-
Notifications
You must be signed in to change notification settings - Fork 13.3k
Implement lock-free queue for scheduler message queue #9710
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Changes from all commits
6d5f94d
e3b6576
81a1fb2
c381d77
d0064f9
1c1caaa
c3d50ce
fd67f06
a53c141
22d1305
d199dc2
File filter
Filter by extension
Conversations
Jump to
Diff view
Diff view
There are no files selected for viewing
Original file line number | Diff line number | Diff line change |
---|---|---|
@@ -0,0 +1,211 @@ | ||
/* Multi-producer/multi-consumer bounded queue | ||
* 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. | ||
*/ | ||
|
||
use unstable::sync::UnsafeArc; | ||
use unstable::atomics::{AtomicUint,Relaxed,Release,Acquire}; | ||
use option::*; | ||
use vec; | ||
use clone::Clone; | ||
use kinds::Send; | ||
use num::{Exponential,Algebraic,Round}; | ||
|
||
struct Node<T> { | ||
sequence: AtomicUint, | ||
value: Option<T>, | ||
} | ||
|
||
struct State<T> { | ||
pad0: [u8, ..64], | ||
buffer: ~[Node<T>], | ||
mask: uint, | ||
pad1: [u8, ..64], | ||
enqueue_pos: AtomicUint, | ||
pad2: [u8, ..64], | ||
dequeue_pos: AtomicUint, | ||
pad3: [u8, ..64], | ||
} | ||
|
||
struct Queue<T> { | ||
priv state: UnsafeArc<State<T>>, | ||
} | ||
|
||
impl<T: Send> State<T> { | ||
fn with_capacity(capacity: uint) -> State<T> { | ||
let capacity = if capacity < 2 || (capacity & (capacity - 1)) != 0 { | ||
if capacity < 2 { | ||
2u | ||
} else { | ||
// use next power of 2 as capacity | ||
2f64.pow(&((capacity as f64).log2().ceil())) as uint | ||
There was a problem hiding this comment. Choose a reason for hiding this commentThe reason will be displayed to describe this comment to others. Learn more. This can use uint::next_power_of_two. |
||
} | ||
} else { | ||
capacity | ||
}; | ||
let buffer = do vec::from_fn(capacity) |i:uint| { | ||
Node{sequence:AtomicUint::new(i),value:None} | ||
}; | ||
State{ | ||
pad0: [0, ..64], | ||
buffer: buffer, | ||
mask: capacity-1, | ||
pad1: [0, ..64], | ||
enqueue_pos: AtomicUint::new(0), | ||
pad2: [0, ..64], | ||
dequeue_pos: AtomicUint::new(0), | ||
pad3: [0, ..64], | ||
} | ||
} | ||
|
||
fn push(&mut self, value: T) -> bool { | ||
let mask = self.mask; | ||
let mut pos = self.enqueue_pos.load(Relaxed); | ||
loop { | ||
let node = &mut self.buffer[pos & mask]; | ||
let seq = node.sequence.load(Acquire); | ||
let diff: int = seq as int - pos as int; | ||
|
||
if diff == 0 { | ||
let enqueue_pos = self.enqueue_pos.compare_and_swap(pos, pos+1, Relaxed); | ||
if enqueue_pos == pos { | ||
node.value = Some(value); | ||
node.sequence.store(pos+1, Release); | ||
break | ||
} else { | ||
pos = enqueue_pos; | ||
} | ||
} else if (diff < 0) { | ||
return false | ||
} else { | ||
pos = self.enqueue_pos.load(Relaxed); | ||
} | ||
} | ||
true | ||
} | ||
|
||
fn pop(&mut self) -> Option<T> { | ||
let mask = self.mask; | ||
let mut pos = self.dequeue_pos.load(Relaxed); | ||
loop { | ||
let node = &mut self.buffer[pos & mask]; | ||
let seq = node.sequence.load(Acquire); | ||
let diff: int = seq as int - (pos + 1) as int; | ||
if diff == 0 { | ||
let dequeue_pos = self.dequeue_pos.compare_and_swap(pos, pos+1, Relaxed); | ||
if dequeue_pos == pos { | ||
let value = node.value.take(); | ||
node.sequence.store(pos + mask + 1, Release); | ||
There was a problem hiding this comment. Choose a reason for hiding this commentThe reason will be displayed to describe this comment to others. Learn more. What happens when the sequence number overflows? There was a problem hiding this comment. Choose a reason for hiding this commentThe reason will be displayed to describe this comment to others. Learn more. According to the author it fails on overflow. I believe this is not a problem for 64 bits, but a serious problem for 32 bits, thus AtomicUint should not be used. I'm not sure if LLVM atomic intrinsics support 64 bit atomic operations on all supported 32 bit architectures, but if they do, an AtomicU64 could fix this. There was a problem hiding this comment. Choose a reason for hiding this commentThe reason will be displayed to describe this comment to others. Learn more. I asked @Aatch on irc if adding AtomicU64 was feasible. It seems technically possible, but not likely to happen any time soon because it involves potentially a lot of work. In this case I think we must find another MPMC lock-free queue for SleeperList. We should probably back out the change to SleeperList until then. There was a problem hiding this comment. Choose a reason for hiding this commentThe reason will be displayed to describe this comment to others. Learn more. Ok, this stuff is hard. I might be completely wrong about the fails on overflow part. I think the author was talking about overflowing the bounds of the queue, not integer overflow. Back to needs more investigation. There was a problem hiding this comment. Choose a reason for hiding this commentThe reason will be displayed to describe this comment to others. Learn more. What happens in the overflow case is the difference between seq and pos+1 becomes negative and pop fails because it thinks the queue is empty so no progress can be made after overflow has occurred. I have an idea for how to detect this case and allow it to continue to the CAS operation, but I'm not sure if it is safe and there might be other problems. I've emailed the author of the algorithm to ask for advice. There was a problem hiding this comment. Choose a reason for hiding this commentThe reason will be displayed to describe this comment to others. Learn more. I've attempt to fix this issue in #10258. |
||
return value | ||
} else { | ||
pos = dequeue_pos; | ||
} | ||
} else if diff < 0 { | ||
return None | ||
} else { | ||
pos = self.dequeue_pos.load(Relaxed); | ||
} | ||
} | ||
} | ||
} | ||
There was a problem hiding this comment. Choose a reason for hiding this commentThe reason will be displayed to describe this comment to others. Learn more. This is a pretty gorgeous data structure. |
||
|
||
impl<T: Send> Queue<T> { | ||
pub fn with_capacity(capacity: uint) -> Queue<T> { | ||
Queue{ | ||
state: UnsafeArc::new(State::with_capacity(capacity)) | ||
} | ||
} | ||
|
||
pub fn push(&mut self, value: T) -> bool { | ||
unsafe { (*self.state.get()).push(value) } | ||
} | ||
|
||
pub fn pop(&mut self) -> Option<T> { | ||
unsafe { (*self.state.get()).pop() } | ||
} | ||
} | ||
|
||
impl<T: Send> Clone for Queue<T> { | ||
fn clone(&self) -> Queue<T> { | ||
Queue { | ||
state: self.state.clone() | ||
} | ||
} | ||
} | ||
|
||
#[cfg(test)] | ||
mod tests { | ||
use prelude::*; | ||
use option::*; | ||
use task; | ||
use comm; | ||
use super::Queue; | ||
|
||
#[test] | ||
fn test() { | ||
let nthreads = 8u; | ||
let nmsgs = 1000u; | ||
let mut q = Queue::with_capacity(nthreads*nmsgs); | ||
assert_eq!(None, q.pop()); | ||
|
||
for _ in range(0, nthreads) { | ||
let (port, chan) = comm::stream(); | ||
chan.send(q.clone()); | ||
do task::spawn_sched(task::SingleThreaded) { | ||
let mut q = port.recv(); | ||
for i in range(0, nmsgs) { | ||
assert!(q.push(i)); | ||
} | ||
} | ||
} | ||
|
||
let mut completion_ports = ~[]; | ||
for _ in range(0, nthreads) { | ||
let (completion_port, completion_chan) = comm::stream(); | ||
completion_ports.push(completion_port); | ||
let (port, chan) = comm::stream(); | ||
chan.send(q.clone()); | ||
do task::spawn_sched(task::SingleThreaded) { | ||
let mut q = port.recv(); | ||
let mut i = 0u; | ||
loop { | ||
match q.pop() { | ||
None => {}, | ||
Some(_) => { | ||
i += 1; | ||
if i == nmsgs { break } | ||
} | ||
} | ||
} | ||
completion_chan.send(i); | ||
} | ||
} | ||
|
||
for completion_port in completion_ports.iter() { | ||
assert_eq!(nmsgs, completion_port.recv()); | ||
} | ||
} | ||
} |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
These can be less than 64 bytes, you can save a bit of memory by using 64 minus the size of the padded object.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Rust has no sizeof that works at compile time.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Yeah, I hit that too with my padding. Since it's hard-coded, though, you can safely change the size to 56, knowing that buffer and mask will fill the remaining space. I admit it's a pretty minor issue, I wish that my first comment on your patch had been something more helpful than this :)
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
That won't work on 32 bit machines though.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
@toffaletti you could do something with
#[cfg(target_word_size = "32")]
and#[cfg(target_word_size = "64")]
to conditionally chose a definition of the structs.There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
I'd also have to conditionally define the constructor. I chose to keep the code simple instead of saving the 16-32 bytes per cpu core.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
You're right, I should have said 60 bytes. I forgot about the possibility of the cache line boundary being in between buffer and mask.