On Sat, 2014-01-04 at 11:27 -0500, Alan Stern wrote:
> On Fri, 3 Jan 2014, James Bottomley wrote:
>
> > > I'm still concerned about one thing. The previous patch does this in
> > > scsi_alloc_target():
> > >
> > > > found:
> > > > - found_target->reap_ref++;
> > > > + if (!kref_get_unless_zero(&found_target->reap_ref))
> > > > + /*
> > > > + * release routine already fired. Target is dead, but
> > > > + * STARGET_DEL may not yet be set (set in the release
> > > > + * routine), so set here as well, just in case
> > > > + */
> > > > + found_target->state = STARGET_DEL;
> > > > spin_unlock_irqrestore(shost->host_lock, flags);
> > >
> > > As a result, the two comments in this patch aren't right:
> > >
> > > > @@ -384,9 +385,15 @@ static void scsi_target_reap_ref_release(struct kref *kref)
> > > > struct scsi_target *starget
> > > > = container_of(kref, struct scsi_target, reap_ref);
> > > >
> > > > - transport_remove_device(&starget->dev);
> > > > - device_del(&starget->dev);
> > > > - starget->state = STARGET_DEL;
> > > > + /*
> > > > + * if we get here and the target is still in the CREATED state that
> > > > + * means it was allocated but never made visible (because a scan
> > > > + * turned up no LUNs), so don't call device_del() on it.
> > > > + */
> > > > + if (starget->state == STARGET_RUNNING) {
> > > > + transport_remove_device(&starget->dev);
> > > > + device_del(&starget->dev);
> > > > + }
> > >
> > > Here the state could already be STARGET_DEL, even though the target is
> > > still visible.
> >
> > Well, I agree with the theory. In practise, there are only a few
> > machine instructions between the kref going to zero and us reaching that
> > point, because kref_release will jump into this routine next, so the
> > condition would be very hard to see.
>
> It's true that the window is very small and not at all likely to be
> hit. Still, I prefer eliminating such things entirely.
>
> > However, I suppose it's easy to
> > close by checking for != STARGET_CREATED and there's no reason not to do
> > that, so I'll change it.
>
> Checking for != STARGET_CREATED is also wrong in principle. The state
> could already be STARGET_DEL even though the target was never made
> visible.
>
> The basic problem is that you are relying on the state to be an
> accurate description of the target's visibility, but
> scsi_alloc_target() changes the state without changing the visibility.
> I really think you should get rid of that extra assignment in
> scsi_alloc_target().
OK, Agreed, but that means modifying the 1/2 patch with the below. This
should make the proposed diff to 2/2 correct.
James
---
diff --git a/drivers/scsi/scsi_scan.c b/drivers/scsi/scsi_scan.c
index ef3f958..5fad646 100644
--- a/drivers/scsi/scsi_scan.c
+++ b/drivers/scsi/scsi_scan.c
@@ -417,7 +417,7 @@ static struct scsi_target *scsi_alloc_target(struct device *parent,
+ shost->transportt->target_size;
struct scsi_target *starget;
struct scsi_target *found_target;
- int error;
+ int error, ref_got;
starget = kzalloc(size, GFP_KERNEL);
if (!starget) {
@@ -466,15 +466,15 @@ static struct scsi_target *scsi_alloc_target(struct device *parent,
return starget;
found:
- if (!kref_get_unless_zero(&found_target->reap_ref))
- /*
- * release routine already fired. Target is dead, but
- * STARGET_DEL may not yet be set (set in the release
- * routine), so set here as well, just in case
- */
- found_target->state = STARGET_DEL;
+ /*
+ * release routine already fired if kref is zero, so if we can still
+ * take the reference, the target must be alive. If we can't, it must
+ * be dying and we need to wait for a new target
+ */
+ ref_got = kref_get_unless_zero(&found_target->reap_ref);
+
spin_unlock_irqrestore(shost->host_lock, flags);
- if (found_target->state != STARGET_DEL) {
+ if (ref_got) {
put_device(dev);
return found_target;
}
@@ -482,8 +482,8 @@ static struct scsi_target *scsi_alloc_target(struct device *parent,
* Unfortunately, we found a dying target; need to wait until it's
* dead before we can get a new one. There is an anomaly here. We
* *should* call scsi_target_reap() to balance the kref_get() of the
- * reap_ref above. However, since the target is in state STARGET_DEL,
- * it's already invisible and the reap_ref is irrelevant. If we call
+ * reap_ref above. However, since the target being released, it's
+ * already invisible and the reap_ref is irrelevant. If we call
* scsi_target_reap() we might spuriously do another device_del() on
* an already invisible target.
*/
--
To unsubscribe from this list: send the line "unsubscribe linux-scsi" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
