summaryrefslogtreecommitdiff
path: root/drivers/infiniband/hw/mlx5
diff options
context:
space:
mode:
Diffstat (limited to 'drivers/infiniband/hw/mlx5')
-rw-r--r--drivers/infiniband/hw/mlx5/data_direct.c2
-rw-r--r--drivers/infiniband/hw/mlx5/gsi.c15
-rw-r--r--drivers/infiniband/hw/mlx5/main.c113
-rw-r--r--drivers/infiniband/hw/mlx5/mlx5_ib.h7
-rw-r--r--drivers/infiniband/hw/mlx5/mr.c11
-rw-r--r--drivers/infiniband/hw/mlx5/umr.c6
6 files changed, 129 insertions, 25 deletions
diff --git a/drivers/infiniband/hw/mlx5/data_direct.c b/drivers/infiniband/hw/mlx5/data_direct.c
index b9ba84afaae2..b81ac5709b56 100644
--- a/drivers/infiniband/hw/mlx5/data_direct.c
+++ b/drivers/infiniband/hw/mlx5/data_direct.c
@@ -35,7 +35,7 @@ static int mlx5_data_direct_vpd_get_vuid(struct mlx5_data_direct_dev *dev)
vpd_data = pci_vpd_alloc(pdev, &vpd_size);
if (IS_ERR(vpd_data)) {
- pci_err(pdev, "Unable to read VPD, err=%ld\n", PTR_ERR(vpd_data));
+ pci_err(pdev, "Unable to read VPD, err=%pe\n", vpd_data);
return PTR_ERR(vpd_data);
}
diff --git a/drivers/infiniband/hw/mlx5/gsi.c b/drivers/infiniband/hw/mlx5/gsi.c
index b804f2dd5628..d5487834ed25 100644
--- a/drivers/infiniband/hw/mlx5/gsi.c
+++ b/drivers/infiniband/hw/mlx5/gsi.c
@@ -131,8 +131,9 @@ int mlx5_ib_create_gsi(struct ib_pd *pd, struct mlx5_ib_qp *mqp,
gsi->cq = ib_alloc_cq(pd->device, gsi, attr->cap.max_send_wr, 0,
IB_POLL_SOFTIRQ);
if (IS_ERR(gsi->cq)) {
- mlx5_ib_warn(dev, "unable to create send CQ for GSI QP. error %ld\n",
- PTR_ERR(gsi->cq));
+ mlx5_ib_warn(dev,
+ "unable to create send CQ for GSI QP. error %pe\n",
+ gsi->cq);
ret = PTR_ERR(gsi->cq);
goto err_free_wrs;
}
@@ -147,8 +148,9 @@ int mlx5_ib_create_gsi(struct ib_pd *pd, struct mlx5_ib_qp *mqp,
gsi->rx_qp = ib_create_qp(pd, &hw_init_attr);
if (IS_ERR(gsi->rx_qp)) {
- mlx5_ib_warn(dev, "unable to create hardware GSI QP. error %ld\n",
- PTR_ERR(gsi->rx_qp));
+ mlx5_ib_warn(dev,
+ "unable to create hardware GSI QP. error %pe\n",
+ gsi->rx_qp);
ret = PTR_ERR(gsi->rx_qp);
goto err_destroy_cq;
}
@@ -294,8 +296,9 @@ static void setup_qp(struct mlx5_ib_gsi_qp *gsi, u16 qp_index)
qp = create_gsi_ud_qp(gsi);
if (IS_ERR(qp)) {
- mlx5_ib_warn(dev, "unable to create hardware UD QP for GSI: %ld\n",
- PTR_ERR(qp));
+ mlx5_ib_warn(dev,
+ "unable to create hardware UD QP for GSI: %pe\n",
+ qp);
return;
}
diff --git a/drivers/infiniband/hw/mlx5/main.c b/drivers/infiniband/hw/mlx5/main.c
index d456e4fde3e1..fc1e86f6c409 100644
--- a/drivers/infiniband/hw/mlx5/main.c
+++ b/drivers/infiniband/hw/mlx5/main.c
@@ -13,6 +13,7 @@
#include <linux/dma-mapping.h>
#include <linux/slab.h>
#include <linux/bitmap.h>
+#include <linux/log2.h>
#include <linux/sched.h>
#include <linux/sched/mm.h>
#include <linux/sched/task.h>
@@ -883,6 +884,51 @@ static void fill_esw_mgr_reg_c0(struct mlx5_core_dev *mdev,
resp->reg_c0.mask = mlx5_eswitch_get_vport_metadata_mask();
}
+/*
+ * Calculate maximum SQ overhead across all QP types.
+ * Other QP types (REG_UMR, UC, RC, UD/SMI/GSI, XRC_TGT)
+ * have smaller overhead than the types calculated below,
+ * so they are implicitly included.
+ */
+static u32 mlx5_ib_calc_max_sq_overhead(void)
+{
+ u32 max_overhead_xrc, overhead_ud_lso, a, b;
+
+ /* XRC_INI */
+ max_overhead_xrc = sizeof(struct mlx5_wqe_xrc_seg);
+ max_overhead_xrc += sizeof(struct mlx5_wqe_ctrl_seg);
+ a = sizeof(struct mlx5_wqe_atomic_seg) +
+ sizeof(struct mlx5_wqe_raddr_seg);
+ b = sizeof(struct mlx5_wqe_umr_ctrl_seg) +
+ sizeof(struct mlx5_mkey_seg) +
+ MLX5_IB_SQ_UMR_INLINE_THRESHOLD / MLX5_IB_UMR_OCTOWORD;
+ max_overhead_xrc += max(a, b);
+
+ /* UD with LSO */
+ overhead_ud_lso = sizeof(struct mlx5_wqe_ctrl_seg);
+ overhead_ud_lso += sizeof(struct mlx5_wqe_eth_pad);
+ overhead_ud_lso += sizeof(struct mlx5_wqe_eth_seg);
+ overhead_ud_lso += sizeof(struct mlx5_wqe_datagram_seg);
+
+ return max(max_overhead_xrc, overhead_ud_lso);
+}
+
+static u32 mlx5_ib_calc_max_qp_wr(struct mlx5_ib_dev *dev)
+{
+ struct mlx5_core_dev *mdev = dev->mdev;
+ u32 max_wqe_bb_units = 1 << MLX5_CAP_GEN(mdev, log_max_qp_sz);
+ u32 max_wqe_size;
+ /* max QP overhead + 1 SGE, no inline, no special features */
+ max_wqe_size = mlx5_ib_calc_max_sq_overhead() +
+ sizeof(struct mlx5_wqe_data_seg);
+
+ max_wqe_size = roundup_pow_of_two(max_wqe_size);
+
+ max_wqe_size = ALIGN(max_wqe_size, MLX5_SEND_WQE_BB);
+
+ return (max_wqe_bb_units * MLX5_SEND_WQE_BB) / max_wqe_size;
+}
+
static int mlx5_ib_query_device(struct ib_device *ibdev,
struct ib_device_attr *props,
struct ib_udata *uhw)
@@ -1041,7 +1087,7 @@ static int mlx5_ib_query_device(struct ib_device *ibdev,
props->max_mr_size = ~0ull;
props->page_size_cap = ~(min_page_size - 1);
props->max_qp = 1 << MLX5_CAP_GEN(mdev, log_max_qp);
- props->max_qp_wr = 1 << MLX5_CAP_GEN(mdev, log_max_qp_sz);
+ props->max_qp_wr = mlx5_ib_calc_max_qp_wr(dev);
max_rq_sg = MLX5_CAP_GEN(mdev, max_wqe_sz_rq) /
sizeof(struct mlx5_wqe_data_seg);
max_sq_desc = min_t(int, MLX5_CAP_GEN(mdev, max_wqe_sz_sq), 512);
@@ -1793,7 +1839,8 @@ static void deallocate_uars(struct mlx5_ib_dev *dev,
}
static int mlx5_ib_enable_lb_mp(struct mlx5_core_dev *master,
- struct mlx5_core_dev *slave)
+ struct mlx5_core_dev *slave,
+ struct mlx5_ib_lb_state *lb_state)
{
int err;
@@ -1805,6 +1852,7 @@ static int mlx5_ib_enable_lb_mp(struct mlx5_core_dev *master,
if (err)
goto out;
+ lb_state->force_enable = true;
return 0;
out:
@@ -1813,16 +1861,22 @@ out:
}
static void mlx5_ib_disable_lb_mp(struct mlx5_core_dev *master,
- struct mlx5_core_dev *slave)
+ struct mlx5_core_dev *slave,
+ struct mlx5_ib_lb_state *lb_state)
{
mlx5_nic_vport_update_local_lb(slave, false);
mlx5_nic_vport_update_local_lb(master, false);
+
+ lb_state->force_enable = false;
}
int mlx5_ib_enable_lb(struct mlx5_ib_dev *dev, bool td, bool qp)
{
int err = 0;
+ if (dev->lb.force_enable)
+ return 0;
+
mutex_lock(&dev->lb.mutex);
if (td)
dev->lb.user_td++;
@@ -1844,6 +1898,9 @@ int mlx5_ib_enable_lb(struct mlx5_ib_dev *dev, bool td, bool qp)
void mlx5_ib_disable_lb(struct mlx5_ib_dev *dev, bool td, bool qp)
{
+ if (dev->lb.force_enable)
+ return;
+
mutex_lock(&dev->lb.mutex);
if (td)
dev->lb.user_td--;
@@ -2994,14 +3051,16 @@ int mlx5_ib_dev_res_cq_init(struct mlx5_ib_dev *dev)
pd = ib_alloc_pd(ibdev, 0);
if (IS_ERR(pd)) {
ret = PTR_ERR(pd);
- mlx5_ib_err(dev, "Couldn't allocate PD for res init, err=%d\n", ret);
+ mlx5_ib_err(dev, "Couldn't allocate PD for res init, err=%pe\n",
+ pd);
goto unlock;
}
cq = ib_create_cq(ibdev, NULL, NULL, NULL, &cq_attr);
if (IS_ERR(cq)) {
ret = PTR_ERR(cq);
- mlx5_ib_err(dev, "Couldn't create CQ for res init, err=%d\n", ret);
+ mlx5_ib_err(dev, "Couldn't create CQ for res init, err=%pe\n",
+ cq);
ib_dealloc_pd(pd);
goto unlock;
}
@@ -3045,7 +3104,9 @@ int mlx5_ib_dev_res_srq_init(struct mlx5_ib_dev *dev)
s0 = ib_create_srq(devr->p0, &attr);
if (IS_ERR(s0)) {
ret = PTR_ERR(s0);
- mlx5_ib_err(dev, "Couldn't create SRQ 0 for res init, err=%d\n", ret);
+ mlx5_ib_err(dev,
+ "Couldn't create SRQ 0 for res init, err=%pe\n",
+ s0);
goto unlock;
}
@@ -3057,7 +3118,9 @@ int mlx5_ib_dev_res_srq_init(struct mlx5_ib_dev *dev)
s1 = ib_create_srq(devr->p0, &attr);
if (IS_ERR(s1)) {
ret = PTR_ERR(s1);
- mlx5_ib_err(dev, "Couldn't create SRQ 1 for res init, err=%d\n", ret);
+ mlx5_ib_err(dev,
+ "Couldn't create SRQ 1 for res init, err=%pe\n",
+ s1);
ib_destroy_srq(s0);
}
@@ -3118,6 +3181,7 @@ 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 +3210,37 @@ 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 +3249,10 @@ 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);
}
@@ -3523,7 +3614,7 @@ static void mlx5_ib_unbind_slave_port(struct mlx5_ib_dev *ibdev,
lockdep_assert_held(&mlx5_ib_multiport_mutex);
- mlx5_ib_disable_lb_mp(ibdev->mdev, mpi->mdev);
+ mlx5_ib_disable_lb_mp(ibdev->mdev, mpi->mdev, &ibdev->lb);
mlx5_core_mp_event_replay(ibdev->mdev,
MLX5_DRIVER_EVENT_AFFILIATION_REMOVED,
@@ -3620,7 +3711,7 @@ static bool mlx5_ib_bind_slave_port(struct mlx5_ib_dev *ibdev,
MLX5_DRIVER_EVENT_AFFILIATION_DONE,
&key);
- err = mlx5_ib_enable_lb_mp(ibdev->mdev, mpi->mdev);
+ err = mlx5_ib_enable_lb_mp(ibdev->mdev, mpi->mdev, &ibdev->lb);
if (err)
goto unbind;
diff --git a/drivers/infiniband/hw/mlx5/mlx5_ib.h b/drivers/infiniband/hw/mlx5/mlx5_ib.h
index 7ffc7ee92cf0..09d82d5f95e3 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 {
@@ -1109,6 +1111,7 @@ struct mlx5_ib_lb_state {
u32 user_td;
int qps;
bool enabled;
+ bool force_enable;
};
struct mlx5_ib_pf_eq {
@@ -1802,6 +1805,10 @@ mlx5_umem_mkc_find_best_pgsz(struct mlx5_ib_dev *dev, struct ib_umem *umem,
bitmap = GENMASK_ULL(max_log_entity_size_cap, min_log_entity_size_cap);
+ /* In KSM mode HW requires IOVA and mkey's page size to be aligned */
+ if (access_mode == MLX5_MKC_ACCESS_MODE_KSM && iova)
+ bitmap &= GENMASK_ULL(__ffs64(iova), 0);
+
return ib_umem_find_best_pgsz(umem, bitmap, iova);
}
diff --git a/drivers/infiniband/hw/mlx5/mr.c b/drivers/infiniband/hw/mlx5/mr.c
index 1317f2cb38a4..325fa04cbe8a 100644
--- a/drivers/infiniband/hw/mlx5/mr.c
+++ b/drivers/infiniband/hw/mlx5/mr.c
@@ -1652,8 +1652,7 @@ reg_user_mr_dmabuf(struct ib_pd *pd, struct device *dma_device,
fd, access_flags);
if (IS_ERR(umem_dmabuf)) {
- mlx5_ib_dbg(dev, "umem_dmabuf get failed (%ld)\n",
- PTR_ERR(umem_dmabuf));
+ mlx5_ib_dbg(dev, "umem_dmabuf get failed (%pe)\n", umem_dmabuf);
return ERR_CAST(umem_dmabuf);
}
@@ -1717,11 +1716,11 @@ 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;
generated by cgit (git 2.34.1) at 2025-10-04 20:59:06 +0000