NB: Skip to the bottom of this answer for an update.
Here’s what’s worked for me so far:
StackTrace GetStackTrace (Thread targetThread)
{
StackTrace stackTrace = null;
var ready = new ManualResetEventSlim();
new Thread (() =>
{
// Backstop to release thread in case of deadlock:
ready.Set();
Thread.Sleep (200);
try { targetThread.Resume(); } catch { }
}).Start();
ready.Wait();
targetThread.Suspend();
try { stackTrace = new StackTrace (targetThread, true); }
catch { /* Deadlock */ }
finally
{
try { targetThread.Resume(); }
catch { stackTrace = null; /* Deadlock */ }
}
return stackTrace;
}
If it deadlocks, the deadlock is automatically freed and you get back a null trace. (You can then call it again.)
I should add that after a few days of testing, I’ve only once been able to create a deadlock on my Core i7 machine. Deadlocks are common, though, on single-core VM when the CPU runs at 100%.
Update: This approach works only for .NET Framework. In .NET Core and .NET 5+, Suspend and Resume cannot be called, so you must use an alternative approach such as Microsoft’s ClrMD library. Add a NuGet reference to the Microsoft.Diagnostics.Runtime package; then you can call DataTarget.AttachToProcess to obtain information about threads and stacks. Note that you cannot sample your own process, so you must start another process, but that is not difficult. Here is a basic Console demo that illustrates the process, using a redirected stdout to send the stack traces back to the host:
using Microsoft.Diagnostics.Runtime;
using System.Diagnostics;
using System.Reflection;
if (args.Length == 3 &&
int.TryParse (args [0], out int pid) &&
int.TryParse (args [1], out int threadID) &&
int.TryParse (args [2], out int sampleInterval))
{
// We're being called from the Process.Start call below.
ThreadSampler.Start (pid, threadID, sampleInterval);
}
else
{
// Start ThreadSampler in another process, with 100ms sampling interval
var startInfo = new ProcessStartInfo (
Path.ChangeExtension (Assembly.GetExecutingAssembly().Location, ".exe"),
Process.GetCurrentProcess().Id + " " + Thread.CurrentThread.ManagedThreadId + " 100")
{
RedirectStandardOutput = true,
CreateNoWindow = true
};
var proc = Process.Start (startInfo);
proc.OutputDataReceived += (sender, args) =>
Console.WriteLine (args.Data != "" ? " " + args.Data : "New stack trace:");
proc.BeginOutputReadLine();
// Do some work to test the stack trace sampling
Demo.DemoStackTrace();
// Kill the worker process when we're done.
proc.Kill();
}
class Demo
{
public static void DemoStackTrace()
{
for (int i = 0; i < 10; i++)
{
Method1();
Method2();
Method3();
}
}
static void Method1()
{
Foo();
}
static void Method2()
{
Foo();
}
static void Method3()
{
Foo();
}
static void Foo() => Thread.Sleep (100);
}
static class ThreadSampler
{
public static void Start (int pid, int threadID, int sampleInterval)
{
DataTarget target = DataTarget.AttachToProcess (pid, false);
ClrRuntime runtime = target.ClrVersions [0].CreateRuntime();
while (true)
{
// Flush cached data, otherwise we'll get old execution info.
runtime.FlushCachedData();
foreach (ClrThread thread in runtime.Threads)
if (thread.ManagedThreadId == threadID)
{
Console.WriteLine(); // Signal new stack trace
foreach (var frame in thread.EnumerateStackTrace().Take (100))
if (frame.Kind == ClrStackFrameKind.ManagedMethod)
Console.WriteLine (" " + frame.ToString());
break;
}
Thread.Sleep (sampleInterval);
}
}
}
This is the mechanism that LINQPad 6+ uses to show live execution tracking in queries (with additional checks, metadata probing and a more elaborate IPC).