Booting Linux off of Google Drive
Competitiveness is a vice of mine. When I heard that a friend got Linux to boot off of NFS, I had to one-up her. I had to prove that I could create something harder, something better, faster, stronger.
Like all good projects, this began with an Idea.
My mind reached out and grabbed wispy tendrils from the æther, forcing the disparate concepts to coalesce. The Mass gained weight in my hands, and a dark, swirling colour promising doom to those who gazed into it for long.
On the brink of insanity, my tattered mind unable to comprehend the twisted interplay of millennia of arcane programmer-time and the ragged screech of madness, I reached into the Mass and steeled myself to the ground lest I be pulled in, and found my magnum opus.
Booting Linux off of a Google Drive root.
But How?
I wanted this to remain self-contained, so I couldn't have a second machine act as a "helper". My mind went immediately to FUSE—a program that acts as a filesystem driver in userspace (with cooperation from the kernel).
I just had to get FUSE programs installed in the Linux kernel initramfs and configure networking. How bad could it be?
The Linux Boot Process
The Linux boot process is, technically speaking, very funny. Allow me to pretend I understand for a moment1:
- The firmware (BIOS/UEFI) starts up and loads the bootloader
- The bootloader loads the kernel
- The kernel unpacks a temporary filesystem into RAM which has the tools to mount the real filesystem
- The kernel mounts the real filesystem and switches the process to the init system running on the new filesystem
As strange as the third step may seem, it's very helpful! We can mount a FUSE filesystem in that step and boot normally.
A Proof of Concept
The initramfs needs to have both network support as well as the proper FUSE binaries. Thankfully, Dracut makes it easy enough to build a custom initramfs.
I decide to build this on top of Arch Linux because it's relatively lightweight and I'm familiar with how it works, as opposed to something like Alpine.
$ git clone https://github.com/dracutdevs/dracut
$ podman run -it --name arch -v ./dracut:/dracut docker.io/archlinux:latest bash
In the container, I installed some packages (including the linux
package because I need a functioning kernel), compiled dracut
from source, and wrote a simple module script in modules.d/90fuse/module-setup.sh
:
#!/bin/bash
check() {
require_binaries fusermount fuseiso mkisofs || return 1
return 0
}
depends() {
return 0
}
install() {
inst_multiple fusermount fuseiso mkisofs
return 0
}
That's it. That's all the code I had to write. Buoyed by my newfound confidence, I powered ahead, building the EFI image.
$ ./dracut.sh --kver 6.9.6-arch1-1 \
--uefi efi_firmware/EFI/BOOT/BOOTX64.efi \
--force -l -N --no-hostonly-cmdline \
--modules "base bash fuse shutdown network" \
--add-drivers "target_core_mod target_core_file e1000" \
--kernel-cmdline "ip=dhcp rd.shell=1 console=ttyS0"
$ qemu-kvm -bios ./FV/OVMF.fd -m 4G \
-drive format=raw,file=fat:rw:./efi_firmware \
-netdev user,id=network0 -device e1000,netdev=network0 -nographic
...
...
dracut Warning: dracut: FATAL: No or empty root= argument
dracut Warning: dracut: Refusing to continue
Generating "/run/initramfs/rdsosreport.txt"
You might want to save "/run/initramfs/rdsosreport.txt" to a USB stick or /boot
after mounting them and attach it to a bug report.
To get more debug information in the report,
reboot with "rd.debug" added to the kernel command line.
Dropping to debug shell.
dracut:/#
Hacker voice I'm in. Now to enable networking and mount a test root. I have already extracted an Arch Linux root into a S3 bucket running locally, so this should be pretty easy, right? I just have to manually set up networking routes and load the drivers.
dracut:/# modprobe fuse
dracut:/# modprobe e1000
dracut:/# ip link set lo up
dracut:/# ip link set eth0 up
dracut:/# dhclient eth0
dhcp: PREINIT eth0 up
dhcp: BOUND setting up eth0
dracut:/# ip route add default via 10.0.2.2 dev eth0 proto dhcp src 10.0.2.15
dracut:/# s3fs -o url=http://192.168.2.209:9000 -o use_path_request_style fuse /sysroot
dracut:/# ls /sysroot
bin dev home lib64 opt root sbin sys usr
boot etc lib mnt proc run srv tmp var
dracut:/# switch_root /sysroot /sbin/init
switch_root: failed to execute /lib/systemd/systemd: Input/output error
dracut:/# ls
sh: ls: command not found
Honestly, I don't know what I expected. Seems like everything is just... gone. Alas, not even tab completion can save me. At this point, I was stuck. I had no idea what to do. I spent days just looking around, poking at the switch_root
source code, all for naught. Until I remembered a link Anthony had sent me: How to shrink root filesystem without booting a livecd. In there, there was a command called pivot_root
that switch_root
seems to call internally. Let's try that out.
dracut:/# logout
...
[ 430.817269] ---[ end Kernel panic - not syncing: Attempted to kill init! exitcode=0x00000100 ]---
...
dracut:/# cd /sysroot
dracut:/sysroot# mkdir oldroot
dracut:/sysroot# pivot_root . oldroot
pivot_root: failed to change root from `.' to `oldroot': Invalid argument
Apparently, pivot_root
is not allowed to pivot roots if the root being switched is in the initramfs. Unfortunate. The Stack Exchange answer tells me to use switch_root
, which doesn't work either. However, part of that answer sticks out to me:
initramfs is rootfs: you can neither pivot_root rootfs, nor unmount it. Instead delete everything out of rootfs to free up the space (find -xdev / -exec rm '{}' ';'), overmount rootfs with the new root (cd /newmount; mount --move . /; chroot .), attach stdin/stdout/stderr to the new /dev/console, and exec the new init.
Would it be possible to manually switch the root without a specialized system call? What if I just chroot?
...
dracut:/# mount --rbind /sys /sysroot/sys
dracut:/# mount --rbind /dev /sysroot/dev
dracut:/# mount -t proc /proc /sysroot/proc
dracut:/# chroot /sysroot /sbin/init
Explicit --user argument required to run as user manager.
Oh, I need to run the chroot
command as PID 1 so Systemd can start up properly. I can actually tweak the initramfs's init script and just put my startup commands in there, and replace the switch_root
call with exec chroot /sbin/init
.
I put this in modules.d/99base/init.sh
in the Dracut source after the udev rules are loaded and bypassed the root
variable checks earlier.
modprobe fuse
modprobe e1000
ip link set lo up
ip link set eth0 up
dhclient eth0
ip route add default via 10.0.2.2 dev eth0 proto dhcp src 10.0.2.15
s3fs -o url=http://192.168.2.209:9000 -o use_path_request_style fuse /sysroot
mount --rbind /sys /sysroot/sys
mount --rbind /dev /sysroot/dev
mount -t proc /proc /sysroot/proc
I also added exec chroot /sysroot /sbin/init
at the end instead of the switch_root
command.
Rebuilding the EFI image and...
![A screenshot of a Linux login screen](https://ersei.net/user/pages/03.blog/40.fuse-root/itworks.png)
I sit there, in front of my computer, staring. It can't have been that easy, can it? Surely, this is a profane act, and the spirit of Dennis Ritchie ought't've stopped me, right?
Nobody stopped me, so I kept going.
I log in with the very secure password root
as root
, and it unceremoniously drops me into a shell.
[root@archlinux ~]# mount
s3fs on / type fuse.s3fs (rw,nosuid,nodev,relatime,user_id=0,group_id=0)
...
[root@archlinux ~]#
At last, Linux booted off of an S3 bucket. I was compelled to share my achievement with others—all I needed was a fetch program to include in the screenshot:
[root@archlinux ~]# pacman -Sy fastfetch
:: Synchronizing package databases...
core.db failed to download
error: failed retrieving file 'core.db' from geo.mirror.pkgbuild.com : Could not resolve host: geo.mirror.pkgbuild.com
warning: fatal error from geo.mirror.pkgbuild.com, skipping for the remainder of this transaction
error: failed retrieving file 'core.db' from mirror.rackspace.com : Could not resolve host: mirror.rackspace.com
warning: fatal error from mirror.rackspace.com, skipping for the remainder of this transaction
error: failed retrieving file 'core.db' from mirror.leaseweb.net : Could not resolve host: mirror.leaseweb.net
warning: fatal error from mirror.leaseweb.net, skipping for the remainder of this transaction
error: failed to synchronize all databases (invalid url for server)
[root@archlinux ~]#
Uh, seems like DNS isn't working, and I'm missing dig
and other debugging tools.
Wait a minute! My root filesystem is on S3! I can just mount it somewhere else with functional networking, chroot
in, and install all my utilities!
Some debugging later, it seems like systemd-resolved doesn't want to run because it Failed to connect stdout to the journal socket, ignoring: Permission denied
. I'm not about to try to debug systemd because it's too complicated and I'm lazy, so instead I'll just use Cloudflare's.
[root@archlinux ~]# echo "nameserver 1.1.1.1" > /etc/resolv.conf
[root@archlinux ~]# pacman -Sy fastfetch
:: Synchronizing package databases...
core is up to date
extra is up to date
...
[root@archlinux ~]# fastfetch
![Fastfetch showing the system running in QEMU](https://ersei.net/user/pages/03.blog/40.fuse-root/fastfetch.png)
I look around, making sure that nobody had tried to stop me. My window was intact, my security system had not tripped, the various canaries I had set up around the house had not been touched. I was safe to continue.
I was ready to have it run on Google Drive.
Google Gets Involved
There's a project already that does Google Drive over FUSE for me already: google-drive-ocamlfuse. Thankfully, I have a Google account lying around that I haven't touched in years ready to go! I follow the instructions, accept the terms of service I didn't read, create all the oauth2 secrets, enable the APIs, install google-drive-ocamlfuse
from the AUR into my Arch Linux VM, patch some PKGBUILD
s (it's been a while), and lo and behold! I have mounted Google Drive! Mounting Drive and a few very longrsync
runs later, I have Arch Linux on Google Drive.
Just kidding, it's never that easy. Here's a non-exhausive list of problems I ran into:
- Symlinks to symlinks don't work (very important for stuff in
/usr/lib
) - Hardlinks don't work
- It's so slowwwww
- Relative symlinks don't work at all
- No dangling symlinks (important for stuff that links to
/proc
and isn't mounted, or stuff that just hasn't copied over yet) - Symlinks outside of Google Drive don't work
- Permissions don't work (neither do attributes)
- Did I mention it's SLOW
With how many problems there are with symlinks, I have half a mind to change the FUSE driver code to just create a file that ends in .internalsymlink
to fix all of that, Google Drive compatibility be damned.
But, I have challenged myself to do this without modifying anything important (no kernel tweaking, no FUSE driver tweaking), so I'll just have to live with it and manually create the symlinks that rsync
fails to make with a hacky sed
command to the rsync
error logs.
In the meantime, I added the token files generated from my laptop into the initramfs, as well as the Google Drive FUSE binary and SSL certificates, and tweaked a few settings2 to make my life slighty easier.
...
inst ./gdfuse-config /.gdfuse/default/config
inst ./gdfuse-state /.gdfuse/default/state
find /etc/ssl -type f -or -type l | while read file; do inst "$file"; done
find /etc/ca-certificates -type f -or -type l | while read file; do inst "$file"; done
...
![A screenshot of Google Drive showing the root of a typical Linux filesystem](https://ersei.net/user/pages/03.blog/40.fuse-root/google-drive-root.png)
It's nice to see that timestamps kinda work, at least. Now all that's left is to wait for the agonizingly slow boot!
chroot: /sbin/init: File not found
Perhaps they did not bother to stop me because they knew I would fail.
I know the file exists since, well, it exists, so why is it not found? Simple: Linux is kinda weird and if the binary you call depends on a library that's not found, then you'll get "File not found".
dracut:/# ldd /sysroot/bin/bash
linux-vdso.so.1 (0x00007e122b196000)
libreadline.so.8 => /usr/lib/libreadline.so.8 (0x00007e122b01a000)
libc.so.6 => /usr/lib/libc.so.6 (0x00007e122ae2e000)
libncursesw.so.6 => /usr/lib/libncursesw.so.6 (0x00007e122adbf000)
/lib64/ld-linux-x86-64.so.2 => /usr/lib64/ld-linux-x86-64.so.2 (0x00007e122b198000)
However, these symlinks don't actually exist! Remember how earlier we noted that relative symlinks don't work? Well, that's come back to bite me. The Kernel is looking for files in /sysroot
inside /sysroot/sysroot
. Luckily, this is an easy enough fix: we just need to have /sysroot
linked to /sysroot/sysroot
without links:
dracut:/# mkdir /sysroot/sysroot
dracut:/# mount --rbind /sysroot /sysroot/sysroot
Now time to boot!
It took five minutes for Arch to rebuild the dynamic linker cache, another minute per systemd unit, and then, nothing. The startup halted in its tracks.
[ TIME ] Timed out waiting for device /dev/ttyS0.
[DEPEND] Dependency failed for Serial Getty on ttyS0.
Guess I have to increase the timeout and reboot. In /etc/systemd/system/dev-ttyS0.device
, I put:
[Unit]
Description=Serial device ttyS0
DefaultDependencies=no
Before=sysinit.target
JobTimeoutSec=infinity
Luckily, it did not take infinite time to boot.
![A Linux login prompt](https://ersei.net/user/pages/03.blog/40.fuse-root/gdrive-booted.png)
I'm so close to victory I can taste it! I just have to increase another timeout. I set LOGIN_TIMEOUT
to 0
in /etc/login.defs
in Google Drive, and tried logging in again.
Thankfully, there's a cache, so subsequent file reads aren't nearly as slow.
![Fastfetch in Google Drive root, showing that the root partition is mounted as fuse.google-drive-ocaml](https://ersei.net/user/pages/03.blog/40.fuse-root/gdrive-fastfetch.png)
Here I am, laurel crown perched upon my head, my chimera of Linux and Google Drive lurching around.
But I'm not satisfied yet. Nobody had stopped me because they want me to succeed. I have to take this further. I need this to work on real hardware.
Now Do It On Real Hardware
Fortunately for me, I switched servers and now have an extra laptop with no storage just lying around! A wonderful victim3 for my test!
There are a few changes I'll have to make:
- Use the right ethernet driver and not the default
e1000
- Do not use a serial display
- Change the network settings to match my house's network topology
All I need is the r8169
driver for my ethernet port, and let's throw in a Powerline into the mix, because it's not going to impact the performance in any way that matters, and I don't have an ethernet cord that can reach my room.
I build the unified EFI file, throw it on a USB drive under /BOOT/EFI
, and stick it in my old server. Despite my best attempts, I couldn't figure out what the modprobe directive is for the laptop's built-in keyboard, so I just modprobed hid_usb
and used an external keyboard to set up networking.
![A screenshot of fastfetch and mount on bare metal showing that we're booted off of Google Drive](https://ersei.net/user/pages/03.blog/40.fuse-root/bare-metal-gdrive.png)
This is my magnum opus. My Great Work. This is the mark I will leave on this planet long after I am gone: The Cloud Native Computer.
Nice thing is, I can just grab the screenshot4 from Google Drive and put it here!
Woe! Cloud Native Computer Be Upon Ye
Despite how silly this project is, there are a few less-silly uses I can think of, like booting Linux off of SSH, or perhaps booting Linux off of a Git repository and tracking every change in Git using gitfs. The possibilities are endless, despite the middling usefulness.
If there is anything I know about technology, it's that moving everything to The Cloud is the current trend. As such, I am prepared to commercialize this for any company wishing to leave their unreliable hardware storage behind and move entirely to The Cloud. Please request a quote if you are interested in True Cloud Native Computing.
Unfortunately, I don't know what to do next with this. Maybe I should install Nix?
Thoughts? Comments? Opinions? Feel free to share (relevant) ones with me! Contact me here if you want.
0 Comments