IB/hfi1: Add running average for adaptive pio

The adaptive PIO implementation only considers the current packet size
when deciding between SDMA and pio for a packet.

This causes credit return forces if small and large packets are
interleaved.

Add a running average to avoid costly credit forces so that a large
sequence of small packets is required to go below the threshold that
chooses pio.

Reviewed-by: Michael J. Ruhl <michael.j.ruhl@intel.com>
Signed-off-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>

drivers/infiniband/hw/hfi1/qp.c

@@ -742,6 +742,8 @@ void *qp_priv_alloc(struct rvt_dev_info *rdi, struct rvt_qp *qp)
 		iowait_wakeup,
 		iowait_sdma_drained,
 		hfi1_init_priority);
+	/* Init to a value to start the running average correctly */
+	priv->s_running_pkt_size = piothreshold / 2;
 	return priv;
 }
 

drivers/infiniband/hw/hfi1/verbs.c

@@ -1223,15 +1223,16 @@ static inline send_routine get_send_routine(struct rvt_qp *qp,
 	case IB_QPT_UD:
 		break;
 	case IB_QPT_UC:
-	case IB_QPT_RC: {
+	case IB_QPT_RC:
+		priv->s_running_pkt_size =
+			(tx->s_cur_size + priv->s_running_pkt_size) / 2;
 		if (piothreshold &&
-		    tx->s_cur_size <= min(piothreshold, qp->pmtu) &&
+		    priv->s_running_pkt_size <= min(piothreshold, qp->pmtu) &&
 		    (BIT(ps->opcode & OPMASK) & pio_opmask[ps->opcode >> 5]) &&
 		    iowait_sdma_pending(&priv->s_iowait) == 0 &&
 		    !sdma_txreq_built(&tx->txreq))
 			return dd->process_pio_send;
 		break;
-	}
 	default:
 		break;
 	}

drivers/infiniband/hw/hfi1/verbs.h

@@ -170,6 +170,7 @@ struct hfi1_qp_priv {
 	struct tid_flow_state flow_state;
 	struct tid_rdma_qp_params tid_rdma;
 	struct rvt_qp *owner;
+	u16 s_running_pkt_size;
 	u8 hdr_type; /* 9B or 16B */
 	struct rvt_sge_state tid_ss;       /* SGE state pointer for 2nd leg */
 	atomic_t n_requests;               /* # of TID RDMA requests in the */