[API Proposal]: Assembly.SetEntryAssembly()

ivdiazsa commented 2 weeks ago

Background and motivation

There exist certain specialty applications that act as a launcher for other applications within the same process. These applications typically wait for some event to occur (file system change, database change, time elapsed, etc), load a target application, and then execute it.

This works almost perfectly within .NET where the launcher can do something like spawn a thread and quickly get out of the way. However, this remains a tad complicated in the initial bookkeeping of the BCL / runtime that applications rely on. If, as applications sometimes do, they want to invoke Assembly.GetEntryAssembly, the result they get would obviously be wrong, as it would point to the launcher instead of the application itself. So, this requires changes in order to operate correctly within a launcher.

In order for customers to have that bookkeeping consistency and be able to change the assembly that actually will be run in these scenarios, we would like to propose allowing them to update what the runtime views as the entry assembly, and acts accordingly to that.

API Proposal

namespace System.Reflection;

public class Assembly
{
     public static Assembly? GetEntryAssembly(); // Existing API
+    public static void SetEntryAssembly(Assembly assembly);
}

API Usage

public class Program
{
    static void Main()
    {
        var watcher = new FileSystemWatcher
        {
            Path = "path_to_directory",
            Filter = "*.*",
            NotifyFilter = NotifyFilters.LastWrite
        };

        watcher.Changed += OnChanged;

        watcher.EnableRaisingEvents = true;

        Console.WriteLine("Press 'q' to quit the sample.");
        while (Console.Read() != 'q') ;
    }

    private static void OnChanged(object source, FileSystemEventArgs e)
    {
        new Thread(() =>
        {
            Assembly appToRun = Assembly.LoadFrom("path_to_other_assembly.dll");
            MethodInfo entryPoint = appToRun.EntryPoint;

            if (entryPoint != null)
            {
                Assembly.SetEntryAssembly(appToRun);      
                entryPoint.Invoke(null, new object[] { new string[] { } });
            }
        }).Start();
    }
}

Alternative Designs

One alternative may be to provide an API like the dotnet/arcade RemoteExecutor that makes it easier to launch a new process with copy of the existing runtime but running a different entry method (potentially from another assembly). That is insufficient for some use cases: some platforms (for example WebAssembly or mobile) do not support launching a new process; in other use cases it may be important to preserve part of the runtime state rather than starting from a brand new process.

Risks

The risk of this new API depends on the depth of the implementation. For diagnostics purposes, there are places in the runtime that cache the string path to the initial entry assembly. We would need to update and potentially change how that information is stored. Additionally, we may want to consider tricks that benefit startup should the entry assembly change early enough (within a startup hook, for example).

Other Notes to Consider

Questions:

Should this be unsupported on NativeAOT? Trimming?
Should we allow this to be called multiple times?
What is the impact on collectable assemblies? For example, if you call SetEntryAssembly with an assembly from a collectable ALC and then later set it to something else, should the collectable ALC be able to be collected?

dotnet-policy-service[bot] commented 2 weeks ago

Tagging subscribers to this area: @dotnet/area-system-reflection See info in area-owners.md if you want to be subscribed.

jkotas commented 2 weeks ago

Should this be unsupported on NativeAOT? Trimming?

There is nothing NativeAOT or trim incompatible in this API.

Should we allow this to be called multiple times?

Yes.

For example, if you call SetEntryAssembly with an assembly from a collectable ALC and then later set it to something else, should the collectable ALC be able to be collected?

Yes.

jkotas commented 2 weeks ago

diagnostics purposes, there are places in the runtime that cache the string path to the initial entry assembly. We would need to update and potentially change how that information is stored.

This is best effort. We won't be ever able to fully abstract away the fact that the application was launched through a launcher. For example, https://learn.microsoft.com/dotnet/api/system.diagnostics.process.mainmodule is going to point to the launcher .exe and not to the app.exe, and it is not something that we want to shim. If there are scenarios where diagnostic tools need to know about the current entrypoint, we can always add new diagnostic APIs to handle the new scenario if the existing diagnostic APIs are not sufficient.

jkotas commented 2 weeks ago

The proposal LGTM! Feel free to flip it to api-ready-to-review. (Also, mark it as blocking if you would like it to be reviewed with higher priority.)

stephentoub commented 2 weeks ago

public void SetEntryAssembly(Assembly assembly);

Is this supposed to be static?

MichalStrehovsky commented 1 week ago

Will there be any validation on the assembly used? E.g.:

Does it have to be the corelib-defined RuntimeAssembly, or can it be an assembly from e.g. MetadataLoadContext?
Can it be a reflection-emitted assembly, or an assembly loaded from a byte array (it's very common for apps to do Assembly.GetEntryAssembly().Location and dynamically loaded assemblies will have an empty Location).
Does it need a non-null EntryPoint?

jkotas commented 1 week ago

We may want to check that it is corelib-defined RuntimeAssembly as a sanity check. I do not think we need any other extraneous validation. It is up to the caller of this API to ensure that it does not break the rest of the app.

bartonjs commented 2 hours ago

Looks good as proposed

namespace System.Reflection;

public partial class Assembly
{
    public static void SetEntryAssembly(Assembly assembly);
}

dotnet / runtime