Use-after-free in _curses from unsynchronized access to shared ncurses SCREEN state

[jacobggman]

Crash report

What happened?

_curses releases the GIL around ncurses calls (refresh, getch, resizeterm, …) with no lock, while stock ncurses (built without --enable-reentrant, the Ubuntu/Debian default) is not thread-safe and keeps all screen state in global curscr/newscr row buffers. refresh()→doupdate uses those rows while resizeterm()→wresize frees them (wresize.c:227-240). Two Python threads therefore race into a use-after-free read+write of the shared screen → heap corruption → crash.

---ncurses version:
/usr/local/lib/libncursesw.so
/usr/local/lib/libncursesw.so.6
/usr/local/lib/libncursesw.so.6.4
ncurses 6.4.20240113
---OS:
PRETTY_NAME="Ubuntu 24.04.4 LTS"

#!/usr/bin/env python3
"""
Reliable use-after-free crash in CPython's `curses` module.

Root cause (Modules/_cursesmodule.c): several ncurses calls release the GIL
with no lock of their own -- notably window.refresh() -> wrefresh()->doupdate()
(_cursesmodule.c:4447/4459) and curses.resizeterm(). Stock ncurses (built
without --enable-reentrant, as on Ubuntu/Debian) is NOT thread-safe and keeps
all screen state in process-global structures.

doupdate()'s TransformLine() walks and writes the global curscr/newscr row
buffers (ncurses/tty/tty_update.c:1307-1308,1494,1625), while
resizeterm()->wresize() FREES those same buffers (ncurses/base/wresize.c:227-240).
Two Python threads therefore execute ncurses C concurrently on the shared,
unlocked SCREEN -> use-after-free read and write -> heap corruption ->
SIGSEGV / SIGABRT ("corrupted size vs. prev_size", "double free or
corruption").

Run (Unix, any CPython with _curses + non-reentrant ncurses):
    python3 poc_uaf_curses.py
Crashes reliably (usually within a couple of seconds). Re-run if it survives
one pass; raising NTHREADS/DURATION increases the chance.
"""
import curses
import sys
import threading
import time

NTHREADS = 8        # refresh threads racing on the shared SCREEN
DURATION = 15       # seconds to race before giving up


def main():
    stop = threading.Event()

    def hammer(win, tag):
        win.nodelay(True)
        i = 0
        while not stop.is_set():
            i += 1
            try:
                # dirty the window so doupdate has real work, then refresh
                # (refresh releases the GIL inside wrefresh()->doupdate()).
                win.addstr(0, 0, f"{tag}{i:06d}" + "x" * 40)
                win.refresh()
            except curses.error:
                pass

    def resizer():
        # Frees + reallocates curscr/newscr/stdscr row buffers via wresize()
        # while the refresh threads are inside doupdate() (GIL released).
        flip = False
        while not stop.is_set():
            flip = not flip
            try:
                curses.resizeterm(24 if flip else 25, 80 if flip else 81)
            except curses.error:
                pass

    def run(stdscr):
        curses.curs_set(0)
        stdscr.nodelay(True)
        wins = [curses.newwin(3, 60, (i % 20), 0) for i in range(NTHREADS)]
        threads = [threading.Thread(target=hammer, args=(w, chr(65 + i)),
                                    daemon=True)
                   for i, w in enumerate(wins)]
        threads += [threading.Thread(target=resizer, daemon=True) for _ in range(2)]
        for t in threads:
            t.start()
        t0 = time.monotonic()
        while time.monotonic() - t0 < DURATION and not stop.is_set():
            try:
                stdscr.noutrefresh()
                curses.doupdate()          # another GIL-releasing racer
            except curses.error:
                pass
            time.sleep(0.01)
        stop.set()

    curses.wrapper(run)
    print("Survived this run -- re-run, or raise NTHREADS/DURATION.",
          file=sys.stderr)


if __name__ == "__main__":
    main()

I would like to fix this myself by adding locking to the curses module.

CPython versions tested on:

3.16

Operating systems tested on:

Linux

Output from running 'python -VV' on the command line:

Python 3.16.0a0 (heads/main:c08c89a, Jul 19 2026, 02:39:49) [GCC 13.3.0]

Linked PRs

• gh-154139: Document that curses is not thread-safe #154173

[picnixz]

cc @serhiy-storchaka

[picnixz]

More generally, I'm not sure curses is meant to be thread-safe so I don't know if we accept that issue.

[jacobggman]

More generally, I'm not sure curses is meant to be thread-safe so I don't know if we accept that issue.

Maybe we should update the documentation to note that this package is not thread-safe?

[serhiy-storchaka]

Thanks, the analysis is correct: _curses releases the GIL around the blocking and refresh calls, so with a non-reentrant curses (the default ncurses build on Ubuntu) two threads can race on the shared screen state.

Whether curses is usable from multiple threads depends on the library and how it was built, so this isn't a missing lock in CPython. It's thread-safe when the library is built reentrant and access is serialized via use_window()/use_screen(); otherwise the application must serialize. A module-internal lock wouldn't work: held across a blocking getch() it would stall every other thread — the freeze the GIL release exists to avoid.

CPython 3.16 already exposes window.use() / screen.use() for this (effective when ncurses is built reentrant). The one gap is documentation, which I'll fix with a thread-safety note.