Open Stack

Excerpts from Joshua Harlow's message of 2015-12-01 09:28:18 -0800:
> Sean Dague wrote:
> > On 12/01/2015 08:08 AM, Duncan Thomas wrote:
> >>
> >> On 1 December 2015 at 13:40, Sean Dague<sean at dague.net
> >> <mailto:sean at dague.net>>  wrote:
> >>
> >>
> >>      The current approach means locks block on their own, are processed in
> >>      the order they come in, but deletes aren't possible. The busy lock would
> >>      mean deletes were normal. Some extra cpu spent on waiting, and lock
> >>      order processing would be non deterministic. It's trade offs, but I
> >>      don't know anywhere that we are using locks as queues, so order
> >>      shouldn't matter. The cpu cost on the busy wait versus the lock file
> >>      cleanliness might be worth making. It would also let you actually see
> >>      what's locked from the outside pretty easily.
> >>
> >>
> >> The cinder locks are very much used as queues in places, e.g. making
> >> delete wait until after an image operation finishes. Given that cinder
> >> can already bring a node into resource issues while doing lots of image
> >> operations concurrently (such as creating lots of bootable volumes at
> >> once) I'd be resistant to anything that makes it worse to solve a
> >> cosmetic issue.
> >
> > Is that really a queue? Don't do X while Y is a lock. Do X, Y, Z, in
> > order after W is done is a queue. And what you've explains above about
> > Don't DELETE while DOING OTHER ACTION, is really just the queue model.
> >
> > What I mean by treating locks as queues was depending on X, Y, Z
> > happening in that order after W. With a busy wait approach they might
> > happen as Y, Z, X or X, Z, B, Y. They will all happen after W is done.
> > But relative to each other, or to new ops coming in, no real order is
> > enforced.
> >
> 
> So ummm, just so people know the fasteners lock code (and the stuff that 
> has existed for file locks in oslo.concurrency and prior to that 
> oslo-incubator...) never has guaranteed the aboved sequencing.
> 
> How it works (and has always worked) is the following:
> 
> 1. A lock object is created 
> (https://github.com/harlowja/fasteners/blob/master/fasteners/process_lock.py#L85)
> 2. That lock object acquire is performed 
> (https://github.com/harlowja/fasteners/blob/master/fasteners/process_lock.py#L125)
> 3. At that point do_open is called to ensure the file exists (if it 
> exists already it is opened in append mode, so no overwrite happen) and 
> the lock object has a reference to the file descriptor of that file 
> (https://github.com/harlowja/fasteners/blob/master/fasteners/process_lock.py#L112)
> 4. A retry loop starts, that repeats until either a provided timeout is 
> elapsed or the lock is acquired, the retry logic u can skip over but the 
> code that the retry loop calls is 
> https://github.com/harlowja/fasteners/blob/master/fasteners/process_lock.py#L92
> 
> The retry loop (really this loop @ 
> https://github.com/harlowja/fasteners/blob/master/fasteners/_utils.py#L87) 
> will idle for a given delay between the next attempt to lock the file, 
> so that means there is no queue like sequencing, and that if for example 
> entity A (who created lock object at t0) sleeps for 50 seconds between 
> delays and entity B (who created lock object at t1) and sleeps for 5 
> seconds between delays would prefer entity B getting it (since entity B 
> has a smaller retry delay).
> 
> So just fyi, I wouldn't be depending on these for queuing/ordering as is...
> 

Agreed, this form of fcntl locking is basically equivalent to
O_CREAT|O_EXCL locks as Sean described, since we never use the blocking
form. I'm not sure why though. The main reason one uses fcntl/flock is
to go ahead and block so waiters queue up efficiently. I'd tend to agree
with Sean that if we're going to busy wait, just using creation locks
will be simpler.

That said, I think what is missing is the metadata for efficiently
cleaning up stale locks. That can be done with fcntl or creation locks,
but with fcntl you have the kernel telling you for sure if the locking
process is still alive when you want to clean up and take the lock. With
creation, you need to write that information into the lock, and remove
it, and then have a way to make sure the process is alive and knows it
has the lock, and that is not exactly simple. For this reason only, I
suggest staying with fcntl.

Beyond that, perhaps what is needed is a tool in oslo_concurrency or
fasteners which one can use to prune stale locks based on said metadata.
Once that exists, a cron job running that is the simplest answer. Or if
need be, let the daemons spawn processes periodically to do that (you
can't use a greenthread, since you may be cleaning up your own locks and
fcntl will gladly let a process re-lock something it already has locked).