In order to send GSI MADs on different P_Keys, mlx5 needs different QPs to
be created, each with a different P_Key set when the QP is modified to the
INIT state.
Create QPs for each non-zero P_Key in the P_Key table.
Signed-off-by: Haggai Eran <haggaie@xxxxxxxxxxxx>
---
drivers/infiniband/hw/mlx5/gsi.c | 170 ++++++++++++++++++++++++++++++++++++++-
1 file changed, 167 insertions(+), 3 deletions(-)
diff --git a/drivers/infiniband/hw/mlx5/gsi.c b/drivers/infiniband/hw/mlx5/gsi.c
index 711655400ca5..91bd20eb59b7 100644
--- a/drivers/infiniband/hw/mlx5/gsi.c
+++ b/drivers/infiniband/hw/mlx5/gsi.c
@@ -40,6 +40,12 @@ struct mlx5_ib_gsi_qp {
enum ib_sig_type sq_sig_type;
/* Serialize qp state modifications */
struct mutex mutex;
+ int num_qps;
+ /* Protects access to the tx_qps. Post send operations synchronize
+ * with tx_qp creation in setup_qp().
+ */
+ spinlock_t lock;
+ struct ib_qp **tx_qps;
};
static struct mlx5_ib_gsi_qp *gsi_qp(struct ib_qp *qp)
@@ -47,6 +53,11 @@ static struct mlx5_ib_gsi_qp *gsi_qp(struct ib_qp *qp)
return container_of(qp, struct mlx5_ib_gsi_qp, ibqp);
}
+static bool mlx5_ib_deth_sqpn_cap(struct mlx5_ib_dev *dev)
+{
+ return MLX5_CAP_GEN(dev->mdev, set_deth_sqpn);
+}
+
struct ib_qp *mlx5_ib_gsi_create_qp(struct ib_pd *pd,
struct ib_qp_init_attr *init_attr)
{
@@ -54,6 +65,8 @@ struct ib_qp *mlx5_ib_gsi_create_qp(struct ib_pd *pd,
struct mlx5_ib_gsi_qp *gsi;
struct ib_qp_init_attr hw_init_attr = *init_attr;
const u8 port_num = init_attr->port_num;
+ const int num_pkeys = pd->device->attrs.max_pkeys;
+ const int num_qps = mlx5_ib_deth_sqpn_cap(dev) ? num_pkeys : 0;
int ret;
mlx5_ib_dbg(dev, "creating GSI QP\n");
@@ -69,6 +82,12 @@ struct ib_qp *mlx5_ib_gsi_create_qp(struct ib_pd *pd,
if (!gsi)
return ERR_PTR(-ENOMEM);
+ gsi->tx_qps = kcalloc(num_qps, sizeof(*gsi->tx_qps), GFP_KERNEL);
+ if (!gsi->tx_qps) {
+ ret = -ENOMEM;
+ goto err_free;
+ }
+
mutex_init(&gsi->mutex);
mutex_lock(&dev->devr.mutex);
@@ -77,8 +96,10 @@ struct ib_qp *mlx5_ib_gsi_create_qp(struct ib_pd *pd,
mlx5_ib_warn(dev, "GSI QP already exists on port %d\n",
port_num);
ret = -EBUSY;
- goto err_free;
+ goto err_free_tx;
}
+ gsi->num_qps = num_qps;
+ spin_lock_init(&gsi->lock);
gsi->cap = init_attr->cap;
gsi->sq_sig_type = init_attr->sq_sig_type;
@@ -91,7 +112,7 @@ struct ib_qp *mlx5_ib_gsi_create_qp(struct ib_pd *pd,
mlx5_ib_warn(dev, "unable to create hardware GSI QP. error %ld\n",
PTR_ERR(gsi->rx_qp));
ret = PTR_ERR(gsi->rx_qp);
- goto err_free;
+ goto err_free_tx;
}
dev->devr.ports[init_attr->port_num - 1].gsi = gsi;
@@ -100,8 +121,10 @@ struct ib_qp *mlx5_ib_gsi_create_qp(struct ib_pd *pd,
return &gsi->ibqp;
-err_free:
+err_free_tx:
mutex_unlock(&dev->devr.mutex);
+ kfree(gsi->tx_qps);
+err_free:
kfree(gsi);
return ERR_PTR(ret);
}
@@ -111,6 +134,7 @@ int mlx5_ib_gsi_destroy_qp(struct ib_qp *qp)
struct mlx5_ib_dev *dev = to_mdev(qp->device);
struct mlx5_ib_gsi_qp *gsi = gsi_qp(qp);
const int port_num = gsi->port_num;
+ int qp_index;
int ret;
mlx5_ib_dbg(dev, "destroying GSI QP\n");
@@ -125,12 +149,143 @@ int mlx5_ib_gsi_destroy_qp(struct ib_qp *qp)
}
dev->devr.ports[port_num - 1].gsi = NULL;
mutex_unlock(&dev->devr.mutex);
+ gsi->rx_qp = NULL;
+
+ for (qp_index = 0; qp_index < gsi->num_qps; ++qp_index) {
+ if (!gsi->tx_qps[qp_index])
+ continue;
+ WARN_ON_ONCE(ib_destroy_qp(gsi->tx_qps[qp_index]));
+ gsi->tx_qps[qp_index] = NULL;
+ }
+ kfree(gsi->tx_qps);
kfree(gsi);
return 0;
}
+static struct ib_qp *create_gsi_ud_qp(struct mlx5_ib_gsi_qp *gsi)
+{
+ struct ib_pd *pd = gsi->rx_qp->pd;
+ struct ib_qp_init_attr init_attr = {
+ .event_handler = gsi->rx_qp->event_handler,
+ .qp_context = gsi->rx_qp->qp_context,
+ .send_cq = gsi->rx_qp->send_cq,
+ .recv_cq = gsi->rx_qp->recv_cq,
+ .cap = {
+ .max_send_wr = gsi->cap.max_send_wr,
+ .max_send_sge = gsi->cap.max_send_sge,
+ .max_inline_data = gsi->cap.max_inline_data,
+ },
+ .sq_sig_type = gsi->sq_sig_type,
+ .qp_type = IB_QPT_UD,
+ .create_flags = mlx5_ib_create_qp_sqpn_qp1(),
+ };
+
+ return ib_create_qp(pd, &init_attr);
+}
+
+static int modify_to_rts(struct mlx5_ib_gsi_qp *gsi, struct ib_qp *qp,
+ u16 qp_index)
+{
+ struct mlx5_ib_dev *dev = to_mdev(qp->device);
+ struct ib_qp_attr attr;
+ int mask;
+ int ret;
+
+ mask = IB_QP_STATE | IB_QP_PKEY_INDEX | IB_QP_QKEY | IB_QP_PORT;
+ attr.qp_state = IB_QPS_INIT;
+ attr.pkey_index = qp_index;
+ attr.qkey = IB_QP1_QKEY;
+ attr.port_num = gsi->port_num;
+ ret = ib_modify_qp(qp, &attr, mask);
+ if (ret) {
+ mlx5_ib_err(dev, "could not change QP%d state to INIT: %d\n",
+ qp->qp_num, ret);
+ return ret;
+ }
+
+ attr.qp_state = IB_QPS_RTR;
+ ret = ib_modify_qp(qp, &attr, IB_QP_STATE);
+ if (ret) {
+ mlx5_ib_err(dev, "could not change QP%d state to RTR: %d\n",
+ qp->qp_num, ret);
+ return ret;
+ }
+
+ attr.qp_state = IB_QPS_RTS;
+ attr.sq_psn = 0;
+ ret = ib_modify_qp(qp, &attr, IB_QP_STATE | IB_QP_SQ_PSN);
+ if (ret) {
+ mlx5_ib_err(dev, "could not change QP%d state to RTS: %d\n",
+ qp->qp_num, ret);
+ return ret;
+ }
+
+ return 0;
+}
+
+static void setup_qp(struct mlx5_ib_gsi_qp *gsi, u16 qp_index)
+{
+ struct ib_device *device = gsi->rx_qp->device;
+ struct mlx5_ib_dev *dev = to_mdev(device);
+ struct ib_qp *qp;
+ unsigned long flags;
+ u16 pkey;
+ int ret;
+
+ ret = ib_query_pkey(device, gsi->port_num, qp_index, &pkey);
+ if (ret) {
+ mlx5_ib_warn(dev, "unable to read P_Key at port %d, index %d\n",
+ gsi->port_num, qp_index);
+ return;
+ }
+
+ if (!pkey) {
+ mlx5_ib_dbg(dev, "invalid P_Key at port %d, index %d. Skipping.\n",
+ gsi->port_num, qp_index);
+ return;
+ }
+
+ spin_lock_irqsave(&gsi->lock, flags);
+ qp = gsi->tx_qps[qp_index];
+ spin_unlock_irqrestore(&gsi->lock, flags);
+ if (qp) {
+ mlx5_ib_dbg(dev, "already existing GSI TX QP at port %d, index %d. Skipping\n",
+ gsi->port_num, qp_index);
+ return;
+ }
+
+ 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));
+ return;
+ }
+
+ ret = modify_to_rts(gsi, qp, qp_index);
+ if (ret)
+ goto err_destroy_qp;
+
+ spin_lock_irqsave(&gsi->lock, flags);
+ WARN_ON_ONCE(gsi->tx_qps[qp_index]);
+ gsi->tx_qps[qp_index] = qp;
+ spin_unlock_irqrestore(&gsi->lock, flags);
+
+ return;
+
+err_destroy_qp:
+ WARN_ON_ONCE(qp);
+}
+
+static void setup_qps(struct mlx5_ib_gsi_qp *gsi)
+{
+ u16 qp_index;
+
+ for (qp_index = 0; qp_index < gsi->num_qps; ++qp_index)
+ setup_qp(gsi, qp_index);
+}
+
int mlx5_ib_gsi_modify_qp(struct ib_qp *qp, struct ib_qp_attr *attr,
int attr_mask)
{
@@ -142,6 +297,15 @@ int mlx5_ib_gsi_modify_qp(struct ib_qp *qp, struct ib_qp_attr *attr,
mutex_lock(&gsi->mutex);
ret = ib_modify_qp(gsi->rx_qp, attr, attr_mask);
+ if (ret) {
+ mlx5_ib_warn(dev, "unable to modify GSI rx QP: %d\n", ret);
+ goto unlock;
+ }
+
+ if (to_mqp(gsi->rx_qp)->state == IB_QPS_RTS)
+ setup_qps(gsi);
+
+unlock:
mutex_unlock(&gsi->mutex);
return ret;
--
1.7.11.2
--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
