[PATCH 0/2] blk-mq: fix kobject lifetime issue

[PATCH 0/2] blk-mq: fix kobject lifetime issue

[Ming Lei]

Hi Jens,

The 1st patch fixes the kobject lifetime issue which is triggerd when
DEBUG_KOBJECT_RELEASE is enabled.

The 2nd patch can be thought as one follow-up cleanup.


Ming Lei (2):
  blk-mq: not embed .mq_kobj and ctx->kobj into queue instance
  blk-mq: alloc q->queue_ctx as normal array

 block/blk-mq-sysfs.c   | 59 +++++++++++++++++++++++++++++++++++++++-----------
 block/blk-mq.c         | 19 ++++++++++------
 block/blk-mq.h         |  4 ++--
 include/linux/blkdev.h |  4 ++--
 4 files changed, 62 insertions(+), 24 deletions(-)

Cc: Guenter Roeck <linux@roeck-us.net>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>


-- 
2.9.5

[PATCH 1/2] blk-mq: not embed .mq_kobj and ctx->kobj into queue instance

[Ming Lei]

Even though .mq_kobj, ctx->kobj and q->kobj share same lifetime
from block layer's view, actually they don't because userspace may
grab one kobject anytime via sysfs, so each kobject's lifetime has
to be independent, then the objects(mq_kobj, ctx) which hosts its
own kobject have to be allocated dynamically.

This patch fixes kernel panic issue during booting when DEBUG_KOBJECT_RELEASE
is enabled.

Reported-by: Guenter Roeck <linux@roeck-us.net>
Cc: Guenter Roeck <linux@roeck-us.net>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: stable@vger.kernel.org
Signed-off-by: Ming Lei <ming.lei@redhat.com>
---
 block/blk-mq-sysfs.c   | 59 +++++++++++++++++++++++++++++++++++++++-----------
 block/blk-mq.c         | 13 ++++++-----
 block/blk-mq.h         |  4 ++--
 include/linux/blkdev.h |  4 ++--
 4 files changed, 58 insertions(+), 22 deletions(-)

diff --git a/block/blk-mq-sysfs.c b/block/blk-mq-sysfs.c
index 3d25b9c419e9..bab236955f56 100644
--- a/block/blk-mq-sysfs.c
+++ b/block/blk-mq-sysfs.c
@@ -13,8 +13,22 @@
 #include "blk-mq.h"
 #include "blk-mq-tag.h"
 
+struct blk_mq_kobj {
+	struct kobject kobj;
+};
+
 static void blk_mq_sysfs_release(struct kobject *kobj)
 {
+	struct blk_mq_kobj *mq_kobj = container_of(kobj, struct blk_mq_kobj,
+						   kobj);
+	kfree(mq_kobj);
+}
+
+static void blk_mq_ctx_sysfs_release(struct kobject *kobj)
+{
+	struct blk_mq_ctx *ctx = container_of(kobj, struct blk_mq_ctx, kobj);
+
+	kfree(ctx);
 }
 
 static void blk_mq_hw_sysfs_release(struct kobject *kobj)
@@ -213,7 +227,7 @@ static struct kobj_type blk_mq_ktype = {
 static struct kobj_type blk_mq_ctx_ktype = {
 	.sysfs_ops	= &blk_mq_sysfs_ops,
 	.default_attrs	= default_ctx_attrs,
-	.release	= blk_mq_sysfs_release,
+	.release	= blk_mq_ctx_sysfs_release,
 };
 
 static struct kobj_type blk_mq_hw_ktype = {
@@ -245,7 +259,7 @@ static int blk_mq_register_hctx(struct blk_mq_hw_ctx *hctx)
 	if (!hctx->nr_ctx)
 		return 0;
 
-	ret = kobject_add(&hctx->kobj, &q->mq_kobj, "%u", hctx->queue_num);
+	ret = kobject_add(&hctx->kobj, q->mq_kobj, "%u", hctx->queue_num);
 	if (ret)
 		return ret;
 
@@ -268,8 +282,8 @@ void blk_mq_unregister_dev(struct device *dev, struct request_queue *q)
 	queue_for_each_hw_ctx(q, hctx, i)
 		blk_mq_unregister_hctx(hctx);
 
-	kobject_uevent(&q->mq_kobj, KOBJ_REMOVE);
-	kobject_del(&q->mq_kobj);
+	kobject_uevent(q->mq_kobj, KOBJ_REMOVE);
+	kobject_del(q->mq_kobj);
 	kobject_put(&dev->kobj);
 
 	q->mq_sysfs_init_done = false;
@@ -286,23 +300,42 @@ void blk_mq_sysfs_deinit(struct request_queue *q)
 	int cpu;
 
 	for_each_possible_cpu(cpu) {
-		ctx = per_cpu_ptr(q->queue_ctx, cpu);
+		ctx = *per_cpu_ptr(q->queue_ctx, cpu);
 		kobject_put(&ctx->kobj);
 	}
-	kobject_put(&q->mq_kobj);
+	kobject_put(q->mq_kobj);
 }
 
-void blk_mq_sysfs_init(struct request_queue *q)
+int blk_mq_sysfs_init(struct request_queue *q)
 {
 	struct blk_mq_ctx *ctx;
 	int cpu;
+	struct blk_mq_kobj *mq_kobj;
+
+	mq_kobj = kzalloc(sizeof(struct blk_mq_kobj), GFP_KERNEL);
+	if (!mq_kobj)
+		return -ENOMEM;
 
-	kobject_init(&q->mq_kobj, &blk_mq_ktype);
+	kobject_init(&mq_kobj->kobj, &blk_mq_ktype);
 
 	for_each_possible_cpu(cpu) {
-		ctx = per_cpu_ptr(q->queue_ctx, cpu);
+		ctx = kzalloc_node(sizeof(*ctx), GFP_KERNEL, cpu_to_node(cpu));
+		if (!ctx)
+			goto fail;
+		*per_cpu_ptr(q->queue_ctx, cpu) = ctx;
 		kobject_init(&ctx->kobj, &blk_mq_ctx_ktype);
 	}
+	q->mq_kobj = &mq_kobj->kobj;
+	return 0;
+
+ fail:
+	for_each_possible_cpu(cpu) {
+		ctx = *per_cpu_ptr(q->queue_ctx, cpu);
+		if (ctx)
+			kobject_put(&ctx->kobj);
+	}
+	kobject_put(&mq_kobj->kobj);
+	return -ENOMEM;
 }
 
 int __blk_mq_register_dev(struct device *dev, struct request_queue *q)
@@ -313,11 +346,11 @@ int __blk_mq_register_dev(struct device *dev, struct request_queue *q)
 	WARN_ON_ONCE(!q->kobj.parent);
 	lockdep_assert_held(&q->sysfs_lock);
 
-	ret = kobject_add(&q->mq_kobj, kobject_get(&dev->kobj), "%s", "mq");
+	ret = kobject_add(q->mq_kobj, kobject_get(&dev->kobj), "%s", "mq");
 	if (ret < 0)
 		goto out;
 
-	kobject_uevent(&q->mq_kobj, KOBJ_ADD);
+	kobject_uevent(q->mq_kobj, KOBJ_ADD);
 
 	queue_for_each_hw_ctx(q, hctx, i) {
 		ret = blk_mq_register_hctx(hctx);
@@ -334,8 +367,8 @@ int __blk_mq_register_dev(struct device *dev, struct request_queue *q)
 	while (--i >= 0)
 		blk_mq_unregister_hctx(q->queue_hw_ctx[i]);
 
-	kobject_uevent(&q->mq_kobj, KOBJ_REMOVE);
-	kobject_del(&q->mq_kobj);
+	kobject_uevent(q->mq_kobj, KOBJ_REMOVE);
+	kobject_del(q->mq_kobj);
 	kobject_put(&dev->kobj);
 	return ret;
 }
diff --git a/block/blk-mq.c b/block/blk-mq.c
index 3b823891b3ef..3589ee601f37 100644
--- a/block/blk-mq.c
+++ b/block/blk-mq.c
@@ -2299,7 +2299,7 @@ static void blk_mq_init_cpu_queues(struct request_queue *q,
 	unsigned int i, j;
 
 	for_each_possible_cpu(i) {
-		struct blk_mq_ctx *__ctx = per_cpu_ptr(q->queue_ctx, i);
+		struct blk_mq_ctx *__ctx = *per_cpu_ptr(q->queue_ctx, i);
 		struct blk_mq_hw_ctx *hctx;
 
 		__ctx->cpu = i;
@@ -2385,7 +2385,7 @@ static void blk_mq_map_swqueue(struct request_queue *q)
 			set->map[0].mq_map[i] = 0;
 		}
 
-		ctx = per_cpu_ptr(q->queue_ctx, i);
+		ctx = *per_cpu_ptr(q->queue_ctx, i);
 		for (j = 0; j < set->nr_maps; j++) {
 			hctx = blk_mq_map_queue_type(q, j, i);
 
@@ -2731,18 +2731,19 @@ struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
 	if (!q->poll_cb)
 		goto err_exit;
 
-	q->queue_ctx = alloc_percpu(struct blk_mq_ctx);
+	q->queue_ctx = alloc_percpu(struct blk_mq_ctx *);
 	if (!q->queue_ctx)
 		goto err_exit;
 
 	/* init q->mq_kobj and sw queues' kobjects */
-	blk_mq_sysfs_init(q);
+	if (blk_mq_sysfs_init(q))
+		goto err_percpu;
 
 	q->nr_queues = nr_hw_queues(set);
 	q->queue_hw_ctx = kcalloc_node(q->nr_queues, sizeof(*(q->queue_hw_ctx)),
 						GFP_KERNEL, set->numa_node);
 	if (!q->queue_hw_ctx)
-		goto err_percpu;
+		goto err_sys_init;
 
 	blk_mq_realloc_hw_ctxs(set, q);
 	if (!q->nr_hw_queues)
@@ -2794,6 +2795,8 @@ struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
 
 err_hctxs:
 	kfree(q->queue_hw_ctx);
+err_sys_init:
+	blk_mq_sysfs_deinit(q);
 err_percpu:
 	free_percpu(q->queue_ctx);
 err_exit:
diff --git a/block/blk-mq.h b/block/blk-mq.h
index facb6e9ddce4..84898793c230 100644
--- a/block/blk-mq.h
+++ b/block/blk-mq.h
@@ -108,7 +108,7 @@ static inline struct blk_mq_hw_ctx *blk_mq_map_queue(struct request_queue *q,
 /*
  * sysfs helpers
  */
-extern void blk_mq_sysfs_init(struct request_queue *q);
+extern int blk_mq_sysfs_init(struct request_queue *q);
 extern void blk_mq_sysfs_deinit(struct request_queue *q);
 extern int __blk_mq_register_dev(struct device *dev, struct request_queue *q);
 extern int blk_mq_sysfs_register(struct request_queue *q);
@@ -129,7 +129,7 @@ static inline enum mq_rq_state blk_mq_rq_state(struct request *rq)
 static inline struct blk_mq_ctx *__blk_mq_get_ctx(struct request_queue *q,
 					   unsigned int cpu)
 {
-	return per_cpu_ptr(q->queue_ctx, cpu);
+	return *per_cpu_ptr(q->queue_ctx, cpu);
 }
 
 /*
diff --git a/include/linux/blkdev.h b/include/linux/blkdev.h
index 1d185f1fc333..9e3892bd67fd 100644
--- a/include/linux/blkdev.h
+++ b/include/linux/blkdev.h
@@ -407,7 +407,7 @@ struct request_queue {
 	const struct blk_mq_ops	*mq_ops;
 
 	/* sw queues */
-	struct blk_mq_ctx __percpu	*queue_ctx;
+	struct blk_mq_ctx __percpu	**queue_ctx;
 	unsigned int		nr_queues;
 
 	unsigned int		queue_depth;
@@ -456,7 +456,7 @@ struct request_queue {
 	/*
 	 * mq queue kobject
 	 */
-	struct kobject mq_kobj;
+	struct kobject *mq_kobj;
 
 #ifdef  CONFIG_BLK_DEV_INTEGRITY
 	struct blk_integrity integrity;
-- 
2.9.5

Re: [PATCH 1/2] blk-mq: not embed .mq_kobj and ctx->kobj into queue instance

[jianchao.wang]

On 11/16/18 11:28 AM, Ming Lei wrote:
...
>  
> +struct blk_mq_kobj {
> +	struct kobject kobj;
> +};
> +
>  static void blk_mq_sysfs_release(struct kobject *kobj)
>  {
> +	struct blk_mq_kobj *mq_kobj = container_of(kobj, struct blk_mq_kobj,
> +						   kobj);
> +	kfree(mq_kobj);
> +}
> +
...
>  
> -void blk_mq_sysfs_init(struct request_queue *q)
> +int blk_mq_sysfs_init(struct request_queue *q)
>  {
>  	struct blk_mq_ctx *ctx;
>  	int cpu;
> +	struct blk_mq_kobj *mq_kobj;
> +
> +	mq_kobj = kzalloc(sizeof(struct blk_mq_kobj), GFP_KERNEL);
> +	if (!mq_kobj)
> +		return -ENOMEM;
>  
> -	kobject_init(&q->mq_kobj, &blk_mq_ktype);
> +	kobject_init(&mq_kobj->kobj, &blk_mq_ktype);
>  
>  	for_each_possible_cpu(cpu) {
> -		ctx = per_cpu_ptr(q->queue_ctx, cpu);
> +		ctx = kzalloc_node(sizeof(*ctx), GFP_KERNEL, cpu_to_node(cpu));
> +		if (!ctx)
> +			goto fail;
> +		*per_cpu_ptr(q->queue_ctx, cpu) = ctx;
>  		kobject_init(&ctx->kobj, &blk_mq_ctx_ktype);
>  	}
> +	q->mq_kobj = &mq_kobj->kobj;
> +	return 0;
> +
> + fail:
> +	for_each_possible_cpu(cpu) {
> +		ctx = *per_cpu_ptr(q->queue_ctx, cpu);
> +		if (ctx)
> +			kobject_put(&ctx->kobj);
> +	}
> +	kobject_put(&mq_kobj->kobj);
> +	return -ENOMEM;
>  }


blk_mq_kobj looks meaningless, why do we need it, or do I miss something ?
And maybe we should allocate ctx in blk_mq_init_allocated_queue.

Thanks
Jianchao

Re: [PATCH 1/2] blk-mq: not embed .mq_kobj and ctx->kobj into queue instance

[Ming Lei]

On Fri, Nov 16, 2018 at 02:11:07PM +0800, jianchao.wang wrote:
> 
> 
> On 11/16/18 11:28 AM, Ming Lei wrote:
> ...
> >  
> > +struct blk_mq_kobj {
> > +	struct kobject kobj;
> > +};
> > +
> >  static void blk_mq_sysfs_release(struct kobject *kobj)
> >  {
> > +	struct blk_mq_kobj *mq_kobj = container_of(kobj, struct blk_mq_kobj,
> > +						   kobj);
> > +	kfree(mq_kobj);
> > +}
> > +
> ...
> >  
> > -void blk_mq_sysfs_init(struct request_queue *q)
> > +int blk_mq_sysfs_init(struct request_queue *q)
> >  {
> >  	struct blk_mq_ctx *ctx;
> >  	int cpu;
> > +	struct blk_mq_kobj *mq_kobj;
> > +
> > +	mq_kobj = kzalloc(sizeof(struct blk_mq_kobj), GFP_KERNEL);
> > +	if (!mq_kobj)
> > +		return -ENOMEM;
> >  
> > -	kobject_init(&q->mq_kobj, &blk_mq_ktype);
> > +	kobject_init(&mq_kobj->kobj, &blk_mq_ktype);
> >  
> >  	for_each_possible_cpu(cpu) {
> > -		ctx = per_cpu_ptr(q->queue_ctx, cpu);
> > +		ctx = kzalloc_node(sizeof(*ctx), GFP_KERNEL, cpu_to_node(cpu));
> > +		if (!ctx)
> > +			goto fail;
> > +		*per_cpu_ptr(q->queue_ctx, cpu) = ctx;
> >  		kobject_init(&ctx->kobj, &blk_mq_ctx_ktype);
> >  	}
> > +	q->mq_kobj = &mq_kobj->kobj;
> > +	return 0;
> > +
> > + fail:
> > +	for_each_possible_cpu(cpu) {
> > +		ctx = *per_cpu_ptr(q->queue_ctx, cpu);
> > +		if (ctx)
> > +			kobject_put(&ctx->kobj);
> > +	}
> > +	kobject_put(&mq_kobj->kobj);
> > +	return -ENOMEM;
> >  }
> 
> 
> blk_mq_kobj looks meaningless, why do we need it, or do I miss something ?

Right, it should have been allocated directly.

> And maybe we should allocate ctx in blk_mq_init_allocated_queue.

Looks either way is fine.


Thanks,
Ming

Re: [PATCH 1/2] blk-mq: not embed .mq_kobj and ctx->kobj into queue instance

[Sasha Levin]

On Fri, Nov 16, 2018 at 11:28:25AM +0800, Ming Lei wrote:
>Even though .mq_kobj, ctx->kobj and q->kobj share same lifetime
>from block layer's view, actually they don't because userspace may
>grab one kobject anytime via sysfs, so each kobject's lifetime has
>to be independent, then the objects(mq_kobj, ctx) which hosts its
>own kobject have to be allocated dynamically.
>
>This patch fixes kernel panic issue during booting when DEBUG_KOBJECT_RELEASE
>is enabled.
>
>Reported-by: Guenter Roeck <linux@roeck-us.net>
>Cc: Guenter Roeck <linux@roeck-us.net>
>Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
>Cc: stable@vger.kernel.org
>Signed-off-by: Ming Lei <ming.lei@redhat.com>

What does this patch depend on? It doesn't apply to Linus's tree nor to
the block tree.

Also, could you please cc lkml with patches?

--
Thanks,
Sasha

Re: [PATCH 1/2] blk-mq: not embed .mq_kobj and ctx->kobj into queue instance

[Ming Lei]

On Fri, Nov 16, 2018 at 01:52:05AM -0500, Sasha Levin wrote:
> On Fri, Nov 16, 2018 at 11:28:25AM +0800, Ming Lei wrote:
> > Even though .mq_kobj, ctx->kobj and q->kobj share same lifetime
> > from block layer's view, actually they don't because userspace may
> > grab one kobject anytime via sysfs, so each kobject's lifetime has
> > to be independent, then the objects(mq_kobj, ctx) which hosts its
> > own kobject have to be allocated dynamically.
> > 
> > This patch fixes kernel panic issue during booting when DEBUG_KOBJECT_RELEASE
> > is enabled.
> > 
> > Reported-by: Guenter Roeck <linux@roeck-us.net>
> > Cc: Guenter Roeck <linux@roeck-us.net>
> > Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
> > Cc: stable@vger.kernel.org
> > Signed-off-by: Ming Lei <ming.lei@redhat.com>
> 
> What does this patch depend on? It doesn't apply to Linus's tree nor to
> the block tree.
> 
> Also, could you please cc lkml with patches?

It depends on for-4.21/block.

thanks,
Ming

Re: [PATCH 1/2] blk-mq: not embed .mq_kobj and ctx->kobj into queue instance

[Sasha Levin]

On Fri, Nov 16, 2018 at 03:49:06PM +0800, Ming Lei wrote:
>On Fri, Nov 16, 2018 at 01:52:05AM -0500, Sasha Levin wrote:
>> On Fri, Nov 16, 2018 at 11:28:25AM +0800, Ming Lei wrote:
>> > Even though .mq_kobj, ctx->kobj and q->kobj share same lifetime
>> > from block layer's view, actually they don't because userspace may
>> > grab one kobject anytime via sysfs, so each kobject's lifetime has
>> > to be independent, then the objects(mq_kobj, ctx) which hosts its
>> > own kobject have to be allocated dynamically.
>> >
>> > This patch fixes kernel panic issue during booting when DEBUG_KOBJECT_RELEASE
>> > is enabled.
>> >
>> > Reported-by: Guenter Roeck <linux@roeck-us.net>
>> > Cc: Guenter Roeck <linux@roeck-us.net>
>> > Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
>> > Cc: stable@vger.kernel.org
>> > Signed-off-by: Ming Lei <ming.lei@redhat.com>
>>
>> What does this patch depend on? It doesn't apply to Linus's tree nor to
>> the block tree.
>>
>> Also, could you please cc lkml with patches?
>
>It depends on for-4.21/block.

And in particular on commit 392546aed22 ("blk-mq: separate number of
hardware queues from nr_cpu_ids") which isn't tagged for stable nor
looks like stable material, so this proposed patch doesn't backport
cleanly to any of the stable trees.

--
Thanks,
Sasha

[PATCH 2/2] blk-mq: alloc q->queue_ctx as normal array

[Ming Lei]

Now q->queue_ctx is just one read-mostly table for query the
'blk_mq_ctx' instance from one cpu index, it isn't necessary
to allocate it as percpu variable. One simple array may be
more efficient.

Cc: Guenter Roeck <linux@roeck-us.net>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Signed-off-by: Ming Lei <ming.lei@redhat.com>
---
 block/blk-mq-sysfs.c   |  6 +++---
 block/blk-mq.c         | 12 +++++++-----
 block/blk-mq.h         |  2 +-
 include/linux/blkdev.h |  2 +-
 4 files changed, 12 insertions(+), 10 deletions(-)

diff --git a/block/blk-mq-sysfs.c b/block/blk-mq-sysfs.c
index bab236955f56..dc4ac733a125 100644
--- a/block/blk-mq-sysfs.c
+++ b/block/blk-mq-sysfs.c
@@ -300,7 +300,7 @@ void blk_mq_sysfs_deinit(struct request_queue *q)
 	int cpu;
 
 	for_each_possible_cpu(cpu) {
-		ctx = *per_cpu_ptr(q->queue_ctx, cpu);
+		ctx = q->queue_ctx[cpu];
 		kobject_put(&ctx->kobj);
 	}
 	kobject_put(q->mq_kobj);
@@ -322,7 +322,7 @@ int blk_mq_sysfs_init(struct request_queue *q)
 		ctx = kzalloc_node(sizeof(*ctx), GFP_KERNEL, cpu_to_node(cpu));
 		if (!ctx)
 			goto fail;
-		*per_cpu_ptr(q->queue_ctx, cpu) = ctx;
+		q->queue_ctx[cpu] = ctx;
 		kobject_init(&ctx->kobj, &blk_mq_ctx_ktype);
 	}
 	q->mq_kobj = &mq_kobj->kobj;
@@ -330,7 +330,7 @@ int blk_mq_sysfs_init(struct request_queue *q)
 
  fail:
 	for_each_possible_cpu(cpu) {
-		ctx = *per_cpu_ptr(q->queue_ctx, cpu);
+		ctx = q->queue_ctx[cpu];
 		if (ctx)
 			kobject_put(&ctx->kobj);
 	}
diff --git a/block/blk-mq.c b/block/blk-mq.c
index 3589ee601f37..20c485e1817a 100644
--- a/block/blk-mq.c
+++ b/block/blk-mq.c
@@ -2299,7 +2299,7 @@ static void blk_mq_init_cpu_queues(struct request_queue *q,
 	unsigned int i, j;
 
 	for_each_possible_cpu(i) {
-		struct blk_mq_ctx *__ctx = *per_cpu_ptr(q->queue_ctx, i);
+		struct blk_mq_ctx *__ctx = q->queue_ctx[i];
 		struct blk_mq_hw_ctx *hctx;
 
 		__ctx->cpu = i;
@@ -2385,7 +2385,7 @@ static void blk_mq_map_swqueue(struct request_queue *q)
 			set->map[0].mq_map[i] = 0;
 		}
 
-		ctx = *per_cpu_ptr(q->queue_ctx, i);
+		ctx = q->queue_ctx[i];
 		for (j = 0; j < set->nr_maps; j++) {
 			hctx = blk_mq_map_queue_type(q, j, i);
 
@@ -2541,7 +2541,7 @@ void blk_mq_release(struct request_queue *q)
 	 */
 	blk_mq_sysfs_deinit(q);
 
-	free_percpu(q->queue_ctx);
+	kfree(q->queue_ctx);
 }
 
 struct request_queue *blk_mq_init_queue(struct blk_mq_tag_set *set)
@@ -2731,7 +2731,9 @@ struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
 	if (!q->poll_cb)
 		goto err_exit;
 
-	q->queue_ctx = alloc_percpu(struct blk_mq_ctx *);
+	q->queue_ctx = kmalloc_array_node(nr_cpu_ids,
+					  sizeof(struct blk_mq_ctx *),
+					  GFP_KERNEL, set->numa_node);
 	if (!q->queue_ctx)
 		goto err_exit;
 
@@ -2798,7 +2800,7 @@ struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
 err_sys_init:
 	blk_mq_sysfs_deinit(q);
 err_percpu:
-	free_percpu(q->queue_ctx);
+	kfree(q->queue_ctx);
 err_exit:
 	q->mq_ops = NULL;
 	return ERR_PTR(-ENOMEM);
diff --git a/block/blk-mq.h b/block/blk-mq.h
index 84898793c230..97829388e1db 100644
--- a/block/blk-mq.h
+++ b/block/blk-mq.h
@@ -129,7 +129,7 @@ static inline enum mq_rq_state blk_mq_rq_state(struct request *rq)
 static inline struct blk_mq_ctx *__blk_mq_get_ctx(struct request_queue *q,
 					   unsigned int cpu)
 {
-	return *per_cpu_ptr(q->queue_ctx, cpu);
+	return q->queue_ctx[cpu];
 }
 
 /*
diff --git a/include/linux/blkdev.h b/include/linux/blkdev.h
index 9e3892bd67fd..9b6ddc5c7a40 100644
--- a/include/linux/blkdev.h
+++ b/include/linux/blkdev.h
@@ -407,7 +407,7 @@ struct request_queue {
 	const struct blk_mq_ops	*mq_ops;
 
 	/* sw queues */
-	struct blk_mq_ctx __percpu	**queue_ctx;
+	struct blk_mq_ctx	**queue_ctx;
 	unsigned int		nr_queues;
 
 	unsigned int		queue_depth;
-- 
2.9.5