This document shows how to build Fossil into OCI compatible containers and how to use those containers in interesting ways. We start off using the original and still most popular container development and runtime platform, Docker, but since you have more options than that, we will show some of these options later on.
1. Quick Start
Fossil ships a Dockerfile
at the top of its source tree,
here, which you can build like so:
$ docker build -t fossil .
If the image built successfully, you can create a container from it and test that it runs:
$ docker run --name fossil -p 9999:8080/tcp fossil
This shows us remapping the internal TCP listening port as 9999 on the
host. This feature of OCI runtimes means there’s little point to using
the “fossil server --port
” feature inside the container. We can let
Fossil default to 8080 internally, then remap it to wherever we want it
on the host instead.
Our stock Dockerfile
configures Fossil with the default feature set,
so you may wish to modify the Dockerfile
to add configuration options,
add APK packages to support those options, and so forth.
The Fossil Makefile
provides two convenience targets,
“make container-image
” and “make container-run
”. The first creates a
versioned container image, and the second does that and then launches a
fresh container based on that image. You can pass extra arguments to the
first command via the Makefile’s DBFLAGS
variable and to the second
with the DCFLAGS
variable. (DB is short for “docker build
”, and DC
is short for “docker create
”, a sub-step of the “run” target.)
To get the custom port setting as in
second command above, say:
$ make container-run DCFLAGS='-p 9999:8080/tcp'
Contrast the raw “docker
” commands above, which create an
unversioned image called fossil:latest
and from that a container
simply called fossil
. The unversioned names are more convenient for
interactive use, while the versioned ones are good for CI/CD type
applications since they avoid a conflict with past versions; it lets you
keep old containers around for quick roll-backs while replacing them
with fresh ones.
2. Repository Storage Options
If you want the container to serve an existing repository, there are at least two right ways to do it.
The wrong way is to use the Dockerfile COPY
command, because by baking
the repo into the image at build time, it will become one of the image’s
base layers. The end result is that each time you build a container from
that image, the repo will be reset to its build-time state. Worse,
restarting the container will do the same thing, since the base image
layers are immutable. This is almost certainly not what you
want.
The correct ways put the repo into the container created from the image, not in the image itself.
2.1 Storing the Repo Inside the Container
The simplest method is to stop the container if it was running, then say:
$ docker cp /path/to/my-project.fossil fossil:/museum/repo.fossil
$ docker start fossil
$ docker exec fossil chown -R 499 /museum
That copies the local Fossil repo into the container where the server expects to find it, so that the “start” command causes it to serve from that copied-in file instead. Since it lives atop the immutable base layers, it persists as part of the container proper, surviving restarts.
Notice that the copy command changes the name of the repository
database. The container configuration expects it to be called
repo.fossil
, which it almost certainly was not out on the host system.
This is because there is only one repository inside this container, so
we don’t have to name it after the project it contains, as is
traditional. A generic name lets us hard-code the server start command.
If you skip the “chown” command above and put “http://localhost:9999/
”
into your browser, expecting to see the copied-in repo’s home page, you
will get an opaque “Not Found” error. This is because the user and group
ID of the file will be that of your local user on the container’s host
machine, which is unlikely to map to anything in the container’s
/etc/passwd
and /etc/group
files, effectively preventing the server
from reading the copied-in repository file. 499 is the default “fossil
”
user ID inside the container, causing Fossil to run with that user’s
privileges after it enters the chroot. (See below for how to
change this default.) You don’t have to restart the server after fixing
this with chmod
: simply reload the browser, and Fossil will try again.
2.2 Storing the Repo Outside the Container
The simple storage method above has a problem: containers are designed to be killed off at the slightest cause, rebuilt, and redeployed. If you do that with the repo inside the container, it gets destroyed, too. The solution is to replace the “run” command above with the following:
$ docker run \
--publish 9999:8080 \
--name fossil-bind-mount \
--volume ~/museum:/museum \
fossil
Because this bind mount maps a host-side directory (~/museum
) into the
container, you don’t need to docker cp
the repo into the container at
all. It still expects to find the repository as repo.fossil
under that
directory, but now both the host and the container can see that repo DB.
Instead of a bind mount, you could instead set up a separate
volume, at which point you
would need to docker cp
the repo file into the container.
Either way, files in these mounted directories have a lifetime
independent of the container(s) they’re mounted into. When you need to
rebuild the container or its underlying image — such as to upgrade to a
newer version of Fossil — the external directory remains behind and gets
remapped into the new container when you recreate it with --volume/-v
.
2.2.1 WAL Mode Interactions
You might be aware that OCI containers allow mapping a single file into the repository rather than a whole directory. Since Fossil repositories are specially-formatted SQLite databases, you might be wondering why we don’t say things like:
--volume ~/museum/my-project.fossil:/museum/repo.fossil
That lets us have a convenient file name for the project outside the
container while letting the configuration inside the container refer to
the generic “/museum/repo.fossil
” name. Why should we have to name
the repo generically on the outside merely to placate the container?
The reason is, you might be serving that repo with WAL mode
enabled. If you map the repo DB alone into the container, the Fossil
instance inside the container will write the -journal
and -wal
files
alongside the mapped-in repository inside the container. That’s fine as
far as it goes, but if you then try using the same repo DB from outside
the container while there’s an active WAL, the Fossil instance outside
won’t know about it. It will think it needs to write its own
-journal
and -wal
files outside the container, creating a high
risk of database corruption.
If we map a whole directory, both sides see the same set of WAL files. Testing gives us a reasonable level of confidence that using WAL across a container boundary is safe when used in this manner.
3. Security
3.1 Why Not Chroot?
Prior to 2023.03.26, the stock Fossil container relied on the chroot jail feature to wall away the shell and other tools provided by BusyBox. It included that as a bare-bones operating system inside the container on the off chance that someone might need it for debugging, but the thing is, Fossil is self-contained, needing none of that power in the main-line use cases.
Our weak “you might need it” justification collapsed when we realized
you could restore this basic shell environment with a one-line change to
the Dockerfile
, as shown below.
3.2 Dropping Unnecessary Capabilities
The example commands above create the container with a default set of
Linux kernel capabilities. Although Docker strips away almost
all of the traditional root capabilities by default, and Fossil doesn’t
need any of those it does take away, Docker does leave some enabled that
Fossil doesn’t actually need. You can tighten the scope of capabilities
by adding “--cap-drop
” options to your container creation commands.
Specifically:
AUDIT_WRITE
: Fossil doesn’t write to the kernel’s auditing log, and we can’t see any reason you’d want to be able to do that as an administrator shelled into the container, either. Auditing is something done on the host, not from inside each individual container.CHOWN
: The Fossil server never even callschown(2)
, and our image build process sets up all file ownership properly, to the extent that this is possible under the limitations of our automation.Curiously, stripping this capability doesn’t affect your ability to run commands like “
chown -R fossil:fossil /museum
” when you’re using bind mounts or external volumes — as we recommend above — because it’s the host OS’s kernel capabilities that affect the underlyingchown(2)
call in that case, not those of the container.If for some reason you did have to change file ownership of in-container files, it’s best to do that by changing the
Dockerfile
to suit, then rebuilding the container, since that bakes the need for the change into your reproducible build process. If you had to do it without rebuilding the container, there’s a workaround for the fact that capabilities are a create-time change, baked semi-indelibly into the container configuration.FSETID
: Fossil doesn’t use the SUID and SGID bits itself, and our build process doesn’t set those flags on any of the files. Although the second fact means we can’t see any harm from leaving this enabled, we also can’t see any good reason to allow it, so we strip it.KILL
: The only place Fossil callskill(2)
is in the backoffice, and then only for processes it created on earlier runs; it doesn’t need the ability to kill processes created by other users. You might wish for this ability as an administrator shelled into the container, but you can pass the “docker exec --user
” option to run commands within your container as the legitimate owner of the process, removing the need for this capability.MKNOD
: As of 2023.03.26, the stock container uses the runtime’s default/dev
node tree. Prior to this, we had to create/dev/null
and/dev/urandom
inside the chroot jail, but even then, these device nodes were created at build time and were never changed at run time, so we didn’t need this run-time capability even then.NET_BIND_SERVICE
: With containerized deployment, Fossil never needs the ability to bind the server to low-numbered TCP ports, not even if you’re running the server in production with TLS enabled and want the service bound to port 443. It’s perfectly fine to let the Fossil instance inside the container bind to its default port (8080) because you can rebind it on the host with the “docker create --publish 443:8080
” option. It’s the container’s host that needs this ability, not the container itself.(Even the container runtime might not need that capability if you’re terminating TLS with a front-end proxy. You’re more likely to say something like “
-p localhost:12345:8080
” and then configure the reverse proxy to translate external HTTPS calls into HTTP directed at this internal port 12345.)NET_RAW
: Fossil itself doesn’t use raw sockets, and while you could swap out the run layer for something more functional that does make use of raw sockets, there’s little call for it. The best reason I can come up with is to be able to run utilities likeping
andtraceroute
, but since we aren’t doing anything clever with the networking configuration, there’s no particularly compelling reason to run these from inside the container. If you need to ping something, do it on the host.If we did not take this hard-line stance, an attacker that broke into the container and gained root privileges might use raw sockets to do a wide array of bad things to any network the container is bound to.
SETFCAP, SETPCAP
: There isn’t much call for file permission granularity beyond the classic Unix ones inside the container, so we drop root’s ability to change them.
All together, we recommend adding the following options to your
“docker run
” commands, as well as to any “docker create
” command
that will be followed by “docker start
”:
--cap-drop AUDIT_WRITE \
--cap-drop CHOWN \
--cap-drop FSETID \
--cap-drop KILL \
--cap-drop MKNOD \
--cap-drop NET_BIND_SERVICE \
--cap-drop NET_RAW \
--cap-drop SETFCAP \
--cap-drop SETPCAP
In the next section, we’ll show a case where you create a container without ever running it, making these options pointless.
4. Extracting a Static Binary
Our 2-stage build process uses Alpine Linux only as a build host. Once we’ve got everything reduced to a single static Fossil binary, we throw all the rest of it away.
A secondary benefit falls out of this process for free: it’s arguably the easiest way to build a purely static Fossil binary for Linux. Most modern Linux distros make this surprisingly difficult, but Alpine’s back-to-basics nature makes static builds work the way they used to, back in the day. If that’s all you’re after, you can do so as easily as this:
$ docker build -t fossil .
$ docker create --name fossil-static-tmp fossil
$ docker cp fossil-static-tmp:/bin/fossil .
$ docker container rm fossil-static-tmp
The result is six or seven megs, depending on the CPU architecture you build for. It’s built stripped.
5. Customization Points
5.1 Fossil Version
The default version of Fossil fetched in the build is the version in the checkout directory at the time you run it. You could override it to get a release build like so:
$ docker build -t fossil --build-arg FSLVER=version-2.20 .
Or equivalently, using Fossil’s Makefile
convenience target:
$ make container-image DBFLAGS='--build-arg FSLVER=version-2.20'
While you could instead use the generic
“release
” tag here, it’s better to use a specific version number
since container builders cache downloaded files, hoping to
reuse them across builds. If you ask for “release
” before a new
version is tagged and then immediately after, you might expect to get
two different tarballs, but because the underlying source tarball URL
remains the same when you do that, you’ll end up reusing the
old tarball from cache. This will occur
even if you pass the “docker build --no-cache
” option.
This is why we default to pulling the Fossil tarball by checkin ID
rather than let it default to the generic “trunk
” tag: so the URL will
change each time you update your Fossil source tree, forcing the builder to
pull a fresh tarball.
5.2 User & Group IDs
The “fossil
” user and group IDs inside the container default to 499.
Why? Regular user IDs start at 500 or 1000 on most Unix type systems,
leaving those below it for system users like this Fossil daemon owner.
Since it’s typical for these to start at 0 and go upward, we started at
500 and went down one instead to reduce the chance of a conflict to as
close to zero as we can manage.
To change it to something else, say:
$ make container-image DBFLAGS='--build-arg UID=501'
This is particularly useful if you’re putting your repository on a
separate volume since the IDs “leak” out into the host environment via
file permissions. You may therefore wish them to mean something on both
sides of the container barrier rather than have “499” appear on the host
in “ls -l
” output.
5.3 Container Engine
Although the Fossil container build system defaults to Docker, we allow
for use of any OCI container system that implements the same interfaces.
We go into more details about this below, but
for now, it suffices to point out that you can switch to Podman while
using our Makefile
convenience targets unchanged by saying:
$ make CENGINE=podman container-run
5.4 Fossil Configuration Options
You can use this same mechanism to enable non-default Fossil configuration options in your build. For instance, to turn on the JSON API and the TH1 docs extension:
$ make container-image \
DBFLAGS='--build-arg FSLCFG="--json --with-th1-docs"'
If you also wanted the Tcl evaluation extension, that brings us to the next point.
5.5 Elaborating the Run Layer
If you want a basic shell environment for temporary debugging of the
running container, that’s easily added. Simply change this line in the
Dockerfile
…
FROM scratch AS run
…to this:
FROM busybox AS run
Rebuild and redeploy to give your Fossil container a BusyBox-based shell environment that you can get into via:
$ docker exec -it -u fossil $(make container-version) sh
(That command assumes you built it via “make container
” and are
therefore using its versioning scheme.)
Another case where you might need to replace this bare-bones “run
”
layer with something more functional is that you’re setting up email
alerts and need some way to integrate with the host’s
MTA. There are a number of alternatives in that linked document, so
for the sake of discussion, we’ll say you’ve chosen Method
2, which requires a Tcl interpreter and its SQLite
extension to push messages into the outbound email queue DB, presumably
bind-mounted into the container.
You can do that by replacing STAGEs 2 and 3 in the stock Dockerfile
with this:
## ---------------------------------------------------------------------
## STAGE 2: Pare that back to the bare essentials, plus Tcl.
## ---------------------------------------------------------------------
FROM alpine AS run
ARG UID=499
ENV PATH "/sbin:/usr/sbin:/bin:/usr/bin"
COPY --from=builder /tmp/fossil /bin/
COPY tools/email-sender.tcl /bin/
RUN set -x \
&& echo "fossil:x:${UID}:${UID}:User:/museum:/false" >> /etc/passwd \
&& echo "fossil:x:${UID}:fossil" >> /etc/group \
&& install -d -m 700 -o fossil -g fossil log museum \
&& apk add --no-cache tcl sqlite-tcl
Build it and test that it works like so:
$ make container-run &&
echo 'puts [info patchlevel]' |
docker exec -i $(make container-version) tclsh
8.6.12
You should remove the PATH
override in the “RUN”
stage, since it’s written for the case where everything is in /bin
.
With these additions, we need the longer PATH
shown above to have
ready access to them all.
Another useful case to consider is that you’ve installed a server extension and you need an interpreter for that script. The first option above won’t work except in the unlikely case that it’s written for one of the bare-bones script interpreters that BusyBox ships.1
Let’s say the extension is written in Python. While you could handle it the same way we do with the Tcl example above, Python is more popular, giving us more options. Let’s inject a Python environment into the stock Fossil container via a suitable “distroless” image instead:
## ---------------------------------------------------------------------
## STAGE 2: Pare that back to the bare essentials, plus Python.
## ---------------------------------------------------------------------
FROM cgr.dev/chainguard/python:latest
USER root
ARG UID=499
ENV PATH "/sbin:/usr/sbin:/bin:/usr/bin"
COPY --from=builder /tmp/fossil /bin/
COPY --from=builder /bin/busybox.static /bin/busybox
RUN [ "/bin/busybox", "--install", "/bin" ]
RUN set -x \
&& echo "fossil:x:${UID}:${UID}:User:/museum:/false" >> /etc/passwd \
&& echo "fossil:x:${UID}:fossil" >> /etc/group \
&& install -d -m 700 -o fossil -g fossil log museum
You will also have to add busybox-static
to the APK package list in
STAGE 1 for the RUN
script at the end of that stage to work, since the
Chainguard Python image lacks a shell, on purpose. The need to
install root-level binaries is why we change USER
temporarily here.
Build it and test that it works like so:
$ make container-run &&
docker exec -i $(make container-version) python --version
3.11.2
The compensation for the hassle of using Chainguard over something more
general purpose like Alpine + “apk add python
”
is huge: we no longer leave a package manager sitting around inside the
container, waiting for some malefactor to figure out how to abuse it.
Beware that there’s a limit to this über-jail’s ability to save you when you go and provide a more capable OS layer like this. The container layer should stop an attacker from accessing any files out on the host that you haven’t explicitly mounted into the container’s namespace, but it can’t stop them from making outbound network connections or modifying the repo DB inside the container.
6. Lightweight Alternatives to Docker
Those afflicted with sticker shock at seeing the size of a Docker Desktop installation — 1.65 GB here — might’ve immediately “noped” out of the whole concept of containers. The first thing to realize is that when it comes to actually serving simple containers like the ones shown above is that Docker Engine suffices, at about a quarter of the size.
Yet on a small server — say, a $4/month ten gig Digital Ocean droplet — that’s still a big chunk of your storage budget. It takes ~60:1 overhead merely to run a Fossil server container? Once again, I wouldn’t blame you if you noped right on out of here, but if you will be patient, you will find that there are ways to run Fossil inside a container even on entry-level cloud VPSes. These are well-suited to running Fossil; you don’t have to resort to raw Fossil service to succeed, leaving the benefits of containerization to those with bigger budgets.
For the sake of simple examples in this section, we’ll assume you’re integrating Fossil into a larger web site, such as with our Debian + nginx + TLS plan. This is why all of the examples below create the container with this option:
--publish 127.0.0.1:9999:8080
The assumption is that there’s a reverse proxy running somewhere that
redirects public web hits to localhost port 9999, which in turn goes to
port 8080 inside the container. This use of port
publishing effectively replaces the use of the
“fossil server --localhost
” option.
For the nginx case, you need to add --scgi
to these commands, and you
might also need to specify --baseurl
.
Containers are a fine addition to such a scheme as they isolate the Fossil sections of the site from the rest of the back-end resources, thus greatly reducing the chance that they’ll ever be used to break into the host as a whole.
(If you wanted to be double-safe, you could put the web server into another container, restricting it to reading from the static web site directory and connecting across localhost to back-end dynamic content servers such as Fossil. That’s way outside the scope of this document, but you can find ready advice for that elsewhere. Seeing how we do this with Fossil should help you bridge the gap in extending this idea to the rest of your site.)
6.1 Stripping Docker Engine Down
The core of Docker Engine is its containerd
daemon and the
runc
container runtime. Add to this the out-of-core CLI program
nerdctl
and you have enough of the engine to run Fossil
containers. The big things you’re missing are:
BuildKit: The container build engine, which doesn’t matter if you’re building elsewhere and shipping the images to the target. A good example is using a container registry as an intermediary between the build and deployment hosts.
SwarmKit: A powerful yet simple orchestrator for Docker that you probably aren’t using with Fossil anyway.
In exchange, you get a runtime that’s about half the size of Docker
Engine. The commands are essentially the same as above, but you say
“nerdctl
” instead of “docker
”. You might alias one to the other,
because you’re still going to be using Docker to build and ship your
container images.
6.2 Podman
A lighter-weight rootless drop-in replacement that doesn’t give up the image builder is Podman. Initially created by Red Hat and thus popular on that family of OSes, it will run on any flavor of Linux. It can even be made to run on macOS via Homebrew or on Windows via WSL2.
On Ubuntu 22.04, the installation size is about 38 MiB, roughly a tenth the size of Docker Engine.
For our purposes here, the only thing that changes relative to the examples at the top of this document are the initial command:
$ podman build -t fossil .
$ podman run --name fossil -p 9999:8080/tcp fossil
Your Linux package repo may have a podman-docker
package which
provides a “docker
” script that calls “podman
” for you, eliminating
even the command name difference. With that installed, the make
commands above will work with Podman as-is.
The only difference that matters here is that Podman doesn’t have the
same default Linux kernel capability set as Docker, which
affects the --cap-drop
flags recommended above to:
$ podman create \
--name fossil \
--cap-drop CHOWN \
--cap-drop FSETID \
--cap-drop KILL \
--cap-drop NET_BIND_SERVICE \
--cap-drop SETFCAP \
--cap-drop SETPCAP \
--publish 127.0.0.1:9999:8080 \
localhost/fossil
$ podman start fossil
6.3 systemd-container
If even the Podman stack is too big for you, the next-best option I’m
aware of is the systemd-container
infrastructure on modern Linuxes,
available since version 239 or so. Its runtime tooling requires only
about 1.4 MiB of disk space:
$ sudo apt install systemd-container btrfs-tools
That command assumes the primary test environment for
this guide, Ubuntu 22.04 LTS with systemd
249. For best
results, /var/lib/machines
should be a btrfs volume, because
$REASONS
. For CentOS Stream 9 and other Red Hattish
systems, you will have to make several adjustments, which we’ve
collected below to keep these examples clear.
We’ll assume your Fossil repository stores something called
“myproject
” within ~/museum/myproject/repo.fossil
, named according
to the reasons given above. We’ll make consistent use of
this naming scheme in the examples below so that you will be able to
replace the “myproject
” element of the various file and path names.
If you use the stock Dockerfile
to generate your
base image, nspawn
won’t recognize it as containing an OS unless you
change the “FROM scratch AS os
” line at the top of the second stage
to something like this:
FROM gcr.io/distroless/static-debian11 AS os
Using that as a base image provides all the files nspawn
checks for to
determine whether the container is sufficiently close to a Linux VM for
the following step to proceed:
$ make container
$ docker container export $(make container-version) |
machinectl import-tar - myproject
Next, create /etc/systemd/nspawn/myproject.nspawn
:
[Exec]
WorkingDirectory=/
Parameters=bin/fossil server \
--baseurl https://example.com/myproject \
--create \
--jsmode bundled \
--localhost \
--port 9000 \
--scgi \
--user admin \
museum/repo.fossil
DropCapability= \
CAP_AUDIT_WRITE \
CAP_CHOWN \
CAP_FSETID \
CAP_KILL \
CAP_MKNOD \
CAP_NET_BIND_SERVICE \
CAP_NET_RAW \
CAP_SETFCAP \
CAP_SETPCAP
ProcessTwo=yes
LinkJournal=no
Timezone=no
[Files]
Bind=/home/fossil/museum/myproject:/museum
[Network]
VirtualEthernet=no
If you recognize most of that from the Dockerfile
discussion above,
congratulations, you’ve been paying attention. The rest should also
be clear from context.
Some of this is expected to vary:
The references to
example.com
andmyproject
are stand-ins for your actual web site and repository name.The command given in the
Parameters
directive assumes you’re setting up SCGI proxying via nginx, but with adjustment, it’ll work with the other repository service methods we’ve documented.The path in the host-side part of the
Bind
value must point at the directory containing therepo.fossil
file referenced in said command so that/museum/repo.fossil
refers to your repo out on the host for the reasons given above.
That being done, we also need a generic systemd
unit file called
/etc/systemd/system/fossil@.service
, containing:
[Unit]
Description=Fossil %i Repo Service
Wants=modprobe@tun.service modprobe@loop.service
After=network.target systemd-resolved.service modprobe@tun.service modprobe@loop.service
[Service]
ExecStart=systemd-nspawn --settings=override --read-only --machine=%i bin/fossil
[Install]
WantedBy=multi-user.target
You shouldn’t have to change any of this because we’ve given the
--setting=override
flag, meaning any setting in the nspawn file
overrides the setting passed to systemd-nspawn
. This arrangement
not only keeps the unit file simple, it allows multiple services to
share the base configuration, varying on a per-repo level through
adjustments to their individual *.nspawn
files.
You may then start the service in the normal way:
$ sudo systemctl enable fossil@myproject
$ sudo systemctl start fossil@myproject
You should then find it running on localhost port 9000 per the nspawn
configuration file above, suitable for proxying Fossil out to the
public using nginx via SCGI. If you aren’t using a front-end proxy
and want Fossil exposed to the world via HTTPS, you might say this instead in
the *.nspawn
file:
Parameters=bin/fossil server \
--cert /path/to/cert.pem \
--create \
--jsmode bundled \
--port 443 \
--user admin \
museum/repo.fossil
You would also need to un-drop the CAP_NET_BIND_SERVICE
capability
to allow Fossil to bind to this low-numbered port.
We use the systemd
template file feature to allow multiple Fossil
servers running on a single machine, each on a different TCP port,
as when proxying them out as subdirectories of a larger site.
To add another project, you must first clone the base “machine” layer:
$ sudo machinectl clone myproject otherthing
That will not only create a clone of /var/lib/machines/myproject
as ../otherthing
, it will create a matching otherthing.nspawn
file for you
as a copy of the first one. Adjust its contents to suit, then enable
and start it as above.
6.3.1 Getting It Working on a RHEL Clone
The biggest difference between doing this on OSes like CentOS versus
Ubuntu is that RHEL (thus also its clones) doesn’t ship btrfs in
its kernel, thus ships with no package repositories containing mkfs.btrfs
, which
machinectl
depends on for achieving its various purposes.
Fortunately, there are workarounds.
First, the apt install
command above becomes:
$ sudo dnf install systemd-container
Second, you have to hack around the lack of machinectl import-tar
:
$ rootfs=/var/lib/machines/fossil
$ sudo mkdir -p $rootfs
$ docker container export fossil | sudo tar -xf -C $rootfs -
The parent directory path in the rootfs
variable is important,
because although we aren’t able to use machinectl
on such systems, the
systemd-nspawn
developers assume you’re using them together; when you give
--machine
, it assumes the machinectl
directory scheme. You could
instead use --directory
, allowing you to store the rootfs wherever
you like, but why make things difficult? It’s a perfectly sensible
default, consistent with the LHS rules.
The final element — the machine name — can be anything you like so long as it matches the nspawn file’s base name.
Finally, since you can’t use machinectl clone
, you have to make
a wasteful copy of /var/lib/machines/myproject
when standing up
multiple Fossil repo services on a single machine. (This is one
of the reasons machinectl
depends on btrfs
: cheap copy-on-write
subvolumes.) Because we give the --read-only
flag, you can simply
cp -r
one machine to a new name rather than go through the
export-and-import dance you used to create the first one.
6.3.2 What Am I Missing Out On?
For all the runtime size savings in this method, you may be wondering what you’re missing out on relative to Podman, which takes up roughly 27× more disk space. Short answer: lots. Long answer:
Build system. You’ll have to build and test your containers some other way. This method is only suitable for running them once they’re built.
Orchestration. All of the higher-level things like “compose” files, Docker Swarm mode, and Kubernetes are unavailable to you at this level. You can run multiple instances of Fossil, but on a single machine only and with a static configuration.
Image layer sharing. When you update an image using one of the above methods, Docker and Podman are smart enough to copy only changed layers. Furthermore, when you base multiple containers on a single image, they don’t make copies of the base layers; they can share them, because base layers are immutable, thus cannot cross-contaminate.
Because we use
systemd-nspawn --read-only
, we get some of this benefit, particularly when usingmachinectl
with/var/lib/machines
as a btrfs volume. Even so, the disk space and network I/O optimizations go deeper in the Docker and Podman worlds.Tooling. Hand-creating and modifying those
systemd
files sucks compared to “podman container create ...
” This is but one of many affordances you will find in the runtimes aimed at daily-use devops warriors.Network virtualization. In the scheme above, we turn off the
systemd
private networking support because in its default mode, it wants to hide containerized services entirely. While there are ways to expose Fossil’s single network service port under that scheme, it adds a lot of administration complexity. In the big-boy container runtimes,docker create --publish
fixes all this up in a single option, whereassystemd-nspawn --port
does approximately none of that despite the command’s superficial similarity.From a purely functional point of view, this isn’t a huge problem if you consider the inbound service direction only, being external connections to the Fossil service we’re providing. Since we do want this Fossil service to be exposed — else why are we running it? — we get all the control we need via
fossil server --localhost
and similar options.The complexity of the
systemd
networking infrastructure’s interactions with containers make more sense when you consider the outbound path. Consider what happens if you enable Fossil’s optional TH1 docs feature plus its Tcl evaluation feature. That would enable anyone with the rights to commit to your repository the ability to make arbitrary network connections on the Fossil host. Then, let us say you have a client-server DBMS server on that same host, bound to localhost for private use by other services on the machine. Now that DBMS is open to access by a rogue Fossil committer because the host’s loopback interface is mapped directly into the container’s network namespace.Proper network virtualization would protect you in this instance.
This author expects that the set of considerations is broader than
presented here, but that it suffices to make our case as it is: if you
can afford the space of Podman or Docker, we strongly recommend using
either of them over the much lower-level systemd-container
infrastructure. You’re getting a considerable amount of value for the
higher runtime cost; it isn’t pointless overhead.
(Incidentally, these are essentially the same reasons why we no longer
talk about the crun
tool underpinning Podman in this document. It’s
even more limited than nspawn
, making it even more difficult to administer while
providing no runtime size advantage. The runc
tool underpinning
Docker is even worse on this score, being scarcely easier to use than
crun
while having a much larger footprint.)
6.3.3 Violated Assumptions
The systemd-container
infrastructure has a bunch of hard-coded
assumptions baked into it. We papered over these problems above,
but if you’re using these tools for other purposes on the machine
you’re serving Fossil from, you may need to know which assumptions
our container violates and the resulting consequences.
Some of it we discussed above already, but there’s one big class of problems we haven’t covered yet. It stems from the fact that our stock container starts a single static executable inside a bare-bones container rather than “boot” an OS image. That causes a bunch of commands to fail:
machinectl poweroff
will fail because the container isn’t running dbus.machinectl start
will try to find an/sbin/init
program in the rootfs, which we haven’t got. We could rename/bin/fossil
to/sbin/init
and then hack the chroot scheme to match, but ick. (This, incidentally, is why we setProcessTwo=yes
above even though Fossil is perfectly capable of running as PID 1, a fact we depend on in the other methods above.)machinectl shell
will fail because there is no login daemon running, which we purposefully avoided adding by creating a “FROM scratch
” container. (If you need a shell, say:sudo systemd-nspawn --machine=myproject /bin/sh
)machinectl status
won’t give you the container logs because we disabled the shared journal, which was in turn necessary because we don’t runsystemd
inside the container, just outside.
If these are problems for you, you may wish to build a
fatter container using debootstrap
or similar. (External
tutorial.)