boco/src/qemu.rs

//! See [`boot`]

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

use crate::{
    forward::PortForward,
    registry::{registry_dir, vmlinuz_path},
    virtiofs::{self, VirtiofsShare},
};

#[derive(Args, Debug)]
pub struct BootCmd {
    /// 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 `boco.cmd=<b64>`). Overrides the CMD inferred at build time.
    #[clap(long)]
    cmd: Option<String>,

    /// Share a host directory into the VM via virtio-fs. 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>,
}

/// Invoke `qemu` to boot a boco image. See [`BootCmd`] for options
///
/// This uses qemu's direct-kernel-boot with `-kernel`, so it does not require a bootloader.
/// The boco image is just an ext4 rootfs.
pub(crate) fn boot(
    BootCmd {
        name,
        memory,
        forward,
        cmd,
        mount,
    }: BootCmd,
) -> Result<()> {
    let reg_dir = registry_dir(&name)?;
    let vmlinuz_path = vmlinuz_path()?;
    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. Expected vmlinuz at {}",
            vmlinuz_path.display()
        );
    }

    let mut cmdline: Vec<String> = vec![
        "console=ttyS0,115200".into(),
        "rw".into(),
        "earlyprintk=serial".into(),
        "nokaslr".into(),
        "init=/boco/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 `boco build` first.");
    }

    if let Some(cmd) = &cmd {
        let argv = ["/bin/sh".to_string(), "-c".to_string(), cmd.clone()];
        let data: Vec<u8> = argv.join("\0").into_bytes();
        let encoded = base64::engine::general_purpose::STANDARD.encode(&data);
        cmdline.push(format!("boco.cmd={encoded}"));
    }

    // Build virtio-fs shares for each `--mount`.  Each share gets its own
    // virtiofsd daemon under `podman unshare` (rootless UID mapping) and a
    // matching `-chardev socket` + `-device vhost-user-fs-pci` pair on the
    // QEMU side. The guest destination path is passed on the kernel cmdline
    // as `boco.mount=<tag>:<base64(guest_path)>` and resolved by /boco/init.
    let mut shares: Vec<VirtiofsShare> = Vec::with_capacity(mount.len());
    let mut fsdev_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.to_string()),
            None => (spec.as_str(), spec.to_string()),
        };
        let share = virtiofs::spawn_share(i, host_path, guest_path)?;
        let tag = share.tag.clone();
        let chardev = format!("char{i}");
        let socket = share
            .socket_path()
            .to_str()
            .context("socket path is not UTF-8")?
            .to_string();
        let dest_b64 =
            base64::engine::general_purpose::STANDARD.encode(share.guest_path.as_bytes());

        fsdev_args.push("-chardev".into());
        fsdev_args.push(format!("socket,id={chardev},path={socket}"));
        fsdev_args.push("-device".into());
        fsdev_args.push(format!(
            "vhost-user-fs-pci,queue-size=1024,chardev={chardev},tag={tag}"
        ));
        cmdline.push(format!("boco.mount={tag}:{dest_b64}"));

        shares.push(share);
    }

    // vhost-user devices require shared guest RAM so virtiofsd (a separate
    // process) can DMA into it.
    let memory_args: Vec<String> = if shares.is_empty() {
        vec!["-m".into(), memory.clone()]
    } else {
        vec![
            "-m".into(),
            memory.clone(),
            "-object".into(),
            format!("memory-backend-memfd,id=mem,size={memory},share=on"),
            "-numa".into(),
            "node,memdev=mem".into(),
        ]
    };

    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(&memory_args)
        .arg("-kernel")
        .arg(vmlinuz_path.as_os_str())
        .args(&fs_args)
        .args(&fsdev_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")?;

    // Ensure virtiofsd daemons are torn down even on error returns below.
    drop(shares);

    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(())
}