Re: [PATCH v2] Fix i_mutex handling in nfsd readdir

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

On Thu, Apr 23, 2009 at 03:40:23PM +0900, hooanon05@yahoo.co.jp wrote:
> 
> "J. Bruce Fields":
> > > Isn't it better to test it BEFORE fh_compose()?
> 	:::
> > Yes, I think you're right.
> 
> Then here you are.

The nfsv4 readdir callback needs a similar fix.

Also, it looks to me like this results in us encoding an entry for this
deleted file in the readdir reply, but with an empty filehandle.  From a
quick glance at the rfc it's not clear to me whether this is really
legal.  I suspect it may cause odd behavior on clients.  At the least it
would seem cleaner to check for this condition early enough that we can
just skip the entry entirely.

--b.

> 
> J. R. Okajima
> 
> ----------------------------------------------------------------------
> 
> commit c98c6c4a207d602bd9498ea5f1d2993a00e98445
> Author: J. R. Okajima <hooanon05@yahoo.co.jp>
> Date:   Thu Apr 23 15:38:43 2009 +0900
> 
>     NFSD: test d_inode before fh_compose()
>     
>     After 2f9092e1020246168b1309b35e085ecd7ff9ff72 "Fix i_mutex vs. readdir
>     handling in nfsd" (and 14f7dd63 "Copy XFS readdir hack into nfsd code"),
>     an entry may be removed between the first mutex_unlock and the second
>     mutex_lock. In this case, lookup_one_len() in compose_entry_fh() will
>     return a negative dentry.
>     It is better to test inode (positive/negative) BEFORE fh_compose().
>     
>     Signed-off-by: J. R. Okajima <hooanon05@yahoo.co.jp>
> 
> diff --git a/fs/nfsd/nfs3xdr.c b/fs/nfsd/nfs3xdr.c
> index 17d0dd9..1b5543b 100644
> --- a/fs/nfsd/nfs3xdr.c
> +++ b/fs/nfsd/nfs3xdr.c
> @@ -851,8 +851,8 @@ compose_entry_fh(struct nfsd3_readdirres *cd, struct svc_fh *fhp,
>  	if (IS_ERR(dchild))
>  		return 1;
>  	if (d_mountpoint(dchild) ||
> -	    fh_compose(fhp, exp, dchild, &cd->fh) != 0 ||
> -	    !dchild->d_inode)
> +	    !dchild->d_inode ||
> +	    fh_compose(fhp, exp, dchild, &cd->fh) != 0)
>  		rv = 1;
>  	dput(dchild);
>  	return rv;