Invoker and booster instance make up a process pair that are meant to exit
at the same time. To arrange this a) when invoker is about to exit, it sends
a terminating signal to booster instance, and b) when booster instance
exits, booster daemon sends a terminating signal to invoker. Overall this
has worked well enough - save some hiccups from potential race conditions -
but it is not compatible with setup where invoker and booster daemon are
running in different namespaces and sending signals is not possible.
There is already an unix domain socket that is kept open for the lifetime of
boosted application - used for transferring application information from
invoker to booster daemon during startup and exit status from booster daemon
to invoker at exit time. This socket can be utilized also for detecting when
peer process exits.
Normally when application (booster instance) exits, booster daemon receives
SIGCHLD, collects application exit reason, forwards it to invoker via
booster socket, and invoker then makes exit with the same exit status as
what application used.
Augment this by having booster daemon watch over booster sockets and
terminate booster instance upon eof on socket. This accomplishes that
application gets killed if/when invoker dies.
Additionally all booster instances are terminated if booster daemon exits
due to SIGTERM.
In general, instead of simply closing booster socket at each end at exit
time, an orderly disconnect is done via: shutdown write end of the socket,
read data until eof is received, then close socket. If this is accomplished
successfully within reasonable time limit, there is no need to send signals
- both peers know that the other end is going to make an appropriate exit.
Previously booster daemon and invoker made an attempt to reproduce
application getting killed by some signal such as as SIGSEGV also at the
invoker side. As this produces false positive crash reports and complicates
things (some of the signals are terminal and can't be handled in
asynchronous manner) this is no longer done - only standard TERM and KILL
signals are used for terminating peers and even then it is done as a last
resort.
Invoker signal handler used non async signal safe functions, those
have been removed.
To ease ad-hoc debugging, logging is automatically switched from syslog to
stderr when booster/invoker is executed from interactive command line.
Signed-off-by: Simo Piiroinen <simo.piiroinen@jolla.com>
Simo Piiroinen
Juha Lintula