Re: [PATCH] scsi_dh_alua: avoid crash during alua_bus_detach()

["Ewan D. Milne" <emilne@redhat.com>]

On Sat, 2020-09-26 at 15:01 -0700, Bart Van Assche wrote:
> On 2020-09-24 03:45, Hannes Reinecke wrote:
> > alua_bus_detach() might be running concurrently with
> > alua_rtpg_work(),
> > so we might trip over h->sdev == NULL and call BUG_ON().
> > The correct way of handling it would be to not set h->sdev to NULL
> > in alua_bus_detach(), and call rcu_synchronize() before the final
> > delete to ensure that all concurrent threads have left the critical
> > section.
> > Then we can get rid of the BUG_ON(), and replace it with a simple
> > if condition.
> > 
> > Cc: Brian Bunker <brian@purestorage.com>
> > Signed-off-by: Hannes Reinecke <hare@suse.de>
> > ---
> >  drivers/scsi/device_handler/scsi_dh_alua.c | 9 +++++----
> >  1 file changed, 5 insertions(+), 4 deletions(-)
> > 
> > diff --git a/drivers/scsi/device_handler/scsi_dh_alua.c
> > b/drivers/scsi/device_handler/scsi_dh_alua.c
> > index f32da0ca529e..308bda2e9c00 100644
> > --- a/drivers/scsi/device_handler/scsi_dh_alua.c
> > +++ b/drivers/scsi/device_handler/scsi_dh_alua.c
> > @@ -658,8 +658,8 @@ static int alua_rtpg(struct scsi_device *sdev,
> > struct alua_port_group *pg)
> >  					rcu_read_lock();
> >  					list_for_each_entry_rcu(h,
> >  						&tmp_pg->dh_list, node)
> > {
> > -						/* h->sdev should
> > always be valid */
> > -						BUG_ON(!h->sdev);
> > +						if (!h->sdev)
> > +							continue;
> >  						h->sdev->access_state =
> > desc[0];
> >  					}
> >  					rcu_read_unlock();
> > @@ -705,7 +705,8 @@ static int alua_rtpg(struct scsi_device *sdev,
> > struct alua_port_group *pg)
> >  			pg->expiry = 0;
> >  			rcu_read_lock();
> >  			list_for_each_entry_rcu(h, &pg->dh_list, node)
> > {
> > -				BUG_ON(!h->sdev);
> > +				if (!h->sdev)
> > +					continue;
> >  				h->sdev->access_state =
> >  					(pg->state &
> > SCSI_ACCESS_STATE_MASK);
> >  				if (pg->pref)
> > @@ -1147,7 +1148,6 @@ static void alua_bus_detach(struct
> > scsi_device *sdev)
> >  	spin_lock(&h->pg_lock);
> >  	pg = rcu_dereference_protected(h->pg, lockdep_is_held(&h-
> > >pg_lock));
> >  	rcu_assign_pointer(h->pg, NULL);
> > -	h->sdev = NULL;
> >  	spin_unlock(&h->pg_lock);
> >  	if (pg) {
> >  		spin_lock_irq(&pg->lock);
> > @@ -1156,6 +1156,7 @@ static void alua_bus_detach(struct
> > scsi_device *sdev)
> >  		kref_put(&pg->kref, release_port_group);
> >  	}
> >  	sdev->handler_data = NULL;
> > +	synchronize_rcu();
> >  	kfree(h);
> >  }
> 
> Hi Hannes,
> 
> Do you agree that the changes in alua_bus_detach() make the changes
> in
> alua_rtpg() superfluous?

I agree that the "if (!h->sdev)   continue;" should not be needed in
alua_rtpg() if the h->sdev remains valid while in the list.

I'm a little concerned about adding the synchronize_rcu() as this is
called in the scsi_device_dev_release_usercontext() path, with a lot
of LUNs it could take a while to remove all the devices, see e.g.:

f983622ae605 scsi: core: Avoid calling synchronize_rcu() for each
               device in scsi_host_block()

It doesn't look like we ever NULL sdev->handler on detach even though
we do a module_put() on the DH.  But we have already called the
release() function so perhaps this doesn't cause a problem in
practice.

-Ewan

> 
> How about freezing command processing for 'sdev' while detaching a
> device handler instead of inserting a synchronize_rcu() call in
> alua_bus_detach()? I'm concerned that the alua_bus_detach() changes
> are
> not sufficient to fix all possible races between detaching a device
> handler and the following code from the SCSI error handler:
> 
> 	if (sdev->handler && sdev->handler->check_sense) {
> 		int rc;
> 
> 		rc = sdev->handler->check_sense(sdev, &sshdr);
> 		if (rc != SCSI_RETURN_NOT_HANDLED)
> 			return rc;
> 		/* handler does not care. Drop down to default handling
> */
> 	}
> 
> Thanks,
> 
> Bart.
>