On 22/11/16 00:08, Andrea Bolognani wrote: > On Mon, 2016-11-21 at 13:12 +1100, Alexey Kardashevskiy wrote: >>>>> 1) switch to PCI Express on newer machine types, and >>>>> expose some sort of capability through QMP so that >>>>> libvirt can know about the switch >>>>> >>>>> [...] >>>>> Option 1) would break horribly with existing libvirt >>>>> versions, and so would Option 2) if we default to using >>>> >>>> How exactly 1) will break libvirt? Migrating from pseries-2.7 to >>>> pseries-2.8 does not work anyway, and machines are allowed to behave >>>> different from version to version, what distinct difference will using >>>> "pseries-pcie-X.Y" make? >>> >>> Existing libvirt versions assume that pseries guests have >> >> If libvirt is using just "pseries" (without a version), then having a >> virtual PCIe-PCI bridge (and "pci.0" always available by default) will do it. > > Please don't. Any device that is included in the guest > by default and can't be turned off makes libvirt's life > significantly more difficult (see below). > >> If libvirt is using a specific version of pseries, then it already knows >> that <=2.7 has pci.0 as a root, pcie.0 otherwise. libvirt has a knowledge >> what QEMU version has what, right? > > It doesn't yet, that's the point :) > > We *could* add such knowledge to libvirt, but *existing* > libvirt versions would still not know about it, which means > that upgrading QEMU withough upgrading libvirt will result > in failure to create new guests. > > >> In what scenario will an additional machine type help? > > Because then libvirt could learn that > > pseries-x.y <-> pci.0 > pseries-pcie-x.y <-> pcie.0 > > the same way it already knows that > > pc-i440fx-x.y <-> pci.0 > pc-q35-x.y <-> pcie.0 > > and choosing between one or the other would be, I think, > much easier for the user as well. > >>> a legacy PCI root bus, and will base their PCI address >>> allocation / PCI topology decisions on that fact: they >>> will, for example, use legacy PCI bridges. >>> >>> So if you used a new QEMU binary with a libvirt version >>> that doesn't know about the change, new guests would end up >>> using the wrong controllers. Existing guests would not be >>> affected as they would stick with the older machine types, >>> of course. >>> >>>> I believe after we introduced the very first >>>> pseries-pcie-X.Y, we will just stop adding new pseries-X.Y. >>> >>> Isn't i440fx still being updated despite the fact that q35 >>> exists? Granted, there are a lot more differences between >>> those two machine types than just the root bus type. >> >> I do not know about i440<->q35 but in pseries the difference is going to be >> very simple. >> >> For example, we did not change the machine type when we switched from >> default OHCI to XHCI, switching from PCI to PCIe does not sound like we >> need a whole new machine type for this either. > > The change from OHCI to XHCI only affected the *default* USB > controller, which libvirt tries its best not to use anyway: > instead, it will prefer to use '-M ...,usb=off' along with > '-device ...' and set both the controller model and its PCI > address explicitly, partially to shield its users from such > changes in QEMU.
Ok. Always likes this approach really. We should start moving to this direction with PHB - stop adding the default PHB at all when -nodefaults is passed (or -machine pseries,pci=off ?) and let libvirt manage PHBs itself (and provide another spapr-phb type like spapr-pcie-host-bridge or add a "pcie_root_bus_type" property to the existing PHB type). What will be wrong with this approach? > > Let's not forget that libvirt is a management layer, and as > such it needs to have as much control as possible over the > virtual hardware presented to the guest, mostly for guest ABI > compatibility purposes. Defaulting to XHCI instead of OHCI, > same as pretty much all other defaults, is good for people who > run QEMU directly, but libvirt needs to be able to control > such knobs itself in order to effectively perform the task it > was designed for. > > Moreover, we're talking about a more fundamental change here: > the PCI Root Bus is not just any random device, it's the one > fundation upon which the entire PCI hierarchy is built. Based > on whether the machine exposes a PCI Express Root Bus or a > legacy PCI Express Root Bus, libvirt will create very > different PCI hierarchies, eg. using several ioh3420 devices > instead of a single pci-bridge device. > > I'm still not sure if that enough to warrant an entirely new > machine type, but it definitely has a more far-reaching impact > than simply flipping the default USB controller from OHCI to > XHCI. > >>> Even if no newer pseries-x.y were to be added after >>> introducing pseries-pcie, you could still easily create >>> guests that use either root bus, so no loss in functionality. >> >> I could do this with the existing pseries if the machine had a "root but >> type" property. > > That was indeed one of my original proposals ;) > > Just note, once again, that the default for this property > would have to be "off" or we would run into the same issues > described above. > > >  Even though, as I mentioned earlier in the thread, > performing version checks on machine types is frowned > upon as it turns into a minefield as soon as backports > and downstream machine types enter the picture... It's > much better to expose some flag through QMP for libvirt > to introspect > -- > Andrea Bolognani / Red Hat / Virtualization > -- Alexey