Agent Skills: Tokio Knowledge Patch

Tokio Knowledge Patch

Use this patch when upgrading Tokio, selecting a safe patch release, or writing version-sensitive runtime, synchronization, I/O, networking, process, signal, tokio-util, or tokio-stream code. Check the detailed topic reference before depending on unstable APIs or patch-level fixes.

Reference index

| Reference | Topics | | --- | --- | | compatibility.md | Rust and dependency floors, safe patch releases, build configuration, target support | | runtime-and-tasks.md | Runtime construction, local execution, task APIs, scheduling, hooks, metrics, tracing | | synchronization.md | watch, broadcast, mpsc, oneshot, Notify, semaphores, permits, and single-assignment state | | io-and-filesystem.md | Async I/O types, files, pipes, AIO, io_uring filesystem support, and standard output | | networking-process-signals.md | Sockets, networking targets, child processes, and signals | | tokio-util-and-stream.md | Companion-crate dependency floors, task utilities, codecs, cancellation, and stream adapters |

Breaking changes and upgrade gates

Select safe patch releases

• Use at least 1.42.1 when a broadcast channel can carry a Send but !Sync value; 1.42.0 clones such values without synchronization.
• On the 1.43 line, use at least 1.43.2 for pidfd-backed child waiting, 1.43.3 for receiverless broadcast::Sender state, and 1.43.4 for closed-and-drained mpsc::Receiver::try_recv behavior.
• Use at least 1.45.1 on wasm32-unknown-unknown; 1.45.0 can panic when time-based metrics call Instant::now().
• Use at least 1.46.1 when task hooks inspect TaskMeta::spawned_at for tasks created with tokio::spawn.
• On the 1.47 line, use 1.47.2 for join!/try_join! macro hygiene, 1.47.4 for closed-channel recv_many, and 1.47.5 for corrected mpsc length, permit wakeups, outstanding-permit state, and nonzero RwLock reader limits.
• Use at least 1.51.1 when semaphore closure after forgotten permits, Linux UDP pending errors, or cancelled io_uring opens matter.
• Do not remain on 1.52.0 when using spawn_blocking; use at least 1.52.1 to avoid the sharded blocking-queue hang.
• Use 1.51.4 or 1.52.4 when a before_park callback can schedule work, so the runtime does not skip the driver.
• Use 1.53.1 for Tokio signal support on Windows with the Rust 1.71 toolchain; 1.53.0 accidentally used a newer standard-library API.

See compatibility.md for the complete upgrade and target matrix.

Enforce compiler and dependency floors

| Package floor | Requirement | | --- | --- | | Tokio 1.39.0 | Rust 1.70 or newer | | Tokio 1.48.0 | Rust 1.71 or newer | | tokio-util 0.7.12 | Rust 1.70 or newer | | tokio-util 0.7.17 | Rust 1.71 or newer | | tokio-util 0.7.18 | Tokio 1.44.0 or newer | | tokio-stream 0.1.13 | Rust 1.56 or newer | | tokio-stream 0.1.15 | Rust 1.63 or newer | | tokio-stream 0.1.16 | Rust 1.70 or newer | | tokio-stream 0.1.14 | Tokio 1.15 or newer |

Adjust rejected or changed code

• Put standard-library sockets into nonblocking mode before any Tokio from_std conversion. Since 1.44.0, passing a blocking socket panics.

let listener = std::net::TcpListener::bind("127.0.0.1:0")?;
listener.set_nonblocking(true)?;
let listener = tokio::net::TcpListener::from_std(listener)?;

• Keep runtime::Builder::event_interval nonzero; passing 0 panics as of 1.50.0.
• Keep in-place blocking operations out of futures and destructors driven by a LocalSet as of 1.46.0. Move blocking work to spawn_blocking.
• Replace deprecated TcpStream::set_linger and TcpSocket::set_linger calls. Use TcpStream::set_zero_linger() from 1.50.0 for abortive close behavior.
• Await, retain, or explicitly discard JoinHandle::abort_handle() and Notify::notified() results; their types are #[must_use] as of 1.40.0 and 1.41.0 respectively.
• Do not interpret signal-stream None as shutdown. Signal listeners remain open as of 1.50.0; arrange explicit cancellation.
• Expect task-owned state to be dropped before its JoinHandle completes as of 1.50.0.
• Expect a panicked task's formatted JoinError to include the panic message as of 1.40.0.
• Reuse or synchronize one Tokio standard-output handle when write ordering matters; multiple handles can reorder writes.

Update unstable configuration

• Select taskdump and io_uring with Cargo features as of 1.48.0, replacing their former custom --cfg switches.
• Pass LocalOptions by value to runtime::Builder::build_local as of 1.46.0.
• Migrate away from the removed unstable alternate multi-threaded runtime as of 1.45.0.
• Treat io_uring as opportunistic: Tokio disables it on EPERM, and later filesystem operations remain explicitly unstable.

Runtime and task quick reference

Choose local execution deliberately

• Use stable tokio::runtime::LocalRuntime from 1.51.0 for local task execution. It first appeared as unstable in 1.41.0.
• With unstable APIs on 1.48.0, use #[tokio::main(flavor = "local")] for a macro-created local runtime.
• Use stable runtime::id::Id and LocalSet::id() from 1.49.0 to identify runtimes and local sets.
• From 1.51.0, assign runtime names and call tokio::runtime::worker_index() for per-worker diagnostic context.

Preserve cooperative scheduling

• watch receives and broadcast::Receiver participate in cooperative scheduling as of 1.41.0.
• select! consumes cooperative task budget as of 1.44.0.
• Use task::coop for custom resources from 1.44.0; cooperative and poll_proceed are available from 1.47.0.
• yield_now takes effect immediately inside block_in_place as of 1.42.0.

Use task and runtime instrumentation

• Stable RuntimeMetrics::global_queue_depth and task Id APIs arrive in 1.41.0.
• Stable per-worker busy duration and park/unpark counts arrive in 1.45.0.
• Unstable spawned-task totals, combined worker park/unpark counts, and worker thread IDs arrive in 1.39.0.
• The runtime metrics API adds task schedule latency in 1.53.0.
• Do not use num_alive_tasks as an exact concurrent invariant; samples are not strongly consistent as of 1.49.0.

Read runtime-and-tasks.md for lifecycle and poll hooks, tracing, task dumps, caller-aware timeouts, and driver handoff.

Synchronization quick reference

• Use tokio::sync::SetOnce from 1.47.0 for asynchronously observable single-assignment state.
• Use Notify::notified_owned() and OwnedNotified from 1.47.0 when a notification future must not borrow its Notify.
• Await broadcast::Sender::closed() from 1.44.0 to stop a producer after all receivers disappear; use broadcast::WeakSender when an observer must not keep the channel open.
• Use mpsc::Receiver::blocking_recv_many from 1.41.0 for blocking batched receives.
• Derive Default for state containing watch::Sender<T> as of 1.39.0 when T satisfies the implementation's bounds.
• Use Tokio mpsc values across unwind-safety bounds as of 1.40.0.

Read synchronization.md before depending on close, permit, wakeup, or fairness semantics.

I/O, networking, process, and signal quick reference

I/O and files

• Use io::util::SimplexStream from 1.40.0, or tokio_util::io::simplex from tokio-util 0.7.18.
• Use public tokio::io::Chain from 1.48.0 when a chained reader's concrete type must be named. From 1.53.0, a zero-length read no longer advances it.
• Inspect a file's buffer limit with File::max_buf_size() from 1.48.0; file clones preserve that limit.
• Convert OwnedFd or OwnedHandle directly into tokio::fs::File from 1.53.0.
• Use Unix pipe try_io and AioSource::register_borrowed from 1.52.0 for immediate pipe operations and borrowed AIO registration.

Networking, processes, and signals

• Use Unix SocketAddr conversions from 1.41.0 and as_abstract_name() from 1.48.0; 1.53.0 expands its inspection methods.
• Use TcpStream::{quickack,set_quickack} from 1.48.0 where TCP quick acknowledgements are exposed, and configure IPv6 TCLASS from 1.49.0.
• Require 1.46.0 when successful macOS TcpStream::shutdown matters.
• Configure Unix child process groups with stable Command::process_group from 1.40.0.
• Use Command::spawn_with from 1.45.0; its callback may consume captured values through FnOnce as of 1.48.0.
• Treat Child::start_kill() after normal exit as a successful cleanup race as of 1.44.0.

Read io-and-filesystem.md and networking-process-signals.md for backend behavior and target-specific support.

Companion-crate quick reference

tokio-util

• Compose cancellation with run_until_cancelled, its owned form, or FutureExt adapters; cancellation wins a simultaneous-ready tie from 0.7.16.
• Use AbortOnDropHandle from 0.7.12 and detach it from 0.7.16.
• Use stable JoinMap from 0.7.16 and JoinQueue from 0.7.17.
• Recheck capacity assumptions: from 0.7.16, Framed::with_capacity applies its capacity to the read buffer too.

tokio-stream

• Use WatchStream::from_changes from 0.1.12 to skip the current value and wait for later watch-channel changes.
• Batch StreamMap results with next_many or poll_next_many from 0.1.16.
• Recover an incomplete timed chunk with ChunksTimeout::into_remainder from 0.1.18.
• Use meaningful receiver-stream size_hint bounds from 0.1.18.

Read tokio-util-and-stream.md for all companion-crate behavior and release floors.