1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417
//! A queue for moving data from one future/task into another.
//!
//! This is a "single-producer, single-consumer" queue that splits into separate
//! `Push` and `Pop` endpoints -- at any given time, there is at most one of
//! each alive in your program, ensuring that pushes and pops are not coming
//! from multiple directions.
//!
//! You create a queue by calling `Queue::new` and passing it a reference to its
//! backing storage. To actually use the queue, however, you must call
//! `Queue::split` to break it into two endpoints, `Push` and `Pop`. As their
//! names suggest, `Push` can be used to push things into the queue, and `Pop`
//! can be used to pop things out of it.
//!
//! The `Push` and `Pop` can then be handed off to separate code paths, so long
//! as they don't outlive the `Queue` and its storage. (The compiler will ensure
//! this.)
//!
//! This queue uses the Rust type system to ensure that only one code path is
//! attempting to push or pop the queue at any given time, because both `Push`
//! and `Pop` endpoints borrow the central `Queue`, and an `async` operation to
//! push or pop through an endpoint borrows that endpoint in turn. Neither
//! `Push` nor `Pop` can be cloned or copied, ensuring the single-producer
//! single-consumer property.
//!
//! # Implementation
//!
//! This is implemented as a concurrent lock-free Lamport queue. This has two
//! implications:
//!
//! 1. If you can arrange the lifetimes correctly (i.e. make the queue static)
//! it is actually safe to operate either Push or Pop from an ISR.
//! 2. It fills up at N-1 elements because one empty slot is used as a sentinel
//! to distinguish full from empty.
//!
//! The adaptations to modern memory ordering semantics are taken from Nhat Minh
//! Lê et al's paper "Correct and Efficient Bounded FIFO Queues," though this
//! implementation does not use _all_ of the optimizations they identified.
use core::cell::UnsafeCell;
use core::mem::MaybeUninit;
use core::sync::atomic::{AtomicUsize, Ordering};
use crate::exec::Notify;
/// A single-producer, single-consumer queue. The `Queue` struct contains the
/// controlling information for the queue overall, and _borrows_ the storage.
///
/// See the module docs for details.
#[derive(Debug)]
pub struct Queue<'s, T> {
storage: &'s mut [UnsafeCell<MaybeUninit<T>>],
/// Index of next slot in `storage` to write during `push`. Must fall in the
/// range `0..N`. Read by both sides, written by pushers.
head: AtomicUsize,
/// Index of next slot in `storage` to read during `pop`. Must fall in the
/// range `0..N`. Read by both sides, written by poppers.
tail: AtomicUsize,
/// Signals blocked pushers that an element has been popped.
popped: Notify,
/// Signals blocked poppers that an element has been pushed.
pushed: Notify,
}
/// This type is easily sharable across threads, because there are no useful
/// operations that can be performed using only a shared reference.
unsafe impl<T> Sync for Queue<'_, T> where T: Send {}
impl<'s, T> Queue<'s, T> {
/// Creates a queue, borrowing the uninitialized `storage` (which will be
/// arbitrarily overwritten).
pub fn new(storage: &'s mut [MaybeUninit<T>]) -> Self {
// Safety: the cast we're about to do is memory-layout-compatible
// because:
// - MaybeUninit<T> has the same memory layout as T
// - UnsafeCell<T> has the same memory layout as T
// - Thus, UnsafeCell<MaybeUninit<T>> has the same memory layout
// as MaybeUninit<T>.
//
// We can do these shenanigans because we have exclusive access to the
// memory backing `storage`, and the caller thinks of it as
// `MaybeUninit`, meaning they aren't making assumptions about its
// contents or dropping it when we're done.
let storage: *mut [MaybeUninit<T>] = storage;
let storage: *mut [UnsafeCell<MaybeUninit<T>>] = storage as *mut _;
let storage: &'s mut [UnsafeCell<MaybeUninit<T>>] = unsafe {
&mut *storage
};
Self {
storage,
head: AtomicUsize::new(0),
tail: AtomicUsize::new(0),
pushed: Notify::new(),
popped: Notify::new(),
}
}
/// Creates a push and pop endpoint for this queue. Note that an exclusive
/// borrow of the queue exists as long as either endpoint exists, ensuring
/// that at most one of each endpoint exists at any point in the program.
///
/// You can, however, drop the first pair of endpoints and make a new pair
/// later -- that's fine.
pub fn split(&mut self) -> (Push<'_, T>, Pop<'_, T>) {
(
Push { q: self, _marker: crate::NotSyncMarker::default() },
Pop { q: self, _marker: crate::NotSyncMarker::default() },
)
}
fn next_index(&self, i: usize) -> usize {
// This produced better code than using remainder on ARMv7-M last
// I checked.
let ni = i.wrapping_add(1);
if ni == self.storage.len() { 0 } else { ni }
}
}
/// It's entirely possible to drop a non-empty Queue in correct code, unlike
/// (say) a `lilos::list`, so we provide a Drop impl that goes through and
/// cleans up queued elements.
impl<T> Drop for Queue<'_, T> {
fn drop(&mut self) {
let h = self.head.load(Ordering::SeqCst);
let mut t = self.head.load(Ordering::SeqCst);
while h != t {
let unsafecell_ptr = self.storage[t].get();
// Safety: this is unsafe because we're accessing the UnsafeCell
// contents. We can do this because since we're &mut Self we
// have exclusive control over our storage.
let maybeuninit = unsafe { &mut *unsafecell_ptr };
// Safety: we're dropping the contents of a MaybeUninit, which
// implies asserting that it's been initialized. We know it's been
// initialized because of the `h != t` condition on our loop.
unsafe {
maybeuninit.assume_init_drop();
}
t = self.next_index(t);
}
}
}
/// Queue endpoint for pushing data. Access to a `Push` _only_ gives you the
/// right to push data and enquire about push-related properties.
///
/// See the module docs for more details.
#[derive(Debug)]
pub struct Push<'a, T> {
q: &'a Queue<'a, T>,
_marker: crate::NotSyncMarker,
}
impl<'q, T> Push<'q, T> {
// Implementation note: Push "owns" the head and does not need to be careful
// with its memory ordering, while the tail is "foreign" and must be
// synchronized.
/// Checks if there is room to push at least one item. Because the `Push`
/// endpoint has exclusive control over data movement into the queue, if
/// this returns `true`, the condition will remain true until a `push` or
/// `try_push` happens through `self`, or `self` is dropped.
///
/// If this returns `false`, of course, room may appear at any time if the
/// other end of the queue is popped.
pub fn can_push(&self) -> bool {
let h = self.q.head.load(Ordering::Relaxed);
let t = self.q.tail.load(Ordering::Acquire);
self.q.next_index(h) != t
}
/// Checks if there is room to push at least one item, and if so, returns a
/// `Permit` that entitles its holder to that queue slot.
///
/// If the queue is full, returns `None`.
///
/// This operation is slightly less useful than [`Push::reserve`] because
/// `try_push` is slightly simpler to use, and doesn't risk data loss.
/// (Whereas `reserve` is specifically designed to fix a class of data loss
/// bugs.)
pub fn try_reserve(&mut self) -> Option<Permit<'q, T>> {
let h = self.q.head.load(Ordering::Relaxed);
let t = self.q.tail.load(Ordering::Acquire);
let h_next = self.q.next_index(h);
if h_next == t {
// We're full.
return None;
}
let unsafecell_ptr = self.q.storage[h].get();
// Safety: this is dereferencing a raw pointer into the unsafecell,
// which we can do because (1) the cell being between h and t implies
// that it is not aliased, and (2) because we have &mut Self we know
// we're not racing any other pushes for this slot. (Pops won't touch
// this slot.)
let maybeuninit = unsafe { &mut *unsafecell_ptr };
Some(Permit {
maybeuninit,
head: &self.q.head,
h_next,
pushed: &self.q.pushed,
})
}
/// Produces a future that resolves when there is enough space in the queue
/// to push one element. It resolves into a [`Permit`], which entitles the
/// holder to pushing an element without needing to check or `await`. This
/// is a deliberate design choice -- it means you can cancel the future
/// without losing the element you were trying to push.
///
/// The returned `Permit` borrows `self` exclusively. This means you must
/// use the `Permit`, or drop it, before you can request another. This
/// prevents a deadlock, where you wait for a second permit that will never
/// emerge.
///
/// The future produced by `reserve` also borrows `self` exclusively. This
/// means you can't simultaneously have two futures waiting for permits from
/// the same `Push`. This wouldn't necessarily be a bad thing, but we need
/// to maintain the exclusive borrow in order to pass it through to the
/// `Permit`.
///
/// # Cancellation
///
/// **Cancel Safety:** Strict.
///
/// This does basically nothing if cancelled (it is intrinsically
/// cancel-safe).
pub async fn reserve<'s>(&'s mut self) -> Permit<'s, T> {
self.q.popped.until(|| self.try_reserve()).await
}
/// Attempts to stuff `value` into the queue.
///
/// If there is space, ownership of `value` moves into the queue and this
/// returns `Ok(())`.
///
/// If there is not space, this returns `Err(value)` -- that is, ownership
/// of `value` is handed back to you.
pub fn try_push(&mut self, value: T) -> Result<(), T> {
let h = self.q.head.load(Ordering::Relaxed);
let t = self.q.tail.load(Ordering::Acquire);
let h_next = self.q.next_index(h);
if h_next == t {
// We're full.
return Err(value);
}
let unsafecell_ptr = self.q.storage[h].get();
// Safety: this is dereferencing a raw pointer into the unsafecell,
// which we can do because (1) the cell being between h and t implies
// that it is not aliased, and (2) because we have &mut Self we know
// we're not racing any other pushes for this slot. (Pops won't touch
// this slot.)
let maybeuninit = unsafe { &mut *unsafecell_ptr };
maybeuninit.write(value);
// We can store instead of compare-exchange here because we are the only
// Push manipulating this field (see: &mut Self).
self.q.head.store(h_next, Ordering::Release);
self.q.pushed.notify();
Ok(())
}
/// Produces a future that resolves when `value` has been pushed into the
/// queue, and not before. This implies that at least one free slot must
/// open in the queue -- either by never having been filled, or by being
/// freed up by a pop -- for the future to resolve.
///
/// Note that the future maintains an exclusive borrow over `self` until
/// that happens -- so just as there can only be one `Push` endpoint for a
/// queue at any given time, there can only be one outstanding `push` future
/// for that endpoint. This means we don't have to define the order of
/// competing pushes, moving concerns about fairness to compile time.
///
/// # Cancellation
///
/// **Cancel Safety:** Weak.
///
/// The future returned by `push` takes ownership of `value` immediately. If
/// the future is dropped before `value` makes it into the queue, `value` is
/// dropped along with it. While this behavior is well-defined and
/// predictable (thus my "weak cancel-safe" label) it is probably not what
/// you want, and so this operation is currently deprecated. Please use
/// [`Push::reserve`] instead.
#[deprecated = "please use Push::reserve to avoid cancellation bugs"]
pub async fn push(&mut self, value: T) {
let mut value = Some(value);
self.q.popped.until(move || {
match self.try_push(value.take().unwrap()) {
Ok(()) => true,
Err(revalue) => {
value = Some(revalue);
false
}
}
}).await
}
}
/// Queue endpoint for popping data. Access to a `Pop` _only_ gives you the
/// right to pop data and enquire about pop-related properties.
///
/// See the module docs for more details.
#[derive(Debug)]
pub struct Pop<'a, T> {
q: &'a Queue<'a, T>,
_marker: crate::NotSyncMarker,
}
impl<T> Pop<'_, T> {
// Implementation note: Pop "owns" the tail and does not need to be careful
// with its memory ordering, while the head is "foreign" and must be
// synchronized.
/// Checks if there is at least one item available to pop from the queue.
///
/// Because the `Pop` endpoint has exclusive control over data movement out
/// of the queue, if this returns `true`, the condition will remain true
/// until a `pop` or `try_pop` happens through `self`, or `self` is dropped.
///
/// If this returns `false`, of course, new items may appear at any time if
/// the other end of the queue is pushed.
pub fn can_pop(&self) -> bool {
let t = self.q.tail.load(Ordering::Relaxed);
let h = self.q.head.load(Ordering::Acquire);
h != t
}
/// Pops an element out of the queue, if the queue is not empty.
///
/// If the queue is empty, returns `None`.
pub fn try_pop(&mut self) -> Option<T> {
let t = self.q.tail.load(Ordering::Relaxed);
let h = self.q.head.load(Ordering::Acquire);
if h == t {
// We're empty.
return None;
}
let t_next = self.q.next_index(t);
let unsafecell_ptr = self.q.storage[t].get();
// Safety: we're dereferencing the raw pointer into the UnsafeCell,
// which we can do because (1) this cell is between t and h, so it's not
// aliased by any pushing, and (2) we have a &mut Self, so it's also by
// definition not aliased by any popping.
let maybeuninit = unsafe { &mut *unsafecell_ptr };
// Safety: we're reading the possibly-uninitialized contents of the
// MaybeUninit, which we can do because the cell is between t and h, and
// thus has been initialized by a previous push. We will bump tail just
// below to switch the cell's state to uninitialized after reading.
let result = unsafe { maybeuninit.assume_init_read() };
self.q.tail.store(t_next, Ordering::Release);
self.q.popped.notify();
Some(result)
}
/// Produces a future that resolves to the next element that can be popped
/// from the queue. If the queue is not empty, this will happen immediately
/// on the first poll of the future; otherwise, it will happen when the
/// future is polled after data has been pushed into the queue.
///
/// Note that the future maintains an exclusive borrow over `self` until
/// that happens -- so just as there can only be one `Pop` endpoint for a
/// queue at any given time, there can only be one outstanding `pop` future
/// for that endpoint. This means we don't have to define the order of
/// competing pops, moving concerns about fairness to compile time.
///
/// # Cancellation
///
/// **Cancel Safety:** Strict.
///
/// The future returned by this function has no side effects until it
/// resolves to a popped element. If you drop it before it has resolved,
/// no data is lost.
pub async fn pop(&mut self) -> T {
self.q.pushed.until(move || self.try_pop()).await
}
}
/// A "permit" giving its holder the right to push one element into a queue,
/// without waiting.
///
/// This is produced by [`Push::try_reserve`]/[`Push::reserve`] when they find
/// space available in the queue. The key property here is that producing a
/// `Permit` does not, by itself, change the queue state -- you can turn around
/// and `Drop` the `Permit` without losing data or pushing anything. This makes
/// cancel safety much easier to reason about.
#[derive(Debug)]
pub struct Permit<'q, T> {
maybeuninit: &'q mut MaybeUninit<T>,
head: &'q AtomicUsize,
pushed: &'q Notify,
h_next: usize,
}
impl<T> Permit<'_, T> {
/// Pushes `value` to the queue, consuming this `Permit`.
///
/// The `push` operation is guaranteed to succeed synchronously, since
/// holding a `Permit` is proof that the next cell at the head of the queue
/// is available for use.
pub fn push(self, value: T) {
self.maybeuninit.write(value);
self.head.store(self.h_next, Ordering::Release);
self.pushed.notify();
}
}