ILockProvider.cs

using System;
using System.Collections.Generic;
using System.Text;
using System.Threading;
using System.Threading.Tasks;

namespace Polly.Utilities
{
	/// <summary>
	/// Defines operations for locks used by Polly policies.
	/// </summary>
	public interface ILockProviderAsync
	{
		/// <summary>
		/// Waits to acquire the lock.
		/// </summary>
		/// <param name="key">A string key being used by the execution.</param>
		/// <param name="context">The Polly execution context consuming this lock.</param>
		/// <param name="cancellationToken">A cancellation token to cancel waiting to acquire the lock.</param>
		/// <throws>OperationCanceledException, if the passed <paramref name="cancellationToken"/> is signaled before the lock is acquired.</throws>
		/// <throws>InvalidOperationException, invalid lock state</throws>
		ValueTask<IDisposable> AcquireLockAsync(string key, Context context, CancellationToken cancellationToken);
	}

	/// <summary>
	/// Defines operations for locks used by Polly policies.
	/// </summary>
	public interface ILockProvider
	{
		/// <summary>
		/// Waits to acquire the lock.
		/// </summary>
		/// <param name="key">A string key being used by the execution.</param>
		/// <param name="context">The Polly execution context consuming this lock.</param>
		/// <throws>InvalidOperationException, invalid lock state</throws>
		IDisposable AcquireLock(string key, Context context);
	}

	/// <summary>
	/// Lock provider that locks on a key per process. The locking mechanism is designed to be able
	/// to be requested and released on different threads if needed.
	/// </summary>
	public class ProcessLockProviderAsync : ILockProviderAsync
	{
		// TODO: Pass via IOC or other method instead of hard-coding static
		internal static readonly int[] keyLocks = new int[1024];

		private class ProcessLockAsync : IDisposable
		{
			private uint hash;
			private bool gotLock;

			[System.Runtime.CompilerServices.MethodImpl(System.Runtime.CompilerServices.MethodImplOptions.AggressiveInlining)]
			public void Dispose()
			{
				if (gotLock)
				{
					gotLock = false;
					ProcessLockProviderAsync.keyLocks[hash] = 0;
				}
                // else we do not care, it can be disposed in an error case and we will simply ignore that the key locks were not touched
			}

			[System.Runtime.CompilerServices.MethodImpl(System.Runtime.CompilerServices.MethodImplOptions.AggressiveInlining)]
			public async ValueTask<IDisposable> AcquireLockAsync(string key, Context context, CancellationToken cancellationToken)
			{
				// Monitor.Enter and Monitor.Exit are tied to a specific thread and are
				//  slower than this spin lock, which does not care about threads and will execute very
				//  quickly, regardless of lock contention
				// https://stackoverflow.com/questions/11001760/monitor-enter-and-monitor-exit-in-different-threads

				// Get a hash based on the key, use this to lock on a specific int in the array. The array is designed
				// to be small enough to not use very much memory, but large enough to avoid collisions.
				// Even if there is a collision, it will be resolved very quickly.
				hash = (uint)key.GetHashCode() % (uint)ProcessLockProviderAsync.keyLocks.Length;

				// To get the lock, we must change the int at hash index from a 0 to a 1. If the value is
				//  already a 1, we don't get the lock. The return value must be 0 (the original value of the int).
				// it is very unlikely to have any contention here, but if so, the spin cycle should be very short.
				// Parameter index 1 (value of 1) is the value to change to if the existing value (Parameter index 2) is 0.
				while (!cancellationToken.IsCancellationRequested && Interlocked.CompareExchange(ref ProcessLockProviderAsync.keyLocks[hash], 1, 0) == 1)
				{
					// give up a clock cycle, we want to get back and try to get the lock again very quickly
					await Task.Yield();
				}

				if (cancellationToken.IsCancellationRequested)
				{
					throw new OperationCanceledException(cancellationToken);
				}

				gotLock = true;
				return this;
			}
		}

		[System.Runtime.CompilerServices.MethodImpl(System.Runtime.CompilerServices.MethodImplOptions.AggressiveInlining)]
		public ValueTask<IDisposable> AcquireLockAsync(string key, Context context, CancellationToken cancellationToken)
		{
			return new ProcessLockAsync().AcquireLockAsync(key, context, cancellationToken);
		}
	}

	/// <summary>
	/// Lock provider that locks on a key per process. The locking mechanism is designed to be able
	/// to be requested and released on different threads if needed.
	/// </summary>
	public class ProcessLockProvider : ILockProvider
	{
		private class ProcessLock : IDisposable
		{
			private uint hash;
			private bool gotLock;

			[System.Runtime.CompilerServices.MethodImpl(System.Runtime.CompilerServices.MethodImplOptions.AggressiveInlining)]
			public void Dispose()
			{
				if (gotLock)
				{
					gotLock = false;
					ProcessLockProviderAsync.keyLocks[hash] = 0;
				}
                // if constructor had exception, we will not get in Dispose as the object will never be created, if the constructor succeeds, gotLock will always be true
                // we still use the gotLock bool in case of multiple dispose calls
			}

			[System.Runtime.CompilerServices.MethodImpl(System.Runtime.CompilerServices.MethodImplOptions.AggressiveInlining)]
			public ProcessLockAsync(string key, Context context)
			{
				hash = (uint)key.GetHashCode() % (uint)ProcessLockProviderAsync.keyLocks.Length;
				while (Interlocked.CompareExchange(ref ProcessLockProviderAsync.keyLocks[hash], 1, 0) == 1)
				{
					Task.Yield().GetAwaiter().GetResult();
				}
				gotLock = true;
			}
		}

		/// <inheritdoc />
		public IDisposable AcquireLock(string key, Context context)
		{
			return new ProcessLock(key, context);
		}
	}
}

@reisenberger I went ahead and implemented the factory pattern. Let me know what you think! I went back to a simple bool gotLock because the factory creates a new instance each time and the contract returned (IDisposable) does not allow attempting to re-acquire the lock. Also, the returned IDisposable is not guaranteed to be thread safe, so calling dispose from multiple threads is unsupported.

@jjxtra Again, thanks for everything on the locking!

Some notes just on what turned out different in the main Policy implementation, compared to what discussed previously and what we have in this gist:

• I was wrong in this case about the benefit of using a struct for the releasers. The compiler does have a special transform for when a using statement directly references a struct (ie using (<some-struct>) { }), to avoid boxing, but here we are returning an IDisposable from a method due to the lock interfaces, so it gets boxed to IDisposable first anyway 🙁 . So I switched back to class in the v0.1.0 basic lock implementation.
• There were a couple of places in my early spike which used the non-thread-safe members on ConcurrentDictionary<,> incl. outside the lock - switched to thread-safe.
• I'm thinking that AcquireLockAsync needs to return IAsyncDisposable, not just ValueTask<IDisposable>. If the lock-releasing is across network I/O (like to Redis), we want the implementation of that release to be able to be fully async. IAsyncDisposable uses ValueTask as the return type of its release method, so when we are just wrapping a sync implementation, we still get the benefit of ValueTask.

Assuming you push on forward to bring the stripe-locking idea into Polly.Contrib.DuplicateRequestCollapser:

• We probably wouldn't need the async implementation ProcessLockProviderAsync? There's nothing async in the work it does (setting aside the need to yield asynchronously given it's async), so we could just use a shallow async-over-sync wrapper. (Just running it async when it doesn't have any need for async work will create extra allocations for the state machines, async continuations etc.)
• I dug into the source code around Task.Yield(), to be sure I understood what Task.Yield().GetAwaiter().GetResult() would do. Looks like it doesn't offer any yielding behaviour if used sync like that, .GetAwaiter() returns a YieldAwaitable, GetResult() on that is a no-op. Looks like the benefit of Task.Yield() only comes if it's genuinely used in an async context; the await then schedules the async continuation back onto one or other TaskScheduler, which looks like the mechanism by which await Task.Yield() gets its value - that async continuation has to compete with other async tasks in the scheduler. NB Moot now that we can now drop .Net Standard 1.x (:+1: ), just sharing what I found cos I had dug into this.

Thanks again for everything you are doing on the locking. 👍

You are very welcome. Would I add the striping to InstanceScopedLockProvider.cs I assume then? I'll do a PR for it...

…

-- Jeff

On Thu, Feb 6, 2020 at 5:07 PM Dylan Reisenberger ***@***.***> wrote: @jjxtra <https://github.com/jjxtra> Again, thanks for everything on the locking! Some notes just on what turned out different in the main Policy implementation <https://github.com/Polly-Contrib/Polly.Contrib.DuplicateRequestCollapser/tree/master/src/Polly.Contrib.DuplicateRequestCollapser>, compared to what discussed previously and what we have in this gist: - I was wrong in this case about the benefit of using a struct for the releasers. The compiler does have a special transform for when a using statement directly references a struct (ie using (<some-struct>) { }), to avoid boxing, but here we are returning an IDisposable from a method due to the lock interfaces, so it gets boxed to IDisposable first anyway 🙁 . So I switched back to class in the v0.1.0 basic lock implementation. - There were a couple of places in my early spike which used the non-thread-safe members on ConcurrentDictionary<,> incl. outside the lock - switched to thread-safe. - I'm thinking that AcquireLockAsync needs to return IAsyncDisposable <https://github.com/Polly-Contrib/Polly.Contrib.DuplicateRequestCollapser/blob/master/src/Polly.Contrib.DuplicateRequestCollapser/IAsyncLockProvider.cs>, not just ValueTask<IDisposable>. If the lock-releasing is across network I/O (like to Redis), we want the implementation of that release to be able to be fully async. IAsyncDisposable uses ValueTask as the return type of its release method <https://github.com/Polly-Contrib/Polly.Contrib.DuplicateRequestCollapser/blob/master/src/Polly.Contrib.DuplicateRequestCollapser/AsyncWrapperLockProvider.cs#L38>, so when we are just wrapping a sync implementation, we still get the benefit of ValueTask. Assuming you push on forward to bring the stripe-locking idea into Polly.Contrib.DuplicateRequestCollapser: - We probably wouldn't need the async implementation ProcessLockProviderAsync? There's nothing async in the work it does (setting aside the need to yield asynchronously given it's async), so we could just use a shallow async-over-sync wrapper <https://github.com/Polly-Contrib/Polly.Contrib.DuplicateRequestCollapser/blob/master/src/Polly.Contrib.DuplicateRequestCollapser/AsyncWrapperLockProvider.cs>. (Just running it async when it doesn't have any need for async work will create extra allocations for the state machines, async continuations etc.) - I dug into the source code around Task.Yield(), to be sure I understood what Task.Yield().GetAwaiter().GetResult() would do. Looks like it doesn't offer any yielding behaviour if used sync like that, .GetAwaiter() returns a YieldAwaitable <https://github.com/dotnet/corefx/blob/26a505faf6d27283f3e9bca33371e19c91f88082/src/Common/src/CoreLib/System/Runtime/CompilerServices/YieldAwaitable.cs#L45>, GetResult() on that is a no-op <https://github.com/dotnet/corefx/blob/26a505faf6d27283f3e9bca33371e19c91f88082/src/Common/src/CoreLib/System/Runtime/CompilerServices/YieldAwaitable.cs#L162>. Looks like the benefit of Task.Yield() only comes if it's genuinely used in an async context; the await then schedules the async continuation back onto one or other TaskScheduler <https://github.com/dotnet/corefx/blob/26a505faf6d27283f3e9bca33371e19c91f88082/src/Common/src/CoreLib/System/Runtime/CompilerServices/YieldAwaitable.cs#L75-L116>, which looks like the mechanism by which await Task.Yield() gets its value - that async continuation has to compete with other async tasks in the scheduler. NB Moot now that we can now drop .Net Standard 1.x (👍 ), just sharing what I found cos I had dug into this. Thanks again for everything you are doing on the locking. 👍 — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <https://gist.github.com/f6116180b2ef5c1550e60567af506c2a?email_source=notifications&email_token=AAJWQGCMJALY6FHELYNLCNTRBSQ5LA5CNFSM4KIE6FYKYY3PNVWWK3TUL52HS4DFVNDWS43UINXW23LFNZ2KUY3PNVWWK3TUL5UWJTQAGBNL4#gistcomment-3168958>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAJWQGDE5RZOLBGT3KE7HCTRBSQ5LANCNFSM4KIE6FYA> .

Yep, PR onto Polly.Contrib.DuplicateRequestCollapser.

My recommendation would be a PR to add a new ISyncLockingProvider, not replace/modify InstanceScopedLockProvider, leaving both as options. Curating a widely-used OSS project really brings home (if it is not already obvious) that different users have different needs. Some will benefit from a striped-lock provider; some will have no need. Giving users options (provided not confusing), rather than pre-deciding on one option to suit all users, can work well. 👍

Thanks again for helping drive this forward!

You are welcome!