[erlang-bugs] escript file operations fail on halt

[Michael Truog]

Hi,

There is an odd type of failure when:
1) async threads are enabled by default for the Erlang VM
2) an escript is used to spawn the Erlang VM
3) erlang:halt/1 is used to terminate the escript with a known error code

The erlang:halt/1 and erlang:halt/2 code here:
https://github.com/erlang/otp/blob/maint/erts/emulator/beam/bif.c#L3937
Makes the default flush parameter false!  The default flush parameter is currently undocumented.  So, when an escript performs a file operation that depends on the async thread pool (based on the internal Erlang code and configuration) and then attempts to do erlang:halt(integer()), the file operations may not complete or perhaps only partially complete.  In my particular use case, I can observe a rename file operation getting stuck inbetween the actual completion of the rename (and I am not using anything but a normal/default Linux filesystem, not NFS).

It seems important to change the default erlang:halt/1 behaviour for escript usage so that flush is true (I understand fail-fast probably means normal Erlang VM usage shouldn't have flush default to true).  An alternative is a new escript function that sets the flush option for the user (which is probably an easier solution to agree on) (e.g., escript:exit/1).

Thanks,
Michael