IB/hfi1: Integrate TID RDMA READ protocol into RC protocol

This patch integrates the TID RDMA READ protocol into the IB RC protocol.
This protocol is an end-to-end protocol between the hfi1 drivers on two
OPA nodes that converts a qualified RDMA READ request into a TID RDMA
READ request to avoid data copying on the requester side. The following
codes are added in this patch:
- Send the TID RDMA READ request;
- Complete the TID RDMA READ send request;
- Send the TID RDMA READ response;
- Complete the TID RDMA READ request;

Reviewed-by: Mike Marciniszyn <mike.marciniszyn@intel.com>
Signed-off-by: Kaike Wan <kaike.wan@intel.com>
Signed-off-by: Dennis Dalessandro <dennis.dalessandro@intel.com>
Signed-off-by: Doug Ledford <dledford@redhat.com>

1 files changed, 271 insertions, 24 deletions

diff --git a/drivers/infiniband/hw/hfi1/rc.c b/drivers/infiniband/hw/hfi1/rc.c
index e478a0b93eb9..a5aacf8e5b93 100644
--- a/drivers/infiniband/hw/hfi1/rc.c
+++ b/drivers/infiniband/hw/hfi1/rc.c
@@ -112,12 +112,14 @@ static int make_rc_ack(struct hfi1_ibdev *dev, struct rvt_qp *qp,
 {
 	struct rvt_ack_entry *e;
 	u32 hwords;
-	u32 len;
-	u32 bth0, bth2;
+	u32 len = 0;
+	u32 bth0 = 0, bth2 = 0;
 	u32 bth1 = qp->remote_qpn | (HFI1_CAP_IS_KSET(OPFN) << IB_BTHE_E_SHIFT);
 	int middle = 0;
 	u32 pmtu = qp->pmtu;
 	struct hfi1_qp_priv *priv = qp->priv;
+	bool last_pkt;
+	u32 delta;
 
 	lockdep_assert_held(&qp->s_lock);
 	/* Don't send an ACK if we aren't supposed to. */
@@ -190,6 +192,26 @@ static int make_rc_ack(struct hfi1_ibdev *dev, struct rvt_qp *qp,
 			hwords++;
 			qp->s_ack_rdma_psn = e->psn;
 			bth2 = mask_psn(qp->s_ack_rdma_psn++);
+		} else if (e->opcode == TID_OP(READ_REQ)) {
+			/*
+			 * If a TID RDMA read response is being resent and
+			 * we haven't seen the duplicate request yet,
+			 * then stop sending the remaining responses the
+			 * responder has seen until the requester re-sends it.
+			 */
+			len = e->rdma_sge.sge_length;
+			if (len && !e->rdma_sge.mr) {
+				qp->s_tail_ack_queue = qp->r_head_ack_queue;
+				goto bail;
+			}
+			/* Copy SGE state in case we need to resend */
+			ps->s_txreq->mr = e->rdma_sge.mr;
+			if (ps->s_txreq->mr)
+				rvt_get_mr(ps->s_txreq->mr);
+			qp->s_ack_rdma_sge.sge = e->rdma_sge;
+			qp->s_ack_rdma_sge.num_sge = 1;
+			qp->s_ack_state = TID_OP(READ_RESP);
+			goto read_resp;
 		} else {
 			/* COMPARE_SWAP or FETCH_ADD */
 			ps->s_txreq->ss = NULL;
@@ -227,6 +249,28 @@ static int make_rc_ack(struct hfi1_ibdev *dev, struct rvt_qp *qp,
 		bth2 = mask_psn(qp->s_ack_rdma_psn++);
 		break;
 
+	case TID_OP(READ_RESP):
+read_resp:
+		e = &qp->s_ack_queue[qp->s_tail_ack_queue];
+		ps->s_txreq->ss = &qp->s_ack_rdma_sge;
+		delta = hfi1_build_tid_rdma_read_resp(qp, e, ohdr, &bth0,
+						      &bth1, &bth2, &len,
+						      &last_pkt);
+		if (delta == 0)
+			goto error_qp;
+		hwords += delta;
+		if (last_pkt) {
+			e->sent = 1;
+			/*
+			 * Increment qp->s_tail_ack_queue through s_ack_state
+			 * transition.
+			 */
+			qp->s_ack_state = OP(RDMA_READ_RESPONSE_LAST);
+		}
+		break;
+	case TID_OP(READ_REQ):
+		goto bail;
+
 	default:
 normal:
 		/*
@@ -256,7 +300,14 @@ normal:
 	ps->s_txreq->hdr_dwords = hwords;
 	hfi1_make_ruc_header(qp, ohdr, bth0, bth1, bth2, middle, ps);
 	return 1;
-
+error_qp:
+	spin_unlock_irqrestore(&qp->s_lock, ps->flags);
+	spin_lock_irqsave(&qp->r_lock, ps->flags);
+	spin_lock(&qp->s_lock);
+	rvt_error_qp(qp, IB_WC_WR_FLUSH_ERR);
+	spin_unlock(&qp->s_lock);
+	spin_unlock_irqrestore(&qp->r_lock, ps->flags);
+	spin_lock_irqsave(&qp->s_lock, ps->flags);
 bail:
 	qp->s_ack_state = OP(ACKNOWLEDGE);
 	/*
@@ -283,16 +334,20 @@ int hfi1_make_rc_req(struct rvt_qp *qp, struct hfi1_pkt_state *ps)
 	struct hfi1_qp_priv *priv = qp->priv;
 	struct hfi1_ibdev *dev = to_idev(qp->ibqp.device);
 	struct ib_other_headers *ohdr;
-	struct rvt_sge_state *ss;
+	struct rvt_sge_state *ss = NULL;
 	struct rvt_swqe *wqe;
-	u32 hwords;
-	u32 len;
-	u32 bth0 = 0, bth2;
+	struct hfi1_swqe_priv *wpriv;
+	struct tid_rdma_request *req = NULL;
+	/* header size in 32-bit words LRH+BTH = (8+12)/4. */
+	u32 hwords = 5;
+	u32 len = 0;
+	u32 bth0 = 0, bth2 = 0;
 	u32 bth1 = qp->remote_qpn | (HFI1_CAP_IS_KSET(OPFN) << IB_BTHE_E_SHIFT);
 	u32 pmtu = qp->pmtu;
 	char newreq;
 	int middle = 0;
 	int delta;
+	struct tid_rdma_flow *flow = NULL;
 
 	lockdep_assert_held(&qp->s_lock);
 	ps->s_txreq = get_txreq(ps->dev, qp);
@@ -334,8 +389,8 @@ int hfi1_make_rc_req(struct rvt_qp *qp, struct hfi1_pkt_state *ps)
 		}
 		clear_ahg(qp);
 		wqe = rvt_get_swqe_ptr(qp, qp->s_last);
-		rvt_send_complete(qp, wqe, qp->s_last != qp->s_acked ?
-			IB_WC_SUCCESS : IB_WC_WR_FLUSH_ERR);
+		hfi1_trdma_send_complete(qp, wqe, qp->s_last != qp->s_acked ?
+					 IB_WC_SUCCESS : IB_WC_WR_FLUSH_ERR);
 		/* will get called again */
 		goto done_free_tx;
 	}
@@ -354,6 +409,7 @@ int hfi1_make_rc_req(struct rvt_qp *qp, struct hfi1_pkt_state *ps)
 
 	/* Send a request. */
 	wqe = rvt_get_swqe_ptr(qp, qp->s_cur);
+check_s_state:
 	switch (qp->s_state) {
 	default:
 		if (!(ib_rvt_state_ops[qp->state] & RVT_PROCESS_NEXT_SEND_OK))
@@ -375,9 +431,13 @@ int hfi1_make_rc_req(struct rvt_qp *qp, struct hfi1_pkt_state *ps)
 			/*
 			 * If a fence is requested, wait for previous
 			 * RDMA read and atomic operations to finish.
+			 * However, there is no need to guard against
+			 * TID RDMA READ after TID RDMA READ.
 			 */
 			if ((wqe->wr.send_flags & IB_SEND_FENCE) &&
-			    qp->s_num_rd_atomic) {
+			    qp->s_num_rd_atomic &&
+			    (wqe->wr.opcode != IB_WR_TID_RDMA_READ ||
+			     priv->pending_tid_r_segs < qp->s_num_rd_atomic)) {
 				qp->s_flags |= RVT_S_WAIT_FENCE;
 				goto bail;
 			}
@@ -526,6 +586,75 @@ no_flow_control:
 				qp->s_cur = 0;
 			break;
 
+		case IB_WR_TID_RDMA_READ:
+			wpriv = wqe->priv;
+			req = wqe_to_tid_req(wqe);
+			delta = cmp_psn(qp->s_psn, wqe->psn);
+
+			/*
+			 * Don't allow more operations to be started
+			 * than the QP limits allow. We could get here under
+			 * three conditions; (1) It's a new request; (2) We are
+			 * sending the second or later segment of a request,
+			 * but the qp->s_state is set to OP(RDMA_READ_REQUEST)
+			 * when the last segment of a previous request is
+			 * received just before this; (3) We are re-sending a
+			 * request.
+			 */
+			if (qp->s_num_rd_atomic >= qp->s_max_rd_atomic) {
+				qp->s_flags |= RVT_S_WAIT_RDMAR;
+				goto bail;
+			}
+			if (newreq) {
+				struct tid_rdma_flow *flow =
+					&req->flows[req->setup_head];
+
+				/*
+				 * Set up s_sge as it is needed for TID
+				 * allocation. However, if the pages have been
+				 * walked and mapped, skip it. An earlier try
+				 * has failed to allocate the TID entries.
+				 */
+				if (!flow->npagesets) {
+					qp->s_sge.sge = wqe->sg_list[0];
+					qp->s_sge.sg_list = wqe->sg_list + 1;
+					qp->s_sge.num_sge = wqe->wr.num_sge;
+					qp->s_sge.total_len = wqe->length;
+					qp->s_len = wqe->length;
+					req->isge = 0;
+					req->clear_tail = req->setup_head;
+					req->flow_idx = req->setup_head;
+					req->state = TID_REQUEST_ACTIVE;
+				}
+			} else if (delta == 0) {
+				/* Re-send a request */
+				req->cur_seg = 0;
+				req->comp_seg = 0;
+				req->ack_pending = 0;
+				req->flow_idx = req->clear_tail;
+				req->state = TID_REQUEST_RESEND;
+			}
+			req->s_next_psn = qp->s_psn;
+			/* Read one segment at a time */
+			len = min_t(u32, req->seg_len,
+				    wqe->length - req->seg_len * req->cur_seg);
+			delta = hfi1_build_tid_rdma_read_req(qp, wqe, ohdr,
+							     &bth1, &bth2,
+							     &len);
+			if (delta <= 0) {
+				/* Wait for TID space */
+				goto bail;
+			}
+			if (newreq && !(qp->s_flags & RVT_S_UNLIMITED_CREDIT))
+				qp->s_lsn++;
+			hwords += delta;
+			ss = &wpriv->ss;
+			/* Check if this is the last segment */
+			if (req->cur_seg >= req->total_segs &&
+			    ++qp->s_cur == qp->s_size)
+				qp->s_cur = 0;
+			break;
+
 		case IB_WR_ATOMIC_CMP_AND_SWP:
 		case IB_WR_ATOMIC_FETCH_AND_ADD:
 			/*
@@ -571,11 +700,13 @@ no_flow_control:
 		default:
 			goto bail;
 		}
-		qp->s_sge.sge = wqe->sg_list[0];
-		qp->s_sge.sg_list = wqe->sg_list + 1;
-		qp->s_sge.num_sge = wqe->wr.num_sge;
-		qp->s_sge.total_len = wqe->length;
-		qp->s_len = wqe->length;
+		if (wqe->wr.opcode != IB_WR_TID_RDMA_READ) {
+			qp->s_sge.sge = wqe->sg_list[0];
+			qp->s_sge.sg_list = wqe->sg_list + 1;
+			qp->s_sge.num_sge = wqe->wr.num_sge;
+			qp->s_sge.total_len = wqe->length;
+			qp->s_len = wqe->length;
+		}
 		if (newreq) {
 			qp->s_tail++;
 			if (qp->s_tail >= qp->s_size)
@@ -583,6 +714,8 @@ no_flow_control:
 		}
 		if (wqe->wr.opcode == IB_WR_RDMA_READ)
 			qp->s_psn = wqe->lpsn + 1;
+		else if (wqe->wr.opcode == IB_WR_TID_RDMA_READ)
+			qp->s_psn = req->s_next_psn;
 		else
 			qp->s_psn++;
 		break;
@@ -699,6 +832,99 @@ no_flow_control:
 		if (qp->s_cur == qp->s_size)
 			qp->s_cur = 0;
 		break;
+	case TID_OP(READ_RESP):
+		if (wqe->wr.opcode != IB_WR_TID_RDMA_READ)
+			goto bail;
+		/* This is used to restart a TID read request */
+		req = wqe_to_tid_req(wqe);
+		wpriv = wqe->priv;
+		/*
+		 * Back down. The field qp->s_psn has been set to the psn with
+		 * which the request should be restart. It's OK to use division
+		 * as this is on the retry path.
+		 */
+		req->cur_seg = delta_psn(qp->s_psn, wqe->psn) / priv->pkts_ps;
+
+		/*
+		 * The following function need to be redefined to return the
+		 * status to make sure that we find the flow. At the same
+		 * time, we can use the req->state change to check if the
+		 * call succeeds or not.
+		 */
+		req->state = TID_REQUEST_RESEND;
+		hfi1_tid_rdma_restart_req(qp, wqe, &bth2);
+		if (req->state != TID_REQUEST_ACTIVE) {
+			/*
+			 * Failed to find the flow. Release all allocated tid
+			 * resources.
+			 */
+			hfi1_kern_exp_rcv_clear_all(req);
+			hfi1_kern_clear_hw_flow(priv->rcd, qp);
+
+			hfi1_trdma_send_complete(qp, wqe, IB_WC_LOC_QP_OP_ERR);
+			goto bail;
+		}
+		req->state = TID_REQUEST_RESEND;
+		len = min_t(u32, req->seg_len,
+			    wqe->length - req->seg_len * req->cur_seg);
+		flow = &req->flows[req->flow_idx];
+		len -= flow->sent;
+		req->s_next_psn = flow->flow_state.ib_lpsn + 1;
+		delta = hfi1_build_tid_rdma_read_packet(wqe, ohdr, &bth1,
+							&bth2, &len);
+		if (delta <= 0) {
+			/* Wait for TID space */
+			goto bail;
+		}
+		hwords += delta;
+		ss = &wpriv->ss;
+		/* Check if this is the last segment */
+		if (req->cur_seg >= req->total_segs &&
+		    ++qp->s_cur == qp->s_size)
+			qp->s_cur = 0;
+		qp->s_psn = req->s_next_psn;
+		break;
+	case TID_OP(READ_REQ):
+		req = wqe_to_tid_req(wqe);
+		delta = cmp_psn(qp->s_psn, wqe->psn);
+		/*
+		 * If the current WR is not TID RDMA READ, or this is the start
+		 * of a new request, we need to change the qp->s_state so that
+		 * the request can be set up properly.
+		 */
+		if (wqe->wr.opcode != IB_WR_TID_RDMA_READ || delta == 0 ||
+		    qp->s_cur == qp->s_tail) {
+			qp->s_state = OP(RDMA_READ_REQUEST);
+			if (delta == 0 || qp->s_cur == qp->s_tail)
+				goto check_s_state;
+			else
+				goto bail;
+		}
+
+		/* Rate limiting */
+		if (qp->s_num_rd_atomic >= qp->s_max_rd_atomic) {
+			qp->s_flags |= RVT_S_WAIT_RDMAR;
+			goto bail;
+		}
+
+		wpriv = wqe->priv;
+		/* Read one segment at a time */
+		len = min_t(u32, req->seg_len,
+			    wqe->length - req->seg_len * req->cur_seg);
+		delta = hfi1_build_tid_rdma_read_req(qp, wqe, ohdr, &bth1,
+						     &bth2, &len);
+		if (delta <= 0) {
+			/* Wait for TID space */
+			goto bail;
+		}
+		hwords += delta;
+		ss = &wpriv->ss;
+		/* Check if this is the last segment */
+		if (req->cur_seg >= req->total_segs &&
+		    ++qp->s_cur == qp->s_size)
+			qp->s_cur = 0;
+		qp->s_psn = req->s_next_psn;
+		break;
 	}
 	qp->s_sending_hpsn = bth2;
 	delta = delta_psn(bth2, wqe->psn);
@@ -1148,8 +1374,8 @@ void hfi1_restart_rc(struct rvt_qp *qp, u32 psn, int wait)
 					hfi1_kern_clear_hw_flow(priv->rcd, qp);
 				}
 
-				rvt_send_complete(qp, wqe,
-						  IB_WC_RETRY_EXC_ERR);
+				hfi1_trdma_send_complete(qp, wqe,
+							 IB_WC_RETRY_EXC_ERR);
 				rvt_error_qp(qp, IB_WC_WR_FLUSH_ERR);
 			}
 			return;
@@ -1189,7 +1415,8 @@ static void reset_sending_psn(struct rvt_qp *qp, u32 psn)
 	for (;;) {
 		wqe = rvt_get_swqe_ptr(qp, n);
 		if (cmp_psn(psn, wqe->lpsn) <= 0) {
-			if (wqe->wr.opcode == IB_WR_RDMA_READ)
+			if (wqe->wr.opcode == IB_WR_RDMA_READ ||
+			    wqe->wr.opcode == IB_WR_TID_RDMA_READ)
 				qp->s_sending_psn = wqe->lpsn + 1;
 			else
 				qp->s_sending_psn = psn + 1;
@@ -1238,8 +1465,9 @@ void hfi1_rc_send_complete(struct rvt_qp *qp, struct hfi1_opa_header *opah)
 	}
 
 	opcode = ib_bth_get_opcode(ohdr);
-	if (opcode >= OP(RDMA_READ_RESPONSE_FIRST) &&
-	    opcode <= OP(ATOMIC_ACKNOWLEDGE)) {
+	if ((opcode >= OP(RDMA_READ_RESPONSE_FIRST) &&
+	     opcode <= OP(ATOMIC_ACKNOWLEDGE)) ||
+	    opcode == TID_OP(READ_RESP)) {
 		WARN_ON(!qp->s_rdma_ack_cnt);
 		qp->s_rdma_ack_cnt--;
 		return;
@@ -1255,8 +1483,12 @@ void hfi1_rc_send_complete(struct rvt_qp *qp, struct hfi1_opa_header *opah)
 	if ((psn & IB_BTH_REQ_ACK) && qp->s_acked != qp->s_tail &&
 	    !(qp->s_flags &
 		(RVT_S_TIMER | RVT_S_WAIT_RNR | RVT_S_WAIT_PSN)) &&
-		(ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK))
-		rvt_add_retry_timer(qp);
+		(ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK)) {
+		if (opcode == TID_OP(READ_REQ))
+			rvt_add_retry_timer_ext(qp, priv->timeout_shift);
+		else
+			rvt_add_retry_timer(qp);
+	}
 
 	while (qp->s_last != qp->s_acked) {
 		u32 s_last;
@@ -1265,6 +1497,7 @@ void hfi1_rc_send_complete(struct rvt_qp *qp, struct hfi1_opa_header *opah)
 		if (cmp_psn(wqe->lpsn, qp->s_sending_psn) >= 0 &&
 		    cmp_psn(qp->s_sending_psn, qp->s_sending_hpsn) <= 0)
 			break;
+		trdma_clean_swqe(qp, wqe);
 		rvt_qp_wqe_unreserve(qp, wqe);
 		s_last = qp->s_last;
 		trace_hfi1_qp_send_completion(qp, wqe, s_last);
@@ -1317,6 +1550,7 @@ struct rvt_swqe *do_rc_completion(struct rvt_qp *qp,
 	    cmp_psn(qp->s_sending_psn, qp->s_sending_hpsn) > 0) {
 		u32 s_last;
 
+		trdma_clean_swqe(qp, wqe);
 		rvt_put_swqe(wqe);
 		rvt_qp_wqe_unreserve(qp, wqe);
 		s_last = qp->s_last;
@@ -1393,6 +1627,7 @@ int do_rc_ack(struct rvt_qp *qp, u32 aeth, u32 psn, int opcode,
 {
 	struct hfi1_ibport *ibp;
 	enum ib_wc_status status;
+	struct hfi1_qp_priv *qpriv = qp->priv;
 	struct rvt_swqe *wqe;
 	int ret = 0;
 	u32 ack_psn;
@@ -1439,6 +1674,8 @@ int do_rc_ack(struct rvt_qp *qp, u32 aeth, u32 psn, int opcode,
 		 */
 		if ((wqe->wr.opcode == IB_WR_RDMA_READ &&
 		     (opcode != OP(RDMA_READ_RESPONSE_LAST) || diff != 0)) ||
+		    (wqe->wr.opcode == IB_WR_TID_RDMA_READ &&
+		     (opcode != TID_OP(READ_RESP) || diff != 0)) ||
 		    ((wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP ||
 		      wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD) &&
 		     (opcode != OP(ATOMIC_ACKNOWLEDGE) || diff != 0))) {
@@ -1492,7 +1729,13 @@ int do_rc_ack(struct rvt_qp *qp, u32 aeth, u32 psn, int opcode,
 	switch (aeth >> IB_AETH_NAK_SHIFT) {
 	case 0:         /* ACK */
 		this_cpu_inc(*ibp->rvp.rc_acks);
-		if (qp->s_acked != qp->s_tail) {
+		if (wqe->wr.opcode == IB_WR_TID_RDMA_READ) {
+			if (wqe_to_tid_req(wqe)->ack_pending)
+				rvt_mod_retry_timer_ext(qp,
+							qpriv->timeout_shift);
+			else
+				rvt_stop_rc_timers(qp);
+		} else if (qp->s_acked != qp->s_tail) {
 			/*
 			 * We are expecting more ACKs so
 			 * mod the retry timer.
@@ -1581,7 +1824,10 @@ int do_rc_ack(struct rvt_qp *qp, u32 aeth, u32 psn, int opcode,
 			ibp->rvp.n_other_naks++;
 class_b:
 			if (qp->s_last == qp->s_acked) {
-				rvt_send_complete(qp, wqe, status);
+				if (wqe->wr.opcode == IB_WR_TID_RDMA_READ)
+					hfi1_kern_read_tid_flow_free(qp);
+
+				hfi1_trdma_send_complete(qp, wqe, status);
 				rvt_error_qp(qp, IB_WC_WR_FLUSH_ERR);
 			}
 			break;
@@ -1622,6 +1868,7 @@ static void rdma_seq_err(struct rvt_qp *qp, struct hfi1_ibport *ibp, u32 psn,
 
 	while (cmp_psn(psn, wqe->lpsn) > 0) {
 		if (wqe->wr.opcode == IB_WR_RDMA_READ ||
+		    wqe->wr.opcode == IB_WR_TID_RDMA_READ ||
 		    wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP ||
 		    wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD)
 			break;