[Linux-cluster] Cluster maintenance mode ?

[David Teigland]

On Wed, Mar 28, 2007 at 03:59:56PM +0200, Marc Grimme wrote:
> On Wednesday 28 March 2007 15:38:10 you wrote:
> > On Wed, Mar 28, 2007 at 10:24:37AM +0200, Marc Grimme wrote:
> > > I.e. is it possible to detect if the fs was not cleanly unmounted?
> > > Are there any possibilities?
> >
> > I believe fsck checks whether the journal was shut down cleanly (part of
> > unmount) and cleans it up if it wasn't.  Would you like an option for fsck
> > to quit if it finds the journal was cleanly shut down by unmount?
> That would be an option. 
> 
> But basically it would be great to have something to detect a "corrupt" 
> filesystem or at least to check if the filesystem was not cleanly unmounted.
> 
> Background is that we had some clusters that crashed because of corrupt 
> gfs-filesystems. But we couldn't see from the dump that those filesystems 
> were corrupt and needed to be fixed.

Ah, I see what you mean -- there's no way to tell that a corrupted fs is
corrupted (apart from mounting it and eventually tripping over the
corruption again.)  So, when a node discovers that the fs is corrupted, it
could set some kind of CORRUPTED flag on the fs.  Mount could then check
for this flag and return an error.  Or we could possibly automate
something like:

  if CORRUPTED flag is set
    try to activate lv exclusively
    if that works, start fsck
  else if CORRUPTED flag is not set
    mount fs

It'll take some thought to figure out how to properly flag an fs when gfs
detects corruption.

Dave