3 //! ## The threading model
5 //! An executing Rust program consists of a collection of native OS threads,
6 //! each with their own stack and local state. Threads can be named, and
7 //! provide some built-in support for low-level synchronization.
9 //! Communication between threads can be done through
10 //! [channels], Rust's message-passing types, along with [other forms of thread
11 //! synchronization](../../std/sync/index.html) and shared-memory data
12 //! structures. In particular, types that are guaranteed to be
13 //! threadsafe are easily shared between threads using the
14 //! atomically-reference-counted container, [`Arc`].
16 //! Fatal logic errors in Rust cause *thread panic*, during which
17 //! a thread will unwind the stack, running destructors and freeing
18 //! owned resources. While not meant as a 'try/catch' mechanism, panics
19 //! in Rust can nonetheless be caught (unless compiling with `panic=abort`) with
20 //! [`catch_unwind`](../../std/panic/fn.catch_unwind.html) and recovered
21 //! from, or alternatively be resumed with
22 //! [`resume_unwind`](../../std/panic/fn.resume_unwind.html). If the panic
23 //! is not caught the thread will exit, but the panic may optionally be
24 //! detected from a different thread with [`join`]. If the main thread panics
25 //! without the panic being caught, the application will exit with a
26 //! non-zero exit code.
28 //! When the main thread of a Rust program terminates, the entire program shuts
29 //! down, even if other threads are still running. However, this module provides
30 //! convenient facilities for automatically waiting for the termination of a
31 //! thread (i.e., join).
33 //! ## Spawning a thread
35 //! A new thread can be spawned using the [`thread::spawn`][`spawn`] function:
40 //! thread::spawn(move || {
45 //! In this example, the spawned thread is "detached," which means that there is
46 //! no way for the program to learn when the spawned thread completes or otherwise
49 //! To learn when a thread completes, it is necessary to capture the [`JoinHandle`]
50 //! object that is returned by the call to [`spawn`], which provides
51 //! a `join` method that allows the caller to wait for the completion of the
57 //! let thread_join_handle = thread::spawn(move || {
61 //! let res = thread_join_handle.join();
64 //! The [`join`] method returns a [`thread::Result`] containing [`Ok`] of the final
65 //! value produced by the spawned thread, or [`Err`] of the value given to
66 //! a call to [`panic!`] if the thread panicked.
68 //! Note that there is no parent/child relationship between a thread that spawns a
69 //! new thread and the thread being spawned. In particular, the spawned thread may or
70 //! may not outlive the spawning thread, unless the spawning thread is the main thread.
72 //! ## Configuring threads
74 //! A new thread can be configured before it is spawned via the [`Builder`] type,
75 //! which currently allows you to set the name and stack size for the thread:
78 //! # #![allow(unused_must_use)]
81 //! thread::Builder::new().name("thread1".to_string()).spawn(move || {
82 //! println!("Hello, world!");
86 //! ## The `Thread` type
88 //! Threads are represented via the [`Thread`] type, which you can get in one of
91 //! * By spawning a new thread, e.g., using the [`thread::spawn`][`spawn`]
92 //! function, and calling [`thread`][`JoinHandle::thread`] on the [`JoinHandle`].
93 //! * By requesting the current thread, using the [`thread::current`] function.
95 //! The [`thread::current`] function is available even for threads not spawned
96 //! by the APIs of this module.
98 //! ## Thread-local storage
100 //! This module also provides an implementation of thread-local storage for Rust
101 //! programs. Thread-local storage is a method of storing data into a global
102 //! variable that each thread in the program will have its own copy of.
103 //! Threads do not share this data, so accesses do not need to be synchronized.
105 //! A thread-local key owns the value it contains and will destroy the value when the
106 //! thread exits. It is created with the [`thread_local!`] macro and can contain any
107 //! value that is `'static` (no borrowed pointers). It provides an accessor function,
108 //! [`with`], that yields a shared reference to the value to the specified
109 //! closure. Thread-local keys allow only shared access to values, as there would be no
110 //! way to guarantee uniqueness if mutable borrows were allowed. Most values
111 //! will want to make use of some form of **interior mutability** through the
112 //! [`Cell`] or [`RefCell`] types.
114 //! ## Naming threads
116 //! Threads are able to have associated names for identification purposes. By default, spawned
117 //! threads are unnamed. To specify a name for a thread, build the thread with [`Builder`] and pass
118 //! the desired thread name to [`Builder::name`]. To retrieve the thread name from within the
119 //! thread, use [`Thread::name`]. A couple examples of where the name of a thread gets used:
121 //! * If a panic occurs in a named thread, the thread name will be printed in the panic message.
122 //! * The thread name is provided to the OS where applicable (e.g., `pthread_setname_np` in
123 //! unix-like platforms).
127 //! The default stack size for spawned threads is 2 MiB, though this particular stack size is
128 //! subject to change in the future. There are two ways to manually specify the stack size for
131 //! * Build the thread with [`Builder`] and pass the desired stack size to [`Builder::stack_size`].
132 //! * Set the `RUST_MIN_STACK` environment variable to an integer representing the desired stack
133 //! size (in bytes). Note that setting [`Builder::stack_size`] will override this.
135 //! Note that the stack size of the main thread is *not* determined by Rust.
137 //! [channels]: crate::sync::mpsc
138 //! [`join`]: JoinHandle::join
139 //! [`Result`]: crate::result::Result
140 //! [`Ok`]: crate::result::Result::Ok
141 //! [`Err`]: crate::result::Result::Err
142 //! [`thread::current`]: current
143 //! [`thread::Result`]: Result
144 //! [`unpark`]: Thread::unpark
145 //! [`thread::park_timeout`]: park_timeout
146 //! [`Cell`]: crate::cell::Cell
147 //! [`RefCell`]: crate::cell::RefCell
148 //! [`with`]: LocalKey::with
150 #![stable(feature = "rust1", since = "1.0.0")]
151 #![deny(unsafe_op_in_unsafe_fn)]
153 #[cfg(all(test, not(target_os = "emscripten")))]
157 use crate::cell::UnsafeCell;
158 use crate::ffi::{CStr, CString};
162 use crate::num::NonZeroU64;
163 use crate::num::NonZeroUsize;
165 use crate::panicking;
167 use crate::sync::Arc;
168 use crate::sys::thread as imp;
169 use crate::sys_common::mutex;
170 use crate::sys_common::thread;
171 use crate::sys_common::thread_info;
172 use crate::sys_common::thread_parker::Parker;
173 use crate::sys_common::{AsInner, IntoInner};
174 use crate::time::Duration;
176 ////////////////////////////////////////////////////////////////////////////////
177 // Thread-local storage
178 ////////////////////////////////////////////////////////////////////////////////
183 #[stable(feature = "rust1", since = "1.0.0")]
184 pub use self::local::{AccessError, LocalKey};
186 // The types used by the thread_local! macro to access TLS keys. Note that there
187 // are two types, the "OS" type and the "fast" type. The OS thread local key
188 // type is accessed via platform-specific API calls and is slow, while the fast
189 // key type is accessed via code generated via LLVM, where TLS keys are set up
190 // by the elf linker. Note that the OS TLS type is always available: on macOS
191 // the standard library is compiled with support for older platform versions
192 // where fast TLS was not available; end-user code is compiled with fast TLS
193 // where available, but both are needed.
195 #[unstable(feature = "libstd_thread_internals", issue = "none")]
196 #[cfg(target_thread_local)]
198 pub use self::local::fast::Key as __FastLocalKeyInner;
199 #[unstable(feature = "libstd_thread_internals", issue = "none")]
201 pub use self::local::os::Key as __OsLocalKeyInner;
202 #[unstable(feature = "libstd_thread_internals", issue = "none")]
203 #[cfg(all(target_arch = "wasm32", not(target_feature = "atomics")))]
205 pub use self::local::statik::Key as __StaticLocalKeyInner;
207 // This is only used to make thread locals with `const { .. }` initialization
208 // expressions unstable. If and/or when that syntax is stabilized with thread
209 // locals this will simply be removed.
211 #[unstable(feature = "thread_local_const_init", issue = "84223")]
212 pub const fn require_unstable_const_init_thread_local() {}
214 ////////////////////////////////////////////////////////////////////////////////
216 ////////////////////////////////////////////////////////////////////////////////
218 /// Thread factory, which can be used in order to configure the properties of
221 /// Methods can be chained on it in order to configure it.
223 /// The two configurations available are:
225 /// - [`name`]: specifies an [associated name for the thread][naming-threads]
226 /// - [`stack_size`]: specifies the [desired stack size for the thread][stack-size]
228 /// The [`spawn`] method will take ownership of the builder and create an
229 /// [`io::Result`] to the thread handle with the given configuration.
231 /// The [`thread::spawn`] free function uses a `Builder` with default
232 /// configuration and [`unwrap`]s its return value.
234 /// You may want to use [`spawn`] instead of [`thread::spawn`], when you want
235 /// to recover from a failure to launch a thread, indeed the free function will
236 /// panic where the `Builder` method will return a [`io::Result`].
243 /// let builder = thread::Builder::new();
245 /// let handler = builder.spawn(|| {
249 /// handler.join().unwrap();
252 /// [`stack_size`]: Builder::stack_size
253 /// [`name`]: Builder::name
254 /// [`spawn`]: Builder::spawn
255 /// [`thread::spawn`]: spawn
256 /// [`io::Result`]: crate::io::Result
257 /// [`unwrap`]: crate::result::Result::unwrap
258 /// [naming-threads]: ./index.html#naming-threads
259 /// [stack-size]: ./index.html#stack-size
260 #[stable(feature = "rust1", since = "1.0.0")]
263 // A name for the thread-to-be, for identification in panic messages
264 name: Option<String>,
265 // The size of the stack for the spawned thread in bytes
266 stack_size: Option<usize>,
270 /// Generates the base configuration for spawning a thread, from which
271 /// configuration methods can be chained.
278 /// let builder = thread::Builder::new()
279 /// .name("foo".into())
280 /// .stack_size(32 * 1024);
282 /// let handler = builder.spawn(|| {
286 /// handler.join().unwrap();
288 #[stable(feature = "rust1", since = "1.0.0")]
289 pub fn new() -> Builder {
290 Builder { name: None, stack_size: None }
293 /// Names the thread-to-be. Currently the name is used for identification
294 /// only in panic messages.
296 /// The name must not contain null bytes (`\0`).
298 /// For more information about named threads, see
299 /// [this module-level documentation][naming-threads].
306 /// let builder = thread::Builder::new()
307 /// .name("foo".into());
309 /// let handler = builder.spawn(|| {
310 /// assert_eq!(thread::current().name(), Some("foo"))
313 /// handler.join().unwrap();
316 /// [naming-threads]: ./index.html#naming-threads
317 #[stable(feature = "rust1", since = "1.0.0")]
318 pub fn name(mut self, name: String) -> Builder {
319 self.name = Some(name);
323 /// Sets the size of the stack (in bytes) for the new thread.
325 /// The actual stack size may be greater than this value if
326 /// the platform specifies a minimal stack size.
328 /// For more information about the stack size for threads, see
329 /// [this module-level documentation][stack-size].
336 /// let builder = thread::Builder::new().stack_size(32 * 1024);
339 /// [stack-size]: ./index.html#stack-size
340 #[stable(feature = "rust1", since = "1.0.0")]
341 pub fn stack_size(mut self, size: usize) -> Builder {
342 self.stack_size = Some(size);
346 /// Spawns a new thread by taking ownership of the `Builder`, and returns an
347 /// [`io::Result`] to its [`JoinHandle`].
349 /// The spawned thread may outlive the caller (unless the caller thread
350 /// is the main thread; the whole process is terminated when the main
351 /// thread finishes). The join handle can be used to block on
352 /// termination of the spawned thread, including recovering its panics.
354 /// For a more complete documentation see [`thread::spawn`][`spawn`].
358 /// Unlike the [`spawn`] free function, this method yields an
359 /// [`io::Result`] to capture any failure to create the thread at
362 /// [`io::Result`]: crate::io::Result
366 /// Panics if a thread name was set and it contained null bytes.
373 /// let builder = thread::Builder::new();
375 /// let handler = builder.spawn(|| {
379 /// handler.join().unwrap();
381 #[stable(feature = "rust1", since = "1.0.0")]
382 pub fn spawn<F, T>(self, f: F) -> io::Result<JoinHandle<T>>
388 unsafe { self.spawn_unchecked(f) }
391 /// Spawns a new thread without any lifetime restrictions by taking ownership
392 /// of the `Builder`, and returns an [`io::Result`] to its [`JoinHandle`].
394 /// The spawned thread may outlive the caller (unless the caller thread
395 /// is the main thread; the whole process is terminated when the main
396 /// thread finishes). The join handle can be used to block on
397 /// termination of the spawned thread, including recovering its panics.
399 /// This method is identical to [`thread::Builder::spawn`][`Builder::spawn`],
400 /// except for the relaxed lifetime bounds, which render it unsafe.
401 /// For a more complete documentation see [`thread::spawn`][`spawn`].
405 /// Unlike the [`spawn`] free function, this method yields an
406 /// [`io::Result`] to capture any failure to create the thread at
411 /// Panics if a thread name was set and it contained null bytes.
415 /// The caller has to ensure that the spawned thread does not outlive any
416 /// references in the supplied thread closure and its return type.
417 /// This can be guaranteed in two ways:
419 /// - ensure that [`join`][`JoinHandle::join`] is called before any referenced
421 /// - use only types with `'static` lifetime bounds, i.e., those with no or only
422 /// `'static` references (both [`thread::Builder::spawn`][`Builder::spawn`]
423 /// and [`thread::spawn`][`spawn`] enforce this property statically)
428 /// #![feature(thread_spawn_unchecked)]
431 /// let builder = thread::Builder::new();
434 /// let thread_x = &x;
436 /// let handler = unsafe {
437 /// builder.spawn_unchecked(move || {
438 /// println!("x = {}", *thread_x);
442 /// // caller has to ensure `join()` is called, otherwise
443 /// // it is possible to access freed memory if `x` gets
444 /// // dropped before the thread closure is executed!
445 /// handler.join().unwrap();
448 /// [`io::Result`]: crate::io::Result
449 #[unstable(feature = "thread_spawn_unchecked", issue = "55132")]
450 pub unsafe fn spawn_unchecked<'a, F, T>(self, f: F) -> io::Result<JoinHandle<T>>
456 let Builder { name, stack_size } = self;
458 let stack_size = stack_size.unwrap_or_else(thread::min_stack);
460 let my_thread = Thread::new(name.map(|name| {
461 CString::new(name).expect("thread name may not contain interior null bytes")
463 let their_thread = my_thread.clone();
465 let my_packet: Arc<UnsafeCell<Option<Result<T>>>> = Arc::new(UnsafeCell::new(None));
466 let their_packet = my_packet.clone();
468 let output_capture = crate::io::set_output_capture(None);
469 crate::io::set_output_capture(output_capture.clone());
472 if let Some(name) = their_thread.cname() {
473 imp::Thread::set_name(name);
476 crate::io::set_output_capture(output_capture);
478 // SAFETY: the stack guard passed is the one for the current thread.
479 // This means the current thread's stack and the new thread's stack
480 // are properly set and protected from each other.
481 thread_info::set(unsafe { imp::guard::current() }, their_thread);
482 let try_result = panic::catch_unwind(panic::AssertUnwindSafe(|| {
483 crate::sys_common::backtrace::__rust_begin_short_backtrace(f)
485 // SAFETY: `their_packet` as been built just above and moved by the
486 // closure (it is an Arc<...>) and `my_packet` will be stored in the
487 // same `JoinInner` as this closure meaning the mutation will be
488 // safe (not modify it and affect a value far away).
489 unsafe { *their_packet.get() = Some(try_result) };
492 Ok(JoinHandle(JoinInner {
495 // `imp::Thread::new` takes a closure with a `'static` lifetime, since it's passed
496 // through FFI or otherwise used with low-level threading primitives that have no
497 // notion of or way to enforce lifetimes.
499 // As mentioned in the `Safety` section of this function's documentation, the caller of
500 // this function needs to guarantee that the passed-in lifetime is sufficiently long
501 // for the lifetime of the thread.
503 // Similarly, the `sys` implementation must guarantee that no references to the closure
504 // exist after the thread has terminated, which is signaled by `Thread::join`
507 Some(imp::Thread::new(
509 mem::transmute::<Box<dyn FnOnce() + 'a>, Box<dyn FnOnce() + 'static>>(
515 packet: Packet(my_packet),
520 ////////////////////////////////////////////////////////////////////////////////
522 ////////////////////////////////////////////////////////////////////////////////
524 /// Spawns a new thread, returning a [`JoinHandle`] for it.
526 /// The join handle provides a [`join`] method that can be used to join the spawned
527 /// thread. If the spawned thread panics, [`join`] will return an [`Err`] containing
528 /// the argument given to [`panic!`].
530 /// If the join handle is dropped, the spawned thread will implicitly be *detached*.
531 /// In this case, the spawned thread may no longer be joined.
532 /// (It is the responsibility of the program to either eventually join threads it
533 /// creates or detach them; otherwise, a resource leak will result.)
535 /// This call will create a thread using default parameters of [`Builder`], if you
536 /// want to specify the stack size or the name of the thread, use this API
539 /// As you can see in the signature of `spawn` there are two constraints on
540 /// both the closure given to `spawn` and its return value, let's explain them:
542 /// - The `'static` constraint means that the closure and its return value
543 /// must have a lifetime of the whole program execution. The reason for this
544 /// is that threads can outlive the lifetime they have been created in.
546 /// Indeed if the thread, and by extension its return value, can outlive their
547 /// caller, we need to make sure that they will be valid afterwards, and since
548 /// we *can't* know when it will return we need to have them valid as long as
549 /// possible, that is until the end of the program, hence the `'static`
551 /// - The [`Send`] constraint is because the closure will need to be passed
552 /// *by value* from the thread where it is spawned to the new thread. Its
553 /// return value will need to be passed from the new thread to the thread
554 /// where it is `join`ed.
555 /// As a reminder, the [`Send`] marker trait expresses that it is safe to be
556 /// passed from thread to thread. [`Sync`] expresses that it is safe to have a
557 /// reference be passed from thread to thread.
561 /// Panics if the OS fails to create a thread; use [`Builder::spawn`]
562 /// to recover from such errors.
566 /// Creating a thread.
571 /// let handler = thread::spawn(|| {
575 /// handler.join().unwrap();
578 /// As mentioned in the module documentation, threads are usually made to
579 /// communicate using [`channels`], here is how it usually looks.
581 /// This example also shows how to use `move`, in order to give ownership
582 /// of values to a thread.
586 /// use std::sync::mpsc::channel;
588 /// let (tx, rx) = channel();
590 /// let sender = thread::spawn(move || {
591 /// tx.send("Hello, thread".to_owned())
592 /// .expect("Unable to send on channel");
595 /// let receiver = thread::spawn(move || {
596 /// let value = rx.recv().expect("Unable to receive from channel");
597 /// println!("{}", value);
600 /// sender.join().expect("The sender thread has panicked");
601 /// receiver.join().expect("The receiver thread has panicked");
604 /// A thread can also return a value through its [`JoinHandle`], you can use
605 /// this to make asynchronous computations (futures might be more appropriate
611 /// let computation = thread::spawn(|| {
612 /// // Some expensive computation.
616 /// let result = computation.join().unwrap();
617 /// println!("{}", result);
620 /// [`channels`]: crate::sync::mpsc
621 /// [`join`]: JoinHandle::join
622 /// [`Err`]: crate::result::Result::Err
623 #[stable(feature = "rust1", since = "1.0.0")]
624 pub fn spawn<F, T>(f: F) -> JoinHandle<T>
630 Builder::new().spawn(f).expect("failed to spawn thread")
633 /// Gets a handle to the thread that invokes it.
637 /// Getting a handle to the current thread with `thread::current()`:
642 /// let handler = thread::Builder::new()
643 /// .name("named thread".into())
645 /// let handle = thread::current();
646 /// assert_eq!(handle.name(), Some("named thread"));
650 /// handler.join().unwrap();
652 #[stable(feature = "rust1", since = "1.0.0")]
653 pub fn current() -> Thread {
654 thread_info::current_thread().expect(
655 "use of std::thread::current() is not possible \
656 after the thread's local data has been destroyed",
660 /// Cooperatively gives up a timeslice to the OS scheduler.
662 /// This calls the underlying OS scheduler's yield primitive, signaling
663 /// that the calling thread is willing to give up its remaining timeslice
664 /// so that the OS may schedule other threads on the CPU.
666 /// A drawback of yielding in a loop is that if the OS does not have any
667 /// other ready threads to run on the current CPU, the thread will effectively
668 /// busy-wait, which wastes CPU time and energy.
670 /// Therefore, when waiting for events of interest, a programmer's first
671 /// choice should be to use synchronization devices such as [`channel`]s,
672 /// [`Condvar`]s, [`Mutex`]es or [`join`] since these primitives are
673 /// implemented in a blocking manner, giving up the CPU until the event
674 /// of interest has occurred which avoids repeated yielding.
676 /// `yield_now` should thus be used only rarely, mostly in situations where
677 /// repeated polling is required because there is no other suitable way to
678 /// learn when an event of interest has occurred.
685 /// thread::yield_now();
688 /// [`channel`]: crate::sync::mpsc
689 /// [`join`]: JoinHandle::join
690 /// [`Condvar`]: crate::sync::Condvar
691 /// [`Mutex`]: crate::sync::Mutex
692 #[stable(feature = "rust1", since = "1.0.0")]
694 imp::Thread::yield_now()
697 /// Determines whether the current thread is unwinding because of panic.
699 /// A common use of this feature is to poison shared resources when writing
700 /// unsafe code, by checking `panicking` when the `drop` is called.
702 /// This is usually not needed when writing safe code, as [`Mutex`es][Mutex]
703 /// already poison themselves when a thread panics while holding the lock.
705 /// This can also be used in multithreaded applications, in order to send a
706 /// message to other threads warning that a thread has panicked (e.g., for
707 /// monitoring purposes).
714 /// struct SomeStruct;
716 /// impl Drop for SomeStruct {
717 /// fn drop(&mut self) {
718 /// if thread::panicking() {
719 /// println!("dropped while unwinding");
721 /// println!("dropped while not unwinding");
728 /// let a = SomeStruct;
733 /// let b = SomeStruct;
738 /// [Mutex]: crate::sync::Mutex
740 #[stable(feature = "rust1", since = "1.0.0")]
741 pub fn panicking() -> bool {
742 panicking::panicking()
745 /// Puts the current thread to sleep for at least the specified amount of time.
747 /// The thread may sleep longer than the duration specified due to scheduling
748 /// specifics or platform-dependent functionality. It will never sleep less.
750 /// This function is blocking, and should not be used in `async` functions.
752 /// # Platform-specific behavior
754 /// On Unix platforms, the underlying syscall may be interrupted by a
755 /// spurious wakeup or signal handler. To ensure the sleep occurs for at least
756 /// the specified duration, this function may invoke that system call multiple
764 /// // Let's sleep for 2 seconds:
765 /// thread::sleep_ms(2000);
767 #[stable(feature = "rust1", since = "1.0.0")]
768 #[rustc_deprecated(since = "1.6.0", reason = "replaced by `std::thread::sleep`")]
769 pub fn sleep_ms(ms: u32) {
770 sleep(Duration::from_millis(ms as u64))
773 /// Puts the current thread to sleep for at least the specified amount of time.
775 /// The thread may sleep longer than the duration specified due to scheduling
776 /// specifics or platform-dependent functionality. It will never sleep less.
778 /// This function is blocking, and should not be used in `async` functions.
780 /// # Platform-specific behavior
782 /// On Unix platforms, the underlying syscall may be interrupted by a
783 /// spurious wakeup or signal handler. To ensure the sleep occurs for at least
784 /// the specified duration, this function may invoke that system call multiple
786 /// Platforms which do not support nanosecond precision for sleeping will
787 /// have `dur` rounded up to the nearest granularity of time they can sleep for.
789 /// Currently, specifying a zero duration on Unix platforms returns immediately
790 /// without invoking the underlying [`nanosleep`] syscall, whereas on Windows
791 /// platforms the underlying [`Sleep`] syscall is always invoked.
792 /// If the intention is to yield the current time-slice you may want to use
793 /// [`yield_now`] instead.
795 /// [`nanosleep`]: https://linux.die.net/man/2/nanosleep
796 /// [`Sleep`]: https://docs.microsoft.com/en-us/windows/win32/api/synchapi/nf-synchapi-sleep
801 /// use std::{thread, time};
803 /// let ten_millis = time::Duration::from_millis(10);
804 /// let now = time::Instant::now();
806 /// thread::sleep(ten_millis);
808 /// assert!(now.elapsed() >= ten_millis);
810 #[stable(feature = "thread_sleep", since = "1.4.0")]
811 pub fn sleep(dur: Duration) {
812 imp::Thread::sleep(dur)
815 /// Blocks unless or until the current thread's token is made available.
817 /// A call to `park` does not guarantee that the thread will remain parked
818 /// forever, and callers should be prepared for this possibility.
820 /// # park and unpark
822 /// Every thread is equipped with some basic low-level blocking support, via the
823 /// [`thread::park`][`park`] function and [`thread::Thread::unpark`][`unpark`]
824 /// method. [`park`] blocks the current thread, which can then be resumed from
825 /// another thread by calling the [`unpark`] method on the blocked thread's
828 /// Conceptually, each [`Thread`] handle has an associated token, which is
829 /// initially not present:
831 /// * The [`thread::park`][`park`] function blocks the current thread unless or
832 /// until the token is available for its thread handle, at which point it
833 /// atomically consumes the token. It may also return *spuriously*, without
834 /// consuming the token. [`thread::park_timeout`] does the same, but allows
835 /// specifying a maximum time to block the thread for.
837 /// * The [`unpark`] method on a [`Thread`] atomically makes the token available
838 /// if it wasn't already. Because the token is initially absent, [`unpark`]
839 /// followed by [`park`] will result in the second call returning immediately.
841 /// In other words, each [`Thread`] acts a bit like a spinlock that can be
842 /// locked and unlocked using `park` and `unpark`.
844 /// Notice that being unblocked does not imply any synchronization with someone
845 /// that unparked this thread, it could also be spurious.
846 /// For example, it would be a valid, but inefficient, implementation to make both [`park`] and
847 /// [`unpark`] return immediately without doing anything.
849 /// The API is typically used by acquiring a handle to the current thread,
850 /// placing that handle in a shared data structure so that other threads can
851 /// find it, and then `park`ing in a loop. When some desired condition is met, another
852 /// thread calls [`unpark`] on the handle.
854 /// The motivation for this design is twofold:
856 /// * It avoids the need to allocate mutexes and condvars when building new
857 /// synchronization primitives; the threads already provide basic
858 /// blocking/signaling.
860 /// * It can be implemented very efficiently on many platforms.
866 /// use std::sync::{Arc, atomic::{Ordering, AtomicBool}};
867 /// use std::time::Duration;
869 /// let flag = Arc::new(AtomicBool::new(false));
870 /// let flag2 = Arc::clone(&flag);
872 /// let parked_thread = thread::spawn(move || {
873 /// // We want to wait until the flag is set. We *could* just spin, but using
874 /// // park/unpark is more efficient.
875 /// while !flag2.load(Ordering::Acquire) {
876 /// println!("Parking thread");
878 /// // We *could* get here spuriously, i.e., way before the 10ms below are over!
879 /// // But that is no problem, we are in a loop until the flag is set anyway.
880 /// println!("Thread unparked");
882 /// println!("Flag received");
885 /// // Let some time pass for the thread to be spawned.
886 /// thread::sleep(Duration::from_millis(10));
888 /// // Set the flag, and let the thread wake up.
889 /// // There is no race condition here, if `unpark`
890 /// // happens first, `park` will return immediately.
891 /// // Hence there is no risk of a deadlock.
892 /// flag.store(true, Ordering::Release);
893 /// println!("Unpark the thread");
894 /// parked_thread.thread().unpark();
896 /// parked_thread.join().unwrap();
899 /// [`unpark`]: Thread::unpark
900 /// [`thread::park_timeout`]: park_timeout
901 #[stable(feature = "rust1", since = "1.0.0")]
903 // SAFETY: park_timeout is called on the parker owned by this thread.
905 current().inner.parker.park();
909 /// Use [`park_timeout`].
911 /// Blocks unless or until the current thread's token is made available or
912 /// the specified duration has been reached (may wake spuriously).
914 /// The semantics of this function are equivalent to [`park`] except
915 /// that the thread will be blocked for roughly no longer than `dur`. This
916 /// method should not be used for precise timing due to anomalies such as
917 /// preemption or platform differences that might not cause the maximum
918 /// amount of time waited to be precisely `ms` long.
920 /// See the [park documentation][`park`] for more detail.
921 #[stable(feature = "rust1", since = "1.0.0")]
922 #[rustc_deprecated(since = "1.6.0", reason = "replaced by `std::thread::park_timeout`")]
923 pub fn park_timeout_ms(ms: u32) {
924 park_timeout(Duration::from_millis(ms as u64))
927 /// Blocks unless or until the current thread's token is made available or
928 /// the specified duration has been reached (may wake spuriously).
930 /// The semantics of this function are equivalent to [`park`][park] except
931 /// that the thread will be blocked for roughly no longer than `dur`. This
932 /// method should not be used for precise timing due to anomalies such as
933 /// preemption or platform differences that might not cause the maximum
934 /// amount of time waited to be precisely `dur` long.
936 /// See the [park documentation][park] for more details.
938 /// # Platform-specific behavior
940 /// Platforms which do not support nanosecond precision for sleeping will have
941 /// `dur` rounded up to the nearest granularity of time they can sleep for.
945 /// Waiting for the complete expiration of the timeout:
948 /// use std::thread::park_timeout;
949 /// use std::time::{Instant, Duration};
951 /// let timeout = Duration::from_secs(2);
952 /// let beginning_park = Instant::now();
954 /// let mut timeout_remaining = timeout;
956 /// park_timeout(timeout_remaining);
957 /// let elapsed = beginning_park.elapsed();
958 /// if elapsed >= timeout {
961 /// println!("restarting park_timeout after {:?}", elapsed);
962 /// timeout_remaining = timeout - elapsed;
965 #[stable(feature = "park_timeout", since = "1.4.0")]
966 pub fn park_timeout(dur: Duration) {
967 // SAFETY: park_timeout is called on the parker owned by this thread.
969 current().inner.parker.park_timeout(dur);
973 ////////////////////////////////////////////////////////////////////////////////
975 ////////////////////////////////////////////////////////////////////////////////
977 /// A unique identifier for a running thread.
979 /// A `ThreadId` is an opaque object that has a unique value for each thread
980 /// that creates one. `ThreadId`s are not guaranteed to correspond to a thread's
981 /// system-designated identifier. A `ThreadId` can be retrieved from the [`id`]
982 /// method on a [`Thread`].
989 /// let other_thread = thread::spawn(|| {
990 /// thread::current().id()
993 /// let other_thread_id = other_thread.join().unwrap();
994 /// assert!(thread::current().id() != other_thread_id);
997 /// [`id`]: Thread::id
998 #[stable(feature = "thread_id", since = "1.19.0")]
999 #[derive(Eq, PartialEq, Clone, Copy, Hash, Debug)]
1000 pub struct ThreadId(NonZeroU64);
1003 // Generate a new unique thread ID.
1004 fn new() -> ThreadId {
1005 // It is UB to attempt to acquire this mutex reentrantly!
1006 static GUARD: mutex::StaticMutex = mutex::StaticMutex::new();
1007 static mut COUNTER: u64 = 1;
1010 let guard = GUARD.lock();
1012 // If we somehow use up all our bits, panic so that we're not
1013 // covering up subtle bugs of IDs being reused.
1014 if COUNTER == u64::MAX {
1015 drop(guard); // in case the panic handler ends up calling `ThreadId::new()`, avoid reentrant lock acquire.
1016 panic!("failed to generate unique thread ID: bitspace exhausted");
1022 ThreadId(NonZeroU64::new(id).unwrap())
1026 /// This returns a numeric identifier for the thread identified by this
1029 /// As noted in the documentation for the type itself, it is essentially an
1030 /// opaque ID, but is guaranteed to be unique for each thread. The returned
1031 /// value is entirely opaque -- only equality testing is stable. Note that
1032 /// it is not guaranteed which values new threads will return, and this may
1033 /// change across Rust versions.
1035 #[unstable(feature = "thread_id_value", issue = "67939")]
1036 pub fn as_u64(&self) -> NonZeroU64 {
1041 ////////////////////////////////////////////////////////////////////////////////
1043 ////////////////////////////////////////////////////////////////////////////////
1045 /// The internal representation of a `Thread` handle
1047 name: Option<CString>, // Guaranteed to be UTF-8
1053 #[stable(feature = "rust1", since = "1.0.0")]
1054 /// A handle to a thread.
1056 /// Threads are represented via the `Thread` type, which you can get in one of
1059 /// * By spawning a new thread, e.g., using the [`thread::spawn`][`spawn`]
1060 /// function, and calling [`thread`][`JoinHandle::thread`] on the
1062 /// * By requesting the current thread, using the [`thread::current`] function.
1064 /// The [`thread::current`] function is available even for threads not spawned
1065 /// by the APIs of this module.
1067 /// There is usually no need to create a `Thread` struct yourself, one
1068 /// should instead use a function like `spawn` to create new threads, see the
1069 /// docs of [`Builder`] and [`spawn`] for more details.
1071 /// [`thread::current`]: current
1077 // Used only internally to construct a thread object without spawning
1078 // Panics if the name contains nuls.
1079 pub(crate) fn new(name: Option<CString>) -> Thread {
1080 Thread { inner: Arc::new(Inner { name, id: ThreadId::new(), parker: Parker::new() }) }
1083 /// Atomically makes the handle's token available if it is not already.
1085 /// Every thread is equipped with some basic low-level blocking support, via
1086 /// the [`park`][park] function and the `unpark()` method. These can be
1087 /// used as a more CPU-efficient implementation of a spinlock.
1089 /// See the [park documentation][park] for more details.
1094 /// use std::thread;
1095 /// use std::time::Duration;
1097 /// let parked_thread = thread::Builder::new()
1099 /// println!("Parking thread");
1101 /// println!("Thread unparked");
1105 /// // Let some time pass for the thread to be spawned.
1106 /// thread::sleep(Duration::from_millis(10));
1108 /// println!("Unpark the thread");
1109 /// parked_thread.thread().unpark();
1111 /// parked_thread.join().unwrap();
1113 #[stable(feature = "rust1", since = "1.0.0")]
1115 pub fn unpark(&self) {
1116 self.inner.parker.unpark();
1119 /// Gets the thread's unique identifier.
1124 /// use std::thread;
1126 /// let other_thread = thread::spawn(|| {
1127 /// thread::current().id()
1130 /// let other_thread_id = other_thread.join().unwrap();
1131 /// assert!(thread::current().id() != other_thread_id);
1133 #[stable(feature = "thread_id", since = "1.19.0")]
1134 pub fn id(&self) -> ThreadId {
1138 /// Gets the thread's name.
1140 /// For more information about named threads, see
1141 /// [this module-level documentation][naming-threads].
1145 /// Threads by default have no name specified:
1148 /// use std::thread;
1150 /// let builder = thread::Builder::new();
1152 /// let handler = builder.spawn(|| {
1153 /// assert!(thread::current().name().is_none());
1156 /// handler.join().unwrap();
1159 /// Thread with a specified name:
1162 /// use std::thread;
1164 /// let builder = thread::Builder::new()
1165 /// .name("foo".into());
1167 /// let handler = builder.spawn(|| {
1168 /// assert_eq!(thread::current().name(), Some("foo"))
1171 /// handler.join().unwrap();
1174 /// [naming-threads]: ./index.html#naming-threads
1175 #[stable(feature = "rust1", since = "1.0.0")]
1176 pub fn name(&self) -> Option<&str> {
1177 self.cname().map(|s| unsafe { str::from_utf8_unchecked(s.to_bytes()) })
1180 fn cname(&self) -> Option<&CStr> {
1181 self.inner.name.as_deref()
1185 #[stable(feature = "rust1", since = "1.0.0")]
1186 impl fmt::Debug for Thread {
1187 fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
1188 f.debug_struct("Thread")
1189 .field("id", &self.id())
1190 .field("name", &self.name())
1191 .finish_non_exhaustive()
1195 ////////////////////////////////////////////////////////////////////////////////
1197 ////////////////////////////////////////////////////////////////////////////////
1199 /// A specialized [`Result`] type for threads.
1201 /// Indicates the manner in which a thread exited.
1203 /// The value contained in the `Result::Err` variant
1204 /// is the value the thread panicked with;
1205 /// that is, the argument the `panic!` macro was called with.
1206 /// Unlike with normal errors, this value doesn't implement
1207 /// the [`Error`](crate::error::Error) trait.
1209 /// Thus, a sensible way to handle a thread panic is to either:
1211 /// 1. propagate the panic with [`std::panic::resume_unwind`]
1212 /// 2. or in case the thread is intended to be a subsystem boundary
1213 /// that is supposed to isolate system-level failures,
1214 /// match on the `Err` variant and handle the panic in an appropriate way
1216 /// A thread that completes without panicking is considered to exit successfully.
1220 /// Matching on the result of a joined thread:
1223 /// use std::{fs, thread, panic};
1225 /// fn copy_in_thread() -> thread::Result<()> {
1226 /// thread::spawn(|| {
1227 /// fs::copy("foo.txt", "bar.txt").unwrap();
1232 /// match copy_in_thread() {
1233 /// Ok(_) => println!("copy succeeded"),
1234 /// Err(e) => panic::resume_unwind(e),
1239 /// [`Result`]: crate::result::Result
1240 /// [`std::panic::resume_unwind`]: crate::panic::resume_unwind
1241 #[stable(feature = "rust1", since = "1.0.0")]
1242 pub type Result<T> = crate::result::Result<T, Box<dyn Any + Send + 'static>>;
1244 // This packet is used to communicate the return value between the spawned thread
1245 // and the rest of the program. Memory is shared through the `Arc` within and there's
1246 // no need for a mutex here because synchronization happens with `join()` (the
1247 // caller will never read this packet until the thread has exited).
1249 // This packet itself is then stored into a `JoinInner` which in turns is placed
1250 // in `JoinHandle` and `JoinGuard`. Due to the usage of `UnsafeCell` we need to
1251 // manually worry about impls like Send and Sync. The type `T` should
1252 // already always be Send (otherwise the thread could not have been created) and
1253 // this type is inherently Sync because no methods take &self. Regardless,
1254 // however, we add inheriting impls for Send/Sync to this type to ensure it's
1255 // Send/Sync and that future modifications will still appropriately classify it.
1256 struct Packet<T>(Arc<UnsafeCell<Option<Result<T>>>>);
1258 unsafe impl<T: Send> Send for Packet<T> {}
1259 unsafe impl<T: Sync> Sync for Packet<T> {}
1261 /// Inner representation for JoinHandle
1262 struct JoinInner<T> {
1263 native: Option<imp::Thread>,
1268 impl<T> JoinInner<T> {
1269 fn join(&mut self) -> Result<T> {
1270 self.native.take().unwrap().join();
1271 unsafe { (*self.packet.0.get()).take().unwrap() }
1275 /// An owned permission to join on a thread (block on its termination).
1277 /// A `JoinHandle` *detaches* the associated thread when it is dropped, which
1278 /// means that there is no longer any handle to thread and no way to `join`
1281 /// Due to platform restrictions, it is not possible to [`Clone`] this
1282 /// handle: the ability to join a thread is a uniquely-owned permission.
1284 /// This `struct` is created by the [`thread::spawn`] function and the
1285 /// [`thread::Builder::spawn`] method.
1289 /// Creation from [`thread::spawn`]:
1292 /// use std::thread;
1294 /// let join_handle: thread::JoinHandle<_> = thread::spawn(|| {
1295 /// // some work here
1299 /// Creation from [`thread::Builder::spawn`]:
1302 /// use std::thread;
1304 /// let builder = thread::Builder::new();
1306 /// let join_handle: thread::JoinHandle<_> = builder.spawn(|| {
1307 /// // some work here
1311 /// A thread being detached and outliving the thread that spawned it:
1314 /// use std::thread;
1315 /// use std::time::Duration;
1317 /// let original_thread = thread::spawn(|| {
1318 /// let _detached_thread = thread::spawn(|| {
1319 /// // Here we sleep to make sure that the first thread returns before.
1320 /// thread::sleep(Duration::from_millis(10));
1321 /// // This will be called, even though the JoinHandle is dropped.
1322 /// println!("♫ Still alive ♫");
1326 /// original_thread.join().expect("The thread being joined has panicked");
1327 /// println!("Original thread is joined.");
1329 /// // We make sure that the new thread has time to run, before the main
1330 /// // thread returns.
1332 /// thread::sleep(Duration::from_millis(1000));
1335 /// [`thread::Builder::spawn`]: Builder::spawn
1336 /// [`thread::spawn`]: spawn
1337 #[stable(feature = "rust1", since = "1.0.0")]
1338 pub struct JoinHandle<T>(JoinInner<T>);
1340 #[stable(feature = "joinhandle_impl_send_sync", since = "1.29.0")]
1341 unsafe impl<T> Send for JoinHandle<T> {}
1342 #[stable(feature = "joinhandle_impl_send_sync", since = "1.29.0")]
1343 unsafe impl<T> Sync for JoinHandle<T> {}
1345 impl<T> JoinHandle<T> {
1346 /// Extracts a handle to the underlying thread.
1351 /// use std::thread;
1353 /// let builder = thread::Builder::new();
1355 /// let join_handle: thread::JoinHandle<_> = builder.spawn(|| {
1356 /// // some work here
1359 /// let thread = join_handle.thread();
1360 /// println!("thread id: {:?}", thread.id());
1362 #[stable(feature = "rust1", since = "1.0.0")]
1363 pub fn thread(&self) -> &Thread {
1367 /// Waits for the associated thread to finish.
1369 /// This function will return immediately if the associated thread has already finished.
1371 /// In terms of [atomic memory orderings], the completion of the associated
1372 /// thread synchronizes with this function returning. In other words, all
1373 /// operations performed by that thread [happen
1374 /// before](https://doc.rust-lang.org/nomicon/atomics.html#data-accesses) all
1375 /// operations that happen after `join` returns.
1377 /// If the associated thread panics, [`Err`] is returned with the parameter given
1380 /// [`Err`]: crate::result::Result::Err
1381 /// [atomic memory orderings]: crate::sync::atomic
1385 /// This function may panic on some platforms if a thread attempts to join
1386 /// itself or otherwise may create a deadlock with joining threads.
1391 /// use std::thread;
1393 /// let builder = thread::Builder::new();
1395 /// let join_handle: thread::JoinHandle<_> = builder.spawn(|| {
1396 /// // some work here
1398 /// join_handle.join().expect("Couldn't join on the associated thread");
1400 #[stable(feature = "rust1", since = "1.0.0")]
1401 pub fn join(mut self) -> Result<T> {
1406 impl<T> AsInner<imp::Thread> for JoinHandle<T> {
1407 fn as_inner(&self) -> &imp::Thread {
1408 self.0.native.as_ref().unwrap()
1412 impl<T> IntoInner<imp::Thread> for JoinHandle<T> {
1413 fn into_inner(self) -> imp::Thread {
1414 self.0.native.unwrap()
1418 #[stable(feature = "std_debug", since = "1.16.0")]
1419 impl<T> fmt::Debug for JoinHandle<T> {
1420 fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
1421 f.debug_struct("JoinHandle").finish_non_exhaustive()
1425 fn _assert_sync_and_send() {
1426 fn _assert_both<T: Send + Sync>() {}
1427 _assert_both::<JoinHandle<()>>();
1428 _assert_both::<Thread>();
1431 /// Returns an estimate of the default amount of parallelism a program should use.
1433 /// Parallelism is a resource. A given machine provides a certain capacity for
1434 /// parallelism, i.e., a bound on the number of computations it can perform
1435 /// simultaneously. This number often corresponds to the amount of CPUs or
1436 /// computer has, but it may diverge in various cases.
1438 /// Host environments such as VMs or container orchestrators may want to
1439 /// restrict the amount of parallelism made available to programs in them. This
1440 /// is often done to limit the potential impact of (unintentionally)
1441 /// resource-intensive programs on other programs running on the same machine.
1445 /// The purpose of this API is to provide an easy and portable way to query
1446 /// the default amount of parallelism the program should use. Among other things it
1447 /// does not expose information on NUMA regions, does not account for
1448 /// differences in (co)processor capabilities, and will not modify the program's
1449 /// global state in order to more accurately query the amount of available
1452 /// The value returned by this function should be considered a simplified
1453 /// approximation of the actual amount of parallelism available at any given
1454 /// time. To get a more detailed or precise overview of the amount of
1455 /// parallelism available to the program, you may wish to use
1456 /// platform-specific APIs as well. The following platform limitations currently
1457 /// apply to `available_parallelism`:
1460 /// - It may undercount the amount of parallelism available on systems with more
1461 /// than 64 logical CPUs. However, programs typically need specific support to
1462 /// take advantage of more than 64 logical CPUs, and in the absence of such
1463 /// support, the number returned by this function accurately reflects the
1464 /// number of logical CPUs the program can use by default.
1465 /// - It may overcount the amount of parallelism available on systems limited by
1466 /// process-wide affinity masks, or job object limitations.
1469 /// - It may overcount the amount of parallelism available when limited by a
1470 /// process-wide affinity mask, or when affected by cgroup limits.
1473 /// - It may overcount the amount of parallelism available when running in a VM
1474 /// with CPU usage limits (e.g. an overcommitted host).
1478 /// This function will, but is not limited to, return errors in the following
1481 /// - If the amount of parallelism is not known for the target platform.
1482 /// - If the program lacks permission to query the amount of parallelism made
1483 /// available to it.
1488 /// # #![allow(dead_code)]
1489 /// #![feature(available_parallelism)]
1490 /// use std::{io, thread};
1492 /// fn main() -> io::Result<()> {
1493 /// let count = thread::available_parallelism()?.get();
1494 /// assert!(count >= 1_usize);
1498 #[doc(alias = "available_concurrency")] // Alias for a previous name we gave this API on unstable.
1499 #[doc(alias = "hardware_concurrency")] // Alias for C++ `std::thread::hardware_concurrency`.
1500 #[doc(alias = "num_cpus")] // Alias for a popular ecosystem crate which provides similar functionality.
1501 #[unstable(feature = "available_parallelism", issue = "74479")]
1502 pub fn available_parallelism() -> io::Result<NonZeroUsize> {
1503 imp::available_parallelism()