[PATCH v2] nvme-rdma: Use mr pool

[PATCH v2] nvme-rdma: Use mr pool

[Israel Rukshin]

Currently, blk_mq_tagset_iter() iterate over initial hctx tags only.
In case scheduler is used, it doesn't iterate the hctx scheduler tags
and the static request aren't been updated.
For example, while using NVMe over Fabrics RDMA host, this cause us not to
reinit the scheduler requests and thus not re-register all the memory regions
during the tagset re-initialization in the reconnect flow.

This may lead to a memory registration error:
"MEMREG for CQE 0xffff88044c14dce8 failed with status memory
management operation error (6)"

With this commit we don't need to reinit the requests.

Signed-off-by: Israel Rukshin <israelr at mellanox.com>
Reviewed-by: Max Gurtovoy <maxg at mellanox.com>
---

Changes from v1:
 - rebase the patch over Sagi patch series "Fix request completion holes"

---
 drivers/nvme/host/rdma.c | 94 +++++++++++++++++++-----------------------------
 1 file changed, 36 insertions(+), 58 deletions(-)

diff --git a/drivers/nvme/host/rdma.c b/drivers/nvme/host/rdma.c
index 8ed1343..944517d 100644
--- a/drivers/nvme/host/rdma.c
+++ b/drivers/nvme/host/rdma.c
@@ -15,6 +15,7 @@
 #include <linux/module.h>
 #include <linux/init.h>
 #include <linux/slab.h>
+#include <rdma/mr_pool.h>
 #include <linux/err.h>
 #include <linux/string.h>
 #include <linux/atomic.h>
@@ -269,32 +270,6 @@ static int nvme_rdma_create_qp(struct nvme_rdma_queue *queue, const int factor)
 	return ret;
 }
 
-static int nvme_rdma_reinit_request(void *data, struct request *rq)
-{
-	struct nvme_rdma_ctrl *ctrl = data;
-	struct nvme_rdma_device *dev = ctrl->device;
-	struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
-	int ret = 0;
-
-	if (WARN_ON_ONCE(!req->mr))
-		return 0;
-
-	ib_dereg_mr(req->mr);
-
-	req->mr = ib_alloc_mr(dev->pd, IB_MR_TYPE_MEM_REG,
-			ctrl->max_fr_pages);
-	if (IS_ERR(req->mr)) {
-		ret = PTR_ERR(req->mr);
-		req->mr = NULL;
-		goto out;
-	}
-
-	req->mr->need_inval = false;
-
-out:
-	return ret;
-}
-
 static void nvme_rdma_exit_request(struct blk_mq_tag_set *set,
 		struct request *rq, unsigned int hctx_idx)
 {
@@ -304,9 +279,6 @@ static void nvme_rdma_exit_request(struct blk_mq_tag_set *set,
 	struct nvme_rdma_queue *queue = &ctrl->queues[queue_idx];
 	struct nvme_rdma_device *dev = queue->device;
 
-	if (req->mr)
-		ib_dereg_mr(req->mr);
-
 	nvme_rdma_free_qe(dev->dev, &req->sqe, sizeof(struct nvme_command),
 			DMA_TO_DEVICE);
 }
@@ -330,21 +302,9 @@ static int nvme_rdma_init_request(struct blk_mq_tag_set *set,
 	if (ret)
 		return ret;
 
-	req->mr = ib_alloc_mr(dev->pd, IB_MR_TYPE_MEM_REG,
-			ctrl->max_fr_pages);
-	if (IS_ERR(req->mr)) {
-		ret = PTR_ERR(req->mr);
-		goto out_free_qe;
-	}
-
 	req->queue = queue;
 
 	return 0;
-
-out_free_qe:
-	nvme_rdma_free_qe(dev->dev, &req->sqe, sizeof(struct nvme_command),
-			DMA_TO_DEVICE);
-	return -ENOMEM;
 }
 
 static int nvme_rdma_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
@@ -444,6 +404,8 @@ static void nvme_rdma_destroy_queue_ib(struct nvme_rdma_queue *queue)
 	struct nvme_rdma_device *dev = queue->device;
 	struct ib_device *ibdev = dev->dev;
 
+	ib_mr_pool_destroy(queue->qp, &queue->qp->rdma_mrs);
+
 	rdma_destroy_qp(queue->cm_id);
 	ib_free_cq(queue->ib_cq);
 
@@ -495,8 +457,26 @@ static int nvme_rdma_create_queue_ib(struct nvme_rdma_queue *queue)
 		goto out_destroy_qp;
 	}
 
+	if (idx == 0)
+		queue->ctrl->max_fr_pages = min_t(u32, NVME_RDMA_MAX_SEGMENTS,
+				ibdev->attrs.max_fast_reg_page_list_len);
+
+	ret = ib_mr_pool_init(queue->qp, &queue->qp->rdma_mrs,
+			      queue->queue_size,
+			      IB_MR_TYPE_MEM_REG,
+			      queue->ctrl->max_fr_pages);
+	if (ret) {
+		dev_err(queue->ctrl->ctrl.device,
+			"failed to initialize MR pool sized %d for QID %d\n",
+			queue->queue_size, idx);
+		goto out_destroy_ring;
+	}
+
 	return 0;
 
+out_destroy_ring:
+	nvme_rdma_free_ring(ibdev, queue->rsp_ring, queue->queue_size,
+			    sizeof(struct nvme_completion), DMA_FROM_DEVICE);
 out_destroy_qp:
 	rdma_destroy_qp(queue->cm_id);
 out_destroy_ib_cq:
@@ -764,9 +744,6 @@ static int nvme_rdma_configure_admin_queue(struct nvme_rdma_ctrl *ctrl,
 
 	ctrl->device = ctrl->queues[0].device;
 
-	ctrl->max_fr_pages = min_t(u32, NVME_RDMA_MAX_SEGMENTS,
-		ctrl->device->dev->attrs.max_fast_reg_page_list_len);
-
 	if (new) {
 		ctrl->ctrl.admin_tagset = nvme_rdma_alloc_tagset(&ctrl->ctrl, true);
 		if (IS_ERR(ctrl->ctrl.admin_tagset))
@@ -777,10 +754,6 @@ static int nvme_rdma_configure_admin_queue(struct nvme_rdma_ctrl *ctrl,
 			error = PTR_ERR(ctrl->ctrl.admin_q);
 			goto out_free_tagset;
 		}
-	} else {
-		error = nvme_reinit_tagset(&ctrl->ctrl, ctrl->ctrl.admin_tagset);
-		if (error)
-			goto out_free_queue;
 	}
 
 	error = nvme_rdma_start_queue(ctrl, 0);
@@ -858,10 +831,6 @@ static int nvme_rdma_configure_io_queues(struct nvme_rdma_ctrl *ctrl, bool new)
 			goto out_free_tag_set;
 		}
 	} else {
-		ret = nvme_reinit_tagset(&ctrl->ctrl, ctrl->ctrl.tagset);
-		if (ret)
-			goto out_free_io_queues;
-
 		blk_mq_update_nr_hw_queues(&ctrl->tag_set,
 			ctrl->ctrl.queue_count - 1);
 	}
@@ -1070,6 +1039,11 @@ static void nvme_rdma_unmap_data(struct nvme_rdma_queue *queue,
 	if (!blk_rq_bytes(rq))
 		return;
 
+	if (req->mr) {
+		ib_mr_pool_put(queue->qp, &queue->qp->rdma_mrs, req->mr);
+		req->mr = NULL;
+	}
+
 	ib_dma_unmap_sg(ibdev, req->sg_table.sgl,
 			req->nents, rq_data_dir(rq) ==
 				    WRITE ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
@@ -1126,12 +1100,18 @@ static int nvme_rdma_map_sg_fr(struct nvme_rdma_queue *queue,
 	struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
 	int nr;
 
+	req->mr = ib_mr_pool_get(queue->qp, &queue->qp->rdma_mrs);
+	if (unlikely(!req->mr))
+		return -EAGAIN;
+
 	/*
 	 * Align the MR to a 4K page size to match the ctrl page size and
 	 * the block virtual boundary.
 	 */
 	nr = ib_map_mr_sg(req->mr, req->sg_table.sgl, count, NULL, SZ_4K);
 	if (unlikely(nr < count)) {
+		ib_mr_pool_put(queue->qp, &queue->qp->rdma_mrs, req->mr);
+		req->mr = NULL;
 		if (nr < 0)
 			return nr;
 		return -EINVAL;
@@ -1171,7 +1151,6 @@ static int nvme_rdma_map_data(struct nvme_rdma_queue *queue,
 
 	req->num_sge = 1;
 	req->inline_data = false;
-	req->mr->need_inval = false;
 	req->send_completed = false;
 	req->resp_completed = false;
 
@@ -1225,7 +1204,7 @@ static void nvme_rdma_send_done(struct ib_cq *cq, struct ib_wc *wc)
 
 	spin_lock_irqsave(&req->lock, flags);
 	req->send_completed = true;
-	end = req->resp_completed && !req->mr->need_inval;
+	end = req->resp_completed && !(req->mr && req->mr->need_inval);
 	spin_unlock_irqrestore(&req->lock, flags);
 	if (end)
 		nvme_end_request(rq, req->cqe.status, req->cqe.result);
@@ -1352,7 +1331,7 @@ static int nvme_rdma_process_nvme_rsp(struct nvme_rdma_queue *queue,
 	if ((wc->wc_flags & IB_WC_WITH_INVALIDATE) &&
 	    wc->ex.invalidate_rkey == req->mr->rkey) {
 		req->mr->need_inval = false;
-	} else if (req->mr->need_inval) {
+	} else if (req->mr && req->mr->need_inval) {
 		ret = nvme_rdma_inv_rkey(queue, req);
 		if (unlikely(ret < 0)) {
 			dev_err(queue->ctrl->ctrl.device,
@@ -1364,7 +1343,7 @@ static int nvme_rdma_process_nvme_rsp(struct nvme_rdma_queue *queue,
 
 	spin_lock_irqsave(&req->lock, flags);
 	req->resp_completed = true;
-	end = req->send_completed && !req->mr->need_inval;
+	end = req->send_completed && !(req->mr && req->mr->need_inval);
 	spin_unlock_irqrestore(&req->lock, flags);
 	if (end) {
 		if (rq->tag == tag)
@@ -1677,7 +1656,7 @@ static blk_status_t nvme_rdma_queue_rq(struct blk_mq_hw_ctx *hctx,
 			sizeof(struct nvme_command), DMA_TO_DEVICE);
 
 	err = nvme_rdma_post_send(queue, sqe, req->sge, req->num_sge,
-			req->mr->need_inval ? &req->reg_wr.wr : NULL, true);
+			req->mr ? &req->reg_wr.wr : NULL, true);
 	if (unlikely(err)) {
 		nvme_rdma_unmap_data(queue, rq);
 		goto err;
@@ -1825,7 +1804,6 @@ static void nvme_rdma_reset_ctrl_work(struct work_struct *work)
 	.submit_async_event	= nvme_rdma_submit_async_event,
 	.delete_ctrl		= nvme_rdma_delete_ctrl,
 	.get_address		= nvmf_get_address,
-	.reinit_request		= nvme_rdma_reinit_request,
 };
 
 static inline bool
-- 
1.8.3.1

[PATCH v2] nvme-rdma: Use mr pool

[Christoph Hellwig]

> @@ -495,8 +457,26 @@ static int nvme_rdma_create_queue_ib(struct nvme_rdma_queue *queue)
>  		goto out_destroy_qp;
>  	}
>  
> +	if (idx == 0)
> +		queue->ctrl->max_fr_pages = min_t(u32, NVME_RDMA_MAX_SEGMENTS,
> +				ibdev->attrs.max_fast_reg_page_list_len);
> +

Hmm.  Why is this code moved here from nvme_rdma_configure_admin_queue?
I really don't like per-controller initialization in per-queue functions
but if there is no good way around it it needs a very good explanation,
and that explanation should be in a comment place next to the code.

> +	req->mr = ib_mr_pool_get(queue->qp, &queue->qp->rdma_mrs);
> +	if (unlikely(!req->mr))
> +		return -EAGAIN;
> +

Shouldn't the MR pool be sized that this always succeeds?  If so a
WARN_ON_ONCE() here might be useful.

[PATCH v2] nvme-rdma: Use mr pool

[Sagi Grimberg]

>> @@ -495,8 +457,26 @@ static int nvme_rdma_create_queue_ib(struct nvme_rdma_queue *queue)
>>   		goto out_destroy_qp;
>>   	}
>>   
>> +	if (idx == 0)
>> +		queue->ctrl->max_fr_pages = min_t(u32, NVME_RDMA_MAX_SEGMENTS,
>> +				ibdev->attrs.max_fast_reg_page_list_len);
>> +
> 
> Hmm.  Why is this code moved here from nvme_rdma_configure_admin_queue?
> I really don't like per-controller initialization in per-queue functions
> but if there is no good way around it it needs a very good explanation,
> and that explanation should be in a comment place next to the code.

I agree, I'd even prefer it calculated here locally for the mr pool and
set it in its original place (perhaps add a simple helper for it).

>> +	req->mr = ib_mr_pool_get(queue->qp, &queue->qp->rdma_mrs);
>> +	if (unlikely(!req->mr))
>> +		return -EAGAIN;
>> +
> 
> Shouldn't the MR pool be sized that this always succeeds?  If so a
> WARN_ON_ONCE() here might be useful.

Agree.