commonware_runtime/tokio/

runtime.rs

#[cfg(feature = "external")]
use crate::Pacer;
#[cfg(not(feature = "iouring-network"))]
use crate::network::tokio::{Config as TokioNetworkConfig, Network as TokioNetwork};
#[cfg(feature = "iouring-storage")]
use crate::storage::iouring::{Config as IoUringConfig, Storage as IoUringStorage};
#[cfg(not(feature = "iouring-storage"))]
use crate::storage::tokio::{Config as TokioStorageConfig, Storage as TokioStorage};
use crate::{
    BlobLayout, BlobVersion, BufferPool, BufferPoolConfig, Clock, Error, Execution, Handle,
    METRICS_PREFIX, Name, SinkOf, StreamOf, child_label,
    network::metered::Network as MeteredNetwork,
    prefixed_name,
    process::metered::Metrics as MeteredProcess,
    signal::Signal,
    storage::metered::Storage as MeteredStorage,
    telemetry::metrics::{
        CounterFamily, GaugeFamily, Metric, Register, Registered, Registry, add_attribute, raw,
        task::Label, validate_label,
    },
    utils::{self, Panicker, signal::Stopper, supervision::Tree},
};
#[cfg(feature = "iouring-network")]
use crate::{
    iouring,
    network::iouring::{Config as IoUringNetworkConfig, Network as IoUringNetwork},
};
use commonware_macros::{select, stability};
#[stability(BETA)]
use commonware_parallel::Rayon;
use commonware_utils::{NZUsize, sync::Mutex, sys_rng};
use governor::clock::{Clock as GClock, ReasonablyRealtime};
use rand_core::{Rng, TryCryptoRng, TryRng};
#[stability(BETA)]
use rayon::ThreadPoolBuilder;
use std::{
    convert::Infallible,
    env,
    future::Future,
    net::{IpAddr, SocketAddr},
    num::NonZeroUsize,
    ops::RangeInclusive,
    panic::{AssertUnwindSafe, catch_unwind, resume_unwind},
    path::PathBuf,
    sync::Arc,
    time::{Duration, SystemTime},
};
use tokio::{
    runtime::{Builder, Handle as RuntimeHandle},
    sync::Notify,
};

#[cfg(feature = "iouring-network")]
cfg_if::cfg_if! {
    if #[cfg(test)] {
        // Use a smaller ring in tests to reduce `io_uring_setup` failures
        // under parallel test load due to mlock/resource limits.
        const IOURING_NETWORK_SIZE: u32 = 128;
    } else {
        const IOURING_NETWORK_SIZE: u32 = 1024;
    }
}

#[derive(Debug)]
struct Metrics {
    tasks_spawned: CounterFamily<Label>,
    tasks_running: GaugeFamily<Label>,
}

impl Metrics {
    pub fn init(registry: &mut impl Register) -> Self {
        Self {
            tasks_spawned: registry.register(
                "tasks_spawned",
                "Total number of tasks spawned",
                raw::Family::default(),
            ),
            tasks_running: registry.register(
                "tasks_running",
                "Number of tasks currently running",
                raw::Family::default(),
            ),
        }
    }
}

#[derive(Clone, Debug)]
pub struct NetworkConfig {
    /// If Some, explicitly sets TCP_NODELAY on the socket.
    /// Otherwise uses system default.
    ///
    /// Defaults to `Some(true)`.
    tcp_nodelay: Option<bool>,

    /// Whether to set `SO_LINGER` to zero on the socket.
    ///
    /// When enabled, causes an immediate RST on close, avoiding
    /// `TIME_WAIT` state. This is useful in adversarial environments to
    /// reclaim socket resources immediately when closing connections to
    /// misbehaving peers.
    ///
    /// Defaults to `true`.
    zero_linger: bool,

    /// Timeout for establishing an outbound TCP connection.
    ///
    /// Defaults to 10 seconds.
    connect_timeout: Duration,

    /// Read/write timeout for network operations.
    ///
    /// Bounds the full `Sink::send` and `Stream::recv` calls rather than each
    /// individual socket syscall. Larger batched writes may therefore require a
    /// larger timeout.
    ///
    /// Defaults to 60 seconds.
    read_write_timeout: Duration,
}

impl Default for NetworkConfig {
    fn default() -> Self {
        Self {
            tcp_nodelay: Some(true),
            zero_linger: true,
            connect_timeout: Duration::from_secs(10),
            read_write_timeout: Duration::from_secs(60),
        }
    }
}

/// Configuration for the `tokio` runtime.
#[derive(Clone)]
pub struct Config {
    /// Number of threads to use for handling async tasks.
    ///
    /// Worker threads are always active (waiting for work).
    ///
    /// Tokio sets the default value to the number of logical CPUs.
    worker_threads: usize,

    /// Number of scheduler ticks between global queue polls.
    ///
    /// When unset, Tokio uses its default behavior for the multi-thread
    /// scheduler. Smaller values reduce the delay before tasks woken from
    /// outside a worker, such as io_uring completion notifications, are polled
    /// from the global queue again.
    global_queue_interval: Option<u32>,

    /// Maximum number of threads to use for blocking tasks.
    ///
    /// Unlike worker threads, blocking threads are created as needed and
    /// exit if left idle for too long.
    ///
    /// Tokio sets the default value to 512 to avoid hanging on lower-level
    /// operations that require blocking (like `fs` and writing to `Stdout`).
    max_blocking_threads: usize,

    /// Stack size to use for runtime-owned threads.
    ///
    /// Defaults to the system stack size when the current platform exposes it,
    /// and otherwise falls back to Rust's default spawned-thread stack size.
    ///
    /// See [utils::thread::system_thread_stack_size].
    thread_stack_size: usize,

    /// Whether or not to catch panics.
    catch_panics: bool,

    /// Base directory for all storage operations, created at start if missing
    /// and held for the run.
    storage_directory: PathBuf,

    /// Blob layouts accepted by storage.
    ///
    /// Defaults to [BlobLayout::ALL] and must be non-empty. New blobs use the latest layout in
    /// the range. Existing blobs outside it fail to open with [Error::BlobLayoutMismatch], so
    /// restrict the range to what a rollback target can read before the upgraded binary first
    /// opens storage.
    storage_blob_layouts: RangeInclusive<BlobLayout>,

    /// Network configuration.
    network_cfg: NetworkConfig,

    /// Explicit buffer pool configuration for network I/O, if provided.
    network_buffer_pool_cfg: Option<BufferPoolConfig>,

    /// Explicit buffer pool configuration for storage I/O, if provided.
    storage_buffer_pool_cfg: Option<BufferPoolConfig>,
}

impl Config {
    /// Returns a new [Config] with default values.
    pub fn new() -> Self {
        let rng = sys_rng().next_u64();
        let storage_directory = env::temp_dir().join(format!("commonware_tokio_runtime_{rng}"));
        Self {
            worker_threads: 2,
            global_queue_interval: None,
            max_blocking_threads: 512,
            thread_stack_size: utils::thread::system_thread_stack_size(),
            catch_panics: false,
            storage_directory,
            storage_blob_layouts: BlobLayout::ALL,
            network_cfg: NetworkConfig::default(),
            network_buffer_pool_cfg: None,
            storage_buffer_pool_cfg: None,
        }
    }

    // Setters
    /// See [Config]
    pub const fn with_worker_threads(mut self, n: usize) -> Self {
        self.worker_threads = n;
        self
    }
    /// See [Config]
    pub const fn with_global_queue_interval(mut self, n: u32) -> Self {
        self.global_queue_interval = Some(n);
        self
    }
    /// See [Config]
    pub const fn with_max_blocking_threads(mut self, n: usize) -> Self {
        self.max_blocking_threads = n;
        self
    }
    /// See [Config]
    pub const fn with_thread_stack_size(mut self, n: usize) -> Self {
        self.thread_stack_size = n;
        self
    }
    /// See [Config]
    pub const fn with_catch_panics(mut self, b: bool) -> Self {
        self.catch_panics = b;
        self
    }
    /// See [Config]
    pub const fn with_connect_timeout(mut self, timeout: Duration) -> Self {
        self.network_cfg.connect_timeout = timeout;
        self
    }
    /// See [Config]
    pub const fn with_read_write_timeout(mut self, d: Duration) -> Self {
        self.network_cfg.read_write_timeout = d;
        self
    }
    /// See [Config]
    pub const fn with_tcp_nodelay(mut self, n: Option<bool>) -> Self {
        self.network_cfg.tcp_nodelay = n;
        self
    }
    /// See [Config]
    pub const fn with_zero_linger(mut self, l: bool) -> Self {
        self.network_cfg.zero_linger = l;
        self
    }
    /// See [Config]
    pub fn with_storage_directory(mut self, p: impl Into<PathBuf>) -> Self {
        self.storage_directory = p.into();
        self
    }
    /// See [Config]
    ///
    /// # Panics
    ///
    /// Panics if `layouts` is empty.
    pub fn with_storage_blob_layouts(mut self, layouts: RangeInclusive<BlobLayout>) -> Self {
        assert!(
            !layouts.is_empty(),
            "storage blob layouts must be non-empty"
        );
        self.storage_blob_layouts = layouts;
        self
    }
    /// See [Config]
    pub fn with_network_buffer_pool_config(mut self, cfg: BufferPoolConfig) -> Self {
        self.network_buffer_pool_cfg = Some(cfg);
        self
    }
    /// See [Config]
    pub fn with_storage_buffer_pool_config(mut self, cfg: BufferPoolConfig) -> Self {
        self.storage_buffer_pool_cfg = Some(cfg);
        self
    }

    // Getters
    /// See [Config]
    pub const fn worker_threads(&self) -> usize {
        self.worker_threads
    }
    /// See [Config]
    pub const fn global_queue_interval(&self) -> Option<u32> {
        self.global_queue_interval
    }
    /// See [Config]
    pub const fn max_blocking_threads(&self) -> usize {
        self.max_blocking_threads
    }
    /// See [Config]
    pub const fn thread_stack_size(&self) -> usize {
        self.thread_stack_size
    }
    /// See [Config]
    pub const fn catch_panics(&self) -> bool {
        self.catch_panics
    }
    /// See [Config]
    pub const fn connect_timeout(&self) -> Duration {
        self.network_cfg.connect_timeout
    }
    /// See [Config]
    pub const fn read_write_timeout(&self) -> Duration {
        self.network_cfg.read_write_timeout
    }
    /// See [Config]
    pub const fn tcp_nodelay(&self) -> Option<bool> {
        self.network_cfg.tcp_nodelay
    }
    /// See [Config]
    pub const fn zero_linger(&self) -> bool {
        self.network_cfg.zero_linger
    }
    /// See [Config]
    pub const fn storage_directory(&self) -> &PathBuf {
        &self.storage_directory
    }
    /// See [Config]
    pub const fn storage_blob_layouts(&self) -> &RangeInclusive<BlobLayout> {
        &self.storage_blob_layouts
    }

    /// Returns the network buffer pool config, deriving pool parallelism from
    /// `worker_threads` if not explicitly configured.
    fn resolved_network_buffer_pool_config(&self) -> BufferPoolConfig {
        self.network_buffer_pool_cfg.clone().unwrap_or_else(|| {
            BufferPoolConfig::for_network().with_parallelism(NZUsize!(self.worker_threads))
        })
    }

    /// Returns the storage buffer pool config, deriving pool parallelism from
    /// `worker_threads` if not explicitly configured.
    fn resolved_storage_buffer_pool_config(&self) -> BufferPoolConfig {
        self.storage_buffer_pool_cfg.clone().unwrap_or_else(|| {
            BufferPoolConfig::for_storage().with_parallelism(NZUsize!(self.worker_threads))
        })
    }
}

impl Default for Config {
    fn default() -> Self {
        Self::new()
    }
}

/// Runtime based on [Tokio](https://tokio.rs).
pub struct Executor {
    registry: Registry,
    metrics: Arc<Metrics>,
    runtime: RuntimeHandle,
    tasks: Arc<TaskTracker>,
    shutdown: Mutex<Stopper>,
    panicker: Panicker,
    thread_stack_size: usize,
}

/// Closes task admission and tracks wrappers through user-future drop and descendant cleanup.
#[derive(Default)]
struct TaskTracker {
    state: Mutex<TaskTrackerState>,
    idle: Notify,
}

#[derive(Default)]
struct TaskTrackerState {
    active: usize,
    closed: bool,
}

impl TaskTracker {
    fn admit(self: &Arc<Self>) -> Option<TaskGuard> {
        let mut state = self.state.lock();
        if state.closed {
            return None;
        }
        state.active = state.active.checked_add(1).expect("active task overflow");
        Some(TaskGuard(Arc::clone(self)))
    }

    fn close(&self) {
        self.state.lock().closed = true;
    }

    async fn wait(&self) {
        loop {
            // Subscribe before checking so the final task cannot notify between the check and wait.
            let idle = self.idle.notified();
            if self.state.lock().active == 0 {
                return;
            }
            idle.await;
        }
    }
}

struct TaskGuard(Arc<TaskTracker>);

impl Drop for TaskGuard {
    fn drop(&mut self) {
        let mut state = self.0.state.lock();
        state.active = state.active.checked_sub(1).expect("active task underflow");
        if state.active == 0 {
            drop(state);
            self.0.idle.notify_one();
        }
    }
}

/// Implementation of [crate::Runner] for the `tokio` runtime.
pub struct Runner {
    cfg: Config,
}

impl Default for Runner {
    fn default() -> Self {
        Self::new(Config::default())
    }
}

impl Runner {
    /// Initialize a new `tokio` runtime with the given number of threads.
    pub const fn new(cfg: Config) -> Self {
        Self { cfg }
    }
}

impl crate::Runner for Runner {
    type Context = Context;

    fn start<F, Fut>(self, f: F) -> Fut::Output
    where
        F: FnOnce(Self::Context) -> Fut,
        Fut: Future,
    {
        // Create a new registry
        let mut registry = Registry::new();
        let mut runtime_registry = registry.sub_registry(METRICS_PREFIX);

        // Initialize runtime
        let metrics = Arc::new(Metrics::init(&mut runtime_registry));
        let mut builder = Builder::new_multi_thread();
        builder
            .worker_threads(self.cfg.worker_threads)
            .max_blocking_threads(self.cfg.max_blocking_threads)
            .thread_stack_size(self.cfg.thread_stack_size)
            .enable_all();
        if let Some(global_queue_interval) = self.cfg.global_queue_interval {
            builder.global_queue_interval(global_queue_interval);
        }
        let runtime = builder.build().expect("failed to create Tokio runtime");

        // Initialize panicker
        let (panicker, panicked) = Panicker::new(self.cfg.catch_panics);

        // Collect process metrics.
        //
        // We prefer to collect process metrics outside of `Context` because
        // we are using `runtime_registry` rather than the one provided by `Context`.
        let process = MeteredProcess::init(&mut runtime_registry);
        runtime.spawn(process.collect(tokio::time::sleep));

        // Initialize buffer pools
        let network_buffer_pool = BufferPool::new(
            self.cfg.resolved_network_buffer_pool_config(),
            &mut runtime_registry.sub_registry("network_buffer_pool"),
        );
        let storage_buffer_pool = BufferPool::new(
            self.cfg.resolved_storage_buffer_pool_config(),
            &mut runtime_registry.sub_registry("storage_buffer_pool"),
        );

        // Initialize storage
        cfg_if::cfg_if! {
            if #[cfg(feature = "iouring-storage")] {
                let mut iouring_registry = runtime_registry.sub_registry("iouring_storage");
                let storage = MeteredStorage::new(
                    IoUringStorage::start(
                        IoUringConfig {
                            storage_directory: self.cfg.storage_directory.clone(),
                            blob_layouts: self.cfg.storage_blob_layouts.clone(),
                            iouring_config: Default::default(),
                            thread_stack_size: self.cfg.thread_stack_size,
                        },
                        &mut iouring_registry,
                        storage_buffer_pool.clone(),
                    ),
                    &mut runtime_registry,
                );
            } else {
                let storage = MeteredStorage::new(
                    TokioStorage::new(
                        TokioStorageConfig::new(
                            self.cfg.storage_directory.clone(),
                            self.cfg.storage_blob_layouts.clone(),
                        ),
                        storage_buffer_pool.clone(),
                    ),
                    &mut runtime_registry,
                );
            }
        }

        // Make any storage a prior process left in the page cache crash-durable before we open it,
        // so the data read during init is durable. This runs under the hold, after any straggling
        // writes from a previous run have landed, so the flush covers them too.
        if let Err(e) = crate::storage::sync(&self.cfg.storage_directory) {
            panic!(
                "failed to sync storage filesystem at startup ({}): {e}",
                self.cfg.storage_directory.display()
            );
        }

        // Initialize network
        cfg_if::cfg_if! {
            if #[cfg(feature = "iouring-network")] {
                let mut iouring_registry = runtime_registry.sub_registry("iouring_network");
                let config = IoUringNetworkConfig {
                    tcp_nodelay: self.cfg.network_cfg.tcp_nodelay,
                    zero_linger: self.cfg.network_cfg.zero_linger,
                    connect_timeout: self.cfg.network_cfg.connect_timeout,
                    read_write_timeout: self.cfg.network_cfg.read_write_timeout,
                    iouring_config: iouring::Config {
                        // TODO (#1045): make `IOURING_NETWORK_SIZE` configurable
                        size: IOURING_NETWORK_SIZE,
                        max_request_timeout: self.cfg.network_cfg.read_write_timeout,
                        shutdown_timeout: Some(self.cfg.network_cfg.read_write_timeout),
                        ..Default::default()
                    },
                    thread_stack_size: self.cfg.thread_stack_size,
                    ..Default::default()
                };
                let network = MeteredNetwork::new(
                    IoUringNetwork::start(
                        config,
                        &mut iouring_registry,
                        network_buffer_pool.clone(),
                    )
                    .unwrap(),
                    &mut runtime_registry,
                );
            } else {
                let config = TokioNetworkConfig::default()
                    .with_connect_timeout(self.cfg.network_cfg.connect_timeout)
                    .with_read_timeout(self.cfg.network_cfg.read_write_timeout)
                    .with_write_timeout(self.cfg.network_cfg.read_write_timeout)
                    .with_tcp_nodelay(self.cfg.network_cfg.tcp_nodelay)
                    .with_zero_linger(self.cfg.network_cfg.zero_linger);
                let network = MeteredNetwork::new(
                    TokioNetwork::new(config, network_buffer_pool.clone()),
                    &mut runtime_registry,
                );
            }
        }

        // Initialize executor
        let executor = Arc::new(Executor {
            registry,
            metrics,
            runtime: runtime.handle().clone(),
            tasks: Arc::new(TaskTracker::default()),
            shutdown: Mutex::new(Stopper::default()),
            panicker,
            thread_stack_size: self.cfg.thread_stack_size,
        });

        // Get metrics
        let label = Label::root();
        executor.metrics.tasks_spawned.get_or_create(&label).inc();
        let gauge = executor.metrics.tasks_running.get_or_create(&label).clone();

        // Run the future
        let tree = Tree::root();
        let context = Context {
            storage,
            name: label.name(),
            attributes: Vec::new(),
            executor: executor.clone(),
            network,
            network_buffer_pool,
            storage_buffer_pool,
            tree: Arc::clone(&tree),
            execution: Execution::default(),
        };
        let output = catch_unwind(AssertUnwindSafe(|| {
            runtime.block_on(panicked.interrupt(f(context)))
        }));
        executor.tasks.close();
        tree.abort();
        runtime.block_on(executor.tasks.wait());
        gauge.dec();

        match output {
            Ok(output) => output,
            Err(panic) => resume_unwind(panic),
        }
    }
}

cfg_if::cfg_if! {
    if #[cfg(feature = "iouring-storage")] {
        type Storage = MeteredStorage<IoUringStorage>;
    } else {
        type Storage = MeteredStorage<TokioStorage>;
    }
}

cfg_if::cfg_if! {
    if #[cfg(feature = "iouring-network")] {
        type Network = MeteredNetwork<IoUringNetwork>;
    } else {
        type Network = MeteredNetwork<TokioNetwork>;
    }
}

/// Implementation of [crate::Spawner], [crate::Clock],
/// [crate::Network], and [crate::Storage] for the `tokio`
/// runtime.
pub struct Context {
    name: String,
    attributes: Vec<(String, String)>,
    executor: Arc<Executor>,
    storage: Storage,
    network: Network,
    network_buffer_pool: BufferPool,
    storage_buffer_pool: BufferPool,
    tree: Arc<Tree>,
    execution: Execution,
}

impl Context {
    /// Access the [Metrics] of the runtime.
    fn metrics(&self) -> &Metrics {
        &self.executor.metrics
    }
}

impl crate::Spawner for Context {
    fn dedicated(mut self) -> Self {
        self.execution = Execution::Dedicated;
        self
    }

    fn shared(mut self, blocking: bool) -> Self {
        self.execution = Execution::Shared(blocking);
        self
    }

    fn spawn<F, Fut, T>(mut self, f: F) -> Handle<T>
    where
        F: FnOnce(Self) -> Fut + Send + 'static,
        Fut: Future<Output = T> + Send + 'static,
        T: Send + 'static,
    {
        // Get metrics
        let (_, metric) = spawn_metrics!(self);

        // Track supervision before resetting configuration
        let parent = Arc::clone(&self.tree);
        let past = self.execution;
        self.execution = Execution::default();
        let (child, aborted) = Tree::child(&parent);
        if aborted {
            return Handle::closed(metric);
        }
        self.tree = child;

        // Spawn the task
        let executor = self.executor.clone();
        let Some(task_guard) = executor.tasks.admit() else {
            return Handle::closed(metric);
        };
        let future = f(self);
        let (f, handle) = Handle::init(
            future,
            metric,
            executor.panicker.clone(),
            Arc::clone(&parent),
        );
        let f = async move {
            let _task_guard = task_guard;
            f.await;
        };

        if matches!(past, Execution::Dedicated) {
            utils::thread::spawn(executor.thread_stack_size, {
                // Ensure the task can access the tokio runtime
                let handle = executor.runtime.clone();
                move || {
                    handle.block_on(f);
                }
            });
        } else if matches!(past, Execution::Shared(true)) {
            executor.runtime.spawn_blocking({
                // Ensure the task can access the tokio runtime
                let handle = executor.runtime.clone();
                move || {
                    handle.block_on(f);
                }
            });
        } else {
            executor.runtime.spawn(f);
        }

        // Register the task on the parent
        if let Some(aborter) = handle.aborter() {
            parent.register(aborter);
        }

        handle
    }

    async fn stop(self, value: i32, timeout: Option<Duration>) -> Result<(), Error> {
        let stop_resolved = {
            let mut shutdown = self.executor.shutdown.lock();
            shutdown.stop(value)
        };

        // Wait for all tasks to complete or the timeout to fire
        let timeout_future = timeout.map_or_else(
            || futures::future::Either::Right(futures::future::pending()),
            |duration| futures::future::Either::Left(self.sleep(duration)),
        );
        select! {
            result = stop_resolved => {
                result.map_err(|_| Error::Closed)?;
                Ok(())
            },
            _ = timeout_future => Err(Error::Timeout),
        }
    }

    fn stopped(&self) -> Signal {
        self.executor.shutdown.lock().stopped()
    }
}

#[stability(BETA)]
impl crate::Strategizer for Context {
    fn strategy(&self, parallelism: NonZeroUsize) -> Rayon {
        let pool = ThreadPoolBuilder::new()
            .num_threads(parallelism.get())
            .stack_size(self.executor.thread_stack_size)
            .build()
            .expect("failed to create Tokio Rayon thread pool");
        Rayon::with_pool(Arc::new(pool))
    }
}

impl crate::Supervisor for Context {
    fn child(&self, label: &'static str) -> Self {
        let (tree, _) = Tree::child(&self.tree);
        Self {
            name: child_label(&self.name, label),
            attributes: self.attributes.clone(),
            executor: self.executor.clone(),
            storage: self.storage.clone(),
            network: self.network.clone(),
            network_buffer_pool: self.network_buffer_pool.clone(),
            storage_buffer_pool: self.storage_buffer_pool.clone(),
            tree,
            execution: Execution::default(),
        }
    }

    fn with_attribute(mut self, key: &'static str, value: impl std::fmt::Display) -> Self {
        // Validate label format (must match [a-zA-Z][a-zA-Z0-9_]*)
        validate_label(key);

        // Add the attribute to the list of attributes
        add_attribute(&mut self.attributes, key, value);
        self
    }

    fn name(&self) -> Name {
        Name {
            label: self.name.clone(),
            attributes: self.attributes.clone(),
        }
    }
}

impl crate::Metrics for Context {
    fn register<N: Into<String>, H: Into<String>, M: Metric>(
        &self,
        name: N,
        help: H,
        metric: M,
    ) -> Registered<M> {
        let name = name.into();
        let help = help.into();
        let metric = Arc::new(metric);
        self.executor.registry.register(
            prefixed_name(&self.name, &name),
            help,
            self.attributes.clone(),
            metric,
        )
    }

    fn encode(&self) -> String {
        self.executor.registry.encode()
    }
}

impl Clock for Context {
    fn current(&self) -> SystemTime {
        SystemTime::now()
    }

    fn sleep(&self, duration: Duration) -> impl Future<Output = ()> + Send + 'static {
        tokio::time::sleep(duration)
    }

    fn sleep_until(&self, deadline: SystemTime) -> impl Future<Output = ()> + Send + 'static {
        let duration_until_deadline = deadline.duration_since(self.current()).unwrap_or_default();
        tokio::time::sleep(duration_until_deadline)
    }
}

#[cfg(feature = "external")]
impl Pacer for Context {
    fn pace<'a, F, T>(
        &'a self,
        _latency: Duration,
        future: F,
    ) -> impl Future<Output = T> + Send + 'a
    where
        F: Future<Output = T> + Send + 'a,
        T: Send + 'a,
    {
        // Execute the future immediately
        future
    }
}

impl GClock for Context {
    type Instant = SystemTime;

    fn now(&self) -> Self::Instant {
        self.current()
    }
}

impl ReasonablyRealtime for Context {}

impl crate::Network for Context {
    type Listener = <Network as crate::Network>::Listener;

    async fn bind(&self, socket: SocketAddr) -> Result<Self::Listener, Error> {
        self.network.bind(socket).await
    }

    async fn dial(&self, socket: SocketAddr) -> Result<(SinkOf<Self>, StreamOf<Self>), Error> {
        self.network.dial(socket).await
    }
}

impl crate::Resolver for Context {
    async fn resolve(&self, host: &str) -> Result<Vec<IpAddr>, Error> {
        // Uses the host's DNS configuration (e.g. /etc/resolv.conf). This delegates to the
        // system's libc resolver.
        //
        // The `:0` port is required by lookup_host's API but is not used
        // for DNS resolution.
        let addrs = tokio::net::lookup_host(format!("{host}:0"))
            .await
            .map_err(|e| Error::ResolveFailed(e.to_string()))?;
        Ok(addrs.map(|addr| addr.ip()).collect())
    }
}

impl TryRng for Context {
    type Error = Infallible;

    fn try_next_u32(&mut self) -> Result<u32, Self::Error> {
        Ok(sys_rng().next_u32())
    }

    fn try_next_u64(&mut self) -> Result<u64, Self::Error> {
        Ok(sys_rng().next_u64())
    }

    fn try_fill_bytes(&mut self, dest: &mut [u8]) -> Result<(), Self::Error> {
        sys_rng().fill_bytes(dest);
        Ok(())
    }
}

impl TryCryptoRng for Context {}

impl crate::Storage for Context {
    type Blob = <Storage as crate::Storage>::Blob;

    async fn open_versioned(
        &self,
        partition: &str,
        name: &[u8],
        versions: std::ops::RangeInclusive<BlobVersion>,
    ) -> Result<(Self::Blob, u64, BlobVersion), Error> {
        self.storage.open_versioned(partition, name, versions).await
    }

    async fn remove(&self, partition: &str, name: Option<&[u8]>) -> Result<(), Error> {
        self.storage.remove(partition, name).await
    }

    async fn scan(&self, partition: &str) -> Result<Vec<Vec<u8>>, Error> {
        self.storage.scan(partition).await
    }
}

impl crate::BufferPooler for Context {
    fn network_buffer_pool(&self) -> &BufferPool {
        &self.network_buffer_pool
    }

    fn storage_buffer_pool(&self) -> &BufferPool {
        &self.storage_buffer_pool
    }
}

#[cfg(test)]
#[allow(deprecated)]
mod tests {
    use super::*;
    use crate::{
        Blob as _, Metrics, Network, Resolver, Runner as _, Sink, Spawner as _, Storage as _,
        Strategizer as _, Stream, Supervisor as _, telemetry::metrics::raw::Counter,
        tokio::telemetry,
    };
    use bytes::Bytes;
    use commonware_parallel::Strategy as _;
    use std::{
        self,
        collections::HashMap,
        net::{IpAddr, Ipv4Addr, Ipv6Addr},
        str::FromStr,
    };
    use tracing::{Level, error};

    struct TaskDropGate {
        entered: std::sync::mpsc::Sender<()>,
        release: std::sync::mpsc::Receiver<()>,
    }

    impl Drop for TaskDropGate {
        fn drop(&mut self) {
            let _ = self.entered.send(());
            let _ = self.release.recv();
        }
    }

    #[derive(Clone, Copy, Debug)]
    enum RootExit {
        Return,
        FuturePanic,
        ConstructorPanic,
    }

    fn spawn_drop_gated_task(
        context: Context,
        execution: Execution,
        drop_gate: TaskDropGate,
        ready: Option<commonware_utils::channel::oneshot::Sender<()>>,
    ) {
        let child = match execution {
            Execution::Dedicated => context.dedicated(),
            Execution::Shared(blocking) => context.shared(blocking),
        };
        child.spawn(move |context| async move {
            let _context = context;
            let _drop_gate = drop_gate;
            if let Some(ready) = ready {
                ready.send(()).unwrap();
            }
            futures::future::pending::<()>().await;
        });
    }

    fn assert_runner_drains_spawned_task(execution: Execution, root_exit: RootExit) {
        let cfg = Config::new();
        let storage_directory = cfg.storage_directory().clone();
        let (ready_tx, ready_rx) = commonware_utils::channel::oneshot::channel();
        let (drop_entered_tx, drop_entered_rx) = std::sync::mpsc::channel();
        let (drop_release_tx, drop_release_rx) = std::sync::mpsc::channel();
        let (runner_done_tx, runner_done_rx) = std::sync::mpsc::channel();
        let runner = std::thread::spawn(move || {
            let result = std::panic::catch_unwind(AssertUnwindSafe(|| {
                let drop_gate = TaskDropGate {
                    entered: drop_entered_tx,
                    release: drop_release_rx,
                };
                match root_exit {
                    RootExit::ConstructorPanic => {
                        Runner::new(cfg).start(move |context| -> futures::future::Pending<()> {
                            spawn_drop_gated_task(context, execution, drop_gate, None);
                            panic!("root constructor panic after spawning child");
                        })
                    }
                    RootExit::Return | RootExit::FuturePanic => {
                        Runner::new(cfg).start(move |context| async move {
                            spawn_drop_gated_task(context, execution, drop_gate, Some(ready_tx));
                            ready_rx.await.unwrap();
                            assert!(
                                matches!(root_exit, RootExit::Return),
                                "root future panic after spawning child"
                            );
                        })
                    }
                }
            }));
            runner_done_tx.send(result.is_err()).unwrap();
        });

        drop_entered_rx
            .recv_timeout(Duration::from_secs(5))
            .expect("spawned task was not canceled after the root returned");
        let early = runner_done_rx.recv_timeout(Duration::from_millis(250));
        let returned_before_cleanup = match early {
            Ok(panicked) => Some(panicked),
            Err(std::sync::mpsc::RecvTimeoutError::Timeout) => None,
            Err(std::sync::mpsc::RecvTimeoutError::Disconnected) => {
                panic!("Runner::start exited without reporting its result")
            }
        };
        drop_release_tx.send(()).unwrap();
        let panicked = returned_before_cleanup.unwrap_or_else(|| {
            runner_done_rx
                .recv_timeout(Duration::from_secs(5))
                .expect("Runner::start did not return after task cleanup completed")
        });
        runner.join().unwrap();
        let _ = std::fs::remove_dir_all(storage_directory);
        assert!(
            returned_before_cleanup.is_none(),
            "Runner::start returned before {execution:?} task cleanup completed"
        );
        assert_eq!(panicked, !matches!(root_exit, RootExit::Return));
    }

    fn run_with_returned_strategy(retain: bool) -> Option<Rayon> {
        let cfg = Config::new();
        let storage_directory = cfg.storage_directory().clone();
        let (strategy_tx, strategy_rx) = std::sync::mpsc::channel();
        let runner = std::thread::spawn(move || {
            let strategy = Runner::new(cfg).start(move |context| async move {
                let strategy = context.strategy(NZUsize!(2));
                strategy.spawn(1, |_| ()).await;
                retain.then_some(strategy)
            });
            strategy_tx.send(strategy).unwrap();
        });

        let strategy = strategy_rx
            .recv_timeout(Duration::from_secs(5))
            .expect("Runner::start did not return after the strategy completed work");
        runner.join().unwrap();
        let _ = std::fs::remove_dir_all(storage_directory);
        strategy
    }

    #[test]
    fn test_storage_blob_layout_restriction() {
        // The default policy accepts legacy V0 blobs while selecting V1 for new blobs.
        let cfg = Config::new();
        let storage_directory = cfg.storage_directory().clone();
        assert_eq!(cfg.storage_blob_layouts(), &BlobLayout::ALL);

        let partition = "layout_restriction";
        let v0_name = b"v0";
        let partition_directory = storage_directory.join(partition);
        std::fs::create_dir_all(&partition_directory).unwrap();
        let v0_path = partition_directory.join(commonware_formatting::hex(v0_name));
        let v0_bytes = crate::storage::tests::v0_blob_bytes(0, b"payload");
        std::fs::write(&v0_path, &v0_bytes).unwrap();

        Runner::new(cfg).start(|context| async move {
            let (blob, size) = context.open(partition, v0_name).await.unwrap();
            assert_eq!(size, 7);
            let payload = blob
                .read_at(0, 7, crate::ReadOptions::default())
                .await
                .unwrap();
            assert_eq!(payload.coalesce(), b"payload".as_slice());
        });

        // A V1-only policy rejects V0 without modifying it and creates new blobs as V1.
        let cfg = Config::new()
            .with_storage_directory(storage_directory.clone())
            .with_storage_blob_layouts(BlobLayout::V1..=BlobLayout::V1);
        Runner::new(cfg).start(|context| async move {
            let result = context.open(partition, v0_name).await;
            assert!(matches!(
                result,
                Err(Error::BlobLayoutMismatch { expected, found })
                    if expected == (BlobLayout::V1..=BlobLayout::V1)
                        && found == BlobLayout::V0
            ));

            context.open(partition, b"v1").await.unwrap();
        });

        assert_eq!(std::fs::read(&v0_path).unwrap(), v0_bytes);
        let v1_path = partition_directory.join(commonware_formatting::hex(b"v1"));
        let v1 = std::fs::read(&v1_path).unwrap();
        assert_eq!(&v1[..4], &BlobLayout::V1.magic());

        // A V0-only policy rejects the intact header-only V1 blob without healing it and
        // creates new blobs the rollback target can read and write.
        let cfg = Config::new()
            .with_storage_directory(storage_directory.clone())
            .with_storage_blob_layouts(BlobLayout::V0..=BlobLayout::V0);
        Runner::new(cfg).start(|context| async move {
            let result = context.open(partition, b"v1").await;
            assert!(matches!(
                result,
                Err(Error::BlobLayoutMismatch { expected, found })
                    if expected == (BlobLayout::V0..=BlobLayout::V0)
                        && found == BlobLayout::V1
            ));

            let (blob, size) = context.open(partition, b"rollback").await.unwrap();
            assert_eq!(size, 0);
            blob.write_at(0, b"rollback!".as_slice(), crate::WriteOptions::SYNC)
                .await
                .unwrap();
            drop(blob);
            let (blob, size) = context.open(partition, b"rollback").await.unwrap();
            assert_eq!(size, 9);
            let payload = blob
                .read_at(0, 9, crate::ReadOptions::default())
                .await
                .unwrap();
            assert_eq!(payload.coalesce(), b"rollback!".as_slice());
        });

        assert_eq!(std::fs::read(&v1_path).unwrap(), v1);
        let rollback_path = partition_directory.join(commonware_formatting::hex(b"rollback"));
        let rollback = std::fs::read(rollback_path).unwrap();
        assert_eq!(&rollback[..4], &BlobLayout::V0.magic());
        assert_eq!(&rollback[8..], b"rollback!");
        let _ = std::fs::remove_dir_all(storage_directory);
    }

    #[test]
    #[should_panic(expected = "non-empty")]
    fn test_storage_blob_layout_restriction_rejects_empty_range() {
        let _ = Config::new().with_storage_blob_layouts(BlobLayout::V1..=BlobLayout::V0);
    }

    #[test]
    fn test_worker_threads_updates_default_buffer_pool_parallelism() {
        let cfg = Config::new().with_worker_threads(8);

        assert_eq!(cfg.worker_threads, 8);
        let network = cfg.resolved_network_buffer_pool_config();
        assert_eq!(network.parallelism(), NZUsize!(8));
        assert_eq!(
            network.thread_cache_config,
            BufferPoolConfig::for_network().thread_cache_config
        );

        let storage = cfg.resolved_storage_buffer_pool_config();
        assert_eq!(storage.parallelism(), NZUsize!(8));
        assert_eq!(
            storage.thread_cache_config,
            BufferPoolConfig::for_storage().thread_cache_config
        );
    }

    #[test]
    fn test_default_thread_stack_size_uses_system_default() {
        let cfg = Config::new();
        assert_eq!(
            cfg.thread_stack_size(),
            utils::thread::system_thread_stack_size()
        );
    }

    #[test]
    fn test_runner_waits_for_spawned_task_cancellation() {
        for execution in [
            Execution::Shared(false),
            Execution::Shared(true),
            Execution::Dedicated,
        ] {
            for root_exit in [
                RootExit::Return,
                RootExit::FuturePanic,
                RootExit::ConstructorPanic,
            ] {
                assert_runner_drains_spawned_task(execution, root_exit);
            }
        }
    }

    #[test]
    fn test_runner_start_waits_for_previous_run() {
        let cfg = Config::new();
        let storage_directory = cfg.storage_directory().clone();

        // The first run keeps its context, and with it the storage directory,
        // until it is released.
        let (started, first_started) = std::sync::mpsc::channel();
        let (release, released) = futures::channel::oneshot::channel();
        let first_cfg = cfg.clone();
        let first = std::thread::spawn(move || {
            Runner::new(first_cfg).start(|context| async move {
                started.send(()).unwrap();
                released.await.unwrap();
                drop(context);
            });
        });
        first_started.recv_timeout(Duration::from_secs(10)).unwrap();

        // A second run on the same directory cannot start until the first has
        // returned.
        let (started, second_started) = std::sync::mpsc::channel();
        let second = std::thread::spawn(move || {
            Runner::new(cfg).start(|_| async move {
                started.send(()).unwrap();
            });
        });
        match second_started.recv_timeout(Duration::from_millis(200)) {
            Err(std::sync::mpsc::RecvTimeoutError::Timeout) => {}
            other => panic!("second run started while the first held the directory: {other:?}"),
        }
        release.send(()).unwrap();
        first.join().unwrap();
        second_started
            .recv_timeout(Duration::from_secs(10))
            .expect("second run did not start after the first returned");
        second.join().unwrap();
        let _ = std::fs::remove_dir_all(storage_directory);
    }

    #[test]
    fn test_runner_owns_runtime_when_context_escapes() {
        let cfg = Config::new();
        let storage_directory = cfg.storage_directory().clone();
        let (ready_tx, ready_rx) = commonware_utils::channel::oneshot::channel();
        let (drop_entered_tx, drop_entered_rx) = std::sync::mpsc::channel();
        let (drop_release_tx, drop_release_rx) = std::sync::mpsc::channel();
        let (runner_returned_tx, runner_returned_rx) = std::sync::mpsc::channel();
        let (context_release_tx, context_release_rx) = std::sync::mpsc::channel();
        let runner = std::thread::spawn(move || {
            let context = Runner::new(cfg).start(move |context| async move {
                context.executor.runtime.spawn(async move {
                    let _drop_gate = TaskDropGate {
                        entered: drop_entered_tx,
                        release: drop_release_rx,
                    };
                    ready_tx.send(()).unwrap();
                    futures::future::pending::<()>().await;
                });
                ready_rx.await.unwrap();
                context
            });
            runner_returned_tx.send(()).unwrap();
            context_release_rx.recv().unwrap();
            drop(context);
        });

        let returned_early = runner_returned_rx
            .recv_timeout(Duration::from_millis(500))
            .is_ok();
        if returned_early {
            drop_release_tx.send(()).unwrap();
            context_release_tx.send(()).unwrap();
            drop_entered_rx
                .recv_timeout(Duration::from_secs(5))
                .expect("escaped Context did not retain the raw runtime task");
        } else {
            drop_entered_rx
                .recv_timeout(Duration::from_secs(5))
                .expect("Runner did not cancel its raw runtime task");
            drop_release_tx.send(()).unwrap();
            runner_returned_rx
                .recv_timeout(Duration::from_secs(5))
                .expect("Runner did not return after raw task cleanup");
            context_release_tx.send(()).unwrap();
        }
        runner.join().unwrap();
        let _ = std::fs::remove_dir_all(storage_directory);
        assert!(
            !returned_early,
            "a returned Context kept the Tokio runtime alive after Runner::start"
        );
    }

    #[test]
    fn test_runner_returns_strategy_after_pool_work() {
        assert!(run_with_returned_strategy(false).is_none());

        let strategy = run_with_returned_strategy(true).unwrap();
        assert_eq!(futures::executor::block_on(strategy.spawn(1, |_| 42)), 42);
    }

    #[test]
    fn test_runner_resumes_strategy_panic_payload_after_pool_work() {
        let cfg = Config::new();
        let storage_directory = cfg.storage_directory().clone();
        let (strategy_tx, strategy_rx) = std::sync::mpsc::channel();
        let runner = std::thread::spawn(move || {
            let result: std::thread::Result<()> =
                std::panic::catch_unwind(AssertUnwindSafe(|| {
                    Runner::new(cfg).start(move |context| async move {
                        let strategy = context.strategy(NZUsize!(2));
                        strategy.spawn(1, |_| ()).await;
                        std::panic::panic_any(strategy);
                    });
                }));
            let strategy = result
                .expect_err("Runner::start did not resume the root panic")
                .downcast::<Rayon>()
                .expect("Runner::start changed the root panic payload");
            strategy_tx.send(*strategy).unwrap();
        });

        let strategy = strategy_rx
            .recv_timeout(Duration::from_secs(5))
            .expect("Runner::start did not resume the strategy panic payload");
        runner.join().unwrap();
        let _ = std::fs::remove_dir_all(storage_directory);
        assert_eq!(futures::executor::block_on(strategy.spawn(1, |_| 42)), 42);
    }

    #[test]
    fn test_thread_stack_size_override() {
        let cfg = Config::new().with_thread_stack_size(4 * 1024 * 1024);
        assert_eq!(cfg.thread_stack_size(), 4 * 1024 * 1024);
    }

    #[test]
    fn test_explicit_buffer_pool_configs_override_worker_threads() {
        // Order does not matter -- explicit configs always win.
        let cfg = Config::new()
            .with_network_buffer_pool_config(
                BufferPoolConfig::for_network().with_parallelism(NZUsize!(2)),
            )
            .with_worker_threads(8)
            .with_storage_buffer_pool_config(
                BufferPoolConfig::for_storage().with_thread_cache_disabled(),
            );

        let network = cfg.resolved_network_buffer_pool_config();
        assert_eq!(network.parallelism(), NZUsize!(2));
        assert_eq!(
            network.thread_cache_config,
            BufferPoolConfig::for_network().thread_cache_config
        );

        let storage = cfg.resolved_storage_buffer_pool_config();
        assert_eq!(storage.parallelism(), NZUsize!(1));
        assert_eq!(
            storage.thread_cache_config,
            BufferPoolConfig::for_storage()
                .with_thread_cache_disabled()
                .thread_cache_config
        );
    }

    #[test]
    fn test_process_rss_metric() {
        let executor = Runner::default();
        executor.start(|context| async move {
            loop {
                // Wait for RSS metric to be available
                let metrics = context.encode();
                if !metrics.contains("runtime_process_rss") {
                    context.sleep(Duration::from_millis(100)).await;
                    continue;
                }

                // Verify the RSS value is eventually populated (greater than 0)
                for line in metrics.lines() {
                    if line.starts_with("runtime_process_rss")
                        && !line.starts_with("runtime_process_rss{")
                    {
                        let parts: Vec<&str> = line.split_whitespace().collect();
                        if parts.len() >= 2 {
                            let rss_value: i64 =
                                parts[1].parse().expect("Failed to parse RSS value");
                            if rss_value > 0 {
                                return;
                            }
                        }
                    }
                }
            }
        });
    }

    #[test]
    fn test_telemetry() {
        let executor = Runner::default();
        executor.start(|context| async move {
            // Define the server address
            let address = SocketAddr::from_str("127.0.0.1:8000").unwrap();

            // Configure telemetry
            telemetry::init(
                context.child("metrics"),
                telemetry::Logs {
                    level: Level::INFO,
                    json: false,
                },
                Some(address),
                None,
            );

            // Register a test metric
            let counter: Counter<u64> = Counter::default();
            let _registered = context.register("test_counter", "Test counter", counter.clone());
            counter.inc();

            // Helper functions to parse HTTP response
            async fn read_line<St: Stream>(stream: &mut St) -> Result<String, Error> {
                let mut line = Vec::new();
                loop {
                    let received = stream.recv(1).await?;
                    let byte = received.coalesce().as_ref()[0];
                    if byte == b'\n' {
                        if line.last() == Some(&b'\r') {
                            line.pop(); // Remove trailing \r
                        }
                        break;
                    }
                    line.push(byte);
                }
                String::from_utf8(line).map_err(|_| Error::ReadFailed)
            }

            async fn read_headers<St: Stream>(
                stream: &mut St,
            ) -> Result<HashMap<String, String>, Error> {
                let mut headers = HashMap::new();
                loop {
                    let line = read_line(stream).await?;
                    if line.is_empty() {
                        break;
                    }
                    let parts: Vec<&str> = line.splitn(2, ": ").collect();
                    if parts.len() == 2 {
                        headers.insert(parts[0].to_string(), parts[1].to_string());
                    }
                }
                Ok(headers)
            }

            async fn read_body<St: Stream>(
                stream: &mut St,
                content_length: usize,
            ) -> Result<String, Error> {
                let received = stream.recv(content_length).await?;
                String::from_utf8(received.coalesce().into()).map_err(|_| Error::ReadFailed)
            }

            // Simulate a client connecting to the server
            let client_handle = context.child("client").spawn(move |context| async move {
                let (mut sink, mut stream) = loop {
                    match context.dial(address).await {
                        Ok((sink, stream)) => break (sink, stream),
                        Err(e) => {
                            // The client may be polled before the server is ready, that's alright!
                            error!(err =?e, "failed to connect");
                            context.sleep(Duration::from_millis(10)).await;
                        }
                    }
                };

                // Send a GET request to the server
                let request = format!(
                    "GET /metrics HTTP/1.1\r\nHost: {address}\r\nConnection: close\r\n\r\n"
                );
                sink.send(Bytes::from(request)).await.unwrap();

                // Read and verify the HTTP status line
                let status_line = read_line(&mut stream).await.unwrap();
                assert_eq!(status_line, "HTTP/1.1 200 OK");

                // Read and parse headers
                let headers = read_headers(&mut stream).await.unwrap();
                println!("Headers: {headers:?}");
                let content_length = headers
                    .get("content-length")
                    .unwrap()
                    .parse::<usize>()
                    .unwrap();

                // Read and verify the body
                let body = read_body(&mut stream, content_length).await.unwrap();
                assert!(body.contains("test_counter_total 1"));
            });

            // Wait for the client task to complete
            client_handle.await.unwrap();
        });
    }

    #[test]
    fn test_resolver() {
        let executor = Runner::default();
        executor.start(|context| async move {
            let addrs = context.resolve("localhost").await.unwrap();
            assert!(!addrs.is_empty());
            for addr in addrs {
                assert!(
                    addr == IpAddr::V4(Ipv4Addr::LOCALHOST)
                        || addr == IpAddr::V6(Ipv6Addr::LOCALHOST)
                );
            }
        });
    }
}