READMEs/README-systemd-nspawn.md

Setting up and administering systemd-nspawn containers

Sai's approach is to run sai-builder from inside the environment it will do builds from. It reaches out to the sai-server instance and git repos, and otherwise there is no ingress into the container or VM from outside.

So with systemd-nspawn, it means the containers should start on the host at boot and run sai-builder inside eg from a systemd service file. You have to:

  • one-time create the containers you want to build in - set them up with a sai user and minimal packageset - build libwebsockets / sai in there - configure sai-builder to run from systemd when the container starts

Systemd only runs on Linux, so this document is about how to do all those steps for Linux hosts and Linux guests. For Windows and OSX, the approach is similar but using an actual machine or qemu / kvm VM.

Create the container on the host

This will create, furnish and enter a container for Ubuntu Xenial (2016-era long-term stable). Other distros are basically the same with minor changes

surprisedRestrictions! Systemd with create a veth device based off your machine name chosen here, it's limited to 11 chars. So two machines here must not have the same first 11 chars in their name!

surprisedRestrictions! Systemd won't let your machine name have an underscore, so no ...x86_64. It's OK with -.

Pick your machine name, in this example xenial-amd64.

Step 1: Create the initial host config for the machine

On the host, mkdir /etc/systemd/nspawn and then create in /etc/systemd/nspawn/xenial-amd64.nspawn

[Exec]
Boot=no
[Network]
VirtualEthernet=no

We will come back and set both of those to yes after we have set the container up manually from the inside.

Step 2: Create the machine

Host # cd /var/lib/machines
Host # debootstrap --arch=amd64 --variant=buildd xenial xenial-amd64 http://archive.ubuntu.com/ubuntu/
Host # systemd-nspawn -M xenial-amd64
Spawning container xenial-amd64 on /var/lib/machines/xenial-amd64.
Press ^] three times within 1s to kill container.
root@myhost:~#

ah that's pretty nice so far, right! We have a shell up inside the container already.

Variations: Fedora Container

Replace the debootstrap part above with

Host # dnf -y --releasever=32 --installroot=`pwd`/fedora32-x8664 install systemd passwd dnf fedora-release

Variations: Debian Container

Replace the debootstrap part above with

Host # debootstrap --arch=amd64 stable buster-amd64

Variations: Centos 8

Create a file "centos8.repo"

[centos8-base]
name=CentOS-8-Base
baseurl=http://mirror.centos.org/centos-8/8.1.1911/BaseOS/x86_64/os
gpgcheck=1

... then...

$ dnf -c centos8.repo --releasever=8 --disablerepo=* --enablerepo=centos8-base --installroot=/var/lib/machines/centos8 groups install 'Minimal Install'

Step 3: Install things into the container

While we're borrowing the host network, install things we want in the container to configure and debug it.

(You might think you can install these at the bootstrap step, but I found this can make trouble both in the right deps not being brought in and it violating the ownership of /usr/lib/.build-id with real root instead of the fake root of the inside of the container.)

Create the sai user and home dir

root@myhost:~# useradd -u883 -gnogroup sai -d/home/sai -m -r

For redhat type systems, use -gnobody instead For BSDs, useradd -u883 -g32766 -d/home/sai -m sai

Specific to Debian Buster / Sid

If a 32-bit install, leave out libc6-i386

Container # apt install sudo make git gcc cmake net-tools libssl-dev vim libc6-i386 dbus systemd-container libmbedtls-dev libev4 libev-dev libuv1 libuv1-dev libmount-dev libsqlite0-dev libgit2-dev inetutils-ping psmisc libgit2-27 libmbedcrypto3 libmbedtls12 libmbedx509-0 libmbedtls-dev libevent-2.1-6 libevent-dev libglib2.0-0 libsqlite3-0 libsqlite3-dev pkg-config libglib2.0-dev libglib2.0-0 libdbus-1-dev libdbus-1-3 g++

Specific to Debian 11

Container # apt install sudo make git gcc cmake net-tools libssl-dev vim libc6-i386 dbus systemd-container libmbedtls-dev libev4 libev-dev libuv1 libuv1-dev libmount-dev libsqlite3-dev libgit2-dev inetutils-ping psmisc libgit2-1.1 libmbedcrypto3 libmbedtls12 libmbedx509-0 libmbedtls-dev libevent-2.1-7 libevent-dev libglib2.0-0 libsqlite3-0 libsqlite3-dev pkg-config libglib2.0-dev libglib2.0-0 libdbus-1-dev libdbus-1-3 g++ file

Specific to Ubuntu

We have to enable the "universe" packageset to get normal things.

Container # apt update && apt install vim && apt edit-sources && apt update

add " universe" to the end of the line in the vim session it pops up... we also need to enable updates so we can maintain the distro image, so add lines similar to this

deb http://archive.ubuntu.com/ubuntu focal-security main universe
deb http://archive.ubuntu.com/ubuntu focal-updates main universe

and save it.

Specific to Xenial

Container # apt install sudo make git gcc cmake net-tools libssl-dev vim libc6-i386 dbus systemd-container libmbedtls-dev glib2.0-dev libevent-dev libev4 libev-dev libuv1 libuv1-dev libmount-dev libsqlite3-dev libgit2-dev inetutils-ping psmisc libsqlite-dev sqlite3 libsqlite3-0 libgit2-24 libgit2-dev libmbedcrypto0 libmbedtls10 libmbedx509-0 libmbedtls-dev libev4 libev-dev libevent-2.0 libevent-dev glib-2.0 pkg-config

Specific to Bionic

Container # apt install sudo make git gcc cmake net-tools libssl-dev vim libc6-i386 dbus systemd-container libmbedtls-dev libev4 libev-dev libuv1 libuv1-dev libmount-dev libsqlite0-dev libgit2-dev inetutils-ping psmisc libgit2-26 libmbedcrypto1 libmbedtls10 libmbedx509-0 libmbedtls-dev libevent-2.1-6 libevent-dev libglib2.0-0 libsqlite3-0 libsqlite3-dev pkg-config libglib2.0-dev libglib2.0-0 libdbus-1-dev libdbus-1-3

Specific to Focal

Container # apt install sudo make git gcc cmake net-tools libssl-dev vim libc6-i386 dbus systemd-container libmbedtls-dev libevent-2.1-7 libev4 libev-dev libuv1 libuv1-dev libmount-dev libsqlite0-dev libgit2-dev inetutils-ping psmisc libgit2-28 libmbedcrypto3 libmbedtls12 libmbedx509-0 libmbedtls-dev libev4 libev-dev libevent-2.1-7 libevent-dev libglib2.0-0 libsqlite3-0 libsqlite3-dev pkg-config libglib2.0-dev libglib2.0-0

Specific to Fedora

Container # dnf install sudo make git gcc gcc-c++ cmake net-tools dhclient openssl-devel vim glibc.i686 dbus mbedtls-devel glib2-devel libevent-devel libev-devel libuv-devel libmount-devel sqlite-devel libgit2 libgit2-devel dbus systemd-container vim libpkgconf procps-ng

Specific to Centos 7

Container #  dnf install sudo make git gcc gcc-c++ cmake net-tools dhclient openssl-devel vim glibc.i686 dbus glib2-devel libevent-devel libev-devel sqlite-devel libgit2 dbus systemd-container vim libpkgconf rpm-build systemd-networkd systemd-resolved

Specific to Centos 8

Centos / RHEL 8 simply doesn't provide some -devel packages like mbedtls, libgit2... most missing -devel packages can be found in the powertools repo

Container #  dnf install sudo make git gcc gcc-c++ cmake net-tools dhclient openssl-devel vim glibc.i686 dbus glib2-devel libevent-devel libev-devel sqlite-devel libgit2 dbus systemd-container vim libpkgconf rpm-build

Specific to Gentoo

Container # emerge dev-db/sqlite libgit2 gcc net-libs/mbedtls dev-libs/openssl cmake libuv libevent dev-libs/glib dev-libs/libev dbus pkgconf libcap

Specific to BSD type unix

There are some assumptions in sai like the scripts will run by /bin/bash and cmake should be on Sai's PATH, it means you may need to symlink some things that don't match to /bin paths, eg, ln -sf /usr/local/bin/cmake /bin/cmake

Step 3: Configure the container private networking

Set up a hostname for the container by editing its /etc/hostname, ideally to a name like machinename-hostname, eg, xenial-myhost if your host is myhost.

Configure the container's systemd-networkd to set up the private network for the container we will switch to next.

In the container, create /etc/systemd/network/50-host0.network with the contents

[Match]
Name=host0
[Network]
DHCP=no
IPForward=1
Gateway=192.168.80.1
Address=192.168.80.9/24

If you have multiple containers, give each one a different 192.168.xx network. Enable systemd pieces to control setting the interface, route and dns.

You can avoid systemd-resolved if you like, rm /etc/resolv.conf to remove its symlink, and then something like echo "nameserver 1.2.3.4" > /etc/resolv.conf Either way at least enable systemd-networkd

Container # ln -sf /run/systemd/resolve/stub-resolv.conf /etc/resolv.conf
Container # systemctl enable systemd-networkd systemd-resolved

Specific to Centos 8

Centos 8 has no systemd-networkd. The base install instead has nmcli, you can set up an ipv4.method=static solution that way using the same network layout as above.

Step 4: Switch over to systemd management of the container

Exit the container session.

On the host, edit /etc/systemd/nspawn/xenial-amd64.nspawn we created earlier and set both Boot and VirtualEthernet to yes now.

Tell the host we want this container up when the host is up, and that we want to start it now.

Host # systemctl enable systemd-nspawn@xenial-amd64
Host # systemctl start systemd-nspawn@xenial-amd64

Then, we can reopen a shell prompt in the now-running container

Host # machinectl -M xenial-amd64 shell
Connected to the local host. Press ^] three times within 1s to exit session.
[root@xenial-amd64 ~]#

You can look around with ps if you have it or systemctl status and see that it was started inside by systemd.

Step 5: Configure the host-side networking

systemd will have created a matching network interface on your host named ve-machinename, where machinename is your (truncated if neccessary) machine name.

We need to explain to Network Manager how to deal with this so the container will get networking services cleanly.

We will call the related NetworkManager connection nm-...

Host # export NMNAME=ve-xenial ; nmcli c add type ethernet autoconnect no con-name nm-$NMNAME ifname $NMNAME ;\
  nmcli c m nm-$NMNAME ipv4.method shared ; \
  nmcli c m nm-$NMNAME ipv4.addresses 192.168.80.1/24 ; \
  nmcli c m nm-$NMNAME connection.zone vm ; \
  nmcli c m nm-$NMNAME connection.autoconnect true ; \
  nmcli c down nm-$NMNAME ; nmcli c up nm-$NMNAME

Step 6: Restart the container

Host # systemctl stop systemd-nspawn@xenial-amd64
Host # systemctl start systemd-nspawn@xenial-amd64
Host # machinectl -M xenial-amd64 shell

If you hit selinux AVCs on the start step, as I did on Fedora 32, you can get around them with three policy additions

Host # echo 'AVC avc:  denied  { search } for  pid=1502 comm="systemd-machine" name="21786" dev="proc" ino=22140 scontext=system_u:system_r:systemd_machined_t:s0 tcontext=system_u:system_r:unconfined_service_t:s0 tclass=dir permissive=0' | audit2allow -M machine-search ; semodule -i machine-search.pp
Host # echo 'AVC avc:  denied  { read } for  pid=1502 comm="systemd-machine" name="cgroup" dev="proc" ino=60301 scontext=system_u:system_r:systemd_machined_t:s0 tcontext=system_u:system_r:unconfined_service_t:s0 tclass=file permissive=0' | audit2allow -M machine-read ; semodule -i machine-read.pp
Host # echo 'AVC avc:  denied  { open } for  pid=1502 comm="systemd-machine" path="/proc/21937/cgroup" dev="proc" ino=46779 scontext=system_u:system_r:systemd_machined_t:s0 tcontext=system_u:system_r:unconfined_service_t:s0 tclass=file permissive=0' | audit2allow -M machine-open ; semodule -i machine-open.pp

In the container, you should see host0 has ben set up

Container # ifconfig host0
host0: flags=4163<UP,BROADCAST,RUNNING,MULTICAST>  mtu 1500
        inet 192.168.80.9  netmask 255.255.255.0  broadcast 192.168.80.255
        inet6 fe80::fc27:1eff:fe53:4d93  prefixlen 64  scopeid 0x20<link>
        ether fe:27:1e:53:4d:93  txqueuelen 1000  (Ethernet)
        RX packets 1117  bytes 102933 (100.5 KiB)
        RX errors 0  dropped 0  overruns 0  frame 0
        TX packets 754  bytes 71331 (69.6 KiB)
        TX errors 0  dropped 0 overruns 0  carrier 0  collisions 0

Whether you can ping things depends on your host routing arrangements for its side of the virtual ethernet. If you pointed the container to your local resolver you need to enable that in the "vm" zone, along with masquerading. You'll only need to do this for the first vm, afterwards it enough to tell nmcli that the connection is in vm zone to inherit the functionality.

Host # firewall-cmd --permanent --new-zone vm
Host # firewall-cmd --permanent --zone vm --add-masquerade
Host # firewall-cmd --permanent --zone vm --add-port 53/udp
Host # firewall-cmd --permanent --zone vm --add-port 53/tcp
Host # systemctl restart firewalld

This might seem complicated but nmcli and firewalld (and systemd) have made all this a lot easier to set up than a few years ago.

Get, build and install lws

We're going to do it as root in the container... I produce and host this code, you may feel its riskier in which case create a new user with sudo rights to handle it.

For redhat type distros, you probably need to add /usr/local/lib to the /etc/ld.so.conf before ldconfig can rgister the new libwebsockets.so

Container # git clone https://libwebsockets.org/repo/libwebsockets
Container # cd libwebsockets && mkdir build && cd build && \
  cmake .. -DLWS_WITH_JOSE=1 -DLWS_UNIX_SOCK=1 -DLWS_WITH_STRUCT_JSON=1 \
   -DLWS_WITH_STRUCT_SQLITE3=1 -DLWS_WITH_GENCRYPTO=1 -DLWS_WITH_SPAWN=1 \
   -DLWS_WITH_SECURE_STREAMS=1 -DLWS_WITH_THREADPOOL=1 \
   -DLWS_WITH_PLUGINS_BUILTIN=1 -DLWS_WITHOUT_DAEMONIZE=0 \
   -DLWS_WITH_SYS_METRICS=1
Container # make -j && make -j install && ldconfig

(You only require LWS_WITHOUT_DAEMONIZE=0 for FreeBSD and OpenBSD rc.d script integration.)

For OSX, you'll need to add something like -DLWS_OPENSSL_INCLUDE_DIRS=/usr/local/opt/openssl@1.1/include -DLWS_OPENSSL_LIBRARIES="/usr/local/opt/openssl/lib/libssl.dylib;/usr/local/opt/openssl/lib/libcrypto.dylib" to the cmake line

For newer M1 based OSX

cmake .. -DLWS_OPENSSL_INCLUDE_DIRS=/opt/homebrew/Cellar/openssl@1.1/1.1.1h/include '-DLWS_OPENSSL_LIBRARIES=/opt/homebrew/Cellar/openssl@1.1/1.1.1h/lib/libssl.dylib;/opt/homebrew/Cellar/openssl@1.1/1.1.1h/lib/libcrypto.dylib' -DLWS_WITH_MINIMAL_EXAMPLES=0 -DLWS_WITH_JOSE=1 -DLWS_UNIX_SOCK=1 -DLWS_WITH_STRUCT_JSON=1 \
    -DLWS_WITH_GENCRYPTO=1 -DLWS_WITH_SPAWN=1 \
   -DLWS_WITH_SECURE_STREAMS=1 -DLWS_WITH_THREADPOOL=1 -DLWS_WITH_JOSE=1 \
   -DLWS_WITH_PLUGINS_BUILTIN=1

Solaris loader paths

Solaris goes its own way to define well-known loader paths, the magic to cover default builds of lws and libgit2 is

# crle -c /var/ld/ld.config -l /lib:/usr/lib:/usr/local/lib:/opt/csw/lib

You also need to build libgit2 similar to as shown below, since it's not packaged.

Windows: build libgit2

For Windows, after following the way to build lws in READMEs/README.build-windows.md in the lws tree, you need to build libgit2

$ git clone https://github.com/libgit2/libgit2
$ cd libgit2 && mkdir build && cd build && cmake .. -DBUILD_CLAR=OFF
$ cmake --build . --config DEBUG

Then, in and Administrator-privs cmd window

$ cmake --install . --config DEBUG

Get, build and install sai

Linux

Container # cd ~
Container # git clone https://warmcat.com/repo/sai
Container # cd sai && mkdir build && cd build && cmake .. && make && make install
Container # cp ../scripts/sai-builder.service /etc/systemd/system
Container # mkdir -p /etc/sai/builder
Container # cp ../scripts/builder-conf /etc/sai/builder/conf
Container # mkdir /var/run/com.warmcat.com.saib.logproxy /var/run/com.warmcat.com.saib.resproxy
Container # chown sai /var/run/com.warmcat.com.saib.logproxy /var/run/com.warmcat.com.saib.resproxy
Container # vim /etc/sai/builder/conf
Container # systemctl enable sai-builder

Windows

$ git clone https://warmcat.com/repo/sai
$ cd sai && mkdir build && cd build
$ cmake .. -DSAI_SERVER=0 -DSAI_LWS_INC_PATH="C:/Program Files (x86)/libwebsockets/include" -DSAI_LWS_LIB_PATH="C:/Program Files (x86)/libwebsockets/lib/websockets.lib"  -DSAI_EXT_PTHREAD_INCLUDE_DIR="C:\Program Files (x86)\pthreads\include" -DSAI_EXT_PTHREAD_LIBRARIES="C:\Program Files (x86)\pthreads\lib\x64\libpthreadGC2.a" -DLWS_OPENSSL_INCLUDE_DIRS="C:\Program Files\OpenSSL\include" -DLWS_OPENSSL_LIBRARIES="C:\Program Files\OpenSSL\lib\libssl.lib;C:\Program Files\OpenSSL\lib\libcrypto.lib" -DSAI_GIT2_LIB_PATH="C:\Program Files (x86)\libgit2\lib\git2.lib" -DSAI_GIT2_INC_PATH="C:\Program Files (x86)\libgit2\include"
$ cmake --build . --config DEBUG

Hack the libgit2 and lws dlls into visibility (or do it a better non hacky way if you know how...) using an Administrator cmd window

# cp "C:\Program Files (x86)\libgit2\bin\git2.dll" C:\Windows\system32
# cp "C:\Program Files (x86)\libwebsockets\bin\websockets.dll"  C:\Windows\system32

Also in the Administrator window, copy and edit the sai-builder config

# mkdir -p C:\ProgramData\sai\builder\
# cp ../scripts/builder-conf C:\ProgramData\sai\builder\conf
# vim C:\ProgramData\sai\builder\conf

OSX

For OSX, you'll need to adapt the cmake to something like

cmake .. -DLWS_OPENSSL_INCLUDE_DIRS=/usr/local/opt/openssl@1.1/include -DLWS_OPENSSL_LIBRARIES="/usr/local/opt/openssl/lib/libssl.dylib;/usr/local/opt/openssl/lib/libcrypto.dylib"

On the newer M1 OSX, honebrew has to be teleported in /opt, the following worked

cmake .. -DSAI_SERVER=0 -DSAI_GIT2_INC_PATH=/opt/homebrew/Cellar/libgit2/1.1.0/include -DSAI_GIT2_LIB_PATH=/opt/homebrew/Cellar/libgit2/1.1.0/lib/libgit2.dylib -DLWS_OPENSSL_INCLUDE_DIRS=/opt/homebrew/Cellar/openssl@1.1/1.1.1h/include '-DLWS_OPENSSL_LIBRARIES=/opt/homebrew/Cellar/openssl@1.1/1.1.1h/lib/libssl.dylib;/opt/homebrew/Cellar/openssl@1.1/1.1.1h/lib/libcrypto.dylib'

Make sai-builder run as a daemon, create the file /Library/LaunchDaemons/com.warmcat.sai-builder.plist

<?xml version="1.0" encoding="UTF-8"?>
<!DOCTYPE plist PUBLIC "-//Apple Computer//DTD PLIST 1.0//EN"
    "http://www.apple.com/DTDs/PropertyList-1.0.dtd">
<plist version="1.0">
<dict>
    <key>Label</key>
    <string>com.warmcat.sai-builder</string>
    <key>ServiceDescription</key>
    <string>Sai Builder Daemon</string>
    <key>ProgramArguments</key>
    <array>             
        <string>/usr/local/bin/sai-builder</string>
    </array>
    <key>RunAtLoad</key>
    <true/>
    <key>StandardErrorPath</key>
    <string>/var/log/sai-builder.log</string>
</dict>
</plist>

% sudo launchctl load /Library/LaunchDaemons/com.warmcat.sai-builder.plist

That's it, he should be running in the background. To stop him

% sudo launchctl stop com.warmcat.sai-builder

Solaris service

Solaris uses its own service manager called SMF. Build sai and then do the following to create and enable the service, which will also start it at subsequent boots.

# mkdir -p /usr/local/svc/method
# cp scripts/usr-local-svc-method-sai_builder-Solaris /usr/local/svc/method/sai_builder
# cp scripts/var-svc-manifest-site-sai_builder.xml-Solaris /var/svc/manifest/site/sai_builder.xml
# svccfg import /var/svc/manifest/site/sai_builder.xml
# svcadm restart svc:/system/manifest-import
# svcadm enable sai_builder

Make the container start at boot

Let's leave the container and go back to the host

Container # exit
Host #

Then, we can configure it to start on boot

Host # systemctl enable systemd-nspawn@xenial-amd64

Administration of containers

The container should be up and running sai-builder, and should come back up subsequent boots too. You can use machinectl to open a shell on the running container like this:

Host # machinectl -M xenial-amd64 shell
Connected to the local host. Press ^] three times within 1s to exit session.
Container # ps fax
    PID TTY      STAT   TIME COMMAND
      1 ?        Ss     0:00 /lib/systemd/systemd
     14 ?        Ss     0:00 /lib/systemd/systemd-journald
     22 ?        Ss     0:00 /lib/systemd/systemd-logind
     23 ?        Ss     0:00 /usr/bin/dbus-daemon --system --address=systemd: --nofork --nopidfile --systemd-activation
     25 ?        Ssl    0:00 /usr/local/bin/sai-builder
     28 pts/0    Ss+    0:00 /sbin/agetty --noclear --keep-baud console 115200 38400 9600 vt220
     42 ?        Ss     0:00 /lib/systemd/systemd-machined
     43 pts/1    Ss     0:00 -bin/sh
     52 pts/1    S      0:00  \_ bash
     77 pts/1    R+     0:00      \_ ps fax
Container #

You can use journalctl etc in there to see what's happening with sai-builder there, etc.

Distro age considerations

Really old distros are not maintained any more for updates, and their tls certs have expired so they can't do https... it's pretty hopeless.

Xenial (16.04) has problems with cmake (not all CTest functionality) and libgit2 (remote fetch broken on its 0.24 version). You can rapidly build both of these by cloning, they have CMake-based flow that's easy to use.

Packageupstream
cmakehttps://github.com/Kitware/CMake.git
libgit2https://github.com/libgit2/libgit2.git

Install into the default /usr/local, and remove the conflicting old distro packages.

Other Linux distros

Other modern Linux distros with systemd can be done in their own containers and also made to start at boot... these are very lightweight compared to full VMs and yet you can administer them using all the usual systemd goodies including randomly opening shells into their universe if needed.

Flashing and testing on embedded devices

Exposing host assets in the containers

For the embedded case you want to flash and test on an external board, you need to expose typically USB tty devices inside the nspawn container safely.

That's not so trivial or as useful as it could be and has some limitations.

Step 1: Bind the related devices

Device binding inside the container is like a one-time thing at container start, wildcards are not supported and if the device doesn't exist, the container won't start. All this makes supporting ephemeral USB serial ports that come and go impossible.

If your USB tty device is always available, then you can elect to expose it to the container by overriding the systemd service definition for your container, eg

Host # mkdir /etc/systemd/system/systemd-nspawn@YOURMACHINE.service.d
Host # vim /etc/systemd/system/systemd-nspawn@YOURMACHINE.service.d/override.conf

and add in there something like

[Service]
DeviceAllow=char-ttyUSB rwm
DeviceAllow=char-usb_device rwm

Step 2: loosen the permissions on the device at the host

I wasn't able to find a way to let the container unprivileged user seem to join the host dialout group. The only workaround was to allow general access to the device node.

Create a file vim /etc/udev/rules.d/50-allow-nspawn-ttyUSB.rules and add in there

KERNEL=="ttyUSB[0-9]*",MODE="0666"
KERNEL=="ttyACM[0-9]*",MODE="0666"

Step 3: Bind the device nodes into the container

Edit the container's nspawn file, like /etc/systemd/nspawn/MYMACHINE.nspawn and in the [Files] section add

Bind=/dev/serial
Bind=/dev/ttyUSB0

This is very awkward because the device nodes like /dev/ttyUSB0 must exist at container start time or it will fail to start, and you can't expose them via the /dev/serial/by-id symlinks either, or use wildcards since it's done as a one-time init at container start.