Re: [PATCH 4/4] nfsd: Pin to vfsmnt instead of mntget

From: "J. Bruce Fields" <bfields@fieldses.org>
To: Kinglong Mee <kinglongmee@gmail.com>
Cc: NeilBrown <neilb@suse.de>,
	linux-fsdevel@vger.kernel.org,
	"linux-nfs@vger.kernel.org" <linux-nfs@vger.kernel.org>,
	Al Viro <viro@zeniv.linux.org.uk>,
	Trond Myklebust <trond.myklebust@primarydata.com>
Subject: Re: [PATCH 4/4] nfsd: Pin to vfsmnt instead of mntget
Date: Fri, 22 May 2015 12:03:52 -0400	[thread overview]
Message-ID: <20150522160352.GC5580@fieldses.org> (raw)
In-Reply-To: <555F4501.2090806@gmail.com>

On Fri, May 22, 2015 at 11:02:25PM +0800, Kinglong Mee wrote:
> On 5/16/2015 7:23 AM, NeilBrown wrote:
> > On Fri, 15 May 2015 17:11:34 -0400 "J. Bruce Fields" <bfields@fieldses.org>
> > wrote:
> > 
> >> On Wed, May 13, 2015 at 02:25:15PM +1000, NeilBrown wrote:
> >>> On Mon, 11 May 2015 21:08:47 +0800 Kinglong Mee <kinglongmee@gmail.com> wrote:
> >>>
> >>>> On 5/8/2015 9:47 PM, J. Bruce Fields wrote:
> >>>>> It could also be useful to have the ability to force an unmount even in
> >>>>> the presence of locks.  That's not a safe default, but an
> >>>>> "allow_force_unmount" export option might be useful.
> >>>
> >>> We already have a mechanism to forcibly drop any locks by writing some magic
> >>> to /proc/fs/nfsd/unlock_{ip,filesystem}.  I don't think we need any more.
> >>
> >> Yeah, I remember thinking this sort of approach would have advantages,
> >> maybe I was wrong, I need to revisit it.
> >>
> >> The unlock_{ip,filesystem} approach requires temporarily shutting down
> >> mountd, doesn't it?
> > 
> > Not necessarily.
> > It does require ensuring that new locks aren't suddenly taken though.
> > 
> > I imagine an early step in the migration process is to "ifconfig down" the
> > virtual interface with the floating ID.  Then you can safely "unlock" and
> > unmount any filesystems are that only accessed via the IP.
> > 
> > But you are right that using the "unlock_*" interface and then unmounting is
> > racy in a way that we are trying to make "unmount" not racy.  So maybe an
> > "allow_force_unmount" would have a place.
> 
> No, unlock_{ip,filesystem} are used for nlmlock, doesn't support nfsv4 resources.

I still prefer the "allow_force_unmount" option, but maybe we should
also fix unlock_{ip,filesystem} to deal with nfsv4.  (Though I think
it's a little less well-defined there due to the possibility of
trunking.)

> Some other interfaces under /sys/kernel/debug/nfsd/forget_* support nfsv4 resources,
> without for an filesystem. It seems will be removed sometime.

We definitely don't want people to depend on those for anything other
than testing clients.

I don't think they'd be practical for this use.  forget_client comes the
closest, but you'd have to figure out the ip address of every client you
want to forget.  If there's a risk people might try to really use that,
then maybe we should go for scarier warnings and/or remove that
particular interface.

--b.