On Mon, 26 Oct 2020 at 07:01, Gerd Hoffmann <kra...@redhat.com> wrote: > On Fri, Oct 23, 2020 at 12:36:05PM +0100, Peter Maydell wrote: > > On Fri, 23 Oct 2020 at 08:10, Gerd Hoffmann <kra...@redhat.com> wrote: > > > > > > Bring new microvm goodies to arm virt too. Wire up > > > -machine usb=on, add sysbus-xhci in case it is enabled. > > > > So my question here is the usual one -- why can't we > > just use a PCI USB controller ? > > Well, pci seems to come with some extra resource needs (see -M pc vs. > -M q35 memory footprint differences discussed some months ago). Thats > why microvm started without pci support, and even now with pcie support > added it is optional (and off by default).
I think PCI is too useful to discard. You can run anything you want (practically) via PCI. If we make it optional then we're going to give ourselves the task of reimplementing memory mapped versions of all the functionality it gives us, all of which is extra code and all of which adds to the amount of stuff on the guest-to-QEMU security boundary. I think to be persuaded that no-PCI was a good idea I'd want to at least see solid figures based on doing this for Arm and on having put a lot of effort into slimming down the PCI handling code/overhead in the guest and still not being able to get near an MMIO setup. That is, try the "improve the guest" approach first. thanks -- PMM
