On Thu, 2025-08-14 at 09:33 -0700, Farhan Ali wrote:
> On 8/14/2025 6:12 AM, Niklas Schnelle wrote:
> > On Wed, 2025-08-13 at 16:56 -0600, Alex Williamson wrote:
> > > On Wed, 13 Aug 2025 14:52:24 -0700
> > > Farhan Ali <alifm@xxxxxxxxxxxxx> wrote:
> > >
> > > > On 8/13/2025 1:30 PM, Alex Williamson wrote:
> > > > > On Wed, 13 Aug 2025 10:08:19 -0700
> > > > > Farhan Ali <alifm@xxxxxxxxxxxxx> wrote:
> > > > >
> > > > > > For zPCI devices we should drive a platform specific function reset
> > > > > > as part of VFIO_DEVICE_RESET. This reset is needed recover a zPCI device
> > > > > > in error state.
> > > > > >
> > > > > > Signed-off-by: Farhan Ali <alifm@xxxxxxxxxxxxx>
> > > > > > ---
> > > > > > arch/s390/pci/pci.c | 1 +
> > > > > > drivers/vfio/pci/vfio_pci_core.c | 4 ++++
> > > > > > drivers/vfio/pci/vfio_pci_priv.h | 5 ++++
> > > > > > drivers/vfio/pci/vfio_pci_zdev.c | 39 ++++++++++++++++++++++++++++++++
> > > > > > 4 files changed, 49 insertions(+)
>
--- snip ---
> > Now for pci_reset_hotplug_slot() via VFIO_DEVICE_PCI_HOT_RESET I'm not
> > sure why that won't work as is. @Farhan do you know?
>
> VFIO_DEVICE_PCI_HOT_RESET would have been sufficient interface for
> majority of PCI devices on s390x as that would drive a bus reset. It was
> sufficient as most devices were single bus devices. But in the latest
> generation of machines (z17) we expose true SR-IOV and an OS can have
> access to both PF and VFs and so these are on the same bus and can have
> different ownership based on what is bound to vfio-pci.
>
Talked to Farhan a bit off list. I think the problem boils down to
this. The s390 PCI support, due to there always being a hypervisor,
does hot and cold plug on a per PCI function basis. And so the hotplug
driver's reset_slot() is just a wrapper around zpci_hot_reset_device()
and also does per PCI function reset.
Now when doing a VFIO_PCI_HOT_RESET this still doesn't give us a usable
per function reset because vfio_pci_ioctl_get_pci_hot_reset_info()'s
grouping assumes, quite naturally, that as a slot reset this will reset
an entire slot and so the ownership checks fail when we use this reset
on e.g. an SR-IOV capable device with PFs + VFs where s390 exposes
their PCI topology even while retaining per function hotplug.
As some more background, we've had the virtual PCI topology for SR-IOV
capable devices since commit 44510d6fa0c0 ("s390/pci: Handling
multifunctions") added in v6.9 but only with z17 will they be generally
available.
Thanks,
Niklas
