@asvetlov @gpshead I closed #23740 and created this because that one had several merge conflicts and unnecessary new content in configure script due to the lack of the "autoconf-archive" package on my system. The C code/documentation changes remain unchanged. Please take a look at this when you have time.

Edit: a gentle request for core review; no rush :)

I have resolved the conflicts for you. You have to pull the changes to your computer, regenerate the clinic and configure scripts and then commit the changes.

You also need to re-add utmp.h to the list of headers.

@tiran Thank you so much for your help. I will look into this very soon.

Edit: I have made the changes :)

@MaxwellDupre Thank you for the review.

@gpshead I have made the requested changes; please review again. You can ignore the other comments that I made. The following are the answers to your questions/requests; sorry for the slightly long response; I just wanted to provide you with all the information that I gathered.

QUESTION1: aren't there potentially some successful dup2 calls to clean up from in this error case?
ANSWER1: I feel like the function should not close the input argument fd upon any sort of failure and let the user take care of what they want to do with fd (please let me know if that is not acceptable). The only other type of cleanup that comes to mind is, say, close()-ing 0 if dup2(fd, 0) succeeds and dup2(fd, 1) fails. However, are closed standard descriptors (0 or 1 or 2) any better than invalid ones? We could save the original standard descriptors using dup() calls and restore them if any of the dup2()s fail but that feels a little absurd because then we have to handle the dup() errors. I looked at a number of implementations (glibc, musl, *BSD, Illumos, util-linux, etc) and none of them actually bother with handling the dup2() errors at all; see justification in this musl mailing list archive (https://musl.openwall.narkive.com/seFOxlLL/add-login-tty#post2). In fact, currently cpython's os.forkpty() (and hence pty.fork() and pty.spawn()) just calls the system's forkpty(), which, in most cases that I saw, internally calls the system's login_tty(), where dup2() errors are ignored.

QUESTION2: add a comment as to why this is opened to check for an error and immediately closed.
ANSWER2: After putting a lot of thought into it, I decided to remove the "archaic code" completely with the following short conversation that I had ~1.5 years earlier as justification: https://marc.info/?l=glibc-help&m=164436296715753&w=3 This means that now the fallback code only uses TIOCSCTTY and avoids the System V way of obtaining a controlling terminal by open()-ing a tty device file. Also, according to https://pubs.opengroup.org/onlinepubs/9699919799/basedefs/V1_chap11.html,

if a session leader has no controlling terminal, and opens a terminal device file that is not already associated with a session without using the O_NOCTTY option (see open()), it is implementation-defined whether the terminal becomes the controlling terminal of the session leader

the bold part if a session leader has no controlling terminal is why the glibc implementation first attempts to close() 0,1,2 before opening the tty device; but that means that further errors cannot be reported until the dup2(fd, 2) succeeds (if at all) since stderr is closed.

Thanks for making the requested changes!

@gpshead: please review the changes made to this pull request.

@gpshead Thank you for the clinic output regeneration :)

@gpshead Thank you for the merge.