Commit 5a5361ce authored by Alexey Marchuk's avatar Alexey Marchuk Committed by Konrad Sztyber
Browse files

accel/mlx5: Correct qp slot check, mkey allocation 



Number of mkeys allocated do not depend on the available
qp slot.
When a task is being executed, number of operations is limited
by qp slot. Take into account that crypto requires at least
2 qp slots

Signed-off-by: default avatarAlexey Marchuk <alexeymar@nvidia.com>
Change-Id: If33da6cbd591a94e6318bcfa582998d22cff4eb3
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/23125


Tested-by: default avatarSPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: default avatarBen Walker <ben@nvidia.com>
Reviewed-by: default avatarShuhei Matsumoto <smatsumoto@nvidia.com>
Community-CI: Mellanox Build Bot
parent 8d8e7cd0

module/accel/mlx5/accel_mlx5.c

+17 −13
Original line number Diff line number Diff line
@@ -296,15 +296,12 @@ accel_mlx5_dev_get_available_slots(struct accel_mlx5_dev *dev, struct accel_mlx5 
static inline uint32_t

accel_mlx5_task_alloc_mkeys(struct accel_mlx5_task *task)

{

	struct accel_mlx5_qp *qp = task->qp;

	struct accel_mlx5_dev *dev = qp->dev;

	uint32_t qp_slot = accel_mlx5_dev_get_available_slots(dev, qp);

	struct accel_mlx5_dev *dev = task->qp->dev;

	uint32_t num_ops;

	int rc;



	assert(task->num_reqs > task->num_completed_reqs);

	num_ops = task->num_reqs - task->num_completed_reqs;

	num_ops = spdk_min(num_ops, qp_slot);

	num_ops = spdk_min(num_ops, ACCEL_MLX5_MAX_MKEYS_IN_TASK);

	if (!num_ops) {

		return 0;
@@ -413,12 +410,15 @@ accel_mlx5_task_process(struct accel_mlx5_task *mlx5_task) 
	struct accel_mlx5_dev *dev = qp->dev;

	/* First RDMA after UMR must have a SMALL_FENCE */

	uint32_t first_rdma_fence = SPDK_MLX5_WQE_CTRL_INITIATOR_SMALL_FENCE;

	uint32_t num_blocks;

	uint32_t num_ops = spdk_min(mlx5_task->num_reqs - mlx5_task->num_completed_reqs,

	uint16_t num_blocks;

	uint16_t num_ops = spdk_min(mlx5_task->num_reqs - mlx5_task->num_completed_reqs,

				    mlx5_task->num_ops);

	uint32_t i;

	uint16_t qp_slot = accel_mlx5_dev_get_available_slots(dev, qp);

	uint16_t i;

	int rc;



	assert(qp_slot > 1);

	num_ops = spdk_min(num_ops, qp_slot >> 1);

	if (spdk_unlikely(!num_ops)) {

		return -EINVAL;

	}
@@ -500,11 +500,9 @@ accel_mlx5_task_continue(struct accel_mlx5_task *task) 
{

	struct accel_mlx5_qp *qp = task->qp;

	struct accel_mlx5_dev *dev = qp->dev;

	uint32_t qp_slot = accel_mlx5_dev_get_available_slots(dev, qp);

	uint32_t num_ops;

	uint16_t qp_slot = accel_mlx5_dev_get_available_slots(dev, qp);



	assert(task->num_reqs > task->num_completed_reqs);

	num_ops = task->num_reqs - task->num_completed_reqs;

	if (task->num_ops == 0) {

		/* No mkeys allocated, try to allocate now */

		if (spdk_unlikely(!accel_mlx5_task_alloc_mkeys(task))) {
@@ -513,9 +511,9 @@ accel_mlx5_task_continue(struct accel_mlx5_task *task) 
			return -ENOMEM;

		}

	}



	num_ops = spdk_min(num_ops, task->num_ops);

	if (spdk_unlikely(num_ops > qp_slot)) {

	/* We need to post at least 1 UMR and 1 RDMA operation */

	if (spdk_unlikely(qp_slot < 2)) {

		/* QP is full, queue this task */

		STAILQ_INSERT_TAIL(&dev->nomem, task, link);

		return -ENOMEM;

	}
@@ -602,6 +600,12 @@ accel_mlx5_task_init(struct accel_mlx5_task *mlx5_task, struct accel_mlx5_dev *d 
		SPDK_DEBUGLOG(accel_mlx5, "no reqs in pool, dev %s\n", dev->dev_ctx->context->device->name);

		return -ENOMEM;

	}

	if (spdk_unlikely(accel_mlx5_dev_get_available_slots(dev, &dev->qp) < 2)) {

		/* Queue is full, queue this task */

		SPDK_DEBUGLOG(accel_mlx5, "dev %s qp %p is full\n", dev->dev_ctx->context->device->name,

			      mlx5_task->qp);

		return -ENOMEM;

	}



	SPDK_DEBUGLOG(accel_mlx5, "task %p, src_iovs %u, dst_iovs %u, num_reqs %u, "

		      "blocks/req %u, blocks %u, inplace %d\n", task, task->s.iovcnt, task->d.iovcnt,