Allow async exceptions to pierce masked waitForProcess

[mitchellwrosen]

Hello!

I believe I found a bug in waitForProcess. When run in the MaskedInterruptible state, an async exception doesn't actually bring the thread down.

For example, the typed-process library exports withProcess which is similar to withCreateProcess, but ends up calling waitForProcess in the cleanup action of a bracket.

Even though waitForProcess is an interruptible foreign call, I think what's happening is:

• Async exception enqueued for delivery to thread making waitForProcess call
• GHC sees it's in an interruptible foreign call, sends SIGPIPE (GHC docs)
• c_waitForProcess returns -1 with errno = EINTR
• Upon returning to Haskell, the pending exception is not delivered, because the masking state is MaskedInterruptible
• c_waitForProcess is retried because it's wrapped by throwErrnoIfMinus1Retry_

So our "interruptible" waitForProcess is only half-interrupted :)

My fix in this patch is to simply poll for any async exceptions before entering c_waitForProcess. Here is some related discussion I had with @snoyberg.

Thanks!

[snoyberg]

This looks good, I feel much better about the usage of allowInterrupt here. For the record: I checked and the new test case does fail when allowInterrupt is not used.

[snoyberg]

The test implementation is pretty fragile.

• The sleep value is low enough that the process could legitimately finish before the waitForProcess begins
• I'd rather not rely on BlockedIndefinitelyOnMVar, but instead use onException to putMVar a value in the case of an exception
• It's worth reporting the ec in the Right case; an exit code of 0 vs 127 will be highly informative, for instance

[mitchellwrosen]

Updated - how's that?

[snoyberg]

Looks good! One last request: can you add a comment to the ChangeLog explaining the change in semantics here?

[mitchellwrosen]

Done

[snoyberg]

Thanks!