IB/hfi1: Add functions for restarting TID RDMA READ request

This patch adds functions to retry TID RDMA READ request. Since TID RDMA
READ request could be retried from any segment boundary, it requires
a number of tracking fields in various structures and those fields
should be reset properly. The qp->s_num_rd_atomic field is reset before
retry and therefore should be incremented for each new or retried
RDMA READ or atomic 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>

drivers/infiniband/hw/hfi1/rc.c

@@ -503,16 +503,14 @@ int hfi1_make_rc_req(struct rvt_qp *qp, struct hfi1_pkt_state *ps)
 			 * Don't allow more operations to be started
 			 * than the QP limits allow.
 			 */
-			if (newreq) {
 			if (qp->s_num_rd_atomic >=
 			    qp->s_max_rd_atomic) {
 				qp->s_flags |= RVT_S_WAIT_RDMAR;
 				goto bail;
 			}
 			qp->s_num_rd_atomic++;
-				if (!(qp->s_flags & RVT_S_UNLIMITED_CREDIT))
+			if (newreq && !(qp->s_flags & RVT_S_UNLIMITED_CREDIT))
 				qp->s_lsn++;
-			}
 			put_ib_reth_vaddr(
 				wqe->rdma_wr.remote_addr,
 				&ohdr->u.rc.reth);
@@ -534,14 +532,12 @@ int hfi1_make_rc_req(struct rvt_qp *qp, struct hfi1_pkt_state *ps)
 			 * Don't allow more operations to be started
 			 * than the QP limits allow.
 			 */
-			if (newreq) {
 			if (qp->s_num_rd_atomic >=
 			    qp->s_max_rd_atomic) {
 				qp->s_flags |= RVT_S_WAIT_RDMAR;
 				goto bail;
 			}
 			qp->s_num_rd_atomic++;
-			}
 
 			/* FALLTHROUGH */
 		case IB_WR_OPFN:
@@ -970,6 +966,43 @@ void hfi1_send_rc_ack(struct hfi1_packet *packet, bool is_fecn)
 	return;
 }
 
+/**
+ * update_num_rd_atomic - update the qp->s_num_rd_atomic
+ * @qp: the QP
+ * @psn: the packet sequence number to restart at
+ * @wqe: the wqe
+ *
+ * This is called from reset_psn() to update qp->s_num_rd_atomic
+ * for the current wqe.
+ * Called at interrupt level with the QP s_lock held.
+ */
+static void update_num_rd_atomic(struct rvt_qp *qp, u32 psn,
+				 struct rvt_swqe *wqe)
+{
+	u32 opcode = wqe->wr.opcode;
+
+	if (opcode == IB_WR_RDMA_READ ||
+	    opcode == IB_WR_ATOMIC_CMP_AND_SWP ||
+	    opcode == IB_WR_ATOMIC_FETCH_AND_ADD) {
+		qp->s_num_rd_atomic++;
+	} else if (opcode == IB_WR_TID_RDMA_READ) {
+		struct tid_rdma_request *req = wqe_to_tid_req(wqe);
+		struct hfi1_qp_priv *priv = qp->priv;
+
+		if (cmp_psn(psn, wqe->lpsn) <= 0) {
+			u32 cur_seg;
+
+			cur_seg = (psn - wqe->psn) / priv->pkts_ps;
+			req->ack_pending = cur_seg - req->comp_seg;
+			priv->pending_tid_r_segs += req->ack_pending;
+			qp->s_num_rd_atomic += req->ack_pending;
+		} else {
+			priv->pending_tid_r_segs += req->total_segs;
+			qp->s_num_rd_atomic += req->total_segs;
+		}
+	}
+}
+
 /**
  * reset_psn - reset the QP state to send starting from PSN
  * @qp: the QP
@@ -984,9 +1017,12 @@ static void reset_psn(struct rvt_qp *qp, u32 psn)
 	u32 n = qp->s_acked;
 	struct rvt_swqe *wqe = rvt_get_swqe_ptr(qp, n);
 	u32 opcode;
+	struct hfi1_qp_priv *priv = qp->priv;
 
 	lockdep_assert_held(&qp->s_lock);
 	qp->s_cur = n;
+	priv->pending_tid_r_segs = 0;
+	qp->s_num_rd_atomic = 0;
 
 	/*
 	 * If we are starting the request from the beginning,
@@ -996,9 +1032,9 @@ static void reset_psn(struct rvt_qp *qp, u32 psn)
 		qp->s_state = OP(SEND_LAST);
 		goto done;
 	}
+	update_num_rd_atomic(qp, psn, wqe);
 
 	/* Find the work request opcode corresponding to the given PSN. */
-	opcode = wqe->wr.opcode;
 	for (;;) {
 		int diff;
 
@@ -1008,8 +1044,11 @@ static void reset_psn(struct rvt_qp *qp, u32 psn)
 			break;
 		wqe = rvt_get_swqe_ptr(qp, n);
 		diff = cmp_psn(psn, wqe->psn);
-		if (diff < 0)
+		if (diff < 0) {
+			/* Point wqe back to the previous one*/
+			wqe = rvt_get_swqe_ptr(qp, qp->s_cur);
 			break;
+		}
 		qp->s_cur = n;
 		/*
 		 * If we are starting the request from the beginning,
@@ -1019,8 +1058,10 @@ static void reset_psn(struct rvt_qp *qp, u32 psn)
 			qp->s_state = OP(SEND_LAST);
 			goto done;
 		}
-		opcode = wqe->wr.opcode;
+
+		update_num_rd_atomic(qp, psn, wqe);
 	}
+	opcode = wqe->wr.opcode;
 
 	/*
 	 * Set the state to restart in the middle of a request.
@@ -1042,6 +1083,10 @@ static void reset_psn(struct rvt_qp *qp, u32 psn)
 		qp->s_state = OP(RDMA_READ_RESPONSE_MIDDLE);
 		break;
 
+	case IB_WR_TID_RDMA_READ:
+		qp->s_state = TID_OP(READ_RESP);
+		break;
+
 	default:
 		/*
 		 * This case shouldn't happen since its only
@@ -1095,6 +1140,14 @@ void hfi1_restart_rc(struct rvt_qp *qp, u32 psn, int wait)
 				wqe = do_rc_completion(qp, wqe, ibp);
 				qp->s_flags &= ~RVT_S_WAIT_ACK;
 			} else {
+				if (wqe->wr.opcode == IB_WR_TID_RDMA_READ) {
+					struct tid_rdma_request *req;
+
+					req = wqe_to_tid_req(wqe);
+					hfi1_kern_exp_rcv_clear_all(req);
+					hfi1_kern_clear_hw_flow(priv->rcd, qp);
+				}
+
 				rvt_send_complete(qp, wqe,
 						  IB_WC_RETRY_EXC_ERR);
 				rvt_error_qp(qp, IB_WC_WR_FLUSH_ERR);
@@ -1108,7 +1161,8 @@ void hfi1_restart_rc(struct rvt_qp *qp, u32 psn, int wait)
 	}
 
 	ibp = to_iport(qp->ibqp.device, qp->port_num);
-	if (wqe->wr.opcode == IB_WR_RDMA_READ)
+	if (wqe->wr.opcode == IB_WR_RDMA_READ ||
+	    wqe->wr.opcode == IB_WR_TID_RDMA_READ)
 		ibp->rvp.n_rc_resends++;
 	else
 		ibp->rvp.n_rc_resends += delta_psn(qp->s_psn, psn);

drivers/infiniband/hw/hfi1/tid_rdma.c

@@ -1622,6 +1622,27 @@ u64 hfi1_access_sw_tid_wait(const struct cntr_entry *entry,
 	return dd->verbs_dev.n_tidwait;
 }
 
+static struct tid_rdma_flow *find_flow_ib(struct tid_rdma_request *req,
+					  u32 psn, u16 *fidx)
+{
+	u16 head, tail;
+	struct tid_rdma_flow *flow;
+
+	head = req->setup_head;
+	tail = req->clear_tail;
+	for ( ; CIRC_CNT(head, tail, MAX_FLOWS);
+	     tail = CIRC_NEXT(tail, MAX_FLOWS)) {
+		flow = &req->flows[tail];
+		if (cmp_psn(psn, flow->flow_state.ib_spsn) >= 0 &&
+		    cmp_psn(psn, flow->flow_state.ib_lpsn) <= 0) {
+			if (fidx)
+				*fidx = tail;
+			return flow;
+		}
+	}
+	return NULL;
+}
+
 static struct tid_rdma_flow *
 __find_flow_ranged(struct tid_rdma_request *req, u16 head, u16 tail,
 		   u32 psn, u16 *fidx)
@@ -2714,3 +2735,64 @@ bool hfi1_handle_kdeth_eflags(struct hfi1_ctxtdata *rcd,
 drop:
 	return ret;
 }
+
+/*
+ * "Rewind" the TID request information.
+ * This means that we reset the state back to ACTIVE,
+ * find the proper flow, set the flow index to that flow,
+ * and reset the flow information.
+ */
+void hfi1_tid_rdma_restart_req(struct rvt_qp *qp, struct rvt_swqe *wqe,
+			       u32 *bth2)
+{
+	struct tid_rdma_request *req = wqe_to_tid_req(wqe);
+	struct tid_rdma_flow *flow;
+	int diff;
+	u32 tididx = 0;
+	u16 fidx;
+
+	if (wqe->wr.opcode == IB_WR_TID_RDMA_READ) {
+		*bth2 = mask_psn(qp->s_psn);
+		flow = find_flow_ib(req, *bth2, &fidx);
+		if (!flow)
+			return;
+	} else {
+		return;
+	}
+
+	diff = delta_psn(*bth2, flow->flow_state.ib_spsn);
+
+	flow->sent = 0;
+	flow->pkt = 0;
+	flow->tid_idx = 0;
+	flow->tid_offset = 0;
+	if (diff) {
+		for (tididx = 0; tididx < flow->tidcnt; tididx++) {
+			u32 tidentry = flow->tid_entry[tididx], tidlen,
+				tidnpkts, npkts;
+
+			flow->tid_offset = 0;
+			tidlen = EXP_TID_GET(tidentry, LEN) * PAGE_SIZE;
+			tidnpkts = rvt_div_round_up_mtu(qp, tidlen);
+			npkts = min_t(u32, diff, tidnpkts);
+			flow->pkt += npkts;
+			flow->sent += (npkts == tidnpkts ? tidlen :
+				       npkts * qp->pmtu);
+			flow->tid_offset += npkts * qp->pmtu;
+			diff -= npkts;
+			if (!diff)
+				break;
+		}
+	}
+
+	if (flow->tid_offset ==
+	    EXP_TID_GET(flow->tid_entry[tididx], LEN) * PAGE_SIZE) {
+		tididx++;
+		flow->tid_offset = 0;
+	}
+	flow->tid_idx = tididx;
+	/* Move flow_idx to correct index */
+	req->flow_idx = fidx;
+
+	req->state = TID_REQUEST_ACTIVE;
+}

drivers/infiniband/hw/hfi1/tid_rdma.h

@@ -207,5 +207,7 @@ void hfi1_rc_rcv_tid_rdma_read_resp(struct hfi1_packet *packet);
 bool hfi1_handle_kdeth_eflags(struct hfi1_ctxtdata *rcd,
 			      struct hfi1_pportdata *ppd,
 			      struct hfi1_packet *packet);
+void hfi1_tid_rdma_restart_req(struct rvt_qp *qp, struct rvt_swqe *wqe,
+			       u32 *bth2);
 
 #endif /* HFI1_TID_RDMA_H */