11:02 a.m.
On Tue, Dec 02, 2014 at 12:08:30PM -0500, Laine Stump wrote:
(Everyone - see the request for opinions/ideas towards the bottom)
The idea behind these patches is the following:
1) most virtual machines only have a single MAC address behind each
interface, and that MAC address is known by libvirt.
2) If we (i.e. libvirt) manually add an entry to the bridge's
forwarding database (fdb) for the MAC address associated with a port
on the bridge, we can turn off learning and unicast_flooding for that
port.
3) kernels starting with 3.15 (and actually working correctly starting
in kernel 3.17) will notice that all of a bridge's ports have flood
and learning turned off, and in that case will turn off promiscuous
mode on all ports. If all but one of the ports have flood/learning
turned off, then promiscuous will be turned off on that port (and left
on for all the other ports)
4) When (4) can be done, there is a measurable performance
advantage. It can also *kind of* help security, as it will prevent a
guest from doing anything useful if it changes its MAC address (but
won't prevent the guest from *sending* packets with a spoofed MAC
address).
NB: These only work with a fixed MAC address, and no vlan tags set in
the guest. Support for both of those will be coming.
So just to be clear on the functional differences between the
impls
With the current impl
1. The bridge has a table mapping MAC <-> TAP devices
which is initially empty
2. Whenever a TAP device sends a packet, it causes an
entry to be added to the MAC <-> TAP table
3. When the bridge needs to forward a packet with a MAC
- If MAC is unknown, it is sent to all TAP devices
- Else it is send to the specific TAP device
4. The physical device is always promiscuous to see all
traffic on the LAN
An result of point 2, is that if the guest changes its
MAC address on the fly, it merely needs to send a packet
and the bridge will learn its new MAC address. That said
our firewall rules can potentially block this dynamic
change.
With the new impl
1. The bridge has a table mapping MAC <-> TAP devices
which is initially empty
2. Libvirt tells the bridge what MAC address is associated
with each TAP device added
3. The bridge always forwards the packets to the correct
TAP device
4. The physical device is never promiscuous, it is
programmed with MACs of all TAP devices
In both cases the guest can send packets with spoofed
MAC addresses. In the old code these packets would make
it onto the LAN and it can receive replies back. In the
new code these packets would make it onto the LAN but
it will never receive replies back.
At a high level the conceptual difference is whether the table
of MAC addresses & TAP devices is statically defined by libvirt
or dynamically defined on the fly.
HERE IS THE REQUEST FOR OPINIONS/IDEAS:
This V2 of the patchset addresses several issues brought up by jferlan
on the original series, and changes the name of the attribute from:
promiscLinks='yes|no'
to
fdb='learningWithFlood|managed'
I'm somewhat more happy with this new naming than the previous but
still looking for better ideas. It is closer to describing what the
new code really does, but "learningWithFlood" seems a bit long and
awkward, while I have been told that "fdb" is too short and
unrecognizeable (I will point out that 1) "fdb" is the same name used
by iproute2's "bridge" command, and 2) another bridge option,
"stp" is
also a three letter acronym that will only be recognized by those
familiar with configuring an L2 bridge device or watching NASCAR on
Saturday afternoons (or whenever it's on - not a fan myself :-))
One thing to remember is that libvirt is supposed to be providing a
higher level configuration language that is independant of specific
implementations. 'fdb' is terminology specific to the Linux
bridge implementation, so I don't think that's appropriate to use
in the XML configuration.
Here is a full list of every idea that either I or someone else has
come up with since I started thinking about this:
promiscLinks='yes|no'
After initially going with this for the v1 of the patchset, I later
decided against it, because it doesn't describe what libvirt is
doing, but only a *possible* side effect on *some* of the ports
connected to the bridge (in practice, it only happens to the physdev
port).
Agreed, promiscLinks seems to be referring to only one small part
of the solution.
fdb='auto|managed'
I like "fdb" as the name of the attribute, because I think it really
gets at what libvirt is doing - it is taking over management of the
bridge's fdb (Forwarding Database), which ends up providing better
performance in several ways.
The problem with this proposal is that the two values are kind of
ambiguous - it's not clear which one is using the bridge module's
built-in management (I had figured this would be "auto"), and which
is telling libvirt to manage it ("managed"). (on the other hand,
the first option is "ignore the issue, let the underlying system
handle it", vs. "libvirt should manage it", so maybe it *is* a
reasonable choice).
Also, see the comment above about the perceived terseness and obscurity
of "fdb".
fdb='learningWithFlood|managed'
This alternate name was suggested by Michael Tsirkin as a way to
unambiguously indicate what was being done in the mode where libvirt
isn't involved in the fdb management. There was some criticism on
IRC that the name is *too* verbose, especially when contrasted with
"fdb".
fdb='learning|managed'
A suggested shortening of "learningWithFlood".
forwardingDatabase='blah'
A way to get around criticism of "fdb". I think this is too verbose,
but maybe I'm biased :-)
I don't like any of these really because they all refer to a specific
implementation choice.
[specify each minor item that separately]
In order to manage the fdb by itself, libvirt disables "learning"
and "unicast_flood" for each tap device, enables "vlan_filtering"
for the bridge itself, then adds an fdb entry for each tap to the
bridge. There was one suggestion that, rather than trying to come
up with a single option that says "do all of these things", we
should instead make each of them separately configured. The problem
with this is that it makes it too easy to screw up the
configuration such that it causes sub-par performance, or simply
doesn't work at all. Part of libvirt's job is making it difficult
to screw up (or at least easier to succeed); for example, libvirt's
virtual networks do a lot of things automatically - create a
bridge, add iptables rules for filtering and NAT, run an instance
of dnsmasq - over time we've offered the option of tweaking more
and more of the details of this setup, but the initial aim was to
provide something that worked with as few required (or even
permitted) tweaks as possible.
I guess what I'm getting at is that I think it would be a mistake
to require turning on several different knobs (which individually
make little/no sense) in order to get the bridge into this higher
performing mode.
Agreed, this is even worse because it is exposing many internal
implementation details.
So - does anyone have an opinion of any of the options offered above,
or any ideas for alternates?
In the meantime, note that while the default is currently
"learningWithFlood" (meaning that that name is never actually directly
used/required anywhere, but is just in the RNG and the enum
definition), the intent of the people who developed this functionality
in the kernel is that eventually it will work so well that libvirt
management of the fdb can silently become the default with no visible
change in behavior.
There is a user visible change in behaviour because this new solution,
as implemented in this patch series, will no longer allow a guest to
change its MAC address on the fly.
Of course if QEMU can provide a notification to libvirt when the guest
changes its MAC address, then libvirt can update the MAC table and so
re-gain functional equivalance with the old solution. I think we would
want to be able to turn that on or off though.
This all leads me to suggest that we should use the following syntax
that is indepedant of the particular implementation choice, and instead
represents the logical behaviour of the feature.
macTable=static|dynamic
The mode of 'static' means that the MAC address table will always
match MACs recorded in libvirt guest XML. This is the new mode
that your patches implement
The mode of 'dynamic' means that the MAC address table will be
able to dynamically update itself as the guest changes MAC addresses.
In the near term, this will use the traditional bridge learning
mode. In the long term, if we can get MAC change notifications from
QEMU, we can switch this over to use the new bridge features.
Regards,
Daniel
--
|: http://berrange.com -o- http://www.flickr.com/photos/dberrange/ :|
|: http://libvirt.org -o- http://virt-manager.org :|
|: http://autobuild.org -o- http://search.cpan.org/~danberr/ :|
|: http://entangle-photo.org -o- http://live.gnome.org/gtk-vnc :|