User namespaces and linux capabilities

Small test program which mounts a squashfs in a private mount namespace. If capabilities are not enabled, user mappings will be at bit wonky and all groups and users other than the current user will appears as nobody/nfsnobody.

In most cases the program should terminate cleanly and not leave any mount or children hanging.

https://www.man7.org/linux/man-pages/man7/user_namespaces.7.html for user namespace documentation.
https://man7.org/linux/man-pages/man7/capabilities.7.html for capability documentation.
https://man7.org/linux/man-pages/man2/prctl.2.html for PR_SET_DUMPABLE

Flow

Check that squashfs and mount point exists
Unshare user and mount namespace
Map user uid to 0 and all other uids and gids mapped 1->1
Unshare user namespace
Map uid 0 back to original user uid and all other uids and gids mapped 1->1
Launch user command

Compiling

gcc virt.c -o virt
gcc set_mappings.c -o set_mappings -lcap

Enabling user namespaces on centos

echo "100" | tee /proc/sys/user/max_user_namespaces

Setting Capabilities

sudo setcap 'cap_setuid,cap_setgid+pie' set_mappings

Running

./virt squashfs_file mount_point cmd <optional cmd args>

Any existing folder is a valid mount point. Mounting is done in a user namespace where the user uid is mapped to 0 so so places like /mnt and /opt are valid.

Why is set_mapping a separate file?

When a file with capabilities is executed the "dumpable" attribute is set to the default value in /proc/sys/fs/suid_dumpable (0 most of the time). This affects the /proc access rights and they are only accesible by root. So the target process can not have the capabilites. Might be fixable with some other program flow (if you can reverse the dumpable for a child).

Won't do this for this simple test.

Sequrity (mostly my speculation)

User namespaces expose a larger part of the kernel code which might have not been as thoroughly tested, so bug risk is higher.

Nortamo / virt_isolation