executor: move current-thread into crate (#1447)

The `CurrentThread` executor is exposed using a feature flag.

Refs: #1264
This commit is contained in:
Carl Lerche
2019-08-15 09:52:25 -07:00
committed by GitHub
parent 8d55f98f6f
commit 9de7083be8
25 changed files with 110 additions and 174 deletions
+825
View File
@@ -0,0 +1,825 @@
//! A single-threaded executor which executes tasks on the same thread from which
//! they are spawned.
//!
//! [`CurrentThread`] is the main type of this crate. It executes tasks on the
//! current thread. The easiest way to start a new [`CurrentThread`] executor
//! is to call [`block_on_all`] with an initial task to seed the executor. All
//! tasks that are being managed by a [`CurrentThread`] executor are able to
//! spawn additional tasks by calling [`spawn`].
//!
//! Application authors will not use this crate directly. Instead, they will use
//! the `tokio` crate. Library authors should only depend on
//! `tokio-current-thread` if they are building a custom task executor.
//!
//! [`CurrentThread`]: struct.CurrentThread.html
//! [`spawn`]: fn.spawn.html
//! [`block_on_all`]: fn.block_on_all.html
mod scheduler;
use self::scheduler::Scheduler;
use crate::park::{Park, ParkThread, Unpark};
use crate::{EnterError, Executor, SpawnError, TypedExecutor};
use std::cell::Cell;
use std::error::Error;
use std::fmt;
use std::future::Future;
use std::pin::Pin;
use std::rc::Rc;
use std::sync::{atomic, Arc};
use std::task::{Context, Poll, Waker};
use std::thread;
use std::time::{Duration, Instant};
/// Executes tasks on the current thread
pub struct CurrentThread<P: Park = ParkThread> {
/// Execute futures and receive unpark notifications.
scheduler: Scheduler<P::Unpark>,
/// Current number of futures being executed.
///
/// The LSB is used to indicate that the runtime is preparing to shut down.
/// Thus, to get the actual number of pending futures, `>>1`.
num_futures: Arc<atomic::AtomicUsize>,
/// Thread park handle
park: P,
/// Handle for spawning new futures from other threads
spawn_handle: Handle,
/// Receiver for futures spawned from other threads
spawn_receiver: crossbeam_channel::Receiver<Pin<Box<dyn Future<Output = ()> + Send + 'static>>>,
/// The thread-local ID assigned to this executor.
id: u64,
}
/// Executes futures on the current thread.
///
/// All futures executed using this executor will be executed on the current
/// thread. As such, `run` will wait for these futures to complete before
/// returning.
///
/// For more details, see the [module level](index.html) documentation.
#[derive(Debug, Clone)]
pub struct TaskExecutor {
// Prevent the handle from moving across threads.
_p: ::std::marker::PhantomData<Rc<()>>,
}
/// Returned by the `turn` function.
#[derive(Debug)]
pub struct Turn {
polled: bool,
}
impl Turn {
/// `true` if any futures were polled at all and `false` otherwise.
pub fn has_polled(&self) -> bool {
self.polled
}
}
/// A `CurrentThread` instance bound to a supplied execution context.
pub struct Entered<'a, P: Park> {
executor: &'a mut CurrentThread<P>,
}
/// Error returned by the `run` function.
#[derive(Debug)]
pub struct RunError {
_p: (),
}
impl fmt::Display for RunError {
fn fmt(&self, fmt: &mut fmt::Formatter<'_>) -> fmt::Result {
write!(fmt, "Run error")
}
}
impl Error for RunError {}
/// Error returned by the `run_timeout` function.
#[derive(Debug)]
pub struct RunTimeoutError {
timeout: bool,
}
impl fmt::Display for RunTimeoutError {
fn fmt(&self, fmt: &mut fmt::Formatter<'_>) -> fmt::Result {
let descr = if self.timeout {
"Run timeout error (timeout)"
} else {
"Run timeout error (not timeout)"
};
write!(fmt, "{}", descr)
}
}
impl Error for RunTimeoutError {}
/// Error returned by the `turn` function.
#[derive(Debug)]
pub struct TurnError {
_p: (),
}
impl fmt::Display for TurnError {
fn fmt(&self, fmt: &mut fmt::Formatter<'_>) -> fmt::Result {
write!(fmt, "Turn error")
}
}
impl Error for TurnError {}
/// Error returned by the `block_on` function.
#[derive(Debug)]
pub struct BlockError<T> {
inner: Option<T>,
}
impl<T> fmt::Display for BlockError<T> {
fn fmt(&self, fmt: &mut fmt::Formatter<'_>) -> fmt::Result {
write!(fmt, "Block error")
}
}
impl<T: fmt::Debug> Error for BlockError<T> {}
/// This is mostly split out to make the borrow checker happy.
struct Borrow<'a, U> {
id: u64,
scheduler: &'a mut Scheduler<U>,
num_futures: &'a atomic::AtomicUsize,
}
trait SpawnLocal {
fn spawn_local(&mut self, future: Pin<Box<dyn Future<Output = ()>>>, already_counted: bool);
}
struct CurrentRunner {
spawn: Cell<Option<*mut dyn SpawnLocal>>,
id: Cell<Option<u64>>,
}
thread_local! {
/// Current thread's task runner. This is set in `TaskRunner::with`
static CURRENT: CurrentRunner = CurrentRunner {
spawn: Cell::new(None),
id: Cell::new(None),
}
}
thread_local! {
/// Unique ID to assign to each new executor launched on this thread.
///
/// The unique ID is used to determine if the currently running executor matches the one
/// referred to by a `Handle` so that direct task dispatch can be used.
static EXECUTOR_ID: Cell<u64> = Cell::new(0)
}
/// Run the executor bootstrapping the execution with the provided future.
///
/// This creates a new [`CurrentThread`] executor, spawns the provided future,
/// and blocks the current thread until the provided future and **all**
/// subsequently spawned futures complete. In other words:
///
/// * If the provided bootstrap future does **not** spawn any additional tasks,
/// `block_on_all` returns once `future` completes.
/// * If the provided bootstrap future **does** spawn additional tasks, then
/// `block_on_all` returns once **all** spawned futures complete.
///
/// See [module level][mod] documentation for more details.
///
/// [`CurrentThread`]: struct.CurrentThread.html
/// [mod]: index.html
pub fn block_on_all<F>(future: F) -> F::Output
where
F: Future,
{
let mut current_thread = CurrentThread::new();
let ret = current_thread.block_on(future);
current_thread.run().unwrap();
ret
}
/// Executes a future on the current thread.
///
/// The provided future must complete or be canceled before `run` will return.
///
/// Unlike [`tokio::spawn`], this function will always spawn on a
/// `CurrentThread` executor and is able to spawn futures that are not `Send`.
///
/// # Panics
///
/// This function can only be invoked from the context of a `run` call; any
/// other use will result in a panic.
///
/// [`tokio::spawn`]: ../fn.spawn.html
pub fn spawn<F>(future: F)
where
F: Future<Output = ()> + 'static,
{
TaskExecutor::current()
.spawn_local(Box::pin(future))
.unwrap();
}
// ===== impl CurrentThread =====
impl CurrentThread<ParkThread> {
/// Create a new instance of `CurrentThread`.
pub fn new() -> Self {
CurrentThread::new_with_park(ParkThread::new())
}
}
impl<P: Park> CurrentThread<P> {
/// Create a new instance of `CurrentThread` backed by the given park
/// handle.
pub fn new_with_park(park: P) -> Self {
let unpark = park.unpark();
let (spawn_sender, spawn_receiver) = crossbeam_channel::unbounded();
let thread = thread::current().id();
let id = EXECUTOR_ID.with(|idc| {
let id = idc.get();
idc.set(id + 1);
id
});
let scheduler = Scheduler::new(unpark);
let waker = scheduler.waker();
let num_futures = Arc::new(atomic::AtomicUsize::new(0));
CurrentThread {
scheduler,
num_futures: num_futures.clone(),
park,
id,
spawn_handle: Handle {
sender: spawn_sender,
num_futures,
waker,
thread,
id,
},
spawn_receiver,
}
}
/// Returns `true` if the executor is currently idle.
///
/// An idle executor is defined by not currently having any spawned tasks.
///
/// Note that this method is inherently racy -- if a future is spawned from a remote `Handle`,
/// this method may return `true` even though there are more futures to be executed.
pub fn is_idle(&self) -> bool {
self.num_futures.load(atomic::Ordering::SeqCst) <= 1
}
/// Spawn the future on the executor.
///
/// This internally queues the future to be executed once `run` is called.
pub fn spawn<F>(&mut self, future: F) -> &mut Self
where
F: Future<Output = ()> + 'static,
{
self.borrow().spawn_local(Box::pin(future), false);
self
}
/// Synchronously waits for the provided `future` to complete.
///
/// This function can be used to synchronously block the current thread
/// until the provided `future` has resolved either successfully or with an
/// error. The result of the future is then returned from this function
/// call.
///
/// Note that this function will **also** execute any spawned futures on the
/// current thread, but will **not** block until these other spawned futures
/// have completed.
///
/// The caller is responsible for ensuring that other spawned futures
/// complete execution.
pub fn block_on<F>(&mut self, future: F) -> F::Output
where
F: Future,
{
let _enter = crate::enter().expect("failed to start `current_thread::Runtime`");
self.enter().block_on(future)
}
/// Run the executor to completion, blocking the thread until **all**
/// spawned futures have completed.
pub fn run(&mut self) -> Result<(), RunError> {
let _enter = crate::enter().expect("failed to start `current_thread::Runtime`");
self.enter().run()
}
/// Run the executor to completion, blocking the thread until all
/// spawned futures have completed **or** `duration` time has elapsed.
pub fn run_timeout(&mut self, duration: Duration) -> Result<(), RunTimeoutError> {
let _enter = crate::enter().expect("failed to start `current_thread::Runtime`");
self.enter().run_timeout(duration)
}
/// Perform a single iteration of the event loop.
///
/// This function blocks the current thread even if the executor is idle.
pub fn turn(&mut self, duration: Option<Duration>) -> Result<Turn, TurnError> {
let _enter = crate::enter().expect("failed to start `current_thread::Runtime`");
self.enter().turn(duration)
}
/// Bind `CurrentThread` instance with an execution context.
fn enter(&mut self) -> Entered<'_, P> {
Entered { executor: self }
}
/// Returns a reference to the underlying `Park` instance.
pub fn get_park(&self) -> &P {
&self.park
}
/// Returns a mutable reference to the underlying `Park` instance.
pub fn get_park_mut(&mut self) -> &mut P {
&mut self.park
}
fn borrow(&mut self) -> Borrow<'_, P::Unpark> {
Borrow {
id: self.id,
scheduler: &mut self.scheduler,
num_futures: &*self.num_futures,
}
}
/// Get a new handle to spawn futures on the executor
///
/// Different to the executor itself, the handle can be sent to different
/// threads and can be used to spawn futures on the executor.
pub fn handle(&self) -> Handle {
self.spawn_handle.clone()
}
}
impl<P: Park> Drop for CurrentThread<P> {
fn drop(&mut self) {
// Signal to Handles that no more futures can be spawned by setting LSB.
//
// NOTE: this isn't technically necessary since the send on the mpsc will fail once the
// receiver is dropped, but it's useful to illustrate how clean shutdown will be
// implemented (e.g., by setting the LSB).
let pending = self.num_futures.fetch_add(1, atomic::Ordering::SeqCst);
// TODO: We currently ignore any pending futures at the time we shut down.
//
// The "proper" fix for this is to have an explicit shutdown phase (`shutdown_on_idle`)
// which sets LSB (as above) do make Handle::spawn stop working, and then runs until
// num_futures.load() == 1.
let _ = pending;
}
}
impl Executor for CurrentThread {
fn spawn(
&mut self,
future: Pin<Box<dyn Future<Output = ()> + Send>>,
) -> Result<(), SpawnError> {
self.borrow().spawn_local(future, false);
Ok(())
}
}
impl<T> TypedExecutor<T> for CurrentThread
where
T: Future<Output = ()> + 'static,
{
fn spawn(&mut self, future: T) -> Result<(), SpawnError> {
self.borrow().spawn_local(Box::pin(future), false);
Ok(())
}
}
impl<P: Park> fmt::Debug for CurrentThread<P> {
fn fmt(&self, fmt: &mut fmt::Formatter<'_>) -> fmt::Result {
fmt.debug_struct("CurrentThread")
.field("scheduler", &self.scheduler)
.field(
"num_futures",
&self.num_futures.load(atomic::Ordering::SeqCst),
)
.finish()
}
}
impl<P: Park + Default> Default for CurrentThread<P> {
fn default() -> Self {
CurrentThread::new_with_park(P::default())
}
}
// ===== impl Entered =====
impl<P: Park> Entered<'_, P> {
/// Spawn the future on the executor.
///
/// This internally queues the future to be executed once `run` is called.
pub fn spawn<F>(&mut self, future: F) -> &mut Self
where
F: Future<Output = ()> + 'static,
{
self.executor.borrow().spawn_local(Box::pin(future), false);
self
}
/// Synchronously waits for the provided `future` to complete.
///
/// This function can be used to synchronously block the current thread
/// until the provided `future` has resolved either successfully or with an
/// error. The result of the future is then returned from this function
/// call.
///
/// Note that this function will **also** execute any spawned futures on the
/// current thread, but will **not** block until these other spawned futures
/// have completed.
///
/// The caller is responsible for ensuring that other spawned futures
/// complete execution.
///
/// # Panics
///
/// This function will panic if the `Park` call returns an error.
pub fn block_on<F>(&mut self, mut future: F) -> F::Output
where
F: Future,
{
// Safety: we shadow the original `future`, so it will never move
// again.
let mut future = unsafe { Pin::new_unchecked(&mut future) };
let waker = self.executor.scheduler.waker();
let mut cx = Context::from_waker(&waker);
loop {
let res = self
.executor
.borrow()
.enter(|| future.as_mut().poll(&mut cx));
match res {
Poll::Ready(e) => return e,
Poll::Pending => {}
}
self.tick();
if self.executor.park.park().is_err() {
panic!("block_on park failed");
}
}
}
/// Run the executor to completion, blocking the thread until **all**
/// spawned futures have completed.
pub fn run(&mut self) -> Result<(), RunError> {
self.run_timeout2(None).map_err(|_| RunError { _p: () })
}
/// Run the executor to completion, blocking the thread until all
/// spawned futures have completed **or** `duration` time has elapsed.
pub fn run_timeout(&mut self, duration: Duration) -> Result<(), RunTimeoutError> {
self.run_timeout2(Some(duration))
}
/// Perform a single iteration of the event loop.
///
/// This function blocks the current thread even if the executor is idle.
pub fn turn(&mut self, duration: Option<Duration>) -> Result<Turn, TurnError> {
let res = if self.executor.scheduler.has_pending_futures() {
self.executor.park.park_timeout(Duration::from_millis(0))
} else {
match duration {
Some(duration) => self.executor.park.park_timeout(duration),
None => self.executor.park.park(),
}
};
if res.is_err() {
return Err(TurnError { _p: () });
}
let polled = self.tick();
Ok(Turn { polled })
}
/// Returns a reference to the underlying `Park` instance.
pub fn get_park(&self) -> &P {
&self.executor.park
}
/// Returns a mutable reference to the underlying `Park` instance.
pub fn get_park_mut(&mut self) -> &mut P {
&mut self.executor.park
}
fn run_timeout2(&mut self, dur: Option<Duration>) -> Result<(), RunTimeoutError> {
if self.executor.is_idle() {
// Nothing to do
return Ok(());
}
let mut time = dur.map(|dur| (Instant::now() + dur, dur));
loop {
self.tick();
if self.executor.is_idle() {
return Ok(());
}
match time {
Some((until, rem)) => {
if self.executor.park.park_timeout(rem).is_err() {
return Err(RunTimeoutError::new(false));
}
let now = Instant::now();
if now >= until {
return Err(RunTimeoutError::new(true));
}
time = Some((until, until - now));
}
None => {
if self.executor.park.park().is_err() {
return Err(RunTimeoutError::new(false));
}
}
}
}
}
/// Returns `true` if any futures were processed
fn tick(&mut self) -> bool {
// Spawn any futures that were spawned from other threads by manually
// looping over the receiver stream
// FIXME: Slightly ugly but needed to make the borrow checker happy
let (mut borrow, spawn_receiver) = (
Borrow {
id: self.executor.id,
scheduler: &mut self.executor.scheduler,
num_futures: &*self.executor.num_futures,
},
&mut self.executor.spawn_receiver,
);
while let Ok(future) = spawn_receiver.try_recv() {
borrow.spawn_local(future, true);
}
// After any pending futures were scheduled, do the actual tick
borrow.scheduler.tick(borrow.id, borrow.num_futures)
}
}
impl<P: Park> fmt::Debug for Entered<'_, P> {
fn fmt(&self, fmt: &mut fmt::Formatter<'_>) -> fmt::Result {
fmt.debug_struct("Entered")
.field("executor", &self.executor)
.finish()
}
}
// ===== impl Handle =====
/// Handle to spawn a future on the corresponding `CurrentThread` instance
#[derive(Clone)]
pub struct Handle {
sender: crossbeam_channel::Sender<Pin<Box<dyn Future<Output = ()> + Send + 'static>>>,
num_futures: Arc<atomic::AtomicUsize>,
/// Waker to the Scheduler
waker: Waker,
thread: thread::ThreadId,
/// The thread-local ID assigned to this Handle's executor.
id: u64,
}
// Manual implementation because the Sender does not implement Debug
impl fmt::Debug for Handle {
fn fmt(&self, fmt: &mut fmt::Formatter<'_>) -> fmt::Result {
fmt.debug_struct("Handle")
.field("shut_down", &self.is_shut_down())
.finish()
}
}
impl Handle {
/// Spawn a future onto the `CurrentThread` instance corresponding to this handle
///
/// # Panics
///
/// This function panics if the spawn fails. Failure occurs if the `CurrentThread`
/// instance of the `Handle` does not exist anymore.
pub fn spawn<F>(&self, future: F) -> Result<(), SpawnError>
where
F: Future<Output = ()> + Send + 'static,
{
if thread::current().id() == self.thread {
let mut e = TaskExecutor::current();
if e.id() == Some(self.id) {
return e.spawn_local(Box::pin(future));
}
}
// NOTE: += 2 since LSB is the shutdown bit
let pending = self.num_futures.fetch_add(2, atomic::Ordering::SeqCst);
if pending % 2 == 1 {
// Bring the count back so we still know when the Runtime is idle.
self.num_futures.fetch_sub(2, atomic::Ordering::SeqCst);
return Err(SpawnError::shutdown());
}
self.sender
.send(Box::pin(future))
.expect("CurrentThread does not exist anymore");
self.waker.wake_by_ref();
Ok(())
}
/// Provides a best effort **hint** to whether or not `spawn` will succeed.
///
/// This function may return both false positives **and** false negatives.
/// If `status` returns `Ok`, then a call to `spawn` will *probably*
/// succeed, but may fail. If `status` returns `Err`, a call to `spawn` will
/// *probably* fail, but may succeed.
///
/// This allows a caller to avoid creating the task if the call to `spawn`
/// has a high likelihood of failing.
pub fn status(&self) -> Result<(), SpawnError> {
if self.is_shut_down() {
return Err(SpawnError::shutdown());
}
Ok(())
}
fn is_shut_down(&self) -> bool {
// LSB of "num_futures" is the shutdown bit
let num_futures = self.num_futures.load(atomic::Ordering::SeqCst);
num_futures % 2 == 1
}
}
// ===== impl TaskExecutor =====
impl TaskExecutor {
/// Returns an executor that executes futures on the current thread.
///
/// The user of `TaskExecutor` must ensure that when a future is submitted,
/// that it is done within the context of a call to `run`.
///
/// For more details, see the [module level](index.html) documentation.
pub fn current() -> TaskExecutor {
TaskExecutor {
_p: ::std::marker::PhantomData,
}
}
/// Get the current executor's thread-local ID.
fn id(&self) -> Option<u64> {
CURRENT.with(|current| current.id.get())
}
/// Spawn a future onto the current `CurrentThread` instance.
pub fn spawn_local(
&mut self,
future: Pin<Box<dyn Future<Output = ()>>>,
) -> Result<(), SpawnError> {
CURRENT.with(|current| match current.spawn.get() {
Some(spawn) => {
unsafe { (*spawn).spawn_local(future, false) };
Ok(())
}
None => Err(SpawnError::shutdown()),
})
}
}
impl Executor for TaskExecutor {
fn spawn(
&mut self,
future: Pin<Box<dyn Future<Output = ()> + Send>>,
) -> Result<(), SpawnError> {
self.spawn_local(future)
}
}
impl<F> TypedExecutor<F> for TaskExecutor
where
F: Future<Output = ()> + 'static,
{
fn spawn(&mut self, future: F) -> Result<(), SpawnError> {
self.spawn_local(Box::pin(future))
}
}
// ===== impl Borrow =====
impl<U: Unpark> Borrow<'_, U> {
fn enter<F, R>(&mut self, f: F) -> R
where
F: FnOnce() -> R,
{
CURRENT.with(|current| {
current.id.set(Some(self.id));
current.set_spawn(self, || f())
})
}
}
impl<U: Unpark> SpawnLocal for Borrow<'_, U> {
fn spawn_local(&mut self, future: Pin<Box<dyn Future<Output = ()>>>, already_counted: bool) {
if !already_counted {
// NOTE: we have a borrow of the Runtime, so we know that it isn't shut down.
// NOTE: += 2 since LSB is the shutdown bit
self.num_futures.fetch_add(2, atomic::Ordering::SeqCst);
}
self.scheduler.schedule(future);
}
}
// ===== impl CurrentRunner =====
impl CurrentRunner {
fn set_spawn<F, R>(&self, spawn: &mut dyn SpawnLocal, f: F) -> R
where
F: FnOnce() -> R,
{
struct Reset<'a>(&'a CurrentRunner);
impl Drop for Reset<'_> {
fn drop(&mut self) {
self.0.spawn.set(None);
self.0.id.set(None);
}
}
let _reset = Reset(self);
let spawn = unsafe { hide_lt(spawn as *mut dyn SpawnLocal) };
self.spawn.set(Some(spawn));
f()
}
}
unsafe fn hide_lt<'a>(p: *mut (dyn SpawnLocal + 'a)) -> *mut (dyn SpawnLocal + 'static) {
use std::mem;
// false positive: https://github.com/rust-lang/rust-clippy/issues/2906
#[allow(clippy::transmute_ptr_to_ptr)]
mem::transmute(p)
}
// ===== impl RunTimeoutError =====
impl RunTimeoutError {
fn new(timeout: bool) -> Self {
RunTimeoutError { timeout }
}
/// Returns `true` if the error was caused by the operation timing out.
pub fn is_timeout(&self) -> bool {
self.timeout
}
}
impl From<EnterError> for RunTimeoutError {
fn from(_: EnterError) -> Self {
RunTimeoutError::new(false)
}
}
// ===== impl BlockError =====
impl<T> BlockError<T> {
/// Returns the error yielded by the future being blocked on
pub fn into_inner(self) -> Option<T> {
self.inner
}
}
impl<T> From<EnterError> for BlockError<T> {
fn from(_: EnterError) -> Self {
BlockError { inner: None }
}
}
@@ -0,0 +1,797 @@
use super::Borrow;
use crate::park::Unpark;
use std::cell::UnsafeCell;
use std::fmt::{self, Debug};
use std::future::Future;
use std::mem;
use std::pin::Pin;
use std::ptr;
use std::sync::atomic::Ordering::{AcqRel, Acquire, Relaxed, Release, SeqCst};
use std::sync::atomic::{AtomicBool, AtomicPtr, AtomicUsize};
use std::sync::{Arc, Weak};
use std::task::{Context, Poll, RawWaker, RawWakerVTable, Waker};
use std::thread;
use std::usize;
/// A generic task-aware scheduler.
///
/// This is used both by `FuturesUnordered` and the current-thread executor.
pub(crate) struct Scheduler<U> {
inner: Arc<Inner<U>>,
nodes: List<U>,
}
// A linked-list of nodes
struct List<U> {
len: usize,
head: *const Node<U>,
tail: *const Node<U>,
}
// Scheduler is implemented using two linked lists. The first linked list tracks
// all items managed by a `Scheduler`. This list is stored on the `Scheduler`
// struct and is **not** thread safe. The second linked list is an
// implementation of the intrusive MPSC queue algorithm described by
// 1024cores.net and is stored on `Inner`. This linked list can push items to
// the back concurrently but only one consumer may pop from the front. To
// enforce this requirement, all popping will be performed via fns on
// `Scheduler` that take `&mut self`.
//
// When a item is submitted to the set a node is allocated and inserted in
// both linked lists. This means that all insertion operations **must** be
// originated from `Scheduler` with `&mut self` The next call to `tick` will
// (eventually) see this node and call `poll` on the item.
//
// Nodes are wrapped in `Arc` cells which manage the lifetime of the node.
// However, `Arc` handles are sometimes cast to `*const Node` pointers.
// Specifically, when a node is stored in at least one of the two lists
// described above, this represents a logical `Arc` handle. This is how
// `Scheduler` maintains its reference to all nodes it manages. Each
// `NotifyHandle` instance is an `Arc<Node>` as well.
//
// When `Scheduler` drops, it clears the linked list of all nodes that it
// manages. When doing so, it must attempt to decrement the reference count (by
// dropping an Arc handle). However, it can **only** decrement the reference
// count if the node is not currently stored in the mpsc channel. If the node
// **is** "queued" in the mpsc channel, then the arc reference count cannot be
// decremented. Once the node is popped from the mpsc channel, then the final
// arc reference count can be decremented, thus freeing the node.
struct Inner<U> {
// Thread unpark handle
unpark: U,
// Tick number
tick_num: AtomicUsize,
// Head/tail of the readiness queue
head_readiness: AtomicPtr<Node<U>>,
tail_readiness: UnsafeCell<*const Node<U>>,
// Used as part of the mpsc queue algorithm
stub: Arc<Node<U>>,
}
unsafe impl<U: Sync + Send> Send for Inner<U> {}
unsafe impl<U: Sync + Send> Sync for Inner<U> {}
struct Node<U> {
// The item
item: UnsafeCell<Option<Task>>,
// The tick at which this node was notified
notified_at: AtomicUsize,
// Next pointer for linked list tracking all active nodes
next_all: UnsafeCell<*const Node<U>>,
// Previous node in linked list tracking all active nodes
prev_all: UnsafeCell<*const Node<U>>,
// Next pointer in readiness queue
next_readiness: AtomicPtr<Node<U>>,
// Whether or not this node is currently in the mpsc queue.
queued: AtomicBool,
// Queue that we'll be enqueued to when notified
queue: Weak<Inner<U>>,
}
/// Returned by `Inner::dequeue`, representing either a dequeue success (with
/// the dequeued node), an empty list, or an inconsistent state.
///
/// The inconsistent state is described in more detail at [1024cores], but
/// roughly indicates that a node will be ready to dequeue sometime shortly in
/// the future and the caller should try again soon.
///
/// [1024cores]: http://www.1024cores.net/home/lock-free-algorithms/queues/intrusive-mpsc-node-based-queue
enum Dequeue<U> {
Data(*const Node<U>),
Empty,
Yield,
Inconsistent,
}
/// Wraps a spawned boxed future
struct Task(Pin<Box<dyn Future<Output = ()>>>);
/// A task that is scheduled. `turn` must be called
pub(crate) struct Scheduled<'a, U> {
task: &'a mut Task,
node: &'a Arc<Node<U>>,
done: &'a mut bool,
}
impl<U> Scheduler<U>
where
U: Unpark,
{
/// Constructs a new, empty `Scheduler`
///
/// The returned `Scheduler` does not contain any items and, in this
/// state, `Scheduler::poll` will return `Ok(Async::Ready(None))`.
pub(crate) fn new(unpark: U) -> Self {
let stub = Arc::new(Node {
item: UnsafeCell::new(None),
notified_at: AtomicUsize::new(0),
next_all: UnsafeCell::new(ptr::null()),
prev_all: UnsafeCell::new(ptr::null()),
next_readiness: AtomicPtr::new(ptr::null_mut()),
queued: AtomicBool::new(true),
queue: Weak::new(),
});
let stub_ptr = &*stub as *const Node<U>;
let inner = Arc::new(Inner {
unpark,
tick_num: AtomicUsize::new(0),
head_readiness: AtomicPtr::new(stub_ptr as *mut _),
tail_readiness: UnsafeCell::new(stub_ptr),
stub,
});
Scheduler {
inner,
nodes: List::new(),
}
}
pub(crate) fn waker(&self) -> Waker {
waker_inner(self.inner.clone())
}
pub(crate) fn schedule(&mut self, item: Pin<Box<dyn Future<Output = ()>>>) {
// Get the current scheduler tick
let tick_num = self.inner.tick_num.load(SeqCst);
let node = Arc::new(Node {
item: UnsafeCell::new(Some(Task::new(item))),
notified_at: AtomicUsize::new(tick_num),
next_all: UnsafeCell::new(ptr::null_mut()),
prev_all: UnsafeCell::new(ptr::null_mut()),
next_readiness: AtomicPtr::new(ptr::null_mut()),
queued: AtomicBool::new(true),
queue: Arc::downgrade(&self.inner),
});
// Right now our node has a strong reference count of 1. We transfer
// ownership of this reference count to our internal linked list
// and we'll reclaim ownership through the `unlink` function below.
let ptr = self.nodes.push_back(node);
// We'll need to get the item "into the system" to start tracking it,
// e.g. getting its unpark notifications going to us tracking which
// items are ready. To do that we unconditionally enqueue it for
// polling here.
self.inner.enqueue(ptr);
}
/// Returns `true` if there are currently any pending futures
pub(crate) fn has_pending_futures(&mut self) -> bool {
// See function definition for why the unsafe is needed and
// correctly used here
unsafe { self.inner.has_pending_futures() }
}
/// Advance the scheduler state, returning `true` if any futures were
/// processed.
///
/// This function should be called whenever the caller is notified via a
/// wakeup.
pub(crate) fn tick(&mut self, eid: u64, num_futures: &AtomicUsize) -> bool {
let mut ret = false;
let tick = self.inner.tick_num.fetch_add(1, SeqCst).wrapping_add(1);
loop {
let node = match unsafe { self.inner.dequeue(Some(tick)) } {
Dequeue::Empty => {
return ret;
}
Dequeue::Yield => {
self.inner.unpark.unpark();
return ret;
}
Dequeue::Inconsistent => {
thread::yield_now();
continue;
}
Dequeue::Data(node) => node,
};
ret = true;
debug_assert!(node != self.inner.stub());
unsafe {
if (*(*node).item.get()).is_none() {
// The node has already been released. However, while it was
// being released, another thread notified it, which
// resulted in it getting pushed into the mpsc channel.
//
// In this case, we just decrement the ref count.
let node = ptr2arc(node);
assert!((*node.next_all.get()).is_null());
assert!((*node.prev_all.get()).is_null());
continue;
};
// We're going to need to be very careful if the `poll`
// function below panics. We need to (a) not leak memory and
// (b) ensure that we still don't have any use-after-frees. To
// manage this we do a few things:
//
// * This "bomb" here will call `release_node` if dropped
// abnormally. That way we'll be sure the memory management
// of the `node` is managed correctly.
//
// * We unlink the node from our internal queue to preemptively
// assume is is complete (will return Ready or panic), in
// which case we'll want to discard it regardless.
//
struct Bomb<'a, U: Unpark> {
borrow: &'a mut Borrow<'a, U>,
node: Option<Arc<Node<U>>>,
}
impl<U: Unpark> Drop for Bomb<'_, U> {
fn drop(&mut self) {
if let Some(node) = self.node.take() {
self.borrow.enter(|| release_node(node))
}
}
}
let node = self.nodes.remove(node);
let mut borrow = Borrow {
id: eid,
scheduler: self,
num_futures,
};
let mut bomb = Bomb {
node: Some(node),
borrow: &mut borrow,
};
let mut done = false;
// Now that the bomb holds the node, create a new scope. This
// scope ensures that the borrow will go out of scope before we
// mutate the node pointer in `bomb` again
{
let node = bomb.node.as_ref().unwrap();
// Get a reference to the inner future. We already ensured
// that the item `is_some`.
let item = (*node.item.get()).as_mut().unwrap();
// Unset queued flag... this must be done before
// polling. This ensures that the item gets
// rescheduled if it is notified **during** a call
// to `poll`.
let prev = (*node).queued.swap(false, SeqCst);
assert!(prev);
// Poll the underlying item with the appropriate `notify`
// implementation. This is where a large bit of the unsafety
// starts to stem from internally. The `notify` instance itself
// is basically just our `Arc<Node>` and tracks the mpsc
// queue of ready items.
//
// Critically though `Node` won't actually access `Task`, the
// item, while it's floating around inside of `Task`
// instances. These structs will basically just use `T` to size
// the internal allocation, appropriately accessing fields and
// deallocating the node if need be.
let borrow = &mut *bomb.borrow;
let mut scheduled = Scheduled {
task: item,
node: bomb.node.as_ref().unwrap(),
done: &mut done,
};
if borrow.enter(|| scheduled.tick()) {
// we have a borrow of the Runtime, so we know it's not shut down
borrow.num_futures.fetch_sub(2, SeqCst);
}
}
if !done {
// The future is not done, push it back into the "all
// node" list.
let node = bomb.node.take().unwrap();
bomb.borrow.scheduler.nodes.push_back(node);
}
}
}
}
}
impl<U: Unpark> Scheduled<'_, U> {
/// Polls the task, returns `true` if the task has completed.
pub(crate) fn tick(&mut self) -> bool {
let waker = unsafe {
// Safety: we don't hold this waker ref longer than
// this `tick` function
waker_ref(self.node)
};
let mut cx = Context::from_waker(&waker);
let ret = match self.task.0.as_mut().poll(&mut cx) {
Poll::Ready(()) => true,
Poll::Pending => false,
};
*self.done = ret;
ret
}
}
impl Task {
pub(crate) fn new(future: Pin<Box<dyn Future<Output = ()> + 'static>>) -> Self {
Task(future)
}
}
impl fmt::Debug for Task {
fn fmt(&self, fmt: &mut fmt::Formatter<'_>) -> fmt::Result {
fmt.debug_struct("Task").finish()
}
}
fn release_node<U>(node: Arc<Node<U>>) {
// The item is done, try to reset the queued flag. This will prevent
// `notify` from doing any work in the item
let prev = node.queued.swap(true, SeqCst);
// Drop the item, even if it hasn't finished yet. This is safe
// because we're dropping the item on the thread that owns
// `Scheduler`, which correctly tracks T's lifetimes and such.
unsafe {
drop((*node.item.get()).take());
}
// If the queued flag was previously set then it means that this node
// is still in our internal mpsc queue. We then transfer ownership
// of our reference count to the mpsc queue, and it'll come along and
// free it later, noticing that the item is `None`.
//
// If, however, the queued flag was *not* set then we're safe to
// release our reference count on the internal node. The queued flag
// was set above so all item `enqueue` operations will not actually
// enqueue the node, so our node will never see the mpsc queue again.
// The node itself will be deallocated once all reference counts have
// been dropped by the various owning tasks elsewhere.
if prev {
mem::forget(node);
}
}
impl<U> Debug for Scheduler<U> {
fn fmt(&self, fmt: &mut fmt::Formatter<'_>) -> fmt::Result {
write!(fmt, "Scheduler {{ ... }}")
}
}
impl<U> Drop for Scheduler<U> {
fn drop(&mut self) {
// When a `Scheduler` is dropped we want to drop all items associated
// with it. At the same time though there may be tons of `Task` handles
// flying around which contain `Node` references inside them. We'll
// let those naturally get deallocated when the `Task` itself goes out
// of scope or gets notified.
while let Some(node) = self.nodes.pop_front() {
release_node(node);
}
// Note that at this point we could still have a bunch of nodes in the
// mpsc queue. None of those nodes, however, have items associated
// with them so they're safe to destroy on any thread. At this point
// the `Scheduler` struct, the owner of the one strong reference
// to `Inner` will drop the strong reference. At that point
// whichever thread releases the strong refcount last (be it this
// thread or some other thread as part of an `upgrade`) will clear out
// the mpsc queue and free all remaining nodes.
//
// While that freeing operation isn't guaranteed to happen here, it's
// guaranteed to happen "promptly" as no more "blocking work" will
// happen while there's a strong refcount held.
}
}
impl<U> Inner<U> {
/// The enqueue function from the 1024cores intrusive MPSC queue algorithm.
fn enqueue(&self, node: *const Node<U>) {
unsafe {
debug_assert!((*node).queued.load(Relaxed));
// This action does not require any coordination
(*node).next_readiness.store(ptr::null_mut(), Relaxed);
// Note that these atomic orderings come from 1024cores
let node = node as *mut _;
let prev = self.head_readiness.swap(node, AcqRel);
(*prev).next_readiness.store(node, Release);
}
}
/// Returns `true` if there are currently any pending futures
///
/// See `dequeue` for an explanation why this function is unsafe.
unsafe fn has_pending_futures(&self) -> bool {
let tail = *self.tail_readiness.get();
let next = (*tail).next_readiness.load(Acquire);
if tail == self.stub() && next.is_null() {
return false;
}
true
}
/// The dequeue function from the 1024cores intrusive MPSC queue algorithm
///
/// Note that this unsafe as it required mutual exclusion (only one thread
/// can call this) to be guaranteed elsewhere.
unsafe fn dequeue(&self, tick: Option<usize>) -> Dequeue<U> {
let mut tail = *self.tail_readiness.get();
let mut next = (*tail).next_readiness.load(Acquire);
if tail == self.stub() {
if next.is_null() {
return Dequeue::Empty;
}
*self.tail_readiness.get() = next;
tail = next;
next = (*next).next_readiness.load(Acquire);
}
if let Some(tick) = tick {
let actual = (*tail).notified_at.load(SeqCst);
// Only dequeue if the node was not scheduled during the current
// tick.
if actual == tick {
// Only doing the check above **should** be enough in
// practice. However, technically there is a potential for
// deadlocking if there are `usize::MAX` ticks while the thread
// scheduling the task is frozen.
//
// If, for some reason, this is not enough, calling `unpark`
// here will resolve the issue.
return Dequeue::Yield;
}
}
if !next.is_null() {
*self.tail_readiness.get() = next;
debug_assert!(tail != self.stub());
return Dequeue::Data(tail);
}
if self.head_readiness.load(Acquire) as *const _ != tail {
return Dequeue::Inconsistent;
}
self.enqueue(self.stub());
next = (*tail).next_readiness.load(Acquire);
if !next.is_null() {
*self.tail_readiness.get() = next;
return Dequeue::Data(tail);
}
Dequeue::Inconsistent
}
fn stub(&self) -> *const Node<U> {
&*self.stub
}
}
impl<U> Drop for Inner<U> {
fn drop(&mut self) {
// Once we're in the destructor for `Inner` we need to clear out the
// mpsc queue of nodes if there's anything left in there.
//
// Note that each node has a strong reference count associated with it
// which is owned by the mpsc queue. All nodes should have had their
// items dropped already by the `Scheduler` destructor above,
// so we're just pulling out nodes and dropping their refcounts.
unsafe {
loop {
match self.dequeue(None) {
Dequeue::Empty => break,
Dequeue::Yield => unreachable!(),
Dequeue::Inconsistent => abort("inconsistent in drop"),
Dequeue::Data(ptr) => drop(ptr2arc(ptr)),
}
}
}
}
}
impl<U> List<U> {
fn new() -> Self {
List {
len: 0,
head: ptr::null_mut(),
tail: ptr::null_mut(),
}
}
/// Appends an element to the back of the list
fn push_back(&mut self, node: Arc<Node<U>>) -> *const Node<U> {
let ptr = arc2ptr(node);
unsafe {
// Point to the current last node in the list
*(*ptr).prev_all.get() = self.tail;
*(*ptr).next_all.get() = ptr::null_mut();
if !self.tail.is_null() {
*(*self.tail).next_all.get() = ptr;
self.tail = ptr;
} else {
// This is the first node
self.tail = ptr;
self.head = ptr;
}
}
self.len += 1;
ptr
}
/// Pop an element from the front of the list
fn pop_front(&mut self) -> Option<Arc<Node<U>>> {
if self.head.is_null() {
// The list is empty
return None;
}
self.len -= 1;
unsafe {
// Convert the ptr to Arc<_>
let node = ptr2arc(self.head);
// Update the head pointer
self.head = *node.next_all.get();
// If the pointer is null, then the list is empty
if self.head.is_null() {
self.tail = ptr::null_mut();
} else {
*(*self.head).prev_all.get() = ptr::null_mut();
}
Some(node)
}
}
/// Remove a specific node
unsafe fn remove(&mut self, node: *const Node<U>) -> Arc<Node<U>> {
let node = ptr2arc(node);
let next = *node.next_all.get();
let prev = *node.prev_all.get();
*node.next_all.get() = ptr::null_mut();
*node.prev_all.get() = ptr::null_mut();
if !next.is_null() {
*(*next).prev_all.get() = prev;
} else {
self.tail = prev;
}
if !prev.is_null() {
*(*prev).next_all.get() = next;
} else {
self.head = next;
}
self.len -= 1;
node
}
}
unsafe fn noop(_: *const ()) {}
// ===== Raw Waker Inner<U> ======
fn waker_inner<U: Unpark>(inner: Arc<Inner<U>>) -> Waker {
let ptr = Arc::into_raw(inner) as *const ();
let vtable = &RawWakerVTable::new(
clone_inner::<U>,
wake_inner::<U>,
wake_by_ref_inner::<U>,
drop_inner::<U>,
);
unsafe { Waker::from_raw(RawWaker::new(ptr, vtable)) }
}
unsafe fn clone_inner<U: Unpark>(data: *const ()) -> RawWaker {
let arc: Arc<Inner<U>> = Arc::from_raw(data as *const Inner<U>);
let clone = arc.clone();
// forget both Arcs so the refcounts don't get decremented
mem::forget(arc);
mem::forget(clone);
let vtable = &RawWakerVTable::new(
clone_inner::<U>,
wake_inner::<U>,
wake_by_ref_inner::<U>,
drop_inner::<U>,
);
RawWaker::new(data, vtable)
}
unsafe fn wake_inner<U: Unpark>(data: *const ()) {
let arc: Arc<Inner<U>> = Arc::from_raw(data as *const Inner<U>);
arc.unpark.unpark();
}
unsafe fn wake_by_ref_inner<U: Unpark>(data: *const ()) {
let arc: Arc<Inner<U>> = Arc::from_raw(data as *const Inner<U>);
arc.unpark.unpark();
// by_ref means we don't own the Node, so forget the Arc
mem::forget(arc);
}
unsafe fn drop_inner<U>(data: *const ()) {
drop(Arc::<Inner<U>>::from_raw(data as *const Inner<U>));
}
// ===== Raw Waker Node<U> ======
unsafe fn waker_ref<U: Unpark>(node: &Arc<Node<U>>) -> Waker {
let ptr = &*node as &Node<U> as *const Node<U> as *const ();
let vtable = &RawWakerVTable::new(
clone_node::<U>,
wake_unreachable,
wake_by_ref_node::<U>,
noop,
);
Waker::from_raw(RawWaker::new(ptr, vtable))
}
unsafe fn wake_unreachable(_data: *const ()) {
unreachable!("waker_ref::wake()");
}
unsafe fn clone_node<U: Unpark>(data: *const ()) -> RawWaker {
let arc: Arc<Node<U>> = Arc::from_raw(data as *const Node<U>);
let clone = arc.clone();
// forget both Arcs so the refcounts don't get decremented
mem::forget(arc);
mem::forget(clone);
let vtable = &RawWakerVTable::new(
clone_node::<U>,
wake_node::<U>,
wake_by_ref_node::<U>,
drop_node::<U>,
);
RawWaker::new(data, vtable)
}
unsafe fn wake_node<U: Unpark>(data: *const ()) {
let arc: Arc<Node<U>> = Arc::from_raw(data as *const Node<U>);
Node::<U>::notify(&arc);
}
unsafe fn wake_by_ref_node<U: Unpark>(data: *const ()) {
let arc: Arc<Node<U>> = Arc::from_raw(data as *const Node<U>);
Node::<U>::notify(&arc);
// by_ref means we don't own the Node, so forget the Arc
mem::forget(arc);
}
unsafe fn drop_node<U>(data: *const ()) {
drop(Arc::<Node<U>>::from_raw(data as *const Node<U>));
}
impl<U: Unpark> Node<U> {
fn notify(me: &Arc<Node<U>>) {
let inner = match me.queue.upgrade() {
Some(inner) => inner,
None => return,
};
// It's our job to notify the node that it's ready to get polled,
// meaning that we need to enqueue it into the readiness queue. To
// do this we flag that we're ready to be queued, and if successful
// we then do the literal queueing operation, ensuring that we're
// only queued once.
//
// Once the node is inserted we be sure to notify the parent task,
// as it'll want to come along and pick up our node now.
//
// Note that we don't change the reference count of the node here,
// we're just enqueueing the raw pointer. The `Scheduler`
// implementation guarantees that if we set the `queued` flag true that
// there's a reference count held by the main `Scheduler` queue
// still.
let prev = me.queued.swap(true, SeqCst);
if !prev {
// Get the current scheduler tick
let tick_num = inner.tick_num.load(SeqCst);
me.notified_at.store(tick_num, SeqCst);
inner.enqueue(&**me);
inner.unpark.unpark();
}
}
}
impl<U> Drop for Node<U> {
fn drop(&mut self) {
// Currently a `Node` is sent across all threads for any lifetime,
// regardless of `T`. This means that for memory safety we can't
// actually touch `T` at any time except when we have a reference to the
// `Scheduler` itself.
//
// Consequently it *should* be the case that we always drop items from
// the `Scheduler` instance, but this is a bomb in place to catch
// any bugs in that logic.
unsafe {
if (*self.item.get()).is_some() {
abort("item still here when dropping");
}
}
}
}
fn arc2ptr<T>(ptr: Arc<T>) -> *const T {
let addr = &*ptr as *const T;
mem::forget(ptr);
addr
}
unsafe fn ptr2arc<T>(ptr: *const T) -> Arc<T> {
let anchor = mem::transmute::<usize, Arc<T>>(0x10);
let addr = &*anchor as *const T;
mem::forget(anchor);
let offset = addr as isize - 0x10;
mem::transmute::<isize, Arc<T>>(ptr as isize - offset)
}
fn abort(s: &str) -> ! {
struct DoublePanic;
impl Drop for DoublePanic {
fn drop(&mut self) {
panic!("panicking twice to abort the program");
}
}
let _bomb = DoublePanic;
panic!("{}", s);
}
+3
View File
@@ -64,6 +64,9 @@ mod global;
pub mod park;
mod typed;
#[cfg(feature = "current-thread")]
pub mod current_thread;
pub use crate::enter::{enter, exit, Enter, EnterError};
pub use crate::error::SpawnError;
pub use crate::executor::Executor;