mjhjmtu

I wrote a shell.nix file to build the development environment for one of my projects. I'm using a shellHook to ensure a postgresql server is started when you drop into the nix-shell.

The shellHook is essentially:

export PGDATA=$PWD/nix/pgdata

pg_ctl start --silent --log $PWD/log/pg.log

Despite the fact that pg_ctl starts a server in the background, if I type Ctrl-C in the shell, the server shuts down. If I set up the same scenario outside of nix-shell, this does not happen.

I'm new to strace, but it looks to me like the postgresql process is receiving SIGINT when I type Ctrl-C in my terminal:

$ strace -p $postgres_pid
strace: Process 20546 attached
select(6, [3 4 5], NULL, NULL, tv_sec=51, tv_usec=289149) = ? ERESTARTNOHAND (To be restarted if no handler)
--- SIGINT si_signo=SIGINT, si_code=SI_KERNEL ---
rt_sigprocmask(SIG_SETMASK, ~[ILL TRAP ABRT BUS FPE SEGV CONT SYS RTMIN RT_1], NULL, 8) = 0
write(2, "LOG: received fast shutdown req"..., 37) = 37
kill(20550, SIGTERM) = 0
...

The postgresql process is attached to the same controlling terminal (pts/12) as my nix-shell process (though this is also true when I run it outside of nix-shell):

$ ps -p $postgres_pid,$nixshell_pid -o pid,ppid,wchan,tty,cmd
 PID PPID WCHAN TT CMD
14608 18292 core_s pts/12 bash --rcfile /tmp/nix-shell-14608-0/rc
16355 1 core_s pts/12 /nix/store/xxxxxx-postgresql-9.6.8/bin/postgres

What's a good next step in debugging this? Should I read up on process groups?

Update: Trying a tip from another question, I found that this fixes the problem:

set -m
pg_ctl start --silent --log $PWD/log/pg.log

The weird thing is, according to $-, the m option was already set. Running echo $- produces imBH both before and after the set -m.

I noticed that in my interactive shells (whether nix-shell or not), $- is imBHs. The s is not present in the shellHook context, and I can't find an explanation of its meaning in the docs for Bash's set builtin. This may not be related though...

The manual says

start mode launches a new server. ... On Unix-like systems, by default, the server's standard output and standard error are sent to pg_ctl's standard output (not standard error). The standard output of pg_ctl should then be redirected to a file or piped to another process such as a log rotating program like rotatelogs; otherwise postgres will write its output to the controlling terminal (from the background) and will not leave the shell's process group. ...Use of either -l or output redirection is recommended.

It seems you need to redirect the output to detatch pg_ctl from the shell.

It seems the problem was that the postgresql server was running as part of the same process group as the shell that launched it via pg_ctl. Typing propagated a SIGINT to all processes in the group.

One way to fix this is to launch postgresql in its own session using setsid.

setsid pg_ctl start --silent --log $PWD/log/pg.log

That said, I still don't know why this only happens in the context of shellHook.