EnvironmentVariable operations with EnvironmentVariableTarget.User fails on Windows Nano Server

[ViktorHofer]

This seems to be a product bug because of an underlying windows api change.

cc @Anipik @stephentoub @danmosemsft

[ViktorHofer]

Setting to 3.0 to make sure that the issue is triaged asap.

[ViktorHofer]

Disabled the test in dotnet/corefx@a969488.

[Anipik]

@ViktorHofer is any exception being thrown here ? do you have the failure\exception message ?

[ViktorHofer]

No exception being thrown in source code. Just assert errors in the different tests that test the behavior.

[ericstj]

@safern and I looked at this previously and were seeing that it's just storing stuff in the registry. I wonder if Nano removed the root key in the registry for user environment variables: https://github.com/dotnet/corefx/blob/954b1fdab8e98060cce3d0d88a34c3b0587a6acb/src/Common/src/CoreLib/System/Environment.Win32.cs#L51

@maryamariyan do you think you can take a look?

[danmoseley]

This may have to move out of 3.0 now given there's no customer reports (?)

@ViktorHofer was it previously passing on a different version of Nano, and the upgrade broke it?

[ViktorHofer]

was it previously passing on a different version of Nano, and the upgrade broke it?

Yes, the upgrade to Windows Nano Server 1809 broke this. I assume the number of customers that use .NET Core on Nano is limited.

[ViktorHofer]

We should look at the api changes in Nano 1809 and/or talk with the Windows team to understand the issue better. This isn't 3.0 specific so I'm fine moving it out but it probably affects all .NET Core versions that support Nano and use that API.

[maryamariyan]

I ran the test on Windows 10, Windows Nano Server 1809 and 1803.

The repro (failure for SetEnvironmentVariable) only happened in 1809.
The reason is HKEY_CURRENT_USER\Environment registry was not present.

After manually adding that in 1809, the test passes for Nano Server 1809 as well.

[danmoseley]

Sounds like we could reasonably work around that in code then (eg create hte key or skip the operation or whatever)

[ericstj]

Skipping the operation is what happens now: that's the bug. We should probably let the Windows folks know about this regression and confirm that removal of the key was intentional.

We could create the key, assuming we can get the ACLs right. On my machine it looks like this key has an explicit ACL (not inherited). We'd want to test what would happen when we create the key from elevated and non-elevated process (if such a thing exists in nano) to ensure we create the key with appropriate ACL.

We should test if we add the key, and create a new process, that the variable will be set in that process. Given Nano removed creation of the key, we need to confirm they didn't remove consumption of the key too.

[ViktorHofer]

Creating the key ourselves is a bit scary, let's first talk with Windows.