On Mon, Nov 16 2020, Trond Myklebust wrote:

> On Mon, 2020-11-16 at 15:43 +1100, NeilBrown wrote:
>> On Mon, Nov 16 2020, Trond Myklebust wrote:
>> 
>> > 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?
>> 
>> ERESTARTSYS is the error that easily causes problems.
>> 
>> Returning 0 causes d_invalidate() to be called which is quite heavy
>> handed in mountpoints.
>
> My point is that it shouldn't get returned for mountpoints. See
> nfs_lookup_revalidate_done().

nfs_lookup_revalidate_done() only checks IS_ROOT(), and while many
mountpoints are IS_ROOT(), not all are (--bind easily makes others).

But that isn't even really relevant here.  The dentry being revalidated
is the underlying directory - that something else is mounted on.
step_into() which follows mount points is called in walk_component()
*after* lookup_fast or lookup_slow which will have revalidated the
dentry.

NeilBrown


>
>> So it is only reasonable to return 0 when we have unambiguous
>> confirmation from the server that the object no longer exists. 
>> ESTALE
>> is unambiguous. EIO might be unambiguous.  ERESTARTSYS, ENOMEM,
>> ETIMEDOUT are transient and don't justify d_invalidate() being
>> called.
>> 
>> (BTW, Commit cc89684c9a26 ("NFS: only invalidate dentrys that are
>> clearly invalid.")
>>  fixed much the same bug 3 years ago).
>>  
>> Thanks,
>> NeilBrown
>> 
>> 
>> > 
>> > >  }
>> > >  
>> > >  static int
>> > 
>> > -- 
>> > Trond Myklebust
>> > Linux NFS client maintainer, Hammerspace
>> > trond.myklebust@hammerspace.com
>
> -- 
> Trond Myklebust
> CTO, Hammerspace Inc
> 4984 El Camino Real, Suite 208
> Los Altos, CA 94022
> ​
> www.hammer.space