boco/src/qemu.rs

//! QEMU: boot a registry entry with qemu-system-x86_64 using direct kernel
//! boot (-kernel/-initrd or -drive). The kernel is shared across all VMs
//! at $XDG_DATA_HOME/slim-rs/registry/vmlinuz. A universal /slim/init script (injected
//! at build time) handles VM bootstrap; /slim/exec runs the container's
//! CMD/ENTRYPOINT. Runtime overrides are delivered as `slim.cmd=<base64>`
//! on the kernel cmdline.

use anyhow::{Context, Result, bail};
use base64::Engine;
use clap::Args;
use std::{fmt::Write as _, process::Command};

use crate::{
    forward::PortForward,
    registry::{registry_base_dir, registry_dir},
};

#[derive(Args, Debug)]
pub struct RunCmd {
    /// Image tag / registry subdir name
    name: String,

    /// Amount of memory to give the VM, in qemu's format.
    #[clap(short, long, default_value = "1024M")]
    memory: String,

    /// Forward ports from host to guest. Example: `tcp:0.0.0.0:80-:8080`
    #[clap(long)]
    forward: Vec<PortForward>,

    /// Override the command to exec in the VM (base64-encoded on the kernel
    /// cmdline as slim.cmd=<b64>). Overrides the CMD inferred at build time.
    #[clap(long)]
    cmd: Option<String>,

    /// Share a host directory into the VM via 9p. Format:
    /// `<host-path>` or `<host-path>:<guest-path>`. When the guest path is
    /// omitted, the host path is used. Non-absolute guest paths are
    /// relative to the image's WORKINGDIR. Can be repeated.
    #[clap(long)]
    mount: Vec<String>,
}

pub(crate) fn run(
    RunCmd {
        name,
        memory,
        forward,
        cmd,
        mount,
    }: RunCmd,
) -> Result<()> {
    let reg_dir = registry_dir(&name)?;
    let vmlinuz_path = registry_base_dir()?.join("vmlinuz");
    let initrd_path = reg_dir.join("initrd");
    let qcow2_path = reg_dir.join("image.qcow2");
    if !vmlinuz_path.exists() {
        // TODO: guide user in how to set up a kernel
        bail!(
            "No kernel available. Place a vmlinuz at {}",
            vmlinuz_path.display()
        );
    }

    let mut cmdline: Vec<String> = vec![
        "console=ttyS0,115200".into(),
        "rw".into(),
        "earlyprintk=serial".into(),
        "nokaslr".into(),
        "init=/slim/init".into(),
        "devtmpfs.mount=1".into(), // Automatically mount /dev at boot
    ];

    let fs_args;
    let qcow2_arg = format!("file={},format=qcow2,if=virtio", qcow2_path.display());
    if qcow2_path.exists() {
        fs_args = vec!["-drive".to_string(), qcow2_arg];
        cmdline.extend(["root=/dev/vda".to_string(), "rootfstype=ext4".to_string()]);
    } else if initrd_path.exists() {
        let initrd_str = initrd_path.to_str().context("Invalid UTF-8")?.to_string();
        fs_args = vec!["-initrd".to_string(), initrd_str];
        cmdline.push("root=/dev/ram0".into());
    } else {
        bail!("Registry missing rootfs/initrd for '{name}'. Run `slim build` first.");
    }

    if let Some(cmd) = &cmd {
        let encoded = base64::engine::general_purpose::STANDARD.encode(cmd.as_bytes());
        cmdline.push(format!("slim.cmd={encoded}"));
    }

    // Build 9p shares for each --mount.  Tags are short (slim0, slim1, …)
    // because 9p mount_tag has a ~31-byte limit.  The guest destination path
    // is passed on the kernel cmdline as slim.mount=<tag>:<base64(guest_path)>.
    // If no guest path is specified, the host path is used as the guest path.
    let mut virtfs_args: Vec<String> = Vec::new();
    for (i, spec) in mount.iter().enumerate() {
        let (host_path, guest_path) = match spec.split_once(':') {
            Some((h, g)) => (h, g),
            None => (spec.as_str(), spec.as_str()),
        };
        let canonical = std::path::Path::new(host_path)
            .canonicalize()
            .with_context(|| format!("Cannot resolve mount path '{host_path}'"))?;
        let host_str = canonical
            .to_str()
            .context("Mount path is not valid UTF-8")?;
        let tag = format!("slim{i}");
        let dest_b64 = base64::engine::general_purpose::STANDARD.encode(guest_path.as_bytes());

        // QEMU's QemuOpts splits on commas — escape literal commas in the
        // path as ",," per QEMU convention.
        let host_escaped = host_str.replace(',', ",,");
        virtfs_args.push("-virtfs".into());
        virtfs_args.push(format!(
            "local,path={host_escaped},mount_tag={tag},security_model=mapped-xattr"
        ));
        cmdline.push(format!("slim.mount={tag}:{dest_b64}"));
    }

    println!("Booting {name} from registry: {}", reg_dir.display());

    let cmdline = cmdline.join(" ");

    // add NIC and forward any user specified ports
    let mut network = "user,model=virtio-net-pci".to_string();
    for forward in &forward {
        _ = write!(&mut network, ",hostfwd={forward}");
    }

    let mut command = Command::new("qemu-system-x86_64");

    // TODO: make a lot of this configurable. especially -cpus
    command
        .args(["-accel", "kvm"])
        .args(["-cpu", "host", "-smp", "cpus=8"])
        .args(["-m", &memory])
        .arg("-kernel")
        .arg(vmlinuz_path.as_os_str())
        .args(&fs_args)
        .args(&virtfs_args)
        .args(["-snapshot"])
        .args(["-no-reboot"])
        .args(["-append", &cmdline])
        .args(["-nographic"])
        .args(["-nic", &network]);

    println!("{command:?}");

    let status = command
        .status()
        .context("Failed to launch qemu-system-x86_64")?;

    match status.code() {
        Some(0) => println!("QEMU exited cleanly."),
        Some(126) | Some(127) => {
            bail!("Failed to start qemu-system-x86_64 (is it installed and in PATH?)")
        }
        Some(code) => bail!("QEMU exited with code {code}"),
        None => bail!("QEMU terminated by a signal"),
    }
    Ok(())
}