RE: [PATCH 12/13] PCI: hv: Convert to pci_assign_unassigned_bus_resources()

From: Michael Kelley

Date: Fri Nov 20 2026 - 12:47:31 EST


From: Ilpo Järvinen <ilpo.jarvinen@xxxxxxxxxxxxxxx> Sent: Friday, November 20, 2026 4:55 AM
>
> On Wed, 18 Nov 2026, Michael Kelley wrote:
>
> > From: Ilpo Järvinen <ilpo.jarvinen@xxxxxxxxxxxxxxx> Sent: Monday, November 16, 2026 7:24 AM
> > >
> > > The pci-hyperv driver scans the bus and assigns resources using
> > > pci_bus_assign_resources(). That function makes bridge window sizing
> > > disjoint from the assignment phase and is going to be removed.
> > >
> > > Convert pci_bus_assign_resources() to
> > > pci_assign_unassigned_bus_resources(). If no other bridges are involved,
> > > the extra sizing should be no-op.
> >
> > I tested this series on an x86 VM in the Azure public cloud that has
> > 5 NVMe devices and a Mellanox VF NIC. All 6 devices are PCI pass-thru
> > devices in the VM that are initially set up using the pci-hyperv.c driver,
> > so each device has its own root bus and bridge. No issues were seen in
> > a basic smoke test.
> >
> > I added some debug code to pci_assign_unassigned_bus_resources(),
> > and indeed the "for_each_pci_bridge()" loop found nothing as each
> > device was setup. So __pci_bus_assign_resources() was called with an
> > empty add_list, and as expected, the new execution path is equivalent
> > to the old pci_bus_assign_resources() call.
> >
> > Question: Your cover letter says drivers that call pci_assign_resources()
> > directly are not covered by this series. pci-hyperv.c *does* call
> > pci_assign_resources() directly in hv_pci_allocate_bridge_windows().
> > Is that a problem? I don't understand all the details of the PCI resource
> > assignment well enough to judge for myself. :-(
>
> There's a small name mixup here.

Indeed! My bad ....

>
> pci_assign_resources() is to graft a device's resource (or a window) into
> resource tree, whereas that calls pci_add_resource() adding windows to the
> root bus and that seems to occur before there is anything underneath. At
> this point, there's nothing to do for the resource fitting and assignment
> algorithm (it iterates the devices underneath).
>
> ...So I don't see a race here related to pci_assign_resources().
>
> Most entrypoints of the fitting and assignment algorith input bus, and I
> don't think anything in kernel would as to assign resources underneath
> that root bus. Only pci_assign_unassigned_resources() looks for all root
> buses, and that is only called from arch init code so I don't think vfio
> setting up it host bridges matters with that call.
>
> There seems to be, in general (not just hyperv), some races that are
> unrelated to resource setup in host bridge setup and scanning. I'll see
> what I can do but it'll be in a separate series and those seems to fall
> under can-shoot-oneself's-own-foot category.
>
> Thanks for taking a look.

Thank you!

Michael