fakeroot (device node handling)

The fakeroot program is designed to allow non-root users to perform actions that would normally require root privileges as part of the package generation process. It is used by the rootfs_ipkg class for root filesystem creation and by the image class for the creation of filesystem images. Some recipes also use fakeroot to assist with parts of the package installation (usually) or building where root privligeses are expected by the package.

In particular fakeroot deals with:

How fakeroot works

First of all we'll look at an example of how the fakeroot process works when used manually.

If we attempt to create a device node as a normal non-root user then the command will fail, telling is that we do not have permission to create device nodes:

~%> mknod hdc b 22 0
mknod: `hdc': Operation not permitted

Yet the image class is able to create device nodes and include them in the final images, all without the need to have root privileges.

Let's try and create that node again, this time we'll run the commands from within a fakeroot process:

~%> ./tmp/staging/x86_64-linux/bin/fakeroot
~#> mknod hdc b 22 0
~#> ls -l hdc
brw-------  1 root root 22, 0 Aug 18 13:20 hdc
~#>

So it looks like we have successfully managed to create a device node, even though we did not have to give a password for the root user. In reality this device node still doesn't exist, it just looks like it exits. Fakeroot is lying to the shell process and telling it that "yes, this file exists and these are it's properties". We'll talk more about how fakeroot actually works in a minute.

In this case hdc is the cd-rom drive, so let's try and actually mount the cd-rom:

~#> mkdir disk
~#> mount hdc disk
ERROR: ld.so: object 'libfakeroot.so.0' from LD_PRELOAD cannot be preloaded: ignored.
mount: only root can do that
~#>

So even though it appears we have root permissions, and that we created a device node, you see that the system gives an error about libfakeroot and about not being able to run mount because we are not root.

If we exit the fakeroot process and then look at the device node this is what we see:

~#> exit
~%> ls -l hdc
brw-------  1 user user 22, 0 Aug 18 13:20 hdc
~#>

Note that it isn't a device node at all, just an empty file owned by the current user!

So what exactly is fakeroot doing? It's using LD_PRELOAD to load a shared library into program which replaces calls into libc, such as open and stat, and then returns information to make it look like certain commands succeeded without actually performing those commands. So when creating a device node fakeroot will:

  1. Intercept the mknod system call and instead of creating a device node it'll just create an empty file, owned by the user who run fakeroot;

  2. It remembers the fact that mknod was called by root and it remembers the properties of the device node;

  3. When a program, such as ls, calls stat on the file fakeroot remembers that it was device node, owned by root, and modifies that stat information to return this to ls. So ls sees a device node even though one doesn't exist.

When we tried to run mount we received the error "ERROR: ld.so: object 'libfakeroot.so.0' from LD_PRELOAD cannot be preloaded: ignored.". This is due to the fact that mount is an suid root binary, and for security reasons LD_PRELOAD is disabled on suid binaries.

There are some very important points to remember when dealing with fakeroot:

  1. All information regarding devices nodes, uid and gids will be lost when fakeroot exists;

  2. None of the device nodes, uids or gids will appear on disk. However if you tar up a directory from within fakeroot (for example), all of these device, uids and gids will appear correctly in the tar archive;

  3. Any suid binaries will not interact with fakeroot;

  4. Any static binaries will not interact with fakeroot;

Root filesystem, images and fakeroot

Many people have been confused by the generated root filesystem not containing any valid device nodes. This is in fact the expected behaviour.

When you look at a generated root filesystem you'll notice that the device nodes all appear to be incorrectly created:

~%> ls -l tmp/rootfs/dev | grep ttySC
-rw-r--r--  1 root root 0 Aug 16 13:07 ttySC0
-rw-r--r--  1 root root 0 Aug 16 13:07 ttySC1
-rw-r--r--  1 root root 0 Aug 16 13:07 ttySC2
~%>

These are empty files and not device nodes at all.

If we look in the image files generated from that root filesystem then everything is actually ok:

~%> tar -ztvf tmp/deploy/images/titan-titan-20060816030639.rootfs.tar.gz | grep " ./dev/ttySC"
crw-r----- root/root     204,8 2006-08-16 13:07:12 ./dev/ttySC0
crw-r----- root/root     204,9 2006-08-16 13:07:12 ./dev/ttySC1
crw-r----- root/root    204,10 2006-08-16 13:07:12 ./dev/ttySC2
~%>

The images are created from within the same fakeroot process as the creation of the root filesystem and therefore it correctly picks up all of the special files and permissions from fakeroot.

NOTE: This means that you cannot use the root filesystem in tmp/rootfs directly on your target device. You need to use the .tar.gz image and uncompress it, as root, in order to generate a root filesystem which is suitable for use directly on the target (or as an NFS root).

Recipes and fakeroot

Some applications require that you have root permissions to run their installation routine, and this is another area where fakeroot can help. In a recipe the method for a standard task, such as the do_install method for the install task:

do_install() {
    install -d ${D}${bindir} ${D}${sbindir} ${D}${mandir}/man8 \
               ${D}${sysconfdir}/default \
               ${D}${sysconfdir}/init.d \
               ${D}${datadir}/arpwatch

    oe_runmake install DESTDIR=${D}
    oe_runmake install-man DESTDIR=${D}
    ...

can be modified to run within a fakeroot environment by prefixing the method name with fakeroot:

fakeroot do_install() {
    install -d ${D}${bindir} ${D}${sbindir} ${D}${mandir}/man8 \
               ${D}${sysconfdir}/default \
               ${D}${sysconfdir}/init.d \
               ${D}${datadir}/arpwatch

    oe_runmake install DESTDIR=${D}
    oe_runmake install-man DESTDIR=${D}
    ...