Gearman worker in shell hangs as a zombie - shell

Question

I have a Gearman worker in a shell script started with perp in the following way:
runuid -s gds \
/usr/bin/gearman -h 127.0.0.1 -t 1000 -w -f gds-rel \
-- xargs /home/gds/gds-rel-worker.sh < /dev/null 2>/dev/null
The worker only does some input validation and calls another shell script run.sh that invokes bash, curl, Terragrunt, Terraform, Ansible and gcloud to provision and update resources in GCP like this:
./run.sh --release 1.2.3 2>&1 >> /var/log/gds-release
The script is intended to run unattended. The problem I have is that after the job finishes successfully (that's both shell scripts run.sh and gds-rel-worker.sh) the Gearman job remains executing, because the child process becomes zombie (see last line below).
root 144748 1 0 Apr29 ? 00:00:00 perpboot -d /etc/perp
root 144749 144748 0 Apr29 ? 00:00:00 \_ tinylog -k 8 -s 100000 -t -z /var/log/perp/perpd-root
root 144750 144748 0 Apr29 ? 00:00:00 \_ perpd /etc/perp
root 2492482 144750 0 May14 ? 00:00:00 \_ tinylog (gearmand) -k 10 -s 100000000 -t -z /var/log/perp/gearmand
gearmand 2492483 144750 0 May14 ? 00:00:08 \_ /usr/sbin/gearmand -L 127.0.0.1 -p 4730 --verbose INFO --log-file stderr --keepalive --keepalive-idle 120 --keepalive-interval 120 --keepalive-count 3 --round-robin --threads 36 --worker-wakeup 3 --job-retries 1
root 2531800 144750 0 May14 ? 00:00:00 \_ tinylog (gds-rel-worker) -k 10 -s 100000000 -t -z /var/log/perp/gds-rel-worker
gds 2531801 144750 0 May14 ? 00:00:00 \_ /usr/bin/gearman -h 127.0.0.1 -t 1000 -w -f gds-rel -- xargs /home/gds/gds-rel-worker.sh
gds 2531880 2531801 0 May14 ? 00:00:00 \_ [xargs] <defunct>
So far I have traced the problem to run.sh, because if I replace its call with something simpler (e.g. echo "Hello"; sleep 5) the worker does not hang. Unfortunately, I have no clue what is causing the problem. The script run.sh is rather long and complex, but has been working without a problem so far. Tracing the worker process I see this:
getpid() = 2531801
write(2, "gearman: ", 9) = 9
write(2, "gearman_worker_work", 19) = 19
write(2, " : ", 3) = 3
write(2, "gearman_wait(GEARMAN_TIMEOUT) ti"..., 151) = 151
write(2, "\n", 1) = 1
sendto(5, "\0REQ\0\0\0'\0\0\0\0", 12, MSG_NOSIGNAL, NULL, 0) = 12
recvfrom(5, "\0RES\0\0\0\n\0\0\0\0", 8192, MSG_NOSIGNAL, NULL, NULL) = 12
sendto(5, "\0REQ\0\0\0\4\0\0\0\0", 12, MSG_NOSIGNAL, NULL, 0) = 12
poll([{fd=5, events=POLLIN}, {fd=3, events=POLLIN}], 2, 1000) = 1 ([{fd=5, revents=POLLIN}])
sendto(5, "\0REQ\0\0\0'\0\0\0\0", 12, MSG_NOSIGNAL, NULL, 0) = 12
recvfrom(5, "\0RES\0\0\0\6\0\0\0\0\0RES\0\0\0(\0\0\0QH:terra-"..., 8192, MSG_NOSIGNAL, NULL, NULL) = 105
pipe([6, 7]) = 0
pipe([8, 9]) = 0
clone(child_stack=NULL, flags=CLONE_CHILD_CLEARTID|CLONE_CHILD_SETTID|SIGCHLD, child_tidptr=0x7fea38480a50) = 2531880
close(6) = 0
close(9) = 0
write(7, "1.2.3\n", 18) = 6
close(7) = 0
read(8, "which: no terraform-0.14 in (/us"..., 1024) = 80
read(8, "Identity added: /home/gds/.ssh/i"..., 1024) = 54
read(8, 0x7fff6251f5b0, 1024) = ? ERESTARTSYS (To be restarted if SA_RESTART is set)
--- SIGCHLD {si_signo=SIGCHLD, si_code=CLD_EXITED, si_pid=2531880, si_uid=1006, si_status=0, si_utime=0, si_stime=0} ---
read(8,
So the worker continues reading standard output even though the child has finished successfully and presumably closed it. Any ideas how to catch what causes this problem?

Answer 1

I was able to solve it. The script run.sh was starting ssh-agent, which opens a socket and since Gearman redirects all outputs the worker continued reading the open file descriptor even after the script successfully completed.
I found it by examining the open file descriptors for the Gearman worker process after it hang:
# ls -l /proc/2531801/fd/*
lr-x------. 1 gds devops 64 May 17 11:26 /proc/2531801/fd/0 -> /dev/null
l-wx------. 1 gds devops 64 May 17 11:26 /proc/2531801/fd/1 -> 'pipe:[9356665]'
l-wx------. 1 gds devops 64 May 17 11:26 /proc/2531801/fd/2 -> 'pipe:[9356665]'
lr-x------. 1 gds devops 64 May 17 11:26 /proc/2531801/fd/3 -> 'pipe:[9357481]'
l-wx------. 1 gds devops 64 May 17 11:26 /proc/2531801/fd/4 -> 'pipe:[9357481]'
lrwx------. 1 gds devops 64 May 17 11:26 /proc/2531801/fd/5 -> 'socket:[9357482]'
lr-x------. 1 gds devops 64 May 17 11:26 /proc/2531801/fd/8 -> 'pipe:[9369888]'
Then identified the processes using file node for the pipe in file descriptor 8 that German worker continued reading:
# lsof | grep 9369888
gearman 2531801 gds 8r FIFO 0,13 0t0 9369888 pipe
ssh-agent 2531899 gds 9w FIFO 0,13 0t0 9369888 pipe
And finally listed files opened by ssh-agent and found what stands behind file descriptor 3:
# ls -l /proc/2531899/fd/*
lrwx------. 1 root root 64 May 17 11:14 /proc/2531899/fd/0 -> /dev/null
lrwx------. 1 root root 64 May 17 11:14 /proc/2531899/fd/1 -> /dev/null
lrwx------. 1 root root 64 May 17 11:14 /proc/2531899/fd/2 -> /dev/null
lrwx------. 1 root root 64 May 17 11:14 /proc/2531899/fd/3 -> 'socket:[9346577]'
# lsof | grep 9346577
ssh-agent 2531899 gds 3u unix 0xffff89016fd34000 0t0 9346577 /tmp/ssh-0b14coFWhy40/agent.2531898 type=STREAM
As a solution I added kill of the ssh-agent before exit from run.sh script and now there are no jobs hanging due to zombie process.