Re: [PATCH] NFS: only invalidate dentrys that are clearly invalid.

[Trond Myklebust <trondmy@hammerspace.com>]

On Mon, 2020-11-16 at 13:59 +1100, NeilBrown wrote:
> 
> Prior to commit 5ceb9d7fdaaf ("NFS: Refactor
> nfs_lookup_revalidate()")
> and error from nfs_lookup_verify_inode() other than -ESTALE would
> result
> in nfs_lookup_revalidate() returning that error code (-ESTALE is
> mapped
> to zero).
> Since that commit, all errors result in zero being returned.
> 
> When nfs_lookup_revalidate() returns zero, the dentry is invalidated
> and, significantly, if the dentry is a directory that is mounted on,
> that mountpoint is lost.
> 
> If you:
>  - mount an NFS filesystem which contains a directory
>  - mount something (e.g. tmpfs) on that directory
>  - use iptables (or scissors) to block traffic to the server
>  - ls -l the-mounted-on-directory
>  - interrupt the 'ls -l'
> you will find that the directory has been unmounted.
> 
> This can be fixed by returning the actual error code from
> nfs_lookup_verify_inode() rather then zero (except for -ESTALE).
> 
> Fixes: 5ceb9d7fdaaf ("NFS: Refactor nfs_lookup_revalidate()")
> Signed-off-by: NeilBrown <neilb@suse.de>
> ---
>  fs/nfs/dir.c | 8 +++++---
>  1 file changed, 5 insertions(+), 3 deletions(-)
> 
> diff --git a/fs/nfs/dir.c b/fs/nfs/dir.c
> index cb52db9a0cfb..d24acf556e9e 100644
> --- a/fs/nfs/dir.c
> +++ b/fs/nfs/dir.c
> @@ -1350,7 +1350,7 @@ nfs_do_lookup_revalidate(struct inode *dir,
> struct dentry *dentry,
>                          unsigned int flags)
>  {
>         struct inode *inode;
> -       int error;
> +       int error = 0;
>  
>         nfs_inc_stats(dir, NFSIOS_DENTRYREVALIDATE);
>         inode = d_inode(dentry);
> @@ -1372,8 +1372,10 @@ nfs_do_lookup_revalidate(struct inode *dir,
> struct dentry *dentry,
>             nfs_check_verifier(dir, dentry, flags & LOOKUP_RCU)) {
>                 error = nfs_lookup_verify_inode(inode, flags);
>                 if (error) {
> -                       if (error == -ESTALE)
> +                       if (error == -ESTALE) {
>                                 nfs_zap_caches(dir);
> +                               error = 0;
> +                       }
>                         goto out_bad;
>                 }
>                 nfs_advise_use_readdirplus(dir);
> @@ -1395,7 +1397,7 @@ nfs_do_lookup_revalidate(struct inode *dir,
> struct dentry *dentry,
>  out_bad:
>         if (flags & LOOKUP_RCU)
>                 return -ECHILD;
> -       return nfs_lookup_revalidate_done(dir, dentry, inode, 0);
> +       return nfs_lookup_revalidate_done(dir, dentry, inode, error);

Which errors do we actually need to return here? As far as I can tell,
the only errors that nfs_lookup_verify_inode() is supposed to return is
ENOMEM, ESTALE, ECHILD, and possibly EIO or ETiMEDOUT.

Why would it be better to return those errors rather than just a 0 when
we need to invalidate the inode, particularly since we already have a
special case in nfs_lookup_revalidate_done() when the dentry is root?

>  }
>  
>  static int

-- 
Trond Myklebust
Linux NFS client maintainer, Hammerspace
trond.myklebust@hammerspace.com