Open Stack

On 11/30/2015 02:15 PM, Sean Dague wrote:
> On 11/30/2015 03:01 PM, Robert Collins wrote:
>> On 1 December 2015 at 08:37, Ben Nemec <openstack at nemebean.com> wrote:
>>> On 11/30/2015 12:42 PM, Joshua Harlow wrote:
>>>> Hi all,
>>>>
>>>> I just wanted to bring up an issue, possible solution and get feedback
>>>> on it from folks because it seems to be an on-going problem that shows
>>>> up not when an application is initially deployed but as on-going
>>>> operation and running of that application proceeds (ie after running for
>>>> a period of time).
>>>>
>>>> The jist of the problem is the following:
>>>>
>>>> A <<pick your favorite openstack project>> has a need to ensure that no
>>>> application on the same machine can manipulate a given resource on that
>>>> same machine, so it uses the lock file pattern (acquire a *local* lock
>>>> file for that resource, manipulate that resource, release that lock
>>>> file) to do actions on that resource in a safe manner (note this does
>>>> not ensure safety outside of that machine, lock files are *not*
>>>> distributed locks).
>>>>
>>>> The api that we expose from oslo is typically accessed via the following:
>>>>
>>>>    oslo_concurrency.lockutils.synchronized(name, lock_file_prefix=None,
>>>> external=False, lock_path=None, semaphores=None, delay=0.01)
>>>>
>>>> or via its underlying library (that I extracted from oslo.concurrency
>>>> and have improved to add more usefulness) @
>>>> http://fasteners.readthedocs.org/
>>>>
>>>> The issue though for <<your favorite openstack project>> is that each of
>>>> these projects now typically has a large amount of lock files that exist
>>>> or have existed and no easy way to determine when those lock files can
>>>> be deleted (afaik no? periodic task exists in said projects to clean up
>>>> lock files, or to delete them when they are no longer in use...) so what
>>>> happens is bugs like https://bugs.launchpad.net/cinder/+bug/1432387
>>>> appear and there is no a simple solution to clean lock files up (since
>>>> oslo.concurrency is really not the right layer to know when a lock can
>>>> or can not be deleted, only the application knows that...)
>>>>
>>>> So then we get a few creative solutions like the following:
>>>>
>>>> - https://review.openstack.org/#/c/241663/
>>>> - https://review.openstack.org/#/c/239678/
>>>> - (and others?)
>>>>
>>>> So I wanted to ask the question, how are people involved in <<your
>>>> favorite openstack project>> cleaning up these files (are they at all?)
>>>>
>>>> Another idea that I have been proposing also is to use offset locks.
>>>>
>>>> This would allow for not creating X lock files, but create a *single*
>>>> lock file per project and use offsets into it as the way to lock. For
>>>> example nova could/would create a 1MB (or larger/smaller) *empty* file
>>>> for locks, that would allow for 1,048,576 locks to be used at the same
>>>> time, which honestly should be way more than enough, and then there
>>>> would not need to be any lock cleanup at all... Is there any reason this
>>>> wasn't initially done back way when this lock file code was created?
>>>> (https://github.com/harlowja/fasteners/pull/10 adds this functionality
>>>> to the underlying library if people want to look it over)
>>>
>>> I think the main reason was that even with a million locks available,
>>> you'd have to find a way to hash the lock names to offsets in the file,
>>> and a million isn't a very large collision space for that.  Having two
>>> differently named locks that hashed to the same offset would lead to
>>> incredibly confusing bugs.
>>>
>>> We could switch to requiring the projects to provide the offsets instead
>>> of hashing a string value, but that's just pushing the collision problem
>>> off onto every project that uses us.
>>>
>>> So that's the problem as I understand it, but where does that leave us
>>> for solutions?  First, there's
>>> https://github.com/openstack/oslo.concurrency/blob/master/oslo_concurrency/lockutils.py#L151
>>> which allows consumers to delete lock files when they're done with them.
>>>  Of course, in that case the onus is on the caller to make sure the lock
>>> couldn't possibly be in use anymore.
>>>
>>> Second, is this actually a problem?  Modern filesystems have absurdly
>>> large limits on the number of files in a directory, so it's highly
>>> unlikely we would ever exhaust that, and we're creating all zero byte
>>> files so there shouldn't be a significant space impact either.  In the
>>> past I believe our recommendation has been to simply create a cleanup
>>> job that runs on boot, before any of the OpenStack services start, that
>>> deletes all of the lock files.  At that point you know it's safe to
>>> delete them, and it prevents your lock file directory from growing forever.
>>
>> Not that high - ext3 (still the default for nova ephemeral
>> partitions!) has a limit of 64k in one directory.
>>
>> That said, I don't disagree - my thinkis is that we should advise
>> putting such files on a tmpfs.
> 
> So, I think the issue really is that the named external locks were
> originally thought to be handling some pretty sensitive critical
> sections. Both cinder / nova have less than 20 such named locks.
> 
> Cinder uses a parametrized version for all volume operations -
> https://github.com/openstack/cinder/blob/7fb767f2d652f070a20fd70d92585d61e56f3a50/cinder/volume/manager.py#L143
> 
> 
> Nova also does something similar in image cache
> https://github.com/openstack/nova/blob/1734ce7101982dd95f8fab1ab4815bd258a33744/nova/virt/libvirt/imagecache.py#L169
> 
> I honestly didn't realize that locks weren't deleting when completed due
> to the implementation details. Honestly, it seems like a busy wait
> try_lock / sleep might be better here than the open blocking, as it
> would let us stay on top of the cleanup, at the cost of a small amount
> of performance when contending for locks. But, honestly, most of these
> aren't performance critical bits, they are safety from corruption places.

I've always wondered if we could just delete the file instead of
unlocking it.  I'm not sure how that would interact with other processes
waiting on the file (they'll have already opened the file), but it's
possible we could do that with fairly minimal changes.  It would just
require some robust testing to make sure it doesn't break anything, and
like I said I've never seen this as a big problem so I didn't pursue it.