Skip to content

Instantly share code, notes, and snippets.

Embed
What would you like to do?
Memory usage guidelines

Memory<T> usage guidelines

This document describes the relationship between Memory<T> and its related classes (MemoryPool<T>, IMemoryOwner<T>, etc.). It also describes best practices when accepting Memory<T> instances in public API surface. Following these guidelines will help developers write clear, bug-free code.

First, a tour of the basic exchange types

  • Span<T> is the basic exchange type that represents contiguous buffers. These buffers may be backed by managed memory (such as T[] or System.String). They may also be backed by unmanaged memory (such as via stackalloc or a raw void*). The Span<T> type is not heapable, meaning that it cannot appear as a field in classes, and it cannot be used across yield or await boundaries.

  • Memory<T> is a wrapper around an object that can generate a Span<T>. For instance, Memory<T> instances can be backed by T[], System.String (readonly), and even SafeHandle instances. Memory<T> cannot be backed by "transient" unmanaged memory; e.g., it is forbidden to back a Memory<T> with stackalloc. The Memory<T> type is heapable, meaning that it can appear as a field in a class, and it can be used across yield and await boundaries.

There are also ReadOnlySpan<T> and ReadOnlyMemory<T> types that correspond to read-only versions of Span<T> and Memory<T>, respectively.

Owners, consumers, and lifetime management

Let's stick a pin in Memory<T> for now and speak about buffers in more general terms. Since buffers can be passed around between APIs, and since buffers can sometimes be accessed from multiple threads, we need to introduce lifetime semantics. There are three core concepts.

The first concept is ownership. The owner of a buffer instance is responsible for lifetime management, including destroying the buffer when it is no longer in use. All buffers have a single owner. Generally the owner is the component which created the buffer or which received the buffer from a factory. Ownership can also be transferred; Component A can relinquish control of the buffer to Component B, at which point Component A may no longer use the buffer, and Component B becomes responsible for destroying the buffer when it is no longer in use.

The second concept is consumption. The consumer of a buffer instance is allowed to use the buffer instance, perhaps writing to or reading from it. Buffers have one consumer at a time unless some external synchronization mechanism is provided.

Importantly, the active consumer of a buffer is not necessarily the buffer's owner. Consider the following pseudocode, where the Buffer type is a stand-in for an arbitrary buffer type.

// Writes 'value' as a human-readable string to the output buffer.
void WriteInt32ToBuffer(int value, Buffer buffer);

// Prints the contents of the buffer to the console.
void PrintBufferToConsole(Buffer buffer);

// Application code
void Main()
{
    var buffer = CreateBuffer();
    try {
        int value = Int32.Parse(Console.ReadLine());
        WriteInt32ToBuffer(value, buffer);
        PrintBufferToConsole(buffer);
    } finally {
        buffer.Destroy();
    }
}

In this pseudocode, the Main method creates the buffer so becomes its owner, and Main is thus responsible for destroying the buffer when it's no longer in use. The buffer only ever has one consumer at a time (first WriteInt32ToBuffer, then PrintBufferToConsole), and neither of the consumers owns the buffer. Note also that "consumer" in this context does not imply a read-only view of the buffer; consumers can modify buffer contents if given a read+write view of the buffer.

The third concept is that of a lease. The lease is the window of time that any given component is allowed to be the consumer of the buffer. In the example above, the WriteInt32ToBuffer method has a lease on (can consume) the buffer between the start of the method call and the time the method returns. Similarly, PrintBufferToConsole has a lease on the buffer while it is executing, and the lease is released when the method unwinds. (There is no API for lease management; a "lease" is simply a conceptual matter.)

Memory<T> and the owner / consumer model

At this point, let's reintroduce Memory<T> into the picture, along with one more type: IMemoryOwner<T>.

The type IMemoryOwner<T> is, as its name suggests, the unit of ownership of the associated Memory<T> instance. If a component has an IMemoryOwner<T> reference, then that component owns the buffer.

Memory<T> is itself the unit of consumption. If a component has a Memory<T> reference, then that component consumes the buffer.

To clarify this point, consider once again the earlier pseudocode, but let's now introduce real types into the system.

// Writes 'value' as a human-readable string to the output buffer.
void WriteInt32ToBuffer(int value, Memory<char> buffer);

// Prints the contents of the buffer to the console.
void PrintBufferToConsole(Memory<char> buffer);

// Application code
void Main()
{
    IMemoryOwner<char> owner = MemoryPool<char>.Shared.Rent();
    try {
        int value = Int32.Parse(Console.ReadLine());
        WriteInt32ToBuffer(value, owner.Memory);
        PrintBufferToConsole(owner.Memory);
    } finally {
        owner.Dispose();
    }

    // Alternatively, with 'using' syntax instead of 'try / finally'

    using (var owner = MemoryPool<char>.Shared.Rent())
    {
        int value = Int32.Parse(Console.ReadLine());
        WriteInt32ToBuffer(value, owner.Memory);
        PrintBufferToConsole(owner.Memory);
    }
}

Again, in this code, the Main method holds the reference to the IMemoryOwner<char> instance, so the Main method is the owner of the buffer. The WriteInt32ToBuffer and PrintBufferToConsole methods accept Memory<T> as a public API, therefore they consume the buffer. (And they only consume it one-at-a-time.)

(The observant reader may note that PrintBufferToConsole should really accept ReadOnlyMemory<char> instead of Memory<char> as a method argument. More on this later.)

A quick note on "ownerless" Memory<T> instances

It is certainly possible to create a Memory<T> without going through IMemoryOwner<T>. One way to do this would be to call one of the Memory<T> constructors directly, passing in a T[]. Or code could call the String.AsMemory extension method to produce a ReadOnlyMemory<char>.

// Writes 'value' as a human-readable string to the output buffer.
void WriteInt32ToBuffer(int value, Memory<char> buffer);

// Prints the contents of the buffer to the console.
void PrintBufferToConsole(Memory<char> buffer);

// Application code
void Main()
{
    Memory<char> memory = new char[64];

    int value = Int32.Parse(Console.ReadLine());
    WriteInt32ToBuffer(value, memory);
    PrintBufferToConsole(memory);
}

In this case, the method which initially creates the Memory<T> instance is the implicit owner of the buffer. Ownership cannot be transferred to any other component because there is no IMemoryOwner<T> to facilitate the transfer. (As an alternative, you can also imagine that the runtime's garbage collector owns the buffer, and all methods shown here just consume of the buffer.)

Usage guidelines

Now that we have the basics down, we can go over the rules necessary for successful usage of Memory<T> and related types.

In the rules below, we'll generally refer just to Memory<T> and Span<T>. The same guidance also applies to ReadOnlyMemory<T> and ReadOnlySpan<T> unless explicitly called out otherwise.

Rule #1: If writing a synchronous API, accept Span<T> instead of Memory<T> as a parameter if possible.

Span<T> is more versatile than Memory<T> and can represent a wider variety of contigious memory buffers. Span<T> also has better performance characteristics than Memory<T>. Finally, Memory<T> is convertible to Span<T>, but there is no Span<T>-to-Memory<T> conversion possible. So if your callers happen to have Memory<T> instance, they'll be able to call your Span<T>-accepting method anyway.

Accepting Span<T> instead of Memory<T> also helps you write a correct consuming method implementation, as you'll automatically get compile-time checks to ensure that you're not attempting to access the buffer beyond your method's lease (more on this later).

Sometimes circumstances will necessitate you taking a Memory<T> parameter instead of a Span<T> parameter, even if you're fully synchronous. Perhaps an API that you depend on has only Memory<T>-based overloads, and you need to flow your input parameter down to that method. This is fine, but be aware of the tradeoffs mentioned in the first paragraph in this rule.

Rule #2: Use ReadOnlySpan<T> or ReadOnlyMemory<T> if the buffer is intended to be immutable.

Consider the PrintBufferToConsole method from the earlier sample code.

void PrintBufferToConsole(Memory<char> buffer);

This method only reads from the buffer; it does not modify the contents of the buffer. The method signature should be changed to the following.

void PrintBufferToConsole(ReadOnlyMemory<char> buffer);

In fact, combining this rule and Rule #1 above, we can do even better and rewrite it as follows.

void PrintBufferToConsole(ReadOnlySpan<char> buffer);

The PrintBufferToConsole method now works with pretty much every buffer type imagineable: T[], stackalloc, and so on. You can even pass a System.String directly into it!

Rule #3: If your method accepts Memory<T> and returns void, you must not use the Memory<T> instance after your method returns.

This relates back to the "lease" concept mentioned earlier. A void-returning method's lease on the Memory<T> begins when the method is entered, and it ends when the method exits.

Consider the following code sample, which calls Log in a loop based on input from the console.

// implementation provided by third party
static void Log(ReadOnlyMemory<char> message);

// user code
public void Main()
{
    using (var owner = MemoryPool<char>.Shared.Rent())
    {
        var memory = owner.Memory;
        var span = memory.Span;
        while (true)
        {
            int value = Int.Parse(Console.ReadLine());
            if (value < 0) { return; }

            int numCharsWritten = value.ToBuffer(span);
            Log(memory.Slice(0, numCharsWritten));
        }
    }
}

If Log is a fully synchronous method, this code will behave as expected, as there will be only one active consumer of the memory instance at any given time.

Now, imagine instead that Log has this implementation.

// !!! INCORRECT IMPLEMENTATION !!!
static void Log(ReadOnlyMemory<char> message)
{
    // Run in background so that we don't block the main thread
    // while performing IO.
    Task.Run(() => {
        File.AppendText(message);
    });
}

In this scenario, Log violates its lease because it's still attempting to use the Memory<T> instance in the background after the original method has returned. The Main method could be mutating the buffer while Log is attempting to read from it, which could result in data corruption.

There are a few ways to resolve this. One way could be for the Log method to return a Task instead of returning void. Another way could be for Log to instead be implemented as follows.

// Acceptable implementation
static void Log(ReadOnlySpan<char> message)
{
    string defensiveCopy = message.ToString();
    // Run in background so that we don't block the main thread
    // while performing IO.
    Task.Run(() => {
        File.AppendText(defensiveCopy);
    });
}

Rule #4: If your method accepts Memory<T> and returns Task, you must not use the Memory<T> instance after the Task transitions to a terminal state.

This is just the async variant of Rule #3. The Log method from the earlier example can be written as follows to be compliant with this rule.

// Acceptable implementation
static Task LogAsync(ReadOnlyMemory<char> message)
{
    return Task.Run(() => {
        File.AppendText(message);
    });
}

Here, "terminal state" means that the Task transitions to a successful, faulted, or canceled state. In other words, "terminal state" means "anything that would cause await to throw or to continue execution."

This guidance holds for methods which return Task, Task<T>, ValueTask<T>, or any similar type.

Rule #5: If your constructor accepts Memory<T> as a parameter, instance methods on the constructed object are assumed to be consumers of the Memory<T> instance.

Consider the following sample code.

class OddValueExtractor {
    public OddValueExtractor(ReadOnlyMemory<int> input);
    public bool TryReadNextOddValue(out int value);
}

void PrintAllOddValues(ReadOnlyMemory<int> input)
{
    var extractor = new OddValueExtractor(input);
    while (extractor.TryReadNextOddValue(out int value))
    {
      Console.WriteLine(value);
    }
}

Here, the OddValueExtractor constructor accepts a Memory<T> as a constructor parameter, so the constructor itself is a consumer of the Memory<T> instance, and all instance methods on the returned value are also consumers of the original Memory<T> instance.

This means that TryReadNextOddValue consumes the Memory<T> instance, even though the instance isn't passed directly to the TryReadNextOddValue method.

Rule #6: If you have a settable Memory<T>-typed property (or equivalent instance method) on your type, instance methods on that object are assumed to be consumers of the Memory<T> instance.

This is really just a variant of Rule #5. This rule exists because property setters or equivalent methods are assumed to capture and persist their inputs, so instance methods on the same object may utilize the captured state.

A sample class which triggers this rule is provided below.

class Person
{
    // settable property
    public Memory<char> FirstName { get; set; }

    // alternatively, equivalent "setter" method
    public SetFirstName(Memory<char> value);

    // alternatively, a public settable field
    public Memory<char> FirstName;
}

Rule #7: If you have an IMemoryOwner<T> reference, you must at some point dispose of it or transfer ownership (but not both).

Since a Memory<T> instance may be backed by either managed or unmanaged memory, it's imperative that the owner call IMemoryOwner<T>.Dispose when all work being performed on the Memory<T> instance is complete. Alternatively, the owner may transfer ownership of the IMemoryOwner<T> instance to a different component, at which point the acquiring component becomes responsible for calling Dispose at the appropriate time (more on this later).

Failure to call the Dispose method may lead to unmanaged memory leaks or other performance degradation.

This rule also applies to code which calls factory methods like MemoryPool<T>.Rent. The caller becomes the owner of the returned IMemoryOwner<T> and is responsible for disposing of the instance when finished.

Rule #8: If you have an IMemoryOwner<T> parameter in your API surface, you are accepting ownership of that instance.

Accepting an instance of this type signals that your component intends to take ownership of this instance. Your component is now responsible for proper disposal per Rule #7.

Any component handing over ownership of the IMemoryOwner<T> instance to a different component should no longer use that instance after the method call completes.

Reminder: If your constructor accepts IMemoryOwner<T> as a parameter, your type should also implement IDisposable, and your Dispose method should call IMemoryOwner<T>.Dispose.

Rule #9: If you're wrapping a synchronous p/invoke method, your API should accept Span<T> as a parameter.

Per Rule #1, Span<T> is generally the correct type to take for synchronous APIs. It is possible to pin Span<T> instances via the fixed keyword, as in the following example.

using System.Runtime.InteropServices;

[DllImport(...)]
private static extern unsafe int ExportedMethod(byte* pbData, int cbData);

public unsafe int ManagedWrapper(Span<byte> data)
{
    fixed (byte* pbData = &MemoryMarshal.GetReference(data))
    {
        int retVal = ExportedMethod(pbData, data.Length);

        /* error checking retVal goes here */

        return retVal;
    }

    // In the above example, 'pbData' can be null; e.g., if
    // the input span is empty. If the exported method absolutely
    // requires that 'pbData' be non-null, even if 'cbData' is 0,
    // consider the following implementation.

    fixed (byte* pbData = &MemoryMarshal.GetReference(data))
    {
        byte dummy = 0;
        int retVal = ExportedMethod((pbData != null) ? pbData : &dummy, data.Length);

        /* error checking retVal goes here */

        return retVal;
    }
}

Rule #10: If you're wrapping an asynchronous p/invoke method, your API should accept Memory<T> as a parameter.

Since you cannot use the fixed keyword across asynchronous operations, the method Memory<T>.Pin is provided to pin Memory<T> instances, regardless of what kind of contiguous memory the instance represents.

The following example shows how to do use this API to perform an asynchronous p/invoke call.

using System.Runtime.InteropServices;

[UnmanagedFunctionPointer(...)]
private delegate void OnCompletedCallback(IntPtr state, int result);

[DllImport(...)]
private static extern unsafe int ExportedAsyncMethod(byte* pbData, int cbData, IntPtr pState, IntPtr lpfnOnCompletedCallback);

private static readonly IntPtr _callbackPtr = GetCompletionCallbackPointer();

public unsafe Task<int> ManagedWrapperAsync(Memory<byte> data)
{
    // setup
    var tcs = new TaskCompletionSource<int>();
    var state = new MyCompletedCallbackState {
        Tcs = tcs
    };
    var pState = (IntPtr)GCHandle.Alloc();

    var memoryHandle = data.Pin();
    state.MemoryHandle = memoryHandle;

    // make the call
    int result;
    try {
        result = ExportedAsyncMethod((byte*)memoryHandle.Pointer, data.Length, pState, _callbackPtr);
    } catch {
        ((GCHandle)pState).Free(); // cleanup since callback won't be invoked
        memoryHandle.Dispose();
        throw;
    }

    if (result != PENDING)
    {
        // Operation completed synchronously; invoke callback manually
        // for result processing and cleanup.
        MyCompletedCallbackImplementation(pState, result);
    }

    return tcs.Task;
}

private static void MyCompletedCallbackImplementation(IntPtr state, int result)
{
    GCHandle handle = (GCHandle)state;
    var actualState = (MyCompletedCallbackState)state;
    handle.Free();
    actualState.MemoryHandle.Dispose();

    /* error checking result goes here */

    if (error) { actualState.Tcs.SetException(...); }
    else { actualState.Tcs.SetResult(result); }
}

private static IntPtr GetCompletionCallbackPointer()
{
    OnCompletedCallback callback = MyCompletedCallbackImplementation;
    GCHandle.Alloc(callback); // keep alive for lifetime of application
    return Marshal.GetFunctionPointerForDelegate(callback);
}

private class MyCompletedCallbackState
{
    public TaskCompletionSource<int> Tcs;
    public MemoryHandle MemoryHandle;
}
@jherby2k

This comment has been minimized.

Copy link

@jherby2k jherby2k commented Apr 12, 2018

Super useful guidance, thanks a bunch!

@ghost

This comment has been minimized.

Copy link

@ghost ghost commented Aug 15, 2018

Thanks, really useful!

I've been experimenting with Span, I'm doing some machine learning research and wondering if the new constructs could allow C# to compete with Python + Numpy. I used your example above to create a p/invoke interface to the Intel MKL (Math Kernal Library), and built a simple Matrix class which wraps a buffer of native memory (allocated using AllocHGlobal). Works fine, but in languages like C++ / python etc the int datatype maps to the platform word size, i.e. Int32 or Int64. In C# it always seems to be an Int32, and the indexer for Span is int. Very large double precision matrix can very easily go over 2GB

So it looks like you can only use a Span to access 2GB of memory even if it's allocated using native code (or even using the AllocHGlobal overload which takes and IntPtr)?

Seems to me if AllocHGlobal has a IntPtr overload, so should Span? Do you know of any work arounds?

Regards
Dave

@antonfirsov

This comment has been minimized.

Copy link

@antonfirsov antonfirsov commented Apr 23, 2021

This stuff is now on docs.microsoft.com, dropping a link here in case someone land on this gist from google:
https://docs.microsoft.com/en-us/dotnet/standard/memory-and-spans/memory-t-usage-guidelines

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