By not retrying, EINTR behaved like EOF. In other words, EINTR had consequences like the write side (kid side) closing the pipe early. Symptoms depended on the _read() caller. When $pipe_reader conflated EINTR with pipe closure, symptoms were application-specific. When _spawn reading the internal "sync pipe" conflated EINTR with pipe closure, the parent would fail to report exec failure. Add a test for the loss of exec failure report. The $pipe_reader problem is much tougher to test, because $pipe_reader only tries read() if select() found data available. Hence, the signal needs to arrive in the narrow window after read() enters the kernel, before read() returns already-available data. The issue report reproduced that on macOS only.
By not retrying, EINTR behaved like EOF. In other words, EINTR had consequences like the write side (kid side) closing the pipe early. Symptoms depended on the _read() caller. When $pipe_reader conflated EINTR with pipe closure, symptoms were application-specific. When _spawn reading the internal "sync pipe" conflated EINTR with pipe closure, the parent would fail to report exec failure. Add a test for the loss of exec failure report. The $pipe_reader problem is much tougher to test, because $pipe_reader only tries read() if select() found data available. Hence, the signal needs to arrive in the narrow window after read() enters the kernel, before read() returns already-available data. The issue report reproduced that on macOS only.
Fixes https://github.com/cpan-authors/IPC-Run/issues/176