Re: [libvirt] [PATCHv3 3/3] qemu: Prefer VFIO for PCI device passthrough

On 10/04/2013 08:55 AM, Peter Krempa wrote: > Prefer using VFIO (if available) to the legacy KVM device passthrough. > > With this patch a PCI passthrough device without the driver configured > will be started with VFIO if it's available on the host. If not legacy > KVM passthrough is checked and error is reported if it's not available. > --- > docs/formatdomain.html.in | 9 ++++----- > src/conf/domain_conf.h | 2 +- > src/qemu/qemu_command.c | 3 ++- > src/qemu/qemu_hostdev.c | 21 +++++++++++++++++++-- > src/qemu/qemu_hostdev.h | 3 ++- > src/qemu/qemu_hotplug.c | 2 +- > src/qemu/qemu_process.c | 15 ++++++++------- > tests/qemuxml2argvtest.c | 11 +++++++++++ > 8 files changed, 48 insertions(+), 18 deletions(-) >

> diff --git a/src/qemu/qemu_hostdev.c b/src/qemu/qemu_hostdev.c > index dbbc2b4..ad408d8 100644 > --- a/src/qemu/qemu_hostdev.c > +++ b/src/qemu/qemu_hostdev.c > @@ -1366,7 +1366,8 @@ qemuHostdevHostSupportsPassthroughLegacy(void) > > bool > qemuHostdevHostVerifySupport(virDomainHostdevDefPtr *hostdevs, > - size_t nhostdevs) > + size_t nhostdevs, > + virQEMUCapsPtr qemuCaps) Aha. I guess if you follow my recommendation in 2/3, you'll need to pass qemuCaps down through the qemuPrepareHostDevices() call chain (and I don't think that's a bad thing).

> { > int supportsPassthroughKVM = -1; > int supportsPassthroughVFIO = -1; > @@ -1387,6 +1388,23 @@ qemuHostdevHostVerifySupport(virDomainHostdevDefPtr *hostdevs, > } > > switch ((virDomainHostdevSubsysPciBackendType) *backend) { > + case VIR_DOMAIN_HOSTDEV_PCI_BACKEND_DEFAULT: > + if (supportsPassthroughVFIO && > + virQEMUCapsGet(qemuCaps, QEMU_CAPS_DEVICE_VFIO_PCI)) { > + *backend = VIR_DOMAIN_HOSTDEV_PCI_BACKEND_VFIO; > + } else if (supportsPassthroughKVM && > + (virQEMUCapsGet(qemuCaps, QEMU_CAPS_PCIDEVICE) || > + virQEMUCapsGet(qemuCaps, QEMU_CAPS_DEVICE))) { > + *backend = VIR_DOMAIN_HOSTDEV_PCI_BACKEND_KVM; > + } else { > + virReportError(VIR_ERR_CONFIG_UNSUPPORTED, "%s", > + _("host doesn't support passthrough of " > + "host PCI devices")); > + return false; > + } > + > + break; > + Ah! And there is the separate case I was asking for in 2/3! But you're changing the "backend" in the data itself - that will lead to incorrect display when someone does a dumpxml (will it cause problems if someone loads the vfio driver? I guess not, since this should only be the "live" data, not the persistent data)