Make all the low-level init code synchronous to avoid weird blocks on `close`.
Also move the net and ctl file descriptor at the beginning of the fd space for
the calf.
The SDK also allow to spamn multiple exec calves, which will all have the same
fd map:
- 0: stdin = /dev/null
- 1: stdout = pipe to parent stdout
- 2: stderr = pipe to parent stderr
- 3: net = socketpair to parent "net" pipe
- 4: ctl = socketpair to parent "ctl" pipe
Signed-off-by: Thomas Gazagnaire <thomas@gazagnaire.org>
This forces us to bind mount /lib but will be replaced by calling the proper
bindings later on.
Signed-off-by: Thomas Gazagnaire <thomas@gazagnaire.org>
It is not necessary to bring up `eth0`, the program does it fine.
This means we can remove shell script, clean up build.
Signed-off-by: Justin Cormack <justin.cormack@docker.com>
Instead of specifying the number of instances, provide a list
of IP addresses for instances. These are passed to the instance
plugin as LogicalID.
Signed-off-by: Rolf Neugebauer <rolf.neugebauer@docker.com>
In #1485 I was still using a local mobylinux/init containing #1436, even though
I had included the necessary files in the swarmd container.
Switch to the current standard init package and drop the unnecessary bind.
Also `git add .gitignore` which I forgot last time too.
Signed-off-by: Ian Campbell <ian.campbell@docker.com>
This is based on https://github.com/docker/swarmkit/pull/1965 which adds a
direct containerd executor to swarmkit. It is very much a work in progress.
With a suitable moby image (such as projects/swarmd/swarmd.yml) something like
this should work:
runc exec swarmd swarmctl service create --image docker.io/library/nginx:alpine --name nginx
runc exec swarmd swarmctl service ls
Signed-off-by: Ian Campbell <ian.campbell@docker.com>
This just sets up the initial cluster via bootstrap.
It does *not* manage state changes correctly afterwards. If one
node crashes (get's killed) it InfraKit will start a new node,
but the new node does not join the cluster (and the old node
is not removed, either).
Signed-off-by: Rolf Neugebauer <rolf.neugebauer@docker.com>
- Script to set up a DfM etcd for bootstrapping a cluser
- Custom/local etcd package for moby
derived from the official image with script to start etcd
- YAML file to create a etcd moby image
- README with current instructions
This has a bunch of stuff, including the discovery URL hard coded.
Signed-off-by: Rolf Neugebauer <rolf.neugebauer@docker.com>
In the riddler change I changed "command" in the yaml to "args"
but did not change the files. In fact we basically used the
default command everywhere so this did not actually break.
Remove the unnecessary "command" lines to simplify yaml.
Revert the command to args change for now as I think I prefer
command, but its easier to switch now. Need to think if the
entrypoint/command distinction matters before finalizing.
Signed-off-by: Justin Cormack <justin.cormack@docker.com>
Generated largely from the specified config; small parts taken from `docker image inspect`,
such as the command line.
Renamed some of the yaml keys to match the OCI spec rather than Docker Compose as
we decided they are more readable, no more underscores.
Add some extra functionality
- tmpfs specification
- fully general mount specification
- no new privileges can be specified now
For nostalgic reasons, using engine-api to talk to the docker cli as
we only need an old API version, and it is nice and easy to vendor...
Signed-off-by: Justin Cormack <justin.cormack@docker.com>
Use a file (that we can drop easily into the rootfs) instead of passing the full
command on the CLI (as config.json needs to be edited otherwise).
Signed-off-by: Thomas Gazagnaire <thomas@gazagnaire.org>
For now, use jbuilder to build the calf as well, this gives us a bit
more control than the mirage tool. We will switch back to the mirage
tool later on if we want to use more fancy backends (such as KVM).
Signed-off-by: Thomas Gazagnaire <thomas@gazagnaire.org>
The protocol allows the client to send concurrent RPCs to the server.
The server replies by keeping the client ID, and the client keeps a
dispatch table of queries to route the retries. By doing things like
that, the server has a strong control over resource allocation, so
a bad client cannot exhaust all the server memory.
Also add some simple tests for the serialization + wire protocol which
are all passing. Proper concurrency/resource exhaustion usage will be
added later.
Signed-off-by: Thomas Gazagnaire <thomas@gazagnaire.org>
Previously, the control plane was using HTTP client/server, that various people
found way too complex to run in a privileged container (for very good reasons).
So switching to a simpler binary protocol, using c-like structures. Will
probably switch to an other serialization protocol later (eg. protobuf
or cap-n-proto).
Signed-off-by: Thomas Gazagnaire <thomas@gazagnaire.org>
