2021-07-15 22:12:21 +02:00
|
|
|
use anyhow::{bail, Context, Result};
|
2021-07-15 01:09:04 +02:00
|
|
|
use caps::Capability;
|
2021-07-20 20:59:34 +02:00
|
|
|
use oci_spec::{LinuxCapabilities, LinuxNamespace, LinuxNamespaceType, Process, Spec};
|
2021-07-15 22:12:21 +02:00
|
|
|
|
2021-07-04 22:44:07 +02:00
|
|
|
use std::{
|
|
|
|
collections::HashMap,
|
2021-07-15 01:09:04 +02:00
|
|
|
convert::TryFrom,
|
|
|
|
ffi::{CString, OsString},
|
2021-07-15 22:12:21 +02:00
|
|
|
fs,
|
2021-07-15 01:09:04 +02:00
|
|
|
os::unix::prelude::OsStrExt,
|
2021-07-04 22:44:07 +02:00
|
|
|
path::{Path, PathBuf},
|
2021-07-15 01:09:04 +02:00
|
|
|
str::FromStr,
|
|
|
|
};
|
|
|
|
|
|
|
|
use crate::{
|
2021-07-19 19:25:01 +02:00
|
|
|
notify_socket::NotifySocket, rootless::detect_rootless, stdio::FileDescriptor, tty, utils,
|
2021-07-04 22:44:07 +02:00
|
|
|
};
|
|
|
|
|
2021-07-15 01:09:04 +02:00
|
|
|
use super::{builder::ContainerBuilder, builder_impl::ContainerBuilderImpl, Container};
|
2021-07-04 22:44:07 +02:00
|
|
|
|
2021-07-15 01:09:04 +02:00
|
|
|
const NAMESPACE_TYPES: &[&str] = &["ipc", "uts", "net", "pid", "mnt", "cgroup"];
|
2021-07-15 22:12:21 +02:00
|
|
|
const TENANT_NOTIFY: &str = "tenant-notify-";
|
|
|
|
const TENANT_TTY: &str = "tenant-tty-";
|
2021-07-04 22:44:07 +02:00
|
|
|
|
2021-07-05 19:40:35 +02:00
|
|
|
/// Builder that can be used to configure the properties of a process
|
|
|
|
/// that will join an existing container sandbox
|
2021-07-04 22:44:07 +02:00
|
|
|
pub struct TenantContainerBuilder {
|
|
|
|
base: ContainerBuilder,
|
|
|
|
env: HashMap<String, String>,
|
|
|
|
cwd: Option<PathBuf>,
|
2021-07-15 01:09:04 +02:00
|
|
|
args: Vec<String>,
|
|
|
|
no_new_privs: Option<bool>,
|
|
|
|
capabilities: Vec<String>,
|
|
|
|
process: Option<PathBuf>,
|
2021-07-04 22:44:07 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
impl TenantContainerBuilder {
|
2021-07-05 19:40:35 +02:00
|
|
|
/// Generates the base configuration for a process that will join
|
|
|
|
/// an existing container sandbox from which configuration methods
|
|
|
|
/// can be chained
|
2021-07-04 22:44:07 +02:00
|
|
|
pub(super) fn new(builder: ContainerBuilder) -> Self {
|
|
|
|
Self {
|
|
|
|
base: builder,
|
|
|
|
env: HashMap::new(),
|
|
|
|
cwd: None,
|
2021-07-15 01:09:04 +02:00
|
|
|
args: Vec::new(),
|
|
|
|
no_new_privs: None,
|
|
|
|
capabilities: Vec::new(),
|
|
|
|
process: None,
|
2021-07-04 22:44:07 +02:00
|
|
|
}
|
|
|
|
}
|
|
|
|
|
2021-07-05 19:40:35 +02:00
|
|
|
/// Sets environment variables for the container
|
2021-07-04 22:44:07 +02:00
|
|
|
pub fn with_env(mut self, env: HashMap<String, String>) -> Self {
|
|
|
|
self.env = env;
|
|
|
|
self
|
|
|
|
}
|
|
|
|
|
2021-07-05 19:40:35 +02:00
|
|
|
/// Sets the working directory of the container
|
2021-07-04 22:44:07 +02:00
|
|
|
pub fn with_cwd<P: Into<PathBuf>>(mut self, path: P) -> Self {
|
|
|
|
self.cwd = Some(path.into());
|
|
|
|
self
|
|
|
|
}
|
|
|
|
|
2021-07-05 19:40:35 +02:00
|
|
|
/// Sets the command the container will be started with
|
2021-07-15 01:09:04 +02:00
|
|
|
pub fn with_container_args(mut self, args: Vec<String>) -> Self {
|
|
|
|
self.args = args;
|
|
|
|
self
|
|
|
|
}
|
|
|
|
|
|
|
|
pub fn with_no_new_privs(mut self, no_new_privs: bool) -> Self {
|
|
|
|
self.no_new_privs = Some(no_new_privs);
|
|
|
|
self
|
|
|
|
}
|
|
|
|
|
|
|
|
pub fn with_capabilities(mut self, capabilities: Vec<String>) -> Self {
|
|
|
|
self.capabilities = capabilities;
|
|
|
|
self
|
|
|
|
}
|
|
|
|
|
|
|
|
pub fn with_process<P: Into<PathBuf>>(mut self, path: P) -> Self {
|
|
|
|
self.process = Some(path.into());
|
2021-07-04 22:44:07 +02:00
|
|
|
self
|
|
|
|
}
|
|
|
|
|
2021-07-05 19:40:35 +02:00
|
|
|
/// Joins an existing container
|
2021-07-04 22:44:07 +02:00
|
|
|
pub fn build(self) -> Result<()> {
|
|
|
|
let container_dir = self.lookup_container_dir()?;
|
2021-07-15 01:09:04 +02:00
|
|
|
let container = self.load_container_state(container_dir.clone())?;
|
|
|
|
let mut spec = self.load_init_spec(&container_dir)?;
|
|
|
|
self.adapt_spec_for_tenant(&mut spec, &container)?;
|
|
|
|
log::debug!("{:#?}", spec);
|
2021-07-04 22:44:07 +02:00
|
|
|
|
2021-07-19 08:22:47 +02:00
|
|
|
let notify_path = Self::setup_notify_listener(&container_dir)?;
|
2021-07-04 22:44:07 +02:00
|
|
|
// convert path of root file system of the container to absolute path
|
|
|
|
let rootfs = fs::canonicalize(&spec.root.path)?;
|
|
|
|
|
|
|
|
// if socket file path is given in commandline options,
|
|
|
|
// get file descriptors of console socket
|
2021-07-15 22:12:21 +02:00
|
|
|
let csocketfd = self.setup_tty_socket(&container_dir)?;
|
2021-07-04 22:44:07 +02:00
|
|
|
|
2021-07-15 01:09:04 +02:00
|
|
|
let use_systemd = self.should_use_systemd(&container);
|
2021-07-04 22:44:07 +02:00
|
|
|
let rootless = detect_rootless(&spec)?;
|
|
|
|
|
|
|
|
let mut builder_impl = ContainerBuilderImpl {
|
|
|
|
init: false,
|
|
|
|
syscall: self.base.syscall,
|
|
|
|
container_id: self.base.container_id,
|
|
|
|
pid_file: self.base.pid_file,
|
|
|
|
console_socket: csocketfd,
|
2021-07-15 22:12:21 +02:00
|
|
|
use_systemd,
|
|
|
|
container_dir,
|
2021-07-04 22:44:07 +02:00
|
|
|
spec,
|
|
|
|
rootfs,
|
|
|
|
rootless,
|
2021-07-19 08:22:47 +02:00
|
|
|
notify_path: notify_path.clone(),
|
2021-07-04 22:44:07 +02:00
|
|
|
container: None,
|
|
|
|
};
|
|
|
|
|
|
|
|
builder_impl.create()?;
|
2021-07-15 01:09:04 +02:00
|
|
|
|
2021-07-15 22:12:21 +02:00
|
|
|
let mut notify_socket = NotifySocket::new(notify_path);
|
2021-07-15 01:09:04 +02:00
|
|
|
notify_socket.notify_container_start()?;
|
2021-07-04 22:44:07 +02:00
|
|
|
Ok(())
|
|
|
|
}
|
|
|
|
|
|
|
|
fn lookup_container_dir(&self) -> Result<PathBuf> {
|
|
|
|
let container_dir = self.base.root_path.join(&self.base.container_id);
|
|
|
|
if !container_dir.exists() {
|
|
|
|
bail!("container {} does not exist", self.base.container_id);
|
|
|
|
}
|
|
|
|
|
|
|
|
Ok(container_dir)
|
|
|
|
}
|
|
|
|
|
|
|
|
fn load_init_spec(&self, container_dir: &Path) -> Result<Spec> {
|
|
|
|
let spec_path = container_dir.join("config.json");
|
|
|
|
|
2021-07-16 00:41:09 +02:00
|
|
|
let spec = oci_spec::Spec::load(spec_path).context("failed to load spec")?;
|
2021-07-04 22:44:07 +02:00
|
|
|
Ok(spec)
|
|
|
|
}
|
2021-07-15 01:09:04 +02:00
|
|
|
|
|
|
|
fn load_container_state(&self, container_dir: PathBuf) -> Result<Container> {
|
|
|
|
let container = Container::load(container_dir)?.refresh_status()?;
|
|
|
|
if !container.can_exec() {
|
|
|
|
bail!(
|
|
|
|
"Cannot exec as container is in state {}",
|
|
|
|
container.status()
|
|
|
|
);
|
|
|
|
}
|
|
|
|
|
|
|
|
Ok(container)
|
|
|
|
}
|
|
|
|
|
|
|
|
fn adapt_spec_for_tenant(&self, spec: &mut Spec, container: &Container) -> Result<()> {
|
|
|
|
if let Some(ref process) = self.process {
|
|
|
|
self.set_process(spec, process)?;
|
|
|
|
} else {
|
|
|
|
self.set_working_dir(spec)?;
|
|
|
|
self.set_args(spec)?;
|
|
|
|
self.set_environment(spec)?;
|
2021-07-15 22:12:21 +02:00
|
|
|
self.set_no_new_privileges(spec);
|
2021-07-15 01:09:04 +02:00
|
|
|
self.set_capabilities(spec)?;
|
|
|
|
}
|
|
|
|
|
|
|
|
if container.pid().is_none() {
|
|
|
|
bail!("Could not retrieve container init pid");
|
|
|
|
}
|
|
|
|
|
|
|
|
let init_process = procfs::process::Process::new(container.pid().unwrap().as_raw())?;
|
|
|
|
self.set_namespaces(spec, init_process.namespaces()?)?;
|
|
|
|
|
|
|
|
Ok(())
|
|
|
|
}
|
|
|
|
|
|
|
|
fn set_process(&self, spec: &mut Spec, process: &Path) -> Result<()> {
|
|
|
|
if !process.exists() {
|
|
|
|
bail!(
|
|
|
|
"Process.json file does not exist at specified path {}",
|
|
|
|
process.display()
|
|
|
|
)
|
|
|
|
}
|
|
|
|
|
2021-07-15 22:12:21 +02:00
|
|
|
let process = utils::open(process)?;
|
2021-07-15 01:09:04 +02:00
|
|
|
let process_spec: Process = serde_json::from_reader(process)?;
|
|
|
|
spec.process = process_spec;
|
2021-07-15 22:12:21 +02:00
|
|
|
Ok(())
|
2021-07-15 01:09:04 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
fn set_working_dir(&self, spec: &mut Spec) -> Result<()> {
|
|
|
|
if let Some(ref cwd) = self.cwd {
|
|
|
|
if cwd.is_relative() {
|
|
|
|
bail!(
|
|
|
|
"Current working directory must be an absolute path, but is {}",
|
|
|
|
cwd.display()
|
|
|
|
);
|
|
|
|
}
|
|
|
|
|
|
|
|
spec.process.cwd = cwd.to_string_lossy().to_string();
|
|
|
|
}
|
|
|
|
|
|
|
|
Ok(())
|
|
|
|
}
|
|
|
|
|
|
|
|
fn set_args(&self, spec: &mut Spec) -> Result<()> {
|
|
|
|
if self.args.is_empty() {
|
|
|
|
bail!("Container command was not specified")
|
|
|
|
}
|
|
|
|
|
|
|
|
spec.process.args = self.args.clone();
|
|
|
|
Ok(())
|
|
|
|
}
|
|
|
|
|
|
|
|
fn set_environment(&self, spec: &mut Spec) -> Result<()> {
|
|
|
|
spec.process.env.append(
|
|
|
|
&mut self
|
|
|
|
.env
|
|
|
|
.iter()
|
|
|
|
.map(|(k, v)| format!("{}={}", k, v))
|
|
|
|
.collect(),
|
|
|
|
);
|
|
|
|
|
|
|
|
Ok(())
|
|
|
|
}
|
|
|
|
|
2021-07-15 22:12:21 +02:00
|
|
|
fn set_no_new_privileges(&self, spec: &mut Spec) {
|
2021-07-15 01:09:04 +02:00
|
|
|
if let Some(no_new_privs) = self.no_new_privs {
|
|
|
|
spec.process.no_new_privileges = no_new_privs;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
fn set_capabilities(&self, spec: &mut Spec) -> Result<()> {
|
|
|
|
if !self.capabilities.is_empty() {
|
2021-07-20 20:59:34 +02:00
|
|
|
let mut caps: Vec<Capability> = Vec::with_capacity(self.capabilities.len());
|
2021-07-15 01:09:04 +02:00
|
|
|
for cap in &self.capabilities {
|
2021-07-20 20:59:34 +02:00
|
|
|
caps.push(Capability::from_str(cap)?);
|
2021-07-15 01:09:04 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
if let Some(ref mut spec_caps) = spec.process.capabilities {
|
|
|
|
spec_caps.ambient.append(&mut caps.clone());
|
|
|
|
spec_caps.bounding.append(&mut caps.clone());
|
|
|
|
spec_caps.effective.append(&mut caps.clone());
|
|
|
|
spec_caps.inheritable.append(&mut caps.clone());
|
2021-07-15 22:12:21 +02:00
|
|
|
spec_caps.permitted.append(&mut caps);
|
2021-07-15 01:09:04 +02:00
|
|
|
} else {
|
|
|
|
spec.process.capabilities = Some(LinuxCapabilities {
|
|
|
|
ambient: caps.clone(),
|
|
|
|
bounding: caps.clone(),
|
|
|
|
effective: caps.clone(),
|
|
|
|
inheritable: caps.clone(),
|
2021-07-15 22:12:21 +02:00
|
|
|
permitted: caps,
|
2021-07-15 01:09:04 +02:00
|
|
|
})
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
Ok(())
|
|
|
|
}
|
|
|
|
|
|
|
|
fn set_namespaces(&self, spec: &mut Spec, init_namespaces: Vec<Namespace>) -> Result<()> {
|
|
|
|
let mut tenant_namespaces = Vec::with_capacity(init_namespaces.len());
|
|
|
|
|
|
|
|
for ns_type in NAMESPACE_TYPES.iter().copied() {
|
|
|
|
if let Some(init_ns) = init_namespaces.iter().find(|n| n.ns_type.eq(ns_type)) {
|
|
|
|
let tenant_ns = LinuxNamespaceType::try_from(ns_type)?;
|
|
|
|
tenant_namespaces.push(LinuxNamespace {
|
|
|
|
typ: tenant_ns,
|
|
|
|
path: Some(init_ns.path.to_string_lossy().to_string()),
|
|
|
|
})
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
let mut linux = spec.linux.as_mut().unwrap();
|
|
|
|
linux.namespaces = tenant_namespaces;
|
|
|
|
Ok(())
|
|
|
|
}
|
|
|
|
|
|
|
|
fn should_use_systemd(&self, container: &Container) -> bool {
|
|
|
|
if let Some(use_systemd) = container.systemd() {
|
|
|
|
return use_systemd;
|
|
|
|
}
|
|
|
|
|
|
|
|
false
|
|
|
|
}
|
2021-07-15 22:12:21 +02:00
|
|
|
|
2021-07-19 08:22:47 +02:00
|
|
|
fn setup_notify_listener(container_dir: &Path) -> Result<PathBuf> {
|
2021-07-15 22:12:21 +02:00
|
|
|
let notify_name = Self::generate_name(&container_dir, TENANT_NOTIFY);
|
|
|
|
let socket_path = container_dir.join(¬ify_name);
|
|
|
|
|
2021-07-19 08:22:47 +02:00
|
|
|
Ok(socket_path)
|
2021-07-15 22:12:21 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
fn setup_tty_socket(&self, container_dir: &Path) -> Result<Option<FileDescriptor>> {
|
|
|
|
let tty_name = Self::generate_name(&container_dir, TENANT_TTY);
|
|
|
|
let csocketfd = if let Some(console_socket) = &self.base.console_socket {
|
|
|
|
Some(tty::setup_console_socket(
|
|
|
|
container_dir,
|
|
|
|
console_socket,
|
|
|
|
&tty_name,
|
|
|
|
)?)
|
|
|
|
} else {
|
|
|
|
None
|
|
|
|
};
|
|
|
|
|
|
|
|
Ok(csocketfd)
|
|
|
|
}
|
|
|
|
|
|
|
|
fn generate_name(dir: &Path, prefix: &str) -> String {
|
|
|
|
loop {
|
|
|
|
let rand = fastrand::i32(..);
|
|
|
|
let name = format!("{}{:x}.sock", prefix, rand);
|
|
|
|
if !dir.join(&name).exists() {
|
|
|
|
return name;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
2021-07-15 01:09:04 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
// Can be removed once https://github.com/eminence/procfs/pull/135 is available
|
|
|
|
trait GetNamespace {
|
|
|
|
fn namespaces(&self) -> Result<Vec<Namespace>>;
|
|
|
|
}
|
|
|
|
|
|
|
|
impl GetNamespace for procfs::process::Process {
|
|
|
|
/// Describes namespaces to which the process with the corresponding PID belongs.
|
|
|
|
/// Doc reference: https://man7.org/linux/man-pages/man7/namespaces.7.html
|
|
|
|
fn namespaces(&self) -> Result<Vec<Namespace>> {
|
|
|
|
let proc_path = PathBuf::from(format!("/proc/{}", self.pid()));
|
|
|
|
let ns = proc_path.join("ns");
|
|
|
|
let mut namespaces = Vec::new();
|
|
|
|
for entry in fs::read_dir(ns)? {
|
|
|
|
let entry = entry?;
|
|
|
|
let path = entry.path();
|
|
|
|
let ns_type = entry.file_name();
|
|
|
|
let cstr = CString::new(path.as_os_str().as_bytes()).unwrap();
|
|
|
|
|
|
|
|
let mut stat = unsafe { std::mem::zeroed() };
|
|
|
|
if unsafe { libc::stat(cstr.as_ptr(), &mut stat) } != 0 {
|
|
|
|
bail!("Unable to stat {:?}", path);
|
|
|
|
}
|
|
|
|
|
|
|
|
namespaces.push(Namespace {
|
|
|
|
ns_type,
|
|
|
|
path,
|
|
|
|
identifier: stat.st_ino,
|
|
|
|
device_id: stat.st_dev,
|
|
|
|
})
|
|
|
|
}
|
|
|
|
|
|
|
|
Ok(namespaces)
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
/// Information about a namespace
|
|
|
|
///
|
|
|
|
/// See also the [Process::namespaces()] method
|
|
|
|
#[derive(Debug, Clone)]
|
|
|
|
pub struct Namespace {
|
|
|
|
/// Namespace type
|
|
|
|
pub ns_type: OsString,
|
|
|
|
/// Handle to the namespace
|
|
|
|
pub path: PathBuf,
|
|
|
|
/// Namespace identifier (inode number)
|
|
|
|
pub identifier: u64,
|
|
|
|
/// Device id of the namespace
|
|
|
|
pub device_id: u64,
|
2021-07-04 22:44:07 +02:00
|
|
|
}
|
2021-07-15 01:09:04 +02:00
|
|
|
|
|
|
|
impl PartialEq for Namespace {
|
|
|
|
fn eq(&self, other: &Self) -> bool {
|
|
|
|
// see https://lore.kernel.org/lkml/87poky5ca9.fsf@xmission.com/
|
|
|
|
self.identifier == other.identifier && self.device_id == other.device_id
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
impl Eq for Namespace {}
|