Add cgroup v2 support

cgroup v1 uses two special files to determine access, where cgroup v2
uses eBPF programs to control access. The code will attach a custom eBPF
program which allows run-time reconfiguration and detach docker's default.

eBPF programs will be detached when the attaching program dies, which
can be dangerous if container-hotplug exits unexpectedly while the program
is running, so we instead pin it (so it stays when the program exits) and
unpin it after the docker container is down. In this case we might have
garbage eBPF programs pinned when container-hotplug exits unexpectedly but
it is safe.

Cargo.toml

@@ -30,6 +30,7 @@ bollard = "0.16"
 futures = "0.3"
 rustix = { version = "0.38", features = ["fs", "stdio", "termios"] }
 bitflags = "2"
+aya = { git = "https://github.com/aya-rs/aya.git" }
 
 [build-dependencies]
 anyhow = { version = "1", features = ["backtrace"] }

README.md

@@ -25,10 +25,7 @@ Another concern is providing a container with well known paths for the devices.
 On bare-metal systems this would usually be achieved with a `SYMLINK` directive in a udev rule.
 This program tries to provide a similar functionality for containers, allowing you to specify symlinks for certain devices.
 
-## Limitations
-
-`container-hotplug` needs to be run as root and relies on `cgroup v1`. It does not support `cgroup v2`.
-On distributions with `cgroup v2`, you can switch back to `cgroup v1` by setting the [kernel parameter](https://wiki.ubuntu.com/Kernel/KernelBootParameters) `systemd.unified_cgroup_hierarchy=0`.
+This tool supports both cgroup v1 and v2.
 
 ## Example
 

src/docker/cgroup.rs

@@ -1,4 +1,8 @@
-use anyhow::{ensure, Result};
+use anyhow::{ensure, Context, Result};
+use aya::maps::{HashMap, MapData};
+use aya::programs::{CgroupDevice, Link};
+use std::fs::File;
+use std::mem::ManuallyDrop;
 use std::path::PathBuf;
 
 // The numerical representation below needs to match BPF_DEVCG constants.
@@ -26,6 +30,10 @@ pub trait DeviceAccessController {
         minor: u32,
         access: Access,
     ) -> Result<()>;
+
+    /// Stop performing access control. This may allow all accesses, so should only be used when
+    /// the cgroup is shutdown.
+    fn stop(self: Box<Self>) -> Result<()>;
 }
 
 pub struct DeviceAccessControllerV1 {
@@ -96,4 +104,95 @@ impl DeviceAccessController for DeviceAccessControllerV1 {
 
         Ok(())
     }
+
+    fn stop(self: Box<Self>) -> Result<()> {
+        Ok(())
+    }
+}
+
+#[repr(C)] // This is read as POD by the BPF program.
+#[derive(Clone, Copy)]
+struct Device {
+    device_type: u32,
+    major: u32,
+    minor: u32,
+}
+
+// SAFETY: Device is `repr(C)`` and has no padding.
+unsafe impl aya::Pod for Device {}
+
+pub struct DeviceAccessControllerV2 {
+    map: HashMap<MapData, Device, u32>,
+    pin: PathBuf,
+}
+
+impl DeviceAccessControllerV2 {
+    pub fn new(id: &str) -> Result<Self> {
+        // We want to take control of the device cgroup filtering from docker. To do this, we attach our own
+        // filter program and detach the one by docker.
+        let cgroup_path = format!("/sys/fs/cgroup/system.slice/docker-{id}.scope");
+        let cgroup = File::open(cgroup_path)?;
+
+        let mut bpf = aya::Bpf::load(include_bytes!(concat!(
+            env!("CARGO_MANIFEST_DIR"),
+            "/cgroup_device_filter/target/bpfel-unknown-none/release/cgroup_device_filter"
+        )))?;
+
+        let program: &mut CgroupDevice = bpf
+            .program_mut("check_device")
+            .context("cannot find check_device program")?
+            .try_into()?;
+
+        program.load()?;
+
+        // Iterate existing programs. We'll need to detach them later.
+        // Wrap this inside `ManuallyDrop` to prevent accidental detaching.
+        let existing_programs = ManuallyDrop::new(CgroupDevice::query(&cgroup)?);
+
+        program.attach(&cgroup)?;
+
+        // Pin the program so that if container-hotplug accidentally exits, the filter won't be removed from the docker
+        // container.
+        let pin: PathBuf = format!("/sys/fs/bpf/docker-{id}-device-filter").into();
+        program.pin(&pin)?;
+
+        // Now our new filter is attached, detach all docker filters.
+        for existing_program in ManuallyDrop::into_inner(existing_programs) {
+            existing_program.detach()?;
+        }
+
+        let map: HashMap<_, Device, u32> = bpf
+            .take_map("DEVICE_PERM")
+            .context("cannot find DEVICE_PERM map")?
+            .try_into()?;
+
+        Ok(Self { map, pin })
+    }
+}
+
+impl DeviceAccessController for DeviceAccessControllerV2 {
+    fn set_permission(
+        &mut self,
+        ty: DeviceType,
+        major: u32,
+        minor: u32,
+        access: Access,
+    ) -> Result<()> {
+        let device = Device {
+            device_type: ty as u32,
+            major,
+            minor,
+        };
+        if access.is_empty() {
+            self.map.remove(&device)?;
+        } else {
+            self.map.insert(device, access.bits(), 0)?;
+        }
+        Ok(())
+    }
+
+    fn stop(self: Box<Self>) -> Result<()> {
+        CgroupDevice::from_pin(&self.pin)?.unpin()?;
+        Ok(())
+    }
 }