[Info-vax] New filesystem mentioned

[Dave Froble]

On 5/14/2019 1:07 PM, Mark Berryman wrote:
> On 5/14/19 9:36 AM, Dave Froble wrote:
>> .
>  > .
>  > .
>> So, I don't understand the comment about DLM calls in ZFS?  Locking is
>> not part of a file system, locking is used in the tools that access a
>> file system.  Things such as a database, or RMS, or other such.
>
> That would not be correct.

Well, first you say what I wrote is not correct.

>  At a minimum, access to the metadata of the
> file system must be coordinated.  Using the allocation bitmap as one
> example, you have the following:
>
> For a non-shared filesystem you have a couple of choices:
> 1. You single-thread all access to the bitmap so there is never more
> than one writer.
> 2. You take out a lock in local memory to coordinate updating the bitmap.
>
> For a share filed system, somewhat more work is involved.
>
> Given: node A and node B in a cluster.  Both node A and node B have
> cached the current copy of the allocation bitmap.
>
> Node A needs to make a change.  It requests an exclusive DLM lock to
> change the bitmap.  DLM will notify all other nodes that node A now has
> an exclusive lock and they must now invalidate their cached copy of the
> bitmap.  Node A updates the bitmap, and probably schedules a disk write
> thereof.
>
> What node A now does with the lock depends on the sharing strategy.  If
> metadata is shared by flushing to disk, node A can't convert the
> exclusive lock back to shared mode until after the disk write completes.
>   That will allow another node's request for an exclusive lock to
> complete and they will then read the updated bitmap back into memory.
>
> Or, perhaps there is an algorithm to inform the other nodes of what
> change was made.  In that case, node A can't release the exclusive lock
> until all other nodes inform node A they have made the necessary
> updates.  This would alleviate the need to flush through disk but would
> require some sort of algorithm to decide by whom and when write-backs
> are scheduled.
>
> Now another node needs to modify the bitmap.  That node needs to request
> an exclusive lock and can't proceed until it gets it.  Loop.

And then everything you write just confirms what I wrote.

:-)

> This is just one piece of metadata to which access must be coordinated,
> and coordinated in a manner that just doesn't exist on a non-shared
> filesystem.
>
> So, yes, there would be a number of changes required to take a
> filesystem from just about any non-VMS system and make it work on VMS.
>
> Mark Berryman
>

And on that we both agree.

-- 
David Froble                       Tel: 724-529-0450
Dave Froble Enterprises, Inc.      E-Mail: davef at tsoft-inc.com
DFE Ultralights, Inc.
170 Grimplin Road
Vanderbilt, PA  15486