rayon_core/thread_pool/
mod.rs

1//! Contains support for user-managed thread pools, represented by the
2//! the [`ThreadPool`] type (see that struct for details).
3//!
4//! [`ThreadPool`]: struct.ThreadPool.html
5
6use crate::broadcast::{self, BroadcastContext};
7use crate::join;
8use crate::registry::{Registry, ThreadSpawn, WorkerThread};
9use crate::scope::{do_in_place_scope, do_in_place_scope_fifo};
10use crate::spawn;
11use crate::{scope, Scope};
12use crate::{scope_fifo, ScopeFifo};
13use crate::{ThreadPoolBuildError, ThreadPoolBuilder};
14use std::error::Error;
15use std::fmt;
16use std::sync::Arc;
17
18mod test;
19
20/// Represents a user created [thread-pool].
21///
22/// Use a [`ThreadPoolBuilder`] to specify the number and/or names of threads
23/// in the pool. After calling [`ThreadPoolBuilder::build()`], you can then
24/// execute functions explicitly within this [`ThreadPool`] using
25/// [`ThreadPool::install()`]. By contrast, top level rayon functions
26/// (like `join()`) will execute implicitly within the current thread-pool.
27///
28///
29/// ## Creating a ThreadPool
30///
31/// ```rust
32/// # use rayon_core as rayon;
33/// let pool = rayon::ThreadPoolBuilder::new().num_threads(8).build().unwrap();
34/// ```
35///
36/// [`install()`][`ThreadPool::install()`] executes a closure in one of the `ThreadPool`'s
37/// threads. In addition, any other rayon operations called inside of `install()` will also
38/// execute in the context of the `ThreadPool`.
39///
40/// When the `ThreadPool` is dropped, that's a signal for the threads it manages to terminate,
41/// they will complete executing any remaining work that you have spawned, and automatically
42/// terminate.
43///
44///
45/// [thread-pool]: https://en.wikipedia.org/wiki/Thread_pool
46/// [`ThreadPool`]: struct.ThreadPool.html
47/// [`ThreadPool::new()`]: struct.ThreadPool.html#method.new
48/// [`ThreadPoolBuilder`]: struct.ThreadPoolBuilder.html
49/// [`ThreadPoolBuilder::build()`]: struct.ThreadPoolBuilder.html#method.build
50/// [`ThreadPool::install()`]: struct.ThreadPool.html#method.install
51pub struct ThreadPool {
52    registry: Arc<Registry>,
53}
54
55impl ThreadPool {
56    #[deprecated(note = "Use `ThreadPoolBuilder::build`")]
57    #[allow(deprecated)]
58    /// Deprecated in favor of `ThreadPoolBuilder::build`.
59    pub fn new(configuration: crate::Configuration) -> Result<ThreadPool, Box<dyn Error>> {
60        Self::build(configuration.into_builder()).map_err(Box::from)
61    }
62
63    pub(super) fn build<S>(
64        builder: ThreadPoolBuilder<S>,
65    ) -> Result<ThreadPool, ThreadPoolBuildError>
66    where
67        S: ThreadSpawn,
68    {
69        let registry = Registry::new(builder)?;
70        Ok(ThreadPool { registry })
71    }
72
73    /// Executes `op` within the threadpool. Any attempts to use
74    /// `join`, `scope`, or parallel iterators will then operate
75    /// within that threadpool.
76    ///
77    /// # Warning: thread-local data
78    ///
79    /// Because `op` is executing within the Rayon thread-pool,
80    /// thread-local data from the current thread will not be
81    /// accessible.
82    ///
83    /// # Warning: execution order
84    ///
85    /// If the current thread is part of a different thread pool, it will try to
86    /// keep busy while the `op` completes in its target pool, similar to
87    /// calling [`ThreadPool::yield_now()`] in a loop. Therefore, it may
88    /// potentially schedule other tasks to run on the current thread in the
89    /// meantime. For example
90    ///
91    /// ```rust
92    /// # use rayon_core as rayon;
93    /// fn main() {
94    ///     rayon::ThreadPoolBuilder::new().num_threads(1).build_global().unwrap();
95    ///     let pool = rayon_core::ThreadPoolBuilder::default().build().unwrap();
96    ///     let do_it = || {
97    ///         print!("one ");
98    ///         pool.install(||{});
99    ///         print!("two ");
100    ///     };
101    ///     rayon::join(|| do_it(), || do_it());
102    /// }
103    /// ```
104    ///
105    /// Since we configured just one thread in the global pool, one might
106    /// expect `do_it()` to run sequentially, producing:
107    ///
108    /// ```ascii
109    /// one two one two
110    /// ```
111    ///
112    /// However each call to `install()` yields implicitly, allowing rayon to
113    /// run multiple instances of `do_it()` concurrently on the single, global
114    /// thread. The following output would be equally valid:
115    ///
116    /// ```ascii
117    /// one one two two
118    /// ```
119    ///
120    /// # Panics
121    ///
122    /// If `op` should panic, that panic will be propagated.
123    ///
124    /// ## Using `install()`
125    ///
126    /// ```rust
127    ///    # use rayon_core as rayon;
128    ///    fn main() {
129    ///         let pool = rayon::ThreadPoolBuilder::new().num_threads(8).build().unwrap();
130    ///         let n = pool.install(|| fib(20));
131    ///         println!("{}", n);
132    ///    }
133    ///
134    ///    fn fib(n: usize) -> usize {
135    ///         if n == 0 || n == 1 {
136    ///             return n;
137    ///         }
138    ///         let (a, b) = rayon::join(|| fib(n - 1), || fib(n - 2)); // runs inside of `pool`
139    ///         return a + b;
140    ///     }
141    /// ```
142    pub fn install<OP, R>(&self, op: OP) -> R
143    where
144        OP: FnOnce() -> R + Send,
145        R: Send,
146    {
147        self.registry.in_worker(|_, _| op())
148    }
149
150    /// Executes `op` within every thread in the threadpool. Any attempts to use
151    /// `join`, `scope`, or parallel iterators will then operate within that
152    /// threadpool.
153    ///
154    /// Broadcasts are executed on each thread after they have exhausted their
155    /// local work queue, before they attempt work-stealing from other threads.
156    /// The goal of that strategy is to run everywhere in a timely manner
157    /// *without* being too disruptive to current work. There may be alternative
158    /// broadcast styles added in the future for more or less aggressive
159    /// injection, if the need arises.
160    ///
161    /// # Warning: thread-local data
162    ///
163    /// Because `op` is executing within the Rayon thread-pool,
164    /// thread-local data from the current thread will not be
165    /// accessible.
166    ///
167    /// # Panics
168    ///
169    /// If `op` should panic on one or more threads, exactly one panic
170    /// will be propagated, only after all threads have completed
171    /// (or panicked) their own `op`.
172    ///
173    /// # Examples
174    ///
175    /// ```
176    ///    # use rayon_core as rayon;
177    ///    use std::sync::atomic::{AtomicUsize, Ordering};
178    ///
179    ///    fn main() {
180    ///         let pool = rayon::ThreadPoolBuilder::new().num_threads(5).build().unwrap();
181    ///
182    ///         // The argument gives context, including the index of each thread.
183    ///         let v: Vec<usize> = pool.broadcast(|ctx| ctx.index() * ctx.index());
184    ///         assert_eq!(v, &[0, 1, 4, 9, 16]);
185    ///
186    ///         // The closure can reference the local stack
187    ///         let count = AtomicUsize::new(0);
188    ///         pool.broadcast(|_| count.fetch_add(1, Ordering::Relaxed));
189    ///         assert_eq!(count.into_inner(), 5);
190    ///    }
191    /// ```
192    pub fn broadcast<OP, R>(&self, op: OP) -> Vec<R>
193    where
194        OP: Fn(BroadcastContext<'_>) -> R + Sync,
195        R: Send,
196    {
197        // We assert that `self.registry` has not terminated.
198        unsafe { broadcast::broadcast_in(op, &self.registry) }
199    }
200
201    /// Returns the (current) number of threads in the thread pool.
202    ///
203    /// # Future compatibility note
204    ///
205    /// Note that unless this thread-pool was created with a
206    /// [`ThreadPoolBuilder`] that specifies the number of threads,
207    /// then this number may vary over time in future versions (see [the
208    /// `num_threads()` method for details][snt]).
209    ///
210    /// [snt]: struct.ThreadPoolBuilder.html#method.num_threads
211    /// [`ThreadPoolBuilder`]: struct.ThreadPoolBuilder.html
212    #[inline]
213    pub fn current_num_threads(&self) -> usize {
214        self.registry.num_threads()
215    }
216
217    /// If called from a Rayon worker thread in this thread-pool,
218    /// returns the index of that thread; if not called from a Rayon
219    /// thread, or called from a Rayon thread that belongs to a
220    /// different thread-pool, returns `None`.
221    ///
222    /// The index for a given thread will not change over the thread's
223    /// lifetime. However, multiple threads may share the same index if
224    /// they are in distinct thread-pools.
225    ///
226    /// # Future compatibility note
227    ///
228    /// Currently, every thread-pool (including the global
229    /// thread-pool) has a fixed number of threads, but this may
230    /// change in future Rayon versions (see [the `num_threads()` method
231    /// for details][snt]). In that case, the index for a
232    /// thread would not change during its lifetime, but thread
233    /// indices may wind up being reused if threads are terminated and
234    /// restarted.
235    ///
236    /// [snt]: struct.ThreadPoolBuilder.html#method.num_threads
237    #[inline]
238    pub fn current_thread_index(&self) -> Option<usize> {
239        let curr = self.registry.current_thread()?;
240        Some(curr.index())
241    }
242
243    /// Returns true if the current worker thread currently has "local
244    /// tasks" pending. This can be useful as part of a heuristic for
245    /// deciding whether to spawn a new task or execute code on the
246    /// current thread, particularly in breadth-first
247    /// schedulers. However, keep in mind that this is an inherently
248    /// racy check, as other worker threads may be actively "stealing"
249    /// tasks from our local deque.
250    ///
251    /// **Background:** Rayon's uses a [work-stealing] scheduler. The
252    /// key idea is that each thread has its own [deque] of
253    /// tasks. Whenever a new task is spawned -- whether through
254    /// `join()`, `Scope::spawn()`, or some other means -- that new
255    /// task is pushed onto the thread's *local* deque. Worker threads
256    /// have a preference for executing their own tasks; if however
257    /// they run out of tasks, they will go try to "steal" tasks from
258    /// other threads. This function therefore has an inherent race
259    /// with other active worker threads, which may be removing items
260    /// from the local deque.
261    ///
262    /// [work-stealing]: https://en.wikipedia.org/wiki/Work_stealing
263    /// [deque]: https://en.wikipedia.org/wiki/Double-ended_queue
264    #[inline]
265    pub fn current_thread_has_pending_tasks(&self) -> Option<bool> {
266        let curr = self.registry.current_thread()?;
267        Some(!curr.local_deque_is_empty())
268    }
269
270    /// Execute `oper_a` and `oper_b` in the thread-pool and return
271    /// the results. Equivalent to `self.install(|| join(oper_a,
272    /// oper_b))`.
273    pub fn join<A, B, RA, RB>(&self, oper_a: A, oper_b: B) -> (RA, RB)
274    where
275        A: FnOnce() -> RA + Send,
276        B: FnOnce() -> RB + Send,
277        RA: Send,
278        RB: Send,
279    {
280        self.install(|| join(oper_a, oper_b))
281    }
282
283    /// Creates a scope that executes within this thread-pool.
284    /// Equivalent to `self.install(|| scope(...))`.
285    ///
286    /// See also: [the `scope()` function][scope].
287    ///
288    /// [scope]: fn.scope.html
289    pub fn scope<'scope, OP, R>(&self, op: OP) -> R
290    where
291        OP: FnOnce(&Scope<'scope>) -> R + Send,
292        R: Send,
293    {
294        self.install(|| scope(op))
295    }
296
297    /// Creates a scope that executes within this thread-pool.
298    /// Spawns from the same thread are prioritized in relative FIFO order.
299    /// Equivalent to `self.install(|| scope_fifo(...))`.
300    ///
301    /// See also: [the `scope_fifo()` function][scope_fifo].
302    ///
303    /// [scope_fifo]: fn.scope_fifo.html
304    pub fn scope_fifo<'scope, OP, R>(&self, op: OP) -> R
305    where
306        OP: FnOnce(&ScopeFifo<'scope>) -> R + Send,
307        R: Send,
308    {
309        self.install(|| scope_fifo(op))
310    }
311
312    /// Creates a scope that spawns work into this thread-pool.
313    ///
314    /// See also: [the `in_place_scope()` function][in_place_scope].
315    ///
316    /// [in_place_scope]: fn.in_place_scope.html
317    pub fn in_place_scope<'scope, OP, R>(&self, op: OP) -> R
318    where
319        OP: FnOnce(&Scope<'scope>) -> R,
320    {
321        do_in_place_scope(Some(&self.registry), op)
322    }
323
324    /// Creates a scope that spawns work into this thread-pool in FIFO order.
325    ///
326    /// See also: [the `in_place_scope_fifo()` function][in_place_scope_fifo].
327    ///
328    /// [in_place_scope_fifo]: fn.in_place_scope_fifo.html
329    pub fn in_place_scope_fifo<'scope, OP, R>(&self, op: OP) -> R
330    where
331        OP: FnOnce(&ScopeFifo<'scope>) -> R,
332    {
333        do_in_place_scope_fifo(Some(&self.registry), op)
334    }
335
336    /// Spawns an asynchronous task in this thread-pool. This task will
337    /// run in the implicit, global scope, which means that it may outlast
338    /// the current stack frame -- therefore, it cannot capture any references
339    /// onto the stack (you will likely need a `move` closure).
340    ///
341    /// See also: [the `spawn()` function defined on scopes][spawn].
342    ///
343    /// [spawn]: struct.Scope.html#method.spawn
344    pub fn spawn<OP>(&self, op: OP)
345    where
346        OP: FnOnce() + Send + 'static,
347    {
348        // We assert that `self.registry` has not terminated.
349        unsafe { spawn::spawn_in(op, &self.registry) }
350    }
351
352    /// Spawns an asynchronous task in this thread-pool. This task will
353    /// run in the implicit, global scope, which means that it may outlast
354    /// the current stack frame -- therefore, it cannot capture any references
355    /// onto the stack (you will likely need a `move` closure).
356    ///
357    /// See also: [the `spawn_fifo()` function defined on scopes][spawn_fifo].
358    ///
359    /// [spawn_fifo]: struct.ScopeFifo.html#method.spawn_fifo
360    pub fn spawn_fifo<OP>(&self, op: OP)
361    where
362        OP: FnOnce() + Send + 'static,
363    {
364        // We assert that `self.registry` has not terminated.
365        unsafe { spawn::spawn_fifo_in(op, &self.registry) }
366    }
367
368    /// Spawns an asynchronous task on every thread in this thread-pool. This task
369    /// will run in the implicit, global scope, which means that it may outlast the
370    /// current stack frame -- therefore, it cannot capture any references onto the
371    /// stack (you will likely need a `move` closure).
372    pub fn spawn_broadcast<OP>(&self, op: OP)
373    where
374        OP: Fn(BroadcastContext<'_>) + Send + Sync + 'static,
375    {
376        // We assert that `self.registry` has not terminated.
377        unsafe { broadcast::spawn_broadcast_in(op, &self.registry) }
378    }
379
380    /// Cooperatively yields execution to Rayon.
381    ///
382    /// This is similar to the general [`yield_now()`], but only if the current
383    /// thread is part of *this* thread pool.
384    ///
385    /// Returns `Some(Yield::Executed)` if anything was executed, `Some(Yield::Idle)` if
386    /// nothing was available, or `None` if the current thread is not part this pool.
387    pub fn yield_now(&self) -> Option<Yield> {
388        let curr = self.registry.current_thread()?;
389        Some(curr.yield_now())
390    }
391
392    /// Cooperatively yields execution to local Rayon work.
393    ///
394    /// This is similar to the general [`yield_local()`], but only if the current
395    /// thread is part of *this* thread pool.
396    ///
397    /// Returns `Some(Yield::Executed)` if anything was executed, `Some(Yield::Idle)` if
398    /// nothing was available, or `None` if the current thread is not part this pool.
399    pub fn yield_local(&self) -> Option<Yield> {
400        let curr = self.registry.current_thread()?;
401        Some(curr.yield_local())
402    }
403}
404
405impl Drop for ThreadPool {
406    fn drop(&mut self) {
407        self.registry.terminate();
408    }
409}
410
411impl fmt::Debug for ThreadPool {
412    fn fmt(&self, fmt: &mut fmt::Formatter<'_>) -> fmt::Result {
413        fmt.debug_struct("ThreadPool")
414            .field("num_threads", &self.current_num_threads())
415            .field("id", &self.registry.id())
416            .finish()
417    }
418}
419
420/// If called from a Rayon worker thread, returns the index of that
421/// thread within its current pool; if not called from a Rayon thread,
422/// returns `None`.
423///
424/// The index for a given thread will not change over the thread's
425/// lifetime. However, multiple threads may share the same index if
426/// they are in distinct thread-pools.
427///
428/// See also: [the `ThreadPool::current_thread_index()` method].
429///
430/// [m]: struct.ThreadPool.html#method.current_thread_index
431///
432/// # Future compatibility note
433///
434/// Currently, every thread-pool (including the global
435/// thread-pool) has a fixed number of threads, but this may
436/// change in future Rayon versions (see [the `num_threads()` method
437/// for details][snt]). In that case, the index for a
438/// thread would not change during its lifetime, but thread
439/// indices may wind up being reused if threads are terminated and
440/// restarted.
441///
442/// [snt]: struct.ThreadPoolBuilder.html#method.num_threads
443#[inline]
444pub fn current_thread_index() -> Option<usize> {
445    unsafe {
446        let curr = WorkerThread::current().as_ref()?;
447        Some(curr.index())
448    }
449}
450
451/// If called from a Rayon worker thread, indicates whether that
452/// thread's local deque still has pending tasks. Otherwise, returns
453/// `None`. For more information, see [the
454/// `ThreadPool::current_thread_has_pending_tasks()` method][m].
455///
456/// [m]: struct.ThreadPool.html#method.current_thread_has_pending_tasks
457#[inline]
458pub fn current_thread_has_pending_tasks() -> Option<bool> {
459    unsafe {
460        let curr = WorkerThread::current().as_ref()?;
461        Some(!curr.local_deque_is_empty())
462    }
463}
464
465/// Cooperatively yields execution to Rayon.
466///
467/// If the current thread is part of a rayon thread pool, this looks for a
468/// single unit of pending work in the pool, then executes it. Completion of
469/// that work might include nested work or further work stealing.
470///
471/// This is similar to [`std::thread::yield_now()`], but does not literally make
472/// that call. If you are implementing a polling loop, you may want to also
473/// yield to the OS scheduler yourself if no Rayon work was found.
474///
475/// Returns `Some(Yield::Executed)` if anything was executed, `Some(Yield::Idle)` if
476/// nothing was available, or `None` if this thread is not part of any pool at all.
477pub fn yield_now() -> Option<Yield> {
478    unsafe {
479        let thread = WorkerThread::current().as_ref()?;
480        Some(thread.yield_now())
481    }
482}
483
484/// Cooperatively yields execution to local Rayon work.
485///
486/// If the current thread is part of a rayon thread pool, this looks for a
487/// single unit of pending work in this thread's queue, then executes it.
488/// Completion of that work might include nested work or further work stealing.
489///
490/// This is similar to [`yield_now()`], but does not steal from other threads.
491///
492/// Returns `Some(Yield::Executed)` if anything was executed, `Some(Yield::Idle)` if
493/// nothing was available, or `None` if this thread is not part of any pool at all.
494pub fn yield_local() -> Option<Yield> {
495    unsafe {
496        let thread = WorkerThread::current().as_ref()?;
497        Some(thread.yield_local())
498    }
499}
500
501/// Result of [`yield_now()`] or [`yield_local()`].
502#[derive(Clone, Copy, Debug, PartialEq, Eq)]
503pub enum Yield {
504    /// Work was found and executed.
505    Executed,
506    /// No available work was found.
507    Idle,
508}