Re: [Qemu-devel] [Xen-devel] [PATCH] libxl: change default QEMU machine to pc-i440fx-1.6

[Stefano Stabellini]

On Wed, 28 May 2014, Fabio Fantoni wrote:
> Il 26/05/2014 10:00, Fabio Fantoni ha scritto:
> > Il 25/05/2014 16:14, Stefano Stabellini ha scritto:
> > > On Fri, 23 May 2014, Fabio Fantoni wrote:
> > > > Il 23/05/2014 18:07, Stefano Stabellini ha scritto:
> > > > > Choose pc-i440fx-1.6 instead of pc for HVM guests, so that we know for
> > > > > sure what is the machine that we are emulating.
> > > > > 
> > > > > Use pc-i440fx-1.6 regardless of the xen_platform_pci option. Add the
> > > > > xen-platform device if requested. Choose slot 2 for the xen-platform
> > > > > device for compatibility with current installations. In case of Intel
> > > > > graphic passthrough, slot 2 might be needed by the grafic card.
> > > > > However
> > > > > now that we can specify the slot explicitly, it is easy to change the
> > > > > position of the xen-platform device on the PCI bus if graphic
> > > > > passthrough is enabled.
> > > > > 
> > > > > Move the machine options earlier, before any other emulated devices
> > > > > options. Otherwise the selected PCI slot for the xen-platform device
> > > > > is
> > > > > not available in QEMU.
> > > > > 
> > > > > Specify PIIX4_PM.acpi-pci-hotplug-with-bridge-support=off, because
> > > > > differently from xenfv, the other QEMU machines do not have that
> > > > > option
> > > > > off by default.
> > > > > 
> > > > > This patch does not change the emulated environment in the guest.
> > > > > 
> > > > > Refer to this thread:
> > > > > http://marc.info/?l=xen-devel&m=140023775929625&w=2
> > > > > 
> > > > > Signed-off-by: Stefano Stabellini <address@hidden>
> > > > > 
> > > > > diff --git a/tools/libxl/libxl_dm.c b/tools/libxl/libxl_dm.c
> > > > > index 8abed7b..fef684f 100644
> > > > > --- a/tools/libxl/libxl_dm.c
> > > > > +++ b/tools/libxl/libxl_dm.c
> > > > > @@ -476,6 +476,29 @@ static char **
> > > > > libxl__build_device_model_args_new(libxl__gc *gc,
> > > > >            flexarray_vappend(dm_args, "-k", keymap, NULL);
> > > > >        }
> > > > >    +    flexarray_append(dm_args, "-machine");
> > > > > +    switch (b_info->type) {
> > > > > +    case LIBXL_DOMAIN_TYPE_PV:
> > > > > +        flexarray_append(dm_args, "xenpv");
> > > > > +        for (i = 0; b_info->extra_pv && b_info->extra_pv[i] != NULL;
> > > > > i++)
> > > > > +            flexarray_append(dm_args, b_info->extra_pv[i]);
> > > > > +        break;
> > > > > +    case LIBXL_DOMAIN_TYPE_HVM:
> > > > > +        flexarray_append(dm_args, "pc-i440fx-1.6,accel=xen");
> > 
> > I think that a note in README should be added: qemu >=1.6.1 is required for
> > hvm domUs.
> > 
> > About the other xl parameter to be added I think that should be:
> > qemu_machine_type="i440fx"|"q35" (when also q35 will be supported on xen)
> > qemu_machine_version="x.y" to specify the machine version, useful for
> > debug/testing and other some cases.
> > 
> > > > > +        flexarray_append(dm_args, "-global");
> > > > > +        flexarray_append(dm_args,
> > > > > "PIIX4_PM.acpi-pci-hotplug-with-bridge-support=off");
> > > > I think is good add a comment for remember to remove this workaround
> > > > when pc
> > > > > =2.1 will be the default since qemu 2.1 will fix it.
> > > > https://lists.gnu.org/archive/html/qemu-devel/2014-05/msg04789.html
> > > The workaround is not actually harmful, it doesn't need to be removed
> > > when pc >= 2.1 in QEMU.
> > > 
> > > 
> > > > > +        if (libxl_defbool_val(b_info->u.hvm.xen_platform_pci)) {
> > > > > +            flexarray_append(dm_args, "-device");
> > > > > +            flexarray_append(dm_args, "xen-platform,addr=0x2");
> > > > The fixed pci address to 0x2 probably is a problem with intel gpu
> > > > passthrough:
> > > > http://lists.gnu.org/archive/html/qemu-devel/2014-05/msg03726.html
> > > Right, however we cannot really change the default position of the
> > > xen-platform device on the PCI bus, otherwise we'll end up changing the
> > > emulated environment for all the VMs out there.
> > > 
> > > I'll leave it to Tiejun to move xen-platform to another slot when
> > > graphic passthrough is enabled.
> > 
> 
> I found another case of problem with xen-platform's fixed pci slot.
> I tested this patch and I saw that qemu not start also in other cases, for
> example the domU of my test:
> > qemu-system-i386: -device xen-platform,addr=0x2: PCI: slot 2 function 0 not
> > available for xen-platform, in use by intel-hda
> > qemu-system-i386: -device xen-platform,addr=0x2: Device initialization
> > failed.
> > qemu-system-i386: -device xen-platform,addr=0x2: Device 'xen-platform' could
> > not be initialized
> 
> The domU's xl cfg:
> > name='W7'
> > builder="hvm"
> > #device_model_override="/usr/lib/xen/bin/qemu-gdb"
> > #device_model_override="/usr/bin/qemu-system-x86_64"
> > #bios="ovmf"
> > memory=2048
> > vcpus=2
> > #nestedhvm=1
> > #vif=['model=e1000,bridge=xenbr0']
> > vif=['bridge=xenbr0,mac=00:16:3e:42:ae:8f']
> > disk=['/mnt/vm/disks/W7.disk1.xm,raw,hda,rw',',raw,hdb,ro,cdrom']
> > boot='dc'
> > device_model_version="qemu-xen"
> > viridian=1
> > vnc=0
> > keymap="it"
> > on_crash="destroy"
> > vga="qxl"
> > #videoram=64
> > spice=1
> > spicehost='0.0.0.0'
> > spiceport=6002
> > spicedisable_ticketing=1
> > spicevdagent=1
> > spice_clipboard_sharing=0
> > spice_image_compression="off"
> > spice_streaming_video="filter"
> > spiceusbredirection=4
> > soundhw="hda"

>From the error message, it looks like soundhw is the option that
conflicts with xen-platform,addr=0x2.


> > localtime=1
> > usbversion=2
> 
> Probably there are also other cases that can create a problem with
> xen-platform fixed address, FWIK now new usb controller (with usbversion) is
> the only other with fixed pci address in libxl, all other emulated qemu
> components not.
> And call it before in qemu binary starts notaffect the pci slot order because
> the xen-platform is already before audio.

Actually it looks like that by passing -device xen-platform before the
other options, it already tries to assign slot 2 if possible. I don't
need to specify addr=0x2. That seems to be the best course of action.

Thanks for testing!