[][src]Module heapless::pool

A heap-less, interrupt-safe, lock-free memory pool (*)

(*) Currently, the implementation is only lock-free and Sync on ARMv7-M devices


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.

This example is not tested

use heapless::{pool, pool::singleton::Box};

// instantiate a memory pool of `[u8; 128]` blocks as a global singleton
    // attributes can be used here
    // #[link_section = ".ccram.A"]
    A: [u8; 128]

fn main() -> ! {
    static mut MEMORY: [u8; 1024] = [0; 1024];

    // increase the capacity of the pool by ~8 blocks

    // 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` ..

fn SysTick() {
    // claim a block of memory
    let y = A::alloc().unwrap();

    // .. do stuff with `y` ..

    // return the memory block to the pool


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' for more information). For this reason, Pool only implements Sync when compiling for ARMv7-M.

Also note that ARMv6-M architecture lacks the primitives for CAS loops so this module does not exist for thumbv6m-none-eabi.


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:

This example is not tested
fn pop(&self) -> Option<NonNull<Node<T>>> {
    let fetch_order = ..;
    let set_order = ..;

    // `self.head` has type `AtomicPtr<Node<T>>`
    let mut head = self.head.load(fetch_order);
    loop {
        if let Some(nn_head) = NonNull::new(head) {
            let next = unsafe { (*head).next };

            // <~ preempted

            match self
                .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 pops the stack twice and then pushes 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 Cortex-M devices use LL/SC (Link-local / Store-conditional) operations to implement CAS loops. Let's look at the actual disassembly of pop.

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). On the Cortex-M, STREX will always fail if the processor takes an exception between it and its corresponding LDREX operation (see 2). 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.


  1. Cortex-M3 Devices Generic User Guide (DUI 0552A), Section 2.2.7 "Synchronization primitives"
  1. ARMv7-M Architecture Reference Manual (DDI 0403E.b), Section A3.4 "Synchronization and semaphores"



Pool as a global singleton



A memory block


Unfortunate implementation detail required to use the Pool.grow_exact method


A lock-free memory pool



Initialized type state


Uninitialized type state