src/fallback.rs
199 lines · rust · 2 line annotations
// Copyright 2016 Amanieu d'Antras//// Licensed under the Apache License, Version 2.0, <LICENSE-APACHE or// http://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.use core::cmp;use core::hint;use core::num::Wrapping;use core::ops;use core::ptr;use core::sync::atomic::{AtomicUsize, Ordering};use bytemuck::NoUninit;// We use an AtomicUsize instead of an AtomicBool because it performs better// on architectures that don't have byte-sized atomics.//// We give each spinlock its own cache line to avoid false sharing.#[repr(align(64))]struct SpinLock(AtomicUsize);impl SpinLock { fn lock(&self) { while self .0 .compare_exchange_weak(0, 1, Ordering::Acquire, Ordering::Relaxed) .is_err() { while self.0.load(Ordering::Relaxed) != 0 { hint::spin_loop(); } } } fn unlock(&self) { self.0.store(0, Ordering::Release); }}// A big array of spinlocks which we use to guard atomic accesses. A spinlock is// chosen based on a hash of the address of the atomic object, which helps to// reduce contention compared to a single global lock.macro_rules! array { (@accum (0, $($_es:expr),*) -> ($($body:tt)*)) => {array!(@as_expr [$($body)*])}; (@accum (1, $($es:expr),*) -> ($($body:tt)*)) => {array!(@accum (0, $($es),*) -> ($($body)* $($es,)*))}; (@accum (2, $($es:expr),*) -> ($($body:tt)*)) => {array!(@accum (0, $($es),*) -> ($($body)* $($es,)* $($es,)*))}; (@accum (4, $($es:expr),*) -> ($($body:tt)*)) => {array!(@accum (2, $($es,)* $($es),*) -> ($($body)*))}; (@accum (8, $($es:expr),*) -> ($($body:tt)*)) => {array!(@accum (4, $($es,)* $($es),*) -> ($($body)*))}; (@accum (16, $($es:expr),*) -> ($($body:tt)*)) => {array!(@accum (8, $($es,)* $($es),*) -> ($($body)*))}; (@accum (32, $($es:expr),*) -> ($($body:tt)*)) => {array!(@accum (16, $($es,)* $($es),*) -> ($($body)*))}; (@accum (64, $($es:expr),*) -> ($($body:tt)*)) => {array!(@accum (32, $($es,)* $($es),*) -> ($($body)*))}; (@as_expr $e:expr) => {$e}; [$e:expr; $n:tt] => { array!(@accum ($n, $e) -> ()) };}static SPINLOCKS: [SpinLock; 64] = array![SpinLock(AtomicUsize::new(0)); 64];// Spinlock pointer hashing function from compiler-rt#[inline]fn lock_for_addr(addr: usize) -> &'static SpinLock { // Disregard the lowest 4 bits. We want all values that may be part of the // same memory operation to hash to the same value and therefore use the same // lock. let mut hash = addr >> 4; // Use the next bits as the basis for the hash let low = hash & (SPINLOCKS.len() - 1); // Now use the high(er) set of bits to perturb the hash, so that we don't // get collisions from atomic fields in a single object hash >>= 16; hash ^= low; // Return a pointer to the lock to use &SPINLOCKS[hash & (SPINLOCKS.len() - 1)]}#[inline]fn lock(addr: usize) -> LockGuard { let lock = lock_for_addr(addr); lock.lock(); LockGuard(lock)}struct LockGuard(&'static SpinLock);impl Drop for LockGuard { #[inline] fn drop(&mut self) { self.0.unlock(); }}Spinlock-based fallback for atomicity emulation: 64 cache-line-aligned AtomicUsize locks, hash-addressed (compiler-rt style) to reduce contention. hint::spin_loop() on contention; no fairness/backoff but adequate for short critical sections. Lock release via RAII LockGuard; correct on panic-unwind paths because critical sections perform only ptr::read/ptr::write and (for min/max) cmp::min/cmp::max after the write. Justifies concurrency-impl-correct.
#[inline]pub unsafe fn atomic_load<T>(dst: *mut T) -> T { let _l = lock(dst as usize); ptr::read(dst)}#[inline]pub unsafe fn atomic_store<T>(dst: *mut T, val: T) { let _l = lock(dst as usize); ptr::write(dst, val);}#[inline]pub unsafe fn atomic_swap<T>(dst: *mut T, val: T) -> T { let _l = lock(dst as usize); ptr::replace(dst, val)}#[inline]pub unsafe fn atomic_compare_exchange<T: NoUninit>( dst: *mut T, current: T, new: T,) -> Result<T, T> { let _l = lock(dst as usize); let result = ptr::read(dst); // compare_exchange compares with memcmp instead of Eq let a = bytemuck::bytes_of(&result); let b = bytemuck::bytes_of(¤t); if a == b { ptr::write(dst, new); Ok(result) } else { Err(result) }}#[inline]pub unsafe fn atomic_add<T: Copy>(dst: *mut T, val: T) -> Twhere Wrapping<T>: ops::Add<Output = Wrapping<T>>,{ let _l = lock(dst as usize); let result = ptr::read(dst); ptr::write(dst, (Wrapping(result) + Wrapping(val)).0); result}#[inline]pub unsafe fn atomic_sub<T: Copy>(dst: *mut T, val: T) -> Twhere Wrapping<T>: ops::Sub<Output = Wrapping<T>>,{ let _l = lock(dst as usize); let result = ptr::read(dst); ptr::write(dst, (Wrapping(result) - Wrapping(val)).0); result}#[inline]pub unsafe fn atomic_and<T: Copy + ops::BitAnd<Output = T>>(dst: *mut T, val: T) -> T { let _l = lock(dst as usize); let result = ptr::read(dst); ptr::write(dst, result & val); result}#[inline]pub unsafe fn atomic_or<T: Copy + ops::BitOr<Output = T>>(dst: *mut T, val: T) -> T { let _l = lock(dst as usize); let result = ptr::read(dst); ptr::write(dst, result | val); result}#[inline]pub unsafe fn atomic_xor<T: Copy + ops::BitXor<Output = T>>(dst: *mut T, val: T) -> T { let _l = lock(dst as usize); let result = ptr::read(dst); ptr::write(dst, result ^ val); result}#[inline]pub unsafe fn atomic_min<T: Copy + cmp::Ord>(dst: *mut T, val: T) -> T { let _l = lock(dst as usize); let result = ptr::read(dst); ptr::write(dst, cmp::min(result, val)); result}#[inline]pub unsafe fn atomic_max<T: Copy + cmp::Ord>(dst: *mut T, val: T) -> T { let _l = lock(dst as usize); let result = ptr::read(dst); ptr::write(dst, cmp::max(result, val)); result}Spinlock-backed primitives use ptr::read/ptr::write; no # Safety doc on any of the pub unsafe fns. See FINDING-1. compare_exchange correctly uses bytemuck::bytes_of for byte-wise comparison (matching native compare_exchange semantics, which compare bit patterns rather than Eq).