Skip to content

Commit 68341a9

Browse files
Joshua LockermanJoshua Lockerman
Joshua Lockerman
authored and
Joshua Lockerman
committed
Improve performance of spsc_queue and stream.
This commit makes two main changes. 1. It switches the spsc_queue node caching strategy from keeping a shared counter of the number of nodes in the cache to keeping a consumer only counter of the number of node eligible to be cached. 2. It separate the consumer and producers fields of spsc_queue and stream into a producer cache line and consumer cache line.
1 parent 0e6f4cf commit 68341a9

File tree

5 files changed

+208
-99
lines changed

5 files changed

+208
-99
lines changed

src/libstd/lib.rs

+2
Original file line numberDiff line numberDiff line change
@@ -244,6 +244,7 @@
244244
#![feature(allow_internal_unstable)]
245245
#![feature(align_offset)]
246246
#![feature(asm)]
247+
#![feature(attr_literals)]
247248
#![feature(box_syntax)]
248249
#![feature(cfg_target_has_atomic)]
249250
#![feature(cfg_target_thread_local)]
@@ -290,6 +291,7 @@
290291
#![feature(prelude_import)]
291292
#![feature(rand)]
292293
#![feature(raw)]
294+
#![feature(repr_align)]
293295
#![feature(repr_simd)]
294296
#![feature(rustc_attrs)]
295297
#![cfg_attr(not(stage0), feature(rustc_const_unstable))]

src/libstd/sync/mpsc/cache_aligned.rs

+37
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,37 @@
1+
// Copyright 2017 The Rust Project Developers. See the COPYRIGHT
2+
// file at the top-level directory of this distribution and at
3+
// http://rust-lang.org/COPYRIGHT.
4+
//
5+
// Licensed under the Apache License, Version 2.0 <LICENSE-APACHE or
6+
// http://www.apache.org/licenses/LICENSE-2.0> or the MIT license
7+
// <LICENSE-MIT or http://opensource.org/licenses/MIT>, at your
8+
// option. This file may not be copied, modified, or distributed
9+
// except according to those terms.
10+
11+
use ops::{Deref, DerefMut};
12+
13+
#[derive(Copy, Clone, Default, PartialEq, Eq, PartialOrd, Ord, Hash)]
14+
#[repr(align(64))]
15+
pub(super) struct Aligner;
16+
17+
#[derive(Copy, Clone, Default, PartialEq, Eq, PartialOrd, Ord, Hash)]
18+
pub(super) struct CacheAligned<T>(pub T, pub Aligner);
19+
20+
impl<T> Deref for CacheAligned<T> {
21+
type Target = T;
22+
fn deref(&self) -> &Self::Target {
23+
&self.0
24+
}
25+
}
26+
27+
impl<T> DerefMut for CacheAligned<T> {
28+
fn deref_mut(&mut self) -> &mut Self::Target {
29+
&mut self.0
30+
}
31+
}
32+
33+
impl<T> CacheAligned<T> {
34+
pub(super) fn new(t: T) -> Self {
35+
CacheAligned(t, Aligner)
36+
}
37+
}

src/libstd/sync/mpsc/mod.rs

+2
Original file line numberDiff line numberDiff line change
@@ -297,6 +297,8 @@ mod sync;
297297
mod mpsc_queue;
298298
mod spsc_queue;
299299

300+
mod cache_aligned;
301+
300302
/// The receiving half of Rust's [`channel`][] (or [`sync_channel`]) type.
301303
/// This half can only be owned by one thread.
302304
///

src/libstd/sync/mpsc/spsc_queue.rs

+106-55
Original file line numberDiff line numberDiff line change
@@ -22,50 +22,61 @@ use core::cell::UnsafeCell;
2222

2323
use sync::atomic::{AtomicPtr, AtomicUsize, Ordering};
2424

25+
use super::cache_aligned::CacheAligned;
26+
2527
// Node within the linked list queue of messages to send
2628
struct Node<T> {
2729
// FIXME: this could be an uninitialized T if we're careful enough, and
2830
// that would reduce memory usage (and be a bit faster).
2931
// is it worth it?
3032
value: Option<T>, // nullable for re-use of nodes
33+
cached: bool, // This node goes into the node cache
3134
next: AtomicPtr<Node<T>>, // next node in the queue
3235
}
3336

3437
/// The single-producer single-consumer queue. This structure is not cloneable,
3538
/// but it can be safely shared in an Arc if it is guaranteed that there
3639
/// is only one popper and one pusher touching the queue at any one point in
3740
/// time.
38-
pub struct Queue<T> {
41+
pub struct Queue<T, ProducerAddition=(), ConsumerAddition=()> {
3942
// consumer fields
43+
consumer: CacheAligned<Consumer<T, ConsumerAddition>>,
44+
45+
// producer fields
46+
producer: CacheAligned<Producer<T, ProducerAddition>>,
47+
}
48+
49+
struct Consumer<T, Addition> {
4050
tail: UnsafeCell<*mut Node<T>>, // where to pop from
4151
tail_prev: AtomicPtr<Node<T>>, // where to pop from
52+
cache_bound: usize, // maximum cache size
53+
cached_nodes: AtomicUsize, // number of nodes marked as cachable
54+
addition: Addition,
55+
}
4256

43-
// producer fields
57+
struct Producer<T, Addition> {
4458
head: UnsafeCell<*mut Node<T>>, // where to push to
4559
first: UnsafeCell<*mut Node<T>>, // where to get new nodes from
4660
tail_copy: UnsafeCell<*mut Node<T>>, // between first/tail
47-
48-
// Cache maintenance fields. Additions and subtractions are stored
49-
// separately in order to allow them to use nonatomic addition/subtraction.
50-
cache_bound: usize,
51-
cache_additions: AtomicUsize,
52-
cache_subtractions: AtomicUsize,
61+
addition: Addition,
5362
}
5463

55-
unsafe impl<T: Send> Send for Queue<T> { }
64+
unsafe impl<T: Send, P: Send + Sync, C: Send + Sync> Send for Queue<T, P, C> { }
5665

57-
unsafe impl<T: Send> Sync for Queue<T> { }
66+
unsafe impl<T: Send, P: Send + Sync, C: Send + Sync> Sync for Queue<T, P, C> { }
5867

5968
impl<T> Node<T> {
6069
fn new() -> *mut Node<T> {
6170
Box::into_raw(box Node {
6271
value: None,
72+
cached: false,
6373
next: AtomicPtr::new(ptr::null_mut::<Node<T>>()),
6474
})
6575
}
6676
}
6777

6878
impl<T> Queue<T> {
79+
#[cfg(test)]
6980
/// Creates a new queue.
7081
///
7182
/// This is unsafe as the type system doesn't enforce a single
@@ -84,18 +95,60 @@ impl<T> Queue<T> {
8495
/// no bound. Otherwise, the cache will never grow larger than
8596
/// `bound` (although the queue itself could be much larger.
8697
pub unsafe fn new(bound: usize) -> Queue<T> {
98+
Self::with_additions(bound, (), ())
99+
}
100+
}
101+
102+
impl<T, ProducerAddition, ConsumerAddition> Queue<T, ProducerAddition, ConsumerAddition> {
103+
104+
/// Creates a new queue. With given additional elements in the producer and
105+
/// consumer portions of the queue.
106+
///
107+
/// Due to the performance implications of cache-contention,
108+
/// we wish to keep fields used mainly by the producer on a separate cache
109+
/// line than those used by the consumer.
110+
/// Since cache lines are usually 64 bytes, it is unreasonably expensive to
111+
/// allocate one for small fields, so we allow users to insert additional
112+
/// fields into the cache lines already allocated by this for the producer
113+
/// and consumer.
114+
///
115+
/// This is unsafe as the type system doesn't enforce a single
116+
/// consumer-producer relationship. It also allows the consumer to `pop`
117+
/// items while there is a `peek` active due to all methods having a
118+
/// non-mutable receiver.
119+
///
120+
/// # Arguments
121+
///
122+
/// * `bound` - This queue implementation is implemented with a linked
123+
/// list, and this means that a push is always a malloc. In
124+
/// order to amortize this cost, an internal cache of nodes is
125+
/// maintained to prevent a malloc from always being
126+
/// necessary. This bound is the limit on the size of the
127+
/// cache (if desired). If the value is 0, then the cache has
128+
/// no bound. Otherwise, the cache will never grow larger than
129+
/// `bound` (although the queue itself could be much larger.
130+
pub unsafe fn with_additions(
131+
bound: usize,
132+
producer_addition: ProducerAddition,
133+
consumer_addition: ConsumerAddition,
134+
) -> Self {
87135
let n1 = Node::new();
88136
let n2 = Node::new();
89137
(*n1).next.store(n2, Ordering::Relaxed);
90138
Queue {
91-
tail: UnsafeCell::new(n2),
92-
tail_prev: AtomicPtr::new(n1),
93-
head: UnsafeCell::new(n2),
94-
first: UnsafeCell::new(n1),
95-
tail_copy: UnsafeCell::new(n1),
96-
cache_bound: bound,
97-
cache_additions: AtomicUsize::new(0),
98-
cache_subtractions: AtomicUsize::new(0),
139+
consumer: CacheAligned::new(Consumer {
140+
tail: UnsafeCell::new(n2),
141+
tail_prev: AtomicPtr::new(n1),
142+
cache_bound: bound,
143+
cached_nodes: AtomicUsize::new(0),
144+
addition: consumer_addition
145+
}),
146+
producer: CacheAligned::new(Producer {
147+
head: UnsafeCell::new(n2),
148+
first: UnsafeCell::new(n1),
149+
tail_copy: UnsafeCell::new(n1),
150+
addition: producer_addition
151+
}),
99152
}
100153
}
101154

@@ -109,35 +162,25 @@ impl<T> Queue<T> {
109162
assert!((*n).value.is_none());
110163
(*n).value = Some(t);
111164
(*n).next.store(ptr::null_mut(), Ordering::Relaxed);
112-
(**self.head.get()).next.store(n, Ordering::Release);
113-
*self.head.get() = n;
165+
(**self.producer.head.get()).next.store(n, Ordering::Release);
166+
*(&self.producer.head).get() = n;
114167
}
115168
}
116169

117170
unsafe fn alloc(&self) -> *mut Node<T> {
118171
// First try to see if we can consume the 'first' node for our uses.
119-
// We try to avoid as many atomic instructions as possible here, so
120-
// the addition to cache_subtractions is not atomic (plus we're the
121-
// only one subtracting from the cache).
122-
if *self.first.get() != *self.tail_copy.get() {
123-
if self.cache_bound > 0 {
124-
let b = self.cache_subtractions.load(Ordering::Relaxed);
125-
self.cache_subtractions.store(b + 1, Ordering::Relaxed);
126-
}
127-
let ret = *self.first.get();
128-
*self.first.get() = (*ret).next.load(Ordering::Relaxed);
172+
if *self.producer.first.get() != *self.producer.tail_copy.get() {
173+
let ret = *self.producer.first.get();
174+
*self.producer.0.first.get() = (*ret).next.load(Ordering::Relaxed);
129175
return ret;
130176
}
131177
// If the above fails, then update our copy of the tail and try
132178
// again.
133-
*self.tail_copy.get() = self.tail_prev.load(Ordering::Acquire);
134-
if *self.first.get() != *self.tail_copy.get() {
135-
if self.cache_bound > 0 {
136-
let b = self.cache_subtractions.load(Ordering::Relaxed);
137-
self.cache_subtractions.store(b + 1, Ordering::Relaxed);
138-
}
139-
let ret = *self.first.get();
140-
*self.first.get() = (*ret).next.load(Ordering::Relaxed);
179+
*self.producer.0.tail_copy.get() =
180+
self.consumer.tail_prev.load(Ordering::Acquire);
181+
if *self.producer.first.get() != *self.producer.tail_copy.get() {
182+
let ret = *self.producer.first.get();
183+
*self.producer.0.first.get() = (*ret).next.load(Ordering::Relaxed);
141184
return ret;
142185
}
143186
// If all of that fails, then we have to allocate a new node
@@ -153,27 +196,27 @@ impl<T> Queue<T> {
153196
// sentinel from where we should start popping from. Hence, look at
154197
// tail's next field and see if we can use it. If we do a pop, then
155198
// the current tail node is a candidate for going into the cache.
156-
let tail = *self.tail.get();
199+
let tail = *self.consumer.tail.get();
157200
let next = (*tail).next.load(Ordering::Acquire);
158201
if next.is_null() { return None }
159202
assert!((*next).value.is_some());
160203
let ret = (*next).value.take();
161204

162-
*self.tail.get() = next;
163-
if self.cache_bound == 0 {
164-
self.tail_prev.store(tail, Ordering::Release);
205+
*self.consumer.0.tail.get() = next;
206+
if self.consumer.cache_bound == 0 {
207+
self.consumer.tail_prev.store(tail, Ordering::Release);
165208
} else {
166-
// FIXME: this is dubious with overflow.
167-
let additions = self.cache_additions.load(Ordering::Relaxed);
168-
let subtractions = self.cache_subtractions.load(Ordering::Relaxed);
169-
let size = additions - subtractions;
170-
171-
if size < self.cache_bound {
172-
self.tail_prev.store(tail, Ordering::Release);
173-
self.cache_additions.store(additions + 1, Ordering::Relaxed);
209+
let cached_nodes = self.consumer.cached_nodes.load(Ordering::Relaxed);
210+
if cached_nodes < self.consumer.cache_bound && !(*tail).cached {
211+
self.consumer.cached_nodes.store(cached_nodes, Ordering::Relaxed);
212+
(*tail).cached = true;
213+
}
214+
215+
if (*tail).cached {
216+
self.consumer.tail_prev.store(tail, Ordering::Release);
174217
} else {
175-
(*self.tail_prev.load(Ordering::Relaxed))
176-
.next.store(next, Ordering::Relaxed);
218+
(*self.consumer.tail_prev.load(Ordering::Relaxed))
219+
.next.store(next, Ordering::Relaxed);
177220
// We have successfully erased all references to 'tail', so
178221
// now we can safely drop it.
179222
let _: Box<Node<T>> = Box::from_raw(tail);
@@ -194,17 +237,25 @@ impl<T> Queue<T> {
194237
// This is essentially the same as above with all the popping bits
195238
// stripped out.
196239
unsafe {
197-
let tail = *self.tail.get();
240+
let tail = *self.consumer.tail.get();
198241
let next = (*tail).next.load(Ordering::Acquire);
199242
if next.is_null() { None } else { (*next).value.as_mut() }
200243
}
201244
}
245+
246+
pub fn producer_addition(&self) -> &ProducerAddition {
247+
&self.producer.addition
248+
}
249+
250+
pub fn consumer_addition(&self) -> &ConsumerAddition {
251+
&self.consumer.addition
252+
}
202253
}
203254

204-
impl<T> Drop for Queue<T> {
255+
impl<T, ProducerAddition, ConsumerAddition> Drop for Queue<T, ProducerAddition, ConsumerAddition> {
205256
fn drop(&mut self) {
206257
unsafe {
207-
let mut cur = *self.first.get();
258+
let mut cur = *self.producer.first.get();
208259
while !cur.is_null() {
209260
let next = (*cur).next.load(Ordering::Relaxed);
210261
let _n: Box<Node<T>> = Box::from_raw(cur);

0 commit comments

Comments
 (0)