With the new Discovery model, join can happen even before serf is
initliazed. It could also happen due to misconfiguration of
--cluster-advertise. The local endpoint join must succeed and later when
the serf initializes and joins the cluster, it will push the local db to
the cluster.
Signed-off-by: Madhu Venugopal <madhu@docker.com>
With the recently introduced docker discovery, the self node discovery
notification can reach the overlay driver after the remote node
discovery notification. In scenarios such as 2 node setup, it seems more
likely. In those scenarios, the serfJoin is not triggered and hence the
neighborship is not formed between the 2 nodes.
The fix is to retain the knowledge of the neighbor and reuse it
immediately after the serfInit is done. Since we do the serfJoin just
once, there is no harm in changing the neighIP to a new value even if it
is not used.
Signed-off-by: Madhu Venugopal <madhu@docker.com>
* integrated hostdiscovery package with the new Docker Discovery
* Integrated hostdiscovery package with libnetwork core
* removed libnetwork_discovery tag
* Introduced driver apis for discovery events
* moved overlay driver to make use of the discovery events
* Using Docker Discovery service.
* Changed integration-tests to make use of the new discovery
Signed-off-by: Madhu Venugopal <madhu@docker.com>
Ideally, both overlay and libnetwork core must be changed to support
kv-store connection retry. But this is a stop-gap measure to unblock the
discovery related PRs.
Signed-off-by: Madhu Venugopal <madhu@docker.com>
Currently the driver configuration is pushed through a separate
api. This makes driver configuration possible at any arbitrary
time. This unncessarily complicates the driver implementation.
More importantly the driver does not get access to it's
configuration before it can do the handshake with libnetwork.
This make the internal drivers a little bit different to
external plugins which can get their configuration before the handshake
with libnetwork.
This PR attempts to fix that mismatch between internal drivers and
external plugins.
Signed-off-by: Jana Radhakrishnan <mrjana@docker.com>
Currently the endpoint data model consists of multiple
interfaces per-endpoint. This seems to be an overkill
since there is no real use case for it. Removing it
to remove unnecessary complexity from the code.
Signed-off-by: Jana Radhakrishnan <mrjana@docker.com>
This way we won't vendor test related functions in docker anymore.
It also moves netns related functions to a new ns package to be able to
call the ns init function in tests. I think this also helps with the
overall package isolation.
Signed-off-by: David Calavera <david.calavera@gmail.com>
Make sure to always explicitly set namespace for all
kernel bound network operations irrespective of whether
the operation is performed in init namespace or a user
defined namespace. This already happens for user defined
netns. But doesn't happen for initial netns that libnetwork
runs in.
Signed-off-by: Jana Radhakrishnan <mrjana@docker.com>
- Maps 1 to 1 with container's networking stack
- It holds container's specific nw options which
before were incorrectly owned by Endpoint.
- Sandbox creation no longer coupled with Endpoint Join,
sandbox and endpoint have now separate lifecycle.
- LeaveAll naturally replaced by Sandbox.Delete
- some pkg and file renaming in order to have clear
mapping between structure name and entity ("sandbox")
- Revisited hosts and resolv.conf handling
- Removed from JoinInfo interface capability of setting hosts and resolv.conf paths
- Changed etchosts.Build() to first write the search domains and then the nameservers
Signed-off-by: Alessandro Boch <aboch@docker.com>
The current lazy network sandbox initialization code has a race
in that if multiple go routines race to join the network the second
and subsequent go routines might try to use the sandbox before it is
fully initialized. Fix this by blocking the go routines in once.Do
calls and also take of care of rolling back properly in case of
error.
Signed-off-by: Jana Radhakrishnan <mrjana@docker.com>
When you start a container after some other container has already
been started in the same network, the current container will have
an fdb which points to a wrong vtep to reach the already started
container. This makes the network connectivity to not work. The root
cause of the issue is because of golang does variable capture by
reference in closures and so we cannot use the return values from
range iterators directly. It needs to be copied to a locally scoped
variable and then use that copy as a capture variable in the closure.
Signed-off-by: Jana Radhakrishnan <mrjana@docker.com>
In that commit, AtomicPutCreate takes previous = nil to Atomically create keys
that don't exist. We need a create operation that is atomic to prevent races
between multiple libnetworks creating the same object.
Previously, we just created new KVs with an index of 0 and wrote them to the
datastore. Consul accepts this behaviour and interprets index of 0 as
non-existing, but other data backends do no.
- Add Exists() to the KV interface. SetIndex() should also modify a KV so
that it exists.
- Call SetIndex() from within the GetObject() method on DataStore interface.
- This ensures objects have the updated values for exists and index.
- Add SetValue() to the KV interface. This allows implementers to define
their own method to marshall and unmarshall (as bitseq and allocator have).
- Update existing users of the DataStore (endpoint, network, bitseq,
allocator, ov_network) to new interfaces.
- Fix UTs.
This commit brings in the first implementation of
overlay driver which makes use of vxlan tunneling
protocol to create logical networks across multiple
hosts.
This is very much alpha code and should be used for
demo and testing purposes only.
Signed-off-by: Jana Radhakrishnan <mrjana@docker.com>
