Re: [Qemu-devel] PCI(e): Documentation "io-reserve" and related properties?

[Andrea Bolognani]

On Thu, 2019-06-06 at 14:20 -0400, Michael S. Tsirkin wrote:
> On Thu, Jun 06, 2019 at 06:19:43PM +0200, Kashyap Chamarthy wrote:
> > Hi folks,
> > 
> > Today I learnt about some obscure PCIe-related properties, in context of
> > the adding PCIe root ports to a guest, namely:
> > 
> >     io-reserve
> >     mem-reserve
> >     bus-reserve
> >     pref32-reserve
> >     pref64-reserve
> > 
> > Unfortunately, the commit[*] that added them provided no documentation
> > whatsover.
> > 
> > In my scenario, I was specifically wondering about what does
> > "io-reserve" mean, in what context to use it, etc.  (But documentation
> > about other properties is also welcome.)
> > 
> > Anyone more well-versed in this area care to shed some light?
> > 
> > 
> > [*] 6755e618d0 (hw/pci: add PCI resource reserve capability to legacy
> >     PCI bridge, 2018-08-21)
> 
> So normally bios would reserve just enough io space to satisfy all
> devices behind a bridge. What if you intend to hotplug more devices?
> These properties allow you to ask bios to reserve extra space.

Is it fair to say that setting io-reserve=0 for a pcie-root-port
would be a way to implement the requirements set forth in

  https://bugzilla.redhat.com/show_bug.cgi?id=1408810

? I tested this on aarch64 and it seems to work as expected, but
then again without documentation it's hard to tell.

More specifically, I created an aarch64/virt guest with several
pcie-root-ports and it couldn't boot much further than GRUB when
the number of ports exceeded 24, but as soon as I added the
io-reserve=0 option I could get the same guest to boot fine with
32 or even 64 pcie-root-ports. I'm attaching the boot log for
reference: there are a bunch of messages about the topic but they
would appear to be benign.

Hotplug seemed to work too: I tried with a single virtio-net-pci
and I could access the network. My understanding is that PCIe
devices are required to work without IO space, so this behavior
matches my expectations.

I wonder, though, what would happen if I had something like

  -device pcie-root-port,io-reserve=0,id=pci.1
  -device pcie-pci-bridge,bus=pci.1

Would I be able to hotplug conventional PCI devices into the
pcie-pci-bridge, or would the lack of IO space reservation for
the pcie-root-port cause issues with that?

-- 
Andrea Bolognani / Red Hat / Virtualization

boot.log
Description: Text Data