[ewg] [PATCH] mlx4_ib: Optimize hugetlab pages support

[Eli Cohen]

Since Linux does not merge adjacent pages into a single scatter entry through
calls to dma_map_sg(), we do this for huge pages which are guaranteed to be
comprised of adjacent natural pages of size PAGE_SIZE. This will result in a
significantly lower number of MTT segments used for registering hugetlb memory
regions.

Signed-off-by: Eli Cohen <eli at mellanox.co.il>
---
I tried this patch and improves memory scalability. Without it,
increasing the amount of memory used cause caused decrease in
throughput. With this patch there was no drop at all - I used a test
based on MPI with 2 processes.


 drivers/infiniband/hw/mlx4/mr.c |   67 ++++++++++++++++++++++++++++++++++----
 1 files changed, 60 insertions(+), 7 deletions(-)

diff --git a/drivers/infiniband/hw/mlx4/mr.c b/drivers/infiniband/hw/mlx4/mr.c
index 8e4d26d..641ea96 100644
--- a/drivers/infiniband/hw/mlx4/mr.c
+++ b/drivers/infiniband/hw/mlx4/mr.c
@@ -119,6 +119,38 @@ out:
 	return err;
 }
 
+int mlx4_ib_umem_write_huge_mtt(struct mlx4_ib_dev *dev, struct mlx4_mtt *mtt,
+			   struct ib_umem *umem, int nhpages)
+{
+	struct ib_umem_chunk *chunk;
+	int j, i, k;
+	dma_addr_t *arr;
+	int err;
+
+	arr = kmalloc(nhpages * sizeof *arr, GFP_KERNEL);
+	if (!arr)
+		return -ENOMEM;
+
+	i = 0;
+	k = 0;
+	list_for_each_entry(chunk, &umem->chunk_list, list)
+		for (j = 0; j < chunk->nmap; ++j, ++k) {
+			if (!(k &  ((1 << (HPAGE_SHIFT - PAGE_SHIFT)) - 1))) {
+				if (sg_dma_len(&chunk->page_list[j]) != PAGE_SIZE) {
+					err = -EAGAIN;
+					goto out;
+				}
+				arr[i++] = sg_dma_address(&chunk->page_list[j]);
+			}
+		}
+
+	err = mlx4_write_mtt(dev->dev, mtt, 0, nhpages, arr);
+
+out:
+	kfree(arr);
+	return err;
+}
+
 struct ib_mr *mlx4_ib_reg_user_mr(struct ib_pd *pd, u64 start, u64 length,
 				  u64 virt_addr, int access_flags,
 				  struct ib_udata *udata)
@@ -128,6 +160,8 @@ struct ib_mr *mlx4_ib_reg_user_mr(struct ib_pd *pd, u64 start, u64 length,
 	int shift;
 	int err;
 	int n;
+	int nhuge;
+	int shift_huge;
 
 	mr = kmalloc(sizeof *mr, GFP_KERNEL);
 	if (!mr)
@@ -142,15 +176,34 @@ struct ib_mr *mlx4_ib_reg_user_mr(struct ib_pd *pd, u64 start, u64 length,
 
 	n = ib_umem_page_count(mr->umem);
 	shift = ilog2(mr->umem->page_size);
-
-	err = mlx4_mr_alloc(dev->dev, to_mpd(pd)->pdn, virt_addr, length,
+	if (mr->umem->hugetlb) {
+		nhuge = ALIGN(n << shift, HPAGE_SIZE) >> HPAGE_SHIFT;
+		shift_huge = HPAGE_SHIFT;
+		err = mlx4_mr_alloc(dev->dev, to_mpd(pd)->pdn, virt_addr, length,
+				    convert_access(access_flags), nhuge, shift_huge, &mr->mmr);
+		if (err)
+			goto err_umem;
+
+		err = mlx4_ib_umem_write_huge_mtt(dev, &mr->mmr.mtt, mr->umem, nhuge);
+		if (err) {
+			if (err != -EAGAIN)
+				goto err_mr;
+			else {
+				mlx4_mr_free(to_mdev(pd->device)->dev, &mr->mmr);
+				goto regular_pages;
+			}
+		}
+	} else {
+regular_pages:
+		err = mlx4_mr_alloc(dev->dev, to_mpd(pd)->pdn, virt_addr, length,
 			    convert_access(access_flags), n, shift, &mr->mmr);
-	if (err)
-		goto err_umem;
+		if (err)
+			goto err_umem;
 
-	err = mlx4_ib_umem_write_mtt(dev, &mr->mmr.mtt, mr->umem);
-	if (err)
-		goto err_mr;
+		err = mlx4_ib_umem_write_mtt(dev, &mr->mmr.mtt, mr->umem);
+		if (err)
+			goto err_mr;
+	}
 
 	err = mlx4_mr_enable(dev->dev, &mr->mmr);
 	if (err)
-- 
1.6.0.5