Demystifying Containers – Part II: Container Runtimes

This second blog post (and talk) is primary scoped to container runtimes, where we will start with their historic origins before digging deeper into two dedicated projects: runc and CRI-O. We will initially build up a great foundation about how container runtimes work under the hood by starting with the lower level runtime runc. Afterwards, we will utilize the more advanced runtime CRI-O to run Kubernetes native workloads, but without even running Kubernetes at all.


In the previous part of this series we discussed Linux Kernel Namespaces and everything around to build up a foundation about containers and their basic isolation techniques. Now we want to dive deeper into answering the question: “How to actually run containers?”. We will do so without being overwhelmed by the details of Kubernetes’ features or security related topics, which will be part of further blog posts and talks.

What is a Container Runtime?

Applications and their required or not required use cases are contentiously discussed topics in the UNIX world. The mainUNIX philosophy propagates minimalism and modular software parts which should fit well together in a complete system. Great examples which follow these philosophical aspects are features like the UNIX pipe or text editors like vim. These tools solve one dedicated task as best as they can and are tremendously successful at it. On the other side, there are projects like systemd or cmake, which do not follow the same approach and implement a richer feature set over time. In the end we have multiple views and opinions about answers to questions like ”How should an initialization system look like?” or ”What should a build system do?”. If these multi-opinionated views mix up with historical events, then answering a simple question might need more explanations than it should.

Now, welcome to the world of containers!

Lots of applications can run containers, whereas every application would have a sightly different opinion about what a container runtime should do and support. For example, systemd is able to run containers via systemd-nspawn, and NixOS has integrated container management as well. Not to mention all the other existing container runtimes like CRI-OKata ContainersFirecrackergVisorcontainerdLXCruncNabla Containers and many more. A lot of them are now part of theCloud Native Computing Foundation (CNCF) and their huge landscape, whereas someone might ask: ”Why do so many container runtimes exist?”.

Per usual for our series of blog posts, we should start from the historical beginning.

A Brief History

After the invention of cgroups back in 2008, a project called Linux Containers (LXC) started to pop-up in the wild, which should revolutionize the container world. LXC combined cgroup and namespace technologies to provide an isolated environment for running applications. You may know that we sometimes live in a parallel world. This means that Google started their own containerization project in 2007 called Let Me Contain That For You (LMCTFY), which works mainly at the same level as LXC does. With LMCTFY, Google tried to provide a stable and API driven configuration without users having to understand the details of cgroups and its internals.

If we now look back into 2013 we see that there was a tool written called Docker, which was built on top of the already existing LXC stack. One invention of Docker was that the user is now able to package containers into images to move them between machines. Docker were the fist ones who tried to make containers a standard software unit, as they state in their ”Standard Container Manifesto”.

Some years later they began to work on libcontainer, a Go native way to spawn and manage containers. LMCTFY was abandoned during that time too, whereas the core concepts and major benefits of LMCTFY were ported into libcontainer and Docker.

We are now back in 2015, where projects like Kubernetes hit version 1.0. A lot of stuff was ongoing during that time: The CNCF was founded as part of the Linux Foundation with the target to promote containers. The Open Container Initiative (OCI)was founded 2015 as well, as an open governance structure around the container ecosystem.

Their main target is to create open industry standards around container formats and runtimes. We were now in a state where containers are used, in terms of their popularity, side by side with classic Virtual Machines (VMs). There was a need for a specification of how containers should run, which resulted in the OCI Runtime Specification. Runtime developers should now be able to have a well-defined API to develop their container runtime. The libcontainer project was donated to the OCI during that time, whereas a new tool called runc was born as part of that. With runc it was now possible to directly interact with libcontainer, interpret the OCI Runtime Specification and run containers from it.

As of today, runc is one of the most popular projects in the container ecosystem and is used in a lot of other projects like containerd (used by Docker), CRI-O and podman. Other projects adopted the OCI Runtime Specification as well. For example Kata Containers makes it possible to build and run secure containers including lightweight virtual machines that feel and perform like containers, but provide stronger workload isolation using hardware virtualization technology as a second layer of defense.

Let’s dig more into the OCI Runtime Specification to get a better understanding about how a container runtime works under the hood.

Running Containers


The OCI Runtime Specification provides information about the configuration, execution environment and overall life cycle of a container. A configuration is mainly a JSON file that contains all necessary information to enable the creation of a container on different target platforms like Linux, Windows or Virtual Machines (VMs).

An example specification can be easily generated with runc:

> runc spec
> cat config.json
  "ociVersion": "1.0.0",
  "process": {
    "terminal": true,
    "user": { "uid": 0, "gid": 0 },
    "args": ["sh"],
    "env": [
    "cwd": "/",
    "capabilities": {
      "bounding": ["CAP_AUDIT_WRITE", "CAP_KILL", "CAP_NET_BIND_SERVICE"],
    "rlimits": [ { "type": "RLIMIT_NOFILE", "hard": 1024, "soft": 1024 } ],
    "noNewPrivileges": true
  "root": { "path": "rootfs", "readonly": true },
  "hostname": "runc",
  "mounts": [
      "destination": "/proc",
      "type": "proc",
      "source": "proc"
  "linux": {
    "resources": { "devices": [ { "allow": false, "access": "rwm" } ] },
    "namespaces": [
      { "type": "pid" },
      { "type": "network" },
      { "type": "ipc" },
      { "type": "uts" },
      { "type": "mount" }
    "maskedPaths": [
    "readonlyPaths": [

This file mainly contains all necessary information for runc to get started with running containers. For example, we have attributes about the running process, the defined environment variables, the user and group IDs, needed mount points and the Linux namespaces to be set up. One thing is still missing to get started running containers: We need an appropriate root file-system (rootfs). We already discovered in the past blog post how to obtain it from an already existing container image:

> skopeo copy docker://opensuse/tumbleweed:latest oci:tumbleweed:latest
[output removed]
> sudo umoci unpack --image tumbleweed:latest bundle
[output removed]

Interestingly, the unpacked container image already includes the Runtime Specification we need to run the bundle:

> sudo chown -R $(id -u) bundle
> cat bundle/config.json
  "ociVersion": "1.0.0",
  "process": {
    "terminal": true,
    "user": { "uid": 0, "gid": 0 },
    "args": ["/bin/bash"],
    "env": [
    "cwd": "/",
    "capabilities": { [...] },
    "rlimits": [...]
  "root": { "path": "rootfs" },
  "hostname": "mrsdalloway",
  "mounts": [...],
  "annotations": {
    "org.opencontainers.image.title": "openSUSE Tumbleweed Base Container",
    "org.opencontainers.image.url": "",
    "org.opencontainers.image.vendor": "openSUSE Project",
    "org.opencontainers.image.version": "20190517.6.190",
  "linux": {
    "resources": { "devices": [ { "allow": false, "access": "rwm" } ] },
    "namespaces": [
      { "type": "pid" },
      { "type": "network" },
      { "type": "ipc" },
      { "type": "uts" },
      { "type": "mount" }

There are now some annotations included beside the usual fields we already know from running runc spec. These can be used to add arbitrary metadata to the container, which can be utilized by higher level runtimes to add additional information to the specification.

Let’s create a new container from the bundle with runc. Before actually calling out to runc, we have to setup a receiver terminal to be able to interact with the container. For this, we can use the recvtty tool included in the runc repository:

> go get
> recvtty tty.sock

In another terminal, we now call runc create with specifying the bundle and terminal socket:

> sudo runc create -b bundle --console-socket $(pwd)/tty.sock container

No further output, so what happened now? It seems like we have created a new container in created state:

> sudo runc list
ID          PID         STATUS      BUNDLE      CREATED                          OWNER
container   29772       created     /bundle     2019-05-21T08:35:51.382141418Z   root

The container seems to be not running, but what is running inside?

> sudo runc ps container
root     29772     1  0 10:35 ?        00:00:00 runc init

The runc init command sets up a fresh environment with all necessary namespaces and launches a new initial process. The main process /bin/bash does not run yet inside the container, but we are still able to execute further processes within the container:

> sudo runc exec -t container echo "Hello, world!"
> Hello, world!

The created state of a container provides a nice environment to setup networking for example. To actually do something within the container, we have to bring it into the running state. This can be done via runc start:

> sudo runc start container

In the terminal where the recvtty process is running, a new bash shell session should now pop up:

mrsdalloway:/ $
mrsdalloway:/ $ ps aux
ps aux
root         1  0.0  0.0   5156  4504 pts/0    Ss   10:28   0:00 /bin/bash
root        29  0.0  0.0   6528  3372 pts/0    R+   10:32   0:00 ps aux

Nice, the container seems to be running. We can now utilize runc to inspect the container’s state:

> sudo runc list
ID          PID         STATUS      BUNDLE      CREATED                          OWNER
container   4985        running     /bundle     2019-05-20T12:14:14.232015447Z   root
> sudo runc ps container
root      6521  6511  0 14:25 pts/0    00:00:00 /bin/bash

The runc init process has gone and now only the actual /bin/bash process exists within the container. We can also do some basic life cycle management with the container:

> sudo runc pause container

It should now be impossible to get any output from the running container in the recvtty session. To resume the container, simply call:

> sudo runc resume container

Everything we tried to type before should now pop up in the resumed container terminal. If we need more information about the container, like the CPU and memory usage, then we can retrieve them via the runc events API:

> sudo runc events container

The output is a bit hard to read, so let’s reformat it and strip some fields:

  "type": "stats",
  "id": "container",
  "data": {
    "cpu": {
      "usage": {
        "total": 31442016,
        "percpu": [ 5133429, 5848165, 827530, ... ],
        "kernel": 20000000,
        "user": 0
      "throttling": {}
    "memory": {
      "usage": {
        "limit": 9223372036854771712,
        "usage": 1875968,
        "max": 6500352,
        "failcnt": 0
      "swap": { "limit": 0, "failcnt": 0 },
      "kernel": {
        "limit": 9223372036854771712,
        "usage": 311296,
        "max": 901120,
        "failcnt": 0
      "kernelTCP": { "limit": 9223372036854771712, "failcnt": 0 },
      "raw": {
        "active_anon": 1564672,
    "pids": { "current": 1 },
    "blkio": {},
    "hugetlb": { "1GB": { "failcnt": 0 }, "2MB": { "failcnt": 0 } },
    "intel_rdt": {}

We can see that we are able to retrieve detailed runtime information about the container.

To stop the container, we simply exit the recvtty session. Afterwards the container can be removed with runc delete:

> sudo runc list
ID          PID         STATUS      BUNDLE      CREATED                         OWNER
container   0           stopped     /bundle     2019-05-21T10:28:32.765888075Z  root
> sudo runc delete container
> sudo runc list
ID          PID         STATUS      BUNDLE      CREATED     OWNER

Containers in the stopped state cannot run again, so they have to be recreated from a fresh state. As already mentioned, the extracted bundle contains the necessary config.json file beside the rootfs, which will be used by runc to setup the container. We could for example modify the initial run command of the container by executing:

> cd bundle
> jq '.process.args = ["echo", "Hello, world!"]' config.json | sponge config.json
> sudo runc run container
> Hello, world!

We have nearly every freedom by editing the rootfs or the config.json. So we could tear down the PID namespace isolation between the container and the host:

> jq '.process.args = ["ps", "a"] | del(.linux.namespaces[0])' config.json | sponge config.json
> sudo runc run container
16583 ?        S+     0:00 sudo runc run container
16584 ?        Sl+    0:00 runc run container
16594 pts/0    Rs+    0:00 ps a
[output truncated]

In the end runc is a pretty low level runtime, whereas improper configuration and usage can lead into serious security concerns. Truly, runc has native support for security enhancements like seccompSecurity-Enhanced Linux (SELinux) andAppArmor but these features should be used by higher level runtimes to ensure correct usage in production. It is also worth mentioning that it is possible to run containers in rootless mode via runc to security harden the deployment even further. We will cover these topics in future blog posts as well, but for now that should suffice on that level.

Another drawback in running containers only with runc would be that we have to manually set up the networking to the host to reach out to the internet or other containers. In order to do that we could use the Runtime Specification Hooks feature to set up a default bridge before actually starting the container.

But why don’t we leave this job to a higher level runtime as well? Let’s go for that and move on.

The Kubernetes Container Runtime Interface (CRI)

Back in 2016, the Kubernetes project announced the implementation of the Container Runtime Interface (CRI), which provides a standard API for container runtimes to work with Kubernetes. This interface enables users to exchange the runtime in a cluster with ease.

How does the API work? At the bottom line of every Kubernetes cluster runs a piece of software called the kubelet, which has the main job of keeping container workloads running and healthy. The kubelet connects to a gRPC server on startup and expects a predefined API there. For example, some service definitions of the API look like this:

// Runtime service defines the public APIs for remote container runtimes
service RuntimeService {
    rpc CreateContainer (...) returns (...) {}
    rpc ListContainers  (...) returns (...) {}
    rpc RemoveContainer (...) returns (...) {}
    rpc StartContainer  (...) returns (...) {}
    rpc StopContainer   (...) returns (...) {}

That seems to be pretty much what we already did with runc, managing the container life cycle. If we look further at the API, we see this:

    rpc ListPodSandbox  (...) returns (...) {}
    rpc RemovePodSandbox(...) returns (...) {}
    rpc RunPodSandbox   (...) returns (...) {}
    rpc StopPodSandbox  (...) returns (...) {}

What does “sandbox” mean? Containers should already be some kind of sandbox, right? Yes, but in the Kubernetes worldPods can consist of multiple containers, whereas this abstract hierarchy has to be mapped into a simple list of containers. Because of that, every creation of a Kubernetes Pod starts with the setup of a so called PodSandbox. Every container running inside the Pod is attached to this sandbox, so the containers inside can share common resources, like their network interfaces for example. runc alone does not provide such features out of the box, so we have to use a higher level runtime to achieve our goal.


CRI-O is a higher level container runtime which has been written on purpose to be used with the Kubernetes CRI. The name originates from the combination of the Container Runtime Interface and the Open Container Initiative. Isn’t that simple? CRI-O’s journey started as Kubernetes incubator project back in 2016 under the name Open Container Initiative Daemon (OCID). Version 1.0.0 has been released one year later in 2017 and follows the Kubernetes release cycles from that day on. This means for example, that the Kubernetes version 1.15 can be safely used together with CRI-O 1.15 and so on.

The implementation of CRI-O follows the main UNIX philosophy and tends to be a lightweight alternative to Docker or containerd when it comes to running production-ready workloads inside of Kubernetes. It is not meant to be a developers-facing tool which can be used from the command line. CRI-O has only one major task: Fulfilling the Kubernetes CRI. To achieve that, it utilizes runc for basic container management in the back, whereas the gRPC server provides the API in the front end. Everything in between is done either by CRI-O itself or by core libraries like containers/storage or containers/image. But in the end it doesn’t mean that we cannot play around with it, so let’s give it a try.

I prepared a container image called “crio-playground” to get started with CRI-O in an efficient manner. This image contains all necessary tools, example files and a working CRI-O instance running in the background. To start a privileged container running the crio-playground, simply execute:

> sudo podman run --privileged -h crio-playground -it saschagrunert/crio-playground
crio-playground:~ $

From now on we will use a tool called crictl to interface with CRI-O and its Container Runtime Interface implementation. crictl allows us to use YAML representations of the CRI API requests to send them to CRI-O. For example, we can create a new PodSandbox with the sandbox.yml lying around in the current working directory of the playground:

  name: sandbox
  namespace: default

To create the sandbox in the running crio-playground container, we now execute:

crio-playground:~ $ crictl runp sandbox.yml

Let’s store the identifier of the sandbox as $POD_ID environment variable for later usage as well:

crio-playground:~ $ export POD_ID=5f2b94f74b28c092021ad8eeae4903ada4b1ef306adf5eaa0e985672363d6336

If we now run crictl pods we can see that we finally have one PodSandbox up and running:

crio-playground:~ $ crictl pods
POD ID              CREATED             STATE               NAME                NAMESPACE           ATTEMPT
5f2b94f74b28c       43 seconds ago      Ready               sandbox             default             0

But what’s inside our sandbox? We surely can examine the sandbox further by using runc:

crio-playground:~ $ runc list
ID                                                                 PID         STATUS      BUNDLE                                                                                                             CREATED                          OWNER
5f2b94f74b28c092021ad8eeae4903ada4b1ef306adf5eaa0e985672363d6336   80          running     /run/containers/storage/vfs-containers/5f2b94f74b28c092021ad8eeae4903ada4b1ef306adf5eaa0e985672363d6336/userdata   2019-05-23T13:43:38.798531426Z   root

The sandbox seems to run in a dedicated bundle under /run/containers.

crio-playground:~ $ runc ps $POD_ID
root        80    68  0 13:43 ?        00:00:00 /pause

Interestingly, there is only one process running inside the sandbox, called pause. As the source code of pause indicates, the main task of this process is to keep the environment running and react to incoming signals. Before we actually create our workload within that sandbox, we have to pre-pull the image we want to run. A trivial example would be to run a web server, so let’s retrieve a nginx image by calling:

crio-playground:~ $ crictl pull nginx:alpine
Image is up to date for

Now let’s create a very simple container definition in YAML, like we did for the sandbox:

  name: container
  image: nginx:alpine

And now, let’s kick off the container. For that we have to provide the hash of the sandbox as well as the YAML definitions of the sandbox and container:

crio-playground:~ $ crictl create $POD_ID container.yml sandbox.yml

Seems to work! Let’s store the container identifier as $CONTAINER_ID for later reuse as well:

crio-playground:~ $ export CONTAINER_ID=b205eb2c6abec3e7ade72e0cea09d827968a4c1089483cab06bdf0f4ee82ff0c

What would you expect if we now check out the status of our two running containers while keeping the CRI API in mind? Correct, the container should be in the created state:

crio-playground:~ $ runc list
ID                                                                 PID         STATUS      BUNDLE                                                                                                             CREATED                          OWNER
5f2b94f74b28c092021ad8eeae4903ada4b1ef306adf5eaa0e985672363d6336   80          running     /run/containers/storage/vfs-containers/5f2b94f74b28c092021ad8eeae4903ada4b1ef306adf5eaa0e985672363d6336/userdata   2019-05-23T13:43:38.798531426Z   root
b205eb2c6abec3e7ade72e0cea09d827968a4c1089483cab06bdf0f4ee82ff0c   343         created     /run/containers/storage/vfs-containers/b205eb2c6abec3e7ade72e0cea09d827968a4c1089483cab06bdf0f4ee82ff0c/userdata   2019-05-23T14:08:53.701174406Z   root

And, like in our previous runc example, the container waits in runc init:

crio-playground:~ $ runc ps $CONTAINER_ID
root       343   331  0 14:08 ?        00:00:00 /usr/sbin/runc init

crictl shows the container in created as well:

crio-playground:~ $ crictl ps -a
CONTAINER ID        IMAGE               CREATED             STATE               NAME                ATTEMPT             POD ID
b205eb2c6abec       nginx:alpine        13 minutes ago      Created             container           0                   5f2b94f74b28c

Now we have to start the workload to get it into the running state:

crio-playground:~ $ crictl start $CONTAINER_ID

This should be successful, too. Let’s verify if all processes are running correctly:

crio-playground:~ $ crictl ps
CONTAINER ID        IMAGE               CREATED             STATE               NAME                ATTEMPT             POD ID
b205eb2c6abec       nginx:alpine        15 minutes ago      Running             container           0                   5f2b94f74b28c

Inside the container should now run an nginx web server:

crio-playground:~ $ runc ps $CONTAINER_ID
root       343   331  0 14:08 ?        00:00:00 nginx: master process nginx -g daemon off;
100        466   343  0 14:24 ?        00:00:00 nginx: worker process

But how to reach the web servers content now? We did not expose any ports or other advanced configuration for the container, so it should be fairly isolated from the host. The solution lies down in the container networking. Because we use a bridged network configuration in the crio-playground, we can simply access the containers network address. To get these we can exec into the container and list the network interfaces:

crio-playground:~ $ crictl exec $CONTAINER_ID ip addr
1: lo: <LOOPBACK,UP,LOWER_UP> mtu 65536 qdisc noqueue state UNKNOWN qlen 1000
    link/loopback 00:00:00:00:00:00 brd 00:00:00:00:00:00
    inet scope host lo
       valid_lft forever preferred_lft forever
    inet6 ::1/128 scope host
       valid_lft forever preferred_lft forever
3: eth0: <BROADCAST,MULTICAST,UP,LOWER_UP> mtu 1500 qdisc noqueue state UP
    link/ether 16:04:8c:44:00:59 brd ff:ff:ff:ff:ff:ff
    inet scope global eth0
       valid_lft forever preferred_lft forever
    inet6 fe80::1404:8cff:fe44:59/64 scope link
       valid_lft forever preferred_lft forever

And now just query the inet address for eth0:

crio-playground:~ $ curl
<!DOCTYPE html>
<title>Welcome to nginx!</title>
[output truncated]

Hooray, it works! We successfully run a Kubernetes workload without running Kubernetes!

The overall Kubernetes story about Network Plugins or the Container Network Interface (CNI) is worth another blog post, but that’s a different story and we stop right here with all the magic.


And that’s a wrap for this part of the blog series about the demystification of containers. We discovered the brief history of container runtimes and had the chance to run containers with the low level runtime runc as well as the higher level runtime CRI-O. I can really recommend to have a closer look at the OCI runtime specification and test different configurations within the crio-playground environment. For sure we will see CRI-O in the future again when we talk about container-related topics like security or networking. Besides that, we will have the chance to explore different tools like podmanbuildah or skopeo, which provide more advanced container management solutions. I really hope you enjoyed the read and will continue following my journey into future parts of this series. Feel free to drop me a line anywhere you can find me on the internet. Stay tuned!

You can find all necessary resources about this series on GitHub.

Demystifying Containers – Part I: Kernel Space

This series of blog posts and corresponding talks aims to provide you with a pragmatic view on containers from a historic perspective. Together we will discover modern cloud architectures layer by layer, which means we will start at the Linux Kernel level and end up at writing our own secure cloud native applications.

Simple examples paired with the historic background will guide you from the beginning with a minimal Linux environment up to crafting secure containers, which fit perfectly into todays’ and futures’ orchestration world. In the end it should be much easier to understand how features within the Linux kernel, container tools, runtimes, software defined networks and orchestration software like Kubernetes are designed and how they work under the hood.

Part I: Kernel Space

This first blog post (and talk) is scoped to Linux kernel related topics, which will provide you with the necessary foundation to build up a deep understanding about containers. We will gain an insight about the history of UNIX, Linux and talk about solutions like chroot, namespaces and cgroups combined with hacking our own examples. Besides this we will peel some containers to get a feeling about future topics we will talk about.


If we are talking about containers nowadays, most people tend to think of the big blue whale or the white steering wheel on the blue background.

Let’s put these thoughts aside and ask ourselves: What are containers in detail? If we look at the corresponding documentation of Kubernetes we only find explanations about “Why to use containers?“ and lots of references to Docker. Docker itself explains containers as “a standard unit of software“. Their explanations provide a general overview but do not reveal much of the underlying “magic“. Eventually, people tend to imagine containers as cheap virtual machines (VMs), which technically does not come close to the real world. This could be reasoned since the word “container” does not mean anything precisely at all. The same applies to the word “pod” in the container orchestration ecosystem.

If we strip it down then containers are only isolated groups of processes running on a single host, which fulfill a set of “common” features. Some of these fancy features are built directly into the Linux kernel and mostly all of them have different historical origins.

So containers have to fulfill four major requirements to be acceptable as such:

  1. Not negotiable: They have to run on a single host. Okay, so two computers cannot run a single container.
  2. Clearly: They are groups of processes. You might know that Linux processes live inside a tree structure, so we can say containers must have a root process.
  3. Okay: They need to be isolated, whatever this means in detail.
  4. Not so clear: They have to fulfill common features. Features in general seem to change over time, so we have to point out what the most common features are.

These requirements alone can lead into confusion and the picture is not clear yet. So let’s start from the historical beginning to keep things simple.


Mostly every UNIX operating system has the possibility to change the root directory of the current running process (and its children). This originates from the first occurrence of chroot in UNIX Version 7 (released 1979), from where it continued the journey into the awesome Berkeley Software Distribution (BSD). In Linux you can nowadays chroot(2) as system call (a kernel API function call) or the corresponding standalone wrapper program. Chroot is also referenced as “jail“, because some person used it as a honeypot to monitor a security hacker back in 1991. So chroot is much older than Linux and it has been (mis)used in the early 2000s for the first approaches in running applications as what we would call today “microservices”. Chroot is currently used by a wide range of applications, for example within build services for different distributions. Nowadays the BSD implementation differs a lots from the Linux one, where we will focus on the latter part for now.

What is needed to run an own chroot environment? Not that much, since something like this already works:

> mkdir -p new-root/{bin,lib64}
> cp /bin/bash new-root/bin
> cp /lib64/{*,*,,*,*} new-root/lib64
> sudo chroot new-root

We create a new root directory, copy a bash shell and its dependencies in and run chroot. This jail is pretty useless: All we have at hand is bash and its builtin functions like cd and pwd.

One might think it could be worth running a statically linked binary in a jail and that would be the same as running a container image. It’s absolutely not, and a jail is not really a standalone security feature but more a good addition to our container world.

The current working directory is left unchanged when calling chroot via a syscall, whereas relative paths can still refer to files outside of the new root. This call changes only the root path and nothing else. Beside this, further calls to chroot do not stack and they will override the current jail. Only privileged processes with the capability CAP_SYS_CHROOT are able to call chroot. At the end of the day the root user can easily escape from a jail by running a program like this:

#include <sys/stat.h>
#include <unistd.h>

int main(void)
    mkdir(".out", 0755);
    return execl("/bin/bash", "-i", NULL);

We create a new jail by overwriting the current one and change the working directly to some relative path outside of the chroot environment. Another call to chroot might bring us outside of the jail which can be verified by spawning a new interactive bash shell.

Nowadays chroot is not used by container runtimes any more and was replaced by pivot_root(2), which has the benefit of putting the old mounts into a separate directory on calling. These old mounts could be unmounted afterwards to make the filesystem completely invisible to broken out processes.

To continue with a more useful jail we need an appropriate root filesystem (rootfs). This contains all binaries, libraries and the necessary file structure. But where to get one? What about peeling it from an already existing Open Container Initiative (OCI) container, which can be easily done with the two tools skopeo and umoci:

> skopeo copy docker://opensuse/tumbleweed:latest oci:tumbleweed:latest

[output removed]

> sudo umoci unpack –image tumbleweed:latest bundle

[output removed]

Now with our freshly downloaded and extracted rootfs we can chroot into the jail via:

> sudo chroot bundle/rootfs

It looks like we’re running inside a fully working environment, right? But what did we achieve? We can see that we may sneak-peak outside the jail from a process perspective:

> mkdir /proc
> mount -t proc proc /proc
> ps aux

[output removed]

There is no process isolation available at all. We can even kill programs running outside of the jail, what a metaphor! Let’s peek into the network devices:

> mkdir /sys
> mount -t sysfs sys /sys
> ls /sys/class/net
eth0 lo

There is no network isolation, too. This missing isolation paired with the ability to leave the jail leads into lots of security related concerns, because jails are sometimes used for wrong (security related) purposes. How to solve this? This is where the Linux namespaces join the party.

Linux Namespaces

Namespaces are a Linux kernel feature which were introduced back in 2002 with Linux 2.4.19. The idea behind a namespace is to wrap certain global system resources in an abstraction layer. This makes it appear like the processes within a namespace have their own isolated instance of the resource. The kernels namespace abstraction allows different groups of processes to have different views of the system.

Not all available namespaces were implemented from the beginning. A full support for what we now understand as “container ready” was finished in kernel version 3.8 back in 2013 with the introduction of the user namespace. We end up having currently seven distinct namespaces implemented: mnt, pid, net, ipc, uts, user and cgroup. No worries, we will discuss them in detail. In September 2016 two additional namespaces were proposed (time and syslog) which are not fully implemented yet. Let’s have a look into the namespace API before digging into certain namespaces.


The namespace API of the Linux kernel consists of three main system calls:


The clone(2) API function creates a new child process, in a manner similar to fork(2). Unlike fork(2), the clone(2) API allows the child process to share parts of its execution context with the calling process, such as the memory space, the table of file descriptors, and the table of signal handlers. You can pass different namespace flags to clone(2)to create new namespaces for the child process.


The function unshare(2) allows a process to disassociate parts of the execution context which are currently being shared with others.


The function setns(2) reassociates the calling thread with the provided namespace file descriptor. This function can be used to join an existing namespace.


Besides the available syscalls, the proc filesystem populates additional namespace related files. Since Linux 3.8, each file in /proc/$PID/ns is a “magic“ link which can be used as a handle for performing operations (like setns(2)) to the referenced namespace.

> ls -Gg /proc/self/ns/
total 0
lrwxrwxrwx 1 0 Feb  6 18:32 cgroup -> 'cgroup:[4026531835]'
lrwxrwxrwx 1 0 Feb  6 18:32 ipc -> 'ipc:[4026531839]'
lrwxrwxrwx 1 0 Feb  6 18:32 mnt -> 'mnt:[4026531840]'
lrwxrwxrwx 1 0 Feb  6 18:32 net -> 'net:[4026532008]'
lrwxrwxrwx 1 0 Feb  6 18:32 pid -> 'pid:[4026531836]'
lrwxrwxrwx 1 0 Feb  6 18:32 pid_for_children -> 'pid:[4026531836]'
lrwxrwxrwx 1 0 Feb  6 18:32 user -> 'user:[4026531837]'
lrwxrwxrwx 1 0 Feb  6 18:32 uts -> 'uts:[4026531838]'

This allows us for example to track in which namespaces certain processes reside. Another way to play around with namespaces apart from the programmatic approach is using tools from the util-linux package. This contains dedicated wrapper programs for the mentioned syscalls. One handy tool related to namespaces within this package is lsns. It lists useful information about all currently accessible namespaces or about a single given one. But now let’s finally get our hands dirty.

Available Namespaces

Mount (mnt)

The first namespace we want to try out is the mnt namespace, which was the first implemented one back in 2002. During that time (mostly) no one thought that multiple namespaces would ever be needed, so they decided to call the namespace clone flag CLONE_NEWNS. This leads into a small inconsistency with other namespace clone flags (I see you suffering!). With the mnt namespace Linux is able to isolate a set of mount points by a group of processes.

A great use case of the mnt namespace is to create environments similar to jails, but in a more secure fashion. How to create such a namespace? This can be easily done via an API function call or the unshare command line tool. So we can do this:

> sudo unshare -m
# mkdir mount-dir
# mount -n -o size=10m -t tmpfs tmpfs mount-dir
# df mount-dir
Filesystem     1K-blocks  Used Available Use% Mounted on
tmpfs              10240     0     10240   0% <PATH>/mount-dir
# touch mount-dir/{0,1,2}

Looks like we have a successfully mounted tmpfs, which is not available on the host system level:

> ls mount-dir
> grep mount-dir /proc/mounts

The actual memory being used for the mount point is laying in an abstraction layer called Virtual File System (VFS), which is part of the kernel and where every other filesystem is based on. If the namespace gets destroyed, the mount memory is unrecoverably lost. The mount namespace abstraction gives us the possibility to create entire virtual environments in which we are the root user even without root permissions.

On the host system we are able to see the mount point via the mountinfo file inside of the proc filesystem:

> grep mount-dir /proc/$(pgrep -u root bash)/mountinfo
349 399 0:84 / /mount-dir rw,relatime - tmpfs tmpfs rw,size=1024k

How to work with these mount points on a source code level? Well, programs tend to keep a file handle on the corresponding /proc/$PID/ns/mnt file, which refers to the used namespace. In the end mount namespace related implementation scenarios can be really complex, but they give us the power to create flexible container filesystem trees. The last thing I want to mention is that mounts can have different flavors (shared, slave, private, unbindable), which is best explained within the shared subtree documentation of the Linux kernel.

UNIX Time-sharing System (uts)

The UTS namespace was introduced in Linux 2.6.19 (2006) and allows us to unshare the domain- and hostname from the current host system. Let’s give it a try:

> sudo unshare -u
# hostname
# hostname new-hostname
# hostname

And if we look at the system level nothing has changed, hooray:

> hostname

The UTS namespace is yet another nice addition in containerization, especially when it comes to container networking related topics.

Interprocess Communication (ipc)

IPC namespaces came with Linux 2.6.19 (2006) too and isolate interprocess communication (IPC) resources. In special these are System V IPC objects and POSIX message queues. One use case of this namespace would be to separate the shared memory (SHM) between two processes to avoid misusage. Instead, each process will be able to use the same identifiers for a shared memory segment and produce two distinct regions. When an IPC namespace is destroyed, then all IPC objects in the namespace are automatically destroyed, too.

Process ID (pid)

The PID namespace was introduced in Linux 2.6.24 (2008) and gives processes an independent set of process identifiers (PIDs). This means that processes which reside in different namespaces can own the same PID. In the end a process has two PIDs: the PID inside the namespace, and the PID outside the namespace on the host system. The PID namespaces can be nested, so if a new process is created it will have a PID for each namespace from its current namespace up to the initial PID namespace.

The first process created in a PID namespace gets the number 1 and gains all the same special treatment as the usual init process. For example, all processes within the namespace will be re-parented to the namespace’s PID 1 rather than the host PID 1. In addition the termination of this process will immediately terminate all processes in its PID namespace and any descendants. Let’s create a new PID namespace:

> sudo unshare -fp --mount-proc
# ps aux
root         1  0.4  0.6  18688  6608 pts/0    S    23:15   0:00 -bash
root        39  0.0  0.1  35480  1768 pts/0    R+   23:15   0:00 ps aux

Looks isolated, doesn’t it? The --mount-proc flag is needed to re-mount the proc filesystem from the new namespace. Otherwise we would not see the PID subtree corresponding with the namespace. Another option would be to manually mount the proc filesystem via mount -t proc proc /proc, but this also overrides the mount from the host where it has to be remounted afterwards.

Network (net)

Network namespaces were completed in Linux 2.6.29 (2009) and can be used to virtualize the network stack. Each network namespace contains its own resource properties within /proc/net. Furthermore, a network namespace contains only a loopback interface on initial creation. Let’s create one:

> sudo unshare -n
# ip l
1: lo: <LOOPBACK> mtu 65536 qdisc noop state DOWN mode DEFAULT group default qlen 1000
    link/loopback 00:00:00:00:00:00 brd 00:00:00:00:00:00

Every network interface (physical or virtual) is present exactly once per namespace. It is possible that an interface will be moved between namespaces. Each namespace contains a private set of IP addresses, its own routing table, socket listing, connection tracking table, firewall, and other network-related resources.

Destroying a network namespace destroys any virtual and moves any physical interfaces within it back to the initial network namespace.

A possible use case for the network namespace is creating Software Defined Networks (SDN) via virtual Ethernet (veth) interface pairs. One end of the network pair will be plugged into a bridged interface whereas the other end will be assigned to the target container. This is how pod networks like flannel work in general.

Let’s see how it works. First, we need to create a new network namespace, which can be done via ip, too:

> sudo ip netns add mynet
> sudo ip netns list

So we created a new network namespace called mynet. When ip creates a network namespace, it will create a bind mount for it under /var/run/netns too. This allows the namespace to persist even when no processes are running within it.

With ip netns exec we can inspect and manipulate our network namespace even further:

> sudo ip netns exec mynet ip l
1: lo: <LOOPBACK> mtu 65536 qdisc noop state DOWN mode DEFAULT group default qlen 1000
    link/loopback 00:00:00:00:00:00 brd 00:00:00:00:00:00
> sudo ip netns exec mynet ping
connect: Network is unreachable

The network seems down, let’s bring it up:

> sudo ip netns exec mynet ip link set dev lo up
> sudo ip netns exec mynet ping
PING ( 56(84) bytes of data.
64 bytes from icmp_seq=1 ttl=64 time=0.016 ms

Hooray! Now let’s create a veth pair which should allow communication later on:

> sudo ip link add veth0 type veth peer name veth1
> sudo ip link show type veth
11: veth1@veth0: <BROADCAST,MULTICAST,M-DOWN> mtu 1500 qdisc noop state DOWN mode DEFAULT group default qlen 1000
    link/ether b2:d1:fc:31:9c:d3 brd ff:ff:ff:ff:ff:ff
12: veth0@veth1: <BROADCAST,MULTICAST,M-DOWN> mtu 1500 qdisc noop state DOWN mode DEFAULT group default qlen 1000
    link/ether ca:0f:37:18:76:52 brd ff:ff:ff:ff:ff:ff

Both interfaces are automatically connected, which means that packets sent to veth0 will be received by veth1 and vice versa. Now we associate one end of the veth pair to our network namespace:

> sudo ip link set veth1 netns mynet
> ip link show type veth
12: veth0@if11: <BROADCAST,MULTICAST> mtu 1500 qdisc noop state DOWN mode DEFAULT group default qlen 1000
    link/ether ca:0f:37:18:76:52 brd ff:ff:ff:ff:ff:ff link-netns mynet

Our network interfaces need some addresses for sure:

> sudo ip netns exec mynet ip addr add dev veth1
> sudo ip netns exec mynet ip link set dev veth1 up
> sudo ip addr add dev veth0
> sudo ip link set dev veth0 up

Communicating in both directions should now be possible:

> ping -c1
PING ( 56(84) bytes of data.
64 bytes from icmp_seq=1 ttl=64 time=0.036 ms

--- ping statistics ---
1 packets transmitted, 1 received, 0% packet loss, time 0ms
rtt min/avg/max/mdev = 0.036/0.036/0.036/0.000 ms
> sudo ip netns exec mynet ping -c1
PING ( 56(84) bytes of data.
64 bytes from icmp_seq=1 ttl=64 time=0.020 ms

--- ping statistics ---
1 packets transmitted, 1 received, 0% packet loss, time 0ms
rtt min/avg/max/mdev = 0.020/0.020/0.020/0.000 ms

It works, but we wouldn’t have any internet access from the network namespace. We would need a network bridge or something similar for that and a default route from the namespace. I leave this task up to you, for now let’s go on to the next namespace.

User ID (user)

With Linux 3.5 (2012) the isolation of user and group IDs was finally possible via namespaces. Linux 3.8 (2013) made it possible to create user namespaces even without being actually privileged. The user namespace enables that a user and group IDs of a process can be different inside and outside of the namespace. An interesting use-case is that a process can have a normal unprivileged user ID outside a user namespace while being fully privileged inside.

Let’s give it a try:

> id -u
> unshare -U
> whoami

After the namespace creation, the files /proc/$PID/{u,g}id_map expose the mappings for user and group IDs for the PID. These files can be written only once to define the mappings.

In general each line within these files contain a one to one mapping of a range of contiguous user IDs between two user namespaces and could look like this:

> cat /proc/$PID/uid_map
0 1000 1

The example above translates to: With the starting user ID 0 the namespace maps to a range starting at ID 1000. This applies only to the user with the ID 1000, since the defined length is 1.

If now a process tries to access a file, its user and group IDs are mapped into the initial user namespace for the purpose of permission checking. When a process retrieves file user and group IDs (via stat(2)), the IDs are mapped in the opposite direction.

In the unshare example (we did above) we implicitly call getuid(2) before writing an appropriate user mapping, which will result in an unmapped ID. This unmapped ID is automatically converted to the overflow user ID (65534 or the value in /proc/sys/kernel/overflow{g,u}id).

The file /proc/$PID/setgroups contains either allow or deny to enable or disable the permission to call thesetgroups(2) syscall within the user namespace. The file was added to address an added security issue introduced with the user namespace: It would be possible to an unprivileged process to create a new namespace in which the user had all privileges. This formerly unprivileged user would be able to drop groups via setgroups(2) to gain access to files he previously not had.

In the end the user namespace enables great security additions to the container world, which are essential for running rootless containers.

Control Group (cgroup)

Cgroups started their journey 2008 with Linux 2.6.24 as dedicated Linux kernel feature. The main goal of cgroups is to support resource limiting, prioritization, accounting and controlling. A major redesign started with version 2 in 2013, whereas the cgroup namespace was added with Linux 4.6 (2016) to prevent leaking host information into a namespace. The second version of cgroups were released there too and major features were added since then. One latest example is an Out-of-Memory (OOM) killer which adds an ability to kill a cgroup as a single unit to guarantee the overall integrity of the workload.

Let’s play around with cgroups and create a new one. By default, the kernel exposes cgroups in /sys/fs/cgroup. To create a new cgroup, we simply create a new sub-directory on that location:

> sudo mkdir /sys/fs/cgroup/memory/demo
> ls /sys/fs/cgroup/memory/demo

You can see that there are already some default values exposed there. Now, we are able to set the memory limits for that cgroup. We are also turning off swap to make our example implementation work.

> sudo su
# echo 100000000 > /sys/fs/cgroup/memory/demo/memory.limit_in_bytes
# echo 0 > /sys/fs/cgroup/memory/demo/memory.swappiness

To assign a process to a cgroup we can write the corresponding PID to the cgroup.procs file:

# echo $$ > /sys/fs/cgroup/memory/demo/cgroup.procs

Now we can execute a sample application to consume more than the allowed 100 megabytes of memory. The application I used is written in Rust and looks like this:

pub fn main() {
    let mut vec = vec![];
    loop {
        vec.extend_from_slice(&[1u8; 10_000_000]);
        println!("{}0 MB", vec.len() / 10_000_000);

If we run the program, we see that the PID will be killed because of the set memory constraints. So our host system is still usable.

# rustc
# ./memory
10 MB
20 MB
30 MB
40 MB
50 MB
60 MB
70 MB
80 MB
90 MB

Composing Namespaces

Namespaces are composable, too! This reveals their true power and makes it possible to have isolated pid namespaces which share the same network interface, like it is done in Kubernetes Pods.

To demonstrate this, let’s create a new namespace with an isolated PID:

> sudo unshare -fp --mount-proc
# ps aux
root         1  0.1  0.6  18688  6904 pts/0    S    23:36   0:00 -bash
root        39  0.0  0.1  35480  1836 pts/0    R+   23:36   0:00 ps aux

The setns(2) syscall with its appropriate wrapper program nsenter can now be used to join the namespace. For this we have to find out which namespace we want to join:

> export PID=$(pgrep -u root bash)
> sudo ls -l /proc/$PID/ns

Now, it is easily possible to join the namespace via nsenter:

> sudo nsenter --pid=/proc/$PID/ns/pid unshare --mount-proc
# ps aux
root         1  0.1  0.0  10804  8840 pts/1    S+   14:25   0:00 -bash
root        48  3.9  0.0  10804  8796 pts/3    S    14:26   0:00 -bash
root        88  0.0  0.0   7700  3760 pts/3    R+   14:26   0:00 ps aux

We can now see that we are member of the same PID namespace! It is also possible to enter already running containers via nsenter, but this topic will be covered later on.

Demo Application

A small demo application can be used to create a simple isolated environment via the namespace API:

#define _GNU_SOURCE
#include <errno.h>
#include <sched.h>
#include <stdio.h>
#include <string.h>
#include <sys/mount.h>
#include <sys/msg.h>
#include <sys/stat.h>
#include <sys/types.h>
#include <sys/wait.h>
#include <unistd.h>

#define STACKSIZE (1024 * 1024)
static char stack[STACKSIZE];

void print_err(char const * const reason)
    fprintf(stderr, "Error %s: %s\n", reason, strerror(errno));

int exec(void * args)
    // Remount proc
    if (mount("proc", "/proc", "proc", 0, "") != 0) {
        print_err("mounting proc");
        return 1;

    // Set a new hostname
    char const * const hostname = "new-hostname";
    if (sethostname(hostname, strlen(hostname)) != 0) {
        print_err("setting hostname");
        return 1;

    // Create a message queue
    key_t key = {0};
    if (msgget(key, IPC_CREAT) == -1) {
        print_err("creating message queue");
        return 1;

    // Execute the given command
    char ** const argv = args;
    if (execvp(argv[0], argv) != 0) {
        print_err("executing command");
        return 1;

    return 0;

int main(int argc, char ** argv)
    // Provide some feedback about the usage
    if (argc < 2) {
        fprintf(stderr, "No command specified\n");
        return 1;

    // Namespace flags
                      CLONE_NEWPID | CLONE_NEWUSER | SIGCHLD;

    // Create a new child process
    pid_t pid = clone(exec, stack + STACKSIZE, flags, &argv[1]);

    if (pid < 0) {
        print_err("calling clone");
        return 1;

    // Wait for the process to finish
    int status = 0;
    if (waitpid(pid, &status, 0) == -1) {
        print_err("waiting for pid");
        return 1;

    // Return the exit code
    return WEXITSTATUS(status);

Purpose of the application is to spawn a new child process in different namespaces. Every command provided to the executable will be forwarded to the new child process. The application terminates, when the command execution is done. You can test and verify the implementation via:

> gcc -o namespaces namespaces.c
> ./namespaces ip a
1: lo: <LOOPBACK> mtu 65536 qdisc noop state DOWN group default qlen 1000
    link/loopback 00:00:00:00:00:00 brd 00:00:00:00:00:00
> ./namespaces ps aux
nobody       1  0.0  0.1  36524  1828 pts/0    R+   23:46   0:00 ps aux
> ./namespaces whoami

This is truly not a working container, but it should give you a slight feeling about how container runtimes leverage namespaces to manage containers. Feel free to use this example as a starting point for your own little experiments with the namespace API.

Putting it all Together

Do you remember the rootfs we extracted from the image within the chroot section? We can use a low level container runtime like runc to easily run a container from the rootfs:

> sudo runc run -b bundle container

If we now inspect the system namespaces, we see that runc already created mnt, uts, ipc, pid and net for us:

> sudo lsns | grep bash
4026532499 mnt         1  6409 root   /bin/bash
4026532500 uts         1  6409 root   /bin/bash
4026532504 ipc         1  6409 root   /bin/bash
4026532505 pid         1  6409 root   /bin/bash
4026532511 net         1  6409 root   /bin/bash

I will stop here and we will learn more about container runtimes and what they do, in upcoming blog posts and talks.


I really hope you enjoyed the read and that the mysteries about containers are now a little bit more fathomable. If you run Linux it is easy to play around with different isolation techniques from scratch. In the end a container runtime nicely uses all these isolation features on different abstraction levels to provide a stable and robust development and production platform for containers.

There are lots of topics which were not covered here because I wanted to stay at a stable level of detail. For sure, a great resource for digging deeper into the topic of Linux namespaces is the Linux programmers manual: NAMESPACES(7).

Feel free to drop me a line or get in contact with me for any questions or feedback. The next blog posts will cover container runtimes, security and the overall ecosystem around latest container technologies. Stay tuned!

You can find all necessary resources about this series on GitHub.