71d4e77db8
This commit is a reimplementation of `std::sync` to be based on the system-provided primitives wherever possible. The previous implementation was fundamentally built on top of channels, and as part of the runtime reform it has become clear that this is not the level of abstraction that the standard level should be providing. This rewrite aims to provide as thin of a shim as possible on top of the system primitives in order to make them safe. The overall interface of the `std::sync` module has in general not changed, but there are a few important distinctions, highlighted below: * The condition variable type, `Condvar`, has been separated out of a `Mutex`. A condition variable is now an entirely separate type. This separation benefits users who only use one mutex, and provides a clearer distinction of who's responsible for managing condition variables (the application). * All of `Condvar`, `Mutex`, and `RWLock` are now directly built on top of system primitives rather than using a custom implementation. The `Once`, `Barrier`, and `Semaphore` types are still built upon these abstractions of the system primitives. * The `Condvar`, `Mutex`, and `RWLock` types all have a new static type and constant initializer corresponding to them. These are provided primarily for C FFI interoperation, but are often useful to otherwise simply have a global lock. The types, however, will leak memory unless `destroy()` is called on them, which is clearly documented. * The `Condvar` implementation for an `RWLock` write lock has been removed. This may be added back in the future with a userspace implementation, but this commit is focused on exposing the system primitives first. * The fundamental architecture of this design is to provide two separate layers. The first layer is that exposed by `sys_common` which is a cross-platform bare-metal abstraction of the system synchronization primitives. No attempt is made at making this layer safe, and it is quite unsafe to use! It is currently not exported as part of the API of the standard library, but the stabilization of the `sys` module will ensure that these will be exposed in time. The purpose of this layer is to provide the core cross-platform abstractions if necessary to implementors. The second layer is the layer provided by `std::sync` which is intended to be the thinnest possible layer on top of `sys_common` which is entirely safe to use. There are a few concerns which need to be addressed when making these system primitives safe: * Once used, the OS primitives can never be **moved**. This means that they essentially need to have a stable address. The static primitives use `&'static self` to enforce this, and the non-static primitives all use a `Box` to provide this guarantee. * Poisoning is leveraged to ensure that invalid data is not accessible from other tasks after one has panicked. In addition to these overall blanket safety limitations, each primitive has a few restrictions of its own: * Mutexes and rwlocks can only be unlocked from the same thread that they were locked by. This is achieved through RAII lock guards which cannot be sent across threads. * Mutexes and rwlocks can only be unlocked if they were previously locked. This is achieved by not exposing an unlocking method. * A condition variable can only be waited on with a locked mutex. This is achieved by requiring a `MutexGuard` in the `wait()` method. * A condition variable cannot be used concurrently with more than one mutex. This is guaranteed by dynamically binding a condition variable to precisely one mutex for its entire lifecycle. This restriction may be able to be relaxed in the future (a mutex is unbound when no threads are waiting on the condvar), but for now it is sufficient to guarantee safety. * Condvars now support timeouts for their blocking operations. The implementation for these operations is provided by the system. Due to the modification of the `Condvar` API, removal of the `std::sync::mutex` API, and reimplementation, this is a breaking change. Most code should be fairly easy to port using the examples in the documentation of these primitives. [breaking-change] Closes #17094 Closes #18003
170 lines
5.6 KiB
Rust
170 lines
5.6 KiB
Rust
// Copyright 2014 The Rust Project Developers. See the COPYRIGHT
|
|
// file at the top-level directory of this distribution and at
|
|
// http://rust-lang.org/COPYRIGHT.
|
|
//
|
|
// Licensed under the Apache License, Version 2.0 <LICENSE-APACHE or
|
|
// http://www.apache.org/licenses/LICENSE-2.0> or the MIT license
|
|
// <LICENSE-MIT or http://opensource.org/licenses/MIT>, at your
|
|
// option. This file may not be copied, modified, or distributed
|
|
// except according to those terms.
|
|
|
|
//! A "once initialization" primitive
|
|
//!
|
|
//! This primitive is meant to be used to run one-time initialization. An
|
|
//! example use case would be for initializing an FFI library.
|
|
|
|
use int;
|
|
use mem::drop;
|
|
use sync::atomic;
|
|
use sync::{StaticMutex, MUTEX_INIT};
|
|
|
|
/// A synchronization primitive which can be used to run a one-time global
|
|
/// initialization. Useful for one-time initialization for FFI or related
|
|
/// functionality. This type can only be constructed with the `ONCE_INIT`
|
|
/// value.
|
|
///
|
|
/// # Example
|
|
///
|
|
/// ```rust
|
|
/// use std::sync::{Once, ONCE_INIT};
|
|
///
|
|
/// static START: Once = ONCE_INIT;
|
|
///
|
|
/// START.doit(|| {
|
|
/// // run initialization here
|
|
/// });
|
|
/// ```
|
|
pub struct Once {
|
|
mutex: StaticMutex,
|
|
cnt: atomic::AtomicInt,
|
|
lock_cnt: atomic::AtomicInt,
|
|
}
|
|
|
|
/// Initialization value for static `Once` values.
|
|
pub const ONCE_INIT: Once = Once {
|
|
mutex: MUTEX_INIT,
|
|
cnt: atomic::INIT_ATOMIC_INT,
|
|
lock_cnt: atomic::INIT_ATOMIC_INT,
|
|
};
|
|
|
|
impl Once {
|
|
/// Perform an initialization routine once and only once. The given closure
|
|
/// will be executed if this is the first time `doit` has been called, and
|
|
/// otherwise the routine will *not* be invoked.
|
|
///
|
|
/// This method will block the calling task if another initialization
|
|
/// routine is currently running.
|
|
///
|
|
/// When this function returns, it is guaranteed that some initialization
|
|
/// has run and completed (it may not be the closure specified).
|
|
pub fn doit(&'static self, f: ||) {
|
|
// Optimize common path: load is much cheaper than fetch_add.
|
|
if self.cnt.load(atomic::SeqCst) < 0 {
|
|
return
|
|
}
|
|
|
|
// Implementation-wise, this would seem like a fairly trivial primitive.
|
|
// The stickler part is where our mutexes currently require an
|
|
// allocation, and usage of a `Once` shouldn't leak this allocation.
|
|
//
|
|
// This means that there must be a deterministic destroyer of the mutex
|
|
// contained within (because it's not needed after the initialization
|
|
// has run).
|
|
//
|
|
// The general scheme here is to gate all future threads once
|
|
// initialization has completed with a "very negative" count, and to
|
|
// allow through threads to lock the mutex if they see a non negative
|
|
// count. For all threads grabbing the mutex, exactly one of them should
|
|
// be responsible for unlocking the mutex, and this should only be done
|
|
// once everyone else is done with the mutex.
|
|
//
|
|
// This atomicity is achieved by swapping a very negative value into the
|
|
// shared count when the initialization routine has completed. This will
|
|
// read the number of threads which will at some point attempt to
|
|
// acquire the mutex. This count is then squirreled away in a separate
|
|
// variable, and the last person on the way out of the mutex is then
|
|
// responsible for destroying the mutex.
|
|
//
|
|
// It is crucial that the negative value is swapped in *after* the
|
|
// initialization routine has completed because otherwise new threads
|
|
// calling `doit` will return immediately before the initialization has
|
|
// completed.
|
|
|
|
let prev = self.cnt.fetch_add(1, atomic::SeqCst);
|
|
if prev < 0 {
|
|
// Make sure we never overflow, we'll never have int::MIN
|
|
// simultaneous calls to `doit` to make this value go back to 0
|
|
self.cnt.store(int::MIN, atomic::SeqCst);
|
|
return
|
|
}
|
|
|
|
// If the count is negative, then someone else finished the job,
|
|
// otherwise we run the job and record how many people will try to grab
|
|
// this lock
|
|
let guard = self.mutex.lock();
|
|
if self.cnt.load(atomic::SeqCst) > 0 {
|
|
f();
|
|
let prev = self.cnt.swap(int::MIN, atomic::SeqCst);
|
|
self.lock_cnt.store(prev, atomic::SeqCst);
|
|
}
|
|
drop(guard);
|
|
|
|
// Last one out cleans up after everyone else, no leaks!
|
|
if self.lock_cnt.fetch_add(-1, atomic::SeqCst) == 1 {
|
|
unsafe { self.mutex.destroy() }
|
|
}
|
|
}
|
|
}
|
|
|
|
#[cfg(test)]
|
|
mod test {
|
|
use prelude::*;
|
|
|
|
use task;
|
|
use super::{ONCE_INIT, Once};
|
|
|
|
#[test]
|
|
fn smoke_once() {
|
|
static O: Once = ONCE_INIT;
|
|
let mut a = 0i;
|
|
O.doit(|| a += 1);
|
|
assert_eq!(a, 1);
|
|
O.doit(|| a += 1);
|
|
assert_eq!(a, 1);
|
|
}
|
|
|
|
#[test]
|
|
fn stampede_once() {
|
|
static O: Once = ONCE_INIT;
|
|
static mut run: bool = false;
|
|
|
|
let (tx, rx) = channel();
|
|
for _ in range(0u, 10) {
|
|
let tx = tx.clone();
|
|
spawn(proc() {
|
|
for _ in range(0u, 4) { task::deschedule() }
|
|
unsafe {
|
|
O.doit(|| {
|
|
assert!(!run);
|
|
run = true;
|
|
});
|
|
assert!(run);
|
|
}
|
|
tx.send(());
|
|
});
|
|
}
|
|
|
|
unsafe {
|
|
O.doit(|| {
|
|
assert!(!run);
|
|
run = true;
|
|
});
|
|
assert!(run);
|
|
}
|
|
|
|
for _ in range(0u, 10) {
|
|
rx.recv();
|
|
}
|
|
}
|
|
}
|