4:08 a.m.
On Fri, Jan 05, 2007 at 09:14:30PM +0000, Daniel P. Berrange wrote:
The following series of (2) patches adds a QEMU driver to libvirt. The first patch
provides a daemon for managing QEMU instances, the second provides a driver letting
libvirt manage QEMU via the daemon.
Basic architecture
------------------
The reason for the daemon architecture is two fold:
- At this time, there is no (practical) way to enumerate QEMU instances, or
reliably connect to the monitor console of an existing process. There is
also no way to determine the guest configuration associated with a daemon.
Okay, we admitted that principle in the first round of QEmu patches last
year. The only question I have is about the multiplication of running daemons
for libvirt, as we also have another one already for read only xen hypervisor
access. We could either decide to keep daemon usages very specific (allows
to also easilly restrict their priviledge) or try to unify them. I guess
from a security POV it's still better to keep them separate, and anyway
they are relatively unlikely to be run at the same time (KVM and Xen
on the same Node).
- It is desirable to be able to manage QEMU instances using either
an unprivilegd
local client, or a remote client. The daemon can provide connectivity via UNIX
domain sockets, or IPv4 / IPv6 and layer in suitable authentication / encryption
via TLS and/or SASL protocols.
C.f. my previous mail, yes authentication is key. Could you elaborate in some
way how the remote access and the authentication is set up, see my previous mail
on the remote xend access, we should try to unify and set up a specific
page to document remote accesses.
Anthony Ligouri is working on patches for QEMU with the goal of
addressing the
first point. For example, an extra command line argument will cause QEMU to save
a PID file and create a UNIX socket for its monitor at a well-defined path. More
functionality in the monitor console will allow the guest configuration to be
reverse engineered from a running guest. Even with those patches, however, it will
still be desirable to have a daemon to provide more flexible connectivity, and to
facilitate implementation libvirt APIs which are host (rather than guest) related.
Thus I expect that over time we can simply enhance the daemon to take advantage of
newer capabilities in the QEMU monitor, but keep the same basic libvirt driver
architecture.
Okay, Work in Progress.
Considering some of the other hypervisor technologies out there, in
particular
User Mode Linux, and lhype, it may well become possible to let this QEMU daemon
also provide the management of these guests - allowing re-use of the single driver
backend in the libvirt client library itself.
Which reopen the question, one multi-featured daemon or multiple simpler
(but possibly redundant) daemons ?
XML format
----------
As discussed in the previous mail thread, the XML format for describing guests
with the QEMU backend is the same structure as that for Xen guests, with
following enhancements:
- The 'type' attribute on the top level <domain> tag can take one of the
values 'qemu', 'kqemu' or 'kvm' instead of 'xen'.
This selects between
the different virtualization approaches QEMU can provide.
- The '<type>' attribute within the <os> block of the XML (for now)
is
still expected to the 'hvm' (indicating full virtualization), although
I'm trying to think of a better name, since its not technically hardware
accelerated unless you're using KVM
yeah I don't have a good value to suggest except "unknown" bacause
basically
we don't know a priori what the running OS will be.
- The '<type>' attribute within the <os> block
of the XML can have two
optional 'arch' and 'machine' attributes. The former selects the CPU
architecture to be emulated; the latter the specific machine to have
QEMU emulate (determine those supported by QEMU using 'qemu -M ?').
Okay, I hope we will have enough flexibility in the virNodeInfo model to
express the various combinations, we have a 32 char string for this, I guess
that should be sufficient, but I don't know how to express that in the best way
I will see how the pacth does it. From my recollection of posts on qemu-devel
some of the machines names can be a bit long on specific emulated target. At
least we should be okay for a PC architecture.
- The <kernel>, <initrd>, <cmdline> elements can
be used to specify
an explicit kernel to boot off[1], otherwise it'll do a boot of the
cdrom, harddisk / floppy (based on <boot> element). Well,the kernel
bits are parsed at least. I've not got around to using them when
building the QEMU argv yet.
Okay
- The disk devices are configured in same way as Xen HVM guests. eg
you
have to use hda -> hdd, and/or fda -> fdb. Only hdc can be selected
as a cdrom device.
Good !
- The network configuration is work in progress. QEMU has many ways
to
setup networking. I use the 'type' attribute to select between the
different approachs 'user', 'tap', 'server',
'client', 'mcast' mapping
them directly onto QEMU command line arguments. You can specify a
MAC address as usual too. I need to implement auto-generation of MAC
addresses if omitted. Most of them have extra bits of metadata though
which I've not figured out appropriate XML for yet. Thus when building
the QEMU argv I currently just hardcode 'user' networking.
Okay, since user is the default in QEmu (assuming I remember correctly :-)
- The QEMU binary is determined automatically based on the
requested
CPU architecture, defaulting to i686 if non specified. It is possible
to override the default binary using the <emulator> element within the
<devices> section. This is different to previously discussed, because
recent work by Anthony merging VMI + KVM to give paravirt guests means
that the <loader> element is best kept to refer to the VMI ROM (or other
ROM like files :-) - this is also closer to Xen semantics anyway.
Hum, the ROM, one more parameter, actually we may once need to provide for
mutiple of them at some point if they start mapping non contiguous area.
Connectivity
------------
The namespace under which all connection URIs come is 'qemud'. Thereafter
there are several options. First, two well-known local hypervisor
connections
- qemud:///session
This is a per-user private hypervisor connection. The libvirt daemon and
qemu guest processes just run as whatever UNIX user your client app is
running. This lets unprivileged users use the qemu driver without needing
any kind admin rights. Obviously you can't use KQEMU or KVM accelerators
unless the /dev/ device node is chmod/chown'd to give you access.
The communication goes over a UNIX domain socket which is mode 0600 created
in the abstract namespace at $HOME/.qemud.d/sock.
okay, makes sense. Everything runs under the user privilege and there is no
escalation.
- qemud:///system
This is a system-wide privileged hypervisor connection. There is only one
of these on any given machine. The libvirt_qemud daemon would be started
ahead of time (by an init script), possibly running as root, or maybe under
a dedicated system user account (and the KQEMU/KVM devices chown'd to match).
Would that be hard to allow autostart ? That's what we do for the read-only
xen hypervisor access. Avoiding starting up stuff in init.d when we have no
garantee it will be used, and auto-shutdown when there is no client is IMHO
generally nicer, but that feature can just be added later possibly, main
drawback is that it requires an suid binary.
The admin would optionally also make it listen on IPv4/6 addrs to
allow
remote communication. (see next URI example)
The local communication goes over one of two possible UNIX domain sockets
Both in the abstract namespace under the directory /var/run. The first socket
called 'qemud' is mode 0600, so only privileged apps (ie root) can access
it,
and gives full control capabilities. The other called 'qemud-ro' is mode
0666
and any clients connecting to it will be restricted to only read-only libvirt
operations by the server.
- qemud://hostname:port/
This lets you connect to a daemon over IPv4 or IPv6. If omitted the port is
8123 (will probably change it). This lets you connect to a system daemon
on a remote host - assuming it was configured to listen on IPv4/6 interfaces.
Hum, for that the daemon requires to be started statically too.
Currently there is zero auth or encryption, but I'm planning
to make it
mandortory to use the TLS protocol - using the GNU TLS library. This will give
encryption, and mutual authentication using either x509 certificates or
PGP keys & trustdbs or perhaps both :-) Will probably start off by implementing
PGP since I understand it better.
So if you wanted to remotely manage a server, you'd copy the server's
certificate/public key to the client into a well known location. Similarly
you'd generate a keypair for the client & copy its public key to the
server. Perhaps I'll allow clients without a key to connect in read-only
mode. Need to prototype it first and then write up some ideas.
okay, though there is multiple authentication and encryption libraries,
and picking the Right One may not be possible, there is so many options,
and people may have specific infrastructure in place. Anyway the current state
is no-auth so anything will be better :-)
Server architecture
-------------------
The server is a fairly simple beast. It is single-threaded using non-blocking I/O
and poll() for all operations. It will listen on multiple sockets for incoming
connections. The protocol used for client-server comms is a very simple binary
message format close to the existing libvirt_proxy.
Good so we keep similar implementations. Any possibility of sharing part of
that code, it's always very sensitive areas, both for security and edge case
in the communication.
Client sends a message, server
receives it, performs appropriate operation & sends a reply to the client. The
client (ie libvirt driver) blocks after sending its message until it gets a reply.
The server does non-blocking reads from the client buffering until it has a single
complete message, then processes it and populates the buffer with a reply and does
non-blocking writes to send it back to the client. It won't try to read a further
message from the client until its sent the entire reply back. ie, it is a totally
synchronous message flow - no batching/pipelining of messages.
Honnestly I think that's good enough, I don't see hundreds of QEmu instances
having to be monitored remotely from a single Node. On an monitoring machine
things may be accelerated by multithreading the gathering process to talk to
multiple Nodes in parallel. At least on the server side I prefer to keep things
as straight as possible.
During the time
the server is processes a message it is not dealing with any other I/O, but thus
far all the operations are very fast to implement, so this isn't a serious issue,
and there ways to deal with it if there are operations which turn out to take a
long time. I certainly want to avoid multi-threading in the server at all costs!
completely agree :-)
As well as monitoring the client & client sockets, the poll()
event loop in the
server also captures stdout & stderr from the QEMU processes. Currently we just
dump this to stdout of the daemon, but I expect we can log it somewhere. When we
start accessing the QEMU monitor there will be another fd in the event loop - ie
the pseduo-TTY (or UNIX socket) on which we talk to the monitor.
At some point we will need to look at adding a Console dump API, that will
be doable for Xen too, but it's not urgent since nobody requested it yet :-)
Inactive guests
---------------
Guests created using 'virsh create' (or equiv API) are treated as
'transient'
domains - ie their config files are not saved to disk. This is consistent with
the behaviour in the Xen backend. Guests created using 'virsh define', however,
are saved out to disk in $HOME/.qemud.d for the per-user session daemon. The
system-wide daemon should use /etc/qemud.d, but currently its still /root/.qemud.d
Maybe this should be asked on the qemu-devel list, Fabrice and Co. may have
a preference on where to store config related stuff for QEmu even if it's not
directly part of QEmu.
The config files are simply saved as the libvirt XML blob ensuring
no data
conversion issues. In any case, QEMU doesn't currently have any config file
format we can leverage. The list of inactive guests is loaded at startup of the
daemon. New config files are expected to be created via the API - files manually
created in the directory after initial startup are not seen. Might like to change
this later.
Hum, maybe we could use FAM/gamin if found at configure time, but well
it's just an additional feature, let's just avoid any uneeeded timer.
XML Examples
------------
This is a guest using plain qemu, with x86_64 architecture and a ISA-only
(ie no PCI) machine emulation. I was actually running this on a 32-bit
host :-) VNC is configured to run on port 5906. QEMU can't automatically
choose a VNC port, so if one isn't specified we assign one based on the
domain ID. This should be fixed in QEMU....
<domain type='qemu'>
<name>demo1</name>
<uuid>4dea23b3-1d52-d8f3-2516-782e98a23fa0</uuid>
<memory>131072</memory>
<vcpu>1</vcpu>
<os>
<type arch='x86_64' machine='isapc'>hvm</type>
</os>
<devices>
<disk type='file' device='disk'>
<source file='/home/berrange/fedora/diskboot.img'/>
<target dev='hda'/>
</disk>
<interface type='user'>
<mac address='24:42:53:21:52:45'/>
</interface>
<graphics type='vnc' port='5906'/>
</devices>
</domain>
A second example, this time using KVM acceleration. Note how I specify a
non-default path to QEMU to pick up the KVM build of QEMU. Normally KVM
binary will default to /usr/bin/qemu-kvm - this may change depending on
how distro packaging of KVM turns out - it may even be merged into regular
QEMU binaries.
<domain type='kvm'>
<name>demo2</name>
<uuid>4dea24b3-1d52-d8f3-2516-782e98a23fa0</uuid>
<memory>131072</memory>
<vcpu>1</vcpu>
<os>
<type>hvm</type>
</os>
<devices>
<emulator>/home/berrange/usr/kvm-devel/bin/qemu-system-x86_64</emulator>
<disk type='file' device='disk'>
<source file='/home/berrange/fedora/diskboot.img'/>
<target dev='hda'/>
</disk>
<interface type='user'>
<mac address='24:42:53:21:52:45'/>
</interface>
<graphics type='vnc' port='-1'/>
</devices>
</domain>
Okay, I'm nearing completion of a Relax-NG schemas allowing to validate
XML instances, I will augment to allow the changes, but based on last week
discussion it should not bee too hard and still retain good validation
properties.
Outstanding work
----------------
- TLS support. Need to add TLS encryption & authentication to both the client
and server side for IPv4/6 communications. This will obviously add a dependancy
on libgnutls.so in libvirt & the daemon. I don't consider this a major
problem
since every non-trivial network app these days uses TLS. The other possible impl
of OpenSSL has GPL-compatability issues, so is not considered.
- Change the wire format to use fixed size data types (ie, int8, int16, int32, etc)
instead of the size-dependant int/long types. At same time define some rules for
the byte ordering. Client must match server ordering ? Server must accept
client's
desired ordering ? Everyone must use BE regardless of server/client format ?
I'm
inclined to say client must match server, since it distributes the byte-swapping
overhead to all clients and lets the common case of x86->x86 be a no-op.
Hum, on the other hand if you do the conversion as suggested by IETF rules
it's easier to find the places where the conversion is missing, unless you
forgot to ntoh and hton on both client and server code. Honnestly I would
not take the performance hit in consideration at that level and not now,
the RPC is gonna totally dominate it by order of magnitudes in my opinion.
- Add a protocol version message as first option to let use
protocol at will later
while maintaining compat with older libvirt client libraries.
Yeah, this also ensure you get a functionning server on that port !
- Improve support for describing the various QEMU network
configurations
- Finish boot options - boot device order & explicit kernel
- Open & use connection to QEMU monitor which will let us implement pause/resume,
suspend/restore drivers, and device hotplug / media changes.
- Return sensible data for virNodeInfo - will need to have operating system dependant
code here - parsing /proc for Linux to determine available RAM & CPU speed. Who
knows what for Solaris / BSD ?!? Anyone know of remotely standard ways for doing
this. Accurate host memory reporting is the only really critical data item we need.
The GNOME guys tried that, maybe dig up the gst (gnome system tools) code
base :-)
- There is a fair bit of duplicate in various helper functions
between the daemon,
and various libvirt driver backends. We should probably pull this stuff out into
a separate lib/ directoy, build it into a static library and then link that into
both libvirt, virsh & the qemud daemon as needed.
Yes definitely !
This all sounds excellent, thanks a lot !!!
Daniel
--
Red Hat Virtualization group http://redhat.com/virtualization/
Daniel Veillard | virtualization library http://libvirt.org/
veillard(a)redhat.com | libxml GNOME XML XSLT toolkit http://xmlsoft.org/
http://veillard.com/ | Rpmfind RPM search engine http://rpmfind.net/