Eli Zaretskii <eliz@gnu.org> writes:

> I see that it's a small mess, but I don't think I understand your
> reasoning about setting the filter to t twice: it looks to me that
> both calls will clear the bit of p->infd, because they both will
> trigger this clause:
>
>       if (EQ (p->filter, Qt) && !EQ (p->status, Qlisten))

> Am I missing something?

Argh, no, I was confused.  The code in Emacs 25 is correct, although
(IMO) somewhat misleading.

>   . if the filter is not t, we will now call add_process_read_fd every
>     time, for no good reason

(This is a moot point due to the second problem, but)
add_process_read_fd just sets some bits, which is harmless even if
repeated.

>   . the patch changes how connect_network_socket works in ways that we
>     don't sufficiently understand

Yes, I agree this is a serious problem.

> I would like to leave connect_network_socket alone on the release
> branch, and just fix the problem with set-process-filter.  I think the
> easiest way is simply not to call set_process_filter_masks in
> set-process-filter, and instead restore the way that code worked
> before 9755b753, i.e. let it test the argument FILTER _before_ that
> filter is installed.  Do you see any problems with this fix?

I think that makes sense, patch attached.