- Introduction
- Background and History
- Creating Mocks
- API Generation
- Mocking Simple Methods
- Parameter Verification
- Method Call Counts
- Implementing Handled Methods
- Returning Values
- Passing Constructor Arguments to a Mock
- Mocking Generic Methods and Types
- Mocking Methods with
ref/out/in
Parameters orref readonly
Return Values - Mocking Properties
- Mocking Indexers
- Mocking Events
- Optional Arguments
- Handling Asynchronous Code
dynamic
Types- "Special" Types
- Using Makes
- Managing Multiple Mocks
- Conclusion
New to Rocks? In this page, we'll cover the essentials of what Rocks can do so you can get up to speed on the API with little effort. We'll go through creating mocks and how you handle methods, properties and events. We'll demonstrate how you can test asynchronous code, and use dynamic
and "special" types. We'll show what "makes" are and where they're useful.
If something is unclear after reading the documentation, you can always browse the tests in source to see specific examples of a case that may not be covered in detail here. If you are still unable to determine how something works, feel free to drop a message in the Discord channel, or add an issue here.
There are great mocking libraries out there, like Moq and NSubstitute, so why did I decide to create YAML (yet another mocking library) in 2015? There are essentially two reasons.
The first reason relates to how code generation was done with mocking libraries. Most (if not all) used an approach that ends up using System.Reflection.Emit
to create a mock type on the fly, which requires knowledge of IL, or a library like Castle.DynamicProxy
to facilitate the mock generation with IL. This is not a trivial endeavour. Furthermore, the generated code can't be stepped into during a debugging process. I wanted to write a mocking library with the new Compiler APIs (Roslyn) to see if I could make the code generation process for the mock much easier and allow a developer to step into that code if necessary.
The other reason was being able to pre-generate the mocks for a given assembly, rather than dynamically generate them in a test. This is what the Microsoft Fakes Library can do, but I wanted to be able to do it where I could easily modify a project file and automatically generate those mocks.
This is what Rocks can do. Mocks are created by generating C# code on the fly and compiling it with the Compiler APIs. This makes it trivial to step into the mock code. Before the 5.0.0 version, this code generation step took place at runtime, but with source generators in C# 9, this generation happens as soon as you state that you want to create a mock of a particular type. Moreover, since the mock is generated C# code, any language feature in C# can be supported, such as optional parameters and pointer values. So, feel free to test Rocks out, and see what you think. Even if you don't use it as your primary mocking library, you may see just how easy it to generate code on the fly with the new Compiler APIs. Enjoy!
There's a lot of scenarios that can be encountered when you want to create a mock. In this section we'll cover all of them.
With Rocks 5.0.0 and above, all of the mocks are created using C# 9.0's source generation feature. This means that you must be targeting .NET 5.0 to use Rocks. With source generation, you have a lot of freedom to generate what you want, and this is exactly what Rocks takes advantage of. When you want to create a mock, a number of properties are generated that group the members you can set expectations on. You'll see in this document references in code like .Methods
, .Properties
, and .Indexers
. For example, if the type you want to mock has methods that can have expectations set on them, you'll see .Methods
show up. Similarly, mockable properties can be found from the .Properties
property. You use these properties to set expectations on specific members. This should become clear as you read on and look at the examples.
Note: if you use Rocks and you get an error akin to having a
Microsoft.CodeAnalysis
version mismatch, you may need to update the version of the .NET SDK you have installed.
Creating a mock in Rocks is pretty easy. Let's say you have an interface defined like this:
public interface IAmSimple
{
void TargetAction();
int TargetFunc();
}
Here's how you create the mock, define its expected interactions, use the mock, and verify the expectations:
[assembly: Rock(typeof(IAmSimple), BuildType.Create)]
var expectations = new IAmSimpleCreateExpectations();
expectations.Methods.TargetAction();
expectations.Methods.TargetFunc().ReturnValue(44);
var mock = expectations.Instance();
mock.TargetAction();
var result = mock.TargetFunc();
// result is equal to 44
expectations.Verify();
The mocking process starts by instantiating an IAmSimpleCreateExpectations
object. This is done by adding RockAttribute
in code. When this attribute is added, Rocks will interrogate the type provided in the generic parameter. It will then create a number of properties for each mockable member. In the example, you can call .Methods.TargetAction()
and .Methods.TargetFunc()
to state that you expect these members to be called. For TargetFunc()
, you can also specify the return value via the ReturnValue()
method.
You can also specify BuildType.Make
to create a "make" type. Makes will be covered later on in this document. RockAttribute
can only exist at the assembly level. If you have multiple instances of these attributes defined in a project targeting the same type to mock, only one expectations class will be generated. The names of the generated expectation types follow these patterns:
Rock(typeof(MyType), BuildType.Create)
generatesMyTypeCreateExpectations
Rock(typeof(MyType), BuildType.Make)
generatesMyTypeMakeExpectations
Rock(typeof(MyType), BuildType.Create | BuildType.Make)
generates bothMyTypeCreateExpectations
andMyTypeMakeExpectations
In other words, the naming pattern is {Mock type name}{Mock kind}Expectations
. The expectation type is generated within the same namespace that the mock type resides within (if one exists).
Note: With version 8.2.0 and before, Rocks had
RockCreateAttribute
andRockMakeAttribute
attributes, both generic and non-generic versions. With 8.3.0, they are marked as being obsolete, and they will be removed in 9.0.0.
Once you set your expectations, you can create an instance of the mock from the generated Instance()
method. If the target type has multiple constructors with different parameters, Instace()
methods will generated for each constructor - constructors will be covered in detail in this section.
Note that all mocks generated with Rocks are strict. That is, if you didn't set up an expectation for the TargetAction()
call, the Verify()
call would fail with a VerificationException
. This exception type provided details on why verification failed.
If your method has parameters, there are a couple of ways you can set up expectations on what should be passed into the method. Here's an interface with a method that takes an integer:
public interface IHaveParameterExpectations
{
void Target(int a);
}
You can verify that Target()
will be called with an exact value by passing in that value when you set up the expectation:
[assembly: Rock(typeof(IHaveParameterExpectations), BuildType.Create)]
var expectations = new IHaveParameterExpectationsCreateExpectations();
expectations.Methods.Target(44);
var mock = expectations.Instance();
mock.Target(44);
expectations.Verify();
If you don't care what the value is, you use Arg.Any<>()
:
[assembly: Rock(typeof(IHaveParameterExpectations), BuildType.Create)]
var expectations = new IHaveParameterExpectationsCreateExpectations();
expectations.Methods.Target(Arg.Any<int>());
var mock = expectations.Instance();
mock.Target(44);
expectations.Verify();
If you want to specify logic to validate the given value, you use Arg.Validate<>()
:
[assembly: Rock(typeof(IHaveParameterExpectations), BuildType.Create)]
var expectations = new IHaveParameterExpectationsCreateExpectations();
expectations.Methods.Target(Arg.Validate<int>(a => a > 20 && a < 50));
var mock = expectations.Instance();
mock.Target(44);
expectations.Verify();
You can also specify multiple expectations:
[assembly: Rock(typeof(IHaveParameterExpectations), BuildType.Create)]
var expectations = new IHaveParameterExpectationsCreateExpectations();
expectations.Methods.Target(Arg.Validate<int>(a => a > 20 && a < 50));
expectations.Methods.Target(10);
var mock = expectations.Instance();
mock.Target(44);
mock.Target(10);
expectations.Verify();
You may want to verify that code under test calls a method a specific number of times. You can do that by specifying an expected call count via ExpectedCallCount()
:
[assembly: Rock(typeof(IAmSimple), BuildType.Create)]
var expectations = new IAmSimpleCreateExpectations();
expectations.Methods.TargetAction().ExpectedCallCount(2);
var mock = expectations.Instance();
mock.TargetAction();
mock.TargetAction();
expectations.Verify();
This also works with multiple expectations:
[assembly: Rock(typeof(IHaveParameterExpectations), BuildType.Create)]
var expectations = new IHaveParameterExpectationsCreateExpectations();
expectations.Methods.Target(44).ExpectedCallCount(2);
expectations.Methods.Target(22).ExpectedCallCount(3);
var mock = expectations.Instance();
mock.Target(22);
mock.Target(44);
mock.Target(22);
mock.Target(44);
mock.Target(22);
expectations.Verify();
We haven't covered properties yet, but the same process is in place with them as well.
You can provide a lambda via the Callback()
method that will be called when a method is invoked so you can do things like capture method argument values:
[assembly: Rock(typeof(IHaveParameterExpectations), BuildType.Create)]
var value = 0;
var expectations = new IHaveParameterExpectationsCreateExpectations();
expectations.Methods.Target(Arg.Validate<int>(i => i > 10))
.Callback(a => value = a);
var mock = expectations.Instance();
mock.Target(44);
// value would be equal to 44 here.
expectations.Verify();
A callback can also be used if the member usage should throw an exception as expected behavior:
expectations.Methods.Target(Arg.Validate<int>(i => i > 10))
.Callback(_ => throw new NotSupportedException());
If a method returns a value, you can use ReturnValue()
:
[assembly: Rock(typeof(IAmSimple), BuildType.Create)]
var expectations = new IAmSimpleCreateExpectations();
expectations.Methods.TargetFunc().ReturnValue(44);
var mock = expectations.Instance();
var x = mock.TargetFunc();
// x is equal to 44.
expectations.Verify();
If you want to mock a class with virtual members where the class only has constructors with multiple arguments, you can do it. You pass the constructor arguments to Instance()
. Rocks will generate an override of Instance()
for each constructor that exists on the target type. Here's an example:
public class MockedClass
{
public MockedClass(int value) { }
public virtual void Target() { }
}
// ...
[assembly: Rock(typeof(MockedClass), BuildType.Create)]
var expectations = new MockedClassCreateExpectations();
expectations.Methods.Target();
var mock = expectations.Instance(44);
mock.Target();
expectations.Verify();
If generics are in play, Rocks can handle that by specifying an open generic:
public interface IHaveGenerics<T>
{
void Target<Q>(T a, Q b);
}
// ...
[assembly: Rock(typeof(IHaveGenerics<>), BuildType.Create)]
var expectations = new IHaveGenericsCreateExpectations<string>();
expectations.Methods.Target<int>("a", 44));
var mock = expectations.Instance();
mock.Target("a", 44);
expectations.Verify();
This allows for more flexibility, in that only one set of supporting types are generated for the target type.
Note: With 8.3.0, if you declare a closed generic, Rocks will create a
ROCK14
diagnostic as a warning. In 9.0.0, this will become an error.
Methods with either ref
, out
, or in
parameters are supported. A delegate is created for you so you can easily create a callback method to handle the parameters if you want:
public interface IHaveRefs
{
void Target(ref int a);
}
// ...
[assembly: Rock(typeof(IHaveRefs), BuildType.Create)]
public void MyTestMethod()
{
static void TargetCallback(ref int a) => a = 4;
var expectations = new IHaveRefsCreateExpectations();
expectations.Methods.Target(3).Callback(TargetCallback);
var mock = expectations.Instance();
var value = 3;
mock.Target(ref value);
// value is 4 here.
expectations.Verify();
}
Since the Action
and Func
delegate types do not support ref
and out
, you have to declare the callback method yourself, but with local functions this is straightforward to do. You can use any method declaration technique you'd like - the callback method just needs to match the signature of the target method.
in
parameters do not require any special handling as Rocks doesn't change the value of parameters. ref readonly
also work in Rocks and require no extra work on your behalf. You can return your own values from methods if you want, and Rocks handles the requirement of having a field available to return by reference.
Mocking properties is a breeze in Rocks. Getters
, Setters
, and/or Initializers
properties are generated for you off of the Properties
property:
public interface IHaveAProperty
{
string GetterAndSetter { get; set; }
}
Here's how you set up the expectations:
[assembly: Rock(typeof(IHaveAProperty), BuildType.Create)]
var setupValue = Guid.NewGuid().ToString();
var expectations = new IHaveAPropertyCreateExpectations();
expectations.Properties.Getters.GetterAndSetter();
expectations.Properties.Setters.GetterAndSetter(setupValue);
var mock = expectations.Instance();
mock.GetterAndSetter = setupValue;
var value = mock.GetterAndSetter;
expectations.Verify();
Note that you can also set up callbacks and expected call counts just like you can with methods.
The init
feature was added with C# 9, and required
properties were added with C# 11. Starting with Rocks 7.0.0
, there's a way to set these properties when the mock is created. A type called ConstructorProperties
is created that contains all of the required
and init
properties, and an instance of this type can be given as the first argument to the generated Instance()
methods:
public class RequiredAndInit
{
public virtual void Foo() { }
public string InitData { get; init; }
public required string RequiredData { get; set; }
}
[assembly: Rock(typeof(RequiredAndInit), BuildType.Create)]
var expectations = new RequiredAndInitCreateExpectations();
expectations.Methods.Foo();
var mock = expectations.Instance(new() { InitData = "a", RequiredData = "b" });
This will set InitData
to "a"
and RequiredData
to "b"
on the mock instance.
ConstructorProperties
will contain properties that are both virtual and non-virtual. If there are no required
properties, the constructorProperties
parameter will be nullable.
Note that this will work with init
indexers as well. required
indexers is currently not possible in C#.
Indexers are not something a lot of .NET developers use, but if you do, you can mock them in Rocks. An Indexers
property is created to set indexer get, set, and/or init expectations. Following the naming convention of an indexer in C#, a method named This()
is generated that allows you to set the expectations:
public interface IHaveIndexer
{
int this[int a] { get; set; }
}
// ...
[assembly: Rock(typeof(IHaveIndexer), BuildType.Create)]
var expectations = new IHaveIndexerCreateExpectations();
expectations.Indexers.Getters.This(3);
expectations.Indexers.Setters.This(4, 3);
var mock = expectations.Instance();
var propertyValue = mock[3];
mock[3] = 4;
expectations.Verify();
Note that the setter looks like it's taking an extra parameter - that's because the value is passed in as the first argument. If you're wondering why the value is the first argument, read the Optional Arguments section for the explanation.
If there's an event on the mock, you can raise it as part of a member's usage:
public interface IHaveAnEvent
{
void Target(int a);
event EventHandler TargetEvent;
}
// ...
[assembly: Rock(typeof(IHaveAnEvent), BuildType.Create)]
var expectations = new IHaveAnEventCreateExpectations();
expectations.Methods.Target(1).RaiseTargetEvent(EventArgs.Empty);
var wasEventRaised = false;
var mock = expectations.Instance();
mock.TargetEvent += (s, e) => wasEventRaised = true;
// wasEventRaised is still false
mock.Target(1);
// wasEventRaised is now equal to true
expectations.Verify();
Rocks generates extension methods for every adornments object (which is returned when an expectation is set). The naming pattern is Raise{Event name}
. There is an AddRaiseEvent()
method you can call directly, but the extension methods makes it convenient to pass in the right event name and its' corresponding event argument value.
Note that these extension methods won't be created if the mock type is generic, an open generic is requested, and there are events on the mock type. This is due to the way the extension methods are created. This situation should be relative rare, and AddRaiseEvent()
can still be used in this case. (This limitation is being tracked with this issue - hopefully this will be resolved in the future.)
If your method or indexer has optional arguments, you can handle them just like other arguments. You can also use Arg.IsDefault()
if you want to use the default argument no matter what it is (or if it changes in the future). Here's what that looks like:
public interface IHaveOptionalArguments
{
void Foo(int a, string b = "b", double c = 3.2);
int this[int a, string b = "b"] { get; set; }
}
// ...
[assembly: Rock(typeof(IHaveOptionalArguments), BuildType.Create)]
var returnValue = 3;
var expectations = new IHaveOptionalArgumentsCreateExpectations();
// In this case, we're assuming b will be set to "b",
// and c will be set to 3.2
expectations.Methods.Foo(1);
// With the indexer getter, we assume b will be set to "b"
expectations.Indexers.Getters.This(2).ReturnValue(returnValue);
// Read the explanation after this code snippet ;)
expectations.Indexers.Setters.This(value: 52, a: 3);
var mock = expectations.Instance();
mock.Foo(1);
var value = mock[2];
mock[3] = 52;
expectations.Verify();
Assert.That(value, Is.EqualTo(returnValue));
There is a little bit of an oddity if an indexer's setter has optional or params
arguments. Properties and indexers have methods that are mapped to the get
, set
, or init
accessors. In this case of an indexer's setter, the set_Item()
mapped method has the value
parameter after the optional parameters. While C#'s compiler makes this happen, you can't declare this in C#:
void set_Item(int a, string b = "b", int value) { /* ... */ }
What Rocks does is generate the This()
extension method with the value
as the first parameter. That way, any parameters with default values or declared as params
can show up after value
.
If your mock returns a Task
, Task<T>
, ValueTask
, or a ValueTask<T>
, or you're using that returned value with async/await
, you can create mocks that will allow you to test it:
public interface IAmAsync
{
Task<int> GoAsync();
}
public class UsesAsync
{
private readonly IAmAsync amAsync;
public UsesAsync(IAmAsync amAsync)
{
this.amAsync = amAsync;
}
public async Task<int> RunGoAsync()
{
return await this.amAsync.GoAsync();
}
}
[assembly: Rock(typeof(IAmAsync), BuildType.Create)]
var expectations = new IAmAsyncCreateExpectations();
expectations.Methods.GoAsync().ReturnValue(Task.FromResult(44));
var uses = new UsesAsync(expectations.Instance());
await uses.RunGoAsync().ConfigureAwait(false);
expectations.Verify();
It's not a common thing to see dynamic
used, but Rocks supports it just fine:
public interface IHaveDynamic
{
void Foo(dynamic d);
}
[assembly: Rock(typeof(IHaveDynamic), BuildType.Create)]
var expectations = new IHaveDynamicCreateExpectations();
expectations.Methods.Foo(Arg.Is<dynamic>("b"));
var mock = expectations.Instance();
mock.Foo("b");
expectations.Verify();
Rocks can mock members with certain kinds of types, such as pointer types and ref structs like Span<T>
. So if you have an interface like this:
public unsafe interface IHavePointers
{
void PointerParameter(int* value);
}
You can mock it like this:
var value = 10;
[assembly: Rock(typeof(IHavePointers), BuildType.Create)]
var expectations = new IHavePointersCreateExpectations();
expectations.Methods.PointerParameter(new()).Callback(_ => *_ = 20);
var mock = expectations.Instance();
mock.PointerParameter(&value);
expectations.Verify();
Note that value
would be equal to 20 after PointerParameter()
is called.
Warning: Unmanaged function pointers are supported, as the code generated by Rocks allows you to do pointer comparisons for verification by disabling the
CS8909
warning. However, this may not work as expected, as you may get different pointer values for the same function. Please refer to the following links for more information:
- dotnet/docs#28782 (see "Comment 5")
- dotnet/roslyn#48919
There are times when your mock needs to return a value where you want to ensure that the return value is a specific instance. As you've seen with these Rocks examples, you define a [Rock(..., BuildType.Create)]
attribute instance, create a new expectations instance, set up expectations, and then call Instance()
. If you need a mock with no expectations, you create a "make" by changing the flag value in RockAttribute
:
public interface IValue { }
public interface IProduceValue
{
IValue Produce();
}
public class UsesProducer
{
private readonly IProduceValue producer;
public UsesProducer(IProduceValue producer)
{
this.producer = producer;
}
public IValue GetValue()
{
return this.producer.Produce();
}
}
// ...
[assembly: Rock(typeof(IProduceValue), BuildType.Create)]
[assembly: Rock(typeof(IValue), BuildType.Make)]
var valueMock = new IValueMakeExpectations().Instance();
var produceExpectations = new IProduceValueCreateExpectations();
produceExpectations.Methods.Produce().ReturnValue(valueMock);
var uses = new UsesProducer(produceExpectations.Instance());
var producedValue = uses.GetValue();
// producedValue and valueMock are the same references.
Note that makes do no have any expectations set up on them so they can't be verified. If you call a method on a make that returns a value, it'll return the default value of the return type - the same thing applies for getters on properties and indexers.
You've now seen the majority of cases that Rocks can handle. Remember to peruse the tests within Rocks.IntegrationTests
in case you get stuck. If you'd like, feel free to submit a PR to update this document to improve its contents. If you run into any issues, please submit an issue. Happy coding!