Go Contain Me (SREcon Americas 2018)

My first day at SREcon Americas 2018 was very exciting and inspiring. It started with the Containers from Scratch workshop by Avishai Ish-Shalom and Nati Cohen. They developed a syscall-level workshop about Linux containers that I can highly recommend. It deals with a program containing and isolating itself step by step using Linux systemcalls. In the end, the program would fork to drop the last bit of privileges that is left. That was super fun, although the network namespaces gave me a hard time due to a silly implementation mistake I made.

The workshop code is in Python and worked perfectly. However, I want to improve my Golang skills so I decided to redo the assignments from the workshop in Golang. Shouldn’t be too hard, should it? After all, docker is written in Golang.

This is my write-up of the endeavor.

General Idea

Similar to other popular container solutions, our little program should:

Spoiler alert: We will not make all of this work. The processes namespace gave me a hard time because I was restricting myself to state-of-the art standard libraries and avoiding custom CGO code. Unfortunately, the syscall package is not considered state-of-the-art anymore. So I had to avoid using it. 😔

Who Am I?

Our container will start off as a simple process that will slowly isolate itself from the current environment. First, we want to know who we are, so we fetch the process ID (PID) first.

pid := unix.Getpid()
log.Printf("pid: %v", pid)

We also want a unique identifier for our container. Process IDs are limited and might eventually be re-assigned. So let’s grab a UUID and use that.

id := uuid.New().String()
log.Printf("container id: %v", id)

We can use the container ID to name directories in a non-conflicting way later. It is highly unlikely that any two UUIDs collide in an environment like ours.

Building Fences

Next thing we want to do is to build a fence around our process. We do this for CPU and memory using Linux cgroups. For this we want to write our PID to /sys/fs/cgroup/cpu/go-contain-me/<UUID>/tasks and write the number of CPU shares we want to grant the process into /sys/fs/cgroup/cpu/go-contain-me/<UUID>/cpu.shares.

cgroupCPU := "/sys/fs/cgroup/cpu/go-contain-me/" + id + "/"
log.Println("cpu cgroup: create")
err = os.MkdirAll(cgroupCPU, 0744)
if err != nil {
    log.Fatal(err)
}
log.Println("cpu cgroup: add pid")
err = ioutil.WriteFile(cgroupCPU+"tasks", []byte(strconv.Itoa(pid)), 0644)
if err != nil {
    log.Fatal(err)
}
if len(*cpuShares) > 0 {
    log.Println("cpu cgroup: set shares")
    err := ioutil.WriteFile(cgroupCPU+"cpu.shares",
        []byte(*cpuShares), 0644)
    if err != nil {
        log.Fatal(err)
    }
}

For the memory cgroup we do something similar, yet there is a small difference. We are not supposed to set shares here, but the actual number of bytes we want the process to be limited to. We can also limit the number of swap bytes the process is allowed to consume.

cgroupMemory := "/sys/fs/cgroup/memory/go-contain-me/" + id + "/"
log.Println("memory cgroup: create")
err = os.MkdirAll(cgroupMemory, 0644)
if err != nil {
    log.Fatal(err)
}
log.Println("memory cgroup: add pid")
err = ioutil.WriteFile(cgroupMemory+"tasks",
    []byte(strconv.Itoa(pid)), 0644)
if err != nil {
    log.Fatal(err)
}
if len(*memoryLimit) > 0 {
    log.Println("memory cgroup: set memory limit")
    err := ioutil.WriteFile(cgroupMemory+"memory.limit_in_bytes",
        []byte(*memoryLimit), 0644)
    if err != nil {
        log.Fatal(err)
    }
}
if len(*swapLimit) > 0 {
    log.Println("memory cgroup: set swap limit")
    err := ioutil.WriteFile(cgroupMemory+"memory.memsw.limit_in_bytes",
        []byte(*swapLimit), 0644)
    if err != nil {
        log.Fatal(err)
    }
}

Great, now we have contained ourselves in terms of resource usage.

Overlay Root File System

Now we shall isolate our process further from the host system step by step. Let’s assume we have an extracted userspace image at /root/go-contain-me/images/busybox. You can use docker extract to get your hands on one quickly if needed. 🐳

As multiple containers might be using the same underlying image, we have to make sure we do not write to the image data. However, we still want to be able to make changes to the data, such as adding, modifying, or removing files. But how? The overlay filesystem comes to the rescue! As the name suggests, we can overlay something called an upperdir onto a lowerdir. We would also need a workdir where we store some copy-on-write information, e.g. for files that have been deleted during operation.

So the first order of business for overlaying a filesystem is to make sure all the required directories exists:

newRoot := baseDir + "/containers/" + id + "/rootfs"
workDir := baseDir + "/containers/" + id + "/workdir"
for _, path := range []string{newRoot, workDir} {
    err = os.MkdirAll(path, os.ModePerm)
    if err != nil {
        log.Fatal(err)
    }
}

After that the whole operation is just a regular mount with a less often seen set of options:

log.Printf("mount: overlay")
imageRoot := baseDir + "/images/" + *image
err = unix.Mount("overlay", newRoot, "overlay", uintptr(unix.MS_NODEV),
    "lowerdir="+imageRoot+",upperdir="+newRoot+",workdir="+workDir)
if err != nil {
    log.Fatal(err)
}

The MS_NODEV flag, by the way, prevents special files (devices) to be accessed on this filesystem. We will create those later using the mknod system call.

Moving To A New Mount Namespace

Right now, our mounts show up on the host and pollute it a bit. Luckily, we can isolate our mounts from the host mounts by creating a new namespace for our process (at some point we can start calling the process a container).

log.Printf("newns: mount")
err = unix.Unshare(unix.CLONE_NEWNS)
if err != nil {
    log.Fatal(err)
}

Now we remount the root filesystem in our namespace to assign it to the newly created namespace:

log.Printf("remount: /")
err = unix.Mount("", "/", "", uintptr(unix.MS_PRIVATE|unix.MS_REC), "")
if err != nil {
    log.Fatal(err)
}

We are using flags again:

Special Mounts

Now that we have our isolated mount namespace, it is time to mount some special filesystems there. Let’s use a for loop to avoid writing the same code over and over again.

mounts := []struct {
    source  string
    target  string
    fsType  string
    flags   uint
    options string
}{
    {source: "proc", target: newRoot + "/proc", fsType: "proc"},
    {source: "sysfs", target: newRoot + "/sys", fsType: "sysfs"},
    {
        source:  "tmpfs",
        target:  newRoot + "/dev",
        fsType:  "tmpfs",
        flags:   unix.MS_NOSUID | unix.MS_STRICTATIME,
        options: "mode=755",
    },
    {
        source: "devpts",
        target: newRoot + "/dev/pts",
        fsType: "devpts",
    },
}
for _, mnt := range mounts {
    // ensure mount target exists
    log.Printf("mkdirall: %v", mnt.target)
    err := os.MkdirAll(mnt.target, os.ModePerm)
    if err != nil {
        log.Fatal(err)
    }

    // mount
    log.Printf("mount: %v (%v)", mnt.source, mnt.fsType)
    flags := uintptr(mnt.flags)
    err = unix.Mount(mnt.source, mnt.target, mnt.fsType, flags, mnt.options)
    if err != nil {
        log.Fatal(err)
    }
}

This should leave us with most the most important filesystems in place under our new root. Remember, our new root is still seen as /root/go-contain-me/containers/<UUID>/rootfs. But that is going to change soon.

Essential File Descriptors

We will soon pivot the process’ root to use our container’s rootfs directory as root. See how I just used container now instead of process? This was totally arbitrary. 🙃 But before we lose access to the current filesystem tree, let’s rescue essential file descriptors such as stdin and stdout. Without them functioning, we would not have much fun with our container.

A simple symlink() does the job:

for i, name := range []string{"stdin", "stdout", "stderr"} {
    source := "/proc/self/fd/" + strconv.Itoa(i))
    target := newRoot + "/dev/" + name
    log.Printf("symlink: %v", name)
    err := unix.Symlink(source, target)
    if err != nil {
        log.Fatal(err)
    }
}

Creating Devices

Processes running inside our container may assume that a certain set of special devices is present. One popular example being /dev/null, which is often used to drop data streams into Nirvana. If /dev/null weren’t present those data streams may end up in a regular file. This could, in turn, quickly fill up the filesystem. If there are no quotas on the container’s filesystem, this might affect the host’s filesystem as well. Not cool.

We’ll use the loop approach one more time here:

devices := []struct {
    name  string
    attr  uint32
    major uint32
    minor uint32
}{
    {name: "null", attr: 0666 | unix.S_IFCHR, major: 1, minor: 3},
    {name: "zero", attr: 0666 | unix.S_IFCHR, major: 1, minor: 3},
    {name: "random", attr: 0666 | unix.S_IFCHR, major: 1, minor: 8},
    {name: "urandom", attr: 0666 | unix.S_IFCHR, major: 1, minor: 9},
    {name: "console", attr: 0666 | unix.S_IFCHR, major: 136, minor: 1},
    {name: "tty", attr: 0666 | unix.S_IFCHR, major: 5, minor: 0},
    {name: "full", attr: 0666 | unix.S_IFCHR, major: 1, minor: 7},
}
for _, dev := range devices {
    dt := int(unix.Mkdev(dev.major, dev.minor))
    log.Printf("mknod: %v (%v)", dev.name, dt)
    err := unix.Mknod(newRoot +"/dev/" dev.name, dev.attr, dt)
    if err != nil {
        log.Fatal(err)
    }
}

Isolate The UNIX Time-Sharing Namespace

We are coming closer to pivoting the root. I promise. However, there are still a few more isolation steps we should do. For example, we want the hostname of the container to be isolated from the hostname of the host. One might expect this to fall under the domain of the network namespace. Surprisingly that is not the case. For historical reasons, the namespace for this is the UNIX Time-Sharing namespace or short UTS.

So let’s unshare() this one before setting the hostname:

log.Printf("newns: UNIX time sharing")
err = unix.Unshare(unix.CLONE_NEWUTS)
if err != nil {
    log.Fatal(err)
}
// change hostname in new UTS
log.Printf("set hostname")
err = unix.Sethostname([]byte(id))
if err != nil {
    log.Fatal(err)
}

Isolate The Process Namespace (b0rked)

We also want to isolate the container process namespace from the host. Meaning, that if we run ps on the container, we don’t want to see the processes of the host.

Note: I was not able to get this one to work. The code compiles, the code runs, but then the contained processes run out of memory real quick. Despite having a generous cgroup setting for memory. I did not investigate much time into debugging this. Feel free to drop me a line if you happen to know what the problem is. 🤓

For the sake of completeness, here is my code:

log.Printf("newns: processes")
err = unix.Unshare(unix.CLONE_NEWPID)
if err != nil {
    log.Fatal(err)
}

Isolating The Network

For the network namespace, we make another call to unshare(). This will give us a new namespace that does contain a loopback interface only. Clean and lean!

log.Printf("newns: network")
err = unix.Unshare(unix.CLONE_NEWNET)
if err != nil {
    log.Fatal(err)
}

If you like to dig deeper into network namespacing: Try ip netns help for a start and don’t forget to link the namespace to the container’s default namespace before unsharing!

Pivoting

Phew. That was a long journey. Now we can pivot the root! Hooray! The operation looks more complicated than it is. Basically, we just do the following things:

log.Printf("pivot root")
oldRootBeforePivot := newRoot + "/.old-root"
oldRootAfterPivot := "/.old-root"
err = os.MkdirAll(oldRootBeforePivot, os.ModePerm)
if err != nil {
    log.Fatalf("mkdirall old root: %v", err)
}

unix.PivotRoot(newRoot, oldRootBeforePivot)
if err != nil {
    log.Fatalf("pivot root: %v", err)
}
unix.Chdir("/")
if err != nil {
    log.Fatalf("chdir: %v", err)
}
unix.Unmount(oldRootAfterPivot, unix.MNT_DETACH)
if err != nil {
    log.Fatalf("unmount old root: %v", err)
}
unix.Rmdir(oldRootAfterPivot)
if err != nil {
    log.Fatalf("rmdir old root: %v", err)
}

The Finally

Hold your breath, now comes the final operation before we fully enter container land! We overload the process with the new binary to run. Here we are using sh to get a shell we can interact with.

Ideally, we would do this in a child process after fork() or clone(), but it turns out, forking isn’t too much of a great idea in Golang. I’ll spare you the details, but there are plenty of discussions about this at the usual places.

err = unix.Exec("/bin/sh", []string{"sh"}, []string{})
log.Fatal(err)

Ideally, the line reading log.Fatal(err) is never reached.

Running It!

It’s time to run this thing! Do yourself a favor and run this in a virtual machine. The code is not free of risk and could force you to reboot in case something goes wrong. And we don’t reboot our computers anymore nowadays, do we? 😂

# ./go-contain-me
2018/03/29 04:03:46 pid: 1054
2018/03/29 04:03:46 container id: c16f889c-6a49-49a4-bbb0-add1094993c5
2018/03/29 04:03:46 cpu cgroup: create
2018/03/29 04:03:46 cpu cgroup: add pid
2018/03/29 04:03:46 memory cgroup: create
2018/03/29 04:03:46 memory cgroup: add pid
2018/03/29 04:03:46 memory cgroup: set memory limit
2018/03/29 04:03:46 mount: overlay
2018/03/29 04:03:46 newns: mount
2018/03/29 04:03:46 remount: /
2018/03/29 04:03:46 mkdirall: /root/go-contain-me/containers/c16f889c-6a49-49a4-bbb0-add1094993c5/rootfs/proc
2018/03/29 04:03:46 mount: proc (proc)
2018/03/29 04:03:46 mkdirall: /root/go-contain-me/containers/c16f889c-6a49-49a4-bbb0-add1094993c5/rootfs/sys
2018/03/29 04:03:46 mount: sysfs (sysfs)
2018/03/29 04:03:46 mkdirall: /root/go-contain-me/containers/c16f889c-6a49-49a4-bbb0-add1094993c5/rootfs/dev
2018/03/29 04:03:46 mount: tmpfs (tmpfs)
2018/03/29 04:03:46 mkdirall: /root/go-contain-me/containers/c16f889c-6a49-49a4-bbb0-add1094993c5/rootfs/dev/pts
2018/03/29 04:03:46 mount: devpts (devpts)
2018/03/29 04:03:46 symlink: stdin
2018/03/29 04:03:46 symlink: stdout
2018/03/29 04:03:46 symlink: stderr
2018/03/29 04:03:46 mknod: null (259)
2018/03/29 04:03:46 mknod: zero (259)
2018/03/29 04:03:46 mknod: random (264)
2018/03/29 04:03:46 mknod: urandom (265)
2018/03/29 04:03:46 mknod: console (34817)
2018/03/29 04:03:46 mknod: tty (1280)
2018/03/29 04:03:46 mknod: full (263)
2018/03/29 04:03:46 newns: UNIX time sharing
2018/03/29 04:03:46 set hostname
2018/03/29 04:03:46 newns: network
2018/03/29 04:03:46 pivot root

Inside the container, we can see only our own mounts:

/ # mount
overlay on / type overlay (rw,nodev,relatime,lowerdir=/root/go-contain-me/images/busybox,upperdir=/root/go-contain-me/containers/c16f889c-6a49-49a4-bbb0-add1094993c5/rootfs,workdir=/root/go-contain-me/containers/c16f889c-6a49-49a4-bbb0-add1094993c5/workdir)
proc on /proc type proc (rw,relatime)
sysfs on /sys type sysfs (rw,relatime)
tmpfs on /dev type tmpfs (rw,nosuid,mode=755)
devpts on /dev/pts type devpts (rw,relatime,mode=600,ptmxmode=000)

We also have our own network namespace. All the host’s devices are gone. If we want to add network interfaces, we may use the netns functionality of iputils.

/ # ip a
1: lo: <LOOPBACK> mtu 65536 qdisc noop qlen 1
    link/loopback 00:00:00:00:00:00 brd 00:00:00:00:00:00

The situation is not that good for the process namespace. As I said, I was not able to get it to work reliably. So here we see all the processes of the host as well. Meh.

/ # ps -e
PID   USER     TIME  COMMAND
    1 root      0:00 {systemd} /sbin/init
    2 root      0:00 [kthreadd]
    3 root      0:00 [ksoftirqd/0]
✂️
 1054 root      0:00 sh
 1066 root      0:00 ps -e

Full Source

Here is the full piece of code for your amusement and further experimentation. The code works with a directory structure that looks similar to this:

root@go-contain-me-1:~# tree
.
`-- go-contain-me
    |-- containers
    |   `-- 8f0f5a2d-0ce8-4bd1-887a-2c5b275ee337
    |       |-- rootfs
    |       `-- workdir
    `-- images
        `-- busybox
            `-- (a full user space here)

Compile the program:

$ CGO_ENABLED=0 GOOS=linux go build -a -ldflags '-extldflags "-static"' .

Here’s the source for your interest:

package main

import (
    "flag"
    "io/ioutil"
    "log"
    "os"
    "strconv"

    "github.com/google/uuid"
    "golang.org/x/sys/unix"
)

var (
    baseDir = "/root/go-contain-me"
)

func main() {
    var err error
    cpuShares := flag.String("cpu-shares", "",
        "CPU shares of the container.")
    memoryLimit := flag.String("memory-limit", "256m",
        "Memory limit of the container.")
    swapLimit := flag.String("swap-limit", "",
        "Swap limit of the container.")
    image := flag.String("image", "busybox", "name of the container image")
    flag.Parse()

    pid := unix.Getpid()
    log.Printf("pid: %v", pid)

    // generate container id
    id := uuid.New().String()
    log.Printf("container id: %v", id)

    // CPU cgroup
    cgroupCPU := "/sys/fs/cgroup/cpu/go-contain-me/" + id + "/"
    log.Println("cpu cgroup: create")
    err = os.MkdirAll(cgroupCPU, 0744)
    if err != nil {
        log.Fatal(err)
    }
    log.Println("cpu cgroup: add pid")
    err = ioutil.WriteFile(cgroupCPU+"tasks", []byte(strconv.Itoa(pid)), 0644)
    if err != nil {
        log.Fatal(err)
    }
    if len(*cpuShares) > 0 {
        log.Println("cpu cgroup: set shares")
        err := ioutil.WriteFile(cgroupCPU+"cpu.shares",
            []byte(*cpuShares), 0644)
        if err != nil {
            log.Fatal(err)
        }
    }

    // memory cgroup
    cgroupMemory := "/sys/fs/cgroup/memory/go-contain-me/" + id + "/"
    log.Println("memory cgroup: create")
    err = os.MkdirAll(cgroupMemory, 0644)
    if err != nil {
        log.Fatal(err)
    }
    log.Println("memory cgroup: add pid")
    err = ioutil.WriteFile(cgroupMemory+"tasks",
        []byte(strconv.Itoa(pid)), 0644)
    if err != nil {
        log.Fatal(err)
    }
    if len(*memoryLimit) > 0 {
        log.Println("memory cgroup: set memory limit")
        err := ioutil.WriteFile(cgroupMemory+"memory.limit_in_bytes",
            []byte(*memoryLimit), 0644)
        if err != nil {
            log.Fatal(err)
        }
    }
    if len(*swapLimit) > 0 {
        log.Println("memory cgroup: set swap limit")
        err := ioutil.WriteFile(cgroupMemory+"memory.memsw.limit_in_bytes",
            []byte(*swapLimit), 0644)
        if err != nil {
            log.Fatal(err)
        }
    }

    // create container directories
    newRoot := baseDir + "/containers/" + id + "/rootfs"
    workDir := baseDir + "/containers/" + id + "/workdir"
    for _, path := range []string{newRoot, workDir} {
        err = os.MkdirAll(path, os.ModePerm)
        if err != nil {
            log.Fatal(err)
        }
    }

    // mount rootfs as overlay
    log.Printf("mount: overlay")
    imageRoot := baseDir + "/images/" + *image
    err = unix.Mount("overlay", newRoot, "overlay", uintptr(unix.MS_NODEV),
        "lowerdir="+imageRoot+",upperdir="+newRoot+",workdir="+workDir)
    if err != nil {
        log.Fatal(err)
    }

    // new mount namespace
    log.Printf("newns: mount")
    err = unix.Unshare(unix.CLONE_NEWNS)
    if err != nil {
        log.Fatal(err)
    }

    // remount rootfs in new namespace
    log.Printf("remount: /")
    err = unix.Mount("", "/", "", uintptr(unix.MS_PRIVATE|unix.MS_REC), "")
    if err != nil {
        log.Fatal(err)
    }

    // mount special
    mounts := []struct {
        source  string
        target  string
        fsType  string
        flags   uint
        options string
    }{
        {source: "proc", target: newRoot + "/proc", fsType: "proc"},
        {source: "sysfs", target: newRoot + "/sys", fsType: "sysfs"},
        {
            source:  "tmpfs",
            target:  newRoot + "/dev",
            fsType:  "tmpfs",
            flags:   unix.MS_NOSUID | unix.MS_STRICTATIME,
            options: "mode=755",
        },
        {
            source: "devpts",
            target: newRoot + "/dev/pts",
            fsType: "devpts",
        },
    }
    for _, mnt := range mounts {
        // ensure mount target exists
        log.Printf("mkdirall: %v", mnt.target)
        err := os.MkdirAll(mnt.target, os.ModePerm)
        if err != nil {
            log.Fatal(err)
        }

        // mount
        log.Printf("mount: %v (%v)", mnt.source, mnt.fsType)
        flags := uintptr(mnt.flags)
        err = unix.Mount(mnt.source, mnt.target, mnt.fsType, flags, mnt.options)
        if err != nil {
            log.Fatal(err)
        }
    }

    // essential file descriptors
    for i, name := range []string{"stdin", "stdout", "stderr"} {
        source := "/proc/self/fd/" + strconv.Itoa(i))
        target := newRoot + "/dev/" + name
        log.Printf("symlink: %v", name)
        err := unix.Symlink(source, target)
        if err != nil {
            log.Fatal(err)
        }
    }

    // create devices
    devices := []struct {
        name  string
        attr  uint32
        major uint32
        minor uint32
    }{
        {name: "null", attr: 0666 | unix.S_IFCHR, major: 1, minor: 3},
        {name: "zero", attr: 0666 | unix.S_IFCHR, major: 1, minor: 3},
        {name: "random", attr: 0666 | unix.S_IFCHR, major: 1, minor: 8},
        {name: "urandom", attr: 0666 | unix.S_IFCHR, major: 1, minor: 9},
        {name: "console", attr: 0666 | unix.S_IFCHR, major: 136, minor: 1},
        {name: "tty", attr: 0666 | unix.S_IFCHR, major: 5, minor: 0},
        {name: "full", attr: 0666 | unix.S_IFCHR, major: 1, minor: 7},
    }
    for _, dev := range devices {
        dt := int(unix.Mkdev(dev.major, dev.minor))
        log.Printf("mknod: %v (%v)", dev.name, dt)
        err := unix.Mknod(newRoot + "dev" + dev.name, dev.attr, dt)
        if err != nil {
            log.Fatal(err)
        }
    }
    // new UTS (UNIX Timesharing System) namespace
    log.Printf("newns: UNIX time sharing")
    err = unix.Unshare(unix.CLONE_NEWUTS)
    if err != nil {
        log.Fatal(err)
    }
    // change hostname in new UTS
    log.Printf("set hostname")
    err = unix.Sethostname([]byte(id))
    if err != nil {
        log.Fatal(err)
    }

    /*
         * can't get it to work :,(
        // new process namespace
        log.Printf("newns: processes")
        err = unix.Unshare(unix.CLONE_NEWPID)
        if err != nil {
            log.Fatal(err)
        }
    */

    // new network namespace
    log.Printf("newns: network")
    err = unix.Unshare(unix.CLONE_NEWNET)
    if err != nil {
        log.Fatal(err)
    }

    // pivot root
    log.Printf("pivot root")
    oldRootBeforePivot := newRoot + "/.old-root"
    oldRootAfterPivot := "/.old-root"
    err = os.MkdirAll(oldRootBeforePivot, os.ModePerm)
    if err != nil {
        log.Fatalf("mkdirall old root: %v", err)
    }

    unix.PivotRoot(newRoot, oldRootBeforePivot)
    if err != nil {
        log.Fatalf("pivot root: %v", err)
    }
    unix.Chdir("/")
    if err != nil {
        log.Fatalf("chdir: %v", err)
    }
    unix.Unmount(oldRootAfterPivot, unix.MNT_DETACH)
    if err != nil {
        log.Fatalf("unmount old root: %v", err)
    }
    unix.Rmdir(oldRootAfterPivot)
    if err != nil {
        log.Fatalf("rmdir old root: %v", err)
    }

    err = unix.Exec("/bin/sh", []string{"sh"}, []string{})
    log.Fatal(err)
}

Note: I used path.Join() in a previous version but I decided to remove it. I found that to be very cluttery. So this will not run properly should the POSIX standard ever decide to replace the path separator / with something else. I am willing to take this risk, though. 😉