Re: [libvirt] [PATCH 4/4] qemu: don't be as insistent about adding dmi-to-pci-bridge or pci-bridge

On 04/23/2016 12:46 PM, Cole Robinson wrote: ... and it turns out this patch doesn't go far enough to have any useful effect. The problem is that the code that automatically assigns PCI addresses does its best to always be sure that there is at least one empty hot-pluggable "standard PCI" slot on the system, so even if we don't explicitly add any pci-bridge to a Q35 domain, the address-assignment code will automatically add one anyway; and since the pci-bridge requires a standard PCI slot itself, we still need a dmi-to-pci-bridge anyway (which provides 32 non-hotpluggable standard PCI slots). I had been aiming to re-do how libvirt sets up the PCI controllers for Q35, eventually eliminating the current dmi-to-pci-bridge + pci-bridge in favor of a pure PCIe setup, using pcie-root-ports and pcie-switch-(up|down)stream-ports, and this was going to be the first step towards that. But after discussing that idea more with Alex Williamson on Friday, I think we *shouldn't* do it, but just leave the PCI controller setup essentially as it is (one possible change - allow connecting pci-bridge directly into a pcie-root port, thus eliminating the dmi-to-pci-bridge) What prompted the idea to change to pure PCIe controllers was that qemu allows any PCI device to be plugged into a PCIe slot and it apparently functions just fine, and there had been occasional complaints that plugging everything into a pci-bridge was somehow problematic (at one point someone suggested that virtio-net didn't work correctly on ARM if plugged into a non-Express slot, but I think that was later disproved). The problem with PCIe-only controllers, as Alex pointed out, is that when a non-Express device is plugged into a PCIe slot, the guest OS will see an apparently PCIe device that has no PCIe capabilities, and while so far this hasn't caused any problem, there is no guarantee that it won't - PCIe devices are supposed to have PCIe capabilities. Since the only emulated qemu device that does this is the NEC UHCI USB3 controller, it sees that we'll need to keep setting up the pci-bridge and plugging all the rest of the devices in there. Still, it would be nice to allow *those who really want to* to have a "pure PCIe" controller setup. We could do that if we did two things: 1) modify the PCI address auto-assignment code to not insist on always having at least one hot-pluggable standard PCI slot available. and one of the following: 2a) don't always add a dmi-to-pci-bridge, but instead only do so *if necessary* in order to plug in a pci-bridge (which would only be added if necessary). or 2b) permit auto-assigning a pci-bridge to plug into a port of pcie-root, thus eliminating the need for dmi-to-pci-bridge completely. Does anyone have opinions about (1) or (2b)? (NB: in the long run it might be nice if qemu would automatically advertise PCIe capabilities for any of its emulated devices that was plugged into a PCIe slot - this would make it possible to get rid of all non-Express controllers without violating any of the PCI specs. However, according to Alex this would have 0 practical effect (no better performance / actual capabilities of the devices / use of PCI address space), and it would also require that qemu gave us some way to understand it was going to do this (i.e. introspection) so I don't know how useful it would be to go to all that trouble.)