afprog: manual cloexec

[Kokan]

Parent waits for child to close all fd before exec.

Fixes #3409.

[kira-syslogng]

Build SUCCESS

[alltilla]

Do we still need the CLOEXEC flag on the fds? In other words: can exec() be called from any other place, than the program dest/src? Maybe our java or python bindings do such thing.

We might be able to remove the CLOEXEC flag, because it is replaced by this new logic.

[kira-syslogng]

Build SUCCESS

[Kokan]

Both java and python languages bindings uses threads and not forking. In spite of those modules there is still fork used in syslog-ng.

1. The very first when syslog-ng stars and spawn a supervisor for itself (as needed); which is not an issue as that is created before anything else, so there is interesting open fds.
2. command-hooks also uses fork (but it directly calls system which yields fork+exec)

Not to mention that we could have a lib that does fork+exec without us knowing (sure not likely), imho better to keep them.

[alltilla]

This works perfectly, thanks! 👍

[bazsi]

With the latest iteration, nothing ensures that sync-fd is closed after all inherited sockets, so you would need to to dup2() it to fd 3 to make sure. If we inherited fd 3, that would simply close it before the iteration, but then syncfd is closed last.

[gaborznagy]

Besides Bazsi's comment to use fd 3 approve from my side.

[kira-syslogng]

Build SUCCESS

[alltilla]

Thanks!

[kira-syslogng]

Build SUCCESS

[gaborznagy]

Thanks!