1 files changed, 36 insertions, 17 deletions

diff --git a/drivers/nvme/host/rdma.c b/drivers/nvme/host/rdma.c
index 26709a2ab593..97f668a39ae1 100644
--- a/drivers/nvme/host/rdma.c
+++ b/drivers/nvme/host/rdma.c
@@ -213,6 +213,11 @@ static struct nvme_rdma_qe *nvme_rdma_alloc_ring(struct ib_device *ibdev,
 	if (!ring)
 		return NULL;
 
+	/*
+	 * Bind the CQEs (post recv buffers) DMA mapping to the RDMA queue
+	 * lifetime. It's safe, since any chage in the underlying RDMA device
+	 * will issue error recovery and queue re-creation.
+	 */
 	for (i = 0; i < ib_queue_size; i++) {
 		if (nvme_rdma_alloc_qe(ibdev, &ring[i], capsule_size, dir))
 			goto out_free_ring;
@@ -274,14 +279,9 @@ static int nvme_rdma_create_qp(struct nvme_rdma_queue *queue, const int factor)
 static void nvme_rdma_exit_request(struct blk_mq_tag_set *set,
 		struct request *rq, unsigned int hctx_idx)
 {
-	struct nvme_rdma_ctrl *ctrl = set->driver_data;
 	struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
-	int queue_idx = (set == &ctrl->tag_set) ? hctx_idx + 1 : 0;
-	struct nvme_rdma_queue *queue = &ctrl->queues[queue_idx];
-	struct nvme_rdma_device *dev = queue->device;
 
-	nvme_rdma_free_qe(dev->dev, &req->sqe, sizeof(struct nvme_command),
-			DMA_TO_DEVICE);
+	kfree(req->sqe.data);
 }
 
 static int nvme_rdma_init_request(struct blk_mq_tag_set *set,
@@ -292,15 +292,11 @@ static int nvme_rdma_init_request(struct blk_mq_tag_set *set,
 	struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
 	int queue_idx = (set == &ctrl->tag_set) ? hctx_idx + 1 : 0;
 	struct nvme_rdma_queue *queue = &ctrl->queues[queue_idx];
-	struct nvme_rdma_device *dev = queue->device;
-	struct ib_device *ibdev = dev->dev;
-	int ret;
 
 	nvme_req(rq)->ctrl = &ctrl->ctrl;
-	ret = nvme_rdma_alloc_qe(ibdev, &req->sqe, sizeof(struct nvme_command),
-			DMA_TO_DEVICE);
-	if (ret)
-		return ret;
+	req->sqe.data = kzalloc(sizeof(struct nvme_command), GFP_KERNEL);
+	if (!req->sqe.data)
+		return -ENOMEM;
 
 	req->queue = queue;
 
@@ -779,6 +775,11 @@ static int nvme_rdma_configure_admin_queue(struct nvme_rdma_ctrl *ctrl,
 
 	ctrl->max_fr_pages = nvme_rdma_get_max_fr_pages(ctrl->device->dev);
 
+	/*
+	 * Bind the async event SQE DMA mapping to the admin queue lifetime.
+	 * It's safe, since any chage in the underlying RDMA device will issue
+	 * error recovery and queue re-creation.
+	 */
 	error = nvme_rdma_alloc_qe(ctrl->device->dev, &ctrl->async_event_sqe,
 			sizeof(struct nvme_command), DMA_TO_DEVICE);
 	if (error)
@@ -1719,12 +1720,20 @@ static blk_status_t nvme_rdma_queue_rq(struct blk_mq_hw_ctx *hctx,
 		return nvmf_fail_nonready_command(&queue->ctrl->ctrl, rq);
 
 	dev = queue->device->dev;
+
+	req->sqe.dma = ib_dma_map_single(dev, req->sqe.data,
+					 sizeof(struct nvme_command),
+					 DMA_TO_DEVICE);
+	err = ib_dma_mapping_error(dev, req->sqe.dma);
+	if (unlikely(err))
+		return BLK_STS_RESOURCE;
+
 	ib_dma_sync_single_for_cpu(dev, sqe->dma,
 			sizeof(struct nvme_command), DMA_TO_DEVICE);
 
 	ret = nvme_setup_cmd(ns, rq, c);
 	if (ret)
-		return ret;
+		goto unmap_qe;
 
 	blk_mq_start_request(rq);
 
@@ -1749,10 +1758,16 @@ static blk_status_t nvme_rdma_queue_rq(struct blk_mq_hw_ctx *hctx,
 	}
 
 	return BLK_STS_OK;
+
 err:
 	if (err == -ENOMEM || err == -EAGAIN)
-		return BLK_STS_RESOURCE;
-	return BLK_STS_IOERR;
+		ret = BLK_STS_RESOURCE;
+	else
+		ret = BLK_STS_IOERR;
+unmap_qe:
+	ib_dma_unmap_single(dev, req->sqe.dma, sizeof(struct nvme_command),
+			    DMA_TO_DEVICE);
+	return ret;
 }
 
 static int nvme_rdma_poll(struct blk_mq_hw_ctx *hctx)
@@ -1765,8 +1780,12 @@ static int nvme_rdma_poll(struct blk_mq_hw_ctx *hctx)
 static void nvme_rdma_complete_rq(struct request *rq)
 {
 	struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
+	struct nvme_rdma_queue *queue = req->queue;
+	struct ib_device *ibdev = queue->device->dev;
 
-	nvme_rdma_unmap_data(req->queue, rq);
+	nvme_rdma_unmap_data(queue, rq);
+	ib_dma_unmap_single(ibdev, req->sqe.dma, sizeof(struct nvme_command),
+			    DMA_TO_DEVICE);
 	nvme_complete_rq(rq);
 }