1use core::fmt;
2use core::ops::{Deref, DerefMut};
3
4/// Pads and aligns a value to the length of a cache line.
5///
6/// In concurrent programming, sometimes it is desirable to make sure commonly accessed pieces of
7/// data are not placed into the same cache line. Updating an atomic value invalidates the whole
8/// cache line it belongs to, which makes the next access to the same cache line slower for other
9/// CPU cores. Use `CachePadded` to ensure updating one piece of data doesn't invalidate other
10/// cached data.
11///
12/// # Size and alignment
13///
14/// Cache lines are assumed to be N bytes long, depending on the architecture:
15///
16/// * On x86-64, aarch64, and powerpc64, N = 128.
17/// * On arm, mips, mips64, riscv32, riscv64, sparc, and hexagon, N = 32.
18/// * On m68k, N = 16.
19/// * On s390x, N = 256.
20/// * On all others, N = 64.
21///
22/// Note that N is just a reasonable guess and is not guaranteed to match the actual cache line
23/// length of the machine the program is running on. On modern Intel architectures, spatial
24/// prefetcher is pulling pairs of 64-byte cache lines at a time, so we pessimistically assume that
25/// cache lines are 128 bytes long.
26///
27/// The size of `CachePadded<T>` is the smallest multiple of N bytes large enough to accommodate
28/// a value of type `T`.
29///
30/// The alignment of `CachePadded<T>` is the maximum of N bytes and the alignment of `T`.
31///
32/// # Examples
33///
34/// Alignment and padding:
35///
36/// ```
37/// use crossbeam_utils::CachePadded;
38///
39/// let array = [CachePadded::new(1i8), CachePadded::new(2i8)];
40/// let addr1 = &*array[0] as *const i8 as usize;
41/// let addr2 = &*array[1] as *const i8 as usize;
42///
43/// assert!(addr2 - addr1 >= 32);
44/// assert_eq!(addr1 % 32, 0);
45/// assert_eq!(addr2 % 32, 0);
46/// ```
47///
48/// When building a concurrent queue with a head and a tail index, it is wise to place them in
49/// different cache lines so that concurrent threads pushing and popping elements don't invalidate
50/// each other's cache lines:
51///
52/// ```
53/// use crossbeam_utils::CachePadded;
54/// use std::sync::atomic::AtomicUsize;
55///
56/// struct Queue<T> {
57/// head: CachePadded<AtomicUsize>,
58/// tail: CachePadded<AtomicUsize>,
59/// buffer: *mut T,
60/// }
61/// ```
62#[derive(Clone, Copy, Default, Hash, PartialEq, Eq)]
63// Starting from Intel's Sandy Bridge, spatial prefetcher is now pulling pairs of 64-byte cache
64// lines at a time, so we have to align to 128 bytes rather than 64.
65//
66// Sources:
67// - https://www.intel.com/content/dam/www/public/us/en/documents/manuals/64-ia-32-architectures-optimization-manual.pdf
68// - https://github.com/facebook/folly/blob/1b5288e6eea6df074758f877c849b6e73bbb9fbb/folly/lang/Align.h#L107
69//
70// ARM's big.LITTLE architecture has asymmetric cores and "big" cores have 128-byte cache line size.
71//
72// Sources:
73// - https://www.mono-project.com/news/2016/09/12/arm64-icache/
74//
75// powerpc64 has 128-byte cache line size.
76//
77// Sources:
78// - https://github.com/golang/go/blob/3dd58676054223962cd915bb0934d1f9f489d4d2/src/internal/cpu/cpu_ppc64x.go#L9
79#[cfg_attr(
80 any(
81 target_arch = "x86_64",
82 target_arch = "aarch64",
83 target_arch = "powerpc64",
84 ),
85 repr(align(128))
86)]
87// arm, mips, mips64, riscv64, sparc, and hexagon have 32-byte cache line size.
88//
89// Sources:
90// - https://github.com/golang/go/blob/3dd58676054223962cd915bb0934d1f9f489d4d2/src/internal/cpu/cpu_arm.go#L7
91// - https://github.com/golang/go/blob/3dd58676054223962cd915bb0934d1f9f489d4d2/src/internal/cpu/cpu_mips.go#L7
92// - https://github.com/golang/go/blob/3dd58676054223962cd915bb0934d1f9f489d4d2/src/internal/cpu/cpu_mipsle.go#L7
93// - https://github.com/golang/go/blob/3dd58676054223962cd915bb0934d1f9f489d4d2/src/internal/cpu/cpu_mips64x.go#L9
94// - https://github.com/golang/go/blob/3dd58676054223962cd915bb0934d1f9f489d4d2/src/internal/cpu/cpu_riscv64.go#L7
95// - https://github.com/torvalds/linux/blob/3516bd729358a2a9b090c1905bd2a3fa926e24c6/arch/sparc/include/asm/cache.h#L17
96// - https://github.com/torvalds/linux/blob/3516bd729358a2a9b090c1905bd2a3fa926e24c6/arch/hexagon/include/asm/cache.h#L12
97//
98// riscv32 is assumed not to exceed the cache line size of riscv64.
99#[cfg_attr(
100 any(
101 target_arch = "arm",
102 target_arch = "mips",
103 target_arch = "mips64",
104 target_arch = "riscv32",
105 target_arch = "riscv64",
106 target_arch = "sparc",
107 target_arch = "hexagon",
108 ),
109 repr(align(32))
110)]
111// m68k has 16-byte cache line size.
112//
113// Sources:
114// - https://github.com/torvalds/linux/blob/3516bd729358a2a9b090c1905bd2a3fa926e24c6/arch/m68k/include/asm/cache.h#L9
115#[cfg_attr(target_arch = "m68k", repr(align(16)))]
116// s390x has 256-byte cache line size.
117//
118// Sources:
119// - https://github.com/golang/go/blob/3dd58676054223962cd915bb0934d1f9f489d4d2/src/internal/cpu/cpu_s390x.go#L7
120// - https://github.com/torvalds/linux/blob/3516bd729358a2a9b090c1905bd2a3fa926e24c6/arch/s390/include/asm/cache.h#L13
121#[cfg_attr(target_arch = "s390x", repr(align(256)))]
122// x86, wasm, and sparc64 have 64-byte cache line size.
123//
124// Sources:
125// - https://github.com/golang/go/blob/dda2991c2ea0c5914714469c4defc2562a907230/src/internal/cpu/cpu_x86.go#L9
126// - https://github.com/golang/go/blob/3dd58676054223962cd915bb0934d1f9f489d4d2/src/internal/cpu/cpu_wasm.go#L7
127// - https://github.com/torvalds/linux/blob/3516bd729358a2a9b090c1905bd2a3fa926e24c6/arch/sparc/include/asm/cache.h#L19
128//
129// All others are assumed to have 64-byte cache line size.
130#[cfg_attr(
131 not(any(
132 target_arch = "x86_64",
133 target_arch = "aarch64",
134 target_arch = "powerpc64",
135 target_arch = "arm",
136 target_arch = "mips",
137 target_arch = "mips64",
138 target_arch = "riscv32",
139 target_arch = "riscv64",
140 target_arch = "sparc",
141 target_arch = "hexagon",
142 target_arch = "m68k",
143 target_arch = "s390x",
144 )),
145 repr(align(64))
146)]
147pub struct CachePadded<T> {
148 value: T,
149}
150
151unsafe impl<T: Send> Send for CachePadded<T> {}
152unsafe impl<T: Sync> Sync for CachePadded<T> {}
153
154impl<T> CachePadded<T> {
155 /// Pads and aligns a value to the length of a cache line.
156 ///
157 /// # Examples
158 ///
159 /// ```
160 /// use crossbeam_utils::CachePadded;
161 ///
162 /// let padded_value = CachePadded::new(1);
163 /// ```
164 pub const fn new(t: T) -> CachePadded<T> {
165 CachePadded::<T> { value: t }
166 }
167
168 /// Returns the inner value.
169 ///
170 /// # Examples
171 ///
172 /// ```
173 /// use crossbeam_utils::CachePadded;
174 ///
175 /// let padded_value = CachePadded::new(7);
176 /// let value = padded_value.into_inner();
177 /// assert_eq!(value, 7);
178 /// ```
179 pub fn into_inner(self) -> T {
180 self.value
181 }
182}
183
184impl<T> Deref for CachePadded<T> {
185 type Target = T;
186
187 fn deref(&self) -> &T {
188 &self.value
189 }
190}
191
192impl<T> DerefMut for CachePadded<T> {
193 fn deref_mut(&mut self) -> &mut T {
194 &mut self.value
195 }
196}
197
198impl<T: fmt::Debug> fmt::Debug for CachePadded<T> {
199 fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
200 f&mut DebugStruct<'_, '_>.debug_struct("CachePadded")
201 .field(name:"value", &self.value)
202 .finish()
203 }
204}
205
206impl<T> From<T> for CachePadded<T> {
207 fn from(t: T) -> Self {
208 CachePadded::new(t)
209 }
210}
211