Re: [v3] ceph: if we are blacklisted, __do_request returns directly

[Jeff Layton]

On Tue, 2020-04-21 at 10:13 +0800, Yanhu Cao wrote:
> On Mon, Apr 20, 2020 at 8:16 PM Jeff Layton <jlayton@xxxxxxxxxx> wrote:
> > On Fri, 2020-04-17 at 19:07 +0800, Yanhu Cao wrote:
> > > If we mount cephfs by the recover_session option,
> > > __do_request can return directly until the client automatically reconnects.
> > > 
> > > Signed-off-by: Yanhu Cao <gmayyyha@xxxxxxxxx>
> > > ---
> > >  fs/ceph/mds_client.c | 6 ++++++
> > >  1 file changed, 6 insertions(+)
> > > 
> > > diff --git a/fs/ceph/mds_client.c b/fs/ceph/mds_client.c
> > > index 486f91f9685b..16ac5e5f7f79 100644
> > > --- a/fs/ceph/mds_client.c
> > > +++ b/fs/ceph/mds_client.c
> > > @@ -2708,6 +2708,12 @@ static void __do_request(struct ceph_mds_client *mdsc,
> > > 
> > >       put_request_session(req);
> > > 
> > > +     if (mdsc->fsc->blacklisted &&
> > > +         ceph_test_mount_opt(mdsc->fsc, CLEANRECOVER)) {
> > > +             err = -EBLACKLISTED;
> > > +             goto finish;
> > > +     }
> > > +
> > 
> > Why check for CLEANRECOVER? If we're mounted with recover_session=no
> > wouldn't we want to do the same thing here?
> > 
> > Either way, it's still blacklisted. The only difference is that it won't
> > attempt to automatically recover the session that way.
> 
> I think mds will clear the blacklist. In addition to loading cephfs
> via recover_session=clean, I didn't find a location where
> fsc->blacklisted is set to false. If the client has been blacklisted,
> should it always be blacklisted (fsc->blacklisted=true)? Or is there
> another way to set fsc->blacklised to false?
> 

Basically, this patch is just changing it so that when the client is
blacklisted and the mount is done with recover_session=clean, we'll
shortcut the rest of the __do_request and just return -EBLACKLISTED.

My question is: why do we need to test for recover_session=clean here? 

If the client _knows_ that it is blacklisted, why would it want to
continue with __do_request in the recover_session=no case? Would it make
more sense to always return early in __do_request when the client is
blacklisted?


> > 
> > >       mds = __choose_mds(mdsc, req, &random);
> > >       if (mds < 0 ||
> > >           ceph_mdsmap_get_state(mdsc->mdsmap, mds) < CEPH_MDS_STATE_ACTIVE) {
> > --
> > Jeff Layton <jlayton@xxxxxxxxxx>
> > 

-- 
Jeff Layton <jlayton@xxxxxxxxxx>