One of the greatest understatements I've heard in a newsgroup was made by Patricia Shanahan, in a Java newsgroup in 2001: "Multi-threaded programming needs a little care." Multi-threading is probably one of the worst understood aspects of programming, and these days almost all application programmers need to understand it to some extent. This article acts as an introduction to multi-threading and gives some hints and tips for how to do it safely. Warning: I'm not an expert on the subject, and when the real experts start discussing it in detail, my head starts to spin somewhat. However, I've tried to pay attention to those who know what they're doing, and hopefully the contents of this article form at least part of a multi-threading "best practice".
This article uses the C# type shorthands throughout - int
for Int32 etc. I hope this makes it easier for C# developers
to read, and won't impede any other developers too much. It also only
talks about the C# ways of declaring variables to be volatile and locking
monitors. Developers using other languages can find the equivalents in their
own preferred environment, I'm sure.
The fact that you're reading this article in the first place means you probably have at least some idea of what multi-threading is about: it's basically trying to do more than one thing at a time within a process.
So, what is a thread? A thread (or "thread of execution") is a sort of context in which code is running. Any one thread follows program flow for wherever it is in the code, in the obvious way. Before multi-threading, effectively there was always one thread running for each process in an operating system (and in many systems, there was only one process running anyway). If you think of processes running in parallel in an operating system (e.g. a browser downloading a file and a word processor allowing you to type, both "at the same time"), then apply the same kind of thinking within a single process, that's a reasonable way to visualise threading.
Multi-threading can occur in a "real" sense, in that a multi-processor box may have more than one processor executing instructions for a particular process at a time, or it may be effectively "simulated" by multiple threads executing in sequence: first some code for thread 1 is executed, then some code for thread 2, then back to thread 1 etc. In this situation, if both thread 1 and thread 2 are "compute bound" (all they're doing is computation, without waiting for any input from the network, or file system, or user etc) then that won't actually speed things up at all - in fact, it'll slow things down as the operating system has to switch between threads, and the memory cache probably won't be as effective. However, much of today's computing involves waiting for something to happen, and during that time the processor can be doing something else. Intel's "Hyper-Threading" technology which is on some of its more recent chips (bearing in mind that this article was written in early 2004!) is a sort of hybrid between this "real" and "simulated" threading - for more information, see Intel's web page on the subject [http://www.intel.com/technology/hyperthread/].
.NET has been designed from the start to support multi-threaded operation. There
are two main ways of multi-threading which .NET encourages: starting your own threads
with ThreadStart delegates, and using the ThreadPool class
either directly (using ThreadPool.QueueUserWorkItem) or indirectly using
asynchronous methods (such as Stream.BeginRead, or calling BeginInvoke
on any delegate).
In general, you should create a new thread "manually" for long-running tasks, and use the thread pool only for brief jobs. The thread pool can only run so many jobs at once, and some framework classes use it internally, so you don't want to block it with a lot of tasks which need to block for other things. The examples in this article mostly use manual thread creation. On the other hand, for short-running tasks, particularly those created often, the thread pool is an excellent choice.
Here is virtually the simplest threading example which actually shows something happening:
using System; using System.Threading; public class Test { static void Main() { ThreadStart job = new ThreadStart(ThreadJob); Thread thread = new Thread(job); thread.Start(); for (int i=0; i < 5; i++) { Console.WriteLine ("Main thread: {0}", i); Thread.Sleep(1000); } } static void ThreadJob() { for (int i=0; i < 10; i++) { Console.WriteLine ("Other thread: {0}", i); Thread.Sleep(500); } } } |
The code creates a new thread which runs the ThreadJob
method, and starts it. That thread counts from 0 to 9 fairly fast (about
twice a second) while the main thread counts from 0 to 4 fairly slowly
(about once a second). The way they count at different speeds is by each
of them including a call to Thread.Sleep, which just makes
the current thread sleep (do nothing) for the specified period of time.
Between each count in the main thread we sleep for 1000ms, and between
each count in the other thread we sleep for 500ms. Here are the results
from one test run on my machine:
Main thread: 0 Other thread: 0 Other thread: 1 Main thread: 1 Other thread: 2 Other thread: 3 Main thread: 2 Other thread: 4 Other thread: 5 Main thread: 3 Other thread: 6 Other thread: 7 Main thread: 4 Other thread: 8 Other thread: 9 |
One important thing to note here is that although the above is very
regular, that's by chance. There's nothing to stop the first "Other
thread" line coming first, or the pattern being slightly off -
Thread.Sleep is always going to be somewhat approximate, and
there's no guarantee that the sleeping thread will immediately start
running as soon as the sleep finishes. (It will become able to
run, but another thread may be currently running, and on a single
processor machine that means the thread which has just "woken up" will
have to wait until the thread scheduler decides to give it some processor
time before it next does anything.)
As with all delegates, there's nothing to restrict you to static methods,
or methods within the class that the delegate is used from. You need to
have access to the method, of course, and if you want to specify an
instance method, you have to use a particular instance. Here's another
version of the program above, using an instance method in a different
class. If the Count method had been static, the value of the
job variable would have been new ThreadStart(Counter.Count).
Most examples given in this article use methods within the same
class, but that's just for brevity and simplicity.
using System; using System.Threading; public class Test { static void Main() { Counter foo = new Counter(); ThreadStart job = new ThreadStart(foo.Count); Thread thread = new Thread(job); thread.Start(); for (int i=0; i < 5; i++) { Console.WriteLine ("Main thread: {0}", i); Thread.Sleep(1000); } } } public class Counter { public void Count() { for (int i=0; i < 10; i++) { Console.WriteLine ("Other thread: {0}", i); Thread.Sleep(500); } } } |
Often when you start a thread, you want to give it some parameters -
usually some data it has to process, a queue to wait on, etc. The
ThreadStart delegate doesn't take any parameters, so the
information has to be stored somewhere else if you are going to create a
new thread yourself. Typically, this means creating a new instance of a
class, and using that instance to store the information. Often the class
itself can contain the delegate used for starting the thread. For
example, we might have a program which needs to fetch the contents of
various URLs, and wants to do it in the background. You could write code
like this:
public class UrlFetcher { string url public UrlFetcher (string url) { this.url = url; } public void Fetch() { // use url here } } [... in a different class ...] UrlFetcher fetcher = new UrlFetcher (myUrl); new Thread (new ThreadStart (fetcher.Fetch)).Start(); |
In some cases, you actually just wish to call a method in some class (possibly the currently executing class) with a specific parameter. In that case, you may wish to use a nested class whose purpose is just to make this call - the state is stored in the class, and the delegate used to start the thread just calls the "real" method with the appropriate parameter. (Note that the object on which to call the method in the first place will also be required as state unless the method is static.)
One alternative to starting the thread using a ThreadStart delegate
is to use the thread pool, either using ThreadPool.QueueUserWorkItem
or by calling a delegate asynchronously. Both of these are covered later on in
a more detailed discussion of the thread pool, which
also contains examples. Note that calling a delegate asynchronously allows you
to specify multiple parameters, and those parameters are strongly typed.
One of the enhancements to C# in version 2.0 is anonymous methods.
These allow you to specify blocks of code as methods within other methods, and use those methods
as delegates. You can access variables (including local variables and parameters of the "outside" method)
within the anonymous method. For example, using an anonymous method to fetch a URL using a normal
ThreadStart delegate (and using inference of delegate type too):
ThreadStart starter = delegate { Fetch (myUrl); };
new Thread(starter).Start();
|
(This could have all been expressed within a single step, creating both the thread
and the delegate in the same line of code, but I believe the above is more readable.)
Here's similar code to use a WaitCallback and queue the job in
ThreadPool:
WaitCallback callback = delegate (object state) { Fetch ((string)state); }; ThreadPool.QueueUserWorkItem (callback, myUrl); |
Note the way that the state is declared.
ParameterizedThreadStart
In .NET 2.0, there is a new delegate, ParameterizedThreadStart, which takes
a parameter of type object. You can create a thread using an instance of
this delegate instead of just ThreadStart, and a new overload to Thread.Start
allows you to specify the value to be passed to the new thread. This is simple, but only accepts
a single parameter and isn't type-safe (just like the options when using thread pool threads).
The earlier code could then be rewritten as:
[In some method or other] Thread t = new Thread (new ParameterizedThreadStart(FetchUrl)); t.Start (myUrl); [And the actual method...] static void FetchUrl(object url) { // use url here, probably casting it to a known type before use } |
Well, that's shown you how to create a new thread and start it. For a very few cases, it really is as simple as that - just occasionally, you end up with a thread which doesn't need access to any data other than its own (the counters in this case). Far more commonly, however, you need threads to access the same data, sooner or later - and that's where the problems start. Let's take a very simple program to start with:
using System; using System.Threading; public class Test { static int count=0; static void Main() { ThreadStart job = new ThreadStart(ThreadJob); Thread thread = new Thread(job); thread.Start(); for (int i=0; i < 5; i++) { count++; } thread.Join(); Console.WriteLine ("Final count: {0}", count); } static void ThreadJob() { for (int i=0; i < 5; i++) { count++; } } } |
This is very straightforward - each of the threads just increments the count
variable, and then the main thread displays the final value of count at the end.
The only really new thing here is the call in the main thread to Thread.Join,
which basically pauses the main thread until the other thread has completed.
So, the result should always be Final count: 10, right? Well, no. In fact,
chances are that that will be the result if you run the above code - but it isn't
guaranteed to be. There are two reasons for this - one fairly simple, and one much subtler.
We'll leave the subtle one for the moment, and just consider the simple one.
The statement count++; actually does three things: it reads the current value
of count, increments that number, and then writes the new value back to the
count variable. Now, if one thread gets as far as reading the current value,
then the other thread takes over, does the whole increment operation, and then the first
thread gets control again, its idea of the value of count is out of date - so
it will increment the old value, and write that newly incremented (but wrong) value
back into the variable.
The easiest way of showing this is by separating the three operations and introducing some
Sleep calls into the code, just to make it more likely that the threads will
clash heads, as it were. Note that introducing Sleep calls should never change
the correctness of a program, in terms of threading - any thread can go to sleep at any
time, basically. In other words, you can never rely on two operations both happening without
another thread doing stuff in between. I've also put some diagnostics in to make it clearer
what's happening. The "main" thread's activities appear on the left, while the "other" thread's
activities are on the right. Here's the code:
using System; using System.Threading; public class Test { static int count=0; static void Main() { ThreadStart job = new ThreadStart(ThreadJob); Thread thread = new Thread(job); thread.Start(); for (int i=0; i < 5; i++) { int tmp = count; Console.WriteLine ("Read count={0}", tmp); Thread.Sleep(50); tmp++; Console.WriteLine ("Incremented tmp to {0}", tmp); Thread.Sleep(20); count = tmp; Console.WriteLine ("Written count={0}", tmp); Thread.Sleep(30); } thread.Join(); Console.WriteLine ("Final count: {0}", count); } static void ThreadJob() { for (int i=0; i < 5; i++) { int tmp = count; Console.WriteLine ("\t\t\t\tRead count={0}", tmp); Thread.Sleep(20); tmp++; Console.WriteLine ("\t\t\t\tIncremented tmp to {0}", tmp); Thread.Sleep(10); count = tmp; Console.WriteLine ("\t\t\t\tWritten count={0}", tmp); Thread.Sleep(40); } } } |
... and here's one set of results I saw ...
Read count=0
Read count=0
Incremented tmp to 1
Written count=1
Incremented tmp to 1
Written count=1
Read count=1
Incremented tmp to 2
Read count=1
Written count=2
Read count=2
Incremented tmp to 2
Incremented tmp to 3
Written count=2
Written count=3
Read count=3
Read count=3
Incremented tmp to 4
Incremented tmp to 4
Written count=4
Written count=4
Read count=4
Read count=4
Incremented tmp to 5
Written count=5
Incremented tmp to 5
Written count=5
Read count=5
Incremented tmp to 6
Written count=6
Final count: 6
|
Just looking at the first few lines shows exactly the nasty behaviour described before
the code: the main thread has read the value 0, the other thread has incremented
count to 1, and then the main thread has incremented its "stale" value
from 0 to 1, and written that value to the variable. The same thing happens a few more
times, and the end result is that count is 6, instead of 10.
Monitor.Enter/Exit and the lock statementWhat we need to fix the problem above is to make sure that while one thread is in a read/increment/write operation, no other threads can try to do the same thing. This is where monitors come in. Every object in .NET has a (theoretical) monitor associated with it. A thread can enter (or acquire) a monitor only if no other thread has currently "got" it. Once a thread has acquired a monitor, it can acquire it more times, or exit (or release) it. The monitor is only available to other threads again once it has been exited as many times as it was entered. If a thread tries to acquire a monitor which is owned by another thread, it will block until it is able to acquire it. (There may be more than one thread trying to acquire the monitor, in which case when the current owner thread releases it for the last time, only one of the threads will acquire it - the other one will have to wait for the new owner to release it too.)
In our example, we want exclusive access to the count
variable while we're performing the increment operation. First we need to
decide on an object to use for locking. I'll discuss this choice in more
detail later, but for the moment we'll
introduce a new variable just for the purposes of locking:
countLock. This is initialised to be a reference a new
object, and thereafter is never changed. It's important that it's not
changed - otherwise one thread would be locking on one object's monitor,
and another object might be locking on a different object's monitor, so
they could interfere with each other just like they did before.
We then simply need to put each increment operation in a Monitor.Enter and Monitor.Exit
pair:
using System; using System.Threading; public class Test { static int count=0; static readonly object countLock = new object(); static void Main() { ThreadStart job = new ThreadStart(ThreadJob); Thread thread = new Thread(job); thread.Start(); for (int i=0; i < 5; i++) { Monitor.Enter(countLock); int tmp = count; Console.WriteLine ("Read count={0}", tmp); Thread.Sleep(50); tmp++; Console.WriteLine ("Incremented tmp to {0}", tmp); Thread.Sleep(20); count = tmp; Console.WriteLine ("Written count={0}", tmp); Monitor.Exit(countLock); Thread.Sleep(30); } thread.Join(); Console.WriteLine ("Final count: {0}", count); } static void ThreadJob() { for (int i=0; i < 5; i++) { Monitor.Enter(countLock); int tmp = count; Console.WriteLine ("\t\t\t\tRead count={0}", tmp); Thread.Sleep(20); tmp++; Console.WriteLine ("\t\t\t\tIncremented tmp to {0}", tmp); Thread.Sleep(10); count = tmp; Console.WriteLine ("\t\t\t\tWritten count={0}", tmp); Monitor.Exit(countLock); Thread.Sleep(40); } } } |
The results look a lot better this time:
Read count=0
Incremented tmp to 1
Written count=1
Read count=1
Incremented tmp to 2
Written count=2
Read count=2
Incremented tmp to 3
Written count=3
Read count=3
Incremented tmp to 4
Written count=4
Read count=4
Incremented tmp to 5
Written count=5
Read count=5
Incremented tmp to 6
Written count=6
Read count=6
Incremented tmp to 7
Written count=7
Read count=7
Incremented tmp to 8
Written count=8
Read count=8
Incremented tmp to 9
Written count=9
Read count=9
Incremented tmp to 10
Written count=10
Final count: 10
|
The fact that the increments were strictly alternating here is just due to the sleeps - in a more normal system there could be two increments in one thread, then three in another, etc. The important thing is that they would always be thread-safe: each increment would be isolated from each other increment, with only one being processed at a time.
There's a chance - a tiny chance, but a chance nonetheless - that the code above would hang,
however. If part of the increment operation (one of the calls to Console.WriteLine, for instance)
threw an exception, the thread would still own the monitor, so the other thread would never be able
to acquire it and move on. The obvious solution to this (if you're used to exception handling, at least) is to
put the call to Monitor.Exit in a finally block, with everything after the call to
Monitor.Enter in a try block. Just like the using statement which puts a call to
Dispose in a finally block automatically, C# provides the lock statement to call
Monitor.Enter and Monitor.Exit with a try/finally block automatically. This makes
it much easier to get synchronization right, as you don't end up having to check for "balanced" calls to
Enter and Exit everywhere. It also makes sure that you don't try to release a monitor
you don't own: in the code we had above, if we changed the value of countLock to be a reference to a
different object within the increment operation, we'd have failed to release the monitor we owned, and tried to
release a monitor we didn't own - which would (in theory) have caused a SynchronizationLockException.
(In fact, the exception wouldn't have been thrown because there's a bug in the framework in version
1.0/1.1, but that's another story.)
The lock statement automatically takes a copy of the reference you specify, and calls both
Enter and Exit with it. (In the example above, and everywhere else in this article,
variables used to hold locks are declared as read-only. I have yet to come across a good reason to change
what a particular piece of code locks on.)
So, we can rewrite our previous code into the somewhat clearer and more robust code below:
using System; using System.Threading; public class Test { static int count=0; static readonly object countLock = new object(); static void Main() { ThreadStart job = new ThreadStart(ThreadJob); Thread thread = new Thread(job); thread.Start(); for (int i=0; i < 5; i++) { lock (countLock) { int tmp = count; Console.WriteLine ("Read count={0}", tmp); Thread.Sleep(50); tmp++; Console.WriteLine ("Incremented tmp to {0}", tmp); Thread.Sleep(20); count = tmp; Console.WriteLine ("Written count={0}", tmp); } Thread.Sleep(30); } thread.Join(); Console.WriteLine ("Final count: {0}", count); } static void ThreadJob() { for (int i=0; i < 5; i++) { lock (countLock) { int tmp = count; Console.WriteLine ("\t\t\t\tRead count={0}", tmp); Thread.Sleep(20); tmp++; Console.WriteLine ("\t\t\t\tIncremented tmp to {0}", tmp); Thread.Sleep(10); count = tmp; Console.WriteLine ("\t\t\t\tWritten count={0}", tmp); } Thread.Sleep(40); } } } |
The second major problem of multi-threading is that of deadlocks. Simply put, this is when two threads each hold a monitor that the other one wants. Each blocks, waiting for the monitor that it's waiting for to be released - and so the monitors are never released, and the application hangs (or at least those threads involved in the deadlock hang).
Here's an example of a program exhibiting deadlock:
using System; using System.Threading; public class Test { static readonly object firstLock = new object(); static readonly object secondLock = new object(); static void Main() { new Thread(new ThreadStart(ThreadJob)).Start(); // Wait until we're fairly sure the other thread // has grabbed firstLock Thread.Sleep(500); Console.WriteLine ("Locking secondLock"); lock (secondLock) { Console.WriteLine ("Locked secondLock"); Console.WriteLine ("Locking firstLock"); lock (firstLock) { Console.WriteLine ("Locked firstLock"); } Console.WriteLine ("Released firstLock"); } Console.WriteLine("Released secondLock"); } static void ThreadJob() { Console.WriteLine ("\t\t\t\tLocking firstLock"); lock (firstLock) { Console.WriteLine("\t\t\t\tLocked firstLock"); // Wait until we're fairly sure the first thread // has grabbed secondLock Thread.Sleep(1000); Console.WriteLine("\t\t\t\tLocking secondLock"); lock (secondLock) { Console.WriteLine("\t\t\t\tLocked secondLock"); } Console.WriteLine ("\t\t\t\tReleased secondLock"); } Console.WriteLine("\t\t\t\tReleased firstLock"); } } |
And the results:
Locking firstLock
Locked firstLock
Locking secondLock
Locked secondLock
Locking firstLock
Locking secondLock
|
(You'll need to hit Ctrl-C or something similar to kill the program.) As you can see, each thread
grabs one lock and then tries to grab the other. The calls to Thread.Sleep have been engineered
so that they will try to do so at inopportune times, and deadlock.
Deadlocks can be a real pain in the neck to debug - they're hard to diagnose, they can crop up seemingly randomly (i.e. they're hard to reproduce) and once you've found out why there's a deadlock, it's not always obvious what the best course of action is. Locking strategies always need to be treated very carefully; you can't just start removing all the lock statements from your code, or you'll end up with data races etc.
The solution (which is easier to describe than to achieve) is to make sure that you always take out locks in the
same order: in the code above, you might decide to never acquire secondLock unless you already had
firstLock. Within one class, it's relatively straightforward to achieve this. It's when you have
calls between classes (including delegates, where you don't really know what you're calling) that things
get somewhat hairy. If possible, you should avoid making method calls outside your own class within a lock,
unless you know pretty definitely that that method won't itself need to lock anything.
Monitor methods
If you looked at the documentation for Monitor.Enter and Monitor.Exit, you will
no doubt have seen that there are various other methods in the Monitor class, all of which
can be useful at different times.
Monitor.TryEnter is the easiest one to describe - it simply attempts to acquire a lock,
but doesn't block (or only blocks for a given period of time) if the lock cannot be acquired.
The other methods (Wait, Pulse and
PulseAll) all go together. They're used to signal
between threads. The idea is that one thread calls
Wait, which makes it block until another thread
calls Pulse or PulseAll. The
difference between Pulse and
PulseAll is how many threads are woken up:
Pulse only wakes up a single waiting thread;
PulseAll wakes up all threads waiting on that
monitor. That doesn't mean they'll all instantly start
running, however: in order to call any of these three methods,
the thread has to own the monitor of the object reference it
passes in as a parameter. When calling Wait, the
monitor is released, but then needs to be reacquired before
the thread will actually run. This means blocking again until
the thread which calls Pulse or
PulseAll releases the monitor (which it must have
in order to pulse the monitor in the first place) - and if
multiple threads are woken up, they'll all try to acquire the
monitor, which only one can have at a time, of course. Just to
repeat: calling Wait unlocks the monitor you're
waiting on. This is an important point, because otherwise the
code looks like it'll just deadlock!
The most common use of these methods is in producer/consumer relationships, where one thread is putting work items on a queue, and another thread is taking them off. The consumer thread typically takes items off the list until it's empty, then waits on a lock. The producer thread pulses the lock when it adds an item to the list (or, if you're worried about efficiency, it can pulse the lock only when it adds an item to a previously-empty list). Here's a sample:
using System; using System.Collections; using System.Threading; public class Test { static ProducerConsumer queue; static void Main() { queue = new ProducerConsumer(); new Thread(new ThreadStart(ConsumerJob)).Start(); Random rng = new Random(0); for (int i=0; i < 10; i++) { Console.WriteLine ("Producing {0}", i); queue.Produce(i); Thread.Sleep(rng.Next(1000)); } } static void ConsumerJob() { // Make sure we get a different random seed from the // first thread Random rng = new Random(1); // We happen to know we've only got 10 // items to receive for (int i=0; i < 10; i++) { object o = queue.Consume(); Console.WriteLine ("\t\t\t\tConsuming {0}", o); Thread.Sleep(rng.Next(1000)); } } } public class ProducerConsumer { readonly object listLock = new object(); Queue queue = new Queue(); public void Produce(object o) { lock (listLock) { queue.Enqueue(o); // We always need to pulse, even if the queue wasn't // empty before. Otherwise, if we add several items // in quick succession, we may only pulse once, waking // a single thread up, even if there are multiple threads // waiting for items. Monitor.Pulse(listLock); } } public object Consume() { lock (listLock) { // If the queue is empty, wait for an item to be added // Note that this is a while loop, as we may be pulsed // but not wake up before another thread has come in and // consumed the newly added object. In that case, we'll // have to wait for another pulse. while (queue.Count==0) { // This releases listLock, only reacquiring it // after being woken up by a call to Pulse Monitor.Wait(listLock); } return queue.Dequeue(); } } } |
Here are some results I got:
Producing 0
Consuming 0
Producing 1
Consuming 1
Producing 2
Consuming 2
Producing 3
Consuming 3
Producing 4
Producing 5
Consuming 4
Producing 6
Consuming 5
Consuming 6
Producing 7
Consuming 7
Producing 8
Consuming 8
Producing 9
Consuming 9
|
Now, there's nothing stopping you from having more than one consumer or producer in the above. Everything will play nicely, and each produced object will only be consumed once, and will be consumed (almost) immediately if there are any consumers waiting for work.
The reason for having both Pulse and PulseAll
is for different situations, where you're waiting on different
conditions. If either there'll only be one thread waiting, or (as is the
case above) any thread can consume any produced object, you
can just use Pulse. If there are several threads waiting on
the object, that ends up being more efficient than PulseAll
- there's no point in waking up a bunch of threads if you know that only
one of them is going to be able to make progress, and that it doesn't
matter which you wake up. Sometimes, however, different threads are
waiting on different conditions, but all waiting on the same monitor. In
that case, you need to use PulseAll so that you make sure
that the thread which is waiting for whatever condition has just occurred
is able to notice it and make progress.
Note that using these methods can easily lead to deadlock - if thread A holds locks X and Y, and waits on Y, but thread B needs to acquire lock X before acquiring and then pulsing Y, thread B won't be able to do anything. Only the lock which is waited on is released, not all the locks the waiting thread owns. Usually you should ensure that prior to waiting, a thread only owns the lock it's going to wait on. Sometimes this isn't possible, but in those cases you should think extra carefully about how everything is going to work.
Monitor.Wait/Pulse isn't the only way of waiting for
something to happen in one thread and telling that thread that it's
happened in another. Win32 programmers have been using various other
mechanisms for a long time, and these are exposed by the
AutoResetEvent, ManualResetEvent and
Mutex classes, all of which derive from
WaitHandle. All of these classes are in the
System.Threading namespace. (The Win32
Semaphore mechanism does not have a managed wrapper in .NET
1.1. It's present in .NET 2.0, but if you need to use it before then, you
could either wrap it yourself using P/Invoke, or write your own counting
semaphore class.)
Some people may be surprised to learn that using these classes can be
significantly slower than using the various Monitor methods.
I believe this is because going "out" of managed code into native Win32
calls and back "in" again is expensive compared with the entirely managed
view of things which Monitor provides.
WaitHandle itself only exposes a few useful instance methods/properties:
WaitOne() - used to wait for the handle to be free/signalled.
The exact meaning of this depends on the concrete type being used (Mutex,
AutoResetEvent or ManualResetEvent).
Close()/Dispose() - used to release the resources used by the handle.
Handle - used to get the native handle being wrapped. Most developers
won't need to use this.
In addition, it has two useful static methods which deal with sets of WaitHandles:
WaitAny() - used to wait for any of the handles in a set to be free/signalled.
WaitAll() - used to wait for all of the handles in a set to be free/signalled.
All of the WaitXXX() methods have overloads allowing you
to specify a timeout and whether or not to exit the synchronization
domain. Frankly, it's unclear exactly what a synchronization domain is -
there don't seem to be any clear and free articles about it on the net or
in MSDN. If you know what they are, please mail me at
skeet@pobox.com (skeet@pobox.com) so I can include the
information in this article - or leave it as "definitely beyond the scope
of this article".
The two "event" classes (which are entirely different from .NET events
- don't get the two confused) come as a sort of pair, and are very
similar. You can think of them like doors - when they're in the
"signalled" (or "set") state they're open, and when they're in the
"non-signalled" (or "reset") state, they're closed. A call to
WaitOne() waits for the door to be opened so the thread can
"go through it" in some sense. The difference between the two classes is
that an AutoResetEvent will reset itself to the
non-signalled state immediately after a call to WaitOne() -
it's as if anyone going through the door closes it behind them. With a
ManualResetEvent, you have to tell the thread to reset it
(close the door) when you want to make calls to WaitOne()
block again. Both classes can manually be set or reset at any time, by any
thread, using the Set and Reset methods,
and can be created in the signalled/set or non-signalled/reset state.
(These methods return a boolean value saying whether or not they were
successful, but the documentation doesn't state why they might fail.)
Here's some sample code which simulates 10 runners. Each runner is
passed a ManualResetEvent which is initially non-signalled.
When the runner completes the race, it signals the event. The main thread
uses WaitHandle.WaitAny to wait for the first runner to finish,
and uses the value returned by the method to say who won the race. It
then uses WaitHandle.WaitAll to wait for everyone to finish.
Note that if we'd used AutoResetEvent instead, we'd have to call
Set on the event of the winner, as it would have been reset
when we detected it being set with the call to WaitAny.
using System; using System.Threading; class Test { static void Main() { ManualResetEvent[] events = new ManualResetEvent[10]; for (int i=0; i < events.Length; i++) { events[i] = new ManualResetEvent(false); Runner r = new Runner(events[i], i); new Thread(new ThreadStart(r.Run)).Start(); } int index = WaitHandle.WaitAny(events); Console.WriteLine ("***** The winner is {0} *****", index); WaitHandle.WaitAll(events); Console.WriteLine ("All finished!"); } } class Runner { static readonly object rngLock = new object(); static Random rng = new Random(); ManualResetEvent ev; int id; internal Runner (ManualResetEvent ev, int id) { this.ev = ev; this.id = id; } internal void Run() { for (int i=0; i < 10; i++) { int sleepTime; // Not sure about the thread safety of Random... lock (rngLock) { sleepTime = rng.Next(2000); } Thread.Sleep(sleepTime); Console.WriteLine ("Runner {0} at stage {1}", id, i); } ev.Set(); } } |
Whereas Auto/ManualResetEvent have a lot in common with
using Monitor.Wait/Pulse, Mutex has even
more in common with Monitor.Enter/Exit. A mutex has a
count of the number of times it's been acquired, and a thread which
is the current owner. If the count is zero, it has no owner and it
can be acquired by anyone. If the count is non-zero, the current
owner can acquire it however many times they like without blocking,
but any other thread has to wait until the count becomes zero before
they can acquire it. The WaitXXX() methods are used to
acquire the mutex, and ReleaseMutex() is used by the
owner thread to decrease the count by one. Only the owner can
decrease the count.
So far, so much like Monitor. The difference is that
a Mutex is a cross-process object - the same mutex
can be used in many processes, if you give it a name. A thread in
one process can wait for a thread in another process to release
the mutex, etc. When you construct a named mutex, you should be
careful about making assumptions as to whether or not you will
be able to acquire initial ownership of it. Fortunately, there is
a constructor which allows the code to detect whether the system
has created a whole new mutex or whether it's used an existing one.
If the constructor requested initial ownership, it will only have
been granted it if it created a new mutex - even if the existing
mutex can immediately be acquired.
Mutex names should start with either "Local\" or "Global\" to indicate whether they should be created in the local or global namespace respectively. (I believe that local is the default, but why take the risk? Make it explicit in the name.) If you create a mutex in the global namespace, it is shared with other users logged into the same machine. If you create a mutex in the local namespace, it is specific to the current user. Make sure you pick a suitably unique name so you don't clash with other programs.
To be honest, I think the principle use that mutexes will be put
to in .NET is the one mentioned earlier - detecting that another
instance of an application is already running. Most people don't
need inter-process communication on this kind of level. The other
use is to enable you to block until either one or all of a set of
WaitHandles is released. For other purposes, where
Monitor is good enough, I suggest using that - especially
as C# has the lock statement specifically to support it.
Here's an example of detecting a running application, however:
using System; using System.Threading; class Test { static void Main() { bool firstInstance; using (Mutex mutex = new Mutex(true, @"Global\Jon.Skeet.MutexTestApp", out firstInstance)) { if (!firstInstance) { Console.WriteLine ("Other instance detected; aborting."); return; } Console.WriteLine ("We're the only instance running - yay!"); for (int i=0; i < 10; i++) { Console.WriteLine (i); Thread.Sleep(1000); } } } } |
Run the example in two different console windows - one will count to
ten slowly; the other will abort after it detects that the other
application instance is running. Note the using statement
around the mutex: this should extend across the whole of the application's
execution, otherwise another instance would be able to create a new mutex
with the same name, after the old one had been destroyed. For instance,
suppose you use a local variable without a using statement, like this:
using System; using System.Threading; class Test { static void Main() { bool firstInstance; // Bad code - do not use! Mutex mutex = new Mutex(true, @"Global\Jon.Skeet.MutexTestApp", out firstInstance); if (!firstInstance) { Console.WriteLine ("Other instance detected; aborting."); return; } Console.WriteLine ("We're the only instance running - yay!"); for (int i=0; i < 10; i++) { Console.WriteLine (i); Thread.Sleep(1000); } } } |
In that case, you'd probably find that everything would work fine under debug, where
the GC is very conservative about what it collects. When not running under the debugger,
however, the GC can tell that the mutex variable isn't used after its initial
assignment, so for the main duration of the app, it can be garbage collected at any time
- and that destroys the mutex! The using statement shown earlier is
only one way round this. You could make it a static variable instead, or use
GC.KeepAlive(mutex); at the end of the method to make sure that the GC doesn't
ignore the variable.
Earlier, when introducing the topic of data races, I mentioned that there was a more subtle reason why the first attempt at the code wasn't thread-safe. It's to do with volatility and data caching. Here's a sample which will make explaining the topic somewhat easier:
using System; using System.Threading; public class Test { static bool stop; static void Main() { ThreadStart job = new ThreadStart(ThreadJob); Thread thread = new Thread(job); thread.Start(); // Let the thread start running Thread.Sleep(2000); // Now tell it to stop counting stop = true; } static void ThreadJob() { int count = 0; while (!stop) { Console.WriteLine ("Extra thread: count {0}", count); Thread.Sleep(100); count++; } } } |
Now, this code is fairly simple. We have a boolean variable (stop)
which is polled by the new thread - it will keep counting until it notices that
stop is true. In the main thread, we pause for a couple of seconds
and then set stop to true.
So, the new thread should count for a couple of seconds and then stop, right?
Well, in fact that's what will almost certainly happen if you run the code, but
it's not guaranteed. The while loop in the new thread could keep
running forever, never really checking whether or not the stop
variable has been set to true. If that sounds bizarre to you, welcome to the
weird and wonderful world of memory models.
Memory in modern computers is a very complicated business, with registers, multiple levels of cache, and multiple processors sharing main memory but possibly not caches, etc. The idea that there's just a single chunk of memory which is accessed in a simple way is very handy for programmers, but lousy for performance. In addition, if a processor knows it might have to read a bit of memory "soon", it could decide to read it early, etc. Hardware manufacturers and compiler writers (including JIT-compiler writers) have worked very hard to make fast code easy to write. The memory model of a platform is the specification of what developers can do safely without knowing too much about the details of the hardware the platform is running on. This means (in our case) that you can run .NET code on any CPU which has a CLR, and so long as you follow the rules of the memory model, you should be okay - however "strong" or "weak" the memory model of the hardware itself is. (A "strong" memory model is one which guarantees a lot; a "weak" model is one which doesn't guarantee much at all, often giving better performance but requiring more work on the part of the developer. "x86" processors have a stronger memory model than the CLR itself, which is one reason problems such as seeing stale data are relatively hard to demonstrate.)
The memory model in .NET talks about when reads and writes "actually" happen compared with when they occur in the program's instruction sequence. Reads and writes can be reordered in any way which doesn't violate the rules given by the memory model. As well as "normal" reads and writes there are volatile reads and writes. Every read which occurs after a volatile read in the instruction sequence occurs after the volatile read in the memory model too - they can't be reordered to before the volatile read. A volatile write goes the other way round - every write which occurs before a volatile write in the instruction sequence occurs before the volatile write in the memory model too.
Don't worry if the above doesn't make much sense - the
resource section at the end of this page contain a
few links which should help you out a bit if you want to really
understand it thoroughly, but the rule is pretty simple: when you have
access to shared data, you need to make sure you read fresh data and
write any changes back in a timely manner. There are two ways of doing
this - volatile variables, and using lock again.
A variable which is declared volatile uses volatile reads and writes
for all its accesses. You can only declare a variable to be volatile if
it's one of the following types: a reference type, byte,
sbyte, short, ushort,
int, uint, char, float, or
bool, or an enumeration with a base type of byte,
sbyte, short, ushort,
int, or uint. If you're only interested in sharing
a single piece of data, and it's one of the above types, then using a volatile
variable is probably the easiest way to go. Note, however, that for a reference
type, only the access to the variable itself is volatile - if you write to
something within the instance the reference refers to, that write won't be
volatile. Personally I don't use volatile variables much, preferring the other
approach: locking.
We've already seen how locking is used to limit access to a single thread at a
time. It also has another side effect: a call to Monitor.Enter performs
an implicit volatile read, and a call to Monitor.Exit performs
an implicit volatile write. The two effects combine nicely: if you're reading,
you perform a volatile read, so you know that your next read will be from
main memory - and because you're then in a lock, you know that nothing else will
be trying to change the value. Similarly, if you're writing, you know that nothing else
will be trying to read the value between you writing it and the volatile write,
so nothing will see an old value - assuming all access to the variable
is covered with the same lock, of course. If you lock using one monitor for some access
to a variable, and another monitor for other access to the same variable, the volatility
and the locking won't mesh quite as nicely, and you won't get as strong a guarantee of
freshness of data. Fortunately, there's very little reason why you'd even want to
try this.
So, to get back to our sample program: it's currently flawed because the new thread
could read the value of stop once (perhaps into a register) and then never
bother reading it from main memory. Alternatively, it could always read it from main memory,
but the original thread may never write it there. To fix it, we could either just make
stop volatile, or we could use a lock. The volatile solution is simple
- just add the keyword volatile to the variable declaration, and you're done.
The locking solution requires a bit more effort, and I'd make things slightly easier by
introducing a property to do the locking. So long as you then only refer to the variable
via the property, you don't need to write the lock all over the place. Here's the full code
with a property which locks:
using System.Threading; public class Test { static bool stop; static readonly object stopLock = new object(); static bool Stop { get { lock (stopLock) { return stop; } } set { lock (stopLock) { stop = value; } } } static void Main() { ThreadStart job = new ThreadStart(ThreadJob); Thread thread = new Thread(job); thread.Start(); // Let the thread start running Thread.Sleep(2000); // Now tell it to stop counting Stop = true; } static void ThreadJob() { int count = 0; while (!Stop) { Console.WriteLine ("Extra thread: count {0}", count); Thread.Sleep(100); count++; } } } |
Unfortunately there's no way of getting the compiler to complain if you
access stop directly, so you do need to be careful to always use the property.
As of .NET 1.1, there is another way of achieving a memory barrier: Thread.MemoryBarrier().
In future versions there may well be separate method calls for "write" memory barriers and "read" memory
barriers. I would advise steering well clear of these unless you're an expert - even the experts
seem to argue amongst themselves about what's needed when. (Read the links in the
resource section for more information.)
Just to reiterate: working things out to be as efficient as possible but still absolutely correct is hard. Fortunately, using locks whenever you want to access shared data is relatively easy and correct by the model. Stick to the simple way of doing things and you don't need to worry about all this too much.
This section is here almost as an aside - because if you're writing thread-safe code to start with, atomicity isn't particularly relevant to you. However, it's a good idea to clear up what atomicity is all about, because many people believe it's to do with volatility and the like.
An operation is atomic if it is indivisible - in other words, nothing else can happen in the middle. So, with an atomic write, you can't have another thread reading the value half way through the write, and ending up "seeing" half of the old value and half of the new value. Similarly, with an atomic read, you can't have another thread changing the value half way through the read, ending up (again) with a value which is neither the old value nor the new value.
The CLR guarantees that for types which are no bigger than the size of a native integer,
if the memory is properly aligned (as it is by default - if you specify an explicit layout,
that could change the alignment), reads and writes are atomic. In other words, if one thread
is changing a properly aligned int variable's value from 0 to 5 and another thread
is reading the variable's value, it will only ever see 0 or 5 - never 1 or 4, for instance.
For a long, however, on a 32-bit machine, if one thread is changing the value from 0 to
0x0123456789abcdef, there's no guarantee that another thread won't see the value as 0x0123456700000000
or 0x0000000089abcdef. You'd have to be unlucky - but writing thread-safe code is all about taking
luck out of the equation.
Fortunately, using the techniques I've already mentioned, you rarely need to worry about atomicity at all. Certainly if you use locking, you don't need to worry as you're already making sure that a read and a write can't overlap. If you use volatile variables there may be a slight chance of problems, as although every type which can be volatile can be atomically written and read, if the alignment of the variable is wrong, you could still get non-atomic reads and writes - the volatility doesn't provide any extra guarantees. Just another reason to use locking :)
Interlocked class
Just occasionally, locking is a bit too much effort (and possibly too much of
a performance hit) for doing very simple operations such as counting. The Interlocked
class provides a set of methods for performing atomic changes: exchanges (optionally performing a comparison
first), increments and decrements. The Exchange and CompareExchange methods act on
variables of type int, object or float; the Increment
and Decrement methods act on variables of type int or long.
Frankly I've never used the class myself in production code - I prefer to take the simple approach of
using one tool (locking) to sort out all my volatility, atomicity and race-avoidance problems. However,
that does come at the cost of a bit of performance. While that's never bothered me overly, if you're
writing code which needs to perform at its absolute fastest, you may want to consider using this class
as a fast way of performing the very specific operations it provides. Here's a sample - the first example
I used to illustrate data races, rewritten to be thread-safe using the Interlocked class:
using System; using System.Threading; public class Test { static int count=0; static void Main() { ThreadStart job = new ThreadStart(ThreadJob); Thread thread = new Thread(job); thread.Start(); for (int i=0; i < 5; i++) { Interlocked.Increment(ref count); } thread.Join(); Console.WriteLine ("Final count: {0}", count); } static void ThreadJob() { for (int i=0; i < 5; i++) { Interlocked.Increment(ref count); } } } |
One of the issues which frequently comes up in newsgroups is how to handle threading in a UI. There are two golden rules for Windows Forms:
Invoke, BeginInvoke, EndInvoke or CreateGraphics,
and InvokeRequired.
Text property) from a different
thread, you run a risk of your program hanging or misbehaving in other ways. You may get away with
it in some cases, but only by blind luck. Fortunately, the Invoke, BeginInvoke
and EndInvoke methods have been provided so that you can ask the UI thread to call a method
for you in a safe manner.
Application.DoEvents(), and this is
the natural thing for many VB programmers to wish to do - but I'd advise against it. It means you have to
consider re-entrancy issues etc, which I believe are harder to diagnose and fix than "normal" threading
problems. You have to judge when to call DoEvents, and you can't use anything which might
block (network access, for instance) without risking an unresponsive UI. I believe there are message
pumping issues in terms of COM objects as well, but I don't have details of them (and I frankly wouldn't
understand them fully anyway).
So, if you have a piece of long-running code which you need to execute, you need to create a new thread (or use a thread pool thread if you prefer) to execute it on, and make sure it doesn't directly try to update the UI with its results. The thread creation part is the same as any other threading problem, and we've addressed that before. The interesting bit is going the other way - invoking a method on the UI thread in order to update the UI.
There are two different ways of invoking a method on the UI thread, one synchronous (Invoke)
and one asynchronous (BeginInvoke). They work in much the same way - you specify a delegate
and (optionally) some arguments, and a message goes on the queue for the UI thread to process. If you use
Invoke, the current thread will block until the delegate has been executed. If you use
BeginInvoke, the call will return immediately. If you need to get the return value of a
delegate invoked asynchronously, you can use EndInvoke with the IAsyncResult
returned by BeginInvoke to wait until the delegate has completed and fetch the return value.
There are two options when working out how to get information between the various threads involved.
The first option is to have state in the class itself, setting it in one thread, retrieving and processing it
in the other (updating the display in the UI thread, for example). The second option is to pass the information
as parameters in the delegate. Using state somewhere is necessary if you're creating a new thread rather
than using the thread pool - but that doesn't mean you have to use state to return information to the UI.
On the other hand, creating a delegate with lots of parameters often feels clumsy, and is in some ways less
efficient than using a simple MethodInvoker or EventHandler delegate. These two
delegates are treated in a special (fast) manner by Invoke and BeginInvoke.
MethodInvoker is just a delegate which takes no parameters and returns no value
(like ThreadStart), and EventHandler takes two parameters (a sender and an
EventArgs parameter and returns no value. Note, however, that if you pass an EventHandler
delegate to Invoke or BeginInvoke then even if you specify parameters yourself,
they are ignored - when the method is invoked, the sender will be the control you have invoked it with,
and the EventArgs will be EventArgs.Empty.
Here is an example which shows several of the above concepts. Notes are provided after the code.
using System; using System.Threading; using System.Windows.Forms; using System.Drawing; public class Test : Form { delegate void StringParameterDelegate (string value); Label statusIndicator; Label counter; Button button; /// <summary> /// Lock around target and currentCount /// </summary> readonly object stateLock = new object(); int target; int currentCount; Random rng = new Random(); Test() { Size = new Size (180, 120); Text = "Test"; Label lbl = new Label(); lbl.Text = "Status:"; lbl.Size = new Size (50, 20); lbl.Location = new Point (10, 10); Controls.Add(lbl); lbl = new Label(); lbl.Text = "Count:"; lbl.Size = new Size (50, 20); lbl.Location = new Point (10, 34); Controls.Add(lbl); statusIndicator = new Label(); statusIndicator.Size = new Size (100, 20); statusIndicator.Location = new Point (70, 10); Controls.Add(statusIndicator); counter = new Label(); counter.Size = new Size (100, 20); counter.Location = new Point (70, 34); Controls.Add(counter); button = new Button(); button.Text = "Go"; button.Size = new Size (50, 20); button.Location = new Point (10, 58); Controls.Add(button); button.Click += new EventHandler (StartThread); } void StartThread (object sender, EventArgs e) { button.Enabled = false; lock (stateLock) { target = rng.Next(100); } Thread t = new Thread(new ThreadStart(ThreadJob)); t.IsBackground = true; t.Start(); } void ThreadJob() { MethodInvoker updateCounterDelegate = new MethodInvoker(UpdateCount); int localTarget; lock (stateLock) { localTarget = target; } UpdateStatus("Starting"); lock (stateLock) { currentCount = 0; } Invoke (updateCounterDelegate); // Pause before starting Thread.Sleep(500); UpdateStatus("Counting"); for (int i=0; i < localTarget; i++) { lock (stateLock) { currentCount = i; } // Synchronously show the counter Invoke (updateCounterDelegate); Thread.Sleep(100); } UpdateStatus("Finished"); Invoke (new MethodInvoker(EnableButton)); } void UpdateStatus(string value) { if (InvokeRequired) { // We're not in the UI thread, so we need to call BeginInvoke BeginInvoke(new StringParameterDelegate(UpdateStatus), new object[]{value}); return; } // Must be on the UI thread if we've got this far statusIndicator.Text = value; } void UpdateCount() { int tmpCount; lock (stateLock) { tmpCount = currentCount; } counter.Text = tmpCount.ToString(); } void EnableButton() { button.Enabled = true; } static void Main() { Application.Run (new Test()); } } |
Notes:
UpdateStatus, which uses InvokeRequired to
detect whether or not it needs to "change thread". If it does, it then
calls BeginInvoke to execute the same method again from
the UI thread. This is quite a common way of making a method which
interacts with the UI thread-safe. The choice of
BeginInvoke rather than Invoke here was just
to demonstrate how to invoke a method asynchronously. In real code, you
would decide based on whether you needed to block to wait for the
access to the UI to complete before continuing or not. In practice, I believe
it's quite rare to actually require UI access to complete first, so I tend
to use BeginInvoke instead of Invoke. Another approach
might be to have a property which did the appropriate invoking when
necessary. That's easier to use from the client code, but slightly
harder work in that you would either have to have another method
anyway, or get the MethodInfo for the property setter in
order to construct the delegate to invoke. In this case we actually know
that BeginInvoke is required because we're running in the
worker thread anyway, but I included the code for the sake of completeness.
EndInvoke after the BeginInvoke. Unlike all other
asynchronous methods (see the later section on the topic) you don't
need to call EndInvoke unless you need the return value of the delegate's method.
Of course, BeginInvoke is also different to all of the other asynchronous methods
as it doesn't cause the delegate to be run on a thread pool thread - that would defeat the
whole point in this case!
MethodInvoker delegate to execute
UpdateCount. We call this using Invoke to
make sure that it executes on the UI thread. This time there's no
attempt to detect whether or not an Invoke is required. I
don't believe there's much harm in calling Invoke or
BeginInvoke when it's not required - it'll just take a
little longer than calling the method directly. (If you call
BeginInvoke it will have a different effect than calling
the method directly as it will occur later, rather than in the current
execution flow, of course.) Again, we actually know that we need to
call Invoke here anyway.
MethodInvoker delegate is used to enable the button again
afterwards.
UpdateCount - the UI thread would then try to
acquire the lock as well, and you'd end up with deadlock.
IsBackground=true;) so that
when the UI thread exits, the whole application finishes. In other cases where you have a thread
which should keep running even after the UI thread has quit, you need to be careful not to call
Invoke or BeginInvoke when the UI thread is no longer running - you will
either block permanently (waiting for the message to be taken off the queue, with nothing actually
looking at messages) or receive an exception.
The point of the thread pool is to avoid creating lots of threads for short tasks. Thread creation isn't particularly cheap, and if you start lots of threads, each doing only just enough work to warrant being run on a different thread in the first place, the cost of creation could significantly hamper performance. The thread pool solves that by having a "pool" of threads which have already been created and are just waiting for work items. When they've finished executing a work item, they then wait for the next one, etc. By default, the thread pool has 25 threads per processor. Note that the thread pool isn't just used for whatever asynchronous calls you make - the .NET framework libraries use it as well, and things can go badly wrong (usually resulting in a deadlock) if all the threads are used and some of them depend on other tasks which are scheduled. (For instance, if one thread is waiting for the results of another work item, but that work item is never run because there are no free threads.) This is a good reason to avoid using the thread pool for particularly long-running tasks. Personally I usually stick to creating a new thread for anything but pretty trivial tasks. If the thread's going to be running for more than a few seconds, the cost of creating the thread is going to be relatively insignificant.
Note that none of the samples below have any locking or volatility in to ensure that "fresh" values are seen. I haven't seen any code in any other samples, either. It's not a problem so long as there's a memory barrier in both the calling thread and the thread pool thread, but I haven't seen any guarantees of that. I would expect that there would be a memory barrier involved in each thread just to get the whole thing up and running in the first place, but as I say I haven't seen it guaranteed anywhere.
You can tell whether or not a thread is from the thread pool or not using Thread.IsThreadPoolThread.
Thread pool threads are background threads - they don't prevent the runtime from exiting when all non-background
threads have completed. There are various different ways of using the thread pool. Here's a brief description
of each of them (except timers which have their own section, following this one):
This is probably the simplest way of executing code in a thread pool thread. You simply provide a
WaitCallback delegate (it doesn't return anything, and takes a single parameter of type object),
and optionally the object to pass as the parameter to the callback when it is executed. If you don't specify
a parameter, the callback will be passed null instead. Here's a sample program to demonstrate it:
using System; using System.Threading; public class Test { static void Main() { ThreadPool.QueueUserWorkItem(new WaitCallback(PrintOut), "Hello"); // Give the callback time to execute - otherwise the app // may terminate before it is called Thread.Sleep(1000); } static void PrintOut (object parameter) { Console.WriteLine(parameter); } } |
There is no built-in way of waiting for your callback to be executed, although you can of course
signal the end of the callback using "normal" threading mechanisms (Monitor.Wait/Pulse etc).
It's relatively hard to find documentation on this topic, but when you create a delegate, the compiler
generates three methods for you: Invoke, BeginInvoke and EndInvoke.
Invoke is used to execute the delegate synchronously (i.e. a line of code such as myDelegate();
is actually compiled as myDelegate.Invoke();). The other two methods are for asynchronous execution,
and must always be called as a pair - every BeginInvoke must be matched by a call to EndInvoke
somewhere to guarantee that you don't leak resources. BeginInvoke takes the same parameters as the
delegate itself does, plus another two parameters - an AsyncCallback which is called after the delegate
has executed, and an object parameter which is made available through the AsyncState property of the
IAsyncResult parameter which is passed to the AsyncCallback. (This is typically used to
pass the delegate which is being invoked, to make it easy to call EndInvoke.) The call to EndInvoke
can be made to find the return value of the executed delegate. Don't worry if it sounds confusing - hopefully this
example will make it somewhat simpler:
using System; using System.Threading; public class Test { delegate int TestDelegate(string parameter); static void Main() { TestDelegate d = new TestDelegate(PrintOut); d.BeginInvoke("Hello", new AsyncCallback(Callback), d); // Give the callback time to execute - otherwise the app // may terminate before it is called Thread.Sleep(1000); } static int PrintOut (string parameter) { Console.WriteLine(parameter); return 5; } static void Callback (IAsyncResult ar) { TestDelegate d = (TestDelegate)ar.AsyncState; Console.WriteLine ("Delegate returned {0}", d.EndInvoke(ar)); } } |
The call to BeginInvoke returns an IAsyncResult which can be used to call
EndInvoke, and you don't have to pass a callback delegate to be executed if you don't want
to - just pass null as the last but one parameter to BeginInvoke. (You may still
wish to pass in meaningful last parameter, as that will be available in the returned IAsyncResult's
AsyncState property, just as it would be in the callback case.)
The call to EndInvoke blocks until the delegate has finished executing - it's sort of like
Thread.Join, but for a specific asynchronous delegate execution rather than a specific thread. Of course,
when you call it from a callback delegate, it won't need to block as the callback will only execute after the
delegate has finished anyway. Here's an example using EndInvoke from the original thread instead
of using a callback:
using System; using System.Threading; public class Test { delegate int TestDelegate(string parameter); static void Main() { TestDelegate d = new TestDelegate(PrintOut); IAsyncResult ar = d.BeginInvoke("Hello", null, null); Console.WriteLine ("Main thread continuing to execute..."); int result = d.EndInvoke(ar); Console.WriteLine ("Delegate returned {0}", result); } static int PrintOut (string parameter) { Console.WriteLine(parameter); return 5; } } |
Sometimes having to call EndInvoke is inconvenient - you
often want "fire and forget" semantics where you don't care about the
result or indeed when exactly the delegate has finished executing. Many
articles suggest just calling BeginInvoke and not bothering
with EndInvoke. This may work without leaking
resources - but it's not guaranteed to, and even if it does now, it may
not in the future. Here is a utility class (adapter from a
mailing list post [http://discuss.develop.com/archives/wa.exe?A2=ind0302b&L=ADVANCED-DOTNET&D=0&P=2534]
which allows you to call FireAndForget to
execute a delegate asynchronously without worrying about EndInvoke:
using System; using System.Threading; public class ThreadUtil { /// <summary> /// Delegate to wrap another delegate and its arguments /// </summary> delegate void DelegateWrapper (Delegate d, object[] args); /// <summary> /// An instance of DelegateWrapper which calls InvokeWrappedDelegate, /// which in turn calls the DynamicInvoke method of the wrapped /// delegate. /// </summary> static DelegateWrapper wrapperInstance = new DelegateWrapper (InvokeWrappedDelegate); /// <summary> /// Callback used to call <code>EndInvoke</code> on the asynchronously /// invoked DelegateWrapper. /// </summary> static AsyncCallback callback = new AsyncCallback(EndWrapperInvoke); /// <summary> /// Executes the specified delegate with the specified arguments /// asynchronously on a thread pool thread. /// </summary> public static void FireAndForget (Delegate d, params object[] args) { // Invoke the wrapper asynchronously, which will then // execute the wrapped delegate synchronously (in the // thread pool thread) wrapperInstance.BeginInvoke(d, args, callback, null); } /// <summary> /// Invokes the wrapped delegate synchronously /// </summary> static void InvokeWrappedDelegate (Delegate d, object[] args) { d.DynamicInvoke(args); } /// <summary> /// Calls EndInvoke on the wrapper and Close on the resulting WaitHandle /// to prevent resource leaks. /// </summary> static void EndWrapperInvoke (IAsyncResult ar) { wrapperInstance.EndInvoke(ar); ar.AsyncWaitHandle.Close(); } } |
When you provide FireAndForget with a delegate to execute, it actually
invokes an internal delegate asynchronously, and that delegate executes the one you provided synchronously.
This gives the effective result of the delegate you provided begin executed asynchronously, but allows the
helper class to call EndInvoke on the delegate that it knows about - the Delegate
class itself doesn't provide BeginInvoke or EndInvoke methods, otherwise
this extra step would be unnecessary. Note the call to ar.AsyncWaitHandle.Close(). This prevents
the WaitHandle leaking until garbage collection. The leak wouldn't cause any problems in most
cases (unlike, for instance, file handles leaking), but in situations where FireAndForget
would be called many, many times in quick succession, you could end up with a vast number of handles until
the garbage collector started finalizing them. (This is also a bad thing in terms of performance - you
shouldn't leave things to be finalised when it can be avoided.) The earlier sample code omitted this step
for simplicity, but it's worth being aware of.
There are various methods in the standard library which come in
BeginXXX, EndXXX pairs, such as
Stream.BeginRead and Stream.EndRead. These
almost all follow the same format, which is quite like calling
BeginInvoke and EndInvoke on a delegate: you
call BeginXXX with some "normal" parameters, an
AsyncCallback parameter and a "state" parameter. The
callback is executed asynchronously when the operation (such as reading
from a stream) has completed. The callback can then use
EndXXX to get the results of the operation. There has been
some discussion on the newsgroups as to whether calls such as
Stream.BeginRead use I/O completion ports, which are a very
efficient way of performing I/O asynchronously without using one thread
per operation. It seems likely that they do, so even after the first
callback has started executing on a thread pool thread, if you need to do
more of the same kind of operation (as you frequently will with something
like a network stream, where you probably haven't read the whole thing in
one go) it's a good idea to keep using asynchronous calls rather than the
synchronous forms. Here's an example which downloads this page
asynchronously:
using System; using System.IO; using System.Net; using System.Text; using System.Threading; public class Test { static readonly object finishedLock = new object(); const string PageUrl = @"https://jonskeet.uk/csharp/threads/threadpool.html"; static void Main() { WebRequest request = WebRequest.Create(PageUrl); RequestResponseState state = new RequestResponseState(); state.request = request; // Lock the object we'll use for waiting now, to make // sure we don't (by some fluke) do everything in the other threads // before getting to Monitor.Wait in this one. If we did, the pulse // would effectively get lost! lock (finishedLock) { request.BeginGetResponse(new AsyncCallback(GetResponseCallback), state); Console.WriteLine ("Waiting for response..."); // Wait until everything's finished. Normally you'd want to // carry on doing stuff here, of course. Monitor.Wait(finishedLock); } } static void GetResponseCallback (IAsyncResult ar) { // Fetch our state information RequestResponseState state = (RequestResponseState) ar.AsyncState; // Fetch the response which has been generated state.response = state.request.EndGetResponse(ar); // Store the response stream in the state state.stream = state.response.GetResponseStream(); // Stash an Encoding for the text. I happen to know that // my web server returns text in ISO-8859-1 - which is // handy, as we don't need to worry about getting half // a character in one read and the other half in another. // (Use a Decoder if you want to cope with that.) state.encoding = Encoding.GetEncoding(28591); // Now start reading from it asynchronously state.stream.BeginRead(state.buffer, 0, state.buffer.Length, new AsyncCallback(ReadCallback), state); } static void ReadCallback (IAsyncResult ar) { // Fetch our state information RequestResponseState state = (RequestResponseState) ar.AsyncState; // Find out how much we've read int len = state.stream.EndRead(ar); // Have we finished now? if (len==0) { // Dispose of things we can get rid of ((IDisposable)state.response).Dispose(); ((IDisposable)state.stream).Dispose(); ReportFinished (state.text.ToString()); return; } // Nope - so decode the text and then call BeginRead again state.text.Append(state.encoding.GetString(state.buffer, 0, len)); state.stream.BeginRead(state.buffer, 0, state.buffer.Length, new AsyncCallback(ReadCallback), state); } static void ReportFinished (string page) { Console.WriteLine ("Read text of page. Length={0} characters.", page.Length); // Assume for convenience that the page length is over 50 characters! Console.WriteLine ("First 50 characters:"); Console.WriteLine (page.Substring(0, 50)); Console.WriteLine ("Last 50 characters:"); Console.WriteLine (page.Substring(page.Length-50)); // Tell the main thread we've finished. lock (finishedLock) { Monitor.Pulse(finishedLock); } } class RequestResponseState { // In production code, you may well want to make these properties, // particularly if it's not a private class as it is in this case. internal WebRequest request; internal WebResponse response; internal Stream stream; internal byte[] buffer = new byte[16384]; internal Encoding encoding; internal StringBuilder text = new StringBuilder(); } } |
Note that there is no exception handling code. This is purely to keep the code simple here - it doesn't
mean you don't need exception handling in your real code. Unfortunately it's harder to make sure that
you don't fail to dispose of unmanaged resources in error situations when using asynchronous methods:
the using statement is no use as it only works within one thread - if you did
put using statements around the BeginRead (etc) method calls, you'd end up
disposing of the stream before it had time to finish reading, which would be disastrous.
Of course, you could avoid asynchronous delegates entirely, not specifying any, and just call
EndRead (etc) immediately after BeginRead in the main thread, just as with
BeginInvoke and EndInvoke on delegates. Just like with delegates, you must
call the matching EndXXX method to avoid potential resource leaks. It's less likely that you'll
fail to do so in this case, of course, as the methods tend to return useful information. You should consider
exceptions when devising your strategy for making these calls, however.
There are various different timers available in .NET, each of which basically calls a delegate after a certain
amount of time has passed. This section is a very brief guide to the differences between them.
See the resources section for more in depth links. All the timers implement
IDisposable, so make sure you dispose of them when you're not using them any more.
This is (obviously) a forms-based timer. After creating it, you can set the interval which elapses between ticks
(using the Interval property) and hook up a delegate to its Tick event. Calling
Start and Stop methods (which effectively just change the value of the
Enabled property) do the obvious things. While the timer is running, it generates ticks and fires
the Tick event each time. Note, however, that because it runs entirely on the UI thread,
if you have long-running UI operations, you may "miss" ticks if more than one would normally occur in the time
taken by the long-running operation. Effectively, it keeps track of when its next tick is due, and when it next
gets a chance to run, if the time is up, it fires the event. Personally, I find this the least useful of the timer
classes.
This is a somewhat more powerful timer. Instead of a Tick event, it has the Elapsed
event. As before, there are Start and Stop methods which are similar to changing
the Enabled property. Changing the
Interval or Enabled properties effectively reset the timer. (In other words, if the interval
were set to two seconds, and every second you disabled and then re-enabled the timer, the event would never be fired.)
The AutoReset property determines whether or not the timer should fire the event once and then stop, or
keep going in a fire event / wait cycle.
By default, the event is fire on a thread pool thread. However, if you wish it to be fired on a particular thread,
you can use the SynchronizingObject property which makes it invoke the event however the synchronizing
object wishes it to. For instance, setting the synchronizing object to a UI control makes the event fire on that
control's UI thread. Unlike the previous timer, the events are effectively queued - the timer doesn't wait for one
event to have completed before starting to wait again and then firing off the next event.
This is the timer class I usually prefer, due to its simplicity. When constructing it, you need to pass in a
TimerCallback delegate, a state object which is passed to the delegate when the timer fires, a
"due" time and an interval. The timer will first fire after the due time has elapsed, and thereafter it will fire
after each interval. Either value may be Timeout.Infinite - if the due time is infinite, the timer
will never fire; if the interval is infinite, the timer will fire once (after the due time) and then it won't fire again.
You can change the due time and the interval at any point using the Change method. (For instance, I
sometimes find it useful to leave the interval as infinite, but every time the timer fires, call Change
with a new due time.)
There's nothing fancy about this timer: the timer always fires on a thread pool thread. If you need to use it to update the UI, you need to use the techniques talked about in the Windows Forms section. You don't start and stop it - if you don't want it to fire, just change the due time to be infinite.
Timers are simple enough that for the most part they don't really need examples, in my view.
Just for kicks, here's a sample of System.Threading.Timer. If you really need other examples,
please mail me (skeet@pobox.com) and let me know what exactly you're after. Note that the
article in the resources section has some sample code too.
using System; using System.Threading; public class Test { static void Main() { Console.WriteLine ("Started at {0:HH:mm:ss.fff}", DateTime.Now); // Start in three seconds, then fire every one second using (Timer timer = new Timer(new TimerCallback(Tick), null, 3000, 1000)) { // Wait for 10 seconds Thread.Sleep(10000); // Then go slow for another 10 seconds timer.Change (0, 2000); Thread.Sleep(10000); } // The timer will now have been disposed automatically due to the using // statement, so there won't be any other threads running, and we'll quit. } static void Tick(object state) { Console.WriteLine ("Ticked at {0:HH:mm:ss.fff}", DateTime.Now); } } |
And here are the results of one run of that test:
Started at 15:32:07.473 Ticked at 15:32:10.520 Ticked at 15:32:11.520 Ticked at 15:32:12.520 Ticked at 15:32:13.520 Ticked at 15:32:14.520 Ticked at 15:32:15.520 Ticked at 15:32:16.520 Ticked at 15:32:17.520 Ticked at 15:32:17.536 Ticked at 15:32:19.552 Ticked at 15:32:21.552 Ticked at 15:32:23.552 Ticked at 15:32:25.552 Ticked at 15:32:27.552 |
Note the very small gap between the ticks at 15:32:17 - this was where Change was called with a due time of zero,
which means "fire the delegate now".
This topic was mostly covered in the volatile section earlier, but as this is a particularly common scenario (with a typical question being "How can I shut down a worker thread?" I think it's worth presenting a working pattern. Here's a code skeleton which just needs the work for the worker thread to perform to be filled in (and any member it needs, of course):
using System; using System.Threading; /// <summary> /// Skeleton for a worker thread. Another thread would typically set up /// an instance with some work to do, and invoke the Run method (eg with /// new Thread(new ThreadStart(job.Run)).Start()) /// </summary> public class Worker { /// <summary> /// Lock covering stopping and stopped /// </summary> readonly object stopLock = new object(); /// <summary> /// Whether or not the worker thread has been asked to stop /// </summary> bool stopping = false; /// <summary> /// Whether or not the worker thread has stopped /// </summary> bool stopped = false; /// <summary> /// Returns whether the worker thread has been asked to stop. /// This continues to return true even after the thread has stopped. /// </summary> public bool Stopping { get { lock (stopLock) { return stopping; } } } /// <summary> /// Returns whether the worker thread has stopped. /// </summary> public bool Stopped { get { lock (stopLock) { return stopped; } } } /// <summary> /// Tells the worker thread to stop, typically after completing its /// current work item. (The thread is *not* guaranteed to have stopped /// by the time this method returns.) /// </summary> public void Stop() { lock (stopLock) { stopping = true; } } /// <summary> /// Called by the worker thread to indicate when it has stopped. /// </summary> void SetStopped() { lock (stopLock) { stopped = true; } } /// <summary> /// Main work loop of the class. /// </summary> public void Run() { try { while (!Stopping) { // Insert work here. Make sure it doesn't tight loop! // (If work is arriving periodically, use a queue and Monitor.Wait, // changing the Stop method to pulse the monitor as well as setting // stopping.) // Note that you may also wish to break out *within* the loop // if work items can take a very long time but have points at which // it makes sense to check whether or not you've been asked to stop. // Do this with just: // if (Stopping) // { // return; // } // The finally block will make sure that the stopped flag is set. } } finally { SetStopped(); } } } |
Note the second part of the comment in the Run method - often you will want to set up a
producer/consumer queue (as with the code given in the Monitor methods section,
and that works fine with the pattern above so long as you make sure that when another thread tells
the worker thread to stop, it pulses the monitor on the queue to make sure that the worker thread
wakes up. (You can do this either by adding a "no-op" work item, or by modifying the
the class implementing the queue to add a mechanism just tell worker threads to wake up and return
a null work item.
In .NET v2, where properties can have different access for setters and getters, I'd recommend turning
the SetStopped method into a setter for the Stopped property. I wouldn't
recommend changing the Stop method into a setter, however. This is partly because it needs
to be public, but should only go from false to true, and partly as a gut
instinct in terms of the word stop being more forceful as a noun than just setting a property to true
- it feels to me like properties shouldn't usually have as much of an effect on other threads as this
one does.
The above code is fine for occasional use, but including it in several classes introduces a fair amount of redundancy. It's not very much work to abstract most of the above into a separate class and provide more functionality at the same time. The resulting class is a bit too long to include in an article, but can be found as part of my Miscellaneous Utilities [https://jonskeet.uk/csharp/miscutil] library.
I said earlier that I'd explain why I created a new variable to lock
on. Many books and articles recommend locking on this for
instance methods and typeof(MyTypeName) (the
Type object for whatever type you're writing code in) for
static methods. I believe this is a bad idea, because it means you have
less control over your locks. Other code may well end up locking on the
same object as you do within your code, which makes it far harder to
ensure that you only obtain locks in an appropriate order.
An alternative to locking on this is to lock on the
reference of the object you're about to access - and indeed
ICollection provides a SyncRoot property for
exactly this purpose, providing a "shared" reference for different callers to
lock on. This is a valid design in some situations, but should generally be avoided
- if you keep your locks as private as possible, you should be able to write
thread-safe objects when you wish to. The difficulty I believe ICollection
is trying to solve is providing a single class which is fast in a single-threaded
situation due to not locking internally but which allows easy thread-safe access when
in a multi-threading situation. It also enables a common reference to be used for locking
across several method calls, such as enumerating the whole collection.
(Obviously ICollection itself doesn't provide
any code for any of this as it's just an interface, but it encourages a consistent
design for implementing classes to follow.)
In my experience, this level of complexity is unusual when developing applications
- usually the ability to make each method thread-safe in itself is all that's required,
and that can be achieved through entirely "private" locks, which no other object knows
about. Even if you never expose your fields directly, just accessing them for callers,
you don't usually know whether methods within those objects have decided to lock on this,
so for absolute control you should create a new variable, and immediately instantiate a new object.
This is the object you end up locking on. For "static locks" you should declare a private static variable
(and immediately instantiate a new object) for the same reason - other classes can get a Type
reference to your type too! While it is unlikely that they would lock on your type, it's not impossible -
a private static variable keeps the lock invisible to other classes. In both cases, make the variable
read-only to stop yourself from accidentally changing the value.
One other practice occasionally seen is that of locking on string literals, for example declaring
string someLock = "Lock guarding some data"; and then locking on someLock. This
has the same problem as the other locking strategies, due to string interning - if two classes both
use the same string literal, they'll end up using references to the same string, so they'll be sharing
a lock without realising it. If you really want to make a lock with a description, you can always
create your own Lock class. (This would also have the benefit of making it obvious to
the rest of your code what the appropriate variable is used for - Lock is more
descriptive than object in itself.)
In any one class you may have several locks, dealing with different bits of data which need to remain consistent. The more locks you have, the more finely grained your locking is - but the more complicated it gets making sure that you always take them out in a consistent order. (You also end up using more memory per instance, of course, but that's usually not an issue.) Of course, one of the benefits about having variables just for locking is that they can get their own XML documentation, making it easier to understand what pieces of state each lock is used to cover.
One final note on the issue: not only do many books and articles recommend locking on this:
the C# compiler does it for you automatically if you declare an event without specifying the add/remove
code. My recommendation is to explicitly write the add/remove code, following a pattern something like this:
/// <summary> /// Delegate backing the SomeEvent event. /// </summary> SomeEventHandler someEvent; /// <summary> /// Lock for SomeEvent delegate access. /// </summary> readonly object someEventLock = new object(); /// <summary> /// Description for the event /// </summary> public event SomeEventHandler SomeEvent { add { lock (someEventLock) { someEvent += value; } } remove { lock (someEventLock) { someEvent -= value; } } } /// <summary> /// Raises the SomeEvent event /// </summary> protected virtual OnSomeEvent(EventArgs e) { SomeEventHandler handler; lock (someEventLock) { handler = someEvent; } if (handler != null) { handler (this, e); } } |
(This assumes that the SomeEventHandler delegate has been declared elsewhere.)
Most of the time you can use a single lock for all your events, in my experience. Note the code for
OnSomeEvent. It's important that you don't write it as:
// Bad code! Do not use! protected virtual OnSomeEvent(EventArgs e) { lock (someEventLock) { if (someEvent != null) { someEvent(this, e); } } } |
or:
// Bad code! Do not use! protected virtual OnSomeEvent(EventArgs e) { if (someEvent != null) { someEvent (this, e); } } |
The first ends up firing the event while still holding the lock, which is a bad idea - you don't know what code is going to be run at this stage, and it could well end up wanting to access the event from another thread, leading to deadlock.
The second doesn't lock at all, making it quite possible that the event delegate will change between the test
for nullity and the invocation. That wouldn't be a problem most of the time, but if the delegate variable's value
becomes null, then trying to invoke the delegate will lead to an exception being thrown. (If we
didn't care whether the delegate variable's value was null or not, we wouldn't test it in the first place.)
Credit where credit is due - some of the ideas here are also embodied in
Jeffrey Richter's
Safe Thread Synchronization [http://msdn.microsoft.com/msdnmag/issues/03/01/NET/]
article, and others are in Ian Griffiths'
blog entry on locking with timeouts [http://www.interact-sw.co.uk/iangblog/2004/03/23/locking].
The C# team have also said that were they designing C# again now, they wouldn't have included the lock
keyword, instead making sure that a mechanism for using the using statement for the same
job would be available. (My main problem with the lock keyword is that it's the name I would
almost always naturally use for the variable containing the monitor in a single-lock class.)
Both Java and .NET made the same mistake when it came to locking, I believe. At first it seems like having a monitor for every object is a good idea, but in practice if you care about monitor privacy, you need to have an extra field in classes which need to worry about thread safety anyway. This is basically the point Jeffrey Richter makes in the article referenced above, and the same thought had been going through my head before I saw his elegant explanation of it.
I then also heard about Ian Griffiths' quest for attempts to acquire monitors with a timeout,
in a way which didn't make code really hard to follow. Using Monitor.TryEnter
will work, but then you've got to put the finally block in yourself, and test
the return value from the method. Ian reasoned that usually what you really want is just an
exception if you can't acquire a lock within a reasonable time limit - it indicates that
you've got deadlock, basically.
What intrigued me more than the idea was the implementation - specifically, using the using
statement to neatly acquire the lock and release it at the end of a block. This page (and the referenced code)
attempts to combine the two ideas, to give an alternative to normal locking which is clearer in intention
than using a plain object, which still maintains the neatness of code, and which allows locking with timeouts.
Over time, I've expanded the initial concept to allow deadlock detection in terms of locks having an order imposed on them, and made a few other tweaks. The code is available as part of my Miscellaneous Utility library [https://jonskeet.uk/csharp/miscutil], and the usage page [https://jonskeet.uk/csharp/miscutil/usage/locking.html] has more details of exactly what's available. This page now just covers the basic principles.
Essentially, the three primary types involved are:
SyncLockLockTimeoutExceptionTimeoutException, but that seems to be specific to Windows Services, annoyingly enough
- to my mind, it should have been in the System namespace.)
LockTokenSyncLock when the monitor has been successfully locked
- it must be disposed in order to release the monitor. The reason for making this a struct
rather than a class is that it then only takes up a bit of stack space for each lock operation,
rather than needing to create a whole extra object each time, which would be nasty for performance.
The typical usage is very straightforward:
// Create the lock - almost always as a field // This constructor overload creates a named // lock with a 20 second default timeout. Either the // name or the default timeout can be omitted. SyncLock syncLock = new SyncLock ("Some name"); // Lock it using its default timeout using (syncLock.Lock()) { // Do stuff here } // ... or lock it with a specific timeout using (syncLock.Lock(10000)) { } |
Please mail me (skeet@pobox.com) with any comments or criticisms.
I've written a lot now about how to use threads, but not when to use threads - when it's appropriate to and when it's better to keep everything in the same thread. This is partly because after a while it becomes fairly natural to work out what belongs where, and partly because it's quite tricky to actually describe.
There are a few times when there's absolutely no point in using multiple threads. For instance, if your application is bound by a single resource (e.g. the disk, or the CPU) and all the tasks you would use multiple threads for will all be trying to use that same resource, you'll just be adding contention. For instance, suppose you had an application which collected all the names of files on your disk. Splitting that job into multiple threads isn't likely to help - if anything, it'll make it worse, because it would be asking the file system for lots of different directories all at the same time, which could make the head seek all over the place instead of progressing steadily. Similarly, if you have workflow where each stage relies entirely on the results of the previous stage, you can't use threads effectively. For instance, if you have a program which loads an image, rotates it, then scales it, then turns it into black and white, then saves it again, each stage really needs the previous one to be finished before it can do anything.
Suppose, however, you wanted to read a bunch of files and then process their contents (e.g. calculating the cryptographic hash using an algorithm which takes a lot of processor power) then that might very well benefit from threading - either by having several threads doing both, or one thread dedicated to disk IO and another dedicated to hash calculation. The latter would probably be better, but would probably involve more work as the threads would need to be passing data to each other rather than just doing their own thing. Even though here there is still a dependency between the data being read from disk and the crypto processing, you don't have to read all the data from the disk before you can start processing it.
Similarly, in applications other than Windows Forms applications, it probably doesn't matter if you need to do something which will take a little while - a second or so, for instance. If you're writing a batch program, it doesn't matter in the slightest whether your code is always doing something different or whether some operations take longer than others. That's not to say that batch processing should always be single-threaded, but you don't need to worry about having a "main" thread which must be ready to react to events.
If you're writing an ASP.NET page, then unless you think the operation will take so long that you're willing to start another thread and send a page back to the user which tells them to wait and then refreshes itself (using an HTTP meta refresh tag, for instance) periodically to check whether the "real" page has finished, it's usually not worth changing to a different thread. It won't make the page come up any faster, as you'll just have to tell the "main" thread to wait for the other one to finish, and it'll be considerably more complicated to implement. Of course, the part about not being any faster isn't true if you genuinely can do two independent things at once - querying two different databases, for instance. In that case, threading can occasionally be very useful even in ASP.NET scenarios, although you should consider using the thread pool for such tasks to avoid creating lots of threads which each only run for a short time. Don't forget that there may well be other requests which want to use the same resources, and you won't be doing less work in total by spreading it out over many threads. In short, while creating extra threads (or explicitly using the thread pool) is sometimes useful in ASP.NET, it's usually a last resort rather than a matter of course.
In Windows Forms applications, you really should put anything which takes any significant amount of time (even reading a short file) into a different thread, unless the code is really only for your own use and you don't mind an unresponsive UI. This is important, as while the UI thread is doing something else (reading a file, doing a heavy calculation, etc) it can't be reacting to events like the user trying to close it, or a previously hidden area now becoming exposed. The UI can easily become very unresponsive, which gives a horrible user experience. Here, threading isn't used to get the job done quickly - it's used to get the job done while keeping the user satisfied with responsiveness. You might be surprised just how quickly a user can notice an app becoming unresponsive. Even if the user can't actually do anything but close the application or move the window around while they wait, it gives a much more professional feel to a program if you don't end up with a big white box when you pass another window over it.
The choice between using a "new" thread and using one from the thread pool is a contentious one - I tend to like using new threads for most purposes, and others recommend almost always using thread pool threads. My concern about using the thread pool is that there are only a certain number of threads in it, and it's used in various ways by the framework itself - and it's not always obvious that it's doing so. If you accidentally end up with all the threads in the thread pool waiting for other work items which are scheduled to run in the thread pool, you'll have a very difficult deadlock to debug.
Using new threads, however, is relatively costly if they're only going to run for a short while - creating new threads isn't terribly cheap at the operating system level, whereas the thread pool will of course re-use threads to avoid repeating this cost. One happy medium is to use your own thread pool which is separate from the system one but which will still re-use threads. I have a fairly simple implementation in my Miscellaneous Utility Library [https://jonskeet.uk/csharp/miscutil] which you can tweak if it doesn't quite meet your requirements. It probably won't perform quite as well as the system thread pool which has been finely tuned - but you have a lot more control over what goes in it, how many threads it creates, etc.
Using multiple threads is almost always going to introduce complexity to your application, so the potential benefits should be carefully considered and weighed up before you start writing code. Work out the "boundaries" between threads in detail - what thread needs access to what data when, etc. This can make an enormous difference when it comes to the implementation. After making the decision to use threads and designing the threading scenarios carefully, keep taking care while writing the code - it's very easy to slip up, unfortunately, even with all the tools the .NET framework provides. The results of this work should be an elegant application which performs well and remains responsive whatever it's doing - something to feel justly proud of. Good luck!
While you only occasionally need to make a decision about whether or not to use multiple threads, every time you design a type you need to consider what its thread safety model is going to be. This will very much depend on its intended usage. Some types naturally end up being used only within a single thread, others naturally end up being shared across threads, and some are used specifically for thread handling, and often have their own special characteristics. One of the most important things to do is to document the type's behaviour: which methods are thread-safe, which aren't, etc. Below are some common models. Note that when I talk about methods, I'm including properties as well.
This is the most commonly found model in the framework. Basically, any static methods can be called by any thread at any time with no nasty side-effects. (There may be side-effects, but they shouldn't cause anything to get into an invalid state.) Instance methods, however, should only be called on a single thread at a time, and there should really be a memory barrier (in the calling code) before using an instance that has previous been used by another thread, and another memory barrier (again, in the calling code) after changing an instance that is thereafter going to be used by another thread. In practice, this is probably one of the areas most people are sloppiest in, just passing objects from one thread to another without making sure things are synchronized specifically - and in practice, it's usually going to be fine, because part of "handing over" an object usually involves a memory barrier in both threads anyway. None of that is the responsibility of the type in question with this model, however - instance methods should basically assuming they're running in a single thread, and not worry about locking in order to query/update values. Of course, if you know that one of the members of the type may be shared amongst threads, using that member may involve locking, depending on the thread-safety of the type of that member.
If a type has a thread affinity, that usually means you can only use it
(or most of it) from the thread it was created on. Controls are the most
obvious examples of this - in the Windows Forms
section I've already talked about the way that you're not meant to use
any methods other than Invoke, BeginInvoke,
EndInvoke, InvokeRequired and CreateGraphics
on a control unless you're running in the thread responsible for it.
Designing your own type like this should be a relatively rare occurrence,
unless it's to do with the GUI, or uses something else with thread affinity.
Some methods and properties naturally use thread-local storage - that is
to say, each thread has its own individual value for a variable.
Thread.CurrentPrincipal is a good example of this - compare
it with Thread.CurrentCulture, which is an instance property.
Most methods and properties using thread-local storage are static, just like
Thread.CurrentPrincipal - having one value per instance and
per thread rarely makes sense, and indeed the simplest way of using
thread-local storage is with ThreadStaticAttribute which can
only be applied to static fields (which can then be returned by static
properties, etc).
Sometimes, it's important for a type to be totally thread-safe, so that anyone
can use any instance of it from any thread, with no internal inconsistencies or
unwanted side-effects. This is particularly true of types which are typically
accessed using the factory or singleton pattern. Encoding is a good
example of this - it would be a pain if threads couldn't use Encoding.ASCII
(or any other encoding) without carefully locking things.
Often, types like this are immutable - once they've been created, they don't change, so even without locking, it doesn't matter what you do with them. Assuming instances are acquired in a thread-safe way to start with (which usually involves a lock or a type initializer to ensure safety), immutable types are naturally thread-safe.
Some types can be accessed in some ways by multiple threads, but not in others.
Many collections fall into this category. For instance, from the documentation for
ArrayList:
| An ArrayList can support multiple readers concurrently, as long as the collection is not modified. To guarantee the thread safety of the ArrayList, all operations must be done through the wrapper returned by the Synchronized method. |
So, you can populate an instance in one thread, and thereafter read the contents from multiple threads. The documentation doesn't specify anything about handing over an instance from one thread to another (such as whether the client code has to make sure there's a memory barrier after the last write and another before the first read in another thread), which is unfortunate, but probably means you can't rely on it. This kind of case is where clear documentation is absolutely crucial. If a type can support multiple threads doing one thing and another thread doing something else (multiple reader, single writer tending to be the most common form) then the documentation should state exactly what is entailed - is that level of safety available without any extra locking on the part of the client, what are the guarantees about how soon that any "new" data is seen by other threads, etc.
There are two methods in the Thread class which are often
used for stopping threads - Abort and Interrupt.
I don't recommend using either of these methods, but it's worth knowing
about them - if only for why they should almost always be avoided.
Calling Thread.Abort aborts that thread as soon as possible.
(Aborting a thread which is executing unmanaged code has no effect until the CLR
gets control again.) A ThreadAbortException is thrown, which is
a special exception which can be caught, but will automatically be rethrown at
the end of the catch block. As it keeps being thrown, the exception
will usually terminate the thread. Thread.ResetAbort can be called
(if the caller has appropriate permissions) to stop the thread's abortion. Calling
the method doesn't prevent stop the currently thrown exception, it just stops
the exception from being rethrown at the end of a catch block. (Usually
this distinction is irrelevant, as you'd almost always want to call ResetAbort
from a catch block anyway).
Calling Thread.Interrupt is similar, but somewhat less drastic. This
causes a ThreadInterruptedException exception to be thrown the next
time the thread enters the WaitSleepJoin state, or immediately
if the thread is already in that state. There's nothing particularly special about
the ThreadInterruptedException - it doesn't get rethrown like
ThreadAbortException does. Note that threads can block without entering
the WaitSleepJoin state, however. For example, reading from a blocking
stream (a common situation where you'd like to interrupt a thread)
doesn't make the thread enter the WaitSleepJoin state.
A post on a newsgroup drew my attention to a bug in version 1.0/1.1 of the framework.
If a thread is aborted or interrupted while it is calling Monitor.Wait,
and after the monitor has been pulsed but before the thread has been able to
acquire it again, it returns from the call (with the appropriate exception) without
reacquiring the monitor. This can lead to situations where the code makes it look
like you'll certainly own the monitor, but you're executing it after an abort or
interrupt and you no longer own it. Here's an example:
using System; using System.Threading; class Test { static object someLock = new object(); static void Main() { Console.WriteLine ("Main thread starting"); Thread secondThread = new Thread (new ThreadStart(ThreadJob)); secondThread.Start(); Console.WriteLine ("Main thread sleeping"); Thread.Sleep(500); lock (someLock) { Console.WriteLine ("Main thread acquired lock - pulsing monitor"); Monitor.Pulse(someLock); Console.WriteLine ("Monitor pulsed; interrupting second thread"); secondThread.Interrupt(); Thread.Sleep(1000); Console.WriteLine ("Main thread still owns lock..."); } } static void ThreadJob() { Console.WriteLine ("Second thread starting"); lock (someLock) { Console.WriteLine ("Second thread acquired lock - about to wait"); try { Monitor.Wait(someLock); } catch (Exception e) { Console.WriteLine ("Second thread caught an exception: {0}", e.Message); } } } } |
The results of the above are:
Main thread starting Main thread sleeping Second thread starting Second thread acquired lock - about to wait Main thread acquired lock - pulsing monitor Monitor pulsed; interrupting second thread Second thread caught an exception: Thread has been interrupted from a waiting state. Main thread still owns lock... |
Note the order of the last two lines - the line from the second thread has
been written while the main thread owns the lock, despite being within
the second thread's lock block.
In fact, the above code should throw a SynchronizationLockException
when it implicitly calls Monitor.Exit at the end of the lock
block. As it happens, Monitor.Exit doesn't throw the exception (despite
the documentation's claims to the contrary).
It's hard to know exactly what should really happen here - if you've told
a thread to be interrupted or aborted, you probably don't want it to hang around
for a long time trying to reacquire a monitor (which is exactly what it does if
the monitor hasn't been pulsed before the thread is interrupted). On the other hand,
not being able to rely on a lock actually being owned within a lock block
is nasty. I believe the behaviour has been changed for version 2 of the framework,
but we'll have to wait to see exactly how it's changed.
Thread.Abort/Interrupt should be avoided
I don't use Thread.Abort/Interrupt routinely. I prefer a
graceful shutdown which lets the thread do
anything it wants to, and keeps things orderly. I dislike aborting or interrupting
threads for the following reasons:
WaitSleepJoin
state.
finally blocks will be executed, you don't want to have to put
them all over the place just in case of an abort or interrupt. In almost all
cases, the only time you don't mind a thread dying at any point in its operation
is when the whole application is going down.
This is really just a collection of the various important parts of the rest of the article. An executive summary, if you like.
Control.Invoke
within a lock that the UI thread will require.
Monitor.Wait when have only acquired the lock you're
waiting on. If you absolutely need to have another lock acquired at the same time, make sure
that lock isn't required by the code which will pulse the monitor.
Invoke, BeginInvoke,
EndInvoke, CreateGraphics or InvokeRequired) other than on its UI
thread.
Control.BeginInvoke, asynchronous BeginXXX method calls
should always make sure there is a matching EndXXX call.
EventHandler delegate to Control.Invoke
(or Control.BeginInvoke), any parameters you specify are ignored - the sender is always set to
the control it is executed for, and the EventArgs is always set to EventArgs.Empty.
Control.*Invoke [http://weblogs.asp.net/justin_rogers/articles/126345.aspx]EndXXX methods except in the case
of Control.BeginInvoke.