Fun with file descriptor leaks

Here’s a fun little bash script:

1
2
3
4
5
6

#!/bin/bash
(
  sleep 20 &
)
ps -f $(pidof sleep)
echo "Bye"

Run it, and you’ll notice a few things:

• Because the subshell running sleep dies immediately, sleep gets reparented to init. (Interestingly enough, on newer Ubuntu releases this isn’t PID 1…), so the script doesn’t have any child processes by the time it prints “Bye”.
• After “Bye” is shown, the script exits immediately, returning control to the shell.

Now, call the script pied_piper.sh, and try the following:

1
2
3

./pied_piper.sh | cat
./pied_piper.sh | ts  # Awesome timestamping utility, same problem though
ssh localhost ./pied_piper.sh

Annoying, isn’t it? These commands won’t finish for 20 seconds! The problem is that sleep is keeping its stdout open, which is the input pipe for cat, ts, ssh, or whatever else you’re piping to (this is very annoying on Jenkins jobs as well).

If a third-party product is pissing you off this way - that is, it died, but somehow still keeps its pipe open, you can find the culprit like so:

1

fuser -v /proc/$PID_OF_PROCESS_WITH_OPEN_PIPE/fd/0

This will usually yield a sleep process as the culprit, with the useless parent information of init (as per my example). The only information you have is the precise delay - in my experience, it helps to find all “sleep” commands lurking about, and tinker with the delay amounts: Found a sleep 30? Change it to sleep 29, see if that’s what shows up.

Here’s how to actually fix the problem:

1
2
3
4
5
6

#!/bin/bash
(
  sleep 20 >&- 2>&- <&- &
)
ps -f $(pidof sleep)
echo "Bye"

This will close stdout, stderr and stdin. As a friend pointed out, it’s often safer to do > /dev/null rather than >&-, as some processes will crap out if they don’t have some semblence of an stdout. However, >&- is shorter, faster, and perfectly safe for sleep.

Of course, it’s better to save the PID for this sleep and kill it when appropriate from within the script - otherwise, you might be accumulating many useless sleep processes.