libksane seems to break QProcess::start calls

[Thiago Macieira]

On Thursday, 3 March 2022 11:00:07 PST Tobias Leupold wrote:
> Am Donnerstag, 3. März 2022, 19:06:21 CET schrieb Stefan Brüns:
> > Are you using either Qt5 < 5.15 or a kernel version which does not support
> > CLONE_FD? - then you are relying on SIGCHLD for process exit notification.
> > 
> > CLONE_FD: https://lwn.net/Articles/636646/
> > Qt5: https://codereview.qt-project.org/c/qt/qtbase/+/108456/
> > 
> > sane-backends/backend/plustek-usbhw.c messes with the signal handlers and
> > fails to restore it: `sigaction(..., ..., NULL)`

If CLONE_PIDFD is active, we don't use the SIGCHLD handler to be notified of 
process exit, but SIGCHLD must not be ignored with SIG_IGN. That causes the 
kernel not to notify anything.

https://gitlab.com/sane-project/backends/-/blob/master/backend/plustek-usbhw.c

I see sigaction, but it is only messing with SIGALRM. Yes, it should restore 
previous the handler (probably SIG_DFL) after the alarm is done, but this 
isn't the issue. It also messes with the signal block mask, but I don't see a 
problem there either: it blocks and unblocks SIGALRM.

> I use Qt 5.15.2 and Gentoo's kernel 5.15.16. Is there something I can check
> to figure out if this could cause the problem? Apparently, there's no
> kernel option I can set?

You can strace the process. Add the options -f -bexecve to see the current 
process and its forks & threads, but stop when it executes a child process. 
You should see SIGCHLD line, like this:

kill(669903, SIGTERM)                   = 0
poll([{fd=3, events=0}, {fd=5, events=0}, {fd=6, events=POLLIN}, {fd=8, 
events=POLLIN}, {fd=12, events=POLLIN}], 5, 5000) = 4 ([{fd=5, 
revents=POLLERR}, {fd=6, revents=POLLHUP}, {fd=8, revents=POLLHUP}, {fd=12, 
revents=POLLIN}])
--- SIGCHLD {si_signo=SIGCHLD, si_code=CLD_KILLED, si_pid=669903, si_uid=1000, 
si_status=SIGTERM, si_utime=0, si_stime=0} ---
fcntl(12, F_GETFL)                      = 0x802 (flags O_RDWR|O_NONBLOCK)
waitid(P_PIDFD, 12, {si_signo=SIGCHLD, si_code=CLD_KILLED, si_pid=669903, 
si_uid=1000, si_status=SIGTERM, si_utime=0, si_stime=0}, WNOHANG|WEXITED, 
NULL) = 0

fd 12 is the pidfd in this case.

Anyway, search for SIGCHLD in the trace and see if anything is setting a 
handler for it. On kernels >= 5.4, Qt doesn't do it, so if you see any mention 
of it aside from deliveries like above, it came from elsewhere.

> Apart from that, the scanner in question actually does use the plustek
> backend! I fear there's no way to fix this in my code?!

Fix the problem where the problem is.

-- 
Thiago Macieira - thiago (AT) macieira.info - thiago (AT) kde.org
   Software Architect - Intel DPG Cloud Engineering