Skip to content

Instantly share code, notes, and snippets.

Embed
What would you like to do?
Memory<T> API documentation and samples

Memory<T> API documentation and samples

This document describes the APIs of Memory<T>, IMemoryOwner<T>, and MemoryManager<T> and their relationships to each other.

See also the Memory<T> usage guidelines document for background information.

First, a brief summary of the basic types

  • Memory<T> is the basic type that represents a contiguous buffer. This type is a struct, which means that developers cannot subclass it and override the implementation. The basic implementation of the type is aware of contigious memory buffers backed by T[] and System.String (in the case of ReadOnlyMemory<char>).

  • IMemoryOwner<T> is the interface that controls the lifetime of Memory<T> instances. The interface implementation how to create and destroy these instances. Instances of this interface are returned by MemoryPool<T>.Rent.

  • MemoryPool<T> is the abstract base class that represents a memory pool, where IMemoryOwner<T> instances can be rented from the pool (via MemoryPool<T>.Rent) and released back to the pool (via IMemoryOwner<T>.Dispose).

  • MemoryManager<T> is the abstract base class used to replace the implementation of Memory<T>. This class is used to extend the knowledge of types that Memory<T> knows about; e.g., to allow Memory<T> to be backed by a SafeHandle. This type is intended for advanced scenarios; most developers will not need to use it.

Some simple examples using built-in types

Say you have a method that needs to use a temporary buffer, but you don't want to incur the cost of an allocation. One way to write this code is to use the existing ArrayPool<T> type.

// workhorse method
static void DoSomething(Memory<byte> scratchBuffer);

public void UserCode()
{
    byte[] rentedArray = ArrayPool<byte>.Shared.Rent(...);
    try {
        DoSomething(rentedArray); // T[] implicitly converted to Memory<T>
    } finally {
        if (rentedArray != null) {
            ArrayPool<byte>.Shared.Return(rentedArray);
        }
    }
}

The above method is almost always allocation-free. When the DoSomething method is finished with the buffer, the buffer is returned back to the shared ArrayPool<byte>.

This method can be replaced with the following simpler code, using MemoryPool<T> instead of ArrayPool<T>.

// workhorse method
static void DoSomething(Memory<byte> scratchBuffer);

public void UserCode()
{
    using (var rental = MemoryPool<byte>.Shared.Rent(...))
    {
        DoSomething(rental.Memory);
    }
}

In fact, you could even abstract away the particular MemoryPool<T> used by this method, allowing the application to use a custom memory pool that is backed by native memory or some other custom implementation.

// workhorse method
static void DoSomething(Memory<byte> scratchBuffer);

public void UserClass : IDisposable
{
    private readonly MemoryPool<T> _pool;

    public UserClass(MemoryPool<T> pool)
    {
        _pool = pool;
    }

    public void UserCode()
    {
        using (var rental = pool.Rent(...))
        {
            DoSomething(rental.Memory);
        }
    }

    void IDisposable.Dispose()
    {
        _pool.Dispose();
    }
}

Creating custom pools

Say that as a security measure, you want to implement a custom MemoryPool<T> that has the following semantics.

  1. You use a different pool than the one used by ArrayPool<T>.Shared to limit secret exposure to the rest of the application.

  2. Any buffers returned to the pool should be cleared before they're next rented out.

The implementation of this custom memory pool might look as follows.

public class CustomMemoryPool<T> : MemoryPool<T>
{
    private readonly ArrayPool<T> _arrayPool = ArrayPool<T>.Create();

    public override int MaxBufferSize => Int32.MaxValue;

    protected override void Dispose(bool disposing)
    {
        if (disposing)
        {
            _arrayPool.Dispose();
        }
    }

    public override IMemoryOwner<T> Rent(int minBufferSize)
    {
        return new CustomMemoryOwner(_arrayPool, minBufferSize);
    }

    private sealed class CustomMemoryOwner : IMemoryOwner<T>
    {
        private readonly ArrayPool<T> _pool;
        private T[] _array;

        public CustomMemoryOwner(ArrayPool<T> pool, int minBufferSize)
        {
            _pool = pool;
            _array = pool.Rent(minBufferSize);
        }

        public Memory<T> Memory
        {
            get
            {
                if (_array == null)
                    throw new ObjectDisposedException(...);

                return new Memory<T>(_array);
            }
        }

        public void Dispose()
        {
            if (_array != null)
            {
                ((Span<T>)_array).Clear();
                _pool.Return(_array);
                _array = null;
            }
        }
    }
}

// workhorse method
static void DoSomething(Memory<byte> scratchBuffer);

static readonly MemoryPool<byte> _customPool = new CustomMemoryPool<byte>();

public void UserCode()
{
    // use custom pool instead of default pool
    using (var rental = _customPool.Rent(...))
    {
        DoSomething(rental.Memory);
    }
}

As a second example, consider the behavior of the existing MemoryPool<T>.Shared.Rent method. This method necessarily always allocates when called. Even though it's not allocating the full buffer during each call to Rent, it is allocating a new lightweight IMemoryOwner<T> wrapper on each call.

Let's say we want to make a custom memory pool which is equivalent to the built-in shared pool but which doesn't allocate on each call to Rent. An implementation of such a pool might look as follows.

public abstract class NonAllocatingPool<T> : MemoryPool<T>
{
    public override int MaxBufferSize => Int32.MaxValue;

    public static new NonAllocatingPool<T>.Impl Shared { get; } = new NonAllocatingPool<T>.Impl();

    protected override void Dispose(bool disposing) { }

    public override IMemoryOwner<T> Rent(int minBufferSize) => RentCore(minBufferSize);

    private Rental RentCore(int minBufferSize) => new MemoryOwner(minBufferSize);

    public sealed class Impl : NonAllocatingPool<T>
    {
        // Typed to return the actual type rather than the
        // interface to avoid boxing, like how List<T>.GetEnumerator()
        // returns List<T>.Enumerator instead of IEnumerator<T>.
        public new Rental Rent(int minBufferSize) => RentCore(minBufferSize);
    }

    // Struct implements the interface so it can be boxed if necessary.
    public struct Rental : IMemoryOwner<T>
    {
        private T[] _array;

        public Rental(int minBufferSize)
        {
            _array = ArrayPool<T>.Shared.Rent(minBufferSize);
        }

        public Memory<T> Memory
        {
            get
            {
                if (_array == null)
                    throw new ObjectDisposedException(...);

                return new Memory<T>(_array);
            }
        }

        public void Dispose()
        {
            if (_array != null)
            {
                ArrayPool<T>.Shared.Return(_array);
                _array = null;
            }
        }
    }
}

// workhorse method
static void DoSomething(Memory<byte> scratchBuffer);

public void UserCode()
{
    // Use non-allocating pool instead of default pool.
    // n.b. it's best to use a 'using' block to avoid inadvertently
    // making copies of the rental struct and exposing the application
    // to double-releasing the rental.
    using (var rental = NonAllocatingPool<byte>.Shared.Rent(...))
    {
        DoSomething(rental.Memory);
    }
}

As shown in these examples, creating a custom pool involves subclassing the MemoryPool<T> type and implementing the IMemoryOwner<T> interface. In particular, these scenarios do not involve subclassing the MemoryManager<T> type.

Replacing the Memory<T> implementation

Rarely, the developer may need to replace the entire implementation of Memory<T> in order to get it to recognize new categories of contiguous buffers. One example would be the scenario where you want to back Memory<T> by an existing SafeHandle, which is not natively supported out-of-box.

To enable these scenarios, subclass the MemoryManager<T> type. This allows replacing the implementation of Memory<T>'s Span property getter and its Pin method. These are advanced scenarios and should only be performed by developers comfortable with manual memory management techniques.

Your SafeHandle-backed MemoryManager<T> type might look like the below.

public unsafe class SafeHandleMemoryManager<T> : MemoryManager<T>
{
    private int _disposeCalled = 0;
    private readonly int _elementCount;
    private readonly SafeHandle _handle;

    public SafeHandleMemoryManager(SafeHandle handle, int elementCount)
    {
        // Important: Add a reference to the SafeHandle so that it's not
        // released without a corresponding call to our Dispose method.
        // It's better for us to leak (due to the owner forgetting to
        // call Dispose) than it is for the GC to reclaim the handle while
        // it might still be in use, potentially leading to heap corruption.

        bool unused = false;
        handle.DangerousAddRef(ref unused);

        _handle = handle;
        _elementCount = elementCount;
    }

    public override int Length => _elementCount;

    public override bool Dispose(bool disposing)
    {
        if (disposing && (Interlocked.Exchange(ref _disposeCalled, 1) == 0))
        {
            _handle.DangerousRelease(); // undo DangerousAddRef from ctor
            _handle.Dispose();
        }
    }

    public override Span<T> GetSpan()
    {
        // n.b. The check below *cannot* detect all use-after-free cases
        // where an incorrectly written consumer attempts to utilize this
        // Memory<T> instance after its lease has ended. The safety of a
        // Memory<T> instance backed by an unmanaged pointer is ultimately
        // dependent on the owner and the consumers following the provided
        // guidance correctly.

        if (_handle.IsClosed)
            throw new ObjectDisposedException(...);

        return new Span<T>((void*)_handle.DangerousGetHandle(), _elementCount);
    }

    public override MemoryHandle Pin(int elementIndex)
    {
        // Check for negative elementIndex and elementIndex greater
        // than our element count. By convention we allow an element
        // index equal to the element count, which returns a pointer
        // to the end of the buffer but which the caller should not
        // attempt to read from or write to.

        if ((uint)elementIndex > (uint)_elementCount)
            throw new ArgumentException("Invalid index.");

        bool unused = false;
        _handle.DangerousAddRef(ref unused); // will be released in Unpin

        // n.b. The Unsafe.Add method below does the correct computation
        // automatically based on the bitness of the current process. If
        // you need to perform the calculation manually by mutiplying
        // sizeof(T) by the element index, be sure to extend the arguments
        // to 64 bits *before* the multiplication if you're running in a
        // 64-bit process, otherwise you could be subject to integer overflow.

        byte* pbData = Unsafe.Add<T>((void*)_handle.DangerousGetHandle(), elementIndex);
        return new MemoryHandle(pbData, owner: this);
    }

    public override void Unpin()
    {
        _handle.DangerousRelease();
    }
}

// factory method
static SafeHandle HeapAlloc(int cb);

// workhorse method
static void DoSomething(Memory<byte> scratchBuffer);

public void UserCode()
{
    using (var manager = new SafeHandleMemoryManager<byte>(HeapAlloc(512), 512))
    {
        // n.b. the SafeHandle will be disposed at the end of this
        // 'using' block.
        DoSomething(manager.Memory);
    }
}

Implementation notes when subclassing MemoryManager<T>

Due to the inherent danger in manual memory management, below are some tips to keep in mind if you have the need to subclass MemoryManager<T>.

Never give your MemoryManger<T>-derived type a finalizer. If an incorrectly-written component loses its reference to the IMemoryOwner<T> and forgets to call Dispose, the garbage collector could kick in unexpectedly, even while a component still has a reference to the Span<T> backed by the custom MemoryManager<T>. In this scenario it is better to leak than to risk heap corruption. (Exception: it is acceptable to create a finalizer for debugging or validation purposes as long as the finalizer no-ops.)

The GetSpan method corresponds to the Memory<T>.Span property getter. Use a constructor on Span<T> if your buffer is backed by a T[], or use MemoryMarshal.CreateSpan if you need to fall back to more complex, unsafe creation semantics.

Finally, consider the relationship between the Pin and Unpin methods. A correctly written Pin method needs to take into account the following.

  • If the underlying buffer is not already pinned, call GCHandle.Alloc to pin the buffer. n.b. the call to GCHandle.Alloc may throw, and the Pin method should be prepared to handle these failures.

  • The parameter passed to Pin is the element index of the requested pointer. When performing any byte offset calculations, the implementation needs to consider the bitness of the current platform so that calculations do not result in integer overflow or underflow. If your underlying buffer is a T[], consider using the Marshal.UnsafeAddrOfPinnedArrayElement method to calculate the correct pointer.

  • The Pin method may cache a single GCHandle for the underlying buffer, or it may return a unique GCHandle per call. If the GCHandle is per-call, then this handle should be given to the MemoryHandle constructor, and MemoryHandle.Dispose will automatically free the GCHandle at the correct time. If the GCHandle is cached per instance, then default(GCHandle) should be passed to the MemoryHandle constructor, and your MemoryManager<T> subclass's Dispose method should free the GCHandle.

  • Pass this for the owner parameter in the MemoryHandle constructor. This will cause MemoryHandle.Dispose to call your MemoryManager<T> subclass's Unpin method at the appropriate time.

  • Despite the name, the Unpin method does not mean "unpin your GCHandle." Rather, it's simply a callback to indicate that the MemoryHandle returned by your Pin method is being destroyed. You generally won't need to put anything into your Unpin method implementation aside from ref counting (if required).

Your MemoryManager<T> subclass may (but is not required to) perform ref counting. If your subclass performs ref counting, this should be for diagnostics or validation, not for correctness. Per the usage guidance, the buffers represented by Memory<T> instances are not inherently thread-safe without external synchronization, so ref counting for correctness does not necessarily help guard against data corruption.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment