On Fri, Oct 15, 2021 at 10:48:41AM +0000, Bertrand Marquis wrote:
> Hi Jan,
>
> > On 15 Oct 2021, at 11:41, Jan Beulich <jbeul...@suse.com> wrote:
> >
> > On 15.10.2021 12:33, Bertrand Marquis wrote:
> >>> On 15 Oct 2021, at 11:24, Jan Beulich <jbeul...@suse.com> wrote:
> >>> On 15.10.2021 11:52, Bertrand Marquis wrote:
> >>>>> On 15 Oct 2021, at 09:32, Roger Pau Monné <roger....@citrix.com> wrote:
> >>>>> On Thu, Oct 14, 2021 at 03:49:50PM +0100, Bertrand Marquis wrote:
> >>>>>> @@ -752,6 +752,19 @@ int pci_add_device(u16 seg, u8 bus, u8 devfn,
> >>>>>>
> >>>>>> check_pdev(pdev);
> >>>>>>
> >>>>>> +#ifdef CONFIG_ARM
> >>>>>> + /*
> >>>>>> + * On ARM PCI devices discovery will be done by Dom0. Add vpci
> >>>>>> handler when
> >>>>>> + * Dom0 inform XEN to add the PCI devices in XEN.
> >>>>>> + */
> >>>>>> + ret = vpci_add_handlers(pdev);
> >>>>>> + if ( ret )
> >>>>>> + {
> >>>>>> + printk(XENLOG_ERR "Setup of vPCI failed: %d\n", ret);
> >>>>>> + goto out;
> >>>>>> + }
> >>>>>> +#endif
> >>>>>
> >>>>> I think vpci_add_handlers should be called after checking that
> >>>>> pdev->domain is != NULL, so I would move this chunk a bit below.
> >>>>
> >>>> On arm this would prevent the dom0less use case or to have the PCI
> >>>> bus enumerated from an other domain.
> >>>> @oleksandr: can you comment on this one, you might have a better
> >>>> answer than me on this ?
> >>>
> >>> Well, without Xen doing the enumeration, some other entity would need
> >>> to do so, including the reporting to Xen. Obviously without a Dom0 it
> >>> would be ambiguous which domain to assign the device to; perhaps it
> >>> should be the caller in this case? That would make that caller domain
> >>> a pseudo-hwdom though, as far as PCI is concerned, which may not be
> >>> desirable according to my (limited) understanding of dom0less.
> >>
> >> This is not really related to this patch but the plan is the following:
> >> - enumeration would have to be done by the firmware or boot loader before
> >> - xen will have some code to detect PCI devices
> >> - dom0less can be used to assign PCI devices to guest
> >>
> >> Anyway does not change the fact that this must be called when domain is
> >> not NULL and I will fix that.
> >
> > Since we now all seem to agree that the NULL would have been a problem,
> > may I ask in how far any of this has actually been tested?
>
> With the whole serie currently on gitlab we have extensively tested passing
> through PCI devices on Arm in several configuration (number of device, MSI,
> MSI-X) and check that PCI was still functional on x86.
>
> With the patches pushed to Xen right now it was checked that:
> - xen compiles properly on arm32, arm64 and x86
> - xen compiles properly with VPCI activated (using a patch) on arm32 and arm64
> - xen on x86 is functionnal (using basic test on QEMU)
> - xen on arm64 is functionnal (with some extensive tests on different targets)
I thinks it's unlikely, but since I haven't checked myself, could you
see if the vpci user-space test harness (tools/tests/vpci) still
builds and functions properly?
Thanks, Roger.
