Don't start the reactor when running Miri

[Kestrer]

It can be useful to run async tests in Miri. However, #[tokio::test] will currently attempt to start the reactor, which will cause an error in Miri as it isn't supported. By only starting the reactor on #[cfg(not(miri))], we will be able to run tests using Miri without having to use futures_executor::block_on or something like that.

[Darksonn]

I suppose something like that could work.

[blasrodri]

@Koxiaet
You’re thinking of ignoring this line when running Miri?

[Kestrer]

@blasrodri Even better, enable_io (and enable_time if Miri doesn't support it) could become a no-op on Miri, which also makes sense and is perhaps simpler

[RalfJung]

FWIW, this should be done with care. cfg(miri) is mainly intended for test cases; we also use it in the standard library in a few places. There are situations in user code where this makes sense, and this might be one of them -- it is just important to be aware of the implication: the cfg'd-out code is not tested by Miri. So any such cfg introduces some risk that when testing things in Miri, one is not testing the things one thinks one is testing.

So when a library does this, it might be worth pointing this out somewhere in the documentation, at least.

[Darksonn]

It's possible that it is better to update tests to not enable IO so we don't run into the miri error on starting epoll. Then we can do it selectively on tests that don't depend on epoll, but are still e.g. threaded.

[blasrodri]

It's possible that it is better to update tests to not enable IO

Do you mean by modifying the proc macro, and changing the enable_all call in case is_test is true on function parse_knobs?

[Darksonn]

No, I meant modifying each test directly rather than modifying the macro. Tests which actually use IO should not be modified of course, as they simply cannot run with miri.

[carllerche]

Until Miri is stabilized as part of the rust stable channel, I am inclined to not factor it in. The runtime builder provides sufficient configuration options to disable the IO driver.

[tiif]

A quick search on miri in the issue tracker led me to this issue, currently the #[tokio::test] issue appeared to be solved.

[nurmohammed840]

@tiif, Just today I added this PR #6900, Without knowing that #[tokio::test] works on Linux, So I closed it.
Unfortunately, at the moment, it does not work on Windows.

[tiif]

Oh yea, you are right. I forgot that #[tokio::test] probably still won't work in MacOS or Windows. This is tracked under rust-lang/miri#1719.

[RalfJung]

The recommendation for non-Linux users that need file system access or epoll support is to run Miri with --target x86_64-unknown-linux-gnu. That should work everywhere and does not require any special tools.