On Thu, 21 Feb 2013 09:32:25 +1100
NeilBrown <neilb@xxxxxxx> wrote:
> On Wed, 20 Feb 2013 11:19:05 -0500 Jeff Layton <jlayton@xxxxxxxxxx> wrote:
>
> > The following set of operations on a NFS client and server will cause
> >
> > server# mkdir a
> > client# cd a
> > server# mv a a.bak
> > client# sleep 30 # (or whatever the dir attrcache timeout is)
> > client# stat .
> > stat: cannot stat ‘.’: Stale NFS file handle
> >
> > Obviously, we should not be getting an ESTALE error back there since the
> > inode still exists on the server. The problem is that the lookup code
> > will call d_revalidate on the dentry that "." refers to, because NFS has
> > FS_REVAL_DOT set.
> >
> > nfs_lookup_revalidate will see that the parent directory has changed and
> > will try to reverify the dentry by redoing a LOOKUP. That of course
> > fails, so the lookup code returns ESTALE.
> >
> > The problem here is that d_revalidate is really a bad fit for this case.
> > What we really want to know at this point is whether the inode is still
> > good or not, but we don't really care what name it goes by or whether
> > the dcache is still valid.
> >
> > Add a new d_op->d_reval_jumped operation and have complete_walk call
> > that instead of d_revalidate. The intent there is to allow for a
> > "weaker" d_revalidate that just checks to see whether the inode is still
> > good. This is also gives us an opportunity to kill off the FS_REVAL_DOT
> > special casing.
> >
> > In a perfect world, this would be a new inode operation instead, but
> > I don't see a way to cleanly handle that for 9p, which needs a
> > dentry in order to get a fid.
>
> The earlier i_op->revalidate inode operation took a 'dentry', not an inode.
> If you look at struct inode_operations, you will see that 8 of them take a
> dentry as their first argument.
>
> Never the less, I would leave it in dentry_operations. It makes it easier to
> use the DCACHE_OP_ optimisation.
>
Good point. I guess my thinking was that we aren't really interested in
the dentry, per-se. But for some filesystems, having the dentry may
make this easier to deal with.
>
> >
> > Cc: NeilBrown <neilb@xxxxxxx>
> > Signed-off-by: Jeff Layton <jlayton@xxxxxxxxxx>
> > ---
> > Documentation/filesystems/Locking | 2 ++
> > Documentation/filesystems/vfs.txt | 32 ++++++++++++++++++++++++++--
> > fs/9p/vfs_dentry.c | 1 +
> > fs/9p/vfs_super.c | 2 +-
> > fs/dcache.c | 3 +++
> > fs/namei.c | 8 ++-----
> > fs/nfs/dir.c | 45 +++++++++++++++++++++++++++++++++++++++
> > fs/nfs/nfs4super.c | 6 +++---
> > fs/nfs/super.c | 6 +++---
> > include/linux/dcache.h | 3 +++
> > include/linux/fs.h | 1 -
> > 11 files changed, 93 insertions(+), 16 deletions(-)
> >
> > diff --git a/Documentation/filesystems/Locking b/Documentation/filesystems/Locking
> > index f48e0c6..9718b667 100644
> > --- a/Documentation/filesystems/Locking
> > +++ b/Documentation/filesystems/Locking
> > @@ -10,6 +10,7 @@ be able to use diff(1).
> > --------------------------- dentry_operations --------------------------
> > prototypes:
> > int (*d_revalidate)(struct dentry *, unsigned int);
> > + int (*d_reval_jumped)(struct dentry *, unsigned int);
> > int (*d_hash)(const struct dentry *, const struct inode *,
> > struct qstr *);
> > int (*d_compare)(const struct dentry *, const struct inode *,
>
> I cannot get excited about the name "d_reval_jumped" .... though once you
> read the explanation in the doco (thanks for that) it makes sense. I guess
> I'll get used to it.
>
Me neither. I think Al mentioned that he's renamed this to
"d_weak_revalidate" in his tree. Neither name really does it for me,
so I'm open to suggestions.
> > /*
> > + * A weaker form of d_revalidate for revalidating just the dentry->d_inode
> > + * when we don't really care about the dentry name. This is called when a
> > + * pathwalk ends on a dentry that was not found via a normal lookup in the
> > + * parent dir (e.g.: ".", "..", procfs symlinks or mountpoint traversals).
> > + *
> > + * In this situation, we just want to verify that the inode itself is OK
> > + * since the dentry might have changed on the server.
> > + */
> > +static int nfs_reval_jumped(struct dentry *dentry, unsigned int flags)
> > +{
> > + int error;
> > + struct inode *inode = dentry->d_inode;
> > +
> > + if (flags & LOOKUP_RCU)
> > + return -ECHILD;
> > +
> > + /*
> > + * I believe we can only get a negative dentry here in the case of a
> > + * procfs-style symlink. Just assume it's correct for now, but we may
> > + * eventually need to do something more here.
> > + */
> > + if (!inode) {
> > + dfprintk(LOOKUPCACHE, "%s: %s/%s has negative inode\n",
> > + __func__, dentry->d_parent->d_name.name,
> > + dentry->d_name.name);
> > + return 1;
> > + }
> > +
> > + if (is_bad_inode(inode)) {
> > + dfprintk(LOOKUPCACHE, "%s: %s/%s has dud inode\n",
> > + __func__, dentry->d_parent->d_name.name,
> > + dentry->d_name.name);
> > + return 0;
> > + }
> > +
> > + error = nfs_revalidate_inode(NFS_SERVER(inode), inode);
> > + dfprintk(LOOKUPCACHE, "NFS: %s: inode %lu is %s\n",
> > + __func__, inode->i_ino, error ? "invalid" : "valid");
> > + if (error)
> > + return 0;
> > + return 1;
> > +}
>
> I wonder if we can delay the "-ECHILD" return a bit.
> Leaving it to after the first two tests should be safe, but doesn't gain us
> anything.
>
> Open-coding the nfs_revalidate_inode as:
> if (!(NFS_I(inode)->cache_validity & NFS_INO_INVALID_ATTR)
> && !nfs_attribute_cache_expired(inode))
> return NFS_STALE(inode) ? 0 : 1;
> error = __nfs_revalidate_inode(server, inode);
>
> and then inserting the -ECHILD code in before the __nfs_revalidate_inode
> should be safe, and means we still benefit from the RCU path in the common
> case.
> Of course, for that to be really useful, nfs_lookup_revalidate would need to
> be changed to only return -ECHILD if it really needed to block, and maybe
> that is too hard, or at least is a job for another day.
>
> Otherwise, looks good - thanks.
>
> Reviewed-by: NeilBrown <neilb@xxxxxxx>
>
>
I don't know that much about rcuwalk mode, but the vfs.txt doc says
this:
If in rcu-walk mode, the filesystem must revalidate the dentry
without blocking or storing to the dentry, d_parent and d_inode
should not be used without care (because they can change and,
in d_inode case, even become NULL under us).
If we assume that d_inode does become NULL after we set the "inode"
pointer, do we still hold a reference to it? Or do we need to ensure
that we take one when we set that pointer?
Also, since this is the last component of the path, I suspect that
we're almost never going to be in rcu-walk mode here, right?
In any case, I think we ought to do that sort of optimization
separately on top of this patch. We probably ought to consider similar
optimization in the d_revalidate routines too. I think we might get
even more gain there anyway.
--
Jeff Layton <jlayton@xxxxxxxxxx>
Attachment:
signature.asc
Description: PGP signature
