diff --git a/lib/nvmf/rdma.c b/lib/nvmf/rdma.c
index e8ad4416a..1c73ce90f 100644
--- a/lib/nvmf/rdma.c
+++ b/lib/nvmf/rdma.c
@@ -1748,19 +1748,6 @@ nvmf_rdma_request_fill_iovs_multi_sgl(struct spdk_nvmf_rdma_transport *rtranspor
 		return -ENOMEM;
 	}
 
-	desc = (struct spdk_nvme_sgl_descriptor *)rdma_req->recv->buf + inline_segment->address;
-
-	for (i = 0; i < num_sgl_descriptors; i++) {
-		num_buffers += SPDK_CEIL_DIV(desc->keyed.length, rtransport->transport.opts.io_unit_size);
-		desc++;
-	}
-	rc = spdk_nvmf_request_get_buffers(req, &rgroup->group, &rtransport->transport,
-					   num_buffers);
-	if (rc != 0) {
-		nvmf_rdma_request_free_data(rdma_req, rtransport);
-		return rc;
-	}
-
 	/* The first WR must always be the embedded data WR. This is how we unwind them later. */
 	current_wr = &rdma_req->data.wr;
 	assert(current_wr != NULL);
@@ -1776,6 +1763,13 @@ nvmf_rdma_request_fill_iovs_multi_sgl(struct spdk_nvmf_rdma_transport *rtranspor
 			goto err_exit;
 		}
 
+		num_buffers = SPDK_CEIL_DIV(desc->keyed.length, rtransport->transport.opts.io_unit_size);
+		rc = spdk_nvmf_request_get_buffers(req, &rgroup->group, &rtransport->transport,
+						   num_buffers);
+		if (rc != 0) {
+			goto err_exit;
+		}
+
 		current_wr->num_sge = 0;
 
 		rc = nvmf_rdma_fill_buffers(rtransport, rgroup, device, req, current_wr,