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
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
//! A heap-less, interrupt-safe, lock-free memory pool (\*)
//!
//! NOTE: This module is not available on targets that do *not* support CAS operations, e.g. ARMv6-M
//!
//! (\*) Currently, the implementation is only lock-free *and* `Sync` on ARMv7-{A,R,M} & ARMv8-M
//! devices
//!
//! # Examples
//!
//! The most common way of using this pool is as a global singleton; the singleton mode gives you
//! automatic deallocation of memory blocks on `drop`.
//!
//! ``` ignore
//! #![no_main]
//! #![no_std]
//!
//! use heapless::{pool, pool::singleton::Box};
//!
//! // instantiate a memory pool of `[u8; 128]` blocks as a global singleton
//! pool!(
//!     // attributes can be used here
//!     // #[link_section = ".ccram.A"]
//!     A: [u8; 128]
//! );
//!
//! #[entry]
//! fn main() -> ! {
//!     static mut MEMORY: [u8; 1024] = [0; 1024];
//!
//!     // increase the capacity of the pool by ~8 blocks
//!     A::grow(MEMORY);
//!
//!     // claim a block of memory
//!     // note that the type is `Box<A>`, and not `Box<[u8; 128]>`
//!     // `A` is the "name" of the pool
//!     let x: Box<A, _> = A::alloc().unwrap();
//!     loop {
//!         // .. do stuff with `x` ..
//!     }
//! }
//!
//! #[exception]
//! fn SysTick() {
//!     // claim a block of memory
//!     let y = A::alloc().unwrap();
//!
//!     // .. do stuff with `y` ..
//!
//!     // return the memory block to the pool
//!     drop(y);
//! }
//! ```
//!
//! # Portability
//!
//! This pool internally uses a Treiber stack which is known to be susceptible to the ABA problem.
//! The only counter measure against the ABA problem that this implementation currently takes is
//! relying on LL/SC (Link-local / Store-conditional) instructions being used to implement CAS loops
//! on the target architecture (see section on ['Soundness'](#soundness) for more information). For
//! this reason, `Pool` only implements `Sync` when compiling for some ARM cores.
//!
//! Also note that ARMv6-M architecture lacks the primitives for CAS loops so this module does *not*
//! exist for `thumbv6m-none-eabi`.
//!
//! # Soundness
//!
//! This pool uses a Treiber stack to keep a list of free memory blocks (nodes). Each of these
//! nodes has a pointer to the next node. To claim a memory block we simply pop a node from the
//! top of the stack and use it as a memory block. The pop operation consists of swapping the
//! current head (top) node with the node below it. The Rust code for the `pop` operation is shown
//! below:
//!
//! ``` ignore
//! fn pop(&self) -> Option<NonNull<Node<T>>> {
//!     let fetch_order = ..;
//!     let set_order = ..;
//!
//!     // `self.head` has type `AtomicPtr<Node<T>>`
//!     // where `struct Node<T> { next: AtomicPtr<Node<T>>, data: UnsafeCell<T> }`
//!     let mut head = self.head.load(fetch_order);
//!     loop {
//!         if let Some(nn_head) = NonNull::new(head) {
//!             let next = unsafe { (*head).next.load(Ordering::Relaxed) };
//!
//!             // <~ preempted
//!
//!             match self
//!                 .head
//!                 .compare_exchange_weak(head, next, set_order, fetch_order)
//!             {
//!                 Ok(_) => break Some(nn_head),
//!                 // head was changed by some interrupt handler / thread
//!                 Err(new_head) => head = new_head,
//!             }
//!         } else {
//!             // stack is observed as empty
//!             break None;
//!         }
//!     }
//! }
//! ```
//!
//! In general, the `pop` operation is susceptible to the ABA problem. If this operation gets
//! preempted by some interrupt handler somewhere between the `head.load` and the
//! `compare_and_exchange_weak`, and that handler modifies the stack in such a way that the head
//! (top) of the stack remains unchanged then resuming the `pop` operation will corrupt the stack.
//!
//! An example: imagine we are doing on `pop` on stack that contains these nodes: `A -> B -> C`,
//! `A` is the head (top), `B` is next to `A` and `C` is next to `B`. The `pop` operation will do a
//! `CAS(&self.head, A, B)` operation to atomically change the head to `B` iff it currently is `A`.
//! Now, let's say a handler preempts the `pop` operation before the `CAS` operation starts and it
//! `pop`s the stack twice and then `push`es back the `A` node; now the state of the stack is `A ->
//! C`. When the original `pop` operation is resumed it will succeed in doing the `CAS` operation
//! setting `B` as the head of the stack. However, `B` was used by the handler as a memory block and
//! no longer is a valid free node. As a result the stack, and thus the allocator, is in a invalid
//! state.
//!
//! However, not all is lost because ARM devices use LL/SC (Link-local / Store-conditional)
//! operations to implement CAS loops. Let's look at the actual disassembly of `pop` for the ARM
//! Cortex-M.
//!
//! ``` text
//! 08000130 <<heapless::pool::Pool<T>>::pop>:
//!  8000130:       6802            ldr     r2, [r0, #0]
//!  8000132:       e00c            b.n     800014e <<heapless::pool::Pool<T>>::pop+0x1e>
//!  8000134:       4611            mov     r1, r2
//!  8000136:       f8d2 c000       ldr.w   ip, [r2]
//!  800013a:       e850 2f00       ldrex   r2, [r0]
//!  800013e:       428a            cmp     r2, r1
//!  8000140:       d103            bne.n   800014a <<heapless::pool::Pool<T>>::pop+0x1a>
//!  8000142:       e840 c300       strex   r3, ip, [r0]
//!  8000146:       b913            cbnz    r3, 800014e <<heapless::pool::Pool<T>>::pop+0x1e>
//!  8000148:       e004            b.n     8000154 <<heapless::pool::Pool<T>>::pop+0x24>
//!  800014a:       f3bf 8f2f       clrex
//!  800014e:       2a00            cmp     r2, #0
//!  8000150:       d1f0            bne.n   8000134 <<heapless::pool::Pool<T>>::pop+0x4>
//!  8000152:       2100            movs    r1, #0
//!  8000154:       4608            mov     r0, r1
//!  8000156:       4770            bx      lr
//! ```
//!
//! LDREX ("load exclusive") is the LL instruction, and STREX ("store exclusive") is the SC
//! instruction (see [1](#references)). On the Cortex-M, STREX will always fail if the processor
//! takes an exception between it and its corresponding LDREX operation (see [2](#references)). If
//! STREX fails then the CAS loop is retried (see instruction @ `0x8000146`). On single core
//! systems, preemption is required to run into the ABA problem and on Cortex-M devices preemption
//! always involves taking an exception. Thus the underlying LL/SC operations prevent the ABA
//! problem on Cortex-M.
//!
//! In the case of multi-core systems if any other core successfully does a STREX op on the head
//! while the current core is somewhere between LDREX and STREX then the current core will fail its
//! STREX operation.
//!
//! # x86_64 support / limitations
//!
//! *NOTE* `Pool` is only `Sync` on `x86_64` if the Cargo feature "x86-sync-pool" is enabled
//!
//! x86_64 support is a gamble. Yes, a gamble. Do you feel lucky enough to use `Pool` on x86_64?
//!
//! As it's not possible to implement *ideal* LL/SC semantics (\*) on x86_64 the architecture is
//! susceptible to the ABA problem described above. To *reduce the chances* of ABA occurring in
//! practice we use version tags (keyword: IBM ABA-prevention tags). Again, this approach does
//! *not* fix / prevent / avoid the ABA problem; it only reduces the chance of it occurring in
//! practice but the chances of it occurring are not reduced to zero.
//!
//! How we have implemented version tags: instead of using an `AtomicPtr` to link the stack `Node`s
//! we use an `AtomicUsize` where the 64-bit `usize` is always comprised of a monotonically
//! increasing 32-bit tag (higher bits) and a 32-bit signed address offset. The address of a node is
//! computed by adding the 32-bit offset to an "anchor" address (the address of a static variable
//! that lives somewhere in the `.bss` linker section). The tag is increased every time a node is
//! popped (removed) from the stack.
//!
//! To see how version tags can prevent ABA consider the example from the previous section. Let's
//! start with a stack in this state: `(~A, 0) -> (~B, 1) -> (~C, 2)`, where `~A` represents the
//! address of node A as a 32-bit offset from the "anchor" and the second tuple element (e.g. `0`)
//! indicates the version of the node. For simplicity, assume a single core system: thread T1 is
//! performing `pop` and before `CAS(&self.head, (~A, 0), (~B, 1))` is executed a context switch
//! occurs and the core resumes T2. T2 pops the stack twice and pushes A back into the stack;
//! because the `pop` operation increases the version the stack ends in the following state: `(~A,
//! 1) -> (~C, 2)`. Now if T1 is resumed the CAS operation will fail because `self.head` is `(~A,
//! 1)` and not `(~A, 0)`.
//!
//! When can version tags fail to prevent ABA? Using the previous example: if T2 performs a `push`
//! followed by a `pop` `(1 << 32) - 1` times before doing its original `pop` - `pop` - `push`
//! operation then ABA will occur because the version tag of node `A` will wraparound to its
//! original value of `0` and the CAS operation in T1 will succeed and corrupt the stack.
//!
//! It does seem unlikely that (1) a thread will perform the above operation and (2) that the above
//! operation will complete within one time slice, assuming time sliced threads. If you have thread
//! priorities then the above operation could occur during the lifetime of many high priorities
//! threads if T1 is running at low priority.
//!
//! Other implementations of version tags use more than 32 bits in their tags (e.g. "Scalable
//! Lock-Free Dynamic Memory Allocation" uses 42-bit tags in its super blocks). In theory, one could
//! use double-word CAS on x86_64 to pack a 64-bit tag and a 64-bit pointer in a double-word but
//! this CAS operation is not exposed in the standard library (and I think it's not available on
//! older x86_64 processors?)
//!
//! (\*) Apparently one can emulate proper LL/SC semantics on x86_64 using hazard pointers (?) --
//! the technique appears to be documented in "ABA Prevention Using Single-Word Instructions", which
//! is not public AFAICT -- but hazard pointers require Thread Local Storage (TLS), which is a
//! non-starter for a `no_std` library like `heapless`.
//!
//! ## x86_64 Limitations
//!
//! Because stack nodes must be located within +- 2 GB of the hidden `ANCHOR` variable, which
//! lives in the `.bss` section, `Pool` may not be able to manage static references created using
//! `Box::leak` -- these heap allocated chunks of memory may live in a very different address space.
//! When the `Pool` is unable to manage a node because of its address it will simply discard it:
//! `Pool::grow*` methods return the number of new memory blocks added to the pool; if these methods
//! return `0` it means the `Pool` is unable to manage the memory given to them.
//!
//! # References
//!
//! 1. [Cortex-M3 Devices Generic User Guide (DUI 0552A)][0], Section 2.2.7 "Synchronization
//! primitives"
//!
//! [0]: http://infocenter.arm.com/help/topic/com.arm.doc.dui0552a/DUI0552A_cortex_m3_dgug.pdf
//!
//! 2. [ARMv7-M Architecture Reference Manual (DDI 0403E.b)][1], Section A3.4 "Synchronization and
//! semaphores"
//!
//! [1]: https://static.docs.arm.com/ddi0403/eb/DDI0403E_B_armv7m_arm.pdf
//!
//! 3. "Scalable Lock-Free Dynamic Memory Allocation" Michael, Maged M.
//!
//! 4. "Hazard pointers: Safe memory reclamation for lock-free objects." Michael, Maged M.

use core::{any::TypeId, mem};
use core::{
    cmp, fmt,
    hash::{Hash, Hasher},
    marker::PhantomData,
    mem::MaybeUninit,
    ops::{Deref, DerefMut},
    ptr,
};

pub use stack::Node;
use stack::{Ptr, Stack};

pub mod singleton;
#[cfg_attr(target_arch = "x86_64", path = "cas.rs")]
#[cfg_attr(not(target_arch = "x86_64"), path = "llsc.rs")]
mod stack;

/// A lock-free memory pool
pub struct Pool<T> {
    stack: Stack<T>,

    // Current implementation is unsound on architectures that don't have LL/SC semantics so this
    // struct is not `Sync` on those platforms
    _not_send_or_sync: PhantomData<*const ()>,
}

// NOTE(any(test)) makes testing easier (no need to enable Cargo features for testing)
#[cfg(any(
    armv7a,
    armv7r,
    armv7m,
    armv8m_main,
    all(target_arch = "x86_64", feature = "x86-sync-pool"),
    test
))]
unsafe impl<T> Sync for Pool<T> {}

unsafe impl<T> Send for Pool<T> {}

impl<T> Pool<T> {
    /// Creates a new empty pool
    pub const fn new() -> Self {
        Pool {
            stack: Stack::new(),

            _not_send_or_sync: PhantomData,
        }
    }

    /// Claims a memory block from the pool
    ///
    /// Returns `None` when the pool is observed as exhausted
    ///
    /// *NOTE:* This method does *not* have bounded execution time because it contains a CAS loop
    pub fn alloc(&self) -> Option<Box<T, Uninit>> {
        if mem::size_of::<T>() == 0 {
            return Some(Box {
                node: Ptr::dangling(),
                _state: PhantomData,
            });
        }

        if let Some(node) = self.stack.try_pop() {
            Some(Box {
                node,
                _state: PhantomData,
            })
        } else {
            None
        }
    }

    /// Returns a memory block to the pool
    ///
    /// *NOTE*: `T`'s destructor (if any) will run on `value` iff `S = Init`
    ///
    /// *NOTE:* This method does *not* have bounded execution time because it contains a CAS loop
    pub fn free<S>(&self, value: Box<T, S>)
    where
        S: 'static,
    {
        if TypeId::of::<S>() == TypeId::of::<Init>() {
            unsafe {
                ptr::drop_in_place(value.node.as_ref().data.get());
            }
        }

        // no operation
        if mem::size_of::<T>() == 0 {
            return;
        }

        self.stack.push(value.node)
    }

    /// Increases the capacity of the pool
    ///
    /// This method might *not* fully utilize the given memory block due to alignment requirements.
    ///
    /// This method returns the number of *new* blocks that can be allocated.
    pub fn grow(&self, memory: &'static mut [u8]) -> usize {
        let sz = mem::size_of::<Node<T>>();

        if sz == 0 {
            // SZT use no memory so a pool of SZT always has maximum capacity
            return usize::max_value();
        }

        let mut p = memory.as_mut_ptr();
        let mut len = memory.len();

        let align = mem::align_of::<Node<T>>();
        let rem = (p as usize) % align;
        if rem != 0 {
            let offset = align - rem;

            if offset >= len {
                // slice is too small
                return 0;
            }

            p = unsafe { p.add(offset) };
            len -= offset;
        }

        let mut n = 0;
        while len >= sz {
            match () {
                #[cfg(target_arch = "x86_64")]
                () => {
                    if let Some(p) = Ptr::new(p as *mut _) {
                        self.stack.push(p);
                    }
                }

                #[cfg(not(target_arch = "x86_64"))]
                () => {
                    self.stack.push(unsafe { Ptr::new_unchecked(p as *mut _) });
                }
            }
            n += 1;

            p = unsafe { p.add(sz) };
            len -= sz;
        }

        n
    }

    /// Increases the capacity of the pool
    ///
    /// Unlike [`Pool.grow`](struct.Pool.html#method.grow) this method fully utilizes the given
    /// memory block
    pub fn grow_exact<A>(&self, memory: &'static mut MaybeUninit<A>) -> usize
    where
        A: AsMut<[Node<T>]>,
    {
        if mem::size_of::<T>() == 0 {
            return usize::max_value();
        }

        let nodes = unsafe { (*memory.as_mut_ptr()).as_mut() };
        let cap = nodes.len();
        for p in nodes {
            match () {
                #[cfg(target_arch = "x86_64")]
                () => {
                    if let Some(p) = Ptr::new(p) {
                        self.stack.push(p);
                    }
                }

                #[cfg(not(target_arch = "x86_64"))]
                () => self.stack.push(core::ptr::NonNull::from(p)),
            }
        }
        cap
    }
}

/// A memory block
pub struct Box<T, STATE = Init> {
    _state: PhantomData<STATE>,
    node: Ptr<Node<T>>,
}

impl<T> Box<T, Uninit> {
    /// Initializes this memory block
    pub fn init(self, val: T) -> Box<T, Init> {
        unsafe {
            ptr::write(self.node.as_ref().data.get(), val);
        }

        Box {
            node: self.node,
            _state: PhantomData,
        }
    }
}

/// Uninitialized type state
pub enum Uninit {}

/// Initialized type state
pub enum Init {}

unsafe impl<T, S> Send for Box<T, S> where T: Send {}

unsafe impl<T, S> Sync for Box<T, S> where T: Sync {}

unsafe impl<T> stable_deref_trait::StableDeref for Box<T> {}

impl<A, T> AsRef<[T]> for Box<A>
where
    A: AsRef<[T]>,
{
    fn as_ref(&self) -> &[T] {
        self.deref().as_ref()
    }
}

impl<A, T> AsMut<[T]> for Box<A>
where
    A: AsMut<[T]>,
{
    fn as_mut(&mut self) -> &mut [T] {
        self.deref_mut().as_mut()
    }
}

impl<T> Deref for Box<T> {
    type Target = T;

    fn deref(&self) -> &T {
        unsafe { &*self.node.as_ref().data.get() }
    }
}

impl<T> DerefMut for Box<T> {
    fn deref_mut(&mut self) -> &mut T {
        unsafe { &mut *self.node.as_ref().data.get() }
    }
}

impl<T> fmt::Debug for Box<T>
where
    T: fmt::Debug,
{
    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
        <T as fmt::Debug>::fmt(self, f)
    }
}

impl<T> fmt::Display for Box<T>
where
    T: fmt::Display,
{
    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
        <T as fmt::Display>::fmt(self, f)
    }
}

impl<T> PartialEq for Box<T>
where
    T: PartialEq,
{
    fn eq(&self, rhs: &Box<T>) -> bool {
        <T as PartialEq>::eq(self, rhs)
    }
}

impl<T> Eq for Box<T> where T: Eq {}

impl<T> PartialOrd for Box<T>
where
    T: PartialOrd,
{
    fn partial_cmp(&self, rhs: &Box<T>) -> Option<cmp::Ordering> {
        <T as PartialOrd>::partial_cmp(self, rhs)
    }
}

impl<T> Ord for Box<T>
where
    T: Ord,
{
    fn cmp(&self, rhs: &Box<T>) -> cmp::Ordering {
        <T as Ord>::cmp(self, rhs)
    }
}

impl<T> Hash for Box<T>
where
    T: Hash,
{
    fn hash<H>(&self, state: &mut H)
    where
        H: Hasher,
    {
        <T as Hash>::hash(self, state)
    }
}

#[cfg(test)]
mod tests {
    use core::{
        mem::{self, MaybeUninit},
        sync::atomic::{AtomicUsize, Ordering},
    };

    use super::{Node, Pool};

    #[test]
    fn grow() {
        static mut MEMORY: [u8; 1024] = [0; 1024];

        static POOL: Pool<[u8; 128]> = Pool::new();

        unsafe {
            POOL.grow(&mut MEMORY);
        }

        for _ in 0..7 {
            assert!(POOL.alloc().is_some());
        }
    }

    #[test]
    fn grow_exact() {
        const SZ: usize = 8;
        static mut MEMORY: MaybeUninit<[Node<[u8; 128]>; SZ]> = MaybeUninit::uninit();

        static POOL: Pool<[u8; 128]> = Pool::new();

        unsafe {
            POOL.grow_exact(&mut MEMORY);
        }

        for _ in 0..SZ {
            assert!(POOL.alloc().is_some());
        }
        assert!(POOL.alloc().is_none());
    }

    #[test]
    fn sanity() {
        const SZ: usize = 2 * mem::size_of::<Node<u8>>() - 1;
        static mut MEMORY: [u8; SZ] = [0; SZ];

        static POOL: Pool<u8> = Pool::new();

        // empty pool
        assert!(POOL.alloc().is_none());

        POOL.grow(unsafe { &mut MEMORY });

        let x = POOL.alloc().unwrap().init(0);
        assert_eq!(*x, 0);

        // pool exhausted
        assert!(POOL.alloc().is_none());

        POOL.free(x);

        // should be possible to allocate again
        assert_eq!(*POOL.alloc().unwrap().init(1), 1);
    }

    #[test]
    fn destructors() {
        static COUNT: AtomicUsize = AtomicUsize::new(0);

        struct X;

        impl X {
            fn new() -> X {
                COUNT.fetch_add(1, Ordering::Relaxed);
                X
            }
        }

        impl Drop for X {
            fn drop(&mut self) {
                COUNT.fetch_sub(1, Ordering::Relaxed);
            }
        }

        static mut MEMORY: [u8; 31] = [0; 31];

        static POOL: Pool<X> = Pool::new();

        POOL.grow(unsafe { &mut MEMORY });

        let x = POOL.alloc().unwrap().init(X::new());
        let y = POOL.alloc().unwrap().init(X::new());
        let z = POOL.alloc().unwrap().init(X::new());

        assert_eq!(COUNT.load(Ordering::Relaxed), 3);

        // this leaks memory
        drop(x);

        assert_eq!(COUNT.load(Ordering::Relaxed), 3);

        // this leaks memory
        mem::forget(y);

        assert_eq!(COUNT.load(Ordering::Relaxed), 3);

        // this runs `X` destructor
        POOL.free(z);

        assert_eq!(COUNT.load(Ordering::Relaxed), 2);
    }
}