It would probably be possible to implement a separate tool for this as well, but I think OOD might want to integrate it into nginx_stage nginx_clean for completeness.

I agree. I'm not sure off the top why it wouldn't clean a PUN for a user that doesn't exist, so I'll have to investigate why it's not cleaning those PUNs. But I agree that it should.

I'm not sure off the top why it wouldn't clean a PUN for a user that doesn't exist

The issue is that the code tries to resolve the username (based on the /var/run/ondemand-nginx/${user}/passenger.pid files that exist on the system) into a struct passwd, to use C terminology. If nothing is found, weird things happen.

I tested this on OOD 3.1.9 using the following methodology:

• Log into one of our testing instances, which creates a PUN for my user
• For the record, I was the only user logged in on the instance.
• Remove the user data by renaming the passwd.db and group.db databases so that they aren't recognized automatically. (See nss_db for more info.)
• Run /opt/ood/nginx_stage/sbin/nginx_stage nginx_clean

The outcome is quite interesting. I get an error message saying:

missing PID file: /var/run/ondemand-nginx/true/passenger.pid

and the nginx_clean exits with code 0. So the error is caught somewhere, but doesn't reflect the original error, being that my username didn't resolve into a valid user account.

I also strace:d the program while testing this and it didn't show any interesting errors on the syscall side. It basically checked what's inside the /var/run/ondemand-nginx/ directory, found my passenger.pid file, and tries to resolve the username into a user, which doesn't return anything. After this, there's an attempt to stat("/var/run/ondemand-nginx/true/passenger.pid"), which returns an ENOENT ("No such file or directory") and proceeds to print the warning message.

I suspect this is where the "all users do exist" assumption starts being a problem:

I'm trying to replicate the circumstances that would get us into this spot. Though I think I need ldap because I can't userdel in a container while the PUN is still active.

I got the hpc-toolset working to do this. Unfortunately, if it works correctly and gets sfoster (an account that doesn't exist anymore) nginx can't actually stop that process because it fails to getpwnam on that same user.

(I added one or two output statements here to see the nginx command)

[root@ondemand nginx_stage]# sbin/nginx_stage nginx_clean -f
looking at user sfoster
sfoster
/opt/ood/ondemand/root/usr/sbin/nginx ["-c", "/var/lib/ondemand-nginx/config/puns/sfoster.conf", "-s", "stop"]
nginx: [emerg] getpwnam("sfoster") failed in /var/lib/ondemand-nginx/config/puns/sfoster.conf:1
[root@ondemand nginx_stage]# vim lib/nginx_stage/generators/nginx_clean_generator.rb 

So I'm not really sure what to do here - save for searching for and kill -9ing the processes still running...

The issue is that PUNs belonging to a UID which doesn't have a valid user attached to it any longer won't be cleaned up.

I'm now wondering why you delete the user instead of just disabling it? Seems like if you delete a user you'd have to do all sorts of cleanup tasks afterwards (like cleaning their HOME and so on). This may just be one more cleanup task you need to take care of.

Thanks for looking into this!

nginx can't actually stop that process because it fails to getpwnam on that same user.

Okay, that's good to know.

I'm now wondering why you delete the user instead of just disabling it?

I won't go into details, but this is how our system is set up to work. Users who aren't authorized to compute won't exist either.

So I'm not really sure what to do here - save for searching for and kill -9ing the processes still running...

There's no need to go to SIGKILL immediately. In fact, when the nginx ... stop command successfully executes, it will simply look up the pid file of the given server configuration, and send a SIGTERM to that process. Below is an excerpt from a strace of nginx_stage nginx_clean -f, where the redacted_user still existed:

1900144 1730970218.810405 openat(AT_FDCWD, "/var/run/ondemand-nginx/redacted_user/passenger.pid", O_RDONLY) = 4
1900144 1730970218.810509 pread64(4, "127797\n", 22, 0) = 7
1900144 1730970218.810593 close(4)      = 0
1900144 1730970218.810692 kill(127797, SIGTERM) = 0
1900144 1730970218.811008 exit_group(0) = ?
1900144 1730970218.811979 +++ exited with 0 +++

I.e., the solution to this issue would in my opinion simply do something similar, without the /opt/ood/ondemand/root/usr/sbin/nginx program:

pid="$(cat /var/run/ondemand-nginx/${redacted_user:?}/passenger.pid)"
/usr/bin/kill -s TERM "$pid"
# Remove the empty PID file directory
rmdir /var/run/ondemand-nginx/${redacted_user:?}
# Remove the PUN config and secret_key_base (kind of hacky, please revisit)
find /var/lib/ondemand-nginx/config/puns/ -name "${redacted_user:?}.*" -delete

The PUN config and secret_key_base seems to be based on usernames only, so this should be easy to integrate at the same time:

I have a PR in to also cleanup non existing PUNs/config files. Looking at the git blame, looks like we've been at it a while, you created #1833 😆.