expose xwayland.Server's pid #207
Conversation
wlroots wants to fork() and waitpid() on the result: https://gitlab.freedesktop.org/wlroots/wlroots/-/commit/871646d22522141c45db2c0bfa1528d595bb69df (this code *almost* does the right thing, but only does SIG_BLOCK in the first fork, not the parent...) applications using wlroots which have SIGCHLD handlers installed may interfere with this, so expose the server pid so they can use WNOWAIT trickery to avoid reaping this pid, to allow wlroots to do so. Maybe this is the wrong patch to fix the described issue: maybe pywayland should block SIGCHLD (via signal.pthread_sigmask(), since applications may be multithreaded, especially if they use asyncio) and then do wlr_xwayland_server_create(), then unblock it for users? But either way, seems like people might want to see the pid here. Signed-off-by: Tycho Andersen <[email protected]>
|
That is what labwc is doing as well, it needs to clean up processes it spawns itself (+ get notified when a primary client terminates) and thus has a global SIGCHLD handler. That indeed breaks lazy initialization of xwayland and therefore labwc uses |
|
Yeah, and actually doing the ptrace_setsigmask() trick won't work, because then the SICHLD will be delivered somewhere else. It really feels like wlroots shouldn't fork here in favor of posix_spawn() or something. |
|
Oh... it turns out they already do a pipe to notify, and the waitpid() that causes the error is really just about zombie prevention, I think. I sent them a patch: https://gitlab.freedesktop.org/wlroots/wlroots/-/merge_requests/4926 |
SIGCHLD here isn't fatal: we have other means of notifying that things were successful or failure, and it causes many compositors to have to do a bunch of extra work: qtile/qtile#5101 flacjacket/pywlroots#207 (comment) djpohly/dwl#212 Signed-off-by: Tycho Andersen <[email protected]>
wlroots wants to fork() and waitpid() on the result:
https://gitlab.freedesktop.org/wlroots/wlroots/-/commit/871646d22522141c45db2c0bfa1528d595bb69df
(this code almost does the right thing, but only does SIG_BLOCK in the first fork, not the parent...)
applications using wlroots which have SIGCHLD handlers installed may interfere with this, so expose the server pid so they can use WNOWAIT trickery to avoid reaping this pid, to allow wlroots to do so.
Maybe this is the wrong patch to fix the described issue: maybe pywayland should block SIGCHLD (via signal.pthread_sigmask(), since applications may be multithreaded, especially if they use asyncio) and then do wlr_xwayland_server_create(), then unblock it for users? But either way, seems like people might want to see the pid here.