> > Incomplete, because some buses make life harder with some additional
> > parsing restraints (bus/dpaa is the first one that comes to mind) or
> > requiring a driver handling the device exists (bus/platform which has
> > *no* in-tree driver, though this is a different story..).
> 
> Thanks for all the input, David.
> 
> In my case it's only PCI bus devices, so I found a relatively simple workaround instead of using the "-a" trick.
> 
> I start by not binding the devices to the vfio-pci kernel driver.
> Then I call rte_eal_init(); and the devices are not detected by the EAL at this point.
> Then I bind the devices to the vfio-pci kernel driver.
> And finally I call rte_bus_scan() and rte_bus_probe(); then the devices are detected at this point.
> 
> Note: I have to pass "--iova-mode pa" to EAL init.
> Otherwise EAL init chooses VA mode, causing my drivers to not load, with the following message:
> "PCI_BUS:   Expecting 'PA' IOVA mode but current mode is 'VA', not initializing"
> 
> For anyone following the thread...
> 
> During experimentation, I observed that if the devices were bound to the vfio-pci kernel driver, and "-a ..." was passed to rte_eal_init() it did - as expected - not load the drivers.
> However, when calling rte_bus_scan() and rte_bus_probe() afterwards, the drivers were also not auto-detected at this point.
> I guess the "-a ..." parameter is sticky, which makes sense.
> 
Yes. The assumption is that, if you don't want to auto-probe, you want to
instead add the devices individually, device by device.

/Bruce