iw_cxgb4: Check for send WR also while posting write with completion WR

Inorder to optimize the NVMEoF read IOPs, iw_cxgb4 posts a FW Write with
Completion WQE that combines an RDMA Write WR and the subsequent RDMA Send
with Invalidate WR.

This patch is an extension to it, where it posts a Write with completion
for RDMA WRITE WR + RDMA SEND WR combination as well.
Reviewed-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: Potnuri Bharat Teja <bharat@chelsio.com>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>

drivers/infiniband/hw/cxgb4/qp.c

@@ -632,6 +632,9 @@ static void build_rdma_write_cmpl(struct t4_sq *sq,
 
 	wcwr->stag_sink = cpu_to_be32(rdma_wr(wr)->rkey);
 	wcwr->to_sink = cpu_to_be64(rdma_wr(wr)->remote_addr);
+	if (wr->next->opcode == IB_WR_SEND)
+		wcwr->stag_inv = 0;
+	else
 		wcwr->stag_inv = cpu_to_be32(wr->next->ex.invalidate_rkey);
 	wcwr->r2 = 0;
 	wcwr->r3 = 0;
@@ -726,6 +729,9 @@ static void post_write_cmpl(struct c4iw_qp *qhp, const struct ib_send_wr *wr)
 
 	/* SEND_WITH_INV swsqe */
 	swsqe = &qhp->wq.sq.sw_sq[qhp->wq.sq.pidx];
+	if (wr->next->opcode == IB_WR_SEND)
+		swsqe->opcode = FW_RI_SEND;
+	else
 		swsqe->opcode = FW_RI_SEND_WITH_INV;
 	swsqe->idx = qhp->wq.sq.pidx;
 	swsqe->complete = 0;
@@ -1133,9 +1139,9 @@ int c4iw_post_send(struct ib_qp *ibqp, const struct ib_send_wr *wr,
 	/*
 	 * Fastpath for NVMe-oF target WRITE + SEND_WITH_INV wr chain which is
 	 * the response for small NVMEe-oF READ requests.  If the chain is
-	 * exactly a WRITE->SEND_WITH_INV and the sgl depths and lengths
-	 * meet the requirements of the fw_ri_write_cmpl_wr work request,
-	 * then build and post the write_cmpl WR.  If any of the tests
+	 * exactly a WRITE->SEND_WITH_INV or a WRITE->SEND and the sgl depths
+	 * and lengths meet the requirements of the fw_ri_write_cmpl_wr work
+	 * request, then build and post the write_cmpl WR. If any of the tests
 	 * below are not true, then we continue on with the tradtional WRITE
 	 * and SEND WRs.
 	 */
@@ -1145,7 +1151,8 @@ int c4iw_post_send(struct ib_qp *ibqp, const struct ib_send_wr *wr,
 	    wr && wr->next && !wr->next->next &&
 	    wr->opcode == IB_WR_RDMA_WRITE &&
 	    wr->sg_list[0].length && wr->num_sge <= T4_WRITE_CMPL_MAX_SGL &&
-	    wr->next->opcode == IB_WR_SEND_WITH_INV &&
+	    (wr->next->opcode == IB_WR_SEND ||
+	    wr->next->opcode == IB_WR_SEND_WITH_INV) &&
 	    wr->next->sg_list[0].length == T4_WRITE_CMPL_MAX_CQE &&
 	    wr->next->num_sge == 1 && num_wrs >= 2) {
 		post_write_cmpl(qhp, wr);