Comment author: berke.durak

Quick workaround is to use Unix.dup.

Comment author: berke.durak

Erratum.

1. Flags in example should be changed to O_WRONLY
2. When that is done the issue no longer manifests itself.
3. However replacing Gc.major () with close_in re-creates the issue.
4. I noticed the issue in code using a mixture Lwt, Lablgtk and Unix, and using Unix.dup fixed my problem.

Comment author: @damiendoligez

The solution is not to increase the refcount. The refcount field is a count of how many heap objects reference this struct channel*, not how many struct channel *s reference the underlying FD.

If you initialize it to nonzero, the finalization function (caml_finalize_channel) will not work properly.

Moreover, the FD is never closed when the channel is GC'd because we are very careful to never automatically close a FD: since closing is a visible side effect, it should only be done explicitly by the user.

Comment author: gerd

Damien is right, and the test is invalid. write() raises EBADF because the descriptor is opened read-only.

Comment author: @damiendoligez

What we probably should do, is document the fact that Unix.in_channel_of_descr takes ownership of the underlying FD (and will close it when it is closed) and that the user must dup it if they don't like that.

Comment author: @xavierleroy

As of now, the GC doesn't close the underlying FD, and the documentation for Unix.{in,out}_channel_of_descr mentions "Closing the channel also closes the underlying file descriptor". Marking this PR as not needing further action.