RDMA/cxgb4: Add DB Overflow Avoidance

Get FULL/EMPTY/DROP events from LLD.  On FULL event, disable normal
user mode DB rings.

Add modify_qp semantics to allow user processes to call into the
kernel to ring doobells without overflowing.

Add DB Full/Empty/Drop stats.

Mark queues when created indicating the doorbell state.

If we're in the middle of db overflow avoidance, then newly created
queues should start out in this mode.

Bump the C4IW_UVERBS_ABI_VERSION to 2 so the user mode library can
know if the driver supports the kernel mode db ringing.

Signed-off-by: Vipul Pandya <vipul@chelsio.com>
Signed-off-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>

1 files changed, 50 insertions, 1 deletions

diff --git a/drivers/infiniband/hw/cxgb4/qp.c b/drivers/infiniband/hw/cxgb4/qp.c
index 5f940aeaab1e..beec66758aec 100644
--- a/drivers/infiniband/hw/cxgb4/qp.c
+++ b/drivers/infiniband/hw/cxgb4/qp.c
@@ -34,6 +34,10 @@
 
 #include "iw_cxgb4.h"
 
+static int db_delay_usecs = 1;
+module_param(db_delay_usecs, int, 0644);
+MODULE_PARM_DESC(db_delay_usecs, "Usecs to delay awaiting db fifo to drain");
+
 static int ocqp_support = 1;
 module_param(ocqp_support, int, 0644);
 MODULE_PARM_DESC(ocqp_support, "Support on-chip SQs (default=1)");
@@ -1128,6 +1132,29 @@ out:
 	return ret;
 }
 
+/*
+ * Called by the library when the qp has user dbs disabled due to
+ * a DB_FULL condition.  This function will single-thread all user
+ * DB rings to avoid overflowing the hw db-fifo.
+ */
+static int ring_kernel_db(struct c4iw_qp *qhp, u32 qid, u16 inc)
+{
+	int delay = db_delay_usecs;
+
+	mutex_lock(&qhp->rhp->db_mutex);
+	do {
+		if (cxgb4_dbfifo_count(qhp->rhp->rdev.lldi.ports[0], 1) < 768) {
+			writel(V_QID(qid) | V_PIDX(inc), qhp->wq.db);
+			break;
+		}
+		set_current_state(TASK_UNINTERRUPTIBLE);
+		schedule_timeout(usecs_to_jiffies(delay));
+		delay = min(delay << 1, 200000);
+	} while (1);
+	mutex_unlock(&qhp->rhp->db_mutex);
+	return 0;
+}
+
 int c4iw_modify_qp(struct c4iw_dev *rhp, struct c4iw_qp *qhp,
 		   enum c4iw_qp_attr_mask mask,
 		   struct c4iw_qp_attributes *attrs,
@@ -1176,6 +1203,15 @@ int c4iw_modify_qp(struct c4iw_dev *rhp, struct c4iw_qp *qhp,
 		qhp->attr = newattr;
 	}
 
+	if (mask & C4IW_QP_ATTR_SQ_DB) {
+		ret = ring_kernel_db(qhp, qhp->wq.sq.qid, attrs->sq_db_inc);
+		goto out;
+	}
+	if (mask & C4IW_QP_ATTR_RQ_DB) {
+		ret = ring_kernel_db(qhp, qhp->wq.rq.qid, attrs->rq_db_inc);
+		goto out;
+	}
+
 	if (!(mask & C4IW_QP_ATTR_NEXT_STATE))
 		goto out;
 	if (qhp->attr.state == attrs->next_state)
@@ -1469,7 +1505,11 @@ struct ib_qp *c4iw_create_qp(struct ib_pd *pd, struct ib_qp_init_attr *attrs,
 	init_waitqueue_head(&qhp->wait);
 	atomic_set(&qhp->refcnt, 1);
 
-	ret = insert_handle(rhp, &rhp->qpidr, qhp, qhp->wq.sq.qid);
+	spin_lock_irq(&rhp->lock);
+	if (rhp->db_state != NORMAL)
+		t4_disable_wq_db(&qhp->wq);
+	ret = insert_handle_nolock(rhp, &rhp->qpidr, qhp, qhp->wq.sq.qid);
+	spin_unlock_irq(&rhp->lock);
 	if (ret)
 		goto err2;
 
@@ -1613,6 +1653,15 @@ int c4iw_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
 			 C4IW_QP_ATTR_ENABLE_RDMA_WRITE |
 			 C4IW_QP_ATTR_ENABLE_RDMA_BIND) : 0;
 
+	/*
+	 * Use SQ_PSN and RQ_PSN to pass in IDX_INC values for
+	 * ringing the queue db when we're in DB_FULL mode.
+	 */
+	attrs.sq_db_inc = attr->sq_psn;
+	attrs.rq_db_inc = attr->rq_psn;
+	mask |= (attr_mask & IB_QP_SQ_PSN) ? C4IW_QP_ATTR_SQ_DB : 0;
+	mask |= (attr_mask & IB_QP_RQ_PSN) ? C4IW_QP_ATTR_RQ_DB : 0;
+
 	return c4iw_modify_qp(rhp, qhp, mask, &attrs, 0);
 }