[PATCH 0/2] NVMe/RDMA CQ pool patches for 5.9

[PATCH 0/2] NVMe/RDMA CQ pool patches for 5.9

[Max Gurtovoy]

This series include 2 patches from Yamin that were removed from the
merge window of 5.8 since it caused conflicts between RDMA and Block
trees. It uses a shared CQ API that was merged to RDMA core layer to
improve performance and reduce resource allocation.

Yamin Friedman (2):
  nvme-rdma: use new shared CQ mechanism
  nvmet-rdma: use new shared CQ mechanism

 drivers/nvme/host/rdma.c   | 77 ++++++++++++++++++++++++++++++----------------
 drivers/nvme/target/rdma.c | 14 ++++-----
 2 files changed, 58 insertions(+), 33 deletions(-)

-- 
1.8.3.1


_______________________________________________
Linux-nvme mailing list
Linux-nvme@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-nvme

[PATCH 1/2] nvme-rdma: use new shared CQ mechanism

[Max Gurtovoy]

From: Yamin Friedman <yaminf@mellanox.com>

Has the driver use shared CQs providing ~10%-20% improvement as seen in
the patch introducing shared CQs. Instead of opening a CQ for each QP
per controller connected, a CQ for each QP will be provided by the RDMA
core driver that will be shared between the QPs on that core reducing
interrupt overhead.

Signed-off-by: Yamin Friedman <yaminf@mellanox.com>
Signed-off-by: Max Gurtovoy <maxg@mellanox.com>
Reviewed-by: Or Gerlitz <ogerlitz@mellanox.com>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
---
 drivers/nvme/host/rdma.c | 77 ++++++++++++++++++++++++++++++++----------------
 1 file changed, 51 insertions(+), 26 deletions(-)

diff --git a/drivers/nvme/host/rdma.c b/drivers/nvme/host/rdma.c
index e881f87..467da08 100644
--- a/drivers/nvme/host/rdma.c
+++ b/drivers/nvme/host/rdma.c
@@ -96,6 +96,7 @@ struct nvme_rdma_queue {
 	int			cm_error;
 	struct completion	cm_done;
 	bool			pi_support;
+	int			cq_size;
 };
 
 struct nvme_rdma_ctrl {
@@ -275,6 +276,7 @@ static int nvme_rdma_create_qp(struct nvme_rdma_queue *queue, const int factor)
 	init_attr.recv_cq = queue->ib_cq;
 	if (queue->pi_support)
 		init_attr.create_flags |= IB_QP_CREATE_INTEGRITY_EN;
+	init_attr.qp_context = queue;
 
 	ret = rdma_create_qp(queue->cm_id, dev->pd, &init_attr);
 
@@ -409,6 +411,14 @@ static int nvme_rdma_dev_get(struct nvme_rdma_device *dev)
 	return NULL;
 }
 
+static void nvme_rdma_free_cq(struct nvme_rdma_queue *queue)
+{
+	if (nvme_rdma_poll_queue(queue))
+		ib_free_cq(queue->ib_cq);
+	else
+		ib_cq_pool_put(queue->ib_cq, queue->cq_size);
+}
+
 static void nvme_rdma_destroy_queue_ib(struct nvme_rdma_queue *queue)
 {
 	struct nvme_rdma_device *dev;
@@ -430,7 +440,7 @@ static void nvme_rdma_destroy_queue_ib(struct nvme_rdma_queue *queue)
 	 * the destruction of the QP shouldn't use rdma_cm API.
 	 */
 	ib_destroy_qp(queue->qp);
-	ib_free_cq(queue->ib_cq);
+	nvme_rdma_free_cq(queue);
 
 	nvme_rdma_free_ring(ibdev, queue->rsp_ring, queue->queue_size,
 			sizeof(struct nvme_completion), DMA_FROM_DEVICE);
@@ -450,13 +460,42 @@ static int nvme_rdma_get_max_fr_pages(struct ib_device *ibdev, bool pi_support)
 	return min_t(u32, NVME_RDMA_MAX_SEGMENTS, max_page_list_len - 1);
 }
 
+static int nvme_rdma_create_cq(struct ib_device *ibdev,
+		struct nvme_rdma_queue *queue)
+{
+	int ret, comp_vector, idx = nvme_rdma_queue_idx(queue);
+	enum ib_poll_context poll_ctx;
+
+	/*
+	 * Spread I/O queues completion vectors according their queue index.
+	 * Admin queues can always go on completion vector 0.
+	 */
+	comp_vector = (idx == 0 ? idx : idx - 1) % ibdev->num_comp_vectors;
+
+	/* Polling queues need direct cq polling context */
+	if (nvme_rdma_poll_queue(queue)) {
+		poll_ctx = IB_POLL_DIRECT;
+		queue->ib_cq = ib_alloc_cq(ibdev, queue, queue->cq_size,
+					   comp_vector, poll_ctx);
+	} else {
+		poll_ctx = IB_POLL_SOFTIRQ;
+		queue->ib_cq = ib_cq_pool_get(ibdev, queue->cq_size,
+					      comp_vector, poll_ctx);
+	}
+
+	if (IS_ERR(queue->ib_cq)) {
+		ret = PTR_ERR(queue->ib_cq);
+		return ret;
+	}
+
+	return 0;
+}
+
 static int nvme_rdma_create_queue_ib(struct nvme_rdma_queue *queue)
 {
 	struct ib_device *ibdev;
 	const int send_wr_factor = 3;			/* MR, SEND, INV */
 	const int cq_factor = send_wr_factor + 1;	/* + RECV */
-	int comp_vector, idx = nvme_rdma_queue_idx(queue);
-	enum ib_poll_context poll_ctx;
 	int ret, pages_per_mr;
 
 	queue->device = nvme_rdma_find_get_device(queue->cm_id);
@@ -467,26 +506,12 @@ static int nvme_rdma_create_queue_ib(struct nvme_rdma_queue *queue)
 	}
 	ibdev = queue->device->dev;
 
-	/*
-	 * Spread I/O queues completion vectors according their queue index.
-	 * Admin queues can always go on completion vector 0.
-	 */
-	comp_vector = (idx == 0 ? idx : idx - 1) % ibdev->num_comp_vectors;
-
-	/* Polling queues need direct cq polling context */
-	if (nvme_rdma_poll_queue(queue))
-		poll_ctx = IB_POLL_DIRECT;
-	else
-		poll_ctx = IB_POLL_SOFTIRQ;
-
 	/* +1 for ib_stop_cq */
-	queue->ib_cq = ib_alloc_cq(ibdev, queue,
-				cq_factor * queue->queue_size + 1,
-				comp_vector, poll_ctx);
-	if (IS_ERR(queue->ib_cq)) {
-		ret = PTR_ERR(queue->ib_cq);
+	queue->cq_size = cq_factor * queue->queue_size + 1;
+
+	ret = nvme_rdma_create_cq(ibdev, queue);
+	if (ret)
 		goto out_put_dev;
-	}
 
 	ret = nvme_rdma_create_qp(queue, send_wr_factor);
 	if (ret)
@@ -512,7 +537,7 @@ static int nvme_rdma_create_queue_ib(struct nvme_rdma_queue *queue)
 	if (ret) {
 		dev_err(queue->ctrl->ctrl.device,
 			"failed to initialize MR pool sized %d for QID %d\n",
-			queue->queue_size, idx);
+			queue->queue_size, nvme_rdma_queue_idx(queue));
 		goto out_destroy_ring;
 	}
 
@@ -523,7 +548,7 @@ static int nvme_rdma_create_queue_ib(struct nvme_rdma_queue *queue)
 		if (ret) {
 			dev_err(queue->ctrl->ctrl.device,
 				"failed to initialize PI MR pool sized %d for QID %d\n",
-				queue->queue_size, idx);
+				queue->queue_size, nvme_rdma_queue_idx(queue));
 			goto out_destroy_mr_pool;
 		}
 	}
@@ -540,7 +565,7 @@ static int nvme_rdma_create_queue_ib(struct nvme_rdma_queue *queue)
 out_destroy_qp:
 	rdma_destroy_qp(queue->cm_id);
 out_destroy_ib_cq:
-	ib_free_cq(queue->ib_cq);
+	nvme_rdma_free_cq(queue);
 out_put_dev:
 	nvme_rdma_dev_put(queue->device);
 	return ret;
@@ -1163,7 +1188,7 @@ static void nvme_rdma_end_request(struct nvme_rdma_request *req)
 static void nvme_rdma_wr_error(struct ib_cq *cq, struct ib_wc *wc,
 		const char *op)
 {
-	struct nvme_rdma_queue *queue = cq->cq_context;
+	struct nvme_rdma_queue *queue = wc->qp->qp_context;
 	struct nvme_rdma_ctrl *ctrl = queue->ctrl;
 
 	if (ctrl->ctrl.state == NVME_CTRL_LIVE)
@@ -1706,7 +1731,7 @@ static void nvme_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc)
 {
 	struct nvme_rdma_qe *qe =
 		container_of(wc->wr_cqe, struct nvme_rdma_qe, cqe);
-	struct nvme_rdma_queue *queue = cq->cq_context;
+	struct nvme_rdma_queue *queue = wc->qp->qp_context;
 	struct ib_device *ibdev = queue->device->dev;
 	struct nvme_completion *cqe = qe->data;
 	const size_t len = sizeof(struct nvme_completion);
-- 
1.8.3.1


_______________________________________________
Linux-nvme mailing list
Linux-nvme@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-nvme

[PATCH 2/2] nvmet-rdma: use new shared CQ mechanism

[Max Gurtovoy]

From: Yamin Friedman <yaminf@mellanox.com>

Has the driver use shared CQs providing ~10%-20% improvement when
multiple disks are used. Instead of opening a CQ for each QP per
controller, a CQ for each core will be provided by the RDMA core driver
that will be shared between the QPs on that core reducing interrupt
overhead.

Signed-off-by: Yamin Friedman <yaminf@mellanox.com>
Signed-off-by: Max Gurtovoy <maxg@mellanox.com>
Reviewed-by: Or Gerlitz <ogerlitz@mellanox.com>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
---
 drivers/nvme/target/rdma.c | 14 +++++++-------
 1 file changed, 7 insertions(+), 7 deletions(-)

diff --git a/drivers/nvme/target/rdma.c b/drivers/nvme/target/rdma.c
index 6731e03..3ccb592 100644
--- a/drivers/nvme/target/rdma.c
+++ b/drivers/nvme/target/rdma.c
@@ -752,7 +752,7 @@ static void nvmet_rdma_read_data_done(struct ib_cq *cq, struct ib_wc *wc)
 {
 	struct nvmet_rdma_rsp *rsp =
 		container_of(wc->wr_cqe, struct nvmet_rdma_rsp, read_cqe);
-	struct nvmet_rdma_queue *queue = cq->cq_context;
+	struct nvmet_rdma_queue *queue = wc->qp->qp_context;
 	u16 status = 0;
 
 	WARN_ON(rsp->n_rdma <= 0);
@@ -1008,7 +1008,7 @@ static void nvmet_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc)
 {
 	struct nvmet_rdma_cmd *cmd =
 		container_of(wc->wr_cqe, struct nvmet_rdma_cmd, cqe);
-	struct nvmet_rdma_queue *queue = cq->cq_context;
+	struct nvmet_rdma_queue *queue = wc->qp->qp_context;
 	struct nvmet_rdma_rsp *rsp;
 
 	if (unlikely(wc->status != IB_WC_SUCCESS)) {
@@ -1258,9 +1258,8 @@ static int nvmet_rdma_create_queue_ib(struct nvmet_rdma_queue *queue)
 	 */
 	nr_cqe = queue->recv_queue_size + 2 * queue->send_queue_size;
 
-	queue->cq = ib_alloc_cq(ndev->device, queue,
-			nr_cqe + 1, queue->comp_vector,
-			IB_POLL_WORKQUEUE);
+	queue->cq = ib_cq_pool_get(ndev->device, nr_cqe + 1,
+				   queue->comp_vector, IB_POLL_WORKQUEUE);
 	if (IS_ERR(queue->cq)) {
 		ret = PTR_ERR(queue->cq);
 		pr_err("failed to create CQ cqe= %d ret= %d\n",
@@ -1322,7 +1321,7 @@ static int nvmet_rdma_create_queue_ib(struct nvmet_rdma_queue *queue)
 err_destroy_qp:
 	rdma_destroy_qp(queue->cm_id);
 err_destroy_cq:
-	ib_free_cq(queue->cq);
+	ib_cq_pool_put(queue->cq, nr_cqe + 1);
 	goto out;
 }
 
@@ -1332,7 +1331,8 @@ static void nvmet_rdma_destroy_queue_ib(struct nvmet_rdma_queue *queue)
 	if (queue->cm_id)
 		rdma_destroy_id(queue->cm_id);
 	ib_destroy_qp(queue->qp);
-	ib_free_cq(queue->cq);
+	ib_cq_pool_put(queue->cq, queue->recv_queue_size + 2 *
+		       queue->send_queue_size + 1);
 }
 
 static void nvmet_rdma_free_queue(struct nvmet_rdma_queue *queue)
-- 
1.8.3.1


_______________________________________________
Linux-nvme mailing list
Linux-nvme@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-nvme

Re: [PATCH 0/2] NVMe/RDMA CQ pool patches for 5.9

[Christoph Hellwig]

On Mon, Jul 13, 2020 at 11:53:28AM +0300, Max Gurtovoy wrote:
> This series include 2 patches from Yamin that were removed from the
> merge window of 5.8 since it caused conflicts between RDMA and Block
> trees. It uses a shared CQ API that was merged to RDMA core layer to
> improve performance and reduce resource allocation.

Thanks,

applied to nvme-5.9.

_______________________________________________
Linux-nvme mailing list
Linux-nvme@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-nvme