[PATCH 0/2] Namespace attachment fixes/simplifications

[PATCH 0/2] Namespace attachment fixes/simplifications

[Keith Busch]

This started off investigating a suggestion Christoph made on the awkward
way the driver removes disks if the controller can't handle IO after
attempting a resume from suspend.

It turns out that it has been a LONG time since the error path from
resume was tested. It definitely was broken, so this turned into a two
part series. I can split 2/2 into a two more if anyone thinks changes
are unrelated.

>From testing it became clear that it is not sufficient to reference count
only the controller anymore. Jens suggested a while back to reference
count the namespaces, but I didn't think there was a need since namespaces
couldn't be deleted while the controller had active references anyway. Now
that namespaces can be deleted independent of the controller, we really
do need a separate reference count.

Keith Busch (2):
  NVMe: Reference count open namespaces
  NVMe: Simplify device resume on io queue failure

 drivers/block/nvme-core.c |   80 ++++++++++++++++++---------------------------
 include/linux/nvme.h      |    2 +-
 2 files changed, 32 insertions(+), 50 deletions(-)

-- 
1.7.10.4

[PATCH 1/2] NVMe: Reference count open namespaces

[Keith Busch]

Dynamic namespace attachment means the namespace may be removed at any
time, so the namespace reference count can not be tied to the device
reference count. This fixes a NULL dereference if an opened namespace
is detached from a controller.

Signed-off-by: Keith Busch <keith.busch at intel.com>
---
 drivers/block/nvme-core.c |   29 ++++++++++++++++++++---------
 include/linux/nvme.h      |    1 +
 2 files changed, 21 insertions(+), 9 deletions(-)

diff --git a/drivers/block/nvme-core.c b/drivers/block/nvme-core.c
index 6f04771..b02ae3d 100644
--- a/drivers/block/nvme-core.c
+++ b/drivers/block/nvme-core.c
@@ -1943,6 +1943,18 @@ static int nvme_compat_ioctl(struct block_device *bdev, fmode_t mode,
 #define nvme_compat_ioctl	NULL
 #endif
 
+static void nvme_free_ns(struct kref *kref)
+{
+	struct nvme_ns *ns = container_of(kref, struct nvme_ns, kref);
+
+	spin_lock(&dev_list_lock);
+	ns->disk->private_data = NULL;
+	spin_unlock(&dev_list_lock);
+
+	put_disk(ns->disk);
+	kfree(ns);
+}
+
 static int nvme_open(struct block_device *bdev, fmode_t mode)
 {
 	int ret = 0;
@@ -1952,21 +1964,25 @@ static int nvme_open(struct block_device *bdev, fmode_t mode)
 	ns = bdev->bd_disk->private_data;
 	if (!ns)
 		ret = -ENXIO;
-	else if (!kref_get_unless_zero(&ns->dev->kref))
+	else if (!kref_get_unless_zero(&ns->kref))
 		ret = -ENXIO;
+	else if (!kref_get_unless_zero(&ns->dev->kref)) {
+		kref_put(&ns->kref, nvme_free_ns);
+		ret = -ENXIO;
+	}
 	spin_unlock(&dev_list_lock);
 
 	return ret;
 }
 
 static void nvme_free_dev(struct kref *kref);
-
 static void nvme_release(struct gendisk *disk, fmode_t mode)
 {
 	struct nvme_ns *ns = disk->private_data;
 	struct nvme_dev *dev = ns->dev;
 
 	kref_put(&dev->kref, nvme_free_dev);
+	kref_put(&ns->kref, nvme_free_ns);
 }
 
 static int nvme_getgeo(struct block_device *bd, struct hd_geometry *geo)
@@ -2126,6 +2142,7 @@ static void nvme_alloc_ns(struct nvme_dev *dev, unsigned nsid)
 	if (!disk)
 		goto out_free_queue;
 
+	kref_init(&ns->kref);
 	ns->ns_id = nsid;
 	ns->disk = disk;
 	ns->lba_shift = 9; /* set to a default value for 512 until disk is validated */
@@ -2360,13 +2377,7 @@ static int nvme_setup_io_queues(struct nvme_dev *dev)
 static void nvme_free_namespace(struct nvme_ns *ns)
 {
 	list_del(&ns->list);
-
-	spin_lock(&dev_list_lock);
-	ns->disk->private_data = NULL;
-	spin_unlock(&dev_list_lock);
-
-	put_disk(ns->disk);
-	kfree(ns);
+	kref_put(&ns->kref, nvme_free_ns);
 }
 
 static int ns_cmp(void *priv, struct list_head *a, struct list_head *b)
diff --git a/include/linux/nvme.h b/include/linux/nvme.h
index b5812c3..992b9c1 100644
--- a/include/linux/nvme.h
+++ b/include/linux/nvme.h
@@ -135,6 +135,7 @@ struct nvme_ns {
 	struct nvme_dev *dev;
 	struct request_queue *queue;
 	struct gendisk *disk;
+	struct kref kref;
 
 	unsigned ns_id;
 	int lba_shift;
-- 
1.7.10.4

[PATCH 2/2] NVMe: Simplify device resume on io queue failure

[Keith Busch]

Releasing io queues and disks was done in a work queue separate from
the controller resumption task to handle failed drives after a resume
from suspend. This is not necesssary since we can resume a device
asynchronously the same as probe.

This patch makes resume use the probe work, and this lets the task delete
gendisks directly if the device is managable but not io capable. Since
deleting disks was the only reason we had the convoluted "reset_workfn",
this patch removes the unnecessary indirection, and places all work
tasks onto the global work queue.

Signed-off-by: Keith Busch <keith.busch at intel.com>
---
 drivers/block/nvme-core.c |   51 ++++++++++-----------------------------------
 include/linux/nvme.h      |    1 -
 2 files changed, 11 insertions(+), 41 deletions(-)

diff --git a/drivers/block/nvme-core.c b/drivers/block/nvme-core.c
index b02ae3d..81b42078 100644
--- a/drivers/block/nvme-core.c
+++ b/drivers/block/nvme-core.c
@@ -79,7 +79,6 @@ MODULE_PARM_DESC(use_cmb_sqes, "use controller's memory buffer for I/O SQes");
 static DEFINE_SPINLOCK(dev_list_lock);
 static LIST_HEAD(dev_list);
 static struct task_struct *nvme_thread;
-static struct workqueue_struct *nvme_workq;
 static wait_queue_head_t nvme_kthread_wait;
 
 static struct class *nvme_class;
@@ -1285,8 +1284,7 @@ static void nvme_abort_req(struct request *req)
 		list_del_init(&dev->node);
 		dev_warn(dev->dev, "I/O %d QID %d timeout, reset controller\n",
 							req->tag, nvmeq->qid);
-		dev->reset_workfn = nvme_reset_failed_dev;
-		queue_work(nvme_workq, &dev->reset_work);
+		schedule_work(&dev->reset_work);
  out:
 		spin_unlock_irqrestore(&dev_list_lock, flags);
 		return;
@@ -2095,8 +2093,7 @@ static int nvme_kthread(void *data)
 				dev_warn(dev->dev,
 					"Failed status: %x, reset controller\n",
 					readl(&dev->bar->csts));
-				dev->reset_workfn = nvme_reset_failed_dev;
-				queue_work(nvme_workq, &dev->reset_work);
+				schedule_work(&dev->reset_work);
 				continue;
 			}
 			for (i = 0; i < dev->queue_count; i++) {
@@ -3047,14 +3044,6 @@ static int nvme_remove_dead_ctrl(void *arg)
 	return 0;
 }
 
-static void nvme_remove_disks(struct work_struct *ws)
-{
-	struct nvme_dev *dev = container_of(ws, struct nvme_dev, reset_work);
-
-	nvme_free_queues(dev, 1);
-	nvme_dev_remove(dev);
-}
-
 static int nvme_dev_resume(struct nvme_dev *dev)
 {
 	int ret;
@@ -3063,10 +3052,10 @@ static int nvme_dev_resume(struct nvme_dev *dev)
 	if (ret)
 		return ret;
 	if (dev->online_queues < 2) {
-		spin_lock(&dev_list_lock);
-		dev->reset_workfn = nvme_remove_disks;
-		queue_work(nvme_workq, &dev->reset_work);
-		spin_unlock(&dev_list_lock);
+		dev_warn(dev->dev, "no io queues created\n");
+		nvme_free_queues(dev, 1);
+		nvme_dev_remove(dev);
+		nvme_free_namespaces(dev);
 	} else {
 		nvme_unfreeze_queues(dev);
 		nvme_dev_add(dev);
@@ -3113,12 +3102,6 @@ static void nvme_reset_failed_dev(struct work_struct *ws)
 	nvme_dev_reset(dev);
 }
 
-static void nvme_reset_workfn(struct work_struct *work)
-{
-	struct nvme_dev *dev = container_of(work, struct nvme_dev, reset_work);
-	dev->reset_workfn(work);
-}
-
 static int nvme_reset(struct nvme_dev *dev)
 {
 	int ret = -EBUSY;
@@ -3128,8 +3111,7 @@ static int nvme_reset(struct nvme_dev *dev)
 
 	spin_lock(&dev_list_lock);
 	if (!work_pending(&dev->reset_work)) {
-		dev->reset_workfn = nvme_reset_failed_dev;
-		queue_work(nvme_workq, &dev->reset_work);
+		schedule_work(&dev->reset_work);
 		ret = 0;
 	}
 	spin_unlock(&dev_list_lock);
@@ -3181,8 +3163,7 @@ static int nvme_probe(struct pci_dev *pdev, const struct pci_device_id *id)
 		goto free;
 
 	INIT_LIST_HEAD(&dev->namespaces);
-	dev->reset_workfn = nvme_reset_failed_dev;
-	INIT_WORK(&dev->reset_work, nvme_reset_workfn);
+	INIT_WORK(&dev->reset_work, nvme_reset_failed_dev);
 	dev->dev = get_device(&pdev->dev);
 	pci_set_drvdata(pdev, dev);
 	result = nvme_set_instance(dev);
@@ -3245,7 +3226,7 @@ static void nvme_reset_notify(struct pci_dev *pdev, bool prepare)
 	if (prepare)
 		nvme_dev_shutdown(dev);
 	else
-		nvme_dev_resume(dev);
+		schedule_work(&dev->probe_work);
 }
 
 static void nvme_shutdown(struct pci_dev *pdev)
@@ -3299,10 +3280,7 @@ static int nvme_resume(struct device *dev)
 	struct pci_dev *pdev = to_pci_dev(dev);
 	struct nvme_dev *ndev = pci_get_drvdata(pdev);
 
-	if (nvme_dev_resume(ndev) && !work_busy(&ndev->reset_work)) {
-		ndev->reset_workfn = nvme_reset_failed_dev;
-		queue_work(nvme_workq, &ndev->reset_work);
-	}
+	schedule_work(&ndev->probe_work);
 	return 0;
 }
 #endif
@@ -3345,13 +3323,9 @@ static int __init nvme_init(void)
 
 	init_waitqueue_head(&nvme_kthread_wait);
 
-	nvme_workq = create_singlethread_workqueue("nvme");
-	if (!nvme_workq)
-		return -ENOMEM;
-
 	result = register_blkdev(nvme_major, "nvme");
 	if (result < 0)
-		goto kill_workq;
+		return result;
 	else if (result > 0)
 		nvme_major = result;
 
@@ -3379,8 +3353,6 @@ static int __init nvme_init(void)
 	__unregister_chrdev(nvme_char_major, 0, NVME_MINORS, "nvme");
  unregister_blkdev:
 	unregister_blkdev(nvme_major, "nvme");
- kill_workq:
-	destroy_workqueue(nvme_workq);
 	return result;
 }
 
@@ -3388,7 +3360,6 @@ static void __exit nvme_exit(void)
 {
 	pci_unregister_driver(&nvme_driver);
 	unregister_blkdev(nvme_major, "nvme");
-	destroy_workqueue(nvme_workq);
 	class_destroy(nvme_class);
 	__unregister_chrdev(nvme_char_major, 0, NVME_MINORS, "nvme");
 	BUG_ON(nvme_thread && !IS_ERR(nvme_thread));
diff --git a/include/linux/nvme.h b/include/linux/nvme.h
index 992b9c1..7725b4c 100644
--- a/include/linux/nvme.h
+++ b/include/linux/nvme.h
@@ -104,7 +104,6 @@ struct nvme_dev {
 	struct list_head namespaces;
 	struct kref kref;
 	struct device *device;
-	work_func_t reset_workfn;
 	struct work_struct reset_work;
 	struct work_struct probe_work;
 	struct work_struct scan_work;
-- 
1.7.10.4

[PATCH 1/2] NVMe: Reference count open namespaces

[Christoph Hellwig]

Shouldn't each namesapce have a reference on the nvme_dev while we're at
it?

Otherwise looks good:

Reviewed-by: Christoph Hellwig <hch at lst.de>

[PATCH 2/2] NVMe: Simplify device resume on io queue failure

[Christoph Hellwig]

On Thu, Oct 01, 2015@05:14:11PM -0600, Keith Busch wrote:
> this patch removes the unnecessary indirection, and places all work
> tasks onto the global work queue.

Can you explain why nvme_workq is removed here?  I don't really
understand why it's realted, but that might be because I don't
even know why it exists in the first place.  Maybe this should be a
separate patch with a separate changelog entry?

Otherwise looks good,

Reviewed-by: Christoph Hellwig <hch at lst.de>

[PATCH 2/2] NVMe: Simplify device resume on io queue failure

[Keith Busch]

On Fri, 2 Oct 2015, Christoph Hellwig wrote:
> On Thu, Oct 01, 2015@05:14:11PM -0600, Keith Busch wrote:
>> this patch removes the unnecessary indirection, and places all work
>> tasks onto the global work queue.
>
> Can you explain why nvme_workq is removed here?  I don't really
> understand why it's realted, but that might be because I don't
> even know why it exists in the first place.  Maybe this should be a
> separate patch with a separate changelog entry?

That's the part I thought may need to be separated. I noticed using the
single threaded work queue for recovery was a bad idea when confirming
the failure injection on my drives to test the rest of the patch.

So yes, the relationship is a bit dubious. I ran out of time to split
before posting yesterday. I'll resend 2/2 as two patches.

> Otherwise looks good,
>
> Reviewed-by: Christoph Hellwig <hch at lst.de>

[PATCH 1/2] NVMe: Reference count open namespaces

[Keith Busch]

On Fri, 2 Oct 2015, Christoph Hellwig wrote:
> Shouldn't each namesapce have a reference on the nvme_dev while we're at
> it?

We just need to make a small change to the device removal. An orderly
removal won't drop the final namespace ref until after the last device ref
is closed first. This relationship isn't necessary and can be simplified
even more if we remove it.

Should I fix and resend this patch, or patch on top of this one?

> Otherwise looks good:
>
> Reviewed-by: Christoph Hellwig <hch at lst.de>

[PATCH 1/2] NVMe: Reference count open namespaces

[Christoph Hellwig]

On Fri, Oct 02, 2015@02:26:22PM +0000, Keith Busch wrote:
> On Fri, 2 Oct 2015, Christoph Hellwig wrote:
> >Shouldn't each namesapce have a reference on the nvme_dev while we're at
> >it?
> 
> We just need to make a small change to the device removal. An orderly
> removal won't drop the final namespace ref until after the last device ref
> is closed first. This relationship isn't necessary and can be simplified
> even more if we remove it.
> 
> Should I fix and resend this patch, or patch on top of this one?

Doing that on top of this patch sounds fine to me.