Stop joining libcontainer executor process into the newly created task
container cgroup, to ensure that the cgroups are fully destroyed on
shutdown, and to make it consistent with other plugin processes.
Previously, executor process is added to the container cgroup so the
executor process resources get aggregated along with user processes in
our metric aggregation.
However, adding executor process to container cgroup adds some
complications with much benefits:
First, it complicates cleanup. We must ensure that the executor is
removed from container cgroup on shutdown. Though, we had a bug where
we missed removing it from the systemd cgroup. Because executor uses
`containerState.CgroupPaths` on launch, which includes systemd, but
`cgroups.GetAllSubsystems` which doesn't.
Second, it may have advese side-effects. When a user process is cpu
bound or uses too much memory, executor should remain functioning
without risk of being killed (by OOM killer) or throttled.
Third, it is inconsistent with other drivers and plugins. Logmon and
DockerLogger processes aren't in the task cgroups. Neither are
containerd processes, though it is equivalent to executor in
responsibility.
Fourth, in my experience when executor process moves cgroup while it's
running, the cgroup aggregation is odd. The cgroup
`memory.usage_in_bytes` doesn't seem to capture the full memory usage of
the executor process and becomes a red-harring when investigating memory
issues.
For all the reasons above, I opted to have executor remain in nomad
agent cgroup and we can revisit this when we have a better story for
plugin process cgroup management.
This commit introduces support for configuring mount propagation when
mounting volumes with the `volume_mount` stanza on Linux targets.
Similar to Kubernetes, we expose 3 options for configuring mount
propagation:
- private, which is equivalent to `rprivate` on Linux, which does not allow the
container to see any new nested mounts after the chroot was created.
- host-to-task, which is equivalent to `rslave` on Linux, which allows new mounts
that have been created _outside of the container_ to be visible
inside the container after the chroot is created.
- bidirectional, which is equivalent to `rshared` on Linux, which allows both
the container to see new mounts created on the host, but
importantly _allows the container to create mounts that are
visible in other containers an don the host_
private and host-to-task are safe, but bidirectional mounts can be
dangerous, as if the code inside a container creates a mount, and does
not clean it up before tearing down the container, it can cause bad
things to happen inside the kernel.
To add a layer of safety here, we require that the user has ReadWrite
permissions on the volume before allowing bidirectional mounts, as a
defense in depth / validation case, although creating mounts should also require
a priviliged execution environment inside the container.
Currently, nomad "plugin" processes (e.g. executor, logmon, docker_logger) are started as CLI
commands to be handled by command CLI framework. Plugin launchers use
`discover.NomadBinary()` to identify the binary and start it.
This has few downsides: The trivial one is that when running tests, one
must re-compile the nomad binary as the tests need to invoke the nomad
executable to start plugin. This is frequently overlooked, resulting in
puzzlement.
The more significant issue with `executor` in particular is in relation
to external driver:
* Plugin must identify the path of invoking nomad binary, which is not
trivial; `discvoer.NomadBinary()` now returns the path to the plugin
rather than to nomad, preventing external drivers from launching
executors.
* The external driver may get a different version of executor than it
expects (specially if we make a binary incompatible change in future).
This commit addresses both downside by having the plugin invocation
handling through an `init()` call, similar to how libcontainer init
handler is done in [1] and recommened by libcontainer [2]. `init()`
will be invoked and handled properly in tests and external drivers.
For external drivers, this change will cause external drivers to launch
the executor that's compiled against.
There a are a couple of downsides to this approach:
* These specific packages (i.e executor, logmon, and dockerlog) need to
be careful in use of `init()`, package initializers. Must avoid having
command execution rely on any other init in the package. I prefixed
files with `z_` (golang processes files in lexical order), but ensured
we don't depend on order.
* The command handling is spread in multiple packages making it a bit
less obvious how plugin starts are handled.
[1] drivers/shared/executor/libcontainer_nsenter_linux.go
[2] eb4aeed24f/libcontainer (using-libcontainer)
Nomad 0.9 incidentally set effective capabilities that is higher than
what's expected of a `nobody` process, and what's set in 0.8.
This change restores the capabilities to ones used in Nomad 0.9.
Implements streamign exec handling in both executors (i.e. universal and
libcontainer).
For creation of TTY, some incidental complexity leaked in. The universal
executor uses github.com/kr/pty for creation of TTYs.
On the other hand, libcontainer expects a console socket and for libcontainer to
create the underlying console object on process start. The caller can then use
`libcontainer.utils.RecvFd()` to get tty master end.
I chose github.com/kr/pty for managing TTYs here. I tried
`github.com/containerd/console` package (which is already imported), but the
package did not work as expected on macOS.
