While there was no strong agreement on my change, Nathaniel and Giampaolo said that it should not harm to merge it :-) So I merged my change.

It has been said multiple times: my change doesn't fully close the race condition. It only reduces the risk that it happens. I rephrased my change to mention that there is still room for the race condition and that polling only *reduces* the risk, it doesn't fully fix the race condition.

--

Changing locking can help. But I'm not sure that it covers *all* cases. For example, what if another function calls os.waitpid() directly for whatever reason and the process completed? The pid recycling issue can also happen in this case, no?

Linux pidfd really fix the issue. If someone wants to enhance subprocess to use the new os.pidfd_open(), I suggest to open a separated issue.

Another approach is to emulate pidfd in userspace. Giampaolo explains that he also compares the process creation time for that (msg356571). I'm not excited to workaround OS issue this way. It requires to write platform specific get_create_time() code. I would prefer to reuse what the OS provides when available: Linux pidfd.

--

This issue title is quite explicit on purpose: "subprocess.Popen.send_signal() should poll the process". I now close this issue: send_signal() does now poll :-)

Because there is no *strong* consensus on my change, I decided to not backport it to stable branches 3.7 and 3.8.

Thanks everyone for this very interesting discussion!