Is 10 seconds enough time to wait for the lock to release?

Also, might want to have the loop print status that it's waiting for lock to release...

@jlk What I've seen is that it all happens in a 2 to 3 second window. Usually the security group script doesn't run for a very long time, so that made me think 10 seconds was enough.

For safety reasons I'll bump it to 15 seconds and wait 500ms between each operation and log while waiting for the lock to be released.

I updated the code and tested it. We came across this mainly when Terraform was being used.

Just checked the logs of two hypervisors:

2016-02-10 10:00:33,519 - Lock on /var/lock/cloudstack_security_group.lock is being held by other process. Waiting for release.

2016-02-10 09:58:39,187 - Lock on /var/lock/cloudstack_security_group.lock is being held by other process. Waiting for release.

They only had to wait one cycle of 500ms before the lock was released.

Cool - LGTM!

To update, we ran more tests on our production platform spread out over 60 hypervisors we saw these lock message pop up in various occasions.

Internally we use Terraform a lot and that seems to hammer on the API

Can you take another look @rhtyd ?

It works on our systems in production

@wido in general alright, I've left few comments that can help us improve this
For master, looks like we'll need to build/publish a new 4.9+ a systemvm template -- in that case would you prefer to use a library included that we can use here like filelock or something else? (https://pypi.python.org/pypi/filelock/)

@rhtyd Changed it to a with statement for Pythonic.

I'd rather not use a external lib since with Basic Networking and KVM this runs on the hypervisor.

@wido thanks, can you comment on how an acquired lock is released or it's not critical to release the lock at all?

@rhtyd The lock does not have to be released. It is done when the script exists. This is usually within 100ms or so.

@wido I tested your original PR and the current updated PR, the locking never happened on my test env (Ubuntu x64).

I read the manpage, and found that when you open a file twice flock will treat both FDs independently while fcntl (the lockf usage) treats them as same FDs. This means that when two processes could run the python script at the same time, no locking would actually occur. We don't get the advisory exclusive lock we want.

I wrote a test program to demonstrate this: (for testing you may replace flock with lockf and see for yourself that it does not lock at all)
https://gist.github.com/rhtyd/5d9983e7d58311ba8e517eefd931a507

So, my suggestion is to:

• use flock instead of lockf, assuming and likely the file exists on local filesystem we can use flock with no issues
• use the previously used open instead of with open, as with open syntax closes file when the statements under it return
• at the end call the lock_file.close() optionally, otherwise on exiting file should be GC-ed and closed

@rhtyd Hmm, that's odd. My logs showed that the lock prevented the script from running twice and our problems were also solved with this.

Before we had race conditions on iptables programming, that is resolved with this patched version.

@rhtyd: I tested and you are partially right. My test script: https://gist.github.com/wido/a236e9f065cef85b0497990e3d4798e6

When using the 'with open() as X' syntax it doesn't work. My Gist however does if I run it twice:

In terminal #1:

wido@wido-desktop:~/Desktop$ ./lock-test.py 
Obtained lock

In terminal #2:

wido@wido-desktop:~/Desktop$ ./lock-test.py 
Failed to obtain lock!
wido@wido-desktop:~/Desktop$

I updated the PR again with this change.

@wido +1

reviewed updated PR, LGTM
cc @swill

@wido is this code currently deployed and working in your prod environment? @rhtyd any suggestions for validating this code before merge? Have you tested this?

@swill I've validated the lock logic in a gist above, I've tested the elusiveness of critical code that runs when a lock is acquired; security groups is implemented using ebtables (rules on bridges) therefore this script is run at hypervisor host (kvm in this case, by the agent) -- the changes ensure that no two executions of this script are allowed at the same time

as for prod testing, @wido can you comment on testing the changes in this PR, thanks

@swill and @rhtyd: The code in the PR is currently in production at PCextreme and running fine.

With a 'with' statement in Python I couldn't get a proper lock. But that's not really a functional change, it's merely a more Python-way of writing code.

I am pretty comfortable with this because @wido is already using this. I don't have an SG environment setup right now for testing, so I would have to build one specifically to test this. Do you guys agree this is ready to merge?

@swill no objection, let's merge this

tag:mergeready