RDMA/cxgb4: DB Drop Recovery for RDMA and LLD queues
Add module option db_fc_threshold which is the count of active QPs
that trigger automatic db flow control mode. Automatically transition
to/from flow control mode when the active qp count crosses
db_fc_theshold.
Add more db debugfs stats
On DB DROP event from the LLD, recover all the iwarp queues.
Signed-off-by: Vipul Pandya <vipul@chelsio.com>
Signed-off-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: Roland Dreier <roland@purestorage.com>
diff --git a/drivers/infiniband/hw/cxgb4/qp.c b/drivers/infiniband/hw/cxgb4/qp.c
index beec667..ba1343e 100644
--- a/drivers/infiniband/hw/cxgb4/qp.c
+++ b/drivers/infiniband/hw/cxgb4/qp.c
@@ -42,6 +42,11 @@
module_param(ocqp_support, int, 0644);
MODULE_PARM_DESC(ocqp_support, "Support on-chip SQs (default=1)");
+int db_fc_threshold = 2000;
+module_param(db_fc_threshold, int, 0644);
+MODULE_PARM_DESC(db_fc_threshold, "QP count/threshold that triggers automatic "
+ "db flow control mode (default = 2000)");
+
static void set_state(struct c4iw_qp *qhp, enum c4iw_qp_state state)
{
unsigned long flag;
@@ -1143,13 +1148,19 @@
mutex_lock(&qhp->rhp->db_mutex);
do {
- if (cxgb4_dbfifo_count(qhp->rhp->rdev.lldi.ports[0], 1) < 768) {
+
+ /*
+ * The interrupt threshold is dbfifo_int_thresh << 6. So
+ * make sure we don't cross that and generate an interrupt.
+ */
+ if (cxgb4_dbfifo_count(qhp->rhp->rdev.lldi.ports[0], 1) <
+ (qhp->rhp->rdev.lldi.dbfifo_int_thresh << 5)) {
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);
+ delay = min(delay << 1, 2000);
} while (1);
mutex_unlock(&qhp->rhp->db_mutex);
return 0;
@@ -1388,6 +1399,14 @@
return ret;
}
+static int enable_qp_db(int id, void *p, void *data)
+{
+ struct c4iw_qp *qp = p;
+
+ t4_enable_wq_db(&qp->wq);
+ return 0;
+}
+
int c4iw_destroy_qp(struct ib_qp *ib_qp)
{
struct c4iw_dev *rhp;
@@ -1405,7 +1424,16 @@
c4iw_modify_qp(rhp, qhp, C4IW_QP_ATTR_NEXT_STATE, &attrs, 0);
wait_event(qhp->wait, !qhp->ep);
- remove_handle(rhp, &rhp->qpidr, qhp->wq.sq.qid);
+ spin_lock_irq(&rhp->lock);
+ remove_handle_nolock(rhp, &rhp->qpidr, qhp->wq.sq.qid);
+ rhp->qpcnt--;
+ BUG_ON(rhp->qpcnt < 0);
+ if (rhp->qpcnt <= db_fc_threshold && rhp->db_state == FLOW_CONTROL) {
+ rhp->rdev.stats.db_state_transitions++;
+ rhp->db_state = NORMAL;
+ idr_for_each(&rhp->qpidr, enable_qp_db, NULL);
+ }
+ spin_unlock_irq(&rhp->lock);
atomic_dec(&qhp->refcnt);
wait_event(qhp->wait, !atomic_read(&qhp->refcnt));
@@ -1419,6 +1447,14 @@
return 0;
}
+static int disable_qp_db(int id, void *p, void *data)
+{
+ struct c4iw_qp *qp = p;
+
+ t4_disable_wq_db(&qp->wq);
+ return 0;
+}
+
struct ib_qp *c4iw_create_qp(struct ib_pd *pd, struct ib_qp_init_attr *attrs,
struct ib_udata *udata)
{
@@ -1508,6 +1544,11 @@
spin_lock_irq(&rhp->lock);
if (rhp->db_state != NORMAL)
t4_disable_wq_db(&qhp->wq);
+ if (++rhp->qpcnt > db_fc_threshold && rhp->db_state == NORMAL) {
+ rhp->rdev.stats.db_state_transitions++;
+ rhp->db_state = FLOW_CONTROL;
+ idr_for_each(&rhp->qpidr, disable_qp_db, NULL);
+ }
ret = insert_handle_nolock(rhp, &rhp->qpidr, qhp, qhp->wq.sq.qid);
spin_unlock_irq(&rhp->lock);
if (ret)
