Re: [PATCH] nfs: fix bdi_unregister() before sb kill

[Jens Axboe <jens.axboe@oracle.com>]

On Fri, Sep 18 2009, Trond Myklebust wrote:
> On Fri, 2009-09-18 at 17:02 +0200, Jens Axboe wrote:
> > On Fri, Sep 18 2009, Jens Axboe wrote:
> > > On Thu, Sep 17 2009, Trond Myklebust wrote:
> > > > On Thu, 2009-09-17 at 21:40 +0200, Jens Axboe wrote:
> > > > > On Thu, Sep 17 2009, Jens Axboe wrote:
> > > > > > Hi,
> > > > > > 
> > > > > > This can cause a hang on NFS umount, since the bdi gets pruned before we
> > > > > > flush any pending dirty IO. Peter, can you check whether it fixes your
> > > > > > issue?
> > > > > 
> > > > > There's another problem with NFS && backing devices. NFS may call
> > > > > bdi_destroy() on a bdi without ever called bdi_init(). This was always a
> > > > > bad idea, now it's an issue.
> > > > > 
> > > > > So, Trond, can I safely add a server->flags private flag to indicate
> > > > > that we have called bdi_init()? Then nfs_free_server() knows when to
> > > > > call bdi_destroy(). Seems like the safest fix, since error handling is
> > > > > currently 'just call nfs_free_server()'.
> > > > 
> > > > Urgh... Is there any reason why we can't just move the call to
> > > > bdi_init() into nfs_bdi_register()? It seems bizarre to have to
> > > > initialise the backing_dev_info twice like this...
> > > 
> > > No reason at all, I don't know why it was implemented that way
> > > originally.
> > > 
> > > > If we do that, then we can just look at the BDI_registered state flag in
> > > > order to figure out if we need to call bdi_unregister()
> > > 
> > > That's not exactly pretty either, diving into internal bdi details to
> > > find out if we did an init/register of the device. BDI_registered is
> > > just a debug flag, it may even go away again shortly.
> > 
> > Trond, we need to make some sort of decision on this very shortly. It
> > definitely needs to be fixed for -rc1, it's causing NFS oopses. So lets
> > make some sort of call on this and get it added, then you/we/I can
> > always pretty it up later.
> > 
> 
> OK... I think the solution is to move the call to bdi_init() into
> nfs_bdi_register(), then move the calls to nfs_bdi_register() into
> nfs_set_super() (with an appropriate call to bdi_destroy() if
> set_anon_super() fails).
> 
> Then we can put bdi_destroy() in place of the call to bdi_unregister()
> in nfs_kill_super().

Yeah, that sounds cleaner.

> I'm not going to attempt a patch, since I don't have a copy of your
> current tree to base it on, but does the above make sense to you?

My current tree in the NFS area is just a one liner to move the
bdi_destroy() in nfs_kill_super(). I'll try and cut a patch later
tonight.

-- 
Jens Axboe