On Thu, 2017-04-20 at 21:59 +0000, Bart Van Assche wrote:
> On Tue, 2017-04-18 at 16:56 -0700, James Bottomley wrote:
> > diff --git a/drivers/scsi/scsi_lib.c b/drivers/scsi/scsi_lib.c
> > index e5a2d590a104..31171204cfd1 100644
> > --- a/drivers/scsi/scsi_lib.c
> > +++ b/drivers/scsi/scsi_lib.c
> > @@ -2611,7 +2611,6 @@ scsi_device_set_state(struct scsi_device
> > *sdev, enum scsi_device_state state)
> > case SDEV_QUIESCE:
> > case SDEV_OFFLINE:
> > case SDEV_TRANSPORT_OFFLINE:
> > - case SDEV_BLOCK:
> > break;
> > default:
> > goto illegal;
> > @@ -2625,6 +2624,7 @@ scsi_device_set_state(struct scsi_device
> > *sdev, enum scsi_device_state state)
> > case SDEV_OFFLINE:
> > case SDEV_TRANSPORT_OFFLINE:
> > case SDEV_CANCEL:
> > + case SDEV_BLOCK:
> > case SDEV_CREATED_BLOCK:
> > break;
> > default:
> > diff --git a/drivers/scsi/scsi_sysfs.c b/drivers/scsi/scsi_sysfs.c
> > index 82dfe07b1d47..e477f95bf169 100644
> > --- a/drivers/scsi/scsi_sysfs.c
> > +++ b/drivers/scsi/scsi_sysfs.c
> > @@ -1282,8 +1282,17 @@ void __scsi_remove_device(struct scsi_device
> > *sdev)
> > return;
> >
> > if (sdev->is_visible) {
> > - if (scsi_device_set_state(sdev, SDEV_CANCEL) != 0)
> > - return;
> > + /*
> > + * If blocked, we go straight to DEL so any
> > commands
> > + * issued during the driver shutdown (like sync
> > cache)
> > + * are errored
> > + */
> > + if (scsi_device_set_state(sdev, SDEV_CANCEL) != 0)
> > {
> > + if (scsi_device_set_state(sdev, SDEV_DEL)
> > != 0)
> > + return;
> > + else
> > + scsi_start_queue(sdev);
> > + }
> >
> > bsg_unregister_queue(sdev->request_queue);
> > device_unregister(&sdev->sdev_dev);
>
> Hello James,
>
> This approach cannot work. A scsi_target_block() call by the
> transport layer can happen concurrently with the
> __scsi_remove_device() call and hence can occur at any time between
> the scsi_start_queue() call by __scsi_remove_device() and the
> sd_shutdown() call, resulting in a deadlock.
How is that possible? Once the device goes into the CANCEL state, it
no longer can be found by starget_for_each_device() because
scsi_device_get() returns NULL ... unless you also have a patch
altering that?
James
> I have been able to trigger this with my tests by simulating a cable
> pull shortly before running "rmmod ib_srp".
>
> That deadlock did not occur with the patch series that makes
> synchronize cache upon shutdown asynchronous. I'm going to resubmit
> that patch series.
>
> Bart.
