Re: readdir cookies

["J. Bruce Fields" <bfields@fieldses.org>]

On Mon, Sep 17, 2012 at 06:03:26PM +0000, Myklebust, Trond wrote:
> > -----Original Message-----
> > From: J. Bruce Fields [mailto:bfields@fieldses.org]
> > Sent: Monday, September 17, 2012 1:50 PM
> > To: Myklebust, Trond
> > Cc: Schumaker, Bryan; linux-nfs@vger.kernel.org
> > Subject: Re: readdir cookies
> > 
> > On Mon, Sep 17, 2012 at 04:56:10PM +0000, Myklebust, Trond wrote:
> > > > -----Original Message-----
> > > > From: J. Bruce Fields [mailto:bfields@fieldses.org]
> > > > Sent: Monday, September 17, 2012 12:51 PM
> > > > To: Myklebust, Trond
> > > > Cc: Schumaker, Bryan; linux-nfs@vger.kernel.org
> > > > Subject: Re: readdir cookies
> > > >
> > > > On Mon, Sep 17, 2012 at 04:05:10PM +0000, Myklebust, Trond wrote:
> > > > > > -----Original Message-----
> > > > > > From: linux-nfs-owner@vger.kernel.org [mailto:linux-nfs-
> > > > > > owner@vger.kernel.org] On Behalf Of J. Bruce Fields
> > > > > > Sent: Monday, September 17, 2012 11:54 AM
> > > > > > To: Myklebust, Trond; Schumaker, Bryan
> > > > > > Cc: linux-nfs@vger.kernel.org
> > > > > > Subject: readdir cookies
> > > > > >
> > > > > > How does the nfs client decide what directory cookies to return?
> > > > > > (As the d_off field to readdir, for example).
> > > > > >
> > > > > > I'd assumed it would return the server's cookie, but testing
> > > > > > shows it returns a simple integer sequence.
> > > > >
> > > > > We cache the cookies and use them to figure out where the readdir
> > > > > cursor
> > > > is after a directory update, but we use positive integers for
> > telldir()/seekdir().
> > > > The reason is that too many servers return cookies that use > 32
> > > > bits (NFSv2-
> > > > incompatible) or are unsigned.
> > > >
> > > > Ok--so the tradeoff is that telldir/seekdir cookies may not be good
> > > > indefinitely?
> > >
> > > Yes. Most apps don't care about telldir()/seekdir(), so it isn't really much of
> > an issue. However, glibc has a bunch of assumptions about the 'd_offs'
> > entries returned by getdents()/getdents64() and will do "sanity checks" on
> > them that often end up failing when we just shove in the raw NFS cookies.
> > 
> > OK, thanks.  When did the client start doing that?
> 
> Years ago. You want to look at commit 00a2642 (NFS: Hide NFS server-generated readdir cookies from userland) which dates back to Linux-2.6.13.
>  
> > (And do you remember any details about the libc problems?  ext4 is
> > returning 64-bit cookies now in some cases, and we've seen a few problems
> > elsewhere, but none with glibc.)
> 
> IIRC: At one point, in glibc-2(.1?), they started using getdents64() instead of getdents() in order to avoid the problem of mismatched 'struct dirent' sizes on 32-bit ix86 platforms. Whenever the getdents64() interface returned a 64-bit cookie (or a 32-bit negative cookie), glibc would generate an EOVERFLOW error.

Got it, thanks.

--b.