Fix panic=abort tests on fuchsia

[c6c7]

This PR goes one step further than #127594 by adding an unstable feature to make determining whether a Fuchsia process aborted more robust. There is certainly more work that needs to go in to landing the unstable feature, but I'm drafting this PR as a means to convey intent for the feature to survey others' thoughts.

r​? @tmandry

[rustbot]

r? @cuviper

rustbot has assigned @cuviper.
They will have a look at your PR within the next two weeks and either review your PR or reassign to another reviewer.

Use r? to explicitly pick a reviewer

[c6c7]

r? @tmandry

[c6c7]

I should state explicitly that a goal of this PR is to enable Fuchsia testing for tests/ui/process/signal-exit-status.rs without needing to copy-paste the magic Fuchsia status code that also appears in library/test/src/test_result.rs (and any futures authors needing to determine whether a Fuchsia process was aborted).

[c6c7]

#58590 references #102032 which ignored some tests because of the inability of those tests easily recognize Fuchsia's process status code in various error cases. This change should consider re-enabling testing for Fuchsia in those cases.

[c6c7]

After a lot of thought, I've changed the approach of this PR significantly. In short, Zircon does not implement signals as the definition of the C Standard Library expects for the implementation of libc::abort(). As a result, I think it's impossible to expose an API on top of a return code from Zircon to determine if a task was aborted or not. The best we can do is smooth the path for retrieving a ZX_TASK_RETCODE.

I will be pushing a new version of this PR shortly.

[c6c7]

The longer story with citations....

There are cases in this repository that attempt to detect whether a process aborted or not. The first is in fn get_result_from_exit_code(..., status: ExitStatus, ...) -> TestResult (defined in test::test_result) which returns whether a test failed based on an ExitStatus. If a platform implements libc::abort abort as defined in ISO C, a panic in a test will result in calling libc::abort which will call libc::raise(SIGABRT). Thus, get_result_from_exit_code will return a test failed when ExitStatus::signal returns Some(SIGABRT). The story does not end here though because Fuchsia and Windows don't support signals, and thus ExitStatus::signal always returns None on those platforms.

(For Windows, it's apparently sufficient to check for STATUS_FAIL_FAST_EXCEPTION, but I don't know the details beyond the existing comments in test::test_result)

For Fuchsia, libc::abort does not call libc::raise(SIGABRT) because Fuchsia in this case deviates from the ISO C definition of libc::abort. Instead, Fuchsia executes the LLVM intrinsic __builtin_trap()1 which generally raises an Zircon kernel exception. Unless the process installed an exception handler, the exception will go uncaught, and the kernel will kill the process with the code ZX_TASK_RETCODE_EXCEPTION_KILL. Thus, for get_result_from_exit_code, the best we can do to check if a process aborted is to observe the return code ZX_TASK_RETCODE_EXCEPTION_KILL.

The second case is tests/ui/process/signal-exit-status.rs which tests whether a process aborted as expected upon calling core::intrinsics::abort(). The goal is very similar to get_result_from_exit_code, but the result is not a TestResult. For Fuchsia, the hindrance for writing this test is not knowing what the return code from the process should be. My first thought was, "Oh, we should just implement ExitStatus::aborted_code(&self) -> Option<i32>." But this is fraught because of there is no return code in Fuchsia that is set if and only if a process aborted. The closest approximation to detecting if a process aborted is to check if the process returned ZX_TASK_RETCODE_EXCEPTION_KILL.

Taking a step back though, it's not actually the primary goal of signal-exit-status to determine that the process returned ZX_TASK_RETCODE_EXCEPTION_KILL and no other possible uncaught exception was raised. There is some potential for a test flake where the Zircon kernel during the signal-exit-status test raises an exception for a reason other than the abort() call; this seems very unlikely though. Thus, it's sufficient for signal-exit-status to merely check for ZX_TASK_RETCODE_EXCEPTION_KILL.

[c6c7]

The latest change modifies the approach to implement task_retcode() instead of aborted_code(). I believe this API more directly corresponds to the semantics around Fuchsia processes and correctly requires callers to reason about whether the ZX_TASK_RETCODE indicates the condition they expect to check.

[c6c7]

I'm not thrilled about adding these constants here since std::sys::pal::unix::process::zircon exists. However, since std::sys::pal::unix::process::zircon is private, I didn't want to cross the bridge of exposing parts of that module in this change.

[bors]

☔ The latest upstream changes (presumably #125443) made this pull request unmergeable. Please resolve the merge conflicts.