1 use crate::job::StackJob;
2 use crate::latch::SpinLatch;
3 use crate::registry::{self, WorkerThread};
4 use crate::unwind;
5 use std::any::Any;
6 
7 use crate::FnContext;
8 
9 #[cfg(test)]
10 mod test;
11 
12 /// Takes two closures and *potentially* runs them in parallel. It
13 /// returns a pair of the results from those closures.
14 ///
15 /// Conceptually, calling `join()` is similar to spawning two threads,
16 /// one executing each of the two closures. However, the
17 /// implementation is quite different and incurs very low
18 /// overhead. The underlying technique is called "work stealing": the
19 /// Rayon runtime uses a fixed pool of worker threads and attempts to
20 /// only execute code in parallel when there are idle CPUs to handle
21 /// it.
22 ///
23 /// When `join` is called from outside the thread pool, the calling
24 /// thread will block while the closures execute in the pool.  When
25 /// `join` is called within the pool, the calling thread still actively
26 /// participates in the thread pool. It will begin by executing closure
27 /// A (on the current thread). While it is doing that, it will advertise
28 /// closure B as being available for other threads to execute. Once closure A
29 /// has completed, the current thread will try to execute closure B;
30 /// if however closure B has been stolen, then it will look for other work
31 /// while waiting for the thief to fully execute closure B. (This is the
32 /// typical work-stealing strategy).
33 ///
34 /// # Examples
35 ///
36 /// This example uses join to perform a quick-sort (note this is not a
37 /// particularly optimized implementation: if you **actually** want to
38 /// sort for real, you should prefer [the `par_sort` method] offered
39 /// by Rayon).
40 ///
41 /// [the `par_sort` method]: ../rayon/slice/trait.ParallelSliceMut.html#method.par_sort
42 ///
43 /// ```rust
44 /// # use rayon_core as rayon;
45 /// let mut v = vec![5, 1, 8, 22, 0, 44];
46 /// quick_sort(&mut v);
47 /// assert_eq!(v, vec![0, 1, 5, 8, 22, 44]);
48 ///
49 /// fn quick_sort<T:PartialOrd+Send>(v: &mut [T]) {
50 ///    if v.len() > 1 {
51 ///        let mid = partition(v);
52 ///        let (lo, hi) = v.split_at_mut(mid);
53 ///        rayon::join(|| quick_sort(lo),
54 ///                    || quick_sort(hi));
55 ///    }
56 /// }
57 ///
58 /// // Partition rearranges all items `<=` to the pivot
59 /// // item (arbitrary selected to be the last item in the slice)
60 /// // to the first half of the slice. It then returns the
61 /// // "dividing point" where the pivot is placed.
62 /// fn partition<T:PartialOrd+Send>(v: &mut [T]) -> usize {
63 ///     let pivot = v.len() - 1;
64 ///     let mut i = 0;
65 ///     for j in 0..pivot {
66 ///         if v[j] <= v[pivot] {
67 ///             v.swap(i, j);
68 ///             i += 1;
69 ///         }
70 ///     }
71 ///     v.swap(i, pivot);
72 ///     i
73 /// }
74 /// ```
75 ///
76 /// # Warning about blocking I/O
77 ///
78 /// The assumption is that the closures given to `join()` are
79 /// CPU-bound tasks that do not perform I/O or other blocking
80 /// operations. If you do perform I/O, and that I/O should block
81 /// (e.g., waiting for a network request), the overall performance may
82 /// be poor.  Moreover, if you cause one closure to be blocked waiting
83 /// on another (for example, using a channel), that could lead to a
84 /// deadlock.
85 ///
86 /// # Panics
87 ///
88 /// No matter what happens, both closures will always be executed.  If
89 /// a single closure panics, whether it be the first or second
90 /// closure, that panic will be propagated and hence `join()` will
91 /// panic with the same panic value. If both closures panic, `join()`
92 /// will panic with the panic value from the first closure.
join<A, B, RA, RB>(oper_a: A, oper_b: B) -> (RA, RB) where A: FnOnce() -> RA + Send, B: FnOnce() -> RB + Send, RA: Send, RB: Send,93 pub fn join<A, B, RA, RB>(oper_a: A, oper_b: B) -> (RA, RB)
94 where
95     A: FnOnce() -> RA + Send,
96     B: FnOnce() -> RB + Send,
97     RA: Send,
98     RB: Send,
99 {
100     #[inline]
101     fn call<R>(f: impl FnOnce() -> R) -> impl FnOnce(FnContext) -> R {
102         move |_| f()
103     }
104 
105     join_context(call(oper_a), call(oper_b))
106 }
107 
108 /// Identical to `join`, except that the closures have a parameter
109 /// that provides context for the way the closure has been called,
110 /// especially indicating whether they're executing on a different
111 /// thread than where `join_context` was called.  This will occur if
112 /// the second job is stolen by a different thread, or if
113 /// `join_context` was called from outside the thread pool to begin
114 /// with.
join_context<A, B, RA, RB>(oper_a: A, oper_b: B) -> (RA, RB) where A: FnOnce(FnContext) -> RA + Send, B: FnOnce(FnContext) -> RB + Send, RA: Send, RB: Send,115 pub fn join_context<A, B, RA, RB>(oper_a: A, oper_b: B) -> (RA, RB)
116 where
117     A: FnOnce(FnContext) -> RA + Send,
118     B: FnOnce(FnContext) -> RB + Send,
119     RA: Send,
120     RB: Send,
121 {
122     #[inline]
123     fn call_a<R>(f: impl FnOnce(FnContext) -> R, injected: bool) -> impl FnOnce() -> R {
124         move || f(FnContext::new(injected))
125     }
126 
127     #[inline]
128     fn call_b<R>(f: impl FnOnce(FnContext) -> R) -> impl FnOnce(bool) -> R {
129         move |migrated| f(FnContext::new(migrated))
130     }
131 
132     registry::in_worker(|worker_thread, injected| unsafe {
133         // Create virtual wrapper for task b; this all has to be
134         // done here so that the stack frame can keep it all live
135         // long enough.
136         let job_b = StackJob::new(call_b(oper_b), SpinLatch::new(worker_thread));
137         let job_b_ref = job_b.as_job_ref();
138         worker_thread.push(job_b_ref);
139 
140         // Execute task a; hopefully b gets stolen in the meantime.
141         let status_a = unwind::halt_unwinding(call_a(oper_a, injected));
142         let result_a = match status_a {
143             Ok(v) => v,
144             Err(err) => join_recover_from_panic(worker_thread, &job_b.latch, err),
145         };
146 
147         // Now that task A has finished, try to pop job B from the
148         // local stack.  It may already have been popped by job A; it
149         // may also have been stolen. There may also be some tasks
150         // pushed on top of it in the stack, and we will have to pop
151         // those off to get to it.
152         while !job_b.latch.probe() {
153             if let Some(job) = worker_thread.take_local_job() {
154                 if job == job_b_ref {
155                     // Found it! Let's run it.
156                     //
157                     // Note that this could panic, but it's ok if we unwind here.
158                     let result_b = job_b.run_inline(injected);
159                     return (result_a, result_b);
160                 } else {
161                     worker_thread.execute(job);
162                 }
163             } else {
164                 // Local deque is empty. Time to steal from other
165                 // threads.
166                 worker_thread.wait_until(&job_b.latch);
167                 debug_assert!(job_b.latch.probe());
168                 break;
169             }
170         }
171 
172         (result_a, job_b.into_result())
173     })
174 }
175 
176 /// If job A panics, we still cannot return until we are sure that job
177 /// B is complete. This is because it may contain references into the
178 /// enclosing stack frame(s).
179 #[cold] // cold path
join_recover_from_panic( worker_thread: &WorkerThread, job_b_latch: &SpinLatch<'_>, err: Box<dyn Any + Send>, ) -> !180 unsafe fn join_recover_from_panic(
181     worker_thread: &WorkerThread,
182     job_b_latch: &SpinLatch<'_>,
183     err: Box<dyn Any + Send>,
184 ) -> ! {
185     worker_thread.wait_until(job_b_latch);
186     unwind::resume_unwinding(err)
187 }
188