Spencer Baugh <sbaugh@janestreet.com> writes:
> While doing this, I had another thought...
>
> Do you know if we have any code which handles the fact that
> delete-process might be called on a process while another thread is
> calling pselect on its fds?
>
> I don't think we do, which suggests that might also be problematic.
> I'll work on a fix which encompasses that as well.

OK, I think this patch fixes the issue.

Many different pieces of code in Emacs can close file descriptors.  We
need to be generically careful to not close file descriptors which some
other thread is currently waiting on.  In my patch, we do this by
letting the waiting_thread itself close those file descriptors, after it
returns from select.

Unfortunately I don't have a reliable test which produces the failure
case, but in this patch I added a fprintf when we close a deferred fd.
This prints whenever some thread would have closed an fd that another
thread was waiting on.  If I run the following program (with emacs -Q
--batch):

(defun my-break-thread ()
  (let ((proc (make-process
               :name "foo"
               :command '("sleep" "1"))))
    (while (process-live-p proc)
      (accept-process-output proc 0.05))))
(while t
  (make-thread #'my-break-thread "thread1")
  (thread-join (make-thread #'my-break-thread "thread2")))

I get about one print every 10 seconds.  Each print is a problem case
which has been averted, so I do think this patch fixes the bug.

(Obviously, we should remove the fprintf before applying the change)