For a while now the signal handling machinery has piggy-backed on ceval's "pending calls" machinery (e.g. Py_AddPendingCall).  This is a bit confusing.  It also increases the risk with unrelated changes to the pending calls code.

In the master branch, the signal handler only uses pending calls to report error on writing into the "wakeup_fd" (fd or socket handle):

commit c08177a1ccad2ed0d50898c2731b518c631aed14
Author: Antoine Pitrou <pitrou@free.fr>
Date:   Wed Jun 28 23:29:29 2017 +0200

    bpo-30703: Improve signal delivery (#2415)

Correct.  The remaining call to Py_AddPendingCall in the signal-handling code is fine.

This issue is only indirectly related.  I suppose you could consider it a follow-up to #30703.  The PR for that issue (GH-2415) switches from using pending calls for signal handlers to using the pending calls machinery without actual pending calls.

So here I want to address taking the next step: deal with pending signals separately from pending calls.  That separation helps simplify efforts to adapt the pending calls machinery for use in arbitrary threads (rather than the main thread).  See #33608.

New changeset fdf282d609fd172d52b59a6f1f062eb701494528 by Eric Snow in branch 'master':
bpo-35423: Stop using the "pending calls" machinery for signals. (gh-10972)
https://github.com/python/cpython/commit/fdf282d609fd172d52b59a6f1f062eb701494528