[Samba] Winbind process stuck at 100% after changing use_mmap to no

[Volker Lendecke]

On Thu, Jul 14, 2016 at 02:39:48PM +0100, Alex Crow wrote:
> >The main hint I would like to give to the MooseFS developers is to get
> >ping_pong -rw working. If you need any assistance in setting that up
> >and getting arguments on why this is important, let us know!
> >
> >With best regards,
> >
> >Volker
> 
> Any arguments that I can pass on would be much appreciated.

The main argument is that Windows applications tend to coordinate access
to files via byte-range locks. They expect data coherency under those
byte range locks, meaning that if a value is written under a brlock,
the brlock is released, and another process reads the value under the
brlock, the value written must be visible. If this is not guaranteed,
then data corruption will likely happen. The main question to the
MooseFS developers would be under which circumstances a value written
via pwrite on node A is visible via pread on node B. ping_pong checks
this sequentiality with fcntl byte range locks, which are the Posix
way of serializing access to ranges of a file. If MooseFS gives other
guarantees, we would like to know about them.

> I've also had a try with RozoFS, which did not work as expected as
> it only supports block-level locks (4096 bytes) so the first lock
> prevent further locks in that block range, which naturally causes
> ping_pong to stall. However I am told by their devs that this should
> not be an issue for CTDB's recovery lock or Samba as they should not
> need to lock ranges of less than a 4k, and they in fact integrate
> CTDB for their customers.

That's right. For the recovery lock this does not matter. But for plain
file access it will. On RozoFS you should set "posix locking = no",
but then you depend on external serialization via ctdb: With "strict
locking = yes" (which is default), Samba does check byte range locks and
this way serializes access to user files on RozoFS. This serialization
however is not coordinated with the cluster file system in question,
so the RozoFS developers might have to tell us about their approach to
data coherency too.

With best regards,

Volker Lendecke