Re: [PATCH 4/4] vl: Prioritize realizations of devices

[Jason Wang <jasowang@redhat.com>]

On Tue, Aug 24, 2021 at 6:37 AM Peter Xu <peterx@redhat.com> wrote:
>
> On Mon, Aug 23, 2021 at 06:05:07PM -0400, Michael S. Tsirkin wrote:
> > On Mon, Aug 23, 2021 at 03:18:51PM -0400, Peter Xu wrote:
> > > On Mon, Aug 23, 2021 at 02:49:12PM -0400, Eduardo Habkost wrote:
> > > > On Wed, Aug 18, 2021 at 03:43:18PM -0400, Peter Xu wrote:
> > > > > QEMU creates -device objects in order as specified by the user's cmdline.
> > > > > However that ordering may not be the ideal order.  For example, some platform
> > > > > devices (vIOMMUs) may want to be created earlier than most of the rest
> > > > > devices (e.g., vfio-pci, virtio).
> > > > >
> > > > > This patch orders the QemuOptsList of '-device's so they'll be sorted first
> > > > > before kicking off the device realizations.  This will allow the device
> > > > > realization code to be able to use APIs like pci_device_iommu_address_space()
> > > > > correctly, because those functions rely on the platfrom devices being realized.
> > > > >
> > > > > Now we rely on vmsd->priority which is defined as MigrationPriority to provide
> > > > > the ordering, as either VM init and migration completes will need such an
> > > > > ordering.  In the future we can move that priority information out of vmsd.
> > > > >
> > > > > Signed-off-by: Peter Xu <peterx@redhat.com>
> > > >
> > > > Can we be 100% sure that changing the ordering of every single
> > > > device being created won't affect guest ABI?  (I don't think we can)
> > >
> > > That's a good question, however I doubt whether there's any real-world guest
> > > ABI for that.  As a developer, I normally specify cmdline parameter in an adhoc
> > > way, so that I assume most parameters are not sensitive to ordering and I can
> > > tune the ordering as wish.  I'm not sure whether that's common for qemu users,
> > > I would expect so, but I may have missed something that I'm not aware of.
> > >
> > > Per my knowledge the only "guest ABI" change is e.g. when we specify "vfio-pci"
> > > to be before "intel-iommu": it'll be constantly broken before this patchset,
> > > while after this series it'll be working.  It's just that I don't think those
> > > "guest ABI" is necessary to be kept, and that's exactly what I want to fix with
> > > the patchset..
> > >
> > > >
> > > > How many device types in QEMU have non-default vmsd priority?
> > >
> > > Not so much; here's the list of priorities and the devices using it:
> > >
> > >        |--------------------+---------|
> > >        | priority           | devices |
> > >        |--------------------+---------|
> > >        | MIG_PRI_IOMMU      |       3 |
> > >        | MIG_PRI_PCI_BUS    |       7 |
> > >        | MIG_PRI_VIRTIO_MEM |       1 |
> > >        | MIG_PRI_GICV3_ITS  |       1 |
> > >        | MIG_PRI_GICV3      |       1 |
> > >        |--------------------+---------|
> >
> > iommu is probably ok. I think virtio mem is ok too,
> > in that it is normally created by virtio-mem-pci ...
>
> Hmm this reminded me whether virtio-mem-pci could have another devfn allocated
> after being moved..
>
> But frankly I still doubt whether we should guarantee that guest ABI on user
> not specifying addr=XXX in pci device parameters - I feel like it's a burden
> that we don't need to carry.
>
> (Btw, trying to keep the order is one thing; declare it guest ABI would be
>  another thing to me)
>
> >
> >
> >
> > > All the rest devices are using the default (0) priority.
> > >
> > > >
> > > > Can we at least ensure devices with the same priority won't be
> > > > reordered, just to be safe?  (qsort() doesn't guarantee that)
> > > >
> > > > If very few device types have non-default vmsd priority and
> > > > devices with the same priority aren't reordered, the risk of
> > > > compatibility breakage would be much smaller.
> > >
> > > I'm also wondering whether it's a good thing to break some guest ABI due to
> > > this change, if possible.
> > >
> > > Let's imagine something breaks after applied, then the only reason should be
> > > that qsort() changed the order of some same-priority devices and it's not the
> > > same as user specified any more.  Then, does it also means there's yet another
> > > ordering requirement that we didn't even notice?
> > >
> > > I doubt whether that'll even happen (or I think there'll be report already, as
> > > in qemu man page there's no requirement on parameter ordering).  In all cases,
> > > instead of "keeping the same priority devices in the same order as the user has
> > > specified", IMHO we should make the broken devices to have different priorities
> > > so the ordering will be guaranteed by qemu internal, rather than how user
> > > specified it.
> >
> > Well giving user control of guest ABI is a reasonable thing to do,
> > it is realize order that users do not really care about.
>
> Makes sense.
>
> >
> > I guess we could move pci slot allocation out of realize
> > so it does not depend on realize order?
>
> Yes that sounds like another approach, but it seems to require more changes.

It looks to me this doesn't solve the issue of using virtio-mmio with vhost?

Thanks

>
> Thanks,
>
> --
> Peter Xu
>