1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
// SPDX-License-Identifier: Apache-2.0 OR MIT

/*
Fallback implementation using global locks.

This implementation uses seqlock for global locks.

This is basically based on global locks in crossbeam-utils's `AtomicCell`,
but seqlock is implemented in a way that does not depend on UB
(see comments in optimistic_read method in atomic! macro for details).

Note that we cannot use a lock per atomic type, since the in-memory representation of the atomic
type and the value type must be the same.
*/

#![cfg_attr(
    any(
        all(
            target_arch = "x86_64",
            not(portable_atomic_no_outline_atomics),
            not(any(target_env = "sgx", miri)),
        ),
        all(
            target_arch = "powerpc64",
            feature = "fallback",
            not(portable_atomic_no_outline_atomics),
            portable_atomic_outline_atomics, // TODO(powerpc64): currently disabled by default
            any(
                all(
                    target_os = "linux",
                    any(
                        target_env = "gnu",
                        all(
                            any(target_env = "musl", target_env = "ohos"),
                            not(target_feature = "crt-static"),
                        ),
                        portable_atomic_outline_atomics,
                    ),
                ),
                target_os = "android",
                target_os = "freebsd",
                all(target_os = "openbsd", portable_atomic_outline_atomics),
            ),
            not(any(miri, portable_atomic_sanitize_thread)),
        ),
        all(
            target_arch = "riscv32",
            not(any(miri, portable_atomic_sanitize_thread)),
            not(portable_atomic_no_asm),
            not(portable_atomic_pre_llvm_19),
            any(
                target_feature = "experimental-zacas",
                portable_atomic_target_feature = "experimental-zacas",
                all(
                    feature = "fallback",
                    not(portable_atomic_no_outline_atomics),
                    any(test, portable_atomic_outline_atomics), // TODO(riscv): currently disabled by default
                    any(target_os = "linux", target_os = "android"),
                ),
            ),
        ),
        all(
            target_arch = "riscv64",
            not(portable_atomic_no_asm),
            not(portable_atomic_pre_llvm_19),
            any(
                target_feature = "experimental-zacas",
                portable_atomic_target_feature = "experimental-zacas",
                all(
                    feature = "fallback",
                    not(portable_atomic_no_outline_atomics),
                    any(test, portable_atomic_outline_atomics), // TODO(riscv): currently disabled by default
                    any(target_os = "linux", target_os = "android"),
                    not(any(miri, portable_atomic_sanitize_thread)),
                ),
            ),
        ),
        all(
            target_arch = "arm",
            any(not(portable_atomic_no_asm), portable_atomic_unstable_asm),
            any(target_os = "linux", target_os = "android"),
            not(portable_atomic_no_outline_atomics),
        ),
    ),
    allow(dead_code)
)]

#[macro_use]
pub(crate) mod utils;

// Use "wide" sequence lock if the pointer width <= 32 for preventing its counter against wrap
// around.
//
// In narrow architectures (pointer width <= 16), the counter is still <= 32-bit and may be
// vulnerable to wrap around. But it's mostly okay, since in such a primitive hardware, the
// counter will not be increased that fast.
//
// Some 64-bit architectures have ABI with 32-bit pointer width (e.g., x86_64 X32 ABI,
// AArch64 ILP32 ABI, mips64 N32 ABI). On those targets, AtomicU64 is available and fast,
// so use it to implement normal sequence lock.
cfg_has_fast_atomic_64! {
    mod seq_lock;
}
cfg_no_fast_atomic_64! {
    #[path = "seq_lock_wide.rs"]
    mod seq_lock;
}

use core::{cell::UnsafeCell, mem, sync::atomic::Ordering};

use seq_lock::{SeqLock, SeqLockWriteGuard};
use utils::CachePadded;

// Some 64-bit architectures have ABI with 32-bit pointer width (e.g., x86_64 X32 ABI,
// AArch64 ILP32 ABI, mips64 N32 ABI). On those targets, AtomicU64 is fast,
// so use it to reduce chunks of byte-wise atomic memcpy.
use seq_lock::{AtomicChunk, Chunk};

// Adapted from https://github.com/crossbeam-rs/crossbeam/blob/crossbeam-utils-0.8.7/crossbeam-utils/src/atomic/atomic_cell.rs#L969-L1016.
#[inline]
#[must_use]
fn lock(addr: usize) -> &'static SeqLock {
    // The number of locks is a prime number because we want to make sure `addr % LEN` gets
    // dispersed across all locks.
    //
    // crossbeam-utils 0.8.7 uses 97 here but does not use CachePadded,
    // so the actual concurrency level will be smaller.
    const LEN: usize = 67;
    const L: CachePadded<SeqLock> = CachePadded::new(SeqLock::new());
    static LOCKS: [CachePadded<SeqLock>; LEN] = [
        L, L, L, L, L, L, L, L, L, L, L, L, L, L, L, L, L, L, L, L, L, L, L, L, L, L, L, L, L, L,
        L, L, L, L, L, L, L, L, L, L, L, L, L, L, L, L, L, L, L, L, L, L, L, L, L, L, L, L, L, L,
        L, L, L, L, L, L, L,
    ];

    // If the modulus is a constant number, the compiler will use crazy math to transform this into
    // a sequence of cheap arithmetic operations rather than using the slow modulo instruction.
    &LOCKS[addr % LEN]
}

macro_rules! atomic {
    ($atomic_type:ident, $int_type:ident, $align:literal) => {
        #[repr(C, align($align))]
        pub(crate) struct $atomic_type {
            v: UnsafeCell<$int_type>,
        }

        impl $atomic_type {
            const LEN: usize = mem::size_of::<$int_type>() / mem::size_of::<Chunk>();

            #[inline]
            unsafe fn chunks(&self) -> &[AtomicChunk; Self::LEN] {
                static_assert!($atomic_type::LEN > 1);
                static_assert!(mem::size_of::<$int_type>() % mem::size_of::<Chunk>() == 0);

                // SAFETY: the caller must uphold the safety contract for `chunks`.
                unsafe { &*(self.v.get() as *const $int_type as *const [AtomicChunk; Self::LEN]) }
            }

            #[inline]
            fn optimistic_read(&self) -> $int_type {
                // Using `MaybeUninit<[usize; Self::LEN]>` here doesn't change codegen: https://godbolt.org/z/86f8s733M
                let mut dst: [Chunk; Self::LEN] = [0; Self::LEN];
                // SAFETY:
                // - There are no threads that perform non-atomic concurrent write operations.
                // - There is no writer that updates the value using atomic operations of different granularity.
                //
                // If the atomic operation is not used here, it will cause a data race
                // when `write` performs concurrent write operation.
                // Such a data race is sometimes considered virtually unproblematic
                // in SeqLock implementations:
                //
                // - https://github.com/Amanieu/seqlock/issues/2
                // - https://github.com/crossbeam-rs/crossbeam/blob/crossbeam-utils-0.8.7/crossbeam-utils/src/atomic/atomic_cell.rs#L1111-L1116
                // - https://rust-lang.zulipchat.com/#narrow/stream/136281-t-lang.2Fwg-unsafe-code-guidelines/topic/avoiding.20UB.20due.20to.20races.20by.20discarding.20result.3F
                //
                // However, in our use case, the implementation that loads/stores value as
                // chunks of usize is enough fast and sound, so we use that implementation.
                //
                // See also atomic-memcpy crate, a generic implementation of this pattern:
                // https://github.com/taiki-e/atomic-memcpy
                let chunks = unsafe { self.chunks() };
                for i in 0..Self::LEN {
                    dst[i] = chunks[i].load(Ordering::Relaxed);
                }
                // SAFETY: integers are plain old data types so we can always transmute to them.
                unsafe { mem::transmute::<[Chunk; Self::LEN], $int_type>(dst) }
            }

            #[inline]
            fn read(&self, _guard: &SeqLockWriteGuard<'static>) -> $int_type {
                // This calls optimistic_read that can return teared value, but the resulting value
                // is guaranteed not to be teared because we hold the lock to write.
                self.optimistic_read()
            }

            #[inline]
            fn write(&self, val: $int_type, _guard: &SeqLockWriteGuard<'static>) {
                // SAFETY: integers are plain old data types so we can always transmute them to arrays of integers.
                let val = unsafe { mem::transmute::<$int_type, [Chunk; Self::LEN]>(val) };
                // SAFETY:
                // - The guard guarantees that we hold the lock to write.
                // - There are no threads that perform non-atomic concurrent read or write operations.
                //
                // See optimistic_read for the reason that atomic operations are used here.
                let chunks = unsafe { self.chunks() };
                for i in 0..Self::LEN {
                    chunks[i].store(val[i], Ordering::Relaxed);
                }
            }
        }

        // Send is implicitly implemented.
        // SAFETY: any data races are prevented by the lock and atomic operation.
        unsafe impl Sync for $atomic_type {}

        impl_default_no_fetch_ops!($atomic_type, $int_type);
        impl_default_bit_opts!($atomic_type, $int_type);
        impl $atomic_type {
            #[inline]
            pub(crate) const fn new(v: $int_type) -> Self {
                Self { v: UnsafeCell::new(v) }
            }

            #[inline]
            pub(crate) fn is_lock_free() -> bool {
                Self::IS_ALWAYS_LOCK_FREE
            }
            pub(crate) const IS_ALWAYS_LOCK_FREE: bool = false;

            #[inline]
            pub(crate) fn get_mut(&mut self) -> &mut $int_type {
                // SAFETY: the mutable reference guarantees unique ownership.
                // (UnsafeCell::get_mut requires Rust 1.50)
                unsafe { &mut *self.v.get() }
            }

            #[inline]
            #[cfg_attr(all(debug_assertions, not(portable_atomic_no_track_caller)), track_caller)]
            pub(crate) fn load(&self, order: Ordering) -> $int_type {
                crate::utils::assert_load_ordering(order);
                let lock = lock(self.v.get() as usize);

                // Try doing an optimistic read first.
                if let Some(stamp) = lock.optimistic_read() {
                    let val = self.optimistic_read();

                    if lock.validate_read(stamp) {
                        return val;
                    }
                }

                // Grab a regular write lock so that writers don't starve this load.
                let guard = lock.write();
                let val = self.read(&guard);
                // The value hasn't been changed. Drop the guard without incrementing the stamp.
                guard.abort();
                val
            }

            #[inline]
            #[cfg_attr(all(debug_assertions, not(portable_atomic_no_track_caller)), track_caller)]
            pub(crate) fn store(&self, val: $int_type, order: Ordering) {
                crate::utils::assert_store_ordering(order);
                let guard = lock(self.v.get() as usize).write();
                self.write(val, &guard)
            }

            #[inline]
            pub(crate) fn swap(&self, val: $int_type, _order: Ordering) -> $int_type {
                let guard = lock(self.v.get() as usize).write();
                let prev = self.read(&guard);
                self.write(val, &guard);
                prev
            }

            #[inline]
            #[cfg_attr(all(debug_assertions, not(portable_atomic_no_track_caller)), track_caller)]
            pub(crate) fn compare_exchange(
                &self,
                current: $int_type,
                new: $int_type,
                success: Ordering,
                failure: Ordering,
            ) -> Result<$int_type, $int_type> {
                crate::utils::assert_compare_exchange_ordering(success, failure);
                let guard = lock(self.v.get() as usize).write();
                let prev = self.read(&guard);
                if prev == current {
                    self.write(new, &guard);
                    Ok(prev)
                } else {
                    // The value hasn't been changed. Drop the guard without incrementing the stamp.
                    guard.abort();
                    Err(prev)
                }
            }

            #[inline]
            #[cfg_attr(all(debug_assertions, not(portable_atomic_no_track_caller)), track_caller)]
            pub(crate) fn compare_exchange_weak(
                &self,
                current: $int_type,
                new: $int_type,
                success: Ordering,
                failure: Ordering,
            ) -> Result<$int_type, $int_type> {
                self.compare_exchange(current, new, success, failure)
            }

            #[inline]
            pub(crate) fn fetch_add(&self, val: $int_type, _order: Ordering) -> $int_type {
                let guard = lock(self.v.get() as usize).write();
                let prev = self.read(&guard);
                self.write(prev.wrapping_add(val), &guard);
                prev
            }

            #[inline]
            pub(crate) fn fetch_sub(&self, val: $int_type, _order: Ordering) -> $int_type {
                let guard = lock(self.v.get() as usize).write();
                let prev = self.read(&guard);
                self.write(prev.wrapping_sub(val), &guard);
                prev
            }

            #[inline]
            pub(crate) fn fetch_and(&self, val: $int_type, _order: Ordering) -> $int_type {
                let guard = lock(self.v.get() as usize).write();
                let prev = self.read(&guard);
                self.write(prev & val, &guard);
                prev
            }

            #[inline]
            pub(crate) fn fetch_nand(&self, val: $int_type, _order: Ordering) -> $int_type {
                let guard = lock(self.v.get() as usize).write();
                let prev = self.read(&guard);
                self.write(!(prev & val), &guard);
                prev
            }

            #[inline]
            pub(crate) fn fetch_or(&self, val: $int_type, _order: Ordering) -> $int_type {
                let guard = lock(self.v.get() as usize).write();
                let prev = self.read(&guard);
                self.write(prev | val, &guard);
                prev
            }

            #[inline]
            pub(crate) fn fetch_xor(&self, val: $int_type, _order: Ordering) -> $int_type {
                let guard = lock(self.v.get() as usize).write();
                let prev = self.read(&guard);
                self.write(prev ^ val, &guard);
                prev
            }

            #[inline]
            pub(crate) fn fetch_max(&self, val: $int_type, _order: Ordering) -> $int_type {
                let guard = lock(self.v.get() as usize).write();
                let prev = self.read(&guard);
                self.write(core::cmp::max(prev, val), &guard);
                prev
            }

            #[inline]
            pub(crate) fn fetch_min(&self, val: $int_type, _order: Ordering) -> $int_type {
                let guard = lock(self.v.get() as usize).write();
                let prev = self.read(&guard);
                self.write(core::cmp::min(prev, val), &guard);
                prev
            }

            #[inline]
            pub(crate) fn fetch_not(&self, _order: Ordering) -> $int_type {
                let guard = lock(self.v.get() as usize).write();
                let prev = self.read(&guard);
                self.write(!prev, &guard);
                prev
            }
            #[inline]
            pub(crate) fn not(&self, order: Ordering) {
                self.fetch_not(order);
            }

            #[inline]
            pub(crate) fn fetch_neg(&self, _order: Ordering) -> $int_type {
                let guard = lock(self.v.get() as usize).write();
                let prev = self.read(&guard);
                self.write(prev.wrapping_neg(), &guard);
                prev
            }
            #[inline]
            pub(crate) fn neg(&self, order: Ordering) {
                self.fetch_neg(order);
            }

            #[inline]
            pub(crate) const fn as_ptr(&self) -> *mut $int_type {
                self.v.get()
            }
        }
    };
}

#[cfg_attr(portable_atomic_no_cfg_target_has_atomic, cfg(any(test, portable_atomic_no_atomic_64)))]
#[cfg_attr(
    not(portable_atomic_no_cfg_target_has_atomic),
    cfg(any(
        test,
        not(any(
            target_has_atomic = "64",
            all(
                target_arch = "riscv32",
                not(any(miri, portable_atomic_sanitize_thread)),
                not(portable_atomic_no_asm),
                any(
                    target_feature = "experimental-zacas",
                    portable_atomic_target_feature = "experimental-zacas",
                ),
            ),
        ))
    ))
)]
cfg_no_fast_atomic_64! {
    atomic!(AtomicI64, i64, 8);
    atomic!(AtomicU64, u64, 8);
}

atomic!(AtomicI128, i128, 16);
atomic!(AtomicU128, u128, 16);

#[cfg(test)]
mod tests {
    use super::*;

    cfg_no_fast_atomic_64! {
        test_atomic_int!(i64);
        test_atomic_int!(u64);
    }
    test_atomic_int!(i128);
    test_atomic_int!(u128);

    // load/store/swap implementation is not affected by signedness, so it is
    // enough to test only unsigned types.
    cfg_no_fast_atomic_64! {
        stress_test!(u64);
    }
    stress_test!(u128);
}