Rewrite sync::mutex as thin layer over native mutexes

Previously, sync::mutex had to split between green and native runtime
systems and thus could not simply use the native mutex facility.

This commit rewrites sync::mutex to link directly to native mutexes; in
the future, the two will probably be coalesced into a single
module (once librustrt is pulled into libstd wholesale).
This commit is contained in:
Aaron Turon 2014-11-14 14:33:51 -08:00
parent 91a2c0d512
commit a68ec98166
3 changed files with 12 additions and 523 deletions

View File

@ -54,7 +54,6 @@ pub mod atomic;
// Concurrent data structures
mod mpsc_intrusive;
pub mod spsc_queue;
pub mod mpsc_queue;
pub mod mpmc_bounded_queue;

View File

@ -1,144 +0,0 @@
/* Copyright (c) 2010-2011 Dmitry Vyukov. All rights reserved.
* Redistribution and use in source and binary forms, with or without
* modification, are permitted provided that the following conditions are met:
*
* 1. Redistributions of source code must retain the above copyright notice,
* this list of conditions and the following disclaimer.
*
* 2. Redistributions in binary form must reproduce the above copyright
* notice, this list of conditions and the following disclaimer in the
* documentation and/or other materials provided with the distribution.
*
* THIS SOFTWARE IS PROVIDED BY DMITRY VYUKOV "AS IS" AND ANY EXPRESS OR IMPLIED
* WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF
* MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO
* EVENT SHALL DMITRY VYUKOV OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
* INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
* LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA,
* OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF
* LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
* NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE,
* EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
*
* The views and conclusions contained in the software and documentation are
* those of the authors and should not be interpreted as representing official
* policies, either expressed or implied, of Dmitry Vyukov.
*/
//! A mostly lock-free multi-producer, single consumer queue.
//!
//! This module implements an intrusive MPSC queue. This queue is incredibly
//! unsafe (due to use of unsafe pointers for nodes), and hence is not public.
#![experimental]
// http://www.1024cores.net/home/lock-free-algorithms
// /queues/intrusive-mpsc-node-based-queue
use core::prelude::*;
use core::atomic;
use core::mem;
use core::cell::UnsafeCell;
// NB: all links are done as AtomicUint instead of AtomicPtr to allow for static
// initialization.
pub struct Node<T> {
pub next: atomic::AtomicUint,
pub data: T,
}
pub struct DummyNode {
pub next: atomic::AtomicUint,
}
pub struct Queue<T> {
pub head: atomic::AtomicUint,
pub tail: UnsafeCell<*mut Node<T>>,
pub stub: DummyNode,
}
impl<T: Send> Queue<T> {
pub fn new() -> Queue<T> {
Queue {
head: atomic::AtomicUint::new(0),
tail: UnsafeCell::new(0 as *mut Node<T>),
stub: DummyNode {
next: atomic::AtomicUint::new(0),
},
}
}
pub unsafe fn push(&self, node: *mut Node<T>) {
(*node).next.store(0, atomic::Release);
let prev = self.head.swap(node as uint, atomic::AcqRel);
// Note that this code is slightly modified to allow static
// initialization of these queues with rust's flavor of static
// initialization.
if prev == 0 {
self.stub.next.store(node as uint, atomic::Release);
} else {
let prev = prev as *mut Node<T>;
(*prev).next.store(node as uint, atomic::Release);
}
}
/// You'll note that the other MPSC queue in std::sync is non-intrusive and
/// returns a `PopResult` here to indicate when the queue is inconsistent.
/// An "inconsistent state" in the other queue means that a pusher has
/// pushed, but it hasn't finished linking the rest of the chain.
///
/// This queue also suffers from this problem, but I currently haven't been
/// able to detangle when this actually happens. This code is translated
/// verbatim from the website above, and is more complicated than the
/// non-intrusive version.
///
/// Right now consumers of this queue must be ready for this fact. Just
/// because `pop` returns `None` does not mean that there is not data
/// on the queue.
pub unsafe fn pop(&self) -> Option<*mut Node<T>> {
let tail = *self.tail.get();
let mut tail = if !tail.is_null() {tail} else {
mem::transmute(&self.stub)
};
let mut next = (*tail).next(atomic::Relaxed);
if tail as uint == &self.stub as *const DummyNode as uint {
if next.is_null() {
return None;
}
*self.tail.get() = next;
tail = next;
next = (*next).next(atomic::Relaxed);
}
if !next.is_null() {
*self.tail.get() = next;
return Some(tail);
}
let head = self.head.load(atomic::Acquire) as *mut Node<T>;
if tail != head {
return None;
}
let stub = mem::transmute(&self.stub);
self.push(stub);
next = (*tail).next(atomic::Relaxed);
if !next.is_null() {
*self.tail.get() = next;
return Some(tail);
}
return None
}
}
impl<T: Send> Node<T> {
pub fn new(t: T) -> Node<T> {
Node {
data: t,
next: atomic::AtomicUint::new(0),
}
}
pub unsafe fn next(&self, ord: atomic::Ordering) -> *mut Node<T> {
mem::transmute::<uint, *mut Node<T>>(self.next.load(ord))
}
}

View File

@ -8,80 +8,22 @@
// option. This file may not be copied, modified, or distributed
// except according to those terms.
//! A proper mutex implementation regardless of the "flavor of task" which is
//! acquiring the lock.
//! A simple native mutex implementation. Warning: this API is likely
//! to change soon.
// # Implementation of Rust mutexes
//
// Most answers to the question of "how do I use a mutex" are "use pthreads",
// but for Rust this isn't quite sufficient. Green threads cannot acquire an OS
// mutex because they can context switch among many OS threads, leading to
// deadlocks with other green threads.
//
// Another problem for green threads grabbing an OS mutex is that POSIX dictates
// that unlocking a mutex on a different thread from where it was locked is
// undefined behavior. Remember that green threads can migrate among OS threads,
// so this would mean that we would have to pin green threads to OS threads,
// which is less than ideal.
//
// ## Using deschedule/reawaken
//
// We already have primitives for descheduling/reawakening tasks, so they're the
// first obvious choice when implementing a mutex. The idea would be to have a
// concurrent queue that everyone is pushed on to, and then the owner of the
// mutex is the one popping from the queue.
//
// Unfortunately, this is not very performant for native tasks. The suspected
// reason for this is that each native thread is suspended on its own condition
// variable, unique from all the other threads. In this situation, the kernel
// has no idea what the scheduling semantics are of the user program, so all of
// the threads are distributed among all cores on the system. This ends up
// having very expensive wakeups of remote cores high up in the profile when
// handing off the mutex among native tasks. On the other hand, when using an OS
// mutex, the kernel knows that all native threads are contended on the same
// mutex, so they're in theory all migrated to a single core (fast context
// switching).
//
// ## Mixing implementations
//
// From that above information, we have two constraints. The first is that
// green threads can't touch os mutexes, and the second is that native tasks
// pretty much *must* touch an os mutex.
//
// As a compromise, the queueing implementation is used for green threads and
// the os mutex is used for native threads (why not have both?). This ends up
// leading to fairly decent performance for both native threads and green
// threads on various workloads (uncontended and contended).
//
// The crux of this implementation is an atomic work which is CAS'd on many
// times in order to manage a few flags about who's blocking where and whether
// it's locked or not.
#![allow(dead_code)]
use core::prelude::*;
use self::Flavor::*;
use alloc::boxed::Box;
use core::atomic;
use core::mem;
use core::cell::UnsafeCell;
use rustrt::local::Local;
use rustrt::mutex;
use rustrt::task::{BlockedTask, Task};
use rustrt::thread::Thread;
use mpsc_intrusive as q;
pub const LOCKED: uint = 1 << 0;
pub const GREEN_BLOCKED: uint = 1 << 1;
pub const NATIVE_BLOCKED: uint = 1 << 2;
pub const BLOCKED: uint = 1 << 1;
/// A mutual exclusion primitive useful for protecting shared data
///
/// This mutex is an implementation of a lock for all flavors of tasks which may
/// be grabbing. A common problem with green threads is that they cannot grab
/// locks (if they reschedule during the lock a contender could deadlock the
/// system), but this mutex does *not* suffer this problem.
///
/// This mutex will properly block tasks waiting for the lock to become
/// available. The mutex can also be statically initialized or created via a
/// `new` constructor.
@ -107,14 +49,6 @@ pub struct Mutex {
lock: Box<StaticMutex>,
}
#[deriving(PartialEq, Show)]
enum Flavor {
Unlocked,
TryLockAcquisition,
GreenAcquisition,
NativeAcquisition,
}
/// The static mutex type is provided to allow for static allocation of mutexes.
///
/// Note that this is a separate type because using a Mutex correctly means that
@ -137,310 +71,35 @@ enum Flavor {
/// // lock is unlocked here.
/// ```
pub struct StaticMutex {
/// Current set of flags on this mutex
state: atomic::AtomicUint,
/// an OS mutex used by native threads
lock: mutex::StaticNativeMutex,
/// Type of locking operation currently on this mutex
flavor: UnsafeCell<Flavor>,
/// uint-cast of the green thread waiting for this mutex
green_blocker: UnsafeCell<uint>,
/// uint-cast of the native thread waiting for this mutex
native_blocker: UnsafeCell<uint>,
/// A concurrent mpsc queue used by green threads, along with a count used
/// to figure out when to dequeue and enqueue.
q: q::Queue<uint>,
green_cnt: atomic::AtomicUint,
}
/// An RAII implementation of a "scoped lock" of a mutex. When this structure is
/// dropped (falls out of scope), the lock will be unlocked.
#[must_use]
pub struct Guard<'a> {
lock: &'a StaticMutex,
guard: mutex::LockGuard<'a>,
}
fn lift_guard(guard: mutex::LockGuard) -> Guard {
Guard { guard: guard }
}
/// Static initialization of a mutex. This constant can be used to initialize
/// other mutex constants.
pub const MUTEX_INIT: StaticMutex = StaticMutex {
lock: mutex::NATIVE_MUTEX_INIT,
state: atomic::INIT_ATOMIC_UINT,
flavor: UnsafeCell { value: Unlocked },
green_blocker: UnsafeCell { value: 0 },
native_blocker: UnsafeCell { value: 0 },
green_cnt: atomic::INIT_ATOMIC_UINT,
q: q::Queue {
head: atomic::INIT_ATOMIC_UINT,
tail: UnsafeCell { value: 0 as *mut q::Node<uint> },
stub: q::DummyNode {
next: atomic::INIT_ATOMIC_UINT,
}
}
lock: mutex::NATIVE_MUTEX_INIT
};
impl StaticMutex {
/// Attempts to grab this lock, see `Mutex::try_lock`
pub fn try_lock<'a>(&'a self) -> Option<Guard<'a>> {
// Attempt to steal the mutex from an unlocked state.
//
// FIXME: this can mess up the fairness of the mutex, seems bad
match self.state.compare_and_swap(0, LOCKED, atomic::SeqCst) {
0 => {
// After acquiring the mutex, we can safely access the inner
// fields.
let prev = unsafe {
mem::replace(&mut *self.flavor.get(), TryLockAcquisition)
};
assert_eq!(prev, Unlocked);
Some(Guard::new(self))
}
_ => None
}
unsafe { self.lock.trylock().map(lift_guard) }
}
/// Acquires this lock, see `Mutex::lock`
pub fn lock<'a>(&'a self) -> Guard<'a> {
// First, attempt to steal the mutex from an unlocked state. The "fast
// path" needs to have as few atomic instructions as possible, and this
// one cmpxchg is already pretty expensive.
//
// FIXME: this can mess up the fairness of the mutex, seems bad
match self.try_lock() {
Some(guard) => return guard,
None => {}
}
// After we've failed the fast path, then we delegate to the different
// locking protocols for green/native tasks. This will select two tasks
// to continue further (one native, one green).
let t: Box<Task> = Local::take();
let can_block = t.can_block();
let native_bit;
if can_block {
self.native_lock(t);
native_bit = NATIVE_BLOCKED;
} else {
self.green_lock(t);
native_bit = GREEN_BLOCKED;
}
// After we've arbitrated among task types, attempt to re-acquire the
// lock (avoids a deschedule). This is very important to do in order to
// allow threads coming out of the native_lock function to try their
// best to not hit a cvar in deschedule.
let mut old = match self.state.compare_and_swap(0, LOCKED,
atomic::SeqCst) {
0 => {
let flavor = if can_block {
NativeAcquisition
} else {
GreenAcquisition
};
// We've acquired the lock, so this unsafe access to flavor is
// allowed.
unsafe { *self.flavor.get() = flavor; }
return Guard::new(self)
}
old => old,
};
// Alright, everything else failed. We need to deschedule ourselves and
// flag ourselves as waiting. Note that this case should only happen
// regularly in native/green contention. Due to try_lock and the header
// of lock stealing the lock, it's also possible for native/native
// contention to hit this location, but as less common.
let t: Box<Task> = Local::take();
t.deschedule(1, |task| {
let task = unsafe { task.cast_to_uint() };
// These accesses are protected by the respective native/green
// mutexes which were acquired above.
let prev = if can_block {
unsafe { mem::replace(&mut *self.native_blocker.get(), task) }
} else {
unsafe { mem::replace(&mut *self.green_blocker.get(), task) }
};
assert_eq!(prev, 0);
loop {
assert_eq!(old & native_bit, 0);
// If the old state was locked, then we need to flag ourselves
// as blocking in the state. If the old state was unlocked, then
// we attempt to acquire the mutex. Everything here is a CAS
// loop that'll eventually make progress.
if old & LOCKED != 0 {
old = match self.state.compare_and_swap(old,
old | native_bit,
atomic::SeqCst) {
n if n == old => return Ok(()),
n => n
};
} else {
assert_eq!(old, 0);
old = match self.state.compare_and_swap(old,
old | LOCKED,
atomic::SeqCst) {
n if n == old => {
// After acquiring the lock, we have access to the
// flavor field, and we've regained access to our
// respective native/green blocker field.
let prev = if can_block {
unsafe {
*self.native_blocker.get() = 0;
mem::replace(&mut *self.flavor.get(),
NativeAcquisition)
}
} else {
unsafe {
*self.green_blocker.get() = 0;
mem::replace(&mut *self.flavor.get(),
GreenAcquisition)
}
};
assert_eq!(prev, Unlocked);
return Err(unsafe {
BlockedTask::cast_from_uint(task)
})
}
n => n,
};
}
}
});
Guard::new(self)
}
// Tasks which can block are super easy. These tasks just call the blocking
// `lock()` function on an OS mutex
fn native_lock(&self, t: Box<Task>) {
Local::put(t);
unsafe { self.lock.lock_noguard(); }
}
fn native_unlock(&self) {
unsafe { self.lock.unlock_noguard(); }
}
fn green_lock(&self, t: Box<Task>) {
// Green threads flag their presence with an atomic counter, and if they
// fail to be the first to the mutex, they enqueue themselves on a
// concurrent internal queue with a stack-allocated node.
//
// FIXME: There isn't a cancellation currently of an enqueue, forcing
// the unlocker to spin for a bit.
if self.green_cnt.fetch_add(1, atomic::SeqCst) == 0 {
Local::put(t);
return
}
let mut node = q::Node::new(0);
t.deschedule(1, |task| {
unsafe {
node.data = task.cast_to_uint();
self.q.push(&mut node);
}
Ok(())
});
}
fn green_unlock(&self) {
// If we're the only green thread, then no need to check the queue,
// otherwise the fixme above forces us to spin for a bit.
if self.green_cnt.fetch_sub(1, atomic::SeqCst) == 1 { return }
let node;
loop {
match unsafe { self.q.pop() } {
Some(t) => { node = t; break; }
None => Thread::yield_now(),
}
}
let task = unsafe { BlockedTask::cast_from_uint((*node).data) };
task.wake().map(|t| t.reawaken());
}
fn unlock(&self) {
// Unlocking this mutex is a little tricky. We favor any task that is
// manually blocked (not in each of the separate locks) in order to help
// provide a little fairness (green threads will wake up the pending
// native thread and native threads will wake up the pending green
// thread).
//
// There's also the question of when we unlock the actual green/native
// locking halves as well. If we're waking up someone, then we can wait
// to unlock until we've acquired the task to wake up (we're guaranteed
// the mutex memory is still valid when there's contenders), but as soon
// as we don't find any contenders we must unlock the mutex, and *then*
// flag the mutex as unlocked.
//
// This flagging can fail, leading to another round of figuring out if a
// task needs to be woken, and in this case it's ok that the "mutex
// halves" are unlocked, we're just mainly dealing with the atomic state
// of the outer mutex.
let flavor = unsafe { mem::replace(&mut *self.flavor.get(), Unlocked) };
let mut state = self.state.load(atomic::SeqCst);
let mut unlocked = false;
let task;
loop {
assert!(state & LOCKED != 0);
if state & GREEN_BLOCKED != 0 {
self.unset(state, GREEN_BLOCKED);
task = unsafe {
*self.flavor.get() = GreenAcquisition;
let task = mem::replace(&mut *self.green_blocker.get(), 0);
BlockedTask::cast_from_uint(task)
};
break;
} else if state & NATIVE_BLOCKED != 0 {
self.unset(state, NATIVE_BLOCKED);
task = unsafe {
*self.flavor.get() = NativeAcquisition;
let task = mem::replace(&mut *self.native_blocker.get(), 0);
BlockedTask::cast_from_uint(task)
};
break;
} else {
assert_eq!(state, LOCKED);
if !unlocked {
match flavor {
GreenAcquisition => { self.green_unlock(); }
NativeAcquisition => { self.native_unlock(); }
TryLockAcquisition => {}
Unlocked => unreachable!(),
}
unlocked = true;
}
match self.state.compare_and_swap(LOCKED, 0, atomic::SeqCst) {
LOCKED => return,
n => { state = n; }
}
}
}
if !unlocked {
match flavor {
GreenAcquisition => { self.green_unlock(); }
NativeAcquisition => { self.native_unlock(); }
TryLockAcquisition => {}
Unlocked => unreachable!(),
}
}
task.wake().map(|t| t.reawaken());
}
/// Loops around a CAS to unset the `bit` in `state`
fn unset(&self, mut state: uint, bit: uint) {
loop {
assert!(state & bit != 0);
let new = state ^ bit;
match self.state.compare_and_swap(state, new, atomic::SeqCst) {
n if n == state => break,
n => { state = n; }
}
}
lift_guard(unsafe { self.lock.lock() })
}
/// Deallocates resources associated with this static mutex.
@ -463,12 +122,6 @@ impl Mutex {
pub fn new() -> Mutex {
Mutex {
lock: box StaticMutex {
state: atomic::AtomicUint::new(0),
flavor: UnsafeCell::new(Unlocked),
green_blocker: UnsafeCell::new(0),
native_blocker: UnsafeCell::new(0),
green_cnt: atomic::AtomicUint::new(0),
q: q::Queue::new(),
lock: unsafe { mutex::StaticNativeMutex::new() },
}
}
@ -494,25 +147,6 @@ impl Mutex {
pub fn lock<'a>(&'a self) -> Guard<'a> { self.lock.lock() }
}
impl<'a> Guard<'a> {
fn new<'b>(lock: &'b StaticMutex) -> Guard<'b> {
if cfg!(debug) {
// once we've acquired a lock, it's ok to access the flavor
assert!(unsafe { *lock.flavor.get() != Unlocked });
assert!(lock.state.load(atomic::SeqCst) & LOCKED != 0);
}
Guard { lock: lock }
}
}
#[unsafe_destructor]
impl<'a> Drop for Guard<'a> {
#[inline]
fn drop(&mut self) {
self.lock.unlock();
}
}
impl Drop for Mutex {
fn drop(&mut self) {
// This is actually safe b/c we know that there is no further usage of