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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078
//! A concurrent work-stealing queue for building schedulers.
//!
//! # Examples
//!
//! Distribute some tasks in a thread pool:
//!
//! ```
//! use work_queue::{Queue, LocalQueue};
//!
//! struct Task(Box<dyn Fn(&mut LocalQueue<Task>) + Send>);
//!
//! let threads = 4;
//!
//! let queue: Queue<Task> = Queue::new(threads, 128);
//!
//! // Push some tasks to the queue.
//! for _ in 0..500 {
//! queue.push(Task(Box::new(|local| {
//! do_work();
//!
//! local.push(Task(Box::new(|_| do_work())));
//! local.push(Task(Box::new(|_| do_work())));
//! })));
//! }
//!
//! // Spawn threads to complete the tasks.
//! let handles: Vec<_> = queue
//! .local_queues()
//! .map(|mut local_queue| {
//! std::thread::spawn(move || {
//! while let Some(task) = local_queue.pop() {
//! task.0(&mut local_queue);
//! }
//! })
//! })
//! .collect();
//!
//! for handle in handles {
//! handle.join().unwrap();
//! }
//! # fn do_work() {}
//! ```
//!
//! # Comparison with crossbeam-deque
//!
//! This crate is similar in purpose to [`crossbeam-deque`](https://docs.rs/crossbeam-deque), which
//! also provides concurrent work-stealing queues. However there are a few notable differences:
//!
//! - This crate is more high level - work stealing is done automatically when calling `pop`
//! instead of you having to manually call it.
//! - As such, we do not support as much customization as `crossbeam-deque` - but the algorithm
//! itself can be optimized better.
//! - Queues have a fixed number of local queues that they support, and this number cannot grow.
//! - Each local queue has a fixed capacity, unlike `crossbeam-deque` which supports local queue
//! growth. This makes our local queues faster.
//!
//! # Implementation
//!
//! This crate's queue implementation is based off [Tokio's current scheduler]. The idea is that
//! each thread holds a fixed-capacity local queue, and there is also an unbounded global queue
//! accessible by all threads. In the general case each worker thread will only interact with its
//! local queue, avoiding lots of synchronization - but if one worker thread happens to have a
//! lot less work than another, it will be spread out evenly due to work stealing.
//!
//! Additionally, each local queue stores a [non-stealable LIFO slot] to optimize for message
//! passing patterns, so that if one task creates another, that created task will be polled
//! immediately, instead of only much later when it reaches the front of the local queue.
//!
//! [Tokio's current scheduler]: https://tokio.rs/blog/2019-10-scheduler
//! [non-stealable LIFO slot]: https://tokio.rs/blog/2019-10-scheduler#optimizing-for-message-passing-patterns
#![warn(missing_debug_implementations, rust_2018_idioms, missing_docs)]
use std::cmp;
use std::collections::hash_map::DefaultHasher;
use std::fmt::{self, Debug, Formatter};
use std::hash::Hasher;
use std::iter::FusedIterator;
use std::mem::{self, MaybeUninit};
use std::ops::Deref;
use std::ptr::{self, NonNull};
#[cfg(not(loom))]
use std::{collections::hash_map::RandomState, hash::BuildHasher};
#[cfg_attr(loom, path = "loom.rs")]
#[cfg_attr(not(loom), path = "std.rs")]
mod facade;
use facade::atomic::{self, AtomicBool, AtomicU16, AtomicU32, AtomicUsize};
use facade::Arc;
use facade::GlobalQueue;
use facade::UnsafeCell;
/// A work queue.
///
/// This implements [`Clone`] and so multiple handles to the queue can be easily created and
/// shared.
#[derive(Debug)]
pub struct Queue<T>(Arc<Shared<T>>);
impl<T> Queue<T> {
/// Create a new work queue.
///
/// `local_queues` is the number of [`LocalQueue`]s yielded by [`Self::local_queues`]. Typically
/// you will have a local queue for each thread on a thread pool.
///
/// `local_queue_size` is the number of items that can be stored in each local queue before it
/// overflows into the global one. You should fine-tune this to your needs.
///
/// # Panics
///
/// This will panic if the local queue size is not a power of two.
///
/// # Examples
///
/// ```
/// use work_queue::Queue;
///
/// let threads = 4;
/// let queue: Queue<i32> = Queue::new(threads, 512);
/// ```
pub fn new(local_queues: usize, local_queue_size: u16) -> Self {
assert_eq!(
local_queue_size.count_ones(),
1,
"Queue size is not a power of two"
);
let mask = local_queue_size - 1;
Self(Arc::new(Shared {
local_queues: (0..local_queues)
.map(|_| LocalQueueInner {
heads: AtomicU32::new(0),
tail: AtomicU16::new(0),
mask,
items: (0..local_queue_size)
.map(|_| UnsafeCell::new(MaybeUninit::uninit()))
.collect(),
})
.collect(),
global_queue: GlobalQueue::new(),
stealing_global: AtomicBool::new(false),
taken_local_queues: AtomicBool::new(false),
searchers: AtomicUsize::new(0),
}))
}
/// Push an item to the global queue. When one of the local queues empties, they can pick this
/// item up.
pub fn push(&self, item: T) {
let _ = self.0.global_queue.push(item);
}
/// Iterate over the local queues of this queue.
///
/// # Panics
///
/// This will panic if called more than one time.
pub fn local_queues(&self) -> LocalQueues<'_, T> {
assert!(!self
.0
.taken_local_queues
.swap(true, atomic::Ordering::Relaxed));
LocalQueues {
shared: self,
index: 0,
#[cfg(not(loom))]
hasher: RandomState::new().build_hasher(),
#[cfg(loom)]
hasher: DefaultHasher::new(),
}
}
/// Get the number of threads that are currently searching for work inside [`pop`](Self::pop).
///
/// If this number is too high, you may wish to avoid calling [`pop`](Self::pop) to reduce
/// contention.
#[must_use]
pub fn searchers(&self) -> usize {
self.0.searchers.load(atomic::Ordering::Relaxed)
}
}
impl<T> Clone for Queue<T> {
fn clone(&self) -> Self {
Self(Arc::clone(&self.0))
}
}
#[derive(Debug)]
struct Shared<T> {
local_queues: Box<[LocalQueueInner<T>]>,
global_queue: GlobalQueue<T>,
/// Whether a thread is currently stealing from the global queue. When `true`, threads
/// should avoid trying to pop from it to reduce contention.
stealing_global: AtomicBool,
/// Whether the local queues have already been yielded to the user and so shouldn't be yielded
/// again.
taken_local_queues: AtomicBool,
/// The number of queues searching for work.
searchers: AtomicUsize,
}
/// The fixed-capacity SP2C queue owned by each local queue.
struct LocalQueueInner<T> {
/// The two heads (fronts) of the queue, packed into one atomic by `pack_heads` and
/// `unpack_heads`.
///
/// The first head, the "stealer" head, always lags behind the second head, the "real" head.
/// Items are popped starting from the real head, but the space between the two heads still
/// cannot be overwritten by the tail, as it's being read by a stealer.
heads: AtomicU32,
/// The back of the queue. Only incremented by the associated queue.
tail: AtomicU16,
/// Bitmask applied to the head and tail to obtain the actual indices, so that the atomics can
/// be incremented and freely overflow outside of the range of the queue itself.
mask: u16,
/// The actual items in the queue.
items: Box<[UnsafeCell<MaybeUninit<T>>]>,
}
unsafe impl<T: Send> Sync for LocalQueueInner<T> {}
impl<T> Debug for LocalQueueInner<T> {
fn fmt(&self, f: &mut Formatter<'_>) -> fmt::Result {
let (protected_head, head) = unpack_heads(self.heads.load(atomic::Ordering::Acquire));
f.debug_struct("LocalQueueInner")
.field("protected_head", &protected_head)
.field("head", &head)
.field("tail", &self.tail)
.field("mask", &format_args!("{:#b}", self.mask))
.finish()
}
}
/// Unpack the `heads` value in a `LocalQueueInner`. Returns a tuple of the stealer head and the
/// real head.
fn unpack_heads(heads: u32) -> (u16, u16) {
((heads >> 16) as u16, heads as u16)
}
/// Pack the `heads` value in a `LocalQueueInner` from its stealer head and real head.
fn pack_heads(stealer: u16, real: u16) -> u32 {
(stealer as u32) << 16 | real as u32
}
/// One of the local queues in a [`Queue`].
///
/// You can create this using [`Queue::local_queues`].
#[derive(Debug)]
pub struct LocalQueue<T> {
/// Special slot that is always popped from first, to optimize for message passing where one
/// task is blocked on another.
lifo_slot: Option<T>,
local: ValidPtr<LocalQueueInner<T>>,
shared: Queue<T>,
/// Random number generator used to find which queue to start work stealing from.
rng: Rng,
}
impl<T> LocalQueue<T> {
/// Load the tail of the local queue.
fn local_tail(&mut self) -> u16 {
// SAFETY: The tail can be loaded without synchronization because only `self` can write to
// it, and we have an `&mut self`.
unsafe { facade::atomic_u16_unsync_load(&self.local.tail) }
}
/// Push an item to the local queue. If the local queue is full, it will move half of its items
/// to the global queue.
pub fn push(&mut self, item: T) {
if let Some(previous) = self.lifo_slot.replace(item) {
self.push_yield(previous);
}
}
/// Push an item to the local queue, skipping the LIFO slot. This can be used to give other
/// tasks a chance to run. Otherwise, there's a risk that one task will completely take over a
/// thread in a push-pop cycle due to the LIFO slot.
pub fn push_yield(&mut self, item: T) {
let tail = self.local_tail();
// We have to use Acquire to make sure that we don't write into memory that is
// currently being read by work stealers.
let mut heads = self.local.heads.load(atomic::Ordering::Acquire);
loop {
let (steal_head, head) = unpack_heads(heads);
// If the local queue is not full, we can simply push to that.
if tail.wrapping_sub(steal_head) < self.local.items.len() as u16 {
let i = tail & self.local.mask;
self.local.items[usize::from(i)]
.with_mut(|slot| unsafe { slot.write(MaybeUninit::new(item)) });
// Release is necessary to make sure the above write is ordered before accesssing
// values.
self.local
.tail
.store(tail.wrapping_add(1), atomic::Ordering::Release);
return;
}
// If no threads are currently stealing, our overflowing local queue will not be
// drained, so we should push half of it to the global queue.
//
// Otherwise (when threads are stealing) we don't want to wait for them to finish, so
// we just push this single item to the global queue (but we don't need to push any
// more since we're about to become less full).
if steal_head == head {
let half = self.local.items.len() as u16 / 2;
// TODO: We could use compare_exchange_weak here, which may potentially improve
// performance.
let res = self.local.heads.compare_exchange(
heads,
pack_heads(head.wrapping_add(half), head.wrapping_add(half)),
// Release is necessary to ensure that any previous writes to `tail` become
// visible after this point. If not made visible, we could get into a situation
// where a thread reads an older value for `tail` than `heads`, and if `heads`
// has advanced beyond the old `tail` by that point it causes all sorts of
// issues.
atomic::Ordering::AcqRel,
// Acquire is necessary because on failure we use the new value to update the
// head (see the Acquire ordering above).
atomic::Ordering::Acquire,
);
// Moving the head failed because another thread has just stolen some items. This
// means the queue is less full, so we can retry pushing to the local queue.
if let Err(new_heads) = res {
heads = new_heads;
continue;
}
// Push half the items in the current queue to the global queue.
for i in 0..half {
let index = head.wrapping_add(i) & self.local.mask;
let item = unsafe {
self.local.items[usize::from(index)]
.with(|slot| slot.read())
.assume_init()
};
let _ = self.shared.0.global_queue.push(item);
}
}
let _ = self.shared.0.global_queue.push(item);
return;
}
}
/// Pop an item from the local queue, or steal from the global and sibling queues if it is
/// empty.
pub fn pop(&mut self) -> Option<T> {
// First try to pop from the LIFO slot.
if let Some(item) = self.lifo_slot.take() {
return Some(item);
}
let tail = self.local_tail();
// First try to pop from the local queue.
let res = atomic_u32_fetch_update(
&self.local.heads,
// No memory orderings are necessary here as this is the only thread that mutates
// the data, and it's not currently mutating the data.
atomic::Ordering::Relaxed,
atomic::Ordering::Relaxed,
|heads| {
let (steal_head, head) = unpack_heads(heads);
if head == tail {
None
} else if steal_head == head {
// There are no current stealers; update both heads.
Some(pack_heads(head.wrapping_add(1), head.wrapping_add(1)))
} else {
// There is currently a stealer; only update the real head, as it's the
// stealer's job to update the stealer head later.
Some(pack_heads(steal_head, head.wrapping_add(1)))
}
},
);
let heads = match res {
// We have successfully popped something from the local queue.
Ok(heads) => {
let (_, head) = unpack_heads(heads);
let i = head & self.local.mask;
return Some(unsafe {
self.local.items[usize::from(i)]
.with(|ptr| ptr.read())
.assume_init()
});
}
// The local queue is empty.
Err(heads) => heads,
};
let (steal_head, head) = unpack_heads(heads);
assert_eq!(head, tail);
// The number of free slots in the queue we can steal into.
let space = self.local.items.len() as u16 - head.wrapping_sub(steal_head);
// Now we will try to steal into this queue from various places.
// TODO: Potentially throttle stealing?
self.shared
.0
.searchers
// No ordering is necessary because we use this as a hint, not for safety.
.fetch_add(1, atomic::Ordering::Relaxed);
struct DecrementSearchers<'a>(&'a AtomicUsize);
impl Drop for DecrementSearchers<'_> {
fn drop(&mut self) {
self.0.fetch_sub(1, atomic::Ordering::Relaxed);
}
}
let _decrement_searchers = DecrementSearchers(&self.shared.0.searchers);
// If there are no threads currently stealing from the global queue, we will steal from it.
if self
.shared
.0
.stealing_global
.compare_exchange(
false,
true,
// No ordering is necessary because we use this as a hint, not for safety.
atomic::Ordering::Relaxed,
atomic::Ordering::Relaxed,
)
.is_ok()
{
let popped_item = self.shared.0.global_queue.pop();
if popped_item.is_some() {
// To avoid having to search for items again after we have completed this one, we
// fill half of our queue with items from the global queue.
let steal = cmp::min(self.local.items.len() as u16 / 2, space);
let mut tail = head;
let end_tail = head.wrapping_add(steal);
// Ensure that the following mutations of the local queue of items will not occur
// before stealers have finished reading.
u32_acquire_fence(&self.local.heads);
while tail != end_tail {
match self.shared.0.global_queue.pop() {
Some(item) => {
let i = tail & self.local.mask;
self.local.items[usize::from(i)]
.with_mut(|slot| unsafe { slot.write(MaybeUninit::new(item)) });
}
None => break,
}
tail = tail.wrapping_add(1);
}
// Release is necessary to make sure the above write is ordered before accesssing
// values.
self.local.tail.store(tail, atomic::Ordering::Release);
}
self.shared
.0
.stealing_global
.store(false, atomic::Ordering::Relaxed);
if let Some(popped_item) = popped_item {
return Some(popped_item);
}
}
// Steal work from sibling queues starting from a random location.
let queues = self.shared.0.local_queues.len();
let start = self.rng.gen_usize_to(queues);
'sibling_queues: for i in 0..queues {
let mut i = start + i;
if i >= queues {
i -= queues;
}
let queue = &self.shared.0.local_queues[i];
if ptr::eq(queue, &*self.local) {
continue;
}
// Acquire is necessary to make sure that the below load of the tail does not occur
// before the load of the head.
let mut queue_heads = queue.heads.load(atomic::Ordering::Acquire);
let (old_queue_head, mut queue_head, steal) = loop {
let (queue_steal_head, queue_head) = unpack_heads(queue_heads);
// If another thread is already stealing from this queue, don't steal from it.
if queue_steal_head != queue_head {
continue 'sibling_queues;
}
// Acquire is necessary so we don't read into items that are currently being
// written by the thread itself.
let queue_tail = queue.tail.load(atomic::Ordering::Acquire);
// The number of items that can be stolen.
let stealable = queue_tail.wrapping_sub(queue_head);
// The number of items we actually want to steal - this is half of their queue,
// rounded up.
let steal = cmp::min(stealable - stealable / 2, space);
if steal == 0 {
continue 'sibling_queues;
}
let new_queue_head = queue_head.wrapping_add(steal);
// TODO: We could use compare_exchange here, which may potentially improve
// performance.
let res = queue.heads.compare_exchange_weak(
queue_heads,
// Only move the real head, as we still need to keep the steal head to read
// from the queue.
pack_heads(queue_head, new_queue_head),
// Release isn't necessary here since the above code doesn't access any memory.
atomic::Ordering::Acquire,
// Acquire is necessary when the compare_exchange fails because the result is
// used to update the values in `queue_heads`; see the Acquire above.
atomic::Ordering::Acquire,
);
match res {
Ok(_) => break (queue_head, new_queue_head, steal),
Err(updated_queue_heads) => queue_heads = updated_queue_heads,
}
};
assert_ne!(steal, 0);
// Ensure that the following mutations of the local queue of items will not occur
// before stealers of our own queue have finished reading.
u32_acquire_fence(&self.local.heads);
// Read the first item separately, as we will be returning it.
let first_item = unsafe {
queue.items[usize::from(old_queue_head & queue.mask)]
.with(|slot| slot.read())
.assume_init()
};
// Copy over the stolen items to our queue.
for i in 1..steal {
let src = &queue.items[usize::from(old_queue_head.wrapping_add(i) & queue.mask)];
let dst =
&self.local.items[usize::from(head.wrapping_add(i - 1) & self.local.mask)];
src.with(|src| dst.with_mut(|dst| unsafe { src.copy_to_nonoverlapping(dst, 1) }))
}
// Update the steal head to match the real head.
loop {
let res = queue.heads.compare_exchange_weak(
pack_heads(old_queue_head, queue_head),
pack_heads(queue_head, queue_head),
// Release is necessary to make sure the above reads are ordered before any
// other thread can write to the values.
atomic::Ordering::Release,
// No ordering is necessary because we're not accessing shared mutable state
// after this point.
atomic::Ordering::Relaxed,
);
match res {
Ok(_) => break,
Err(updated_queue_heads) => {
let (updated_queue_steal_head, update_queue_head) =
unpack_heads(updated_queue_heads);
assert_eq!(updated_queue_steal_head, old_queue_head);
queue_head = update_queue_head;
}
}
}
if steal > 1 {
// Release is necessary to make sure the above writes are ordered before accessing
// values.
self.local
.tail
.store(tail.wrapping_add(steal - 1), atomic::Ordering::Release);
}
return Some(first_item);
}
// Lastly, pop from the global queue without guarding against contention, since there is
// nowhere else we can currently get items from.
self.shared.0.global_queue.pop()
}
/// Get the number of threads that are currently searching for work inside [`pop`](Self::pop).
///
/// If this number is too high, you may wish to avoid calling [`pop`](Self::pop) to reduce
/// contention.
#[must_use]
pub fn searchers(&self) -> usize {
self.shared.searchers()
}
/// Get the global queue that is associated with this local queue.
#[must_use]
pub fn global(&self) -> &Queue<T> {
&self.shared
}
}
/// An iterator over the [`LocalQueue`]s in a [`Queue`]. Created by [`Queue::local_queues`].
#[derive(Debug)]
#[must_use = "iterators are lazy and do nothing unless consumed"]
pub struct LocalQueues<'a, T> {
shared: &'a Queue<T>,
index: usize,
hasher: DefaultHasher,
}
impl<T> Iterator for LocalQueues<'_, T> {
type Item = LocalQueue<T>;
fn next(&mut self) -> Option<Self::Item> {
let inner = self.shared.0.local_queues.get(self.index)?;
self.index += 1;
Some(LocalQueue {
lifo_slot: None,
// SAFETY: The `LocalQueue` stores an `Arc` so this pointer is guaranteed to be valid
// until the type is dropped.
local: unsafe { ValidPtr::new(inner) },
shared: self.shared.clone(),
rng: Rng {
state: {
self.hasher.write_usize(self.index);
self.hasher.finish()
},
},
})
}
fn size_hint(&self) -> (usize, Option<usize>) {
let len = self.len();
(len, Some(len))
}
}
impl<T> ExactSizeIterator for LocalQueues<'_, T> {
fn len(&self) -> usize {
self.shared.0.local_queues.len() - self.index
}
}
impl<T> FusedIterator for LocalQueues<'_, T> {}
/// A `*const T` that is guaranteed to always be valid and non-null.
struct ValidPtr<T: ?Sized>(NonNull<T>);
impl<T: ?Sized> ValidPtr<T> {
unsafe fn new(ptr: *const T) -> Self {
Self(NonNull::new_unchecked(ptr as *mut T))
}
}
impl<T: ?Sized> Clone for ValidPtr<T> {
fn clone(&self) -> Self {
*self
}
}
impl<T: ?Sized> Copy for ValidPtr<T> {}
impl<T: ?Sized> Deref for ValidPtr<T> {
type Target = T;
fn deref(&self) -> &Self::Target {
unsafe { self.0.as_ref() }
}
}
impl<T: ?Sized + Debug> Debug for ValidPtr<T> {
fn fmt(&self, f: &mut Formatter<'_>) -> fmt::Result {
T::fmt(self, f)
}
}
unsafe impl<T: ?Sized + Sync> Send for ValidPtr<T> {}
unsafe impl<T: ?Sized + Sync> Sync for ValidPtr<T> {}
#[cfg(target_pointer_width = "64")]
type DoubleUsize = u128;
#[cfg(target_pointer_width = "32")]
type DoubleUsize = u64;
/// Wyrand RNG.
#[derive(Debug)]
struct Rng {
state: u64,
}
impl Rng {
fn gen_u64(&mut self) -> u64 {
self.state = self.state.wrapping_add(0xA0761D6478BD642F);
let t = u128::from(self.state) * u128::from(self.state ^ 0xE7037ED1A0B428DB);
(t >> 64) as u64 ^ t as u64
}
fn gen_usize(&mut self) -> usize {
self.gen_u64() as usize
}
fn gen_usize_to(&mut self, to: usize) -> usize {
// Adapted from https://www.pcg-random.org/posts/bounded-rands.html
const USIZE_BITS: usize = mem::size_of::<usize>() * 8;
let mut x = self.gen_usize();
let mut m = ((x as DoubleUsize * to as DoubleUsize) >> USIZE_BITS) as usize;
let mut l = x.wrapping_mul(to);
if l < to {
let t = to.wrapping_neg() % to;
while l < t {
x = self.gen_usize();
m = ((x as DoubleUsize * to as DoubleUsize) >> USIZE_BITS) as usize;
l = x.wrapping_mul(to);
}
}
m
}
}
fn atomic_u32_fetch_update<F>(
atomic: &AtomicU32,
set_order: atomic::Ordering,
fetch_order: atomic::Ordering,
mut f: F,
) -> Result<u32, u32>
where
F: FnMut(u32) -> Option<u32>,
{
let mut prev = atomic.load(fetch_order);
while let Some(next) = f(prev) {
match atomic.compare_exchange_weak(prev, next, set_order, fetch_order) {
Ok(x) => return Ok(x),
Err(next_prev) => prev = next_prev,
}
}
Err(prev)
}
fn u32_acquire_fence(atomic: &AtomicU32) {
if cfg!(tsan) {
// ThreadSanitizer doesn't support fences.
atomic.load(atomic::Ordering::Acquire);
} else {
atomic::fence(atomic::Ordering::Acquire);
}
}
#[cfg(all(test, not(loom)))]
mod tests {
use super::*;
use std::collections::HashSet;
#[test]
fn rng() {
let mut rng = Rng { state: 3493858 };
let mut remaining: HashSet<_> = (0..15).collect();
while !remaining.is_empty() {
let value = rng.gen_usize_to(15);
assert!(value < 15, "{} is not less than 15!", value);
remaining.remove(&value);
}
}
#[test]
fn lifo_slot() {
let queue = Queue::new(1, 2);
let mut local = queue.local_queues().next().unwrap();
assert_eq!(local.pop(), None);
assert_eq!(local.pop(), None);
local.push(Box::new(5));
assert_eq!(local.pop(), Some(Box::new(5)));
assert_eq!(local.pop(), None);
}
#[test]
fn push_many() {
let queue = Queue::new(1, 2);
let mut local = queue.local_queues().next().unwrap();
for i in 0..4 {
local.push(Box::new(i));
}
assert_eq!(local.pop(), Some(Box::new(3)));
assert_eq!(local.pop(), Some(Box::new(1)));
assert_eq!(local.pop(), Some(Box::new(0)));
assert_eq!(local.pop(), Some(Box::new(2)));
assert_eq!(local.pop(), None);
}
#[test]
fn wrapping() {
let queue = Queue::new(1, 2);
let mut local = queue.local_queues().next().unwrap();
local.push_yield(Box::new(0));
for i in 0..10 {
local.push_yield(Box::new(i + 1));
assert_eq!(local.pop(), Some(Box::new(i)));
}
assert_eq!(local.pop(), Some(Box::new(10)));
assert_eq!(local.pop(), None);
assert_eq!(local.pop(), None);
}
#[test]
fn steal_global() {
for &size in &[2, 4, 8, 16, 32, 64] {
let queue = Queue::new(4, size);
for i in 0..16 {
queue.push(Box::new(i));
}
let mut local = queue.local_queues().next().unwrap();
for i in 0..16 {
assert_eq!(local.pop(), Some(Box::new(i)));
}
assert_eq!(local.pop(), None);
}
}
#[test]
fn steal_siblings() {
let queue = Queue::new(2, 64);
let mut locals: Vec<_> = queue.local_queues().collect();
locals[0].push_yield(Box::new(4));
locals[0].push_yield(Box::new(5));
locals[1].push(Box::new(1));
locals[1].push(Box::new(0));
queue.push(Box::new(2));
queue.push(Box::new(3));
for i in 0..6 {
assert_eq!(locals[1].pop(), Some(Box::new(i)));
}
}
#[test]
fn many_locals() {
let queue = <Queue<()>>::new(10, 128);
queue.local_queues().for_each(drop);
}
#[test]
fn searchers() {
let queue = Queue::new(2, 64);
let mut locals = queue.local_queues();
let mut local_a = locals.next().unwrap();
let mut local_b = locals.next().unwrap();
assert_eq!(local_a.searchers(), 0);
assert_eq!(local_b.searchers(), 0);
local_a.push(());
local_a.push(());
local_a.pop().unwrap();
local_a.pop().unwrap();
queue.push(());
local_b.pop().unwrap();
assert!(local_b.pop().is_none());
assert_eq!(local_a.searchers(), 0);
assert_eq!(local_b.searchers(), 0);
// This test hangs on Miri.
if cfg!(not(miri)) {
let stop = Arc::new(AtomicBool::new(false));
let handle = std::thread::spawn({
let stop = Arc::clone(&stop);
move || {
while !stop.load(atomic::Ordering::Relaxed) {
local_b.pop();
}
}
});
loop {
let searchers = local_a.searchers();
assert!(searchers < 2);
if searchers == 1 {
break;
}
}
stop.store(true, atomic::Ordering::Relaxed);
handle.join().unwrap();
}
}
#[test]
fn stress() {
let queue = Queue::new(4, 4);
if cfg!(miri) {
for _ in 0..3 {
queue.push(4);
}
} else {
for _ in 0..32 {
queue.push(6);
}
}
let threads: Vec<_> = queue
.local_queues()
.map(|mut queue| {
std::thread::spawn(move || {
while let Some(num) = queue.pop() {
for _ in 0..num {
queue.push(num - 1);
}
}
})
})
.collect();
for thread in threads {
thread.join().unwrap();
}
}
#[test]
fn cobb() {
use std::cell::UnsafeCell;
struct State(Option<Box<[UnsafeCell<LocalQueue<Box<i32>>>]>>);
unsafe impl Sync for State {}
cobb::run_test(cobb::TestCfg {
threads: 4,
iterations: if cfg!(miri) { 100 } else { 1000 },
sub_iterations: if cfg!(miri) { 1 } else { 10 },
setup: || {
let queue = Queue::new(4, 4);
State(Some(
queue
.local_queues()
.map(UnsafeCell::new)
.collect::<Box<[_]>>(),
))
},
test: |State(state), tctx| {
let local_queues = state.as_ref().unwrap();
let queue = unsafe { &mut *local_queues[tctx.thread_index()].get() };
if tctx.thread_index() < 2 {
queue.push(Box::new(5));
} else {
queue.pop();
}
},
teardown: |state| *state = State(None),
..Default::default()
});
}
}
#[cfg(all(test, loom))]
mod loom_tests {
use super::*;
fn locals<T, const N: usize>(queue: &Queue<T>) -> [LocalQueue<T>; N] {
array_init::from_iter(queue.local_queues()).expect("incorrect number of local queues")
}
#[test]
fn pop_none() {
loom::model(|| {
let queue: Queue<()> = Queue::new(2, 1);
let [mut local_1, mut local_2] = locals(&queue);
loom::thread::spawn(move || assert!(local_1.pop().is_none()));
assert!(local_2.pop().is_none());
});
}
#[test]
fn concurrent_steal_global() {
loom::model(|| {
let queue: Queue<Box<i32>> = Queue::new(2, 1);
let [mut local_1, mut local_2] = locals(&queue);
for i in 0..2 {
queue.push(Box::new(i));
}
loom::thread::spawn(move || {
local_1.pop();
local_1.pop();
});
local_2.pop();
});
}
#[test]
fn concurrent_steal_sibling() {
loom::model(|| {
let queue: Queue<Box<i32>> = Queue::new(3, 1);
let [mut local_1, mut local_2, mut local_3] = locals(&queue);
for i in 0..4 {
local_1.push(Box::new(i));
}
loom::thread::spawn(move || {
local_2.pop();
local_2.pop();
});
local_3.pop();
});
}
#[test]
fn global_spsc() {
loom::model(|| {
let queue: Queue<Box<i32>> = Queue::new(1, 4);
let [mut local] = locals(&queue);
loom::thread::spawn(move || {
for i in 0..6 {
queue.push(Box::new(i));
}
});
for _ in 0..6 {
local.pop();
}
});
}
#[test]
fn sibling_spsc_few() {
loom::model(|| {
let queue: Queue<Box<i32>> = Queue::new(2, 4);
let [mut local_1, mut local_2] = locals(&queue);
loom::thread::spawn(move || {
for i in 0..4 {
local_1.push(Box::new(i));
}
});
for _ in 0..4 {
local_2.pop();
}
});
}
#[test]
fn sibling_spsc_many() {
loom::model(|| {
let queue: Queue<Box<i32>> = Queue::new(2, 4);
let [mut local_1, mut local_2] = locals(&queue);
loom::thread::spawn(move || {
for i in 0..8 {
local_1.push(Box::new(i));
}
});
local_2.pop();
});
}
}