tokio/runtime/scheduler/multi_thread/queue.rs
1//! Run-queue structures to support a work-stealing scheduler
2
3use crate::loom::cell::UnsafeCell;
4use crate::loom::sync::Arc;
5use crate::runtime::scheduler::multi_thread::{Overflow, Stats};
6use crate::runtime::task;
7
8use std::mem::{self, MaybeUninit};
9use std::ptr;
10use std::sync::atomic::Ordering::{AcqRel, Acquire, Relaxed, Release};
11
12// Use wider integers when possible to increase ABA resilience.
13//
14// See issue #5041: <https://github.com/tokio-rs/tokio/issues/5041>.
15cfg_has_atomic_u64! {
16 type UnsignedShort = u32;
17 type UnsignedLong = u64;
18 type AtomicUnsignedShort = crate::loom::sync::atomic::AtomicU32;
19 type AtomicUnsignedLong = crate::loom::sync::atomic::AtomicU64;
20}
21cfg_not_has_atomic_u64! {
22 type UnsignedShort = u16;
23 type UnsignedLong = u32;
24 type AtomicUnsignedShort = crate::loom::sync::atomic::AtomicU16;
25 type AtomicUnsignedLong = crate::loom::sync::atomic::AtomicU32;
26}
27
28/// Producer handle. May only be used from a single thread.
29pub(crate) struct Local<T: 'static> {
30 inner: Arc<Inner<T>>,
31}
32
33/// Consumer handle. May be used from many threads.
34pub(crate) struct Steal<T: 'static>(Arc<Inner<T>>);
35
36pub(crate) struct Inner<T: 'static> {
37 /// Concurrently updated by many threads.
38 ///
39 /// Contains two `UnsignedShort` values. The `LSB` byte is the "real" head of
40 /// the queue. The `UnsignedShort` in the `MSB` is set by a stealer in process
41 /// of stealing values. It represents the first value being stolen in the
42 /// batch. The `UnsignedShort` indices are intentionally wider than strictly
43 /// required for buffer indexing in order to provide ABA mitigation and make
44 /// it possible to distinguish between full and empty buffers.
45 ///
46 /// When both `UnsignedShort` values are the same, there is no active
47 /// stealer.
48 ///
49 /// Tracking an in-progress stealer prevents a wrapping scenario.
50 head: AtomicUnsignedLong,
51
52 /// Only updated by producer thread but read by many threads.
53 tail: AtomicUnsignedShort,
54
55 /// Elements
56 buffer: Box<[UnsafeCell<MaybeUninit<task::Notified<T>>>; LOCAL_QUEUE_CAPACITY]>,
57}
58
59unsafe impl<T> Send for Inner<T> {}
60unsafe impl<T> Sync for Inner<T> {}
61
62#[cfg(not(loom))]
63const LOCAL_QUEUE_CAPACITY: usize = 256;
64
65// Shrink the size of the local queue when using loom. This shouldn't impact
66// logic, but allows loom to test more edge cases in a reasonable a mount of
67// time.
68#[cfg(loom)]
69const LOCAL_QUEUE_CAPACITY: usize = 4;
70
71const MASK: usize = LOCAL_QUEUE_CAPACITY - 1;
72
73// Constructing the fixed size array directly is very awkward. The only way to
74// do it is to repeat `UnsafeCell::new(MaybeUninit::uninit())` 256 times, as
75// the contents are not Copy. The trick with defining a const doesn't work for
76// generic types.
77fn make_fixed_size<T>(buffer: Box<[T]>) -> Box<[T; LOCAL_QUEUE_CAPACITY]> {
78 assert_eq!(buffer.len(), LOCAL_QUEUE_CAPACITY);
79
80 // safety: We check that the length is correct.
81 unsafe { Box::from_raw(Box::into_raw(buffer).cast()) }
82}
83
84/// Create a new local run-queue
85pub(crate) fn local<T: 'static>() -> (Steal<T>, Local<T>) {
86 let buffer = std::iter::repeat_with(|| UnsafeCell::new(MaybeUninit::uninit()));
87
88 let inner = Arc::new(Inner {
89 head: AtomicUnsignedLong::new(0),
90 tail: AtomicUnsignedShort::new(0),
91 buffer: make_fixed_size(buffer.take(LOCAL_QUEUE_CAPACITY).collect()),
92 });
93
94 let local = Local {
95 inner: inner.clone(),
96 };
97
98 let remote = Steal(inner);
99
100 (remote, local)
101}
102
103impl<T> Local<T> {
104 /// Returns the number of entries in the queue
105 pub(crate) fn len(&self) -> usize {
106 let (_, head) = unpack(self.inner.head.load(Acquire));
107 // safety: this is the **only** thread that updates this cell.
108 let tail = unsafe { self.inner.tail.unsync_load() };
109 len(head, tail)
110 }
111
112 /// How many tasks can be pushed into the queue
113 pub(crate) fn remaining_slots(&self) -> usize {
114 let (steal, _) = unpack(self.inner.head.load(Acquire));
115 // safety: this is the **only** thread that updates this cell.
116 let tail = unsafe { self.inner.tail.unsync_load() };
117
118 LOCAL_QUEUE_CAPACITY - len(steal, tail)
119 }
120
121 pub(crate) fn max_capacity(&self) -> usize {
122 LOCAL_QUEUE_CAPACITY
123 }
124
125 /// Returns false if there are any entries in the queue
126 ///
127 /// Separate to `is_stealable` so that refactors of `is_stealable` to "protect"
128 /// some tasks from stealing won't affect this
129 pub(crate) fn has_tasks(&self) -> bool {
130 self.len() != 0
131 }
132
133 /// Pushes a batch of tasks to the back of the queue. All tasks must fit in
134 /// the local queue.
135 ///
136 /// # Panics
137 ///
138 /// The method panics if there is not enough capacity to fit in the queue.
139 pub(crate) fn push_back(&mut self, tasks: impl ExactSizeIterator<Item = task::Notified<T>>) {
140 let len = tasks.len();
141 assert!(len <= LOCAL_QUEUE_CAPACITY);
142
143 if len == 0 {
144 // Nothing to do
145 return;
146 }
147
148 let head = self.inner.head.load(Acquire);
149 let (steal, _) = unpack(head);
150
151 // safety: this is the **only** thread that updates this cell.
152 let mut tail = unsafe { self.inner.tail.unsync_load() };
153
154 if tail.wrapping_sub(steal) <= (LOCAL_QUEUE_CAPACITY - len) as UnsignedShort {
155 // Yes, this if condition is structured a bit weird (first block
156 // does nothing, second returns an error). It is this way to match
157 // `push_back_or_overflow`.
158 } else {
159 panic!()
160 }
161
162 for task in tasks {
163 let idx = tail as usize & MASK;
164
165 self.inner.buffer[idx].with_mut(|ptr| {
166 // Write the task to the slot
167 //
168 // Safety: There is only one producer and the above `if`
169 // condition ensures we don't touch a cell if there is a
170 // value, thus no consumer.
171 unsafe {
172 ptr::write((*ptr).as_mut_ptr(), task);
173 }
174 });
175
176 tail = tail.wrapping_add(1);
177 }
178
179 self.inner.tail.store(tail, Release);
180 }
181
182 /// Pushes a task to the back of the local queue, if there is not enough
183 /// capacity in the queue, this triggers the overflow operation.
184 ///
185 /// When the queue overflows, half of the current contents of the queue is
186 /// moved to the given Injection queue. This frees up capacity for more
187 /// tasks to be pushed into the local queue.
188 pub(crate) fn push_back_or_overflow<O: Overflow<T>>(
189 &mut self,
190 mut task: task::Notified<T>,
191 overflow: &O,
192 stats: &mut Stats,
193 ) {
194 let tail = loop {
195 let head = self.inner.head.load(Acquire);
196 let (steal, real) = unpack(head);
197
198 // safety: this is the **only** thread that updates this cell.
199 let tail = unsafe { self.inner.tail.unsync_load() };
200
201 if tail.wrapping_sub(steal) < LOCAL_QUEUE_CAPACITY as UnsignedShort {
202 // There is capacity for the task
203 break tail;
204 } else if steal != real {
205 // Concurrently stealing, this will free up capacity, so only
206 // push the task onto the inject queue
207 overflow.push(task);
208 return;
209 } else {
210 // Push the current task and half of the queue into the
211 // inject queue.
212 match self.push_overflow(task, real, tail, overflow, stats) {
213 Ok(_) => return,
214 // Lost the race, try again
215 Err(v) => {
216 task = v;
217 }
218 }
219 }
220 };
221
222 self.push_back_finish(task, tail);
223 }
224
225 // Second half of `push_back`
226 fn push_back_finish(&self, task: task::Notified<T>, tail: UnsignedShort) {
227 // Map the position to a slot index.
228 let idx = tail as usize & MASK;
229
230 self.inner.buffer[idx].with_mut(|ptr| {
231 // Write the task to the slot
232 //
233 // Safety: There is only one producer and the above `if`
234 // condition ensures we don't touch a cell if there is a
235 // value, thus no consumer.
236 unsafe {
237 ptr::write((*ptr).as_mut_ptr(), task);
238 }
239 });
240
241 // Make the task available. Synchronizes with a load in
242 // `steal_into2`.
243 self.inner.tail.store(tail.wrapping_add(1), Release);
244 }
245
246 /// Moves a batch of tasks into the inject queue.
247 ///
248 /// This will temporarily make some of the tasks unavailable to stealers.
249 /// Once `push_overflow` is done, a notification is sent out, so if other
250 /// workers "missed" some of the tasks during a steal, they will get
251 /// another opportunity.
252 #[inline(never)]
253 fn push_overflow<O: Overflow<T>>(
254 &mut self,
255 task: task::Notified<T>,
256 head: UnsignedShort,
257 tail: UnsignedShort,
258 overflow: &O,
259 stats: &mut Stats,
260 ) -> Result<(), task::Notified<T>> {
261 /// How many elements are we taking from the local queue.
262 ///
263 /// This is one less than the number of tasks pushed to the inject
264 /// queue as we are also inserting the `task` argument.
265 const NUM_TASKS_TAKEN: UnsignedShort = (LOCAL_QUEUE_CAPACITY / 2) as UnsignedShort;
266
267 assert_eq!(
268 tail.wrapping_sub(head) as usize,
269 LOCAL_QUEUE_CAPACITY,
270 "queue is not full; tail = {tail}; head = {head}"
271 );
272
273 // Claim all tasks.
274 //
275 // We are claiming the tasks **before** reading them out of the buffer.
276 // This is safe because only the **current** thread is able to push new
277 // tasks.
278 //
279 // There isn't really any need for memory ordering... Relaxed would
280 // work. This is because all tasks are pushed into the queue from the
281 // current thread (or memory has been acquired if the local queue handle
282 // moved).
283 if self
284 .inner
285 .head
286 .compare_exchange_weak(pack(head, head), pack(tail, tail), Release, Relaxed)
287 .is_err()
288 {
289 // We failed to claim the tasks, losing the race. Return out of
290 // this function and try the full `push` routine again. The queue
291 // may not be full anymore.
292 return Err(task);
293 }
294
295 // Add back the first half of tasks.
296 //
297 // We are doing it this way instead of just taking half of the tasks because we want the
298 // *second* half of the tasks, and if you just incremented `head` by `NUM_TASKS_TAKEN`,
299 // then you would be taking the first half instead of the second half.
300 //
301 // Pushing the second half of the local queue to the injection queue is better because when
302 // we take tasks *out* of the injection queue, we always place them in the first half. This
303 // means that if a task is in the second half, then we know for sure that this task is not
304 // a task we just got from the injection queue. This ensures that when we take a task out
305 // of the injection queue, then it will not be moved back into the injection queue (at
306 // least not until after we have polled it at least once).
307 //
308 // Note that if a concurrent worker tries to steal from us between these two operations and
309 // sees that the worker queue is empty, then that worker may go to sleep, and we do not
310 // notify it about these tasks becoming available for stealing again. Ordinarily this would
311 // be a problem, but it isn't in this case because the worker will be notified about the
312 // tasks we are adding to the injection queue instead, which ensures that the stealer wakes
313 // up again to take the tasks from the injection queue.
314 self.inner
315 .tail
316 .store(tail.wrapping_add(NUM_TASKS_TAKEN), Release);
317
318 /// An iterator that takes elements out of the run queue.
319 struct BatchTaskIter<'a, T: 'static> {
320 buffer: &'a [UnsafeCell<MaybeUninit<task::Notified<T>>>; LOCAL_QUEUE_CAPACITY],
321 head: UnsignedLong,
322 i: UnsignedLong,
323 }
324 impl<'a, T: 'static> Iterator for BatchTaskIter<'a, T> {
325 type Item = task::Notified<T>;
326
327 #[inline]
328 fn next(&mut self) -> Option<task::Notified<T>> {
329 if self.i == UnsignedLong::from(NUM_TASKS_TAKEN) {
330 None
331 } else {
332 let i_idx = self.i.wrapping_add(self.head) as usize & MASK;
333 let slot = &self.buffer[i_idx];
334
335 // safety: Our CAS from before has assumed exclusive ownership
336 // of the task pointers in this range.
337 let task = slot.with(|ptr| unsafe { ptr::read((*ptr).as_ptr()) });
338
339 self.i += 1;
340 Some(task)
341 }
342 }
343 }
344
345 // safety: The CAS above ensures that no consumer will look at these
346 // values again, and we are the only producer.
347 let batch_iter = BatchTaskIter {
348 buffer: &self.inner.buffer,
349 head: head.wrapping_add(NUM_TASKS_TAKEN) as UnsignedLong,
350 i: 0,
351 };
352 overflow.push_batch(batch_iter.chain(std::iter::once(task)));
353
354 // Add 1 to factor in the task currently being scheduled.
355 stats.incr_overflow_count();
356
357 Ok(())
358 }
359
360 /// Pops a task from the local queue.
361 pub(crate) fn pop(&mut self) -> Option<task::Notified<T>> {
362 let mut head = self.inner.head.load(Acquire);
363
364 let idx = loop {
365 let (steal, real) = unpack(head);
366
367 // safety: this is the **only** thread that updates this cell.
368 let tail = unsafe { self.inner.tail.unsync_load() };
369
370 if real == tail {
371 // queue is empty
372 return None;
373 }
374
375 let next_real = real.wrapping_add(1);
376
377 // If `steal == real` there are no concurrent stealers. Both `steal`
378 // and `real` are updated.
379 let next = if steal == real {
380 pack(next_real, next_real)
381 } else {
382 assert_ne!(steal, next_real);
383 pack(steal, next_real)
384 };
385
386 // Attempt to claim a task.
387 let res = self
388 .inner
389 .head
390 .compare_exchange_weak(head, next, AcqRel, Acquire);
391
392 match res {
393 Ok(_) => break real as usize & MASK,
394 Err(actual) => head = actual,
395 }
396 };
397
398 Some(self.inner.buffer[idx].with(|ptr| unsafe { ptr::read(ptr).assume_init() }))
399 }
400}
401
402impl<T> Steal<T> {
403 /// Returns the number of entries in the queue
404 pub(crate) fn len(&self) -> usize {
405 let (_, head) = unpack(self.0.head.load(Acquire));
406 let tail = self.0.tail.load(Acquire);
407 len(head, tail)
408 }
409
410 /// Return true if the queue is empty,
411 /// false if there are any entries in the queue
412 pub(crate) fn is_empty(&self) -> bool {
413 self.len() == 0
414 }
415
416 /// Steals half the tasks from self and place them into `dst`.
417 pub(crate) fn steal_into(
418 &self,
419 dst: &mut Local<T>,
420 dst_stats: &mut Stats,
421 ) -> Option<task::Notified<T>> {
422 // Safety: the caller is the only thread that mutates `dst.tail` and
423 // holds a mutable reference.
424 let dst_tail = unsafe { dst.inner.tail.unsync_load() };
425
426 // To the caller, `dst` may **look** empty but still have values
427 // contained in the buffer. If another thread is concurrently stealing
428 // from `dst` there may not be enough capacity to steal.
429 let (steal, _) = unpack(dst.inner.head.load(Acquire));
430
431 if dst_tail.wrapping_sub(steal) > LOCAL_QUEUE_CAPACITY as UnsignedShort / 2 {
432 // we *could* try to steal less here, but for simplicity, we're just
433 // going to abort.
434 return None;
435 }
436
437 // Steal the tasks into `dst`'s buffer. This does not yet expose the
438 // tasks in `dst`.
439 let mut n = self.steal_into2(dst, dst_tail);
440
441 if n == 0 {
442 // No tasks were stolen
443 return None;
444 }
445
446 dst_stats.incr_steal_count(n as u16);
447 dst_stats.incr_steal_operations();
448
449 // We are returning a task here
450 n -= 1;
451
452 let ret_pos = dst_tail.wrapping_add(n);
453 let ret_idx = ret_pos as usize & MASK;
454
455 // safety: the value was written as part of `steal_into2` and not
456 // exposed to stealers, so no other thread can access it.
457 let ret = dst.inner.buffer[ret_idx].with(|ptr| unsafe { ptr::read((*ptr).as_ptr()) });
458
459 if n == 0 {
460 // The `dst` queue is empty, but a single task was stolen
461 return Some(ret);
462 }
463
464 // Make the stolen items available to consumers
465 dst.inner.tail.store(dst_tail.wrapping_add(n), Release);
466
467 Some(ret)
468 }
469
470 // Steal tasks from `self`, placing them into `dst`. Returns the number of
471 // tasks that were stolen.
472 fn steal_into2(&self, dst: &mut Local<T>, dst_tail: UnsignedShort) -> UnsignedShort {
473 let mut prev_packed = self.0.head.load(Acquire);
474 let mut next_packed;
475
476 let n = loop {
477 let (src_head_steal, src_head_real) = unpack(prev_packed);
478 let src_tail = self.0.tail.load(Acquire);
479
480 // If these two do not match, another thread is concurrently
481 // stealing from the queue.
482 if src_head_steal != src_head_real {
483 return 0;
484 }
485
486 // Number of available tasks to steal
487 let n = src_tail.wrapping_sub(src_head_real);
488 let n = n - n / 2;
489
490 if n == 0 {
491 // No tasks available to steal
492 return 0;
493 }
494
495 // Update the real head index to acquire the tasks.
496 let steal_to = src_head_real.wrapping_add(n);
497 assert_ne!(src_head_steal, steal_to);
498 next_packed = pack(src_head_steal, steal_to);
499
500 // Claim all those tasks. This is done by incrementing the "real"
501 // head but not the steal. By doing this, no other thread is able to
502 // steal from this queue until the current thread completes.
503 let res = self
504 .0
505 .head
506 .compare_exchange_weak(prev_packed, next_packed, AcqRel, Acquire);
507
508 match res {
509 Ok(_) => break n,
510 Err(actual) => prev_packed = actual,
511 }
512 };
513
514 assert!(
515 n <= LOCAL_QUEUE_CAPACITY as UnsignedShort / 2,
516 "actual = {n}"
517 );
518
519 let (first, _) = unpack(next_packed);
520
521 // Take all the tasks
522 for i in 0..n {
523 // Compute the positions
524 let src_pos = first.wrapping_add(i);
525 let dst_pos = dst_tail.wrapping_add(i);
526
527 // Map to slots
528 let src_idx = src_pos as usize & MASK;
529 let dst_idx = dst_pos as usize & MASK;
530
531 // Read the task
532 //
533 // safety: We acquired the task with the atomic exchange above.
534 let task = self.0.buffer[src_idx].with(|ptr| unsafe { ptr::read((*ptr).as_ptr()) });
535
536 // Write the task to the new slot
537 //
538 // safety: `dst` queue is empty and we are the only producer to
539 // this queue.
540 dst.inner.buffer[dst_idx]
541 .with_mut(|ptr| unsafe { ptr::write((*ptr).as_mut_ptr(), task) });
542 }
543
544 let mut prev_packed = next_packed;
545
546 // Update `src_head_steal` to match `src_head_real` signalling that the
547 // stealing routine is complete.
548 loop {
549 let head = unpack(prev_packed).1;
550 next_packed = pack(head, head);
551
552 let res = self
553 .0
554 .head
555 .compare_exchange_weak(prev_packed, next_packed, AcqRel, Acquire);
556
557 match res {
558 Ok(_) => return n,
559 Err(actual) => prev_packed = actual,
560 }
561 }
562 }
563}
564
565impl<T> Clone for Steal<T> {
566 fn clone(&self) -> Steal<T> {
567 Steal(self.0.clone())
568 }
569}
570
571impl<T> Drop for Local<T> {
572 fn drop(&mut self) {
573 if !std::thread::panicking() {
574 assert!(self.pop().is_none(), "queue not empty");
575 }
576 }
577}
578
579/// Calculate the length of the queue using the head and tail.
580/// The `head` can be the `steal` or `real` head.
581fn len(head: UnsignedShort, tail: UnsignedShort) -> usize {
582 tail.wrapping_sub(head) as usize
583}
584
585/// Split the head value into the real head and the index a stealer is working
586/// on.
587fn unpack(n: UnsignedLong) -> (UnsignedShort, UnsignedShort) {
588 let real = n & UnsignedShort::MAX as UnsignedLong;
589 let steal = n >> (mem::size_of::<UnsignedShort>() * 8);
590
591 (steal as UnsignedShort, real as UnsignedShort)
592}
593
594/// Join the two head values
595fn pack(steal: UnsignedShort, real: UnsignedShort) -> UnsignedLong {
596 (real as UnsignedLong) | ((steal as UnsignedLong) << (mem::size_of::<UnsignedShort>() * 8))
597}
598
599#[test]
600fn test_local_queue_capacity() {
601 assert!(LOCAL_QUEUE_CAPACITY - 1 <= u8::MAX as usize);
602}