Re: [PATCH] scsi: scsi_dh_rdac: Avoid crash when a disk attach failed

[Ding Hui]

On 2023/9/6 23:51, Mike Christie wrote:
> On 8/3/23 6:28 AM, Huang Cun wrote:
> > When a disk fails to attach, the struct rdac_dh_data is released,
> > but it is not removed from the ctlr->dh_list. When attaching another
> > disk, the released rdac_dh_data will be accessed and the following
> > BUG_ON() may be observed:
> >
> > [  414.696167] scsi 5:0:0:7: rdac: Attach failed (8)
> > ...
> > [  423.615364] kernel BUG at drivers/scsi/device_handler/scsi_dh_rdac.c:427!
> > [  423.615731] invalid opcode: 0000 [#1] SMP NOPTI
> > ...
> > [  423.623247] Call Trace:
> > [  423.623598]  rdac_bus_attach+0x203/0x4c0
> > [  423.623949]  ? scsi_dh_handler_attach+0x2d/0x90
> > [  423.624300]  scsi_dh_handler_attach+0x2d/0x90
> > [  423.624652]  scsi_sysfs_add_sdev+0x88/0x270
> > [  423.625004]  scsi_probe_and_add_lun+0xc47/0xd50
> > [  423.625354]  scsi_report_lun_scan+0x339/0x3b0
> > [  423.625705]  __scsi_scan_target+0xe9/0x220
> > [  423.626056]  scsi_scan_target+0xf6/0x100
> > [  423.626404]  fc_scsi_scan_rport+0xa5/0xb0
> > [  423.626757]  process_one_work+0x15e/0x3f0
> > [  423.627106]  worker_thread+0x4c/0x440
> > [  423.627453]  ? rescuer_thread+0x350/0x350
> > [  423.627804]  kthread+0xf8/0x130
> > [  423.628153]  ? kthread_destroy_worker+0x40/0x40
> > [  423.628509]  ret_from_fork+0x1f/0x40
> >
> > Fixes: 1a5dc166cd88 ("scsi_dh_rdac: update 'access_state' field")
> > Signed-off-by: Huang Cun <huangcun@xxxxxxxxxxxxxx>
> > Signed-off-by: Ding Hui <dinghui@xxxxxxxxxxxxxx>
> > Cc: Donglin Peng <pengdonglin@xxxxxxxxxxxxxx>
> > ---
> >   drivers/scsi/device_handler/scsi_dh_rdac.c | 2 ++
> >   1 file changed, 2 insertions(+)
> >
> > diff --git a/drivers/scsi/device_handler/scsi_dh_rdac.c b/drivers/scsi/device_handler/scsi_dh_rdac.c
> > index c5538645057a..9d487c2b7708 100644
> > --- a/drivers/scsi/device_handler/scsi_dh_rdac.c
> > +++ b/drivers/scsi/device_handler/scsi_dh_rdac.c
> > @@ -762,8 +762,10 @@ static int rdac_bus_attach(struct scsi_device *sdev)
> >   
> >   clean_ctlr:
> >   	spin_lock(&list_lock);
> > +	list_del_rcu(&h->node);
> >   	kref_put(&h->ctlr->kref, release_controller);
> >   	spin_unlock(&list_lock);
> > +	synchronize_rcu();
> >   
>
> Should this be:
>
> spin_lock(&list_lock);
> list_del_rcu(&h->node);
> spin_unlock(&list_lock);
>
> synchronize_rcu();
>
> kref_put(&h->ctlr->kref, release_controller);
>
>
> ?
>
> If you do the synchronize_rcu after the kref_put, then the kref_put
> could free the rdac_dh_data, while check_ownership is still
> accessing the rdac_dh_data, right?

You are right.

But I think we should keep the kref_put() and release callback be protected by list_lock, and only free
the ctlr after synchronize_rcu().

So how about the additional modify (not yet tested):

--- a/drivers/scsi/device_handler/scsi_dh_rdac.c
+++ b/drivers/scsi/device_handler/scsi_dh_rdac.c
@@ -166,6 +166,7 @@ struct rdac_controller {
 	struct scsi_device	*ms_sdev;
 	struct list_head	ms_head;
 	struct list_head	dh_list;
+	struct rcu_head		rcu;
 };
 
 struct c2_inquiry {
@@ -320,7 +321,7 @@ static void release_controller(struct kref *kref)
 	ctlr = container_of(kref, struct rdac_controller, kref);
 
 	list_del(&ctlr->node);
-	kfree(ctlr);
+	kfree_rcu(ctlr, rcu);
 }
 
 static struct rdac_controller *get_controller(int index, char *array_name,


--
Thanks,
- Ding Hui