[Date Prev][Date Next]   [Thread Prev][Thread Next]   [Thread Index] [Date Index] [Author Index]

Re: [Cluster-devel] [GFS2] Don't use journal lock type



Hi,

On Tue, 2007-08-07 at 09:02 -0500, David Teigland wrote:
> On Tue, Aug 07, 2007 at 02:52:31PM +0100, Steven Whitehouse wrote:
> > Hi,
> > 
> > On Tue, 2007-08-07 at 08:45 -0500, David Teigland wrote:
> > > On Tue, Aug 07, 2007 at 10:26:06AM +0100, Steven Whitehouse wrote:
> > > >  		error = gfs2_glock_nq_num(sdp, sdp->sd_lockstruct.ls_jid,
> > > > -					  &gfs2_journal_glops,
> > > > +					  &gfs2_inode_glops,
> > > >  					  LM_ST_EXCLUSIVE, LM_FLAG_NOEXP,
> > > >  					  &sdp->sd_journal_gh);
> > > 
> > > The lock for journal N is now the same as the lock for the inode at
> > > block N -- that doesn't work.
> > > 
> > 
> > Yes, thats the point of the patch, after all journals are now inodes. If
> > we don't do that then two things happen:
> > 
> >  - Accesses of the journal inodes via the meta fs will not be "in sync"
> > with what the kernel sees
> >  - More importantly, and the reason why this patch was created is that
> > upon recovery of a journal, the cache wasn't being flushed resulting in
> > the blocks being sometimes "seen" again the next time a node recovers
> > the same remote journal.
> 
> The journal with id N does not live in the inode at fs block N.  So,
> you're still using two different locks for the journal, and one of them
> now collides with something else.
> 
I see now! Well spotted, and I'll send a new patch shortly to use the
inode's own lock (which is available in both places anyway),

Steve.



[Date Prev][Date Next]   [Thread Prev][Thread Next]   [Thread Index] [Date Index] [Author Index]