On Wed, Jan 26, 2022 at 5:34 PM Jeff Layton <jlayton@xxxxxxxxxx> wrote:
>
> On Wed, 2022-01-26 at 17:23 +0100, Ilya Dryomov wrote:
> > On Tue, Jan 25, 2022 at 10:10 PM Jeff Layton <jlayton@xxxxxxxxxx> wrote:
> > >
> > > Dan reported that he was unable to write to files that had been
> > > asynchronously created when the client's OSD caps are restricted to a
> > > particular namespace.
> > >
> > > The issue is that the layout for the new inode is only partially being
> > > filled. Ensure that we populate the pool_ns_data and pool_ns_len in the
> > > iinfo before calling ceph_fill_inode.
> > >
> > > Reported-by: Dan van der Ster <dan@xxxxxxxxxxxxxx>
> > > Fixes: 9a8d03ca2e2c ("ceph: attempt to do async create when possible")
> > > Signed-off-by: Jeff Layton <jlayton@xxxxxxxxxx>
> > > ---
> > > fs/ceph/file.c | 7 +++++++
> > > 1 file changed, 7 insertions(+)
> > >
> > > diff --git a/fs/ceph/file.c b/fs/ceph/file.c
> > > index cbe4d5a5cde5..efea321ff643 100644
> > > --- a/fs/ceph/file.c
> > > +++ b/fs/ceph/file.c
> > > @@ -599,6 +599,7 @@ static int ceph_finish_async_create(struct inode *dir, struct dentry *dentry,
> > > struct ceph_inode_info *ci = ceph_inode(dir);
> > > struct inode *inode;
> > > struct timespec64 now;
> > > + struct ceph_string *pool_ns;
> > > struct ceph_mds_client *mdsc = ceph_sb_to_mdsc(dir->i_sb);
> > > struct ceph_vino vino = { .ino = req->r_deleg_ino,
> > > .snap = CEPH_NOSNAP };
> > > @@ -648,11 +649,17 @@ static int ceph_finish_async_create(struct inode *dir, struct dentry *dentry,
> > > in.max_size = cpu_to_le64(lo->stripe_unit);
> > >
> > > ceph_file_layout_to_legacy(lo, &in.layout);
> > > + pool_ns = ceph_try_get_string(lo->pool_ns);
> > > + if (pool_ns) {
> > > + iinfo.pool_ns_len = pool_ns->len;
> > > + iinfo.pool_ns_data = pool_ns->str;
> > > + }
> >
> > Considering that we have a reference from try_prep_async_create(), do
> > we actually need to bother with ceph_try_get_string() here?
> >
>
> Technically, no. We could just do a rcu_dereference_protected there
> since we know that lo is private and can't change. Want me to send a v2?
Yeah, let's not do the reference dance when it isn't needed.
I'd probably use rcu_dereference_raw() to avoid having to think about
omitted READ_ONCE (even if the reasoning is trivial), but up to you.
Thanks,
Ilya
