Re: NFS/d_splice_alias breakage

From: Al Viro
Date: Fri Jun 03 2016 - 00:53:22 EST

On Fri, Jun 03, 2016 at 05:42:53AM +0100, Al Viro wrote:
> On Fri, Jun 03, 2016 at 05:26:48AM +0100, Al Viro wrote:
> > Looks like the right thing to do would be to do d_drop() at no_open:,
> > just before we call nfs_lookup(). If not earlier, actually... How
> > about the following?
> A bit of rationale: dentry in question is negative and attempt to open
> it has just failed; in case it's really negative we did that d_drop()
> anyway (and then unconditionally rehashed it). In case when we proceed to
> nfs_lookup() and it does not fail, we'll have it rehashed there (with the
> right inode). What do we lose from doing d_drop() on *any* error here?
> It's negative, with dubious validity. In the worst case, we had open
> and lookup fail, but it's just us - the sucker really is negative and
> somebody else would be able to revalidate it. If we drop it here (and
> not rehash later), that somebody else will have to allocate an in-core
> dentry before doing lookup or atomic_open. Which is negligible extra
> burden.

I suspect that it got broken in commit 275bb3078 (NFSv4: Move dentry
instantiation into the NFSv4-specific atomic open code). Prior to that
commit, d_drop() was there (error or no error). Looks like 3.10+, indeed.
I agree that we shouldn't drop it on successful open, but it needs to be
done on errors. All of them, not just ENOENT, as in that commit.