Can't call foreign function: clock_gettime

[kpp]

cargo +nightly miri -V # miri 0.1.0 (76138c5 2019-02-17)
rustc -vV # 1.34.0-nightly (146aa60f3 2019-02-18)

How to reproduce:

cd tokio
git log -n1 # commit 4985e0c6084646e0de2a42b361ca897882eff504
$ cargo +nightly miri test
   Compiling tokio v0.1.15 (/home/humbug/tokio)
running 1 test
error[E0080]: constant evaluation error: can't call foreign function: clock_gettime
   --> /home/humbug/.rustup/toolchains/nightly-x86_64-unknown-linux-gnu/lib/rustlib/src/rust/src/libstd/sys/unix/time.rs:361:13
    |
361 |             libc::clock_gettime(clock, &mut t.t)
    |             ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ can't call foreign function: clock_gettime
    |
    = note: inside call to `std::sys::unix::time::inner::now` at /home/humbug/.rustup/toolchains/nightly-x86_64-unknown-linux-gnu/lib/rustlib/src/rust/src/libstd/sys/unix/time.rs:275:26
    = note: inside call to `std::sys::unix::time::inner::Instant::now` at /home/humbug/.rustup/toolchains/nightly-x86_64-unknown-linux-gnu/lib/rustlib/src/rust/src/libstd/time.rs:158:22
    = note: inside call to `std::time::Instant::now` at /home/humbug/tokio/tokio-timer/src/clock/clock.rs:48:17
    = note: inside call to closure at /home/humbug/.rustup/toolchains/nightly-x86_64-unknown-linux-gnu/lib/rustlib/src/rust/src/libstd/thread/local.rs:300:16
    = note: inside call to `<std::thread::LocalKey<T>><std::cell::Cell<std::option::Option<*const tokio_timer::clock::Clock>>>::try_with::<[closure@DefId(54/1:9 ~ tokio_timer[4753]::clock[0]::clock[0]::now[0]::{{closure}}[0])], std::time::Instant>` at /home/humbug/.rustup/toolchains/nightly-x86_64-unknown-linux-gnu/lib/rustlib/src/rust/src/libstd/thread/local.rs:246:9
    = note: inside call to `<std::thread::LocalKey<T>><std::cell::Cell<std::option::Option<*const tokio_timer::clock::Clock>>>::with::<[closure@DefId(54/1:9 ~ tokio_timer[4753]::clock[0]::clock[0]::now[0]::{{closure}}[0])], std::time::Instant>` at /home/humbug/tokio/tokio-timer/src/clock/clock.rs:46:5
    = note: inside call to `tokio_timer::clock::now` at /home/humbug/tokio/tokio-timer/src/timeout.rs:128:40
note: inside call to `<tokio_timer::Timeout<T>><futures::Finished<(), ()>>::new` at src/util/future.rs:64:9
   --> src/util/future.rs:64:9
    |
64  |         Timeout::new(self, timeout)
    |         ^^^^^^^^^^^^^^^^^^^^^^^^^^^
note: inside call to `<futures::Finished<(), ()> as util::future::FutureExt>::timeout` at src/util/future.rs:90:32
   --> src/util/future.rs:90:32
    |
90  |         let timeouted_future = base_future.timeout(Duration::new(0, 0));
    |                                ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
note: inside call to `util::future::test::timeout_polls_at_least_once` at src/util/future.rs:88:5
   --> src/util/future.rs:88:5
    |
88  | /     fn timeout_polls_at_least_once() {
89  | |         let base_future = future::result::<(), ()>(Ok(()));
90  | |         let timeouted_future = base_future.timeout(Duration::new(0, 0));
91  | |         assert!(timeouted_future.wait().is_ok());
92  | |     }
    | |_____^

...
...
...

[bjorn3]

What value should it return? Always zero (makes timeout impossible), current time (not deterministic) or start with zero and increase every executed statement?

[oli-obk]

Even starting with zero and incrementing with every call to clock_gettime counts as "not deterministic", because it will differ from runtime behavior. Before implementing this, we'll need to add a flag to miri that turns nondeterministic behavior off (I think it should be the default). So maybe --deterministic.

[RalfJung]

Yeah, this is currently outside the range of what Miri is intended to support. We can't do I/O anyway so timeouts would hardly make much sense.

[kpp]

Yeah but tokio is an important part of Rust ecosystem but it scares me a little bit with its TSAN suppressions file. I would like to have a opportunity to run Tokio tests with miri (even a subset of them).

[RalfJung]

Sure it'd be great if we could run more things in Miri, just some things are much harder to get running than others. ;) Basically, for now the focus is on libraries that are platform-agnostic and do not need to access any but the most basic OS facilities. Even there we still have some gaps, such as the following program:

fn main() {
    let v: [Vec<i32>; 0] = [];
    let c = v.concat();
    assert_eq!(c, []);
}

(This is an instance of #224.)

---

TSAN suppressions file

Notice that Miri does not support concurrency, and if it did it'd likely not detect data races. So it is not a supplement for TSAN.

Don't get me wrong, I'd love to make Miri test tokio. It's just a lot of work. Don't expect us to implement epoll any time soon...

Does Tokio have a part of the test suite that can run without any I/O? Does that even make sense? That would be the kind of tests that I think you could hope to run in Miri. Another good idea might be to look at the infrastructure crates that tokio relies on, and test those.

[petrochenkov]

Can't miri execute these on host by dlopening libc and executing clock_gettime via dlsymed pointer, then return the result into the interpreter?
That would cover some basic cases at least.
(I haven't thought this through too much, so there may be obvious blockers.)

[oli-obk]

the dlopen scheme is basically what we're considering in #11

The "obvious blockers" are that it's not a priority for @RalfJung and I to spend implementation time on it ;)

[kpp]

Another good idea might be to look at the infrastructure crates that tokio relies on, and test those.

Thanks! I will try!

[RalfJung]

Can't miri execute these on host by dlopening libc and executing clock_gettime via dlsymed pointer, then return the result into the interpreter?

Besides the huge implementation complexity, that would also break cross-interpretaion -- currently Miri can execute MIR for any target on any host.

[petrochenkov]

Well, yes, this "semihosting" is supposed to be a separate limited mode, not something that would work on every program in any configuration.

[RalfJung]

With #683 coming up to tackle #653, it might be time to reconsider this.

My position is that we should have a flag like -Zmiri-allow-communication (or -Zmiri-disable-communication, not sure what the default should be) to enable Miri and the outside world to communicate (beyond mere stdout/stderr). This flag would let the outside world affect Miri's execution (through environment variables or getting the current time, and eventually through reading files or whatnot) and vice versa (through creating/writing files and so on). Do you think that would make sense?

[kpp]

I don't know.

cc @carllerche