On 03/03/2016 02:29 AM, Shyam wrote:
On 03/02/2016 03:10 AM, Avra Sengupta wrote:
Hi,
All fops in NSR, follow a specific workflow as described in this
UML(https://docs.google.com/presentation/d/1lxwox72n6ovfOwzmdlNCZBJ5vQcCaONvZva0aLWKUqk/edit?usp=sharing).
However all locking fops will follow a slightly different workflow as
described below. This is a first proposed draft for handling locks, and
we would like to hear your concerns and queries regarding the same.
This change, to handle locking FOPs differently, is due to what
limitation/problem? (apologies if I missed an earlier thread on the same)
My understanding is that this is due to the fact that the actual FOP
could fail/block (non-blocking/blocking) as there is an existing lock
held, and hence just adding a journal entry and meeting quorum, is not
sufficient for the success of the FOP (it is necessary though to
handle lock preservation in the event of leadership change), rather
acquiring the lock is. Is this understanding right?
Yes it is right, the change in approach for handling locks is to avoid
getting into a deadlock amogst the followers.
Based on the above understanding of mine, and the discussion below,
the intention seems to be to place the locking xlator below the
journal. What if we place this xlator above the journal, but add
requirements that FOPs handled by this xlator needs to reach the journal?
Assuming we adopt this strategy (i.e the locks xlator is above the
journal xlator), a successful lock acquisition by the locks xlator is
not enough to guarantee that the lock is preserved across the replica
group, hence it has to reach the journal and as a result pass through
other replica members journal and locks xlators as well.
If we do the above, what are the advantages and repercussions of the
same?
Why would we want to put the locking xlator above the journal. Is there
a use case for that?
Firstly, we would have to modify the locking xlator to make it pass through.
We would also introduce a small window where we perform the lock
successfully, but have a failure on the journal. We would then have to
release the lock because we failed to journal it. In the previous
approach, if we fail to journal it, we wouldn't even go to the locking
xlator. Logically it makes the locking xlator dependent on the journal's
output, whereas ideally the journal should be dependent on the locking
xlator's output.
Some of the points noted here (like conflicting non-blocking locks
when the previous lock is not yet released) could be handled. Also in
your scheme, what happens to blocking lock requests, the FOP will
block, there is no async return to handle the success/failure of the
same.
Yes the FOP will block on blocking lock requests. I assume that's the
behaviour today. Please correct me if I am wrong.
The downside is that on reconciliation we need to, potentially, undo
some of the locks that are held by the locks xlator (in the new
leader), which is outside the scope of the journal xlator.
Yes we need to do lock cleanup on reconciliation, which is anyways
outside the scope of the journal xlator. The reconciliation daemon will
compare the terms on each replica node, and either acquire or release
locks accordingly.
I also assume we need to do the same for the leases xlator as well, right?
Yes, as long as we handle locking properly leases xlators shouldn't be a
problem.
1. On receiving the lock, the leader will Journal the lock himself, and
then try to actually acquire the lock. At this point in time, if it
fails to acquire the lock, then it will invalidate the journal entry,
and return a -ve ack back to the client. However, if it is successful in
acquiring the lock, it will mark the journal entry as complete, and
forward the fop to the followers.
2. The followers on receiving the fop, will journal it, and then try to
actually acquire the lock. If it fails to acquire the lock, then it will
invalidate the journal entry, and return a -ve ack back to the leader.
If it is successful in acquiring the lock, it will mark the journal
entry as complete,and send a +ve ack to the leader.
3. The leader on receiving all acks, will perform a quorum check. If
quorum meets, it will send a +ve ack to the client. If the quorum fails,
it will send a rollback to the followers.
4. The followers on receiving the rollback, will journal it first, and
then release the acquired lock. It will update the rollback entry in the
journal as complete and send an ack to the leader.
5. The leader on receiving the rollback acks, will journal it's own
rollback, and then release the acquired lock. It will update the
rollback entry in the journal, and send a -ve ack to the client.
Few things to be noted in the above workflow are:
1. It will be a synchronous operation, across the replica volume.
2. Reconciliation will take care of nodes who have missed out the locks.
3. On a client disconnect, there will be a lock-timeout on whose
expiration all locks held by that particular client will be released.
Regards,
Avra
_______________________________________________
Gluster-devel mailing list
Gluster-devel@xxxxxxxxxxx
http://www.gluster.org/mailman/listinfo/gluster-devel
_______________________________________________
Gluster-devel mailing list
Gluster-devel@xxxxxxxxxxx
http://www.gluster.org/mailman/listinfo/gluster-devel
