[PATCH] blk-mq: fix a memory leak bug

[PATCH] blk-mq: fix a memory leak bug

[Wenwen Wang]

From: Wenwen Wang <wenwen@cs.uga.edu>

In blk_mq_init_allocated_queue(), a kernel buffer is allocated through
kcalloc_node() to hold hardware dispatch queues in the request queue 'q',
i.e., 'q->queue_hw_ctx'.  Later on, if the blk-mq device has no scheduler
set, a scheduler will be initialized through elevator_init_mq(). If this
initialization fails, blk_mq_init_allocated_queue() needs to be terminated
with an error code returned to indicate this failure. However, the
allocated buffer is not freed on this execution path, leading to a memory
leak bug. Moreover, the required cleanup work is also missed on this path.

To fix the above issues, free the allocated buffer and invoke the cleanup
functions.

Signed-off-by: Wenwen Wang <wenwen@cs.uga.edu>
---
 block/blk-mq.c | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)

diff --git a/block/blk-mq.c b/block/blk-mq.c
index e5ef40c..04fe077 100644
--- a/block/blk-mq.c
+++ b/block/blk-mq.c
@@ -2845,6 +2845,8 @@ static unsigned int nr_hw_queues(struct blk_mq_tag_set *set)
 struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
 						  struct request_queue *q)
 {
+	int ret = -ENOMEM;
+
 	/* mark the queue as mq asap */
 	q->mq_ops = set->ops;
 
@@ -2906,11 +2908,9 @@ struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
 	blk_mq_map_swqueue(q);
 
 	if (!(set->flags & BLK_MQ_F_NO_SCHED)) {
-		int ret;
-
 		ret = elevator_init_mq(q);
 		if (ret)
-			return ERR_PTR(ret);
+			goto err_hctxs;
 	}
 
 	return q;
@@ -2924,7 +2924,7 @@ struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
 	q->poll_cb = NULL;
 err_exit:
 	q->mq_ops = NULL;
-	return ERR_PTR(-ENOMEM);
+	return ERR_PTR(ret);
 }
 EXPORT_SYMBOL(blk_mq_init_allocated_queue);
 
-- 
2.7.4

Re: [PATCH] blk-mq: fix a memory leak bug

[Ming Lei]

On Sun, Jul 14, 2019 at 3:04 PM Wenwen Wang <wang6495@umn.edu> wrote:
>
> From: Wenwen Wang <wenwen@cs.uga.edu>
>
> In blk_mq_init_allocated_queue(), a kernel buffer is allocated through
> kcalloc_node() to hold hardware dispatch queues in the request queue 'q',
> i.e., 'q->queue_hw_ctx'.  Later on, if the blk-mq device has no scheduler
> set, a scheduler will be initialized through elevator_init_mq(). If this
> initialization fails, blk_mq_init_allocated_queue() needs to be terminated
> with an error code returned to indicate this failure. However, the
> allocated buffer is not freed on this execution path, leading to a memory
> leak bug. Moreover, the required cleanup work is also missed on this path.
>
> To fix the above issues, free the allocated buffer and invoke the cleanup
> functions.
>
> Signed-off-by: Wenwen Wang <wenwen@cs.uga.edu>
> ---
>  block/blk-mq.c | 8 ++++----
>  1 file changed, 4 insertions(+), 4 deletions(-)
>
> diff --git a/block/blk-mq.c b/block/blk-mq.c
> index e5ef40c..04fe077 100644
> --- a/block/blk-mq.c
> +++ b/block/blk-mq.c
> @@ -2845,6 +2845,8 @@ static unsigned int nr_hw_queues(struct blk_mq_tag_set *set)
>  struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
>                                                   struct request_queue *q)
>  {
> +       int ret = -ENOMEM;
> +

The above isn't necessary because the function always returns
ERR_PTR(-ENOMEM) in case of failure.

>         /* mark the queue as mq asap */
>         q->mq_ops = set->ops;
>
> @@ -2906,11 +2908,9 @@ struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
>         blk_mq_map_swqueue(q);
>
>         if (!(set->flags & BLK_MQ_F_NO_SCHED)) {
> -               int ret;
> -
>                 ret = elevator_init_mq(q);
>                 if (ret)
> -                       return ERR_PTR(ret);
> +                       goto err_hctxs;

The above change itself is fine.

However, elevator_init_mq() shouldn't return failure since none should
work any time.
That said 'none' should be fallback to in case that default
mq-deadline can't be initialized.

thanks,
Ming Lei

Re: [PATCH] blk-mq: fix a memory leak bug

[Wenwen Wang]

On Sun, Jul 14, 2019 at 8:26 PM Ming Lei <tom.leiming@gmail.com> wrote:
>
> On Sun, Jul 14, 2019 at 3:04 PM Wenwen Wang <wang6495@umn.edu> wrote:
> >
> > From: Wenwen Wang <wenwen@cs.uga.edu>
> >
> > In blk_mq_init_allocated_queue(), a kernel buffer is allocated through
> > kcalloc_node() to hold hardware dispatch queues in the request queue 'q',
> > i.e., 'q->queue_hw_ctx'.  Later on, if the blk-mq device has no scheduler
> > set, a scheduler will be initialized through elevator_init_mq(). If this
> > initialization fails, blk_mq_init_allocated_queue() needs to be terminated
> > with an error code returned to indicate this failure. However, the
> > allocated buffer is not freed on this execution path, leading to a memory
> > leak bug. Moreover, the required cleanup work is also missed on this path.
> >
> > To fix the above issues, free the allocated buffer and invoke the cleanup
> > functions.
> >
> > Signed-off-by: Wenwen Wang <wenwen@cs.uga.edu>
> > ---
> >  block/blk-mq.c | 8 ++++----
> >  1 file changed, 4 insertions(+), 4 deletions(-)
> >
> > diff --git a/block/blk-mq.c b/block/blk-mq.c
> > index e5ef40c..04fe077 100644
> > --- a/block/blk-mq.c
> > +++ b/block/blk-mq.c
> > @@ -2845,6 +2845,8 @@ static unsigned int nr_hw_queues(struct blk_mq_tag_set *set)
> >  struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
> >                                                   struct request_queue *q)
> >  {
> > +       int ret = -ENOMEM;
> > +
>
> The above isn't necessary because the function always returns
> ERR_PTR(-ENOMEM) in case of failure.
>
> >         /* mark the queue as mq asap */
> >         q->mq_ops = set->ops;
> >
> > @@ -2906,11 +2908,9 @@ struct request_queue *blk_mq_init_allocated_queue(struct blk_mq_tag_set *set,
> >         blk_mq_map_swqueue(q);
> >
> >         if (!(set->flags & BLK_MQ_F_NO_SCHED)) {
> > -               int ret;
> > -
> >                 ret = elevator_init_mq(q);
> >                 if (ret)
> > -                       return ERR_PTR(ret);
> > +                       goto err_hctxs;
>
> The above change itself is fine.
>
> However, elevator_init_mq() shouldn't return failure since none should
> work any time.
> That said 'none' should be fallback to in case that default
> mq-deadline can't be initialized.

Thanks for your comments! I agree that 'none' is the fallback if
'mq-deadline' cannot be initialized.

But, the error-handling branch after elevator_init_mq() is still
necessary, unless elevator_init_mq() always returns zero, which is not
true.

Thanks!
Wenwen