Date: Mon, 23 Apr 2012 09:12:55 -0400
From: Jeff Layton <jlayton@redhat.com>
To: "J. Bruce Fields" <bfields@fieldses.org>
Cc: Miklos Szeredi <miklos@szeredi.hu>, Malahal Naineni <malahal@us.ibm.com>,
        Steve Dickson <SteveD@redhat.com>, linux-fsdevel@vger.kernel.org,
        linux-nfs@vger.kernel.org, linux-kernel@vger.kernel.org,
        viro@zeniv.linux.org.uk, hch@infradead.org,
        michael.brantley@deshaw.com, sven.breuner@itwm.fraunhofer.de,
        chuck.lever@oracle.com, pstaubach@exagrid.com,
        trond.myklebust@fys.uio.no, rees@umich.edu
Subject: Re: [PATCH RFC v3] vfs: make fstatat retry once on ESTALE errors
 from getattr call
Message-ID: <20120423091255.00f926c4@tlielax.poochiereds.net>
In-Reply-To: <20120423130009.GA13681@fieldses.org>
References: <1334316311-22331-1-git-send-email-jlayton@redhat.com>
	<1334749927-26138-1-git-send-email-jlayton@redhat.com>
	<20120420104055.511e15bc@tlielax.poochiereds.net>
	<4F91C49D.8070908@RedHat.com>
	<20120420203725.GA3512@us.ibm.com>
	<20120420171314.73801874@corrin.poochiereds.net>
	<CAJfpegt40cgMJQQo3JuNaaS1w957Y2a_NxVoyvx3bmTMj1TGOA@mail.gmail.com>
	<20120423080012.7c23ef24@tlielax.poochiereds.net>
	<20120423130009.GA13681@fieldses.org>
Mime-Version: 1.0
Content-Type: text/plain; charset=US-ASCII
Sender: linux-nfs-owner@vger.kernel.org

On Mon, 23 Apr 2012 09:00:09 -0400
"J. Bruce Fields" <bfields@fieldses.org> wrote:

> On Mon, Apr 23, 2012 at 08:00:12AM -0400, Jeff Layton wrote:
> > On Sun, 22 Apr 2012 07:40:57 +0200
> > Miklos Szeredi <miklos@szeredi.hu> wrote:
> > 
> > > On Fri, Apr 20, 2012 at 11:13 PM, Jeff Layton <jlayton@redhat.com> wrote:
> > > > On Fri, 20 Apr 2012 15:37:26 -0500
> > > > Malahal Naineni <malahal@us.ibm.com> wrote:
> > > >
> > > >> Steve Dickson [SteveD@redhat.com] wrote:
> > > >> > > 2) if we assume that it is fairly representative of one, how can we
> > > >> > > achieve retrying indefinitely with NFS, or at least some large finite
> > > >> > > amount?
> > > >> > The amount of looping would be peer speculation. If the problem can
> > > >> > not be handled by one simple retry I would say we simply pass the
> > > >> > error up to the app... Its an application issue...
> > > >>
> > > >> As someone said, ESTALE is an incorrect errno for a path based call.
> > > >> How about turning ESTALE into ENOENT after a retry or few retries?
> > > >>
> > > >
> > > > It's not really the same thing. One could envision an application
> > > > that's repeatedly renaming a new file on top of another one. The file
> > > > is never missing from the namespace of the server, but you could still
> > > > end up getting an ESTALE.
> > > >
> > > > That would break other atomicity guarantees in an even worse way, IMO...
> > > 
> > > For directory operations ESTALE *is* equivalent to ENOENT if already
> > > retrying with LOOKUP_REVAL.  Think about it.  Atomic replacement by
> > > another directory with rename(2) is not an excuse here actually.
> > > Local filesystems too can end up with IS_DEAD directory after lookup
> > > in that case.
> > > 
> > 
> > Doesn't that violate POSIX? rename(2) is supposed to be atomic, and I
> > can't see where there's any exception for that for directories.
> 
> Hm, but that only allows atomic replacement of the last component of a
> path.
> 
> Suppose you're looking up a path, you've so far reached intermediate
> directory "D", and the next step of the lookup (of some entry in D)
> returns ESTALE.  Then either:
> 
> 	- D has since been unlinked, and ENOENT is obviously right.
> 	- D was unlinked and then replaced by something else, in which
> 	  case there was still a moment when ENOENT was correct.
> 	- D was replaced atomically by a rename.  But for the rename to
> 	  work it must have been replacing an empty directory, so there
> 	  was still a moment when ENOENT would have been correct.

I don't think so...D should always exist in the namespace, so ENOENT
would not be correct. Just because it was empty doesn't mean that it
didn't exist...

> 	  (Exception: if D was actually a regular file or some other
> 	  non-directory object, then ENOTDIR would be the right error:
> 	  but if you're able to get at least object type atomically with
> 	  a lookup, then you should have noticed this already on lookup
> 	  of D.)
> 
> I think that's what Miklos meant?
> 
> --b.

Here's an example -- suppose we have two directories: /foo
and /bar. /bar is empty. We call:

    rename("/foo","/bar");

...and at the same time, someone is calling:

    stat("/bar");

...the calls race and in this condition the stat() gets ESTALE back
-- /bar got replaced after we did the lookup.

According to POSIX, the name "/bar" should never be absent from the
namespace in this situation, so I'm not sure I understand why returning
ENOENT here would be acceptable.

-- 
Jeff Layton <jlayton@redhat.com>