On 6/9/20 12:27 PM, Daniel P. Berrangé wrote:
On Mon, Jun 08, 2020 at 11:05:00PM -0400, Laine Stump wrote:
On 6/8/20 10:51 AM, Daniel P. Berrangé wrote:
The virtual network has never supported NAT with IPv6 since this feature
didn't exist at the time. NAT has been available since RHEL-7 vintage
though, and it is desirable to be able to use it.
This series enables it with
<forward mode=3D"nat">
<nat ipv6=3D"yes"/>
</forward>
I've had this lurking on my "this is something I should do" list for a long
time, but couldn't decide on the best name in XML (and also figured that the
problem with accept_ra needed to be fixed first), so it never got to the
top. So I'm glad to see you've done it, disappointed in myself that I never
did it :-/
I like your XML knob naming better than what I'd considered. I had thought
of having <forward mode='supernat'> (or some other more reasonable extra
mode), but your proposal is more orthogonal and matches with the existing
ipv6='yes' at the toplevel of <network> (which is used to enable ipv6
traffic between guests on the bridge even when there are no IPv6 addresses
configured for the network.)
I considered mode="nat6" as an alternative, but it would have meant
updating many switch() statements, and is a somewhat misleading as a
name.
Yeah, the way you have it is much better.
</network>
Conceptually this means
- Try to gimme a subnet with IPv4 and DHCP
- Try to gimme a subnet with IPv6 and RAs
Now when we start the virtual network
- If IPv4 is not enabled on host, don't assign addr
What will we use to check for this? Not just "no IP addresses configured", I
guess, since it may be the case that libvirt has just happened to come up
before NM or whoever has started any networks. (or maybe someone wants to
use IPv6 on a libvirt virtual network, but have no IPv6 connectivity beyond
the host).
IIUC, we can simply check whether it is possible to create a socket
with AF_INET or AF_INET6. If the kernel supports it, then this
should suceed, even if network manager isn't running yet.
- Else
- Iterate N=3D1..254 to find a free range for IPv4
- Use 192.168.N.0/24 for subnet
- Use 192.168.N.1 for host IP
- Use 192.168.N.2 -> 192.168.N.254 for guest DHCP
- If IPv6 is not enabled on host, don't assign addr
- Else
- Generate NNNN:NNNN as 4 random bytes
- Use fd00:add:f00d:NNNN:NNNN::0/64 for IPv6 subnet
- Use fd00:add:f00d:NNNN:NNNN::1 for host IP
- Use route advertizement for IPv6 zero-conf
With NNNN:NNNN, even with 1000 guests running, we have just a 0.02%
chance of clashing with a guest for IPv6.
The "live" XML would always reflect the currently assigned addresses
Proactively monitor the address allocations of the host. If we see
a conflicting address appear, take down the dnsmasq intance, generate
a new subnet, bring dnsmasq back online.
Hmm. How would you see this monitoring happening? We couldn't do it with an
external script like I had done for simple "shut down on conflict" without
adding extra functionality to libvirt's network driver. We *could* go back
to the idea of monitoring netlink change messages ourselves within libvirtd
and doing it all internally ourselves. Or maybe the NM script I proposed
could go beyond simply destroying conflicting networks, and also restart any
network that had autoaddr='yes'; to make this fully functional we would need
to finally put in the proper stuff so that tap devices (and the underlying
emulated NICs) would be set offline when their connected network was
destroyed, and then reconnected/set online when the network was re-started.
Getting the networks to behave this way would be useful in general anyway,
even without thinking about the conflicting-networks problem. The one
downside of externally controlling renumbering-on-conflict using an external
script is that it would only work with NetworkManager...
Yeah, I'm trying to remember now why we went the NM hook route, rather
than listening for netlink events. I guess NM is much simpler to hook
into.
You mentioned the NM hook idea (I hadn't even known they had such hooks)
when we were talking about libvirtd using activation sockets (unrelated
to network conflict stuff), and I said that needing to have a thread
always listening for netlink change messages was in conflict with the
idea of having libvirtd use activation sockets and exit after a timeout.
Once we decide that we're responsible for watching all netlink change
messages, we have to have a process that's always running (also we would
have to read the entire state of all interfaces each time that process
was started, and respond to any changes that had taken place since the
last time the process exited).
(Probably a NM hook script is also much easier than monitoring netlink
message, since there's no need to figure out which field in which type
of messages to look at, and you don't have to setup the thread to listen
for it, etc. Just a simple python program that gets called with all
relevant info in text format either in argv, or in the environment.)
I'd honestly not thought about this too much though - just having
an automatically numbered network will already be a huge step forward
compared to current day.
True. First step is having it auto-numbered at startup. 2nd step is
making it auto-renumber when required while running. Either step 1.5 or
step 3 would be saving the auto-number decision to use as first try the
next time it's started.
In particular if we insituted a rule that if we are NOT on a hypervisor,
we count from N=254 -> 0, when picking 192.168.N.0, and count from
N=0 -> 254 when we are on a hypervisor, then we'll trivially avoid the
host/guest clash in simple case, even if network is not yet online.
Don't anyone dare mention nested virt with 3 levels of libvirt...
Seriously though, even without automatic teardown & restart, we'd
be way better off by simply not hardcoding 192.168.N.0 at RPM
install time when the network env is not the same as the run time
network env. eg cloud images
Yep. Especially when you consider that (e.g.) even installing Fedora to
disk from a live image doesn't involve running any of the rpm
postinstall scripts - it just copies the complete image over to the
destination from the live image; so the environment encountered when the
final OS boots is two "environment steps" removed from the environment
that was in effect at the time the rpm postinstall ran.
Ideally we would have to bring the guest network links offline and
then online again to force DHCP re-assignment immediately.
Yeah, I think it really makes sense that when a libvirt network is
destroyed, all the tap devices are set offline, and the emulated NICs are
set offline as well; then when a libvirt network is started, we would go
through all devices that are supposed to be connected to that network,
reconnect the taps, set them online, and set the emulated NIC online. We
currently do the reconnection part when libvirtd is restarted but can't do
it immediately when a *network* is restarted because the network driver has
no access to the list of active guests and their interfaces....
Hmm, we do now maintain the list of ports for each network though, and it
would be possible to expand that to keep the name of the tap device
associated with the port in addition to the other info (e.g. whether or not
the NIC has been set offline via an API call), *but* when a network is
destroyed, all ports registered with that network are also destroyed, so
just expanding the attributes for the ports isn't going to get us where we
need. So, do we want to 1) change it to maintain active ports for a network
when it is destroyed so that they can be easily reactivated when the network
is restarted? Or do we want to 2) change the network driver to make calls to
all registered hypervisor drivers during a net-start to look for all guest
interfaces that think they are connected to the network? The former sounds
much more efficient, but I don't know how "dirty" it seems to maintain state
for something that has been "destroyed"...
Or maybe we instead need to also add a new API for networks
virNetworkReconnect(), which will use newly expanded info in the network
ports list to reconnect all guest interfaces.
Responsibility for enslaving a TAP device into a bridge still lives with
the virt drivers, not the network driver.
The virt drivers could listen for lifecycle events from the network driver
and auto-reconnect.
Yeah, that could work too. *something* so that an entity that has the
proper info about tap devices finds out that the tap device needs to be
re-attached. And since currently all the code from
creating/destroying/attaching tap devices lives in the various
hypervisors, I guess this makes the most sense.
Alternatively the virt driver could listen for netlink events and see the
virbr0 being deleted, and created by the kernel.
On one hand, it seems cleaner to do it by watching for lifecycle events
from the network driver. On the other hand, that wouldn't properly
handle cases where someone was "secretly" attaching to a libvirt-created
bridge with <interface type='bridge'> (e.g. session-mode guests that are
taking advantage of qemu-bridge-helper) or guests that attach to a
non-libvirt-managed bridge that happens to disappear and reappear.
On a different sub-topic - it would be nice to provide some stability to the
subnet used for an autoaddr='yes' network (think of the case where every
time a host is booted, libvirt starts its default network when
192.168.122.0/24 is available, but then a short time later a host interface
is always started on the same subnet - that would mean every time the host
booted the exact same destabilizing dance would take place even though it
would be pretty easy to predict the eventually-used subnet based on past
experience).
Although we historically have avoided automatic changes to libvirt config
files by libvirtd itself as much as possible (the only cases I can think of
are when we're modifying the config to take care of some compatibility
problem after an upgrade), what do you think about having the autoaddr='yes'
networks automatically update the config with the current subnet info?
(maybe this would need to only be done if not starting from a live image or
something, or maybe it should just always be done). This would then be used
as the first guess the next time the network was started. That way we would
avoid the need to delay starting libvirt networks until after host
networking was fully up; the subnet might bounce around a bit that first
time, but once a stable address was found during that first run, it would
then be used from the get-go during all subsequent boots (until/unless
something changed and it had to be changed yet again).
We could stash the previously chosen subnet in /var/cache/libvirt/network
or /var/lib/libvirt/network, no need to modify the inactive XML config.
This is like how dnsmasq "remembers" DHCP leases previously given for guests.
Yeah, even better.
I think /var/cache/libvirt/network sounds more appropriate. What format
do you think it should be stored in? We have have some stuff like that
stored as XML (in particular, qemu capabilities cache is in /var/cache
and stored as XML), and at least one thing stored in /var/run/libvirt as
JSON (saved MAC address/vlan tag of VFs saved when a VF is assigned to a
guest, so that it can be restored later). It was actually *me* who added
that latter - the file originally just held the plain raw MAC address
(in ASCII), but we needed to expand it; I was originally going to put it
there in XML, but there was *something* else going on at the time that
convinced me to put it in JSON instead (some other new file was being
stored in JSON, or ??? I seriously don't remember the details, just that
there was something that influenced that decision).
(BTW, I notice that the specfile creates /var/lib/libvirt/network during
install/upgrade, but it's empty, and I don't see any place in the code
that references it (we do put stuff in /var/run/libvirt/network, but
that's a different beast). Is that just an artifact of when we were
incorrectly putting network status in /var/lib/libvirt instead of
/var/run/libvirt, and we forgot to remove the line in the specfile when
that was fixed? Or did my cursory audit of the code miss something?)
