The filesystem hierarchy has received some cleanups and simplifications that
typical GNU/Linux distributions don't have. First of all, there is no /usr
sub hierarchy split and there is no /bin vs /sbin split.
The /usr split and the /bin vs /sbin split are historical artifacts. Of
course, over the years, people started to pretend that it was a deliberate
design choice and invented lots of interesting reasoning to justify why this is
the way to go, despite the original technical reasons being obsolete for a
really long time now. Please refer to this thread on the BusyBox mailing list
for further discussion.
The directories typically found in /usr have been merged back into the
filesystem root (/usr/bin and /usr/sbin to /bin and /usr/lib
to /lib).
The /usr/share directory containing application data has also been moved to
the filesystem root (i.e. there is a /share).
Many systems have a /usr/libexec directory containing executables not intended
to be run by people but by programs. This has been moved to /lib/libexec.
The /usr directory is still present but now serves its original purpose again.
Storing user home directories. For instance, the /root directory has been
moved to /usr/root.
Since build tools, source code and headers are typically not installed, there
is currently no need to think about where to put /usr/include or /usr/src.
A compressed, read only squashfs image is mounted to the root node of the virtual filesystem hierarchy.
The directories /dev, /proc and /sys do not contain any files and are
used as mount points for devtmpfs, procfs and sysfs respectively.
The directories /tmp, /run and /var are used as mount points for tmpfs
mounts. The sub hierarchy in /var is constructed during system boot.
See below on how changes to /var/lib are persisted and preserved across
reboots.
The directories /boot and /mnt are used as mount points for temporary
mounts. The former specifically for mounting the boot medium containing the
kernel and squasfs image in order to apply updates, the later for other
temporary mounts.
For details on how the system boots and configures the filesystem from kernel command line arguments, see boot.md.
A directory /cfg was added to implement an overlay mount setup. The directory
/cfg/preserve contains the original versions of files built into the
compressed, read only root filesystem.
For instance, the original contents of /etc are actually stored in
/cfg/preserve/etc.
The directory /cfg/overlay is used as a mount point for a read/write
partition that is used to override and persist changes to the base
configuration.
During system boot, an overlay filesystem is mounted to /etc with the lower
directory set to /cfg/preserve/etc and the upper directory set to
/cfg/overlay/etc.
As a result, the /etc directory initially contains files stored in the
squashfs image, but changes can be made. The altered files are stored on the
dedicated partition or device mounted to /cfg/overlay.
A similar setup is used for the /var/lib and the /usr directories. The
/var/lib directory combines /cfg/preserve/var_lib with overlays from
/cfg/overlay/var_lib.
This setup allows for simple management of site local configuration changes, simple backups and a simple "factory reset" (i.e. wiping the overlay partition or device).
On systems where this is possible, BTRFS is used for the overlay partition. BTRFS sub volumes can be used to make snapshots of the changed configuration and if something should break, allows for a simple revert to the last known good state.
The overlay setup can also be disabled (resulting in bind mounts to
/cfg/preserve) or configured to use a tmpfs as backing store.
Some processors support executing op codes for slightly different architectures.
For instance, many 64 bit processors can be set into 32 bit mode and run 32 bit
programs. Such programs then require additional 32 bit versions of shared
libraries, already built for the 64 bit system, creating the necessity for
having two different versions of the /lib directory.
For the time being, it has been decided to not include multiarch support.
All packages are built for a single target architecture. This simplifies both
the build process and the final system as well as reducing the memory footprint
of the system image. A proposal exists for creating a separate /system32 sub
hierarchy on 64 bit targets that require 32 bit binaries.