On Wed, Mar 04, 2026 at 06:50:02PM +0100, Danilo Krummrich wrote: > On Wed Mar 4, 2026 at 6:02 PM CET, Leon Romanovsky wrote: > > On Wed, Mar 04, 2026 at 12:45:51PM -0400, Jason Gunthorpe wrote: > >> On Wed, Mar 04, 2026 at 06:27:11PM +0200, Leon Romanovsky wrote: > >> > On Wed, Mar 04, 2026 at 03:57:57PM +0100, Danilo Krummrich wrote: > >> > > On Wed Mar 4, 2026 at 3:26 PM CET, Leon Romanovsky wrote: > >> > > > On Wed, Mar 04, 2026 at 10:18:52AM -0400, Jason Gunthorpe wrote: > >> > > >> On Wed, Mar 04, 2026 at 10:47:50AM +0200, Leon Romanovsky wrote: > >> > > >> > On Tue, Mar 03, 2026 at 04:15:20PM -0500, Peter Colberg wrote: > >> > > >> > > Add Rust abstractions for the Single Root I/O Virtualization > >> > > >> > > (SR-IOV) > >> > > >> > > capability of a PCI device. Provide a minimal set of wrappers > >> > > >> > > for the > >> > > >> > > SR-IOV C API to enable and disable SR-IOV for a device, and > >> > > >> > > query if > >> > > >> > > a PCI device is a Physical Function (PF) or Virtual Function > >> > > >> > > (VF). > >> > > >> > > >> > > >> > <...> > >> > > >> > > >> > > >> > > For PF drivers written in C, disabling SR-IOV on remove() may > >> > > >> > > be opted > >> > > >> > > into by setting the flag managed_sriov in the pci_driver > >> > > >> > > structure. For > >> > > >> > > PF drivers written in Rust, disabling SR-IOV on unbind() is > >> > > >> > > mandatory. > >> > > >> > > >> > > >> > Why? Could you explain the rationale behind this difference > >> > > >> > between C and > >> > > >> > Rust? Let me remind you that SR‑IOV devices which do not disable > >> > > >> > VFs do so > >> > > >> > for a practical and well‑established reason: maximizing hardware > >> > > >> > utilization. > >> > > >> > >> > > >> Personally I think drivers doing this are wrong. That such a driver > >> > > >> bug was allowed to become UAPI is pretty bad. The rust approach is > >> > > >> better. > >> > > > > >> > > > We already had this discussion. I see this as a perfectly valid > >> > > > use-case. > >> > > > >> > > Can you remind about a specific use-case for this please? (Ideally, > >> > > one that > >> > > can't be solved otherwise.) > >> > > >> > You create X VFs through sriov_configure, unbind PF, bind it to vfio > >> > instead and forward (X + 1) functions to different VMs. > >> > >> No, illegal, and it doesn't even work right. When VFIO FLRs the PF it > >> will blow up the half baked SRIOV and break everything. > > > > The FLR can be disabled. For example, PCI_DEV_FLAGS_NO_FLR_RESET flag > > will do it. > > But this is a quirk and not a feature, no? So, we shouldn't use it as a > baseline > for actual features.
My point is slightly different. I was trying to explain the rationale for preserving VFs after the PF is unbound, a design choice that predates the introduction of the VFIO .srio_configure callback. Thanks
