[PATCH] IB/mlx5: fall back to vmalloc for mlx5_ib_wq

[PATCH] IB/mlx5: fall back to vmalloc for mlx5_ib_wq

[Li Dongyang]

We observed multiple times on our Lustre OSS servers that when
the system memory is fragmented, kmalloc() in create_kernel_qp()
could fail order 4/5 allocations while we still have many free pages.

Fall back to vmalloc to allow the operation to contine, also switch
to kmalloc_array() from kmalloc().

Signed-off-by: Li Dongyang <dongyang.li-FCV4sgi5zeUQrrorzV6ljw@public.gmane.org>
---
 drivers/infiniband/hw/mlx5/qp.c  | 54 +++++++++++++++++++++++++++++-----------
 drivers/infiniband/hw/mlx5/srq.c |  8 ++++--
 2 files changed, 45 insertions(+), 17 deletions(-)

diff --git a/drivers/infiniband/hw/mlx5/qp.c b/drivers/infiniband/hw/mlx5/qp.c
index 0889ff367c86..e662fa5af5bb 100644
--- a/drivers/infiniband/hw/mlx5/qp.c
+++ b/drivers/infiniband/hw/mlx5/qp.c
@@ -959,11 +959,35 @@ static int create_kernel_qp(struct mlx5_ib_dev *dev,
 		goto err_free;
 	}
 
-	qp->sq.wrid = kmalloc(qp->sq.wqe_cnt * sizeof(*qp->sq.wrid), GFP_KERNEL);
-	qp->sq.wr_data = kmalloc(qp->sq.wqe_cnt * sizeof(*qp->sq.wr_data), GFP_KERNEL);
-	qp->rq.wrid = kmalloc(qp->rq.wqe_cnt * sizeof(*qp->rq.wrid), GFP_KERNEL);
-	qp->sq.w_list = kmalloc(qp->sq.wqe_cnt * sizeof(*qp->sq.w_list), GFP_KERNEL);
-	qp->sq.wqe_head = kmalloc(qp->sq.wqe_cnt * sizeof(*qp->sq.wqe_head), GFP_KERNEL);
+	qp->sq.wrid = kmalloc_array(qp->sq.wqe_cnt, sizeof(*qp->sq.wrid),
+				    GFP_KERNEL | __GFP_NOWARN);
+	if (!qp->sq.wrid)
+		qp->sq.wrid = __vmalloc(qp->sq.wqe_cnt * sizeof(*qp->sq.wrid),
+					GFP_KERNEL, PAGE_KERNEL);
+	qp->sq.wr_data = kmalloc_array(qp->sq.wqe_cnt, sizeof(*qp->sq.wr_data),
+				       GFP_KERNEL | __GFP_NOWARN);
+	if (!qp->sq.wr_data)
+		qp->sq.wr_data = __vmalloc(qp->sq.wqe_cnt *
+					   sizeof(*qp->sq.wr_data),
+					   GFP_KERNEL, PAGE_KERNEL);
+	qp->rq.wrid = kmalloc_array(qp->rq.wqe_cnt, sizeof(*qp->rq.wrid),
+				    GFP_KERNEL | __GFP_NOWARN);
+	if (!qp->rq.wrid)
+		qp->rq.wrid = __vmalloc(qp->rq.wqe_cnt * sizeof(*qp->rq.wrid),
+					GFP_KERNEL, PAGE_KERNEL);
+	qp->sq.w_list = kmalloc_array(qp->sq.wqe_cnt, sizeof(*qp->sq.w_list),
+				      GFP_KERNEL | __GFP_NOWARN);
+	if (!qp->sq.w_list)
+		qp->sq.w_list = __vmalloc(qp->sq.wqe_cnt *
+					  sizeof(*qp->sq.w_list),
+					  GFP_KERNEL, PAGE_KERNEL);
+	qp->sq.wqe_head = kmalloc_array(qp->sq.wqe_cnt,
+					sizeof(*qp->sq.wqe_head),
+					GFP_KERNEL | __GFP_NOWARN);
+	if (!qp->sq.wqe_head)
+		qp->sq.wqe_head = __vmalloc(qp->sq.wqe_cnt *
+					    sizeof(*qp->sq.wqe_head),
+					    GFP_KERNEL, PAGE_KERNEL);
 
 	if (!qp->sq.wrid || !qp->sq.wr_data || !qp->rq.wrid ||
 	    !qp->sq.w_list || !qp->sq.wqe_head) {
@@ -975,11 +999,11 @@ static int create_kernel_qp(struct mlx5_ib_dev *dev,
 	return 0;
 
 err_wrid:
-	kfree(qp->sq.wqe_head);
-	kfree(qp->sq.w_list);
-	kfree(qp->sq.wrid);
-	kfree(qp->sq.wr_data);
-	kfree(qp->rq.wrid);
+	kvfree(qp->sq.wqe_head);
+	kvfree(qp->sq.w_list);
+	kvfree(qp->sq.wrid);
+	kvfree(qp->sq.wr_data);
+	kvfree(qp->rq.wrid);
 	mlx5_db_free(dev->mdev, &qp->db);
 
 err_free:
@@ -992,11 +1016,11 @@ static int create_kernel_qp(struct mlx5_ib_dev *dev,
 
 static void destroy_qp_kernel(struct mlx5_ib_dev *dev, struct mlx5_ib_qp *qp)
 {
-	kfree(qp->sq.wqe_head);
-	kfree(qp->sq.w_list);
-	kfree(qp->sq.wrid);
-	kfree(qp->sq.wr_data);
-	kfree(qp->rq.wrid);
+	kvfree(qp->sq.wqe_head);
+	kvfree(qp->sq.w_list);
+	kvfree(qp->sq.wrid);
+	kvfree(qp->sq.wr_data);
+	kvfree(qp->rq.wrid);
 	mlx5_db_free(dev->mdev, &qp->db);
 	mlx5_buf_free(dev->mdev, &qp->buf);
 }
diff --git a/drivers/infiniband/hw/mlx5/srq.c b/drivers/infiniband/hw/mlx5/srq.c
index 43707b101f47..08a91f3ea240 100644
--- a/drivers/infiniband/hw/mlx5/srq.c
+++ b/drivers/infiniband/hw/mlx5/srq.c
@@ -196,7 +196,11 @@ static int create_srq_kernel(struct mlx5_ib_dev *dev, struct mlx5_ib_srq *srq,
 	}
 	mlx5_fill_page_array(&srq->buf, in->pas);
 
-	srq->wrid = kmalloc(srq->msrq.max * sizeof(u64), GFP_KERNEL);
+	srq->wrid = kmalloc_array(srq->msrq.max, sizeof(u64),
+				  GFP_KERNEL | __GFP_NOWARN);
+	if (!srq->wrid)
+		srq->wrid = __vmalloc(srq->msrq.max * sizeof(u64),
+				      GFP_KERNEL, PAGE_KERNEL);
 	if (!srq->wrid) {
 		err = -ENOMEM;
 		goto err_in;
@@ -230,7 +234,7 @@ static void destroy_srq_user(struct ib_pd *pd, struct mlx5_ib_srq *srq)
 
 static void destroy_srq_kernel(struct mlx5_ib_dev *dev, struct mlx5_ib_srq *srq)
 {
-	kfree(srq->wrid);
+	kvfree(srq->wrid);
 	mlx5_buf_free(dev->mdev, &srq->buf);
 	mlx5_db_free(dev->mdev, &srq->db);
 }
-- 
2.14.1

--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" 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] IB/mlx5: fall back to vmalloc for mlx5_ib_wq

[Leon Romanovsky]

[-- Attachment #1: Type: text/plain, Size: 1690 bytes --]

On Wed, Aug 16, 2017 at 09:06:32PM +1000, Li Dongyang wrote:
> We observed multiple times on our Lustre OSS servers that when
> the system memory is fragmented, kmalloc() in create_kernel_qp()
> could fail order 4/5 allocations while we still have many free pages.
>
> Fall back to vmalloc to allow the operation to contine, also switch
> to kmalloc_array() from kmalloc().
>
> Signed-off-by: Li Dongyang <dongyang.li-FCV4sgi5zeUQrrorzV6ljw@public.gmane.org>
> ---
>  drivers/infiniband/hw/mlx5/qp.c  | 54 +++++++++++++++++++++++++++++-----------
>  drivers/infiniband/hw/mlx5/srq.c |  8 ++++--
>  2 files changed, 45 insertions(+), 17 deletions(-)
>
> diff --git a/drivers/infiniband/hw/mlx5/qp.c b/drivers/infiniband/hw/mlx5/qp.c
> index 0889ff367c86..e662fa5af5bb 100644
> --- a/drivers/infiniband/hw/mlx5/qp.c
> +++ b/drivers/infiniband/hw/mlx5/qp.c
> @@ -959,11 +959,35 @@ static int create_kernel_qp(struct mlx5_ib_dev *dev,
>  		goto err_free;
>  	}
>
> -	qp->sq.wrid = kmalloc(qp->sq.wqe_cnt * sizeof(*qp->sq.wrid), GFP_KERNEL);
> -	qp->sq.wr_data = kmalloc(qp->sq.wqe_cnt * sizeof(*qp->sq.wr_data), GFP_KERNEL);
> -	qp->rq.wrid = kmalloc(qp->rq.wqe_cnt * sizeof(*qp->rq.wrid), GFP_KERNEL);
> -	qp->sq.w_list = kmalloc(qp->sq.wqe_cnt * sizeof(*qp->sq.w_list), GFP_KERNEL);
> -	qp->sq.wqe_head = kmalloc(qp->sq.wqe_cnt * sizeof(*qp->sq.wqe_head), GFP_KERNEL);
> +	qp->sq.wrid = kmalloc_array(qp->sq.wqe_cnt, sizeof(*qp->sq.wrid),
> +				    GFP_KERNEL | __GFP_NOWARN);
> +	if (!qp->sq.wrid)
> +		qp->sq.wrid = __vmalloc(qp->sq.wqe_cnt * sizeof(*qp->sq.wrid),
> +					GFP_KERNEL, PAGE_KERNEL);

Please don't open code, the kvmalloc and kvmalloc_array functions.

Thanks for doing it.

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 833 bytes --]

Re: [PATCH] IB/mlx5: fall back to vmalloc for mlx5_ib_wq

[Dongyang Li]

I'm so out of touch with the new interfaces in the kernel.
Will post updated patch.

Thanks
Dongyang
On Wed, 2017-08-16 at 15:12 +0300, Leon Romanovsky wrote:
> On Wed, Aug 16, 2017 at 09:06:32PM +1000, Li Dongyang wrote:
> > We observed multiple times on our Lustre OSS servers that when
> > the system memory is fragmented, kmalloc() in create_kernel_qp()
> > could fail order 4/5 allocations while we still have many free
> > pages.
> > 
> > Fall back to vmalloc to allow the operation to contine, also switch
> > to kmalloc_array() from kmalloc().
> > 
> > Signed-off-by: Li Dongyang <dongyang.li@anu.edu.au>
> > ---
> >  drivers/infiniband/hw/mlx5/qp.c  | 54
> > +++++++++++++++++++++++++++++-----------
> >  drivers/infiniband/hw/mlx5/srq.c |  8 ++++--
> >  2 files changed, 45 insertions(+), 17 deletions(-)
> > 
> > diff --git a/drivers/infiniband/hw/mlx5/qp.c
> > b/drivers/infiniband/hw/mlx5/qp.c
> > index 0889ff367c86..e662fa5af5bb 100644
> > --- a/drivers/infiniband/hw/mlx5/qp.c
> > +++ b/drivers/infiniband/hw/mlx5/qp.c
> > @@ -959,11 +959,35 @@ static int create_kernel_qp(struct
> > mlx5_ib_dev *dev,
> >  		goto err_free;
> >  	}
> > 
> > -	qp->sq.wrid = kmalloc(qp->sq.wqe_cnt * sizeof(*qp-
> > >sq.wrid), GFP_KERNEL);
> > -	qp->sq.wr_data = kmalloc(qp->sq.wqe_cnt * sizeof(*qp-
> > >sq.wr_data), GFP_KERNEL);
> > -	qp->rq.wrid = kmalloc(qp->rq.wqe_cnt * sizeof(*qp-
> > >rq.wrid), GFP_KERNEL);
> > -	qp->sq.w_list = kmalloc(qp->sq.wqe_cnt * sizeof(*qp-
> > >sq.w_list), GFP_KERNEL);
> > -	qp->sq.wqe_head = kmalloc(qp->sq.wqe_cnt * sizeof(*qp-
> > >sq.wqe_head), GFP_KERNEL);
> > +	qp->sq.wrid = kmalloc_array(qp->sq.wqe_cnt, sizeof(*qp-
> > >sq.wrid),
> > +				    GFP_KERNEL | __GFP_NOWARN);
> > +	if (!qp->sq.wrid)
> > +		qp->sq.wrid = __vmalloc(qp->sq.wqe_cnt *
> > sizeof(*qp->sq.wrid),
> > +					GFP_KERNEL, PAGE_KERNEL);
> 
> Please don't open code, the kvmalloc and kvmalloc_array functions.
> 
> Thanks for doing it.