I am proposing the following patch. Signal blocking effectively prevents signals from being processed by the child before it has the time to change disposition.
I re-ran my test script and the result is now correct. Signals sent to child are no longer caught by parent. With a small program I also tested what happens with pended signals if their disposition is changed before they're delivered. They're simply delivered according to the new disposition. Example. If a SIGTERM is sent to the child right after the fork, the signal will be pended. The child will then change the signal disposition to SIG_DFL (terminate). When the child unblocks the signals, SIGTERM is delivered to the child and the child terminates. Additional notes from http://man7.org/linux/man-pages/man7/signal.7.html: - A child created via fork(2) inherits a copy of its parent's signal mask; the signal mask is preserved across execve(2). - A child created via fork(2) initially has an empty pending signal set; the pending signal set is preserved across an execve(2). All this is on RHEL5.11. Regards, 2016-02-17 18:36 GMT-05:00 Philippe Anctil <[email protected]>: > Hi, > > I have looked further into this issue. Part of the solution would be to > reset signal disposition to SIG_DFL, either in the child after the fork, or > around the call to create_client in accept connection for the child to > inherit appropriate disposition from the start. Temporarily blocking > signals could help. I have to verify exactly how it behaves. > > By the way, why is SIG_CHLD set to null_handler in create_client? The > default disposition for this signal is SIG_IGN. > > Regards, > > > 2016-02-16 11:45 GMT-05:00 Philippe Anctil <[email protected]>: > >> Hi, >> >> In fork mode, sending a sigterm signal to a child is caught by the >> parent. I suspect it has something to do with signal_pipe being shared by >> the child and parent after the fork. >> >> Here I sent a signal to the parent while a child process was also >> running. The parent shut down as expected. After sending sigterm, the child >> remains. >> >> Processes before sigterm: >> >> UID PID PPID C STIME TTY STAT TIME CMD >> user 1423 1 0 10:32 ? Ss 0:00 >> /usr/local/stunnel/bin/stunnel.bin /usr/local/stunnel/etc/server.conf >> user 1921 1423 0 10:37 ? S 0:00 \_ >> /usr/local/stunnel/bin/stunnel.bin /usr/local/stunnel/etc/server.conf >> >> selected pid 1423 >> >> Processes after: >> >> UID PID PPID C STIME TTY STAT TIME CMD >> user 1921 1 0 10:37 ? S 0:00 >> /usr/local/stunnel/bin/stunnel.bin /usr/local/stunnel/etc/server.conf >> >> Here I sent the signal to the child. The parent shut down and the child >> completed normally. >> >> Processes before: >> >> UID PID PPID C STIME TTY STAT TIME CMD >> user 1276 1 0 10:31 ? Ss 0:00 >> /usr/local/stunnel/bin/stunnel.bin /usr/local/stunnel/etc/server.conf >> user 1309 1276 0 10:31 ? S 0:00 \_ >> /usr/local/stunnel/bin/stunnel.bin /usr/local/stunnel/etc/server.conf >> >> selected pid 1309 >> >> Processes after: >> >> UID PID PPID C STIME TTY STAT TIME CMD >> user 1309 1 0 10:31 ? S 0:00 >> /usr/local/stunnel/bin/stunnel.bin /usr/local/stunnel/etc/server.conf >> >> I know I should be trying threads. I will eventually. I can only switch >> once I validate it can sustain our level of traffic. Fork has done that >> flawlessly for many years now. >> >> Best regards, >> >> >> -- >> Philippe Anctil >> > > > > -- > Philippe Anctil > -- Philippe Anctil
fork_signal_disposition.patch
Description: Binary data
_______________________________________________ stunnel-users mailing list [email protected] https://www.stunnel.org/cgi-bin/mailman/listinfo/stunnel-users
