1
0
forked from ROMEO/nexosim
nexosim/asynchronix/src/util/task_set.rs
2024-04-29 21:00:12 +02:00

454 lines
18 KiB
Rust

//! Primitive for the efficient management of concurrent tasks.
use std::sync::atomic::Ordering;
use std::sync::Arc;
use diatomic_waker::WakeSource;
use futures_task::{waker_ref, ArcWake, WakerRef};
use crate::loom_exports::sync::atomic::{AtomicU32, AtomicU64};
/// Special value for the `next` field of a task, indicating that the task to
/// which this field belongs is not currently in the list of scheduled tasks.
const SLEEPING: u32 = u32::MAX;
/// Special value for a task index, indicating the absence of task.
const EMPTY: u32 = u32::MAX - 1;
/// Mask for the index of the task pointed to by the head of the list of
/// scheduled tasks.
const INDEX_MASK: u64 = u32::MAX as u64;
/// Mask for the scheduling countdown in the head of the list of scheduled
/// tasks.
const COUNTDOWN_MASK: u64 = !INDEX_MASK;
/// A single increment of the scheduling countdown in the head of the list of
/// scheduled tasks.
const COUNTDOWN_ONE: u64 = 1 << 32;
/// A primitive that simplifies the management of a set of tasks scheduled
/// concurrently.
///
/// A `TaskSet` maintains both a vector-based list of tasks (or more accurately,
/// task waker handles) and a linked list of the subset of tasks that are
/// currently scheduled. The latter is stored in a vector-based Treiber stack
/// which links tasks through indices rather than pointers. Using indices has
/// two advantages: (i) it makes a fully safe implementation possible and (ii)
/// it can take advantage of a single CAS to simultaneously move the head and
/// decrement the outstanding amount of tasks to be scheduled before the parent
/// task is notified.
///
/// This can be used to implement primitives similar to `FuturesOrdered` or
/// `FuturesUnordered` in the `futures` crate.
///
/// The `notify_count` argument of `TaskSet::take_scheduled()` can be set to
/// more than 1 to wake the parent task less frequently. For instance, if
/// `notify_count` is set to the number of pending sub-tasks, the parent task
/// will only be woken once all subtasks have been woken.
pub(crate) struct TaskSet {
/// Set of all tasks, scheduled or not.
///
/// In some cases, the use of `resize()` to shrink the task set may leave
/// inactive tasks at the back of the vector, in which case the length of
/// the vector will exceed `task_count`.
tasks: Vec<Arc<Task>>,
/// Shared Treiber stack head and parent task notifier.
shared: Arc<Shared>,
/// Count of all tasks, scheduled or not.
task_count: usize,
}
impl TaskSet {
/// Creates an initially empty set of tasks associated to the parent task
/// which notifier is provided.
#[allow(clippy::assertions_on_constants)]
pub(crate) fn new(notifier: WakeSource) -> Self {
// Only 32-bit targets and above are supported.
assert!(usize::BITS >= u32::BITS);
Self {
tasks: Vec::new(),
shared: Arc::new(Shared {
head: AtomicU64::new(EMPTY as u64),
notifier,
}),
task_count: 0,
}
}
/// Creates a set of `len` tasks associated to the parent task which
/// notifier is provided.
#[allow(clippy::assertions_on_constants)]
pub(crate) fn with_len(notifier: WakeSource, len: usize) -> Self {
// Only 32-bit targets and above are supported.
assert!(usize::BITS >= u32::BITS);
assert!(len <= EMPTY as usize && len <= SLEEPING as usize);
let len = len as u32;
let shared = Arc::new(Shared {
head: AtomicU64::new(EMPTY as u64),
notifier,
});
let tasks: Vec<_> = (0..len)
.map(|idx| {
Arc::new(Task {
idx,
shared: shared.clone(),
next: AtomicU32::new(SLEEPING),
})
})
.collect();
Self {
tasks,
shared,
task_count: len as usize,
}
}
/// Take all scheduled tasks and returns an iterator over their indices, or
/// if there are no currently scheduled tasks returns `None` and requests a
/// notification to be sent after `notify_count` tasks have been scheduled.
///
/// In all cases, the list of scheduled tasks will be empty right after this
/// call.
///
/// If there were scheduled tasks, no notification is requested because this
/// method is expected to be called repeatedly until it returns `None`.
/// Failure to do so will result in missed notifications.
///
/// If no tasks were scheduled, the notification is guaranteed to be
/// triggered no later than after `notify_count` tasks have been scheduled,
/// though it may in some cases be triggered earlier. If the specified
/// `notify_count` is zero then no notification is requested.
pub(crate) fn take_scheduled(&self, notify_count: usize) -> Option<TaskIterator<'_>> {
let countdown = u32::try_from(notify_count).unwrap();
let mut head = self.shared.head.load(Ordering::Relaxed);
loop {
let new_head = if head & INDEX_MASK == EMPTY as u64 {
(countdown as u64 * COUNTDOWN_ONE) | EMPTY as u64
} else {
EMPTY as u64
};
// Ordering: this Acquire operation synchronizes with all Release
// operations in `Task::wake_by_ref` and ensures that all memory
// operations performed during and before the tasks were scheduled
// become visible.
match self.shared.head.compare_exchange_weak(
head,
new_head,
Ordering::Acquire,
Ordering::Relaxed,
) {
Ok(_) => break,
Err(h) => head = h,
}
}
let index = (head & INDEX_MASK) as u32;
if index == EMPTY {
None
} else {
Some(TaskIterator {
task_list: self,
next_index: index,
})
}
}
/// Discards all scheduled tasks and cancels any request for notification
/// that may be set.
///
/// This method is very cheap if there are no scheduled tasks and if no
/// notification is currently requested.
///
/// All discarded tasks are put in the sleeping (unscheduled) state.
pub(crate) fn discard_scheduled(&self) {
if self.shared.head.load(Ordering::Relaxed) != EMPTY as u64 {
// Dropping the iterator ensures that all tasks are put in the
// sleeping state.
let _ = self.take_scheduled(0);
}
}
/// Set the number of active tasks.
///
/// Note that this method may discard already scheduled tasks.
///
/// # Panic
///
/// This method will panic if `len` is greater than `u32::MAX - 1`.
pub(crate) fn resize(&mut self, len: usize) {
assert!(len <= EMPTY as usize && len <= SLEEPING as usize);
self.task_count = len;
// Add new tasks if necessary.
if len >= self.tasks.len() {
while len > self.tasks.len() {
let idx = self.tasks.len() as u32;
self.tasks.push(Arc::new(Task {
idx,
shared: self.shared.clone(),
next: AtomicU32::new(SLEEPING),
}));
}
return;
}
// Try to shrink the vector of tasks.
//
// The main issue when shrinking the vector of tasks is that stale
// wakers may still be around and may at any moment be scheduled and
// insert their task index in the list of scheduled tasks. If it cannot
// be guaranteed that this will not happen, then the vector of tasks
// cannot be shrunk further, otherwise the iterator for scheduled tasks
// will later fail when reaching a task with an invalid index.
//
// We follow a 2-steps strategy:
//
// 1) remove all tasks currently in the list of scheduled task and set
// them to `SLEEPING` state in case some of them might have an index
// that will be invalidated when the vector of tasks is shrunk;
//
// 2) attempt to iteratively shrink the vector of tasks by removing
// tasks starting from the back of the vector:
// - If a task is in the `SLEEPING` state, then its `next` pointer is
// changed to an arbitrary value other than`SLEEPING`, but the task
// is not inserted in the list of scheduled tasks; this way, the
// task will be effectively rendered inactive. The task can now be
// removed from the vector.
// - If a task is found in a non-`SLEEPING` state (meaning that there
// was a race and the task was scheduled after step 1) then abandon
// further shrinking and leave this task in the vector; the iterator
// for scheduled tasks mitigates such situation by only yielding
// task indices that are within the expected range.
// Step 1: unscheduled tasks that may be scheduled.
self.discard_scheduled();
// Step 2: attempt to remove tasks starting at the back of the vector.
while self.tasks.len() > len {
// There is at least one task since `len()` was non-zero.
let task = self.tasks.last().unwrap();
// Ordering: Relaxed ordering is sufficient since the task is
// effectively discarded.
if task
.next
.compare_exchange(SLEEPING, EMPTY, Ordering::Relaxed, Ordering::Relaxed)
.is_err()
{
// The task could not be removed for now so the set of tasks cannot
// be shrunk further.
break;
}
self.tasks.pop();
}
}
/// Returns `true` if one or more sub-tasks are currently scheduled.
pub(crate) fn has_scheduled(&self) -> bool {
// Ordering: the content of the head is only used as an advisory flag so
// Relaxed ordering is sufficient.
self.shared.head.load(Ordering::Relaxed) & INDEX_MASK != EMPTY as u64
}
/// Returns a reference to the waker associated to the active task with the
/// specified index.
///
/// # Panics
///
/// This method will panic if there is no active task with the provided
/// index.
pub(crate) fn waker_of(&self, idx: usize) -> WakerRef {
assert!(idx < self.task_count);
waker_ref(&self.tasks[idx])
}
pub(crate) fn len(&self) -> usize {
self.task_count
}
}
/// Internals shared between a `TaskSet` and its associated `Task`s.
struct Shared {
/// Head of the Treiber stack for scheduled tasks.
///
/// The lower 32 bits specify the index of the last scheduled task (the
/// actual head), if any, whereas the upper 32 bits specify the countdown of
/// tasks still to be scheduled before the parent task is notified.
head: AtomicU64,
/// A notifier used to wake the parent task.
notifier: WakeSource,
}
/// An asynchronous task associated with the future of a sender.
struct Task {
/// Index of this task.
idx: u32,
/// Index of the next task in the list of scheduled tasks.
next: AtomicU32,
/// Head of the list of scheduled tasks.
shared: Arc<Shared>,
}
impl ArcWake for Task {
fn wake_by_ref(arc_self: &Arc<Self>) {
let mut next = arc_self.next.load(Ordering::Relaxed);
let mut head = loop {
if next == SLEEPING {
// The task appears not to be scheduled yet: prepare its
// insertion in the list of scheduled tasks by setting the next
// task index to the index of the task currently pointed by the
// head.
//
// Ordering: Relaxed ordering is sufficient since the upcoming
// CAS on the head already ensure that all memory operations
// that precede this call to `wake_by_ref` become visible when
// the tasks are stolen.
let head = arc_self.shared.head.load(Ordering::Relaxed);
match arc_self.next.compare_exchange_weak(
SLEEPING,
(head & INDEX_MASK) as u32,
Ordering::Relaxed,
Ordering::Relaxed,
) {
Ok(_) => break head,
Err(n) => next = n,
}
} else {
// The task appears to be already scheduled: confirm this and
// establish proper memory synchronization by performing a no-op
// RMW.
//
// Ordering: the Release ordering synchronizes with the Acquire
// swap operation in `TaskIterator::next` and ensures that all
// memory operations that precede this call to `wake_by_ref`
// will be visible when the task index is yielded.
match arc_self.next.compare_exchange_weak(
next,
next,
Ordering::Release,
Ordering::Relaxed,
) {
Ok(_) => return,
Err(n) => next = n,
}
}
};
// The index to the next task has been set to the index in the head.
// Other concurrent calls to `wake` or `wake_by_ref` will now see the
// task as scheduled so this thread is responsible for moving the head.
loop {
// Attempt a CAS which decrements the countdown if it is not already
// cleared and which sets the head's index to this task's index.
let countdown = head & COUNTDOWN_MASK;
let new_countdown = countdown.wrapping_sub((countdown != 0) as u64 * COUNTDOWN_ONE);
let new_head = new_countdown | arc_self.idx as u64;
// Ordering: this Release operation synchronizes with the Acquire
// operation on the head in `TaskSet::steal_scheduled` and ensures
// that the value of the `next` field as well as all memory
// operations that precede this call to `wake_by_ref` become visible
// when the tasks are stolen.
match arc_self.shared.head.compare_exchange_weak(
head,
new_head,
Ordering::Release,
Ordering::Relaxed,
) {
Ok(_) => {
// If the countdown has just been cleared, it is necessary
// to send a notification.
if countdown == COUNTDOWN_ONE {
arc_self.shared.notifier.notify();
}
return;
}
Err(h) => {
head = h;
// Update the index of the next task to the new value of the
// head.
//
// Why use a swap instead of a simple store? This is to
// maintain a release sequence which includes previous
// atomic operation on this field, and more specifically any
// no-op CAS that could have been performed by a concurrent
// call to wake. This ensures in turn that all memory
// operations that precede a no-op CAS will be visible when
// `next` is Acquired in `TaskIterator::next`.
//
// Ordering: Relaxed ordering is sufficient since
// synchronization is ensured by the upcoming CAS on the
// head.
arc_self
.next
.swap((head & INDEX_MASK) as u32, Ordering::Relaxed);
}
}
}
}
}
/// An iterator over scheduled tasks.
pub(crate) struct TaskIterator<'a> {
task_list: &'a TaskSet,
next_index: u32,
}
impl<'a> Iterator for TaskIterator<'a> {
type Item = usize;
fn next(&mut self) -> Option<Self::Item> {
while self.next_index != EMPTY {
let index = self.next_index as usize;
// Ordering: the Acquire ordering synchronizes with any no-op CAS
// that could have been performed in `Task::wake_by_ref`, ensuring
// that all memory operations that precede such call to
// `Task::wake_by_ref` become visible.
self.next_index = self.task_list.tasks[index]
.next
.swap(SLEEPING, Ordering::Acquire);
// Only yield the index if the task is indeed active.
if index < self.task_list.task_count {
return Some(index);
}
}
None
}
}
impl<'a> Drop for TaskIterator<'a> {
fn drop(&mut self) {
// Put all remaining scheduled tasks in the sleeping state.
//
// Ordering: the task is ignored so it is not necessary to ensure that
// memory operations performed before the task was scheduled are
// visible. For the same reason, it is not necessary to synchronize with
// no-op CAS operations in `Task::wake_by_ref`, which is why separate
// load and store operations are used rather than a more expensive swap
// operation.
while self.next_index != EMPTY {
let index = self.next_index as usize;
self.next_index = self.task_list.tasks[index].next.load(Ordering::Relaxed);
self.task_list.tasks[index]
.next
.store(SLEEPING, Ordering::Relaxed);
}
}
}