EventAggregator, is it thread-safe? - wpf

Is this thread-safe?
The EventAggregator in Prism is a very simple class with only one method. I was surprised when I noticed that there was no lock around the null check and creation of a new type to add to the private _events collection. If two threads called GetEvent simultaneously for the same type (before it exists in _events) it looks like this would result in two entries in the collection.
/// <summary>
/// Gets the single instance of the event managed by this EventAggregator. Multiple calls to this method with the same <typeparamref name="TEventType"/> returns the same event instance.
/// </summary>
/// <typeparam name="TEventType">The type of event to get. This must inherit from <see cref="EventBase"/>.</typeparam>
/// <returns>A singleton instance of an event object of type <typeparamref name="TEventType"/>.</returns>
public TEventType GetEvent<TEventType>() where TEventType : EventBase
{
TEventType eventInstance = _events.FirstOrDefault(evt => evt.GetType() == typeof(TEventType)) as TEventType;
if (eventInstance == null)
{
eventInstance = Activator.CreateInstance<TEventType>();
_events.Add(eventInstance);
}
return eventInstance;
}

No, not thread safe.
The instance member access of the List class itself are NOT thread safe, as per MSDN under Thread safety
The method itself is not thread safe
2 threads could enter the method at the same time
Both try to get the FirstOrDefault
Both get nothing
Both add a new TEventType
I would
switch to one of the System.CollectionConcurrentX collections in .NET 4
http://msdn.microsoft.com/en-us/library/system.collections.concurrent.aspx
or
do your own locking

It depends what "_events" is...
There are some awesome new thread-safe classes in .NET 4 ...
Check out http://msdn.microsoft.com/en-us/library/system.collections.concurrent.aspx

Well, based on that code paste, I'd say no, it's not 100% thread safe.
Of course, you have the source, so you can just add the lock yourself. :)
Actually, as a general rule of thumb, I include the entire CAL project in my solution, at least at the start. It helps a lot in debugging those odd region registration/creation exceptions...

Related

WPF MVVM Opening one View from another

I am just getting started with WPF + MVVM. I think I’ve got the hang of the basics. However, I have a question (hopefully not a stupid one).
I have a View showing a list of customers. I want to edit one of these customers. How do I load the edit view with its separate ViewModel from the List ViewModel.
I’m sure this is a fairly standard scenario, with a fairly straightforward answer, but I’ve spent a chunk of time googling and come up with nothing concrete. Can someone point me in the direction of a straightforward example?
If I’m wrong and it’s not straightforward, what is the best way to do this type of thing?
A common way of doing this (not only in MVVM but it applies well) is to give your list VM access to a so-called service. This service then implements creating and showing the editor (for which it probably uses yet another service).
Example:
/// Always use an interface for the service: it will make it a breeze
/// to test your VM as it decouples it from the actual service implmentation(s)
interface ICustomerEditorService
{
/// Do whatever needed to get the user to edit the Customer passed in,
/// and return the updated one or null if nothing changed.
/// Customer here is likeyly your customer model, or whatever is neede
/// to represent the editable data
Customer EditCustomer( Customer toEdit );
}
class ListViewModel
{
/// service gets passed to constructor, you can use dependency injection
/// like MEF to get this handled easily;
/// when testing, pass a mock here
public ListViewModel( ...., ICustomerEditorService editorService )
{
....
}
private void OnEditButtonClicked()
{
var editedCustomer = editorService.EditCustomer( GetSelectedCustomer() );
//do stuff with editedCustomer
}
}
/// A real implementation
class CustomerEditorService
{
public Customer EditCustomer( Customer toEdit )
{
var vm = new CustomerEditorViewModel( toEdit );
var view = new CustomerEditorView( vm );
if( myWindowManager.ShowDialog( view ) == Cancel )
return null;
return vm.Customer;
}
}

The calling thread cannot access this object because a different thread owns it

I am using the following code.
public partial class SettingApp
{
public SettingApp()
{
InitializeComponent();
Parallel.Invoke(SetDataInTextBox);
}
private void SetDataInTextBox()
{
txtIncAns.Text = Properties.Settings.Default.IncludeAN;
txtIncAuthor.Text = Properties.Settings.Default.IncludeAutt;
txtIncQuo.Text = Properties.Settings.Default.IncludeQU;
txtIncSpBegin.Text = Properties.Settings.Default.IncludeSP;
}
}
The program gives the following error
The calling thread cannot access this object because a different
thread owns it.
Which is the right way
update :
Whether this is right :
public partial class SettingApp
{
private delegate void SetDataInTextBoxDelegate();
public SettingApp()
{
InitializeComponent();
txtIncAns.Dispatcher.BeginInvoke(DispatcherPriority.Normal, new SetDataInTextBoxDelegate(SetDataInTextBox));
}
private void SetDataInTextBox()
{
txtIncAns.Text = Properties.Settings.Default.IncludeAN;
txtIncAuthor.Text = Properties.Settings.Default.IncludeAutt;
txtIncQuo.Text = Properties.Settings.Default.IncludeQU;
txtIncSpBegin.Text = Properties.Settings.Default.IncludeSP;
}
}
Only the UI thread can access UI elements, which I'm guessing is what those txt things are. Parallel.Invoke is in your case not the UI thread, so the exception is thrown when you try to access the .Text property on the controls.
You need to marshal the call across to the UI thread. In WinForms, controls have various ways to help you do this:
if (myControl.InvokeRequired)
{
myControl.Invoke(...);
}
else
{
myControl.Text = "something";
}
MSDN has an article with examples on it here (VS2010):
http://msdn.microsoft.com/en-us/library/757y83z4(v=VS.100).aspx
Update 1:
For WPF the model is similar, but includes the Dispatcher:
myControl.Dispatcher.Invoke(...);
MSDN Forum Post
MSDN Article
Update 2: Of course, it looks like you don't even need to use multi-threaded code here. I would guess the overhead of using the multi-threaded portion is more than the code you eventually call. Simply remove the use of multiple threads from this section and set the properties directly:
public SettingApp()
{
InitializeComponent();
SetDataInTextBox();
}
private void SetDataInTextBox()
{
txtIncAns.Text = Properties.Settings.Default.IncludeAN;
txtIncAuthor.Text = Properties.Settings.Default.IncludeAutt;
txtIncQuo.Text = Properties.Settings.Default.IncludeQU;
txtIncSpBegin.Text = Properties.Settings.Default.IncludeSP;
}
As Adam said, only the UI thread can access UI elements. In the case of WPF, you'd use myControl.Dispatcher.Invoke().
Since these calls are all going to be invoked on the UI thread anyway, you should remove the Parallel.Invoke() and call the method directly.
Just an alternate suggestion. I would move toward databinding your textboxes to properties of a class, even if you don't want to go the full MVVM design, databinding is your friend in WPF. Then if you need the threading, WPF will handle updating controls on UI thread when the property changes, even when the property is changed on another thread.

WPF excessive PropertyChanged events

Typically in the property setter of an object we may want to raise a PropertyChanged event such as,
public event PropertyChangedEventHandler PropertyChanged;
protected void Notify(string property)
{
if (PropertyChanged != null)
{
PropertyChanged(this, new PropertyChangedEventArgs(property));
}
}
public string UserNote
{
get { return _userNote; }
set
{
_userNote = value;
Notify("UserNote");
}
}
In our existing code base I see instances where PropertyChangedEventArgs is being sent null in order to indicate that all properties of the object have changed. This seems inefficient and seems to lead to far more events being triggered than is needed. It also seems to causes issues where objects update each other in a circular fashion.
Is this ever a good practice?
A comment in the code tries to justify it ...
//The purpose of this method is to wire up clients of NotificationBase that are also
//NotificationBases to *their* clients. Consider the following classes:
public class ClassA : NotificationBase
{
public int Foo
{
get { return 123; }
set { Notify("Foo"); }
}
}
public class ClassB : NotificationBase
{
ClassA A = new ClassA();
public ClassB()
{
A.PropertyChanged += AllChanged;
}
public void SetFoo()
{
A.Foo = 456;
}
}
public class ClassC
{
ClassB B = new ClassB();
public ClassC()
{
B.PropertyChanged += delegate { dosomething(); };
B.SetFoo(); // causes "dosomething" above to be called
}
}
/// ClassB.SetFoo calls ClassA.Foo's setter, which calls ClassA.Notify("Foo").
/// The event registration in ClassB's ctor causes ClassB.AllChanged to be called, which calls
/// ClassB.Notify(null) - implying that ALL of ClassB's properties have changed.
/// The event registration in ClassC's ctor causes the "dosomething" delegate to be called.
/// So a Notify in ClassA is routed to ClassC via ClassB's PropertyChanged event.
protected void AllChanged(Object sender, PropertyChangedEventArgs e)
{
Notify(null);
}
Any thoughts much appreciated.
Regards,
Fzzy
This is actually a problem with the design (or its documentation) of PropertyChangedEventArgs. Setting PropertyName to null means "all properties on this object have changed." But unless the class is sealed, or you're using reflection, you can't actually know that all properties on the object have changed. The most you can say is that all of the properties in the object's base class have changed.
This is reason enough to not use this particular convention, in my book, except in the vanishingly small number of cases where I create sealed classes that implement property-change notification.
As a practical matter, what you're really trying to do is just raise one event that tells listeners "a whole bunch of properties on this object have changed, but I'm not going to bother to tell you about them one by one." When you say:
I see instances where PropertyChangedEventArgs is being sent null in order to indicate that all properties of the object have changed. This seems inefficient and seems to lead to far more events being triggered than is needed.
...the actual intent is the exact opposite. If a method changes the Foo, Bar, Baz, and Bat properties on an object, and the object has only four or five properties, raising one event is probably better than raising four. On the other hand, if the object has sixty properties, raising four events is probably better making every one of the object's listeners - even those that aren't looking at those four properties - do whatever they do when the properties that they care about change, because those properties didn't.
The problem is that the property-change notification system, as designed, isn't a fine-grained enough tool for every single job. It's designed to be completely generic, and has no knowledge of a particular application domain built into it.
And that, it seems to me, is what's missing from your design: application domain knowledge.
In your second example, if a Fixture object has (say) ten properties that depend on the value of FixtureStatus, raising ten property-change events may seem a little excessive. Maybe it is. Maybe the object should raise a FixtureStatusChanged event instead. Then classes with knowledge of your application domain can listen to this one event and ignore the PropertyChanged event. (You still raise the PropertyChanged event on the other properties, so that objects that don't know what a FixtureStatusChanged event means can stay current - that is, if it's still necessary for your class to implement INotifyPropertyChanged once you've implemented FixtureStatusChanged.)
A secondary comment: Most classes in the C# universe, if they implement a method that raises the Foo event, call that method OnFoo. This is an important convention: it makes the relationship between the method and the event explicit, and it makes the fact that the code that's calling the method is raising an event easy to recognize. Notify is a weak name for a method in general - notify who? of what? - and in this case it actually obfuscates something that should be made explicit. Property-change notification is tricky enough without your naming convention concealing the fact that it's happening.
Ignoring the other stuff, I'd say the Notify(null) alone is a bad practice. It's not inherently clear what that means, and to a developer working the code 5 years down the line would probably assume that it meant something else unless they happened upon the comments.
I have come across situations wherein computed properties (without setters) need to fire PropertyChangeNotification when some other property i set via a setter.
eg
double Number
{
get { return num;}
set
{
num=value;
OnPropertyChanged("Number");
OnPropertyChanged("TwiceNumber");
}
}
double TwiceNumber
{
get {return _num * 2.0;}
}
As a rule I only do it with get only properties and I don't see why in this case a property firing a change notification on the other is bad. But I think if I do it for any other case I most likely don't know what I am doing!

.net propertychange notification handlers - strings vs. expressions

Using WPF has made me a fan of INotifyPropertyChanged. I like to use a helper that takes an expression and returns the name as a string (see example code below). In lots of applications I see by very proficient programmers, however, I see code that handles the strings raw (see 2nd example below). By proficient I mean MVP types who know how to use Expressions.
To me, the opportunity for having the compiler catch mistakes in addition to easy refactoring makes the Exression approach better. Is there an argument in favor of using raw strings that I am missing?
Cheers,
Berryl
Expression helper example:
public static string GetPropertyName<T>(Expression<Func<T, object>> propertyExpression)
{
Check.RequireNotNull<object>(propertyExpression, "propertyExpression");
switch (propertyExpression.Body.NodeType)
{
case ExpressionType.MemberAccess:
return (propertyExpression.Body as MemberExpression).Member.Name;
case ExpressionType.Convert:
return ((propertyExpression.Body as UnaryExpression).Operand as MemberExpression).Member.Name;
}
var msg = string.Format("Expression NodeType: '{0}' does not refer to a property and is therefore not supported",
propertyExpression.Body.NodeType);
Check.Require(false, msg);
throw new InvalidOperationException(msg);
}
Raw strings example code (in some ViewModelBase type class):
/// <summary>
/// Warns the developer if this object does not have
/// a public property with the specified name. This
/// method does not exist in a Release build.
/// </summary>
[Conditional("DEBUG"), DebuggerStepThrough]
public void VerifyPropertyName(string propertyName) {
// Verify that the property name matches a real,
// public, instance property on this object.
if (TypeDescriptor.GetProperties(this)[propertyName] == null) {
string msg = "Invalid property name: " + propertyName;
if (ThrowOnInvalidPropertyName) throw new Exception(msg);
else Debug.Fail(msg);
}
}
/// <summary>
/// Returns whether an exception is thrown, or if a Debug.Fail() is used
/// when an invalid property name is passed to the VerifyPropertyName method.
/// The default value is false, but subclasses used by unit tests might
/// override this property's getter to return true.
/// </summary>
protected virtual bool ThrowOnInvalidPropertyName { get; private set; }
To me, the opportunity for having the compiler catch mistakes in addition to easy refactoring makes the Exression approach better. Is there an argument in favor of using raw strings that I am missing?
I agree, and personally, use the expression approach in my own code, in most cases.
However, there are two reasons I know of to avoid this:
It is less obvious, especially to less experienced .NET developers. Writing RaisePropertyChanged(() => this.MyProperty ); is not always as obvious to people as RaisePropertyChanged("MyProperty");, and doesn't match framework samples, etc.
There is some performance overhead to using expressions. Personally, I don't feel this is really that meaningful of a reason, since this is usually used in data binding scenarios (which are already slow due to reflection usage), but it is potentially a valid concern.
The benefit of using the TypeDescriptor approach is that it enables dynamic property scenarios based on ICustomTypeDescriptor implementations where the implementation can effectively create dynamic property metadata on the fly for a type that is being described. Consider a DataSet whose "properties" are determined by the result set it is populated with for example.
This is something that expressions does not provide however because it's based on actual type information (a good thing) as opposed to strings.
I wound up spending more time on this than I expected, but did find a solution that has a nice mix of safety/refactorability and performance. Good background reading and an alternate solution using reflection is here. I like Sacha Barber's solution even better (background here.
The idea is to use an expression helper for a property that will participate in change notification, but only take the hit once for it, by storing the resulting PropertyChangedEventArgs in your view model. For example:
private static PropertyChangedEventArgs mobilePhoneNumberChangeArgs =
ObservableHelper.CreateArgs<CustomerModel>(x => x.MobilePhoneNumber);
HTH,
Berryl
Stack walk is slow and lambda expression is even slower. We have solution similar to well known lambda expression but almost as fast as string literal. See
http://zamboch.blogspot.com/2011/03/raising-property-changed-fast-and-safe.html
A CallerMemberName attribute was introduced in .net 4.5
This attribute can only be attached to optional string parameters and if the parameter is not used by caller in the function call then name of the caller will be passed in the string parameter
This removes the need to specify name of property when raising the PropertyChanged event thus it works with refactoring and because the changes are done at compile time there's no difference in performance.
Below is an example of implementation and more info can be found at http://msdn.microsoft.com/en-us/library/system.runtime.compilerservices.callermembernameattribute.aspx and http://msdn.microsoft.com/en-us/library/hh534540.aspx
public class DemoCustomer : INotifyPropertyChanged
{
string _customerName
public string CustomerName
{
get { return _customerNameValue;}
set
{
if (value != _customerNameValue)
{
_customerNameValue = value;
NotifyPropertyChanged();
}
}
}
public event PropertyChangedEventHandler PropertyChanged;
private void NotifyPropertyChanged([CallerMemberName] String propertyName = "")
{
if (PropertyChanged != null)
{
PropertyChanged(this, new PropertyChangedEventArgs(propertyName));
}
}
}

What C-integration problems arise with stackless VM implementations?

By stackless VM I mean implementation which maintains its own stack on the heap instead of using system "C-stack". This has a lot of advantages like continuations and serializable state, but also has some disadvantages when it comes to C-bindings, especially to C-VM-C kind of callbacks (or VM-C-VM).
The question is what exactly these disadvantages are? Could anyone give a good example of a real problem?
It sounds like you're already familiar with some of the disadvantages and the advantages.
Some others:
a) Makes it possible to support proper tail call optimization even if the underlying implementation does not have any support for it
b) Easier to construct things like a language level "stack trace"
c) Easier to add proper continuations, as you pointed out
I recently wrote a simple "Scheme" interpreter in C#, which initially used the .NET stack. I then re-wrote it to use an explicit stack - so perhaps the following will help you:
The first version used the implicit .NET runtime stack...
Initially, it was just a class hierarchy, with different forms (Lambda, Let, etc.) being implementations of the following interface:
// A "form" is an expression that can be evaluted with
// respect to an environment
// e.g.
// "(* x 3)"
// "x"
// "3"
public interface IForm
{
object Evaluate(IEnvironment environment);
}
IEnvironment looked as you'd expect:
/// <summary>
/// Fundamental interface for resolving "symbols" subject to scoping.
/// </summary>
public interface IEnvironment
{
object Lookup(string name);
IEnvironment Extend(string name, object value);
}
For adding "builtins" to my Scheme interpreter, I initially had the following interface:
/// <summary>
/// A function is either a builtin function (i.e. implemented directly in CSharp)
/// or something that's been created by the Lambda form.
/// </summary>
public interface IFunction
{
object Invoke(object[] args);
}
That was when it used the implicit .NET runtime stack. There was definitely less code, but it was impossible to add things like proper tail recursion, and most importantly, it was awkward for my interpreter to be able to provide a "language level" stack trace in the case of a runtime error.
So I rewrote it to have an explicit (heap allocated) stack.
My "IFunction" interface had to change to the following, so that I could implement things like "map" and "apply", which call back into the Scheme interpreter:
/// <summary>
/// A function that wishes to use the thread state to
/// evaluate its arguments. The function should either:
/// a) Push tasks on to threadState.Pending which, when evaluated, will
/// result in the result being placed on to threadState.Results
/// b) Push its result directly on to threadState.Results
/// </summary>
public interface IStackFunction
{
void Evaluate(IThreadState threadState, object[] args);
}
And IForm changed to:
public interface IForm
{
void Evaluate(IEnvironment environment, IThreadState s);
}
Where IThreadState is as follows:
/// <summary>
/// The state of the interpreter.
/// The implementation of a task which takes some arguments,
/// call them "x" and "y", and which returns an argument "z",
/// should follow the following protocol:
/// a) Call "PopResult" to get x and y
/// b) Either
/// i) push "z" directly onto IThreadState using PushResult OR
/// ii) push a "task" on to the stack which will result in "z" being
/// pushed on to the result stack.
///
/// Note that ii) is "recursive" in its definition - that is, a task
/// that is pushed on to the task stack may in turn push other tasks
/// on the task stack which, when evaluated,
/// ... ultimately will end up pushing the result via PushResult.
/// </summary>
public interface IThreadState
{
void PushTask(ITask task);
object PopResult();
void PushResult(object result);
}
And ITask is:
public interface ITask
{
void Execute(IThreadState s);
}
And my main "event" loop is:
ThreadState threadState = new ThreadState();
threadState.PushTask(null);
threadState.PushTask(new EvaluateForm(f, environment));
ITask next = null;
while ((next = threadState.PopTask()) != null)
next.Execute(threadState);
return threadState.PopResult(); // Get what EvaluateForm evaluated to
EvaluateForm is just a task that calls IForm.Evaluate with a specific environment.
Personally, I found this new version much "nicer" to work with from an implementation point of view - easy to get a stack trace, easy to make it implement full continuations (although... I haven't done this as yet - need to make my "stacks" persistent linked-lists rather than using C# Stack, and ITask "returns" the new ThreadState rather than mutating it so that I can have a "call-continuation" task)... etc. etc.
Basically, you're just less dependent on the underlying language implementation.
About the only downside I can find is performance... But in my case, it's just an interpreter so I don't care that much about performance anyway.
I'd also point you to this very nice article on the benefits of re-writing recursive code as iterative code with a stack, by one of the authors of the KAI C++ compiler: Considering Recursion
After e-mail conversation with Steve Dekorte (author of Io programming language) and Konstantin Olenin, I've found a problem and a (partial) solution to it.
Imagine the call from VM to C function, which calls back VM method. During the period of time when VM executes the callback, portion of VM state lays outside of the VM: in the C stack and registers. If you would save VM state at that moment it is guaranteed that you couldn't restore the state correctly next time VM is loaded.
The solution is to model VM as a message-receiving actor: VM can send async notifications to the native code and native code can send async notifications to the VM. That is, in the single-threaded environment, when VM gains control, no additional state is stored outside of it (except data irrelevant to VM runtime).
This does not mean that you can correctly restore VM state in any circumstances, but at least, you can build your own reliable system on top of it.

Resources