[PATCH 0/2] target refcounting infrastructure fixes for usb

[PATCH 0/2] target refcounting infrastructure fixes for usb

[James Bottomley]

After about 3 iterations as an RFC, I'm elevating this to a patch.  If
it works, I'll try for this merge window with a delayed backport to
stable (after the next kernel release to make sure no bugs got
introduced).

This set should fix our target problems with USB by making the target
visibility properly reference counted.

James Bottomley (2):
  [SCSI] fix our current target reap infrastructure.
  [SCSI] dual scan thread bug fix

 drivers/scsi/scsi_scan.c   | 112 ++++++++++++++++++++++++++++-----------------
 drivers/scsi/scsi_sysfs.c  |  20 +++++---
 include/scsi/scsi_device.h |   3 +-
 3 files changed, 86 insertions(+), 49 deletions(-)

-- 
1.8.4.3



--
To unsubscribe from this list: send the line "unsubscribe linux-usb" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

[PATCH 2/2] dual scan thread bug fix

[James Bottomley]

In the highly unusual case where two threads are running concurrently through
the scanning code scanning the same target, we run into the situation where
one may allocate the target while the other is still using it.  In this case,
because the reap checks for STARGET_CREATED and kills the target without
reference counting, the second thread will do the wrong thing on reap.

Fix this by reference counting even creates and doing the STARGET_CREATED
check in the final put.

Signed-off-by: James Bottomley <JBottomley@Parallels.com>
---
 drivers/scsi/scsi_scan.c | 23 ++++++++++++++++-------
 1 file changed, 16 insertions(+), 7 deletions(-)

diff --git a/drivers/scsi/scsi_scan.c b/drivers/scsi/scsi_scan.c
index 5fad646..4109530 100644
--- a/drivers/scsi/scsi_scan.c
+++ b/drivers/scsi/scsi_scan.c
@@ -320,6 +320,7 @@ static void scsi_target_destroy(struct scsi_target *starget)
 	struct Scsi_Host *shost = dev_to_shost(dev->parent);
 	unsigned long flags;
 
+	starget->state = STARGET_DEL;
 	transport_destroy_device(dev);
 	spin_lock_irqsave(shost->host_lock, flags);
 	if (shost->hostt->target_destroy)
@@ -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_CREATED) {
+		transport_remove_device(&starget->dev);
+		device_del(&starget->dev);
+	}
 	scsi_target_destroy(starget);
 }
 
@@ -506,11 +513,13 @@ static struct scsi_target *scsi_alloc_target(struct device *parent,
  */
 void scsi_target_reap(struct scsi_target *starget)
 {
+	/*
+	 * serious problem if this triggers: STARGET_DEL is only set in the if
+	 * the reap_ref drops to zero, so we're trying to do another final put
+	 * on an already released kref
+	 */
 	BUG_ON(starget->state == STARGET_DEL);
-	if (starget->state == STARGET_CREATED)
-		scsi_target_destroy(starget);
-	else
-		scsi_target_reap_ref_put(starget);
+	scsi_target_reap_ref_put(starget);
 }
 
 /**
-- 
1.8.4.3

[PATCH 1/2] fix our current target reap infrastructure

[James Bottomley]

This patch eliminates the reap_ref and replaces it with a proper kref.
On last put of this kref, the target is removed from visibility in
sysfs.  The final call to scsi_target_reap() for the device is done from
__scsi_remove_device() and only if the device was made visible.  This
ensures that the target disappears as soon as the last device is gone
rather than waiting until final release of the device (which is often
too long).

Reviewed-by: Alan Stern <stern-nwvwT67g6+6dFdvTe/nMLpVzexx5G7lz@public.gmane.org>
Signed-off-by: James Bottomley <JBottomley-MU7nAjRaF3makBO8gow8eQ@public.gmane.org>
---
 drivers/scsi/scsi_scan.c   | 99 ++++++++++++++++++++++++++++------------------
 drivers/scsi/scsi_sysfs.c  | 20 +++++++---
 include/scsi/scsi_device.h |  3 +-
 3 files changed, 75 insertions(+), 47 deletions(-)

diff --git a/drivers/scsi/scsi_scan.c b/drivers/scsi/scsi_scan.c
index 307a811..5fad646 100644
--- a/drivers/scsi/scsi_scan.c
+++ b/drivers/scsi/scsi_scan.c
@@ -371,6 +371,31 @@ static struct scsi_target *__scsi_find_target(struct device *parent,
 }
 
 /**
+ * scsi_target_reap_ref_release - remove target from visibility
+ * @kref: the reap_ref in the target being released
+ *
+ * Called on last put of reap_ref, which is the indication that no device
+ * under this target is visible anymore, so render the target invisible in
+ * sysfs.  Note: we have to be in user context here because the target reaps
+ * should be done in places where the scsi device visibility is being removed.
+ */
+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;
+	scsi_target_destroy(starget);
+}
+
+static void scsi_target_reap_ref_put(struct scsi_target *starget)
+{
+	kref_put(&starget->reap_ref, scsi_target_reap_ref_release);
+}
+
+/**
  * scsi_alloc_target - allocate a new or find an existing target
  * @parent:	parent of the target (need not be a scsi host)
  * @channel:	target channel number (zero if no channels)
@@ -392,7 +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) {
@@ -401,7 +426,7 @@ static struct scsi_target *scsi_alloc_target(struct device *parent,
 	}
 	dev = &starget->dev;
 	device_initialize(dev);
-	starget->reap_ref = 1;
+	kref_init(&starget->reap_ref);
 	dev->parent = get_device(parent);
 	dev_set_name(dev, "target%d:%d:%d", shost->host_no, channel, id);
 	dev->bus = &scsi_bus_type;
@@ -441,29 +466,36 @@ static struct scsi_target *scsi_alloc_target(struct device *parent,
 	return starget;
 
  found:
-	found_target->reap_ref++;
+	/*
+	 * 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;
 	}
-	/* Unfortunately, we found a dying target; need to
-	 * wait until it's dead before we can get a new one */
+	/*
+	 * 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 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.
+	 */
 	put_device(&found_target->dev);
-	flush_scheduled_work();
+	/*
+	 * length of time is irrelevant here, we just want to yield the CPU
+	 * for a tick to avoid busy waiting for the target to die.
+	 */
+	msleep(1);
 	goto retry;
 }
 
-static void scsi_target_reap_usercontext(struct work_struct *work)
-{
-	struct scsi_target *starget =
-		container_of(work, struct scsi_target, ew.work);
-
-	transport_remove_device(&starget->dev);
-	device_del(&starget->dev);
-	scsi_target_destroy(starget);
-}
-
 /**
  * scsi_target_reap - check to see if target is in use and destroy if not
  * @starget: target to be checked
@@ -474,28 +506,11 @@ static void scsi_target_reap_usercontext(struct work_struct *work)
  */
 void scsi_target_reap(struct scsi_target *starget)
 {
-	struct Scsi_Host *shost = dev_to_shost(starget->dev.parent);
-	unsigned long flags;
-	enum scsi_target_state state;
-	int empty = 0;
-
-	spin_lock_irqsave(shost->host_lock, flags);
-	state = starget->state;
-	if (--starget->reap_ref == 0 && list_empty(&starget->devices)) {
-		empty = 1;
-		starget->state = STARGET_DEL;
-	}
-	spin_unlock_irqrestore(shost->host_lock, flags);
-
-	if (!empty)
-		return;
-
-	BUG_ON(state == STARGET_DEL);
-	if (state == STARGET_CREATED)
+	BUG_ON(starget->state == STARGET_DEL);
+	if (starget->state == STARGET_CREATED)
 		scsi_target_destroy(starget);
 	else
-		execute_in_process_context(scsi_target_reap_usercontext,
-					   &starget->ew);
+		scsi_target_reap_ref_put(starget);
 }
 
 /**
@@ -1532,6 +1547,10 @@ struct scsi_device *__scsi_add_device(struct Scsi_Host *shost, uint channel,
 	}
 	mutex_unlock(&shost->scan_mutex);
 	scsi_autopm_put_target(starget);
+	/*
+	 * paired with scsi_alloc_target().  Target will be destroyed unless
+	 * scsi_probe_and_add_lun made an underlying device visible
+	 */
 	scsi_target_reap(starget);
 	put_device(&starget->dev);
 
@@ -1612,8 +1631,10 @@ static void __scsi_scan_target(struct device *parent, unsigned int channel,
 
  out_reap:
 	scsi_autopm_put_target(starget);
-	/* now determine if the target has any children at all
-	 * and if not, nuke it */
+	/*
+	 * paired with scsi_alloc_target(): determine if the target has
+	 * any children at all and if not, nuke it
+	 */
 	scsi_target_reap(starget);
 
 	put_device(&starget->dev);
diff --git a/drivers/scsi/scsi_sysfs.c b/drivers/scsi/scsi_sysfs.c
index 9117d0b..665acbf 100644
--- a/drivers/scsi/scsi_sysfs.c
+++ b/drivers/scsi/scsi_sysfs.c
@@ -383,17 +383,14 @@ static void scsi_device_dev_release_usercontext(struct work_struct *work)
 {
 	struct scsi_device *sdev;
 	struct device *parent;
-	struct scsi_target *starget;
 	struct list_head *this, *tmp;
 	unsigned long flags;
 
 	sdev = container_of(work, struct scsi_device, ew.work);
 
 	parent = sdev->sdev_gendev.parent;
-	starget = to_scsi_target(parent);
 
 	spin_lock_irqsave(sdev->host->host_lock, flags);
-	starget->reap_ref++;
 	list_del(&sdev->siblings);
 	list_del(&sdev->same_target_siblings);
 	list_del(&sdev->starved_entry);
@@ -413,8 +410,6 @@ static void scsi_device_dev_release_usercontext(struct work_struct *work)
 	/* NULL queue means the device can't be used */
 	sdev->request_queue = NULL;
 
-	scsi_target_reap(scsi_target(sdev));
-
 	kfree(sdev->inquiry);
 	kfree(sdev);
 
@@ -1071,6 +1066,13 @@ void __scsi_remove_device(struct scsi_device *sdev)
 		sdev->host->hostt->slave_destroy(sdev);
 	transport_destroy_device(dev);
 
+	/*
+	 * Paired with the kref_get() in scsi_sysfs_initialize().  We have
+	 * remoed sysfs visibility from the device, so make the target
+	 * invisible if this was the last device underneath it.
+	 */
+	scsi_target_reap(scsi_target(sdev));
+
 	put_device(dev);
 }
 
@@ -1133,7 +1135,7 @@ void scsi_remove_target(struct device *dev)
 			continue;
 		if (starget->dev.parent == dev || &starget->dev == dev) {
 			/* assuming new targets arrive at the end */
-			starget->reap_ref++;
+			kref_get(&starget->reap_ref);
 			spin_unlock_irqrestore(shost->host_lock, flags);
 			if (last)
 				scsi_target_reap(last);
@@ -1217,6 +1219,12 @@ void scsi_sysfs_device_initialize(struct scsi_device *sdev)
 	list_add_tail(&sdev->same_target_siblings, &starget->devices);
 	list_add_tail(&sdev->siblings, &shost->__devices);
 	spin_unlock_irqrestore(shost->host_lock, flags);
+	/*
+	 * device can now only be removed via __scsi_remove_device() so hold
+	 * the target.  Target will be held in CREATED state until something
+	 * beneath it becomes visible (in which case it moves to RUNNING)
+	 */
+	kref_get(&starget->reap_ref);
 }
 
 int scsi_is_sdev_device(const struct device *dev)
diff --git a/include/scsi/scsi_device.h b/include/scsi/scsi_device.h
index d65fbec..b4f1eff 100644
--- a/include/scsi/scsi_device.h
+++ b/include/scsi/scsi_device.h
@@ -257,7 +257,7 @@ struct scsi_target {
 	struct list_head	siblings;
 	struct list_head	devices;
 	struct device		dev;
-	unsigned int		reap_ref; /* protected by the host lock */
+	struct kref		reap_ref; /* last put renders target invisible */
 	unsigned int		channel;
 	unsigned int		id; /* target id ... replace
 				     * scsi_device.id eventually */
@@ -284,7 +284,6 @@ struct scsi_target {
 #define SCSI_DEFAULT_TARGET_BLOCKED	3
 
 	char			scsi_level;
-	struct execute_work	ew;
 	enum scsi_target_state	state;
 	void 			*hostdata; /* available to low-level driver */
 	unsigned long		starget_data[0]; /* for the transport */
-- 
1.8.4.3



--
To unsubscribe from this list: send the line "unsubscribe linux-usb" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Re: [PATCH 0/2] target refcounting infrastructure fixes for usb

[Sarah Sharp]

Hi James,

Sorry for the extremely late ack.  These patches work fine when applied
against 3.14-rc5, and I can't trigger the oops anymore.  Hans' UAS
patches are going to be queued for 3.15.  Any chance of these patches
making it into 3.15 as well?

Sarah Sharp

On Tue, Jan 21, 2014 at 06:58:39AM -0800, James Bottomley wrote:
> After about 3 iterations as an RFC, I'm elevating this to a patch.  If
> it works, I'll try for this merge window with a delayed backport to
> stable (after the next kernel release to make sure no bugs got
> introduced).
> 
> This set should fix our target problems with USB by making the target
> visibility properly reference counted.
> 
> James Bottomley (2):
>   [SCSI] fix our current target reap infrastructure.
>   [SCSI] dual scan thread bug fix
> 
>  drivers/scsi/scsi_scan.c   | 112 ++++++++++++++++++++++++++++-----------------
>  drivers/scsi/scsi_sysfs.c  |  20 +++++---
>  include/scsi/scsi_device.h |   3 +-
>  3 files changed, 86 insertions(+), 49 deletions(-)
> 
> -- 
> 1.8.4.3
> 
> 
> 
--
To unsubscribe from this list: send the line "unsubscribe linux-usb" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Re: [PATCH 0/2] target refcounting infrastructure fixes for usb

[James Bottomley]

On Thu, 2014-03-06 at 15:18 -0800, Sarah Sharp wrote:
> Hi James,
> 
> Sorry for the extremely late ack.

That's OK ... my original plan was to send them in the last merge window
even without your ack, but that got a bit derailed, so I'll do it in
this one with.

>   These patches work fine when applied
> against 3.14-rc5, and I can't trigger the oops anymore.  Hans' UAS
> patches are going to be queued for 3.15.  Any chance of these patches
> making it into 3.15 as well?

Yes, the plan is to queue for 3.15 and request a delayed backport to
stable in case any of the extensive logic changes triggers another bug
in the field.

James