Re: [PATCH v2 0/2] cephfs: prepare for changes to VFS locking
From: NeilBrown
Date: Tue Nov 17 2026 - 14:58:06 EST
On Tue, 10 Nov 2026, Xiubo Li wrote:
> On Tue, Nov 10, 2026 at 07:20:23AM +1100, NeilBrown wrote:
> > @@ -2230,7 +2226,7 @@ int ceph_readdir_prepopulate(struct ceph_mds_request *req,
> > if (inode_state_read_once(in) & I_NEW)
> > unlock_new_inode(in);
> >
> > - if (d_really_is_negative(dn)) {
> > + if (d_in_lookup(dn) || d_really_is_negative(dn)) {
>
> Nit: An in-lookup dentry is always negative: __d_instantiate() warns on one,
> and __d_add() ends the lookup before it sets the inode. Here dn can only
> be in-lookup if d_alloc_trylock() has just allocated it for us, so
> d_really_is_negative(dn) already covers that case and the old test could
> stay as it was. Feel free to keep it if you prefer it spelled out.
I think I would prefer to keep the d_in_lookup() test but it isn't a
strong preference. In my mind an in-lookup dentry is neither positive or
negtive yet, though the d_inode is certainly NULL.
>
> Otherwise this looks good to me.
>
> Today ceph_readdir() runs with i_rwsem held exclusive (WRAP_DIR_ITER),
> so the -EWOULDBLOCK branch cannot be reached yet. To exercise it, I ran
> a test-only build that drops the parent's shared lock in ceph_lookup()
> and sleeps 300ms while the dentry is in-lookup, which is what the lookup
> side will look like once d_alloc_parallel() moves above i_rwsem. In
> that build, readdir replies met the in-lookup dentries
> (d_alloc_trylock() returned -EWOULDBLOCK 64 times over 4 rounds), and
> afterwards:
> - the first and the second (cached) listing both showed every name,
> - every lookup returned the inode the other client sees,
> - there were no duplicate hashed dentries in the directory (checked
> with drgn),
> - the MDS cap count matched,
> - KASAN and lockdep stayed quiet.
Thanks for this testing - I really appreciate it.
Thanks,
NeilBrown
>
> Reviewed-by: Xiubo Li <xiubo.li@xxxxxxxxx>
> Tested-by: Xiubo Li <xiubo.li@xxxxxxxxx>
>
> Thanks
> Xiubo
>
> NeilBrown <neilb@xxxxxxxxxxx> 于2026年11月9日周一 21:25写道:
> >
> > This v2 drops 3/3 which was deprecating d_add() - there is no rush for
> > that so I'm not going to push it at present.
> > It also fixes the first patch to increment "skipped" when d_alloc_trylock()
> > returns -EWOULDBLOCK.
> >
> > The VFS patches that this series depends on are upstream, so this can
> > land in any tree.
> >
> > Thanks,
> > NeilBrown
> >
> > original intro:
> > I am working on changes to directory locking. The medium term goal is
> > to lift d_alloc_parallel() locking above i_rwsem, so d_alloc_paralle()
> > can be run unlocked, and consequently cannot be called when ->i_rwsem is
> > held.
> >
> > The requires:
> > 1/ that d_alloc_parallel() not be called while i_rwsem is held and also
> > that d_alloc()/d_splice_alias() not be used (unless the filesystems
> > doesn't support creating objects).
> > cephfs uses d_alloc()/d_splice_alias() when priming the dcache from
> > the results of readdir. A new interface d_alloc_trylock()
> > is now available and can be used instead.
> >
> > 2/ a dentry must not be d_drop()ed while an operation is ongoing, as an
> > unlocked d_alloc_parallel() would then be able to create a new
> > dentry with the same name. cephfs calls d_drop() in couple of places
> > before d_splice_alias() but this is no longer needed as d_splice_alias()
> > handles spliced aliases now.
> >
> >
> > [PATCH v2 1/2] cephfs: Use d_alloc_trylock() in
> > [PATCH v2 2/2] cephfs: Don't d_drop() before d_splice_alias()
> >
>