1 | //! Methods for custom fork-join scopes, created by the [`scope()`] |
2 | //! and [`in_place_scope()`] functions. These are a more flexible alternative to [`join()`]. |
3 | //! |
4 | //! [`scope()`]: fn.scope.html |
5 | //! [`in_place_scope()`]: fn.in_place_scope.html |
6 | //! [`join()`]: ../join/join.fn.html |
7 | |
8 | use crate::broadcast::BroadcastContext; |
9 | use crate::job::{ArcJob, HeapJob, JobFifo, JobRef}; |
10 | use crate::latch::{CountLatch, Latch}; |
11 | use crate::registry::{global_registry, in_worker, Registry, WorkerThread}; |
12 | use crate::unwind; |
13 | use std::any::Any; |
14 | use std::fmt; |
15 | use std::marker::PhantomData; |
16 | use std::mem::ManuallyDrop; |
17 | use std::ptr; |
18 | use std::sync::atomic::{AtomicPtr, Ordering}; |
19 | use std::sync::Arc; |
20 | |
21 | #[cfg (test)] |
22 | mod test; |
23 | |
24 | /// Represents a fork-join scope which can be used to spawn any number of tasks. |
25 | /// See [`scope()`] for more information. |
26 | /// |
27 | ///[`scope()`]: fn.scope.html |
28 | pub struct Scope<'scope> { |
29 | base: ScopeBase<'scope>, |
30 | } |
31 | |
32 | /// Represents a fork-join scope which can be used to spawn any number of tasks. |
33 | /// Those spawned from the same thread are prioritized in relative FIFO order. |
34 | /// See [`scope_fifo()`] for more information. |
35 | /// |
36 | ///[`scope_fifo()`]: fn.scope_fifo.html |
37 | pub struct ScopeFifo<'scope> { |
38 | base: ScopeBase<'scope>, |
39 | fifos: Vec<JobFifo>, |
40 | } |
41 | |
42 | struct ScopeBase<'scope> { |
43 | /// thread registry where `scope()` was executed or where `in_place_scope()` |
44 | /// should spawn jobs. |
45 | registry: Arc<Registry>, |
46 | |
47 | /// if some job panicked, the error is stored here; it will be |
48 | /// propagated to the one who created the scope |
49 | panic: AtomicPtr<Box<dyn Any + Send + 'static>>, |
50 | |
51 | /// latch to track job counts |
52 | job_completed_latch: CountLatch, |
53 | |
54 | /// You can think of a scope as containing a list of closures to execute, |
55 | /// all of which outlive `'scope`. They're not actually required to be |
56 | /// `Sync`, but it's still safe to let the `Scope` implement `Sync` because |
57 | /// the closures are only *moved* across threads to be executed. |
58 | marker: PhantomData<Box<dyn FnOnce(&Scope<'scope>) + Send + Sync + 'scope>>, |
59 | } |
60 | |
61 | /// Creates a "fork-join" scope `s` and invokes the closure with a |
62 | /// reference to `s`. This closure can then spawn asynchronous tasks |
63 | /// into `s`. Those tasks may run asynchronously with respect to the |
64 | /// closure; they may themselves spawn additional tasks into `s`. When |
65 | /// the closure returns, it will block until all tasks that have been |
66 | /// spawned into `s` complete. |
67 | /// |
68 | /// `scope()` is a more flexible building block compared to `join()`, |
69 | /// since a loop can be used to spawn any number of tasks without |
70 | /// recursing. However, that flexibility comes at a performance price: |
71 | /// tasks spawned using `scope()` must be allocated onto the heap, |
72 | /// whereas `join()` can make exclusive use of the stack. **Prefer |
73 | /// `join()` (or, even better, parallel iterators) where possible.** |
74 | /// |
75 | /// # Example |
76 | /// |
77 | /// The Rayon `join()` function launches two closures and waits for them |
78 | /// to stop. One could implement `join()` using a scope like so, although |
79 | /// it would be less efficient than the real implementation: |
80 | /// |
81 | /// ```rust |
82 | /// # use rayon_core as rayon; |
83 | /// pub fn join<A,B,RA,RB>(oper_a: A, oper_b: B) -> (RA, RB) |
84 | /// where A: FnOnce() -> RA + Send, |
85 | /// B: FnOnce() -> RB + Send, |
86 | /// RA: Send, |
87 | /// RB: Send, |
88 | /// { |
89 | /// let mut result_a: Option<RA> = None; |
90 | /// let mut result_b: Option<RB> = None; |
91 | /// rayon::scope(|s| { |
92 | /// s.spawn(|_| result_a = Some(oper_a())); |
93 | /// s.spawn(|_| result_b = Some(oper_b())); |
94 | /// }); |
95 | /// (result_a.unwrap(), result_b.unwrap()) |
96 | /// } |
97 | /// ``` |
98 | /// |
99 | /// # A note on threading |
100 | /// |
101 | /// The closure given to `scope()` executes in the Rayon thread-pool, |
102 | /// as do those given to `spawn()`. This means that you can't access |
103 | /// thread-local variables (well, you can, but they may have |
104 | /// unexpected values). |
105 | /// |
106 | /// # Task execution |
107 | /// |
108 | /// Task execution potentially starts as soon as `spawn()` is called. |
109 | /// The task will end sometime before `scope()` returns. Note that the |
110 | /// *closure* given to scope may return much earlier. In general |
111 | /// the lifetime of a scope created like `scope(body)` goes something like this: |
112 | /// |
113 | /// - Scope begins when `scope(body)` is called |
114 | /// - Scope body `body()` is invoked |
115 | /// - Scope tasks may be spawned |
116 | /// - Scope body returns |
117 | /// - Scope tasks execute, possibly spawning more tasks |
118 | /// - Once all tasks are done, scope ends and `scope()` returns |
119 | /// |
120 | /// To see how and when tasks are joined, consider this example: |
121 | /// |
122 | /// ```rust |
123 | /// # use rayon_core as rayon; |
124 | /// // point start |
125 | /// rayon::scope(|s| { |
126 | /// s.spawn(|s| { // task s.1 |
127 | /// s.spawn(|s| { // task s.1.1 |
128 | /// rayon::scope(|t| { |
129 | /// t.spawn(|_| ()); // task t.1 |
130 | /// t.spawn(|_| ()); // task t.2 |
131 | /// }); |
132 | /// }); |
133 | /// }); |
134 | /// s.spawn(|s| { // task s.2 |
135 | /// }); |
136 | /// // point mid |
137 | /// }); |
138 | /// // point end |
139 | /// ``` |
140 | /// |
141 | /// The various tasks that are run will execute roughly like so: |
142 | /// |
143 | /// ```notrust |
144 | /// | (start) |
145 | /// | |
146 | /// | (scope `s` created) |
147 | /// +-----------------------------------------------+ (task s.2) |
148 | /// +-------+ (task s.1) | |
149 | /// | | | |
150 | /// | +---+ (task s.1.1) | |
151 | /// | | | | |
152 | /// | | | (scope `t` created) | |
153 | /// | | +----------------+ (task t.2) | |
154 | /// | | +---+ (task t.1) | | |
155 | /// | (mid) | | | | | |
156 | /// : | + <-+------------+ (scope `t` ends) | |
157 | /// : | | | |
158 | /// |<------+---+-----------------------------------+ (scope `s` ends) |
159 | /// | |
160 | /// | (end) |
161 | /// ``` |
162 | /// |
163 | /// The point here is that everything spawned into scope `s` will |
164 | /// terminate (at latest) at the same point -- right before the |
165 | /// original call to `rayon::scope` returns. This includes new |
166 | /// subtasks created by other subtasks (e.g., task `s.1.1`). If a new |
167 | /// scope is created (such as `t`), the things spawned into that scope |
168 | /// will be joined before that scope returns, which in turn occurs |
169 | /// before the creating task (task `s.1.1` in this case) finishes. |
170 | /// |
171 | /// There is no guaranteed order of execution for spawns in a scope, |
172 | /// given that other threads may steal tasks at any time. However, they |
173 | /// are generally prioritized in a LIFO order on the thread from which |
174 | /// they were spawned. So in this example, absent any stealing, we can |
175 | /// expect `s.2` to execute before `s.1`, and `t.2` before `t.1`. Other |
176 | /// threads always steal from the other end of the deque, like FIFO |
177 | /// order. The idea is that "recent" tasks are most likely to be fresh |
178 | /// in the local CPU's cache, while other threads can steal older |
179 | /// "stale" tasks. For an alternate approach, consider |
180 | /// [`scope_fifo()`] instead. |
181 | /// |
182 | /// [`scope_fifo()`]: fn.scope_fifo.html |
183 | /// |
184 | /// # Accessing stack data |
185 | /// |
186 | /// In general, spawned tasks may access stack data in place that |
187 | /// outlives the scope itself. Other data must be fully owned by the |
188 | /// spawned task. |
189 | /// |
190 | /// ```rust |
191 | /// # use rayon_core as rayon; |
192 | /// let ok: Vec<i32> = vec![1, 2, 3]; |
193 | /// rayon::scope(|s| { |
194 | /// let bad: Vec<i32> = vec![4, 5, 6]; |
195 | /// s.spawn(|_| { |
196 | /// // We can access `ok` because outlives the scope `s`. |
197 | /// println!("ok: {:?}" , ok); |
198 | /// |
199 | /// // If we just try to use `bad` here, the closure will borrow `bad` |
200 | /// // (because we are just printing it out, and that only requires a |
201 | /// // borrow), which will result in a compilation error. Read on |
202 | /// // for options. |
203 | /// // println!("bad: {:?}", bad); |
204 | /// }); |
205 | /// }); |
206 | /// ``` |
207 | /// |
208 | /// As the comments example above suggest, to reference `bad` we must |
209 | /// take ownership of it. One way to do this is to detach the closure |
210 | /// from the surrounding stack frame, using the `move` keyword. This |
211 | /// will cause it to take ownership of *all* the variables it touches, |
212 | /// in this case including both `ok` *and* `bad`: |
213 | /// |
214 | /// ```rust |
215 | /// # use rayon_core as rayon; |
216 | /// let ok: Vec<i32> = vec![1, 2, 3]; |
217 | /// rayon::scope(|s| { |
218 | /// let bad: Vec<i32> = vec![4, 5, 6]; |
219 | /// s.spawn(move |_| { |
220 | /// println!("ok: {:?}" , ok); |
221 | /// println!("bad: {:?}" , bad); |
222 | /// }); |
223 | /// |
224 | /// // That closure is fine, but now we can't use `ok` anywhere else, |
225 | /// // since it is owned by the previous task: |
226 | /// // s.spawn(|_| println!("ok: {:?}", ok)); |
227 | /// }); |
228 | /// ``` |
229 | /// |
230 | /// While this works, it could be a problem if we want to use `ok` elsewhere. |
231 | /// There are two choices. We can keep the closure as a `move` closure, but |
232 | /// instead of referencing the variable `ok`, we create a shadowed variable that |
233 | /// is a borrow of `ok` and capture *that*: |
234 | /// |
235 | /// ```rust |
236 | /// # use rayon_core as rayon; |
237 | /// let ok: Vec<i32> = vec![1, 2, 3]; |
238 | /// rayon::scope(|s| { |
239 | /// let bad: Vec<i32> = vec![4, 5, 6]; |
240 | /// let ok: &Vec<i32> = &ok; // shadow the original `ok` |
241 | /// s.spawn(move |_| { |
242 | /// println!("ok: {:?}" , ok); // captures the shadowed version |
243 | /// println!("bad: {:?}" , bad); |
244 | /// }); |
245 | /// |
246 | /// // Now we too can use the shadowed `ok`, since `&Vec<i32>` references |
247 | /// // can be shared freely. Note that we need a `move` closure here though, |
248 | /// // because otherwise we'd be trying to borrow the shadowed `ok`, |
249 | /// // and that doesn't outlive `scope`. |
250 | /// s.spawn(move |_| println!("ok: {:?}" , ok)); |
251 | /// }); |
252 | /// ``` |
253 | /// |
254 | /// Another option is not to use the `move` keyword but instead to take ownership |
255 | /// of individual variables: |
256 | /// |
257 | /// ```rust |
258 | /// # use rayon_core as rayon; |
259 | /// let ok: Vec<i32> = vec![1, 2, 3]; |
260 | /// rayon::scope(|s| { |
261 | /// let bad: Vec<i32> = vec![4, 5, 6]; |
262 | /// s.spawn(|_| { |
263 | /// // Transfer ownership of `bad` into a local variable (also named `bad`). |
264 | /// // This will force the closure to take ownership of `bad` from the environment. |
265 | /// let bad = bad; |
266 | /// println!("ok: {:?}" , ok); // `ok` is only borrowed. |
267 | /// println!("bad: {:?}" , bad); // refers to our local variable, above. |
268 | /// }); |
269 | /// |
270 | /// s.spawn(|_| println!("ok: {:?}" , ok)); // we too can borrow `ok` |
271 | /// }); |
272 | /// ``` |
273 | /// |
274 | /// # Panics |
275 | /// |
276 | /// If a panic occurs, either in the closure given to `scope()` or in |
277 | /// any of the spawned jobs, that panic will be propagated and the |
278 | /// call to `scope()` will panic. If multiple panics occurs, it is |
279 | /// non-deterministic which of their panic values will propagate. |
280 | /// Regardless, once a task is spawned using `scope.spawn()`, it will |
281 | /// execute, even if the spawning task should later panic. `scope()` |
282 | /// returns once all spawned jobs have completed, and any panics are |
283 | /// propagated at that point. |
284 | pub fn scope<'scope, OP, R>(op: OP) -> R |
285 | where |
286 | OP: FnOnce(&Scope<'scope>) -> R + Send, |
287 | R: Send, |
288 | { |
289 | in_worker(|owner_thread, _| { |
290 | let scope = Scope::<'scope>::new(Some(owner_thread), None); |
291 | scope.base.complete(Some(owner_thread), || op(&scope)) |
292 | }) |
293 | } |
294 | |
295 | /// Creates a "fork-join" scope `s` with FIFO order, and invokes the |
296 | /// closure with a reference to `s`. This closure can then spawn |
297 | /// asynchronous tasks into `s`. Those tasks may run asynchronously with |
298 | /// respect to the closure; they may themselves spawn additional tasks |
299 | /// into `s`. When the closure returns, it will block until all tasks |
300 | /// that have been spawned into `s` complete. |
301 | /// |
302 | /// # Task execution |
303 | /// |
304 | /// Tasks in a `scope_fifo()` run similarly to [`scope()`], but there's a |
305 | /// difference in the order of execution. Consider a similar example: |
306 | /// |
307 | /// [`scope()`]: fn.scope.html |
308 | /// |
309 | /// ```rust |
310 | /// # use rayon_core as rayon; |
311 | /// // point start |
312 | /// rayon::scope_fifo(|s| { |
313 | /// s.spawn_fifo(|s| { // task s.1 |
314 | /// s.spawn_fifo(|s| { // task s.1.1 |
315 | /// rayon::scope_fifo(|t| { |
316 | /// t.spawn_fifo(|_| ()); // task t.1 |
317 | /// t.spawn_fifo(|_| ()); // task t.2 |
318 | /// }); |
319 | /// }); |
320 | /// }); |
321 | /// s.spawn_fifo(|s| { // task s.2 |
322 | /// }); |
323 | /// // point mid |
324 | /// }); |
325 | /// // point end |
326 | /// ``` |
327 | /// |
328 | /// The various tasks that are run will execute roughly like so: |
329 | /// |
330 | /// ```notrust |
331 | /// | (start) |
332 | /// | |
333 | /// | (FIFO scope `s` created) |
334 | /// +--------------------+ (task s.1) |
335 | /// +-------+ (task s.2) | |
336 | /// | | +---+ (task s.1.1) |
337 | /// | | | | |
338 | /// | | | | (FIFO scope `t` created) |
339 | /// | | | +----------------+ (task t.1) |
340 | /// | | | +---+ (task t.2) | |
341 | /// | (mid) | | | | | |
342 | /// : | | + <-+------------+ (scope `t` ends) |
343 | /// : | | | |
344 | /// |<------+------------+---+ (scope `s` ends) |
345 | /// | |
346 | /// | (end) |
347 | /// ``` |
348 | /// |
349 | /// Under `scope_fifo()`, the spawns are prioritized in a FIFO order on |
350 | /// the thread from which they were spawned, as opposed to `scope()`'s |
351 | /// LIFO. So in this example, we can expect `s.1` to execute before |
352 | /// `s.2`, and `t.1` before `t.2`. Other threads also steal tasks in |
353 | /// FIFO order, as usual. Overall, this has roughly the same order as |
354 | /// the now-deprecated [`breadth_first`] option, except the effect is |
355 | /// isolated to a particular scope. If spawns are intermingled from any |
356 | /// combination of `scope()` and `scope_fifo()`, or from different |
357 | /// threads, their order is only specified with respect to spawns in the |
358 | /// same scope and thread. |
359 | /// |
360 | /// For more details on this design, see Rayon [RFC #1]. |
361 | /// |
362 | /// [`breadth_first`]: struct.ThreadPoolBuilder.html#method.breadth_first |
363 | /// [RFC #1]: https://github.com/rayon-rs/rfcs/blob/master/accepted/rfc0001-scope-scheduling.md |
364 | /// |
365 | /// # Panics |
366 | /// |
367 | /// If a panic occurs, either in the closure given to `scope_fifo()` or |
368 | /// in any of the spawned jobs, that panic will be propagated and the |
369 | /// call to `scope_fifo()` will panic. If multiple panics occurs, it is |
370 | /// non-deterministic which of their panic values will propagate. |
371 | /// Regardless, once a task is spawned using `scope.spawn_fifo()`, it |
372 | /// will execute, even if the spawning task should later panic. |
373 | /// `scope_fifo()` returns once all spawned jobs have completed, and any |
374 | /// panics are propagated at that point. |
375 | pub fn scope_fifo<'scope, OP, R>(op: OP) -> R |
376 | where |
377 | OP: FnOnce(&ScopeFifo<'scope>) -> R + Send, |
378 | R: Send, |
379 | { |
380 | in_worker(|owner_thread, _| { |
381 | let scope = ScopeFifo::<'scope>::new(Some(owner_thread), None); |
382 | scope.base.complete(Some(owner_thread), || op(&scope)) |
383 | }) |
384 | } |
385 | |
386 | /// Creates a "fork-join" scope `s` and invokes the closure with a |
387 | /// reference to `s`. This closure can then spawn asynchronous tasks |
388 | /// into `s`. Those tasks may run asynchronously with respect to the |
389 | /// closure; they may themselves spawn additional tasks into `s`. When |
390 | /// the closure returns, it will block until all tasks that have been |
391 | /// spawned into `s` complete. |
392 | /// |
393 | /// This is just like `scope()` except the closure runs on the same thread |
394 | /// that calls `in_place_scope()`. Only work that it spawns runs in the |
395 | /// thread pool. |
396 | /// |
397 | /// # Panics |
398 | /// |
399 | /// If a panic occurs, either in the closure given to `in_place_scope()` or in |
400 | /// any of the spawned jobs, that panic will be propagated and the |
401 | /// call to `in_place_scope()` will panic. If multiple panics occurs, it is |
402 | /// non-deterministic which of their panic values will propagate. |
403 | /// Regardless, once a task is spawned using `scope.spawn()`, it will |
404 | /// execute, even if the spawning task should later panic. `in_place_scope()` |
405 | /// returns once all spawned jobs have completed, and any panics are |
406 | /// propagated at that point. |
407 | pub fn in_place_scope<'scope, OP, R>(op: OP) -> R |
408 | where |
409 | OP: FnOnce(&Scope<'scope>) -> R, |
410 | { |
411 | do_in_place_scope(None, op) |
412 | } |
413 | |
414 | pub(crate) fn do_in_place_scope<'scope, OP, R>(registry: Option<&Arc<Registry>>, op: OP) -> R |
415 | where |
416 | OP: FnOnce(&Scope<'scope>) -> R, |
417 | { |
418 | let thread = unsafe { WorkerThread::current().as_ref() }; |
419 | let scope = Scope::<'scope>::new(thread, registry); |
420 | scope.base.complete(thread, || op(&scope)) |
421 | } |
422 | |
423 | /// Creates a "fork-join" scope `s` with FIFO order, and invokes the |
424 | /// closure with a reference to `s`. This closure can then spawn |
425 | /// asynchronous tasks into `s`. Those tasks may run asynchronously with |
426 | /// respect to the closure; they may themselves spawn additional tasks |
427 | /// into `s`. When the closure returns, it will block until all tasks |
428 | /// that have been spawned into `s` complete. |
429 | /// |
430 | /// This is just like `scope_fifo()` except the closure runs on the same thread |
431 | /// that calls `in_place_scope_fifo()`. Only work that it spawns runs in the |
432 | /// thread pool. |
433 | /// |
434 | /// # Panics |
435 | /// |
436 | /// If a panic occurs, either in the closure given to `in_place_scope_fifo()` or in |
437 | /// any of the spawned jobs, that panic will be propagated and the |
438 | /// call to `in_place_scope_fifo()` will panic. If multiple panics occurs, it is |
439 | /// non-deterministic which of their panic values will propagate. |
440 | /// Regardless, once a task is spawned using `scope.spawn_fifo()`, it will |
441 | /// execute, even if the spawning task should later panic. `in_place_scope_fifo()` |
442 | /// returns once all spawned jobs have completed, and any panics are |
443 | /// propagated at that point. |
444 | pub fn in_place_scope_fifo<'scope, OP, R>(op: OP) -> R |
445 | where |
446 | OP: FnOnce(&ScopeFifo<'scope>) -> R, |
447 | { |
448 | do_in_place_scope_fifo(None, op) |
449 | } |
450 | |
451 | pub(crate) fn do_in_place_scope_fifo<'scope, OP, R>(registry: Option<&Arc<Registry>>, op: OP) -> R |
452 | where |
453 | OP: FnOnce(&ScopeFifo<'scope>) -> R, |
454 | { |
455 | let thread = unsafe { WorkerThread::current().as_ref() }; |
456 | let scope = ScopeFifo::<'scope>::new(thread, registry); |
457 | scope.base.complete(thread, || op(&scope)) |
458 | } |
459 | |
460 | impl<'scope> Scope<'scope> { |
461 | fn new(owner: Option<&WorkerThread>, registry: Option<&Arc<Registry>>) -> Self { |
462 | let base = ScopeBase::new(owner, registry); |
463 | Scope { base } |
464 | } |
465 | |
466 | /// Spawns a job into the fork-join scope `self`. This job will |
467 | /// execute sometime before the fork-join scope completes. The |
468 | /// job is specified as a closure, and this closure receives its |
469 | /// own reference to the scope `self` as argument. This can be |
470 | /// used to inject new jobs into `self`. |
471 | /// |
472 | /// # Returns |
473 | /// |
474 | /// Nothing. The spawned closures cannot pass back values to the |
475 | /// caller directly, though they can write to local variables on |
476 | /// the stack (if those variables outlive the scope) or |
477 | /// communicate through shared channels. |
478 | /// |
479 | /// (The intention is to eventually integrate with Rust futures to |
480 | /// support spawns of functions that compute a value.) |
481 | /// |
482 | /// # Examples |
483 | /// |
484 | /// ```rust |
485 | /// # use rayon_core as rayon; |
486 | /// let mut value_a = None; |
487 | /// let mut value_b = None; |
488 | /// let mut value_c = None; |
489 | /// rayon::scope(|s| { |
490 | /// s.spawn(|s1| { |
491 | /// // ^ this is the same scope as `s`; this handle `s1` |
492 | /// // is intended for use by the spawned task, |
493 | /// // since scope handles cannot cross thread boundaries. |
494 | /// |
495 | /// value_a = Some(22); |
496 | /// |
497 | /// // the scope `s` will not end until all these tasks are done |
498 | /// s1.spawn(|_| { |
499 | /// value_b = Some(44); |
500 | /// }); |
501 | /// }); |
502 | /// |
503 | /// s.spawn(|_| { |
504 | /// value_c = Some(66); |
505 | /// }); |
506 | /// }); |
507 | /// assert_eq!(value_a, Some(22)); |
508 | /// assert_eq!(value_b, Some(44)); |
509 | /// assert_eq!(value_c, Some(66)); |
510 | /// ``` |
511 | /// |
512 | /// # See also |
513 | /// |
514 | /// The [`scope` function] has more extensive documentation about |
515 | /// task spawning. |
516 | /// |
517 | /// [`scope` function]: fn.scope.html |
518 | pub fn spawn<BODY>(&self, body: BODY) |
519 | where |
520 | BODY: FnOnce(&Scope<'scope>) + Send + 'scope, |
521 | { |
522 | let scope_ptr = ScopePtr(self); |
523 | let job = HeapJob::new(move || unsafe { |
524 | // SAFETY: this job will execute before the scope ends. |
525 | let scope = scope_ptr.as_ref(); |
526 | ScopeBase::execute_job(&scope.base, move || body(scope)) |
527 | }); |
528 | let job_ref = self.base.heap_job_ref(job); |
529 | |
530 | // Since `Scope` implements `Sync`, we can't be sure that we're still in a |
531 | // thread of this pool, so we can't just push to the local worker thread. |
532 | // Also, this might be an in-place scope. |
533 | self.base.registry.inject_or_push(job_ref); |
534 | } |
535 | |
536 | /// Spawns a job into every thread of the fork-join scope `self`. This job will |
537 | /// execute on each thread sometime before the fork-join scope completes. The |
538 | /// job is specified as a closure, and this closure receives its own reference |
539 | /// to the scope `self` as argument, as well as a `BroadcastContext`. |
540 | pub fn spawn_broadcast<BODY>(&self, body: BODY) |
541 | where |
542 | BODY: Fn(&Scope<'scope>, BroadcastContext<'_>) + Send + Sync + 'scope, |
543 | { |
544 | let scope_ptr = ScopePtr(self); |
545 | let job = ArcJob::new(move || unsafe { |
546 | // SAFETY: this job will execute before the scope ends. |
547 | let scope = scope_ptr.as_ref(); |
548 | let body = &body; |
549 | let func = move || BroadcastContext::with(move |ctx| body(scope, ctx)); |
550 | ScopeBase::execute_job(&scope.base, func) |
551 | }); |
552 | self.base.inject_broadcast(job) |
553 | } |
554 | } |
555 | |
556 | impl<'scope> ScopeFifo<'scope> { |
557 | fn new(owner: Option<&WorkerThread>, registry: Option<&Arc<Registry>>) -> Self { |
558 | let base = ScopeBase::new(owner, registry); |
559 | let num_threads = base.registry.num_threads(); |
560 | let fifos = (0..num_threads).map(|_| JobFifo::new()).collect(); |
561 | ScopeFifo { base, fifos } |
562 | } |
563 | |
564 | /// Spawns a job into the fork-join scope `self`. This job will |
565 | /// execute sometime before the fork-join scope completes. The |
566 | /// job is specified as a closure, and this closure receives its |
567 | /// own reference to the scope `self` as argument. This can be |
568 | /// used to inject new jobs into `self`. |
569 | /// |
570 | /// # See also |
571 | /// |
572 | /// This method is akin to [`Scope::spawn()`], but with a FIFO |
573 | /// priority. The [`scope_fifo` function] has more details about |
574 | /// this distinction. |
575 | /// |
576 | /// [`Scope::spawn()`]: struct.Scope.html#method.spawn |
577 | /// [`scope_fifo` function]: fn.scope_fifo.html |
578 | pub fn spawn_fifo<BODY>(&self, body: BODY) |
579 | where |
580 | BODY: FnOnce(&ScopeFifo<'scope>) + Send + 'scope, |
581 | { |
582 | let scope_ptr = ScopePtr(self); |
583 | let job = HeapJob::new(move || unsafe { |
584 | // SAFETY: this job will execute before the scope ends. |
585 | let scope = scope_ptr.as_ref(); |
586 | ScopeBase::execute_job(&scope.base, move || body(scope)) |
587 | }); |
588 | let job_ref = self.base.heap_job_ref(job); |
589 | |
590 | // If we're in the pool, use our scope's private fifo for this thread to execute |
591 | // in a locally-FIFO order. Otherwise, just use the pool's global injector. |
592 | match self.base.registry.current_thread() { |
593 | Some(worker) => { |
594 | let fifo = &self.fifos[worker.index()]; |
595 | // SAFETY: this job will execute before the scope ends. |
596 | unsafe { worker.push(fifo.push(job_ref)) }; |
597 | } |
598 | None => self.base.registry.inject(job_ref), |
599 | } |
600 | } |
601 | |
602 | /// Spawns a job into every thread of the fork-join scope `self`. This job will |
603 | /// execute on each thread sometime before the fork-join scope completes. The |
604 | /// job is specified as a closure, and this closure receives its own reference |
605 | /// to the scope `self` as argument, as well as a `BroadcastContext`. |
606 | pub fn spawn_broadcast<BODY>(&self, body: BODY) |
607 | where |
608 | BODY: Fn(&ScopeFifo<'scope>, BroadcastContext<'_>) + Send + Sync + 'scope, |
609 | { |
610 | let scope_ptr = ScopePtr(self); |
611 | let job = ArcJob::new(move || unsafe { |
612 | // SAFETY: this job will execute before the scope ends. |
613 | let scope = scope_ptr.as_ref(); |
614 | let body = &body; |
615 | let func = move || BroadcastContext::with(move |ctx| body(scope, ctx)); |
616 | ScopeBase::execute_job(&scope.base, func) |
617 | }); |
618 | self.base.inject_broadcast(job) |
619 | } |
620 | } |
621 | |
622 | impl<'scope> ScopeBase<'scope> { |
623 | /// Creates the base of a new scope for the given registry |
624 | fn new(owner: Option<&WorkerThread>, registry: Option<&Arc<Registry>>) -> Self { |
625 | let registry = registry.unwrap_or_else(|| match owner { |
626 | Some(owner) => owner.registry(), |
627 | None => global_registry(), |
628 | }); |
629 | |
630 | ScopeBase { |
631 | registry: Arc::clone(registry), |
632 | panic: AtomicPtr::new(ptr::null_mut()), |
633 | job_completed_latch: CountLatch::new(owner), |
634 | marker: PhantomData, |
635 | } |
636 | } |
637 | |
638 | fn heap_job_ref<FUNC>(&self, job: Box<HeapJob<FUNC>>) -> JobRef |
639 | where |
640 | FUNC: FnOnce() + Send + 'scope, |
641 | { |
642 | unsafe { |
643 | self.job_completed_latch.increment(); |
644 | job.into_job_ref() |
645 | } |
646 | } |
647 | |
648 | fn inject_broadcast<FUNC>(&self, job: Arc<ArcJob<FUNC>>) |
649 | where |
650 | FUNC: Fn() + Send + Sync + 'scope, |
651 | { |
652 | let n_threads = self.registry.num_threads(); |
653 | let job_refs = (0..n_threads).map(|_| unsafe { |
654 | self.job_completed_latch.increment(); |
655 | ArcJob::as_job_ref(&job) |
656 | }); |
657 | |
658 | self.registry.inject_broadcast(job_refs); |
659 | } |
660 | |
661 | /// Executes `func` as a job, either aborting or executing as |
662 | /// appropriate. |
663 | fn complete<FUNC, R>(&self, owner: Option<&WorkerThread>, func: FUNC) -> R |
664 | where |
665 | FUNC: FnOnce() -> R, |
666 | { |
667 | let result = unsafe { Self::execute_job_closure(self, func) }; |
668 | self.job_completed_latch.wait(owner); |
669 | self.maybe_propagate_panic(); |
670 | result.unwrap() // only None if `op` panicked, and that would have been propagated |
671 | } |
672 | |
673 | /// Executes `func` as a job, either aborting or executing as |
674 | /// appropriate. |
675 | unsafe fn execute_job<FUNC>(this: *const Self, func: FUNC) |
676 | where |
677 | FUNC: FnOnce(), |
678 | { |
679 | let _: Option<()> = Self::execute_job_closure(this, func); |
680 | } |
681 | |
682 | /// Executes `func` as a job in scope. Adjusts the "job completed" |
683 | /// counters and also catches any panic and stores it into |
684 | /// `scope`. |
685 | unsafe fn execute_job_closure<FUNC, R>(this: *const Self, func: FUNC) -> Option<R> |
686 | where |
687 | FUNC: FnOnce() -> R, |
688 | { |
689 | let result = match unwind::halt_unwinding(func) { |
690 | Ok(r) => Some(r), |
691 | Err(err) => { |
692 | (*this).job_panicked(err); |
693 | None |
694 | } |
695 | }; |
696 | Latch::set(&(*this).job_completed_latch); |
697 | result |
698 | } |
699 | |
700 | fn job_panicked(&self, err: Box<dyn Any + Send + 'static>) { |
701 | // capture the first error we see, free the rest |
702 | if self.panic.load(Ordering::Relaxed).is_null() { |
703 | let nil = ptr::null_mut(); |
704 | let mut err = ManuallyDrop::new(Box::new(err)); // box up the fat ptr |
705 | let err_ptr: *mut Box<dyn Any + Send + 'static> = &mut **err; |
706 | if self |
707 | .panic |
708 | .compare_exchange(nil, err_ptr, Ordering::Release, Ordering::Relaxed) |
709 | .is_ok() |
710 | { |
711 | // ownership now transferred into self.panic |
712 | } else { |
713 | // another panic raced in ahead of us, so drop ours |
714 | let _: Box<Box<_>> = ManuallyDrop::into_inner(err); |
715 | } |
716 | } |
717 | } |
718 | |
719 | fn maybe_propagate_panic(&self) { |
720 | // propagate panic, if any occurred; at this point, all |
721 | // outstanding jobs have completed, so we can use a relaxed |
722 | // ordering: |
723 | let panic = self.panic.swap(ptr::null_mut(), Ordering::Relaxed); |
724 | if !panic.is_null() { |
725 | let value = unsafe { Box::from_raw(panic) }; |
726 | unwind::resume_unwinding(*value); |
727 | } |
728 | } |
729 | } |
730 | |
731 | impl<'scope> fmt::Debug for Scope<'scope> { |
732 | fn fmt(&self, fmt: &mut fmt::Formatter<'_>) -> fmt::Result { |
733 | fmt.debug_struct("Scope" ) |
734 | .field("pool_id" , &self.base.registry.id()) |
735 | .field("panic" , &self.base.panic) |
736 | .field("job_completed_latch" , &self.base.job_completed_latch) |
737 | .finish() |
738 | } |
739 | } |
740 | |
741 | impl<'scope> fmt::Debug for ScopeFifo<'scope> { |
742 | fn fmt(&self, fmt: &mut fmt::Formatter<'_>) -> fmt::Result { |
743 | fmt.debug_struct("ScopeFifo" ) |
744 | .field("num_fifos" , &self.fifos.len()) |
745 | .field("pool_id" , &self.base.registry.id()) |
746 | .field("panic" , &self.base.panic) |
747 | .field("job_completed_latch" , &self.base.job_completed_latch) |
748 | .finish() |
749 | } |
750 | } |
751 | |
752 | /// Used to capture a scope `&Self` pointer in jobs, without faking a lifetime. |
753 | /// |
754 | /// Unsafe code is still required to dereference the pointer, but that's fine in |
755 | /// scope jobs that are guaranteed to execute before the scope ends. |
756 | struct ScopePtr<T>(*const T); |
757 | |
758 | // SAFETY: !Send for raw pointers is not for safety, just as a lint |
759 | unsafe impl<T: Sync> Send for ScopePtr<T> {} |
760 | |
761 | // SAFETY: !Sync for raw pointers is not for safety, just as a lint |
762 | unsafe impl<T: Sync> Sync for ScopePtr<T> {} |
763 | |
764 | impl<T> ScopePtr<T> { |
765 | // Helper to avoid disjoint captures of `scope_ptr.0` |
766 | unsafe fn as_ref(&self) -> &T { |
767 | &*self.0 |
768 | } |
769 | } |
770 | |