It's been a bit of a slow week, I'm afraid.
spectrum-vm
-----------
As mentioned last week, it was becoming extremely inconvenient to test
out custom kernel/crosvm/rootfs builds by modifying a Nix expression
every time. So I wrote spectrum-vm, a small program that takes command
line arguments for kernel, crosvm, and rootfs paths to use, and falls
back to pre-built defaults if any of those aren't specified. This makes
turnaround time on testing different kernel builds much quicker, because
they can be built incrementally outside of nix sandboxes.
To make things more concrete, suppose I have built a kernel in
/home/src/linux/arch/x86_64/boot/bzImage, and I want to test it with the
Spectrum crosvm demo setup. I can run
spectrum-vm -k /home/src/linux/arch/x86_64/boot/bzImage
and I get a VM booted from that kernel, with no need to nix-build
anything.
By default, the launched VM will run Weston Terminal inside a Wayland
compositor. But you can even replace this with a custom command:
spectrum-vm -c 'echo "Hello, world!"'
The implementation of that last bit is quite the hack. I'll leave
figuring out how it works as an exercise to the reader. ;)
spectrum-vm is available now in Spectrum's Nixpkgs as
spectrumPackages.spectrum-vm, as of commit
817a1a9b0de30d8bbc2ad2e80a802d87c56e96ff. Implementing this has meant
that the huge start-vm.nix file that used to exist on the crosvm-demo
branch had to be split up into a bunch of discreet units. This is
really great, because start-vm.nix was starting to grow enormous, and
experimenting with changes was becoming A Problem. Now, that's all nice
and modular under the new spectrumPackages namespace. And _that_ means
that the crosvm-demo branch can finally go away, and Spectrum's Nixpkgs
is no longer split across two branches. (The purpose of the separation
was to avoid there being a horrible start-vm.nix file in the root on
master.) So in addition to a lovely helpful tool, this has also
resulted in much cleaner code. Yay!
One feature I would still like to add is an option that maps to crosvm
run's --shared-dir, to mount a directory into a VM over 9p. That would
make getting code into a VM to test it extremely convenient. I haven't
gotten around to it yet though.
Guest memory sharing
--------------------
Last week I shared a bunch of research I had done on sharing
guest-allocated buffers with other VMs. I was hoping to avoid having to
patch a Wayland compositor to request buffers from the host instead of
allocating them itself, but I've since come to the conclusion that
avoiding that won't be feasible, for now at least.
Basically, the kernel primitives aren't there yet, and it would be silly
for me, with almost zero kernel development experience, to attempt to
write them. There are all sorts of horrible edge cases to think about,
like what happens if the guest-allocated memory moves around. Compared
to this, patching a Wayland compositor doesn't seem too bad. Having to
go through a request/response over a vsock to the host will be a bit of
a performance overhead compared to a smarter solution, but I think this
kind of allocation happens rarely enough that it probably doesn't matter
anyway.
A nice middle ground would have been to use virtio-gpu resources rather
than a custom message over a vsock. It would have been a much cleaner
way to do it, and it might not even involve writing any kernel code, but
the practical benefits of it would be extremely limited when contrasted
with the time it would take me to implement that.
I don't expect the approach I'm taking now of patching the Wayland
compositor to necessarily be permanent. As I linked to last week,
there are kernel efforts happening to add the necessary primitives to
share guest allocations. At some point in the future we might be able
to use them. And at basically any time we could switch to using
virtio-gpu resources if somebody wanted to put in the work.
Patching a compositor really isn't a big deal, though. It's unlikely
Spectrum will support more than one compositor anyway, because we'll
need to customise it to do things like Qubes-style window borders that
are coloured based on security domain. And, even then, if we're using a
wlroots-based compositor, it's probably only wlroots that needs
patching, so in that case any other wlroots-based compositor would
probably work. Speaking of compositors...
Wayfire
-------
If you know what Wayfire is, you're probably surprised to see it as an
item in a Spectrum status update it. If you don't, Wayfire is a Wayland
compositor notable for having a bunch of over-the-top Compiz-style
graphical effects, like wobbly windows and the titular fire. I
recommend checking out the demo videos on https://wayfire.org/ if you
haven't seen them -- it's quite something.
But I bring this up not because I think support for wobbly windows is
imperative to Spectrum's chances of success, but because fundamentally
Wayfire is just an extremely modular Wayland compositor. Virtually
everything in Wayfire is implemented as a plugin. Switching between
windows is a plugin, screen zoom is a plugin, and so on.
This is really cool because in Spectrum we'll want to have, for example,
coloured window decorations like in Qubes. With any other compositor,
we'd have to hack that into the compositor internals somehow. With
Wayfire, it can just be a plugin.
Wayfire is also a pleasantly simple but featureful windowing
experience. You can drag windows around, put two side by side by
dragging them to screen edges, have multiple workspaces, etc. I think
it will be straightforward enough to not be a problem for people coming
to Spectrum regardless of their computing background, and powerful
enough to not get in the way of Serious Computer Business. And anything
missing could of course be implemented as a plugin. ;)
So, I've been doing some work to package Wayfire in Nixpkgs. This is
not all that easy, because Wayfire seems to assume that plugins will be
installed into Wayfire's prefix. This isn't really possible with Nix
unless you compile all the plugins along with Wayfire, so I'm working
some patches for Wayfire to look up plugins in a search path specified
as an environment variable, which on the Nix side we can then wrap in
the idiomatic wayfireWithPackages style. My initial impressions of the
Wayfire codebase are that it is relatively clean, and straightforward
enough to understand. The core compositor is about 13000 SLoC, which is
pleasantly small.
That, unfortunately, is it for this week. I didn't get as much done as
I'd hoped, but I did at least have loads of good thoughts about things
to do later in Spectrum's development, that aren't quite concrete enough
to put into words yet.
Wow, a week isn't very long at all. But still, there's definitely
enough to write about, even if I spent most of the week just reading
mailing list threads. More on that in a bit.
Spectrum Live + nixpkgs
-----------------------
I did another stream. This time, we looked at updating the Chromium OS
packages in Spectrum. The script I used to update them had broken -- it
was stuck on an old version, so first we had to fix the script, and then
fix the package builds. A lot of the Chromium OS packages aren't
actually necessary at the moment, but I made them a while ago when it
looked like they would be. I couldn't get one particular such package,
libchrome, to build any more at all after spending quite a few hours on
it, so I ended up just dropping it, along with libbrillo, which depended
on it. Some other package updates had some complexity, but we managed
to work through them all on the stream.
This stream was much more focused on what I was working on than the last
one. The last one was more about people asking general questions and me
answering them and demonstrating things. This time I stayed on the
package updates for the whole stream. I figure people asked all their
questions last time. But there were lots of questions about specific
issues in the package updates, etc. It was nice to know that I had
explained things sufficiently well that people could follow along with
what I was doing. Viewer numbers were up too, which was nice. We were
on 15 for most of the stream.
The packaging work I mostly did on the stream, and then finished up
afterwards is here:
https://spectrum-os.org/git/nixpkgs/log/?h=2bd191e91f097ff2146099fae8b7c8bd…
And a couple of small upstream Nixpkgs PRs:
https://github.com/NixOS/nixpkgs/pull/85549https://github.com/NixOS/nixpkgs/pull/85551
crosvm still needs to be updated in upstream Nixpkgs. If you want to
help out, this might be a good thing to pick up! You'd need to copy the
update.py change I made in Spectrum's Nixpkgs to the crosvm update.py in
upstream Nixpkgs, and then copy the changes in Spectrum nixpkgs' crosvm
to upstream Nixpkgs's crosvm.
If not, I'll probably get to it in a few days.
crosvm
------
Last week, I said that we might me able to use Virtio Wayland as a
transport for crosvm device <-> crosvm communication across a VM
boundary, but that I'd just had the idea and I might decide it couldn't
work.
Having researched it extensively this week, reading over the code for
the virtio_wl kernel driver with puck and consuming a huge number of
messages on assorted kernel-adjacent mailing lists, I think we can make
it work, and it's our best path forward, but I'm still not sure how
exactly. The main problem is that Virtio Wayland can only send shared
memory from host to guest, not the other way around. So when a Wayland
compositor in a guest hands the Wayland crosvm device, it's not going to
be able to pass it over the VM boundary to the crosvm instance running a
Wayland client.
Sharing a guest memory buffer with the host, or another guest, turns out
not to be a solved problem in general. As an example, here is a very
long mailing list thread about how a generic way to do this could work,
from November/December, which reached no firm conclusions:
https://lore.kernel.org/linux-media/20191105105456.7xbhtistnbp272lj@sirius.…
I expect that for most devices, e.g. the network driver, this won't be a
problem. But we're definitely going to need this for Wayland. One
option would be to modify a Wayland compositor to make it ask the host
for some shared memory, rather than allocating its own shared memory
buffer. That would work, but it's the last resort, because it would
have to be a modification made to any Wayland compositor we might want
to use.
At a high level, what we want to do is have a crosvm device running
inside a device VM tell crosvm that a certain address in guest memory
is some shared memory, and have it handle that as it would shared memory
it had allocated itself.
crosvm does support sending dmabufs from guest to host, because of
virtio-gpu. There's a relatively new kernel API called udmabuf that
converts a memfd into a dmabuf, which sounds like it should solve
exactly this problem. It has no documentation in the kernel, but I did
manage to find a small amount of documentation in a patch that was for
some reason never applied[1]. I then managed to write a small test
program that at least called all the ioctls properly and ended up with a
dmabuf. Actually testing this any further will involve a VM, and
testing this sort of thing in a VM at the moment is a real pain, and I
want to make a small tooling improvement around that before I proceed
any further with testing udmabuf. There are also good reasons to
suspect that the udmabuf idea won't work anyway. puck thinks that the
only dmabufs Virtio Wayland will be able to transfer are those allocated
by virtio-gpu. So, not dmabufs that we just made out of a memfd given
to us by a Wayland compositor.
If udmabuf ends up not working, I'll want to see whether we can use
virtio-gpu to allocate dmabufs. The problem with this is that I don't
know how we'll bridge compositor shared memory to virtio-gpu dmabufs.
It'll presumably involve writing kernel code, which I've been pleased to
avoid so far. I can at least probably use the udmabuf implementation as
a starting point.
The third least bad option would be to modify the virtio_wl kernel
driver so it supports shared memory originating from the guest. This
would be very nice to use from userspace, but I just don't think there's
any way I'll be able to write the required kernel code.
And the fourth least bad option (the worst one) would be to modify the
compositor as mentioned above.
spectrum-vm
-----------
The current way of running a Spectrum test VM is to build and run a file
called start-vm.nix that lives in the root of a non-default branch of
Spectrum's nixpkgs called crosvm-demo. Any time you want to try a
custom kernel, rootfs, or crosvm, you need to modify this file. This is
extremely annoying. Also, as time goes on, start-vm.nix is getting
longer and longer, and more and more difficult to understand. So I
think it's finally time to do something better. The something better is
that we introduce derivations for a Spectrum rootfs, guest kernel, and
crosvm, and then provide a program, spectrum-vm, that by default runs a
VM with those three components. It would also have options for
overriding those components, so if you built a custom kernel you could
just drop that one in instead of the default one, for example. And
finally, it would allow changing the flags passed to crosvm.
I started working on this this morning, and I'm quite excited about it.
It's not going to take very long at all to do, but it's going to be such
a big improvement. It will also mean I no longer have a separate
crosvm-demo branch, to merge Spectrum nixpkgs master into every time I
update it. So I'm going to spend the next couple of days on this, and
then get back into investigating udmabuf with a nice new tool.
So, that's what happened this week. I feel like I just spent most of
the week reading mailing lists and not making much progress, but I've
managed to write quite a lot (again) so I suppose I've accomplished
something!
Now, I've been extremely immersed in this stuff all week, and I suspect
that some of this update, especially the crosvm stuff, will not be clear
to somebody who has not been similarly immersed. And, it's getting
dangerously close to midnight UTC, my self-imposed This Week in Spectrum
deadline. So I'm going to send this as is, even though it's probably
confusing. But! I do want you to understand what's going on. So
please feel free to reply and point out anything I didn't explain very
well and ask me for a better explanation. Just be prepared for me not
to know the answer, since I'm still trying to figure out a lot of this
stuff! As a nice side effect, maybe we'll get a mailing list discussion
going. :)
[1]: https://lore.kernel.org/lkml/20180911134216.9760-14-kraxel@redhat.com/
Before we begin, **if you are subscribed only to the discuss mailing
list**, I'd like to hear from you whether you'd like future weekly
development updates to come to you or not, because I'm not really sure
whether this sort of update belongs on the devel mailing list, or the
discuss one. So for this initial edition, I'm sending to both, but I
probably won't do that going forward.
So, as I wrote in the recent status update on my blog, I'm going to try
weekly(?) mailing updates instead of lengthy blog posts every few
months. I was going to start last week, but I ended up taking most of
the week off because I'd been working every day for a long time and was
starting to find it difficult to concentrate. But I'm back now, and
this week has been a good one.
Spectrum Live
-------------
(This was actually last week, but I felt this deserved a mention.)
A bit more than a week ago now, I did a live stream where I showed
myself working on a fairly simple change to crosvm. Viewer numbers
stabilised at around 11, and we had a lively IRC conversation, which I
was pleasantly surprised by. I ended up spending most of the stream not
working on that change, but instead explaining how various things in the
system will work, and helping getting a couple of people started running
the Spectrum code. By the end of the stream two people had succeeded in
this, and that means two people who are set up to work on Spectrum. One
was even starting work on trying to build a smaller kernel for Spectrum
VMs!
Overall I think this was extremely successful, and I'll try to do it
again soon. I'll announce the stream a bit ahead of time in #spectrum
on Freenode.
crosvm
------
Having previously split crosvm into two separate programs -- one to run
a virtual device, and one to run a device VM -- I further split it into
three programs. This might sound a bit silly, but here's the idea --
one program is the crosvm that runs an application VM, the second will
become part of the crosvm that runs a device VM, and the third will be
the device implementation itself, that will run inside the device VM.
As a reminder, the overall idea here is to put virtual devices into
their own VMs, rather than being run directly on the host.
I also pulled in the latest changes from Google. Of particular note is
commit 146450b, which updates the msg_socket library to support
dynamically sized messages. If you're paying attention, you'll recall
that I just wrote my own serde-based replacement for msg_socket because
I needed to do exactly this. That work might now have been superseded,
but I'm not sure yet. msg_socket2 uses SOCK_STREAM sockets (which can
be used between VMs or VM<->guest), while msg_socket uses
SOCK_SEQPACKET, which can't. I haven't checked whether msg_socket's new
implementation could work without SOCK_SEQPACKET-style preserved message
boundaries -- if it can't, msg_socket2 will stick around. msg_socket2
also properly handles file descriptor lifetimes using the borrow
checker. This means that safe Rust code can't send closed or
non-existent file descriptors over msg_socket2. msg_socket does not do
this, because it just represents file descriptors as integers with no
lifetime information. Whatever happens to msg_socket2, at some point
I'll want to make msg_socket do file descriptor lifetimes properly. In
general, it's not clear from existing documentation how to properly
handle file descriptor lifetimes in Rust, and so I plan a blog post on
this topic soon.
The next problem is to figure out communication between the device VM
and the device crosvm. The protocols currently in use rely on being
able to share memory using file descriptors, which is not possible
between a VM and the host. To avoid using file descriptors, I would
have to change those protocols, possibly on a per-device basis, which
would be a lot of work, and might have to be done again whenever crosvm
adds a new device type. But, I think I might have a better, if
surprising solution. Despite its name, the Virtio Wayland protocol has
almost nothing to do with Wayland. Instead, it's a fairly generic
socket-like protocol that allows sending "virtual file descriptors" of
certain specific types between a VM and its host. So I think I should
be able to send this sort of communication over "Virtio Wayland",
without requiring any changes to the communications being channeled. I
just had this idea, so it's entirely possible it won't work out, but I
think it's a neat unexpected reuse of a cool technology.
Infrastructure
--------------
The web interface for the Spectrum mailing lists has been upgraded from
Django 1.11 (which is now EOL) to Django 2.2 (current LTS; Nixpkgs
default). I sent this upstream as Nixpkgs PR #84896.
Additionally, spectrum-os.org. now has an SPF header, which I thought
was already there but apparently wasn't. This should help with mailing
list deliverability.
That's it for this week. Feel free to reply with any comments or
insights, either on the content or the format. I'll try to be back with
another development update next week.
