[Pvfs2-users] VFS clients on a single machine appear to work

[Murali Vilayannur]

Hi Lorenzo,

> Thank you for the explanation. This looks promising.
> One issue I've noticed, which might be related to flock, is that on occasion, in one of the concurrent threads,
> open( , O_CREAT|O_WRONLY, 0755) will return -1 with errno "Permission denied". I'm running as root and the file and parent
> directory mode are 0755. When using the same test with the ext3 file system I don't get this error.

Hmmm. Could you do a
echo 1 > /proc/sys/pvfs2/debug,
run your tests and send the dmesg logs to me off-list if it is too big?

How reproducible is this error? Perhaps you could share this program for
us. I doubt this is related to locking, I suspect it could be yet another
permissions issue with the VFS code :(
Thanks,
Murali


> Lorenzo
>
>       From: Murali Vilayannur <vilayann at mcs.anl.gov>
>       To: Lorenzo Pinto <lxpinto at convera.com>
>       Cc: pvfs2-users at beowulf-underground.org
>       Subject: Re: [Pvfs2-users] VFS clients on a single machine appear to work with flock
>       Date: Mon, 10 Jul 2006 11:27:54 -0500 (CDT)
>
>  Hi Lorenzo,
>
> > I realize that flock and other flavors of advisory locking are not
> > officially supported by PVFS through VFS. However, with my setup of one
> > PVFS client machine, four I/O servers, and one metadata server
> > (different from the client machine), it does appear that multiple
> > applications and/or threads, on the one client machine, calling flock on
> > a pvfs2 file-system file behave correctly.  This can be tested using the
> > (FC5) flock application. As expected, this does not work with multiple
> > client machines.
>
> A generic flock() mechanism is implemented in the VFS layer
> fs/locks.c. Network file systems that allow clients to run on other
> physically separate machines require more code than what the VFS layer
> does to serialize readers.writers on a single node.
> We havent done that yet for PVFS and hence it wont work on multiple client
> nodes as you indicate below.
>
> > Now the question:  Is the behavior I'm observing by design, or is it
> > some coincidence that should not be relied upon.
>
> So it should work as long as they are all threads/processes on the same
> machine/running on the same kernel.
> Thanks,
> Murali
>
>
>
>
>
>