summaryrefslogtreecommitdiff
diff options
context:
space:
mode:
authorYishai Hadas <yishaih@nvidia.com>2025-08-13 15:36:01 +0300
committerJason Gunthorpe <jgg@nvidia.com>2025-08-25 14:40:29 -0300
commitd9e6e85b7beb2aeb8defac2f705b23532ddb25d4 (patch)
tree69b8997d146e54ad8c4dbb18372b71670ca50d66
parent8e5442f3935d269f15b0b9a0d77cdf1cf3bd006f (diff)
RDMA/mlx5: Enable Data-Direct with Relaxed Ordering
Relaxed Ordering can improve performance in certain scenarios. Enable it in the Data-Direct use case as well. Link: https://patch.msgid.link/r/1221dcdda8061ba5f6bc3519044083c7438b257e.1755088503.git.leon@kernel.org Signed-off-by: Yishai Hadas <yishaih@nvidia.com> Reviewed-by: Gal Shalom <galshalom@Nvidia.com> Signed-off-by: Leon Romanovsky <leonro@nvidia.com> Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
-rw-r--r--drivers/infiniband/hw/mlx5/main.c32
-rw-r--r--drivers/infiniband/hw/mlx5/mlx5_ib.h2
-rw-r--r--drivers/infiniband/hw/mlx5/mr.c8
-rw-r--r--drivers/infiniband/hw/mlx5/umr.c6
4 files changed, 41 insertions, 7 deletions
diff --git a/drivers/infiniband/hw/mlx5/main.c b/drivers/infiniband/hw/mlx5/main.c
index d456e4fde3e1..505349c68e79 100644
--- a/drivers/infiniband/hw/mlx5/main.c
+++ b/drivers/infiniband/hw/mlx5/main.c
@@ -3118,6 +3118,7 @@ mlx5_ib_create_data_direct_resources(struct mlx5_ib_dev *dev)
{
int inlen = MLX5_ST_SZ_BYTES(create_mkey_in);
struct mlx5_core_dev *mdev = dev->mdev;
+ bool ro_supp = false;
void *mkc;
u32 mkey;
u32 pdn;
@@ -3146,14 +3147,37 @@ mlx5_ib_create_data_direct_resources(struct mlx5_ib_dev *dev)
MLX5_SET(mkc, mkc, length64, 1);
MLX5_SET(mkc, mkc, qpn, 0xffffff);
err = mlx5_core_create_mkey(mdev, &mkey, in, inlen);
- kvfree(in);
if (err)
- goto err;
+ goto err_mkey;
dev->ddr.mkey = mkey;
dev->ddr.pdn = pdn;
+
+ /* create another mkey with RO support */
+ if (MLX5_CAP_GEN(dev->mdev, relaxed_ordering_write)) {
+ MLX5_SET(mkc, mkc, relaxed_ordering_write, 1);
+ ro_supp = true;
+ }
+
+ if (MLX5_CAP_GEN(dev->mdev, relaxed_ordering_read)) {
+ MLX5_SET(mkc, mkc, relaxed_ordering_read, 1);
+ ro_supp = true;
+ }
+
+ if (ro_supp) {
+ err = mlx5_core_create_mkey(mdev, &mkey, in, inlen);
+ /* RO is defined as best effort */
+ if (!err) {
+ dev->ddr.mkey_ro = mkey;
+ dev->ddr.mkey_ro_valid = true;
+ }
+ }
+
+ kvfree(in);
return 0;
+err_mkey:
+ kvfree(in);
err:
mlx5_core_dealloc_pd(mdev, pdn);
return err;
@@ -3162,6 +3186,10 @@ err:
static void
mlx5_ib_free_data_direct_resources(struct mlx5_ib_dev *dev)
{
+
+ if (dev->ddr.mkey_ro_valid)
+ mlx5_core_destroy_mkey(dev->mdev, dev->ddr.mkey_ro);
+
mlx5_core_destroy_mkey(dev->mdev, dev->ddr.mkey);
mlx5_core_dealloc_pd(dev->mdev, dev->ddr.pdn);
}
diff --git a/drivers/infiniband/hw/mlx5/mlx5_ib.h b/drivers/infiniband/hw/mlx5/mlx5_ib.h
index 7ffc7ee92cf0..309f256fc27d 100644
--- a/drivers/infiniband/hw/mlx5/mlx5_ib.h
+++ b/drivers/infiniband/hw/mlx5/mlx5_ib.h
@@ -854,6 +854,8 @@ struct mlx5_ib_port_resources {
struct mlx5_data_direct_resources {
u32 pdn;
u32 mkey;
+ u32 mkey_ro;
+ u8 mkey_ro_valid :1;
};
struct mlx5_ib_resources {
diff --git a/drivers/infiniband/hw/mlx5/mr.c b/drivers/infiniband/hw/mlx5/mr.c
index 1317f2cb38a4..d3c82ffa300e 100644
--- a/drivers/infiniband/hw/mlx5/mr.c
+++ b/drivers/infiniband/hw/mlx5/mr.c
@@ -1717,11 +1717,11 @@ reg_user_mr_dmabuf_by_data_direct(struct ib_pd *pd, u64 offset,
goto end;
}
- /* The device's 'data direct mkey' was created without RO flags to
- * simplify things and allow for a single mkey per device.
- * Since RO is not a must, mask it out accordingly.
+ /* If no device's 'data direct mkey' with RO flags exists
+ * mask it out accordingly.
*/
- access_flags &= ~IB_ACCESS_RELAXED_ORDERING;
+ if (!dev->ddr.mkey_ro_valid)
+ access_flags &= ~IB_ACCESS_RELAXED_ORDERING;
crossed_mr = reg_user_mr_dmabuf(pd, &data_direct_dev->pdev->dev,
offset, length, virt_addr, fd,
access_flags, MLX5_MKC_ACCESS_MODE_KSM,
diff --git a/drivers/infiniband/hw/mlx5/umr.c b/drivers/infiniband/hw/mlx5/umr.c
index 7ef35cddce81..4e562e0dd9e1 100644
--- a/drivers/infiniband/hw/mlx5/umr.c
+++ b/drivers/infiniband/hw/mlx5/umr.c
@@ -761,7 +761,11 @@ _mlx5r_umr_update_mr_pas(struct mlx5_ib_mr *mr, unsigned int flags, bool dd,
if (dd) {
cur_ksm->va = cpu_to_be64(rdma_block_iter_dma_address(&biter));
- cur_ksm->key = cpu_to_be32(dev->ddr.mkey);
+ if (mr->access_flags & IB_ACCESS_RELAXED_ORDERING &&
+ dev->ddr.mkey_ro_valid)
+ cur_ksm->key = cpu_to_be32(dev->ddr.mkey_ro);
+ else
+ cur_ksm->key = cpu_to_be32(dev->ddr.mkey);
if (mr->umem->is_dmabuf &&
(flags & MLX5_IB_UPD_XLT_ZAP)) {
cur_ksm->va = 0;