akihirosuda / subuidless

Rootless Containers without `/etc/subuid` and `/etc/subgid`

OCI Seccomp Receiver for running Rootless Containers without /etc/subuid and /etc/subgid

subuidless is an implementaion of OCI Seccomp Receiver for running Rootless Containers without /etc/subuid and /etc/subgid.

subuidlesss emulates ID-related system calls using Seccomp User Notification and XAttrs. Unlike ptrace implementatins of similar projects such as runROOTLESS (PRoot) and remainroot, subuidless can minimize the overhead of system call hooking.

Status

Early POC. Do not use.

Why do we need subuidless?

• It is hard to configure /etc/subuid and /etc/subgid in LDAP environments
• Some container images may require strange UIDs/GIDs that are out of the typical /etc/subuid and /etc/subgid configuration. The typical configuration only allows 65,536 IDs to be available in the container.

Goals and non-goals

Goals:

• Simplicity
• Minimal overhead

Non-goals:

• Provide security boundry across emulated IDs

Requirements

• crun with containers/crun#438
• libseccomp >= v2.5.0
• libprotobuf-c

Note: libseccomp >= v2.5.0 is not available as a dpkg/rpm package in most distros as of July 2020.

To install libseccomp from the source onto a custom prefix (/opt/libseccomp):

$ git clone https://github.com/seccomp/libseccomp.git
$ cd libseccomp
$ git checkout v2.5.0
$ ./autogen.sh
$ ./configure --prefix=/opt/seccomp && make && sudo make install

To install crun:

$ git clone https://github.com/containers/crun.git
$ cd crun
$ hub checkout https://github.com/containers/crun/pull/438
$ ./autogen.sh
$ CFLAGS="-I/opt/libseccomp/include/" LDFLAGS="-L/opt/libseccomp/lib" ./configure && make && sudo make install

Usage

Terminal 1:

$ LIBSECCOMP_PREFIX=/opt/libseccomp ./make.sh
$ mkdir -p ./test/rootfs && docker create --name foo alpine && docker export foo | tar Cx ./test/rootfs && docker rm -f foo
$ ./subuidless ~/.subuidless.sock
Listening on /home/user/.subuidless.sock
...

Terminal 2:

$ RUN_OCI_SECCOMP_RECEIVER=~/.subuidless.sock unshare -r crun run -b ./test foo
/ # cat /proc/self/uid_map
         0       1001          1
/ # touch foo
/ # chown 42:42 foo

Make sure that the chown command succeeds without Invalid argument error, even though no subuid is configured in the uid_map file. The UID ang GID are recorded to the user.rootlesscontainers xattr of the target file.

FIXME: The chowned value are not shown in ls -l currently. Will be shown after the implementaion of stat syscalls. Use getfattr -d -e hex -m user.rootlesscontainers to inspect the xattr value.

Hooked system calls

To be documented, see SCMP_ACT_NOTIFY entries in ./test/config.json.