Gentoo GSoC Weekly Report 05/14

Created on
12/05/18 00:00

Modified on
12/05/18 13:00

Filed under
GSoC 2018

Things done this week

This week was about creating a repository for an init program written in C/C++, testing it out, and trying to load the real init inside Gentoo root. The progress of each subproject are listed below.

Writing a "preinit" in C/C++

I've set up a repository called preinit_angler, which currently implements the test function (print haiku into kernel message so that we can verify them on next boot) described in this article. Implementing the toy init is necessary is because we set up a tiny-platform that we can place other things onto it in this way: maybe password prompts for entering LUKS passphrases, or options to choose different Android snapshots. What's more, this verifies that our toolchain was really generating correctly statically-linked executables that can be directly executed by the kernel. The repository also holds the incomplete code in attempt to implement Android cryptfs decryption by hand.

Attempt to decrypt Android userdata

I've ran into problems tackling the native encryption of Android: after spending a considerable amount of time on reading the code and splitting functions out from vold, I ran into the dead end of ARM's Trusted Execution Environment (TEE, learn more about it here on Wikipedia). The detailed process of my attempt is logged in this article: Analysis of Android cryptfs.

Bypassing Android force encryption and restructuring userdata partition

As we can't use an encrypted userdata for now, and encryption is forced on AOSP for Nexus 6P, we'll need to bypass the policy. The last section in the cryptfs analysis article ("What next?") describes how to do this. In short, we modify vold's fstab so that vold won't encrypt the phone on first boot.

The next part is about "chainloading" the real, full-blown /sbin/init, which belongs to OpenRC. This is done in several stages:

mounting userdata partition, which holds Gentoo's real filesystem root
switching root into the mounted partition, executing OpenRC's init
launching Android's init in chroot (with jail if Android init checks if it's PID1; uses jchroot)

Note that for jchroot to work, several namespace-related config options are needed, specifically:

CONFIG_SYSVIPC
CONFIG_IPC_NS
CONFIG_PID_NS
CONFIG_NET_NS
CONFIG_UTS_NS

Failing to enable these options will result in jchroot (or to say, the clone(2)) failing with EINVAL.

Current boot sequence, filesystem structure and current work status

After some discussion with UnderSampled, OxR463 and heroxbd on IRC, I feel it necessary to explain a few things here.

Boot sequence and where we are now

The boot sequence is of the following stages:

bootloader loads boot.img, executing the kernel;
kernel executes /init, which is preinit, in the initramfs that came with boot.img;
preinit mounts userdata and handles mountpoints correctly, then executes Gentoo's openrc-init;
OpenRC starts, launching a service that uses jchroot to launch Android's init.

We're currently at the preinit stage, trying to launch OpenRC's init.

Filesystem structure

The final filesystem structure would be like this: Gentoo root sits in userdata partition. The path /var/lib/android is special: it holds the Android init as well as Android mountpoints. When we jchroot into /var/lib/android, Android init would mount the partition it needs correctly, except for /data in the Android root. We'll need to change the fstab entry for userdata into a bind mount of /data (from Android's perspective of view; the real path in userdata is /var/lib/android/data) to itself, thus making /data a mountpoint, so as to make vold happy.

Work to be done next week

By the end of next week we shall see: an Android system successfully booting inside a chroot (and namespaces) under OpenRC init, who gets loaded by preinit. Several expected problems/projects:

switch_root into userdata/gentoo temporarily to see if OpenRC loads fine. Fix it if it's not working.
Refactor userdata, deal with the mountpoints correctly, and edit Android fstab accordingly. Android's fstab parser may not support bind mounts; we may have to bind-mount it ourselves and remove that line in fstab completely.
We'll definitely run into problems, thus a serial console through UART (in headphone jack for Nexus'es and Pixels) is necessary; asynchronous console-ramoops does not fulfill the requirement for interactions.

This requires a headphone-jack UART cable, which was made by @imi415 on Telegram. Kudos for @imi415!