Re: [Qemu-devel] [PATCH v6 2/7] hw/misc: add vmcoreinfo device

[Michael S. Tsirkin]

On Tue, Oct 10, 2017 at 03:01:10PM -0300, Eduardo Habkost wrote:
> On Tue, Oct 10, 2017 at 04:06:28PM +0100, Daniel P. Berrange wrote:
> > On Tue, Oct 10, 2017 at 05:00:18PM +0200, Marc-André Lureau wrote:
> > > Hi
> > > 
> > > On Tue, Oct 10, 2017 at 10:31 AM, Daniel P. Berrange
> > > <address@hidden> wrote:
> > > > On Tue, Oct 10, 2017 at 12:44:26AM +0300, Michael S. Tsirkin wrote:
> > > >> On Mon, Oct 09, 2017 at 02:02:18PM +0100, Daniel P. Berrange wrote:
> > > >> > On Mon, Oct 09, 2017 at 02:43:44PM +0200, Igor Mammedov wrote:
> > > >> > > On Mon, 9 Oct 2017 12:03:36 +0100
> > > >> > > "Daniel P. Berrange" <address@hidden> wrote:
> > > >> > >
> > > >> > > > On Mon, Sep 11, 2017 at 06:59:24PM +0200, Marc-André Lureau 
> > > >> > > > wrote:
> > > >> > > > > See docs/specs/vmcoreinfo.txt for details.
> > > >> > > > >
> > > >> > > > > "etc/vmcoreinfo" fw_cfg entry is added when using "-device 
> > > >> > > > > vmcoreinfo".
> > > >> > > >
> > > >> > > > I'm wondering if you considered just adding the entry to fw_cfg 
> > > >> > > > by
> > > >> > > > default, without requiring any -device arg ? Unless I'm 
> > > >> > > > misunderstanding,
> > > >> > > > this doesn't feel like a device to me - its just a well known 
> > > >> > > > bucket
> > > >> > > > in fw_cfg IIUC ?  Obviously its existance would need to be tied 
> > > >> > > > to
> > > >> > > > the latest machine type for ABI reasons though. The benefit of 
> > > >> > > > this
> > > >> > > > is that it would "just work" without us having to plumb it 
> > > >> > > > through to
> > > >> > > > all the downstream applications that use QEMU for mgmt guest 
> > > >> > > > (OpenStack,
> > > >> > > > oVirt, GNOME Boxes, virt-manager, and countless other mgmt apps).
> > > >> > > it follows model set by pvpanic device, it's easier to manage from 
> > > >> > > migration
> > > >> > > POV, one could use it even for old machine types with new qemu 
> > > >> > > (just by adding
> > > >> > > device, it makes instance not backwards migratable to old qemu but 
> > > >> > > should work
> > > >> > > for forward migration) and if user doesn't need it, device could 
> > > >> > > be just omitted
> > > >> > > from CLI.
> > > >> >
> > > >> > Sure but it means that in effect no one will have this functionality 
> > > >> > enabled
> > > >> > for several years. pvpanic has been around a long time and I rarely 
> > > >> > see it
> > > >> > present in configured guests :-(
> > > >> >
> > > >> >
> > > >> > Regards,
> > > >> > Daniel
> > > >>
> > > >> libvirt runs with -nodefaults, right? I'd argue pretty strongly 
> > > >> -nodefaults
> > > >> shouldn't add optional devices anyway.
> > > >
> > > > This isn't really adding a device though is it - it is just a well known
> > > > location in fw_cfg to receive data.
> > > 
> > > Enabling the device on some configurations by default can be done as a
> > > follow-up patch. Can we get this series reviewed & merged?
> > 
> > The problem with the -device approach + turning it on by default is that 
> > there
> > is no way to turn it off again if you don't want it. eg there's way to undo
> > an implicit '-device foo' except via -nodefaults, but since libvirt uses 
> > that
> > already it would negate the effect of enabling it by default 
> > unconditionally.
> 
> It's still possible to add a -machine option that can
> enable/disable automatic creation of the device.
> 
> But I also don't see why it needs to be implemented using -device
> if it's not really a device.  A boolean machine or fw_cfg
> property is good enough for that.

It certainly feels like a device. It has state
(that needs to be migrated), it has a host/guest interface.


> > 
> > Your previous approach of "-global fw_cfg.vmcoreinfo=on" is nicer in this
> > respect, as you can trivially turn it on/off, overriding the default state
> > in both directions.
> 
> Both "-global fw_cfg.vmcoreinfo=on|off" and
> "-machine vmcoreinfo=on|off" sound good enough to me.


Certainly not a fw cfg flag. Can be a machine flag I guess
but then we'd have to open-code each such device.
And don't forget auto - this is what Daniel asks for.

I don't necessarily see this device as so special and think a generic
interface to control what goes into the machine would be better (e.g.
look how you use hacky -global to control fw cfg options, it only works
if there's a single one), but if everyone thinks otherwise and agrees we
should have it in there by default, and a property to disable, fine.

Can be a patch on top though.


> -- 
> Eduardo