Daniel Veillard wrote:
Yeah my first reaction was to provide XML based configuration files
that libvir would be able to understand, but I really think the configurations
may come from very varied sources possibly databases. That's why I would
rather have the library agnostic when it comes to what a configuration
may look like.
I'll have to think about this one. My biggest fear is that if there are
a bunch of management tools capable of creating domains you'll end up
with a bunch of different places storing configs.
If you think about a typical use-case (or at least, my typical use-case
:-)), a user is likely to have a larger number of domains than what are
running (say a domain for a bunch of different distros for testing). It
would be nice if any management tools could see those domains and easily
start them up.
What is needed to start a domain ?
Required
1) name of domain
2) initial memory
3) maximum memory
Optional
4) uuid
5) ssidref
6) poweroff, reboot, crash actions
7) number of vcpus
8) initial vcpu => cpu mapping
Builder Specific Info
Linux (misnomer, this is really any PV domain)
Required
1) kernel
Optional
2) ramdisk
3) root device
4) extra (additional kernel args)
Vmx (another misnomer, really any FV domain)
Required
1) device model path
Optional
2) device model arguments
3) DISPLAY
Device Specific Info
VBD
1) backend ID
2) uname (file:/path/to/file)
3) dev (hda1)
4) mode (r, w, w!)
VIF
Required
1) backend ID
2) MAC
Optional
3) bridge
4) ip
5) script
6) vifname
PIO Range
1) starting address
2) enabling address
Of course, this is subject to change as new features are added but I'm
pretty sure this is pretty complete.
For Xen we know what is needed, at
least at the moment (paths to kernels, images, block devices paths, network
devices and I/Os), but this may change slightly with full virtualization,
and other virtualization engine like QEmu have different ways since they
emulate the bootstrap process too.
Fortunately, since FV uses the qemu device model, we could launch qemu
pretty easily reusing this existing config format (since qemu already is
launched anyway). However, IMHO, the VBD description Xen uses sucks so
I'm not sure we want to adopt it as the universal model.
To me we don't have a good answer to this question so that's one more reason
to keep this as opaque as possible.
opaqueness means easy backwards and forward compatibility so I'm
definitely a fan :-)
I think we could achieve this with the following requirements:
1) Allow domain creation based on an opaque configuration object
2) Allow storing and retrieval of these configuration objects from the
XenStore (using a user-specified path)
I think this isolates the state well enough that it leaves the majority
of the library stateless. Just some random thoughts...
Hum ... you will still need to put some decoding of this object inside
the library to pass enough information to xend to start it (since I believe
that we will need to use the xend API to create domains). This doesn't make
the code agnostic to the format but it makes the API agnostic which is
important too (though the versionning as xen evolves sounds a bit hellish,
we'd better carefully design the versioning in that format :-)
I definitely think it might be worthwhile to design a Xen configuration
schema that we could keep consistent as things move forward. Some
things would obviously map 1-1 like name but other things might require
careful thinking (like block devices).
Thoughts?
Regards,
Anthony Liguori
Daniel