On Tue, Jun 14, 2016 at 10:19:32AM +0300, Marcel Apfelbaum wrote: > Create the iommu device with '-device intel-iommu' instead of > '-machine,iommu=on'. > > The device is part of the machine properties because we wanted > to ensure is created before any other PCI device. > > The alternative is to skip the bus_master_enable_region at > the time the device is created. We can create this region > at machine_done phase. (patch 1) > > Then we need to enable sysbus devices(*) for PC machines (patch 2), > since intel-iommu is a sysbus device. > > Patch 3 moves the IOMMU init proces into iommu's realize function > and allows the device creation in both ways.
This breaks make check on x86: GTESTER check-qtest-ppc64 Broken pipe GTester: last random seed: R02Sf26cb511ed533726cf1afef21a00bde4 /scm/qemu/tests/Makefile.include:668: recipe for target 'check-qtest-ppc64' failed make: *** [check-qtest-ppc64] Error 1 > Finally patch 4 removes the iommu machine property. > > v3 -> v4: > - Rebased on mst/pci tree (Michael). > > v2 -> v3: > - Add machine_done notifier in pci_bus realize and remove it unrealize > (Paolo). > - Add comments for 'cannot_instantiate_with_device_add_yet' (Markus). > - Split adding the -device iommu support and removing the iommu machine > property (Michael). > - Use pci_setup_iommu as before (Peter) > - Mark intel-iommu as not hot-pluggable (Peter) > - Rebased on master > > v1 -> v2: > - Enable bus_master also on init if the guest OS already booted to enable > hotplug (Paolo). > - Add a machine_done notifier to PCIBus instead of adding functionality > for q35 machine_done callback. The main reason is we don't want to > replicate > the code for all platforms that support PCI and is also cleaner this way. > - Added 'cannot_instantiate_with_device_add_yet' to sysbus devices that lead > to crashes if added with -device. > - Rebased on master > > > (*) Creates a new problem since we have now a bunch of > new devices that can be created with -device on Q35: > name "q35-pcihost", bus System > name "sysbus-ohci", bus System, desc "OHCI USB Controller" > name "allwinner-ahci", bus System > name "cfi.pflash01", bus System > name "esp", bus System > name "SUNW,fdtwo", bus System > name "sysbus-ahci", bus System > name "sysbus-fdc", bus System > name "vfio-amd-xgbe", bus System, desc "VFIO AMD XGBE" > name "vfio-calxeda-xgmac", bus System, desc "VFIO Calxeda XGMAC" > name "virtio-mmio", bus System > name "fw_cfg", bus System > name "fw_cfg_io", bus System > name "fw_cfg_mem", bus System > name "generic-sdhci", bus System > name "hpet", bus System > name "i440FX-pcihost", bus System > name "intel-iommu", bus System > name "ioapic", bus System > name "isabus-bridge", bus System > name "kvm-ioapic", bus System > name "kvmclock", bus System > name "kvmvapic", bus System > name "pxb-host", bus System > > Took care of the ones creating immediate issues (like crashes) by marking them > as 'cannot_instantiate_with_device_add_yet'. I didn't mark them all because: > - libvirt will mask them anyway > - some of them have already a "protection" in place > - it is possible that some of them can be actually used with -device on > other platform. > - those are not 'interesting' scenarios. > If somebody spots devices in the list that cannot be added with -device on > any platform > please let me know and I'll mark them. > > > Thanks, > Marcel > > Marcel Apfelbaum (4): > hw/pci: delay bus_master_enable_region initialization > q35: allow dynamic sysbus > hw/iommu: enable iommu with -device > machine: remove iommu property > > hw/core/machine.c | 20 ------------------ > hw/i386/intel_iommu.c | 16 +++++++++++++++ > hw/i386/pc_q35.c | 2 +- > hw/pci-bridge/pci_expander_bridge.c | 2 ++ > hw/pci-host/piix.c | 2 ++ > hw/pci-host/q35.c | 31 +++------------------------- > hw/pci/pci.c | 41 > +++++++++++++++++++++++++++++-------- > include/hw/pci-host/q35.h | 1 - > include/hw/pci/pci_bus.h | 2 ++ > include/sysemu/sysemu.h | 1 + > qemu-options.hx | 3 --- > vl.c | 5 +++++ > 12 files changed, 64 insertions(+), 62 deletions(-) > > -- > 2.4.3