John Hubbard
2025-Oct-02 17:49 UTC
[PATCH v2 1/2] rust: pci: skip probing VFs if driver doesn't support VFs
On 10/2/25 10:40 AM, Danilo Krummrich wrote:> On Thu Oct 2, 2025 at 7:37 PM CEST, Danilo Krummrich wrote: >> On Thu Oct 2, 2025 at 7:05 PM CEST, Jason Gunthorpe wrote: >>> On Thu, Oct 02, 2025 at 06:05:28PM +0200, Danilo Krummrich wrote: >>>> On Thu Oct 2, 2025 at 5:23 PM CEST, Jason Gunthorpe wrote: >>>>> This is not what I've been told, the VF driver has significant >>>>> programming model differences in the NVIDIA model, and supports >>>>> different commands. >>>> >>>> Ok, that means there are some more fundamental differences between the host PF >>>> and the "VM PF" code that we have to deal with. >>> >>> That was my understanding. >>> >>>> But that doesn't necessarily require that the VF parts of the host have to be in >>>> nova-core as well, i.e. with the information we have we can differentiate >>>> between PF, VF and PF in the VM (indicated by a device register). >>> >>> I'm not entirely sure what you mean by this.. >>> >>> The driver to operate the function in "vGPU" mode as indicated by the >>> register has to be in nova-core, since there is only one device ID. >> >> Yes, the PF driver on the host and the PF (from VM perspective) driver in the VM >> have to be that same. But the VF driver on the host can still be a seaparate >> one. >> >>>>> If you look at the VFIO driver RFC it basically does no mediation, it >>>>> isn't intercepting MMIO - the guest sees the BARs directly. Most of >>>>> the code is "profiling" from what I can tell. Some config space >>>>> meddling. >>>> >>>> Sure, there is no mediation in that sense, but it needs quite some setup >>>> regardless, no? >>>> >>>> I thought there is a significant amount of semantics that is different between >>>> booting the PF and the VF on the host. >>> >>> I think it would be good to have Zhi clarify more of this, but from >>> what I understand are at least three activites comingled all together: >>> >>> 1) Boot the PF in "vGPU" mode so it can enable SRIOV >>For this, we could pass a kernel module parameter to nova-core.>> Ok, this might be where the confusion above comes from. When I talk about >> nova-core in vGPU mode I mean nova-core running in the VM on the (from VM >> perspective) PF. >> >> But you seem to mean nova-core running on the host PF with vGPU on top? That of >> course has to be in nova-core. >> >>> 2) Enable SRIOV and profile VFs to allocate HW resources to them >> >> I think that's partially in nova-core and partially in vGPU; nova-core providing >> the abstraction of the corresponding firmware / hardware interfaces and vGPU >> controlling the semantics of the resource handling? >> >> This is what I thought vGPU has a secondary part for where it binds to nova-core >> through the auxiliary bus, i.e. vGPU consisting out of two drivers actually; the >> VFIO parts and a "per VF resource controller". > > Forgot to add: But I think Zhi explained that this is not necessary and can be > controlled by the VFIO driver, i.e. the PCI driver that binds to the VF itself.Yes, this is the direction that I originally (3 whole days ago, haha) had in mind, after talking with Zhi and a few others: nova-core handles PFs, and the VFIO driver handles the VFs, and use the "is virtual" logic to sort them out. Looking forward to Zhi's reaction to the other approach that you and Jason have been debating. This is all very educational to me, as a VFIO newbie. :)> >>> 3) VFIO variant driver to convert the VF into a "VM PF" with whatever >>> mediation and enhancement needed >> >> That should be vGPU only land. >thanks, -- John Hubbard
Jason Gunthorpe
2025-Oct-02 18:05 UTC
[PATCH v2 1/2] rust: pci: skip probing VFs if driver doesn't support VFs
On Thu, Oct 02, 2025 at 10:49:21AM -0700, John Hubbard wrote:> > Forgot to add: But I think Zhi explained that this is not necessary and can be > > controlled by the VFIO driver, i.e. the PCI driver that binds to the VF itself. > > Yes, this is the direction that I originally (3 whole days ago, haha) had in mind, > after talking with Zhi and a few others: nova-core handles PFs, and the VFIO driver > handles the VFs, and use the "is virtual" logic to sort them out.To be clear, no matter what the VFIO driver bound to the VF should not become entangled with any aux devices. The VFIO VF driver uses pci_iov_get_pf_drvdata() to reach into the PF to request the PF's help. Eg for live migration or things of that nature. My point here is that generally we don't put profiling code in the VFIO driver and then use pci_iov_get_pf_drvdata() to access the PF do actually do the profiling. The VF cannot/should not control profiling of itself - that would be a security problem once it is assigned to a VM. So the profiling resides entirely inside the PF world and should operate without VFIO. As I've said this design is compatible with VFs for containers and so on. So it is the strongly preferred design pattern. Jason