[ewg] Re: [PATCH v3] mlx4_ib: Optimize hugetlab pages support

[Yossi Etigin]

Eli Cohen wrote:
> Since Linux may not merge adjacent pages into a single scatter entry through
> calls to dma_map_sg(), we check the special case of hugetlb pages which are
> likely to be mapped to coniguous dma addresses and if they are, take advantage
> of this. This will result in a significantly lower number of MTT segments used
> for registering hugetlb memory regions.
> 

How about the one below - it fixes bugzilla #1569 (fix mapping for size that is not
on page boundary):

---

Since Linux may not merge adjacent pages into a single scatter entry through
calls to dma_map_sg(), we check the special case of hugetlb pages which are
likely to be mapped to coniguous dma addresses and if they are, take advantage
of this. 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>
---
 drivers/infiniband/hw/mlx4/mr.c |   81 ++++++++++++++++++++++++++++++++++----
 1 files changed, 72 insertions(+), 9 deletions(-)

Index: b/drivers/infiniband/hw/mlx4/mr.c
===================================================================
--- a/drivers/infiniband/hw/mlx4/mr.c	2008-11-19 21:32:15.000000000 +0200
+++ b/drivers/infiniband/hw/mlx4/mr.c	2009-03-30 18:29:55.000000000 +0300
@@ -119,6 +119,70 @@ out:
 	return err;
 }
 
+static int handle_hugetlb_user_mr(struct ib_pd *pd, struct mlx4_ib_mr *mr,
+				  u64 start, u64 virt_addr, int access_flags)
+{
+#if defined(CONFIG_HUGETLB_PAGE) && !defined(__powerpc__) && !defined(__ia64__)
+	struct mlx4_ib_dev *dev = to_mdev(pd->device);
+	struct ib_umem_chunk *chunk;
+	unsigned dsize;
+	dma_addr_t daddr;
+	unsigned cur_size = 0;
+	dma_addr_t uninitialized_var(cur_addr);
+	int n;
+	struct ib_umem	*umem = mr->umem;
+	u64 *arr;
+	int err = 0;
+	int i;
+	int j = 0;
+	int off = start & (HPAGE_SIZE - 1);
+
+	n = DIV_ROUND_UP(off + umem->length, HPAGE_SIZE);
+	arr = kmalloc(n * sizeof *arr, GFP_KERNEL);
+	if (!arr)
+		return -ENOMEM;
+
+	list_for_each_entry(chunk, &umem->chunk_list, list)
+		for (i = 0; i < chunk->nmap; ++i) {
+			daddr = sg_dma_address(&chunk->page_list[i]);
+			dsize = sg_dma_len(&chunk->page_list[i]);
+			if (!cur_size) {
+				cur_addr = daddr;
+				cur_size = dsize;
+			} else if (cur_addr + cur_size != daddr) {
+				err = -EINVAL;
+				goto out;
+			} else
+				cur_size += dsize;
+
+			if (cur_size > HPAGE_SIZE) {
+				err = -EINVAL;
+				goto out;
+			} else if (cur_size == HPAGE_SIZE) {
+				cur_size = 0;
+				arr[j++] = cur_addr;
+			}
+		}
+
+	if (cur_size) {
+		arr[j++] = cur_addr;
+	}
+
+	err = mlx4_mr_alloc(dev->dev, to_mpd(pd)->pdn, virt_addr, umem->length,
+			    convert_access(access_flags), n, HPAGE_SHIFT, &mr->mmr);
+	if (err)
+		goto out;
+
+	err = mlx4_write_mtt(dev->dev, &mr->mmr.mtt, 0, n, arr);
+
+out:
+	kfree(arr);
+	return err;
+#else
+	return -ENOSYS;
+#endif
+}
+
 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)
@@ -140,17 +204,20 @@ struct ib_mr *mlx4_ib_reg_user_mr(struct
 		goto err_free;
 	}
 
-	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,
-			    convert_access(access_flags), n, shift, &mr->mmr);
-	if (err)
-		goto err_umem;
-
-	err = mlx4_ib_umem_write_mtt(dev, &mr->mmr.mtt, mr->umem);
-	if (err)
-		goto err_mr;
+	if (!mr->umem->hugetlb ||
+	    handle_hugetlb_user_mr(pd, mr, start, virt_addr, access_flags)) {
+		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,
+				    convert_access(access_flags), n, shift, &mr->mmr);
+		if (err)
+			goto err_umem;
+
+		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)