[cmake] Don't let tests inherit the terminal of ctest as stdin

[guitargeek]

Tests added with ROOTTEST_ADD_TEST are wrapped in coreutils timeout, which moves them into a background process group. When ctest is run from a terminal, the tests inherit it as stdin, and any terminal access from a background process group stops the process with SIGTTOU. The test then hangs until the ctest timeout kills it.

Since Python 3.15, this happens for every doctest: doctest.testmod() creates a pdb.Pdb, whose constructor sets up the new _pyrepl console when stdin is a TTY, which calls tcsetattr() on it. The JupyROOT doctests therefore always timed out when running ctest interactively, while passing in the CI where there is no terminal.

Use /dev/null as stdin for tests that don't specify an input file, like in non-interactive environments.

🤖 Done with the help of AI

[github-actions]

Test Results

    23 files      23 suites   3d 20h 54m 54s ⏱️
 3 885 tests  3 884 ✅ 0 💤 1 ❌
79 647 runs  79 646 ✅ 0 💤 1 ❌

For more details on these failures, see this check.

Results for commit 494a260.

♻️ This comment has been updated with latest results.

[pcanal]

This is likely effective but is that the canonical ctest way to say there is no access to the terminal? Also don't we have the same/similar problem on Windows?

Related question why is "the pdb.Pdb constructor calling tcsetattr()"? Isn't there a missing argument to the python invocation to set 'this is a batch mode'?

[guitargeek]

Hi, thanks for the many questions and again looking at one of my Python 3.15 compatibility PRs! Let me try to unpack the concerns and address them one-by-one.

ctest: As far as I can tell there is no dedicated mechanism. ctest passes its own stdin to the tests (checked with a minimal project run from a pty: isatty(0) is true in the test), and there is no test property for it. The driver runs the command with execute_process, which inherits stdin unless INPUT_FILE is given, so INPUT_FILE /dev/null is the standard way to say "no stdin". It's also what tests already get in CI.

Windows: Not affected. The hang comes from POSIX job control: timeout puts the test in a background process group, where reading from or reconfiguring the terminal raises SIGTTIN/SIGTTOU and stops the process. We don't use timeout on Windows (if(TIMEOUT_BINARY AND NOT MSVC)), and the Windows console has no background groups, so I limited the change to UNIX.

pdb: Since 3.15, pdb uses the new _pyrepl line editor when stdin is a TTY (gh-145378), and Pdb.__init__ sets it up eagerly. The resulting UnixConsole calls tcsetattr() on stdin, which triggers SIGTTOU from a background group even though the settings it writes are unchanged. doctest creates a Pdb on every run to support pdb.set_trace() in examples, so every doctest hits this.

There's no batch-mode flag. The closest is the PYTHON_BASIC_REPL env variable, which pdb also honors. But that only fixes this one symptom: the real problem is that tests are given a terminal they don't own, and detaching stdin fixes that for every test.

[pcanal]

Fair enough. I agree that this is the simplest way forward.

[pcanal]

I recommend

Suggested change

	set(_input INPUT_FILE /dev/null)
	endif()
	set(_input INPUT_FILE /dev/null)
	else() # or elseif(WIN32)
	set(_input INPUT_FILE NUL)
	endif()

for symmetry and to avoid surprises.

[pcanal]

LGTM.