Re: Potential issue with hypervisor_isolated_pci_functions() handling

[Niklas Schnelle <schnelle@linux.ibm.com>]

On Wed, 2025-10-22 at 22:49 +0800, Huacai Chen wrote:
> Hi, Niklas,
> 
> On Wed, Oct 22, 2025 at 7:40 PM Niklas Schnelle <schnelle@linux.ibm.com> wrote:
> > 
> > Hi Huacai, Hi Jan, Hi Bjorn,
> > 
> > I noticed that with commit a02fd05661d7 ("PCI: Extend isolated function
> > probing to LoongArch") LoongArch now also makes use of the isolated
> > function probing. First, nice to see this sees more users and that the
> > interface is useful to you.
> > 
> > Seeing this, I was reminded of an issue I ran into when using this
> > mechanism with SR-IOV capable devices. In that case VFs with devfn > 7
> > and PCI_SLOT(devfn) != 0 wouldn't get probed.
> > This is because if a device is found next_fn() will check whether dev-
> > > multifunction is set. So if for example devfn == 8 is found (fn == 0)
> > dev->multifunction won't be set in pci_scan_slot() and for VFs it's
> > also not set via PCI_HEADER_TYPE_MFD. So we won't check for devfn == 9
> > even though that could very well be there.
> > 
> > Now for s390 this currently doesn't cause issues because for all
> > multifunction devices we have, we either get a VF alone and then since
> > commit 25f39d3dcb48 ("s390/pci: Ignore RID for isolated VFs") use devfn
> > == 0, or we have the parent PF passed-through and then VFs always get
> > hot plugged after SR-IOV enable, which then uses
> > pci_scan_single_device(). And for non VFs we always have devfn == 0
> > and/or devfn == 1 and multifunction via the header. So in a sense the
> > above commit works around the issue, though that wasn't its primary
> > intention.
> > 
> > Did either of you also run into this issue or can reproduce it?
> > 
> > Somewhat related, if ARI is enabled this would also break the isolated
> > function case including on s390 where ARI is used by the platform
> > firmware, but detected as off by Linux because there is no struct
> > pci_dev associated with the PCI bus. If I patch Linux to correctly
> > detect ARI, it no longer finds an isolated PF with devfn == 1.
> LoongArch do have some problems after commit a02fd05661d7 ("PCI:
> Extend isolated function probing to LoongArch"). Please see:
> https://lore.kernel.org/linux-pci/20251014074100.2149737-1-chenhuacai@loongson.cn/
> 
> Now we don't know what happens exactly, so I haven't answered Bjorn's
> question...
> 
> Huacai
> 

Hi Huacai,

I have now send[0] a fix proposal together with correctly exposing ARI
enablement in s390. I also reproduced both the issue I saw with SR-IOV
and with ARI and confirmed that the first patch in the series fixes
them. You're on Cc and I'd highly appreciate if you could test whether
this fixes your issues as well.

Thanks,
Niklas

[0]
https://lore.kernel.org/linux-pci/20251023-ari_no_bus_dev-v4-0-8482e2ed10bd@linux.ibm.com/