MIME-Version: 1.0
In-Reply-To: <4FA345DA4F4AE44899BD2B03EEEC2FA9286B511E@sacexcmbx05-prd.hq.netapp.com>
References: <20130305174954.GG12795@htj.dyndns.org>
	<20130305140312.243cb094@tlielax.poochiereds.net>
	<20130305190923.GI12795@htj.dyndns.org>
	<20130305183941.19ff39ce@tlielax.poochiereds.net>
	<20130305234700.GE1227@htj.dyndns.org>
	<20130306181608.GA18687@redhat.com>
	<20130306185304.GM1227@htj.dyndns.org>
	<CA+55aFwDogteVd=vwGHXDSASnga-nZZnKaQz9aO1yBU2CPKSbA@mail.gmail.com>
	<20130306212452.GO1227@htj.dyndns.org>
	<CA+55aFy49AWptMcvEzJxg95ikf-dC+-vcCPq-cnaCCwje3tyoQ@mail.gmail.com>
	<20130306213636.GP1227@htj.dyndns.org>
	<20130307064140.71c0936b@tlielax.poochiereds.net>
	<CA+55aFxe1_x8wvVYciiX8tuee+VsmXGKqEbONQdFM4s4W2iuPg@mail.gmail.com>
	<4FA345DA4F4AE44899BD2B03EEEC2FA9286B511E@sacexcmbx05-prd.hq.netapp.com>
Date: Thu, 7 Mar 2013 08:25:10 -0800
Message-ID: <CA+55aFySyBP8q5AvPrt-s_ZPdgamfoX-dERw4EJ21EcsDLBjYA@mail.gmail.com>
Subject: Re: LOCKDEP: 3.9-rc1: mount.nfs/4272 still has locks held!
From: Linus Torvalds <torvalds@linux-foundation.org>
To: "Myklebust, Trond" <Trond.Myklebust@netapp.com>
Cc: Jeff Layton <jlayton@redhat.com>, Tejun Heo <tj@kernel.org>,
        Oleg Nesterov <oleg@redhat.com>,
        Mandeep Singh Baines <msb@chromium.org>,
        Ming Lei <ming.lei@canonical.com>,
        "J. Bruce Fields" <bfields@fieldses.org>,
        Linux Kernel Mailing List <linux-kernel@vger.kernel.org>,
        "linux-nfs@vger.kernel.org" <linux-nfs@vger.kernel.org>,
        "Rafael J. Wysocki" <rjw@sisk.pl>,
        Andrew Morton <akpm@linux-foundation.org>,
        Ingo Molnar <mingo@redhat.com>, Al Viro <viro@zeniv.linux.org.uk>
Content-Type: text/plain; charset=UTF-8
Sender: linux-nfs-owner@vger.kernel.org

On Thu, Mar 7, 2013 at 7:59 AM, Myklebust, Trond
<Trond.Myklebust@netapp.com> wrote:
>
> It _shouldn't_ be an interruption unless the filesystem can't make
> progress.

So how can we tell? Calling "freezable_schedule()" if you're not ready
to be frozen is not good. And nobody but the NFS code can know.

You might want to introduce some counter that counts number of
outstanding non-interruptible events, and only call the "freezable"
version if that counter is zero.

A better alternative might be to *never* call the freezable version.
Because those freezable_*() things are really quite disgusting, and
are wrong - they don't actually freeze the process, they say "I don't
care if you freeze me while I sleep", and you might actually wake up
*while* the system is being frozen. I think the whole concept is
broken. Rafaei - comments? The function really is crap, regardless of
any unrelated NFS problems.

So what NFS could do instead is actually check the "do I need to
freeze" flag, and if you need to freeze you consider it an abort - and
do *not* try to continue. Just freeze, and then act as if the machine
got rebooted as far as NFS was concerned. That should work anyway, no?

That does sound a lot more complex, though.

                 Linus