IB/mlx5: Reset flow support for IB kernel ULPs

The driver exposes interfaces that directly relate to HW state.
Upon fatal error, consumers of these interfaces (ULPs) that rely
on completion of all their posted work-request could hang, thereby
introducing dependencies in shutdown order. To prevent this from
happening, we manage the relevant resources (CQs, QPs) that are used
by the device. Upon a fatal error, we now generate simulated
completions for outstanding WQEs that were not completed at the
time the HW was reset.

It includes invoking the completion event handler for all involved
CQs so that the ULPs will poll those CQs. When polled we return
simulated CQEs with IB_WC_WR_FLUSH_ERR return code enabling ULPs
to clean up their  resources and not wait forever for completions
upon receiving remove_one.

The above change requires an extra check in the data path to make
sure that when device is in error state, the simulated CQEs will
be returned and no further WQEs will be posted.

Signed-off-by: Maor Gottlieb <maorg@mellanox.com>
Signed-off-by: Leon Romanovsky <leon@kernel.org>
Signed-off-by: Doug Ledford <dledford@redhat.com>

drivers/infiniband/hw/mlx5/cq.c

@@ -424,6 +424,83 @@ static void get_sig_err_item(struct mlx5_sig_err_cqe *cqe,
 	item->key = be32_to_cpu(cqe->mkey);
 }
 
+static void sw_send_comp(struct mlx5_ib_qp *qp, int num_entries,
+			 struct ib_wc *wc, int *npolled)
+{
+	struct mlx5_ib_wq *wq;
+	unsigned int cur;
+	unsigned int idx;
+	int np;
+	int i;
+
+	wq = &qp->sq;
+	cur = wq->head - wq->tail;
+	np = *npolled;
+
+	if (cur == 0)
+		return;
+
+	for (i = 0;  i < cur && np < num_entries; i++) {
+		idx = wq->last_poll & (wq->wqe_cnt - 1);
+		wc->wr_id = wq->wrid[idx];
+		wc->status = IB_WC_WR_FLUSH_ERR;
+		wc->vendor_err = MLX5_CQE_SYNDROME_WR_FLUSH_ERR;
+		wq->tail++;
+		np++;
+		wc->qp = &qp->ibqp;
+		wc++;
+		wq->last_poll = wq->w_list[idx].next;
+	}
+	*npolled = np;
+}
+
+static void sw_recv_comp(struct mlx5_ib_qp *qp, int num_entries,
+			 struct ib_wc *wc, int *npolled)
+{
+	struct mlx5_ib_wq *wq;
+	unsigned int cur;
+	int np;
+	int i;
+
+	wq = &qp->rq;
+	cur = wq->head - wq->tail;
+	np = *npolled;
+
+	if (cur == 0)
+		return;
+
+	for (i = 0;  i < cur && np < num_entries; i++) {
+		wc->wr_id = wq->wrid[wq->tail & (wq->wqe_cnt - 1)];
+		wc->status = IB_WC_WR_FLUSH_ERR;
+		wc->vendor_err = MLX5_CQE_SYNDROME_WR_FLUSH_ERR;
+		wq->tail++;
+		np++;
+		wc->qp = &qp->ibqp;
+		wc++;
+	}
+	*npolled = np;
+}
+
+static void mlx5_ib_poll_sw_comp(struct mlx5_ib_cq *cq, int num_entries,
+				 struct ib_wc *wc, int *npolled)
+{
+	struct mlx5_ib_qp *qp;
+
+	*npolled = 0;
+	/* Find uncompleted WQEs belonging to that cq and retrun mmics ones */
+	list_for_each_entry(qp, &cq->list_send_qp, cq_send_list) {
+		sw_send_comp(qp, num_entries, wc + *npolled, npolled);
+		if (*npolled >= num_entries)
+			return;
+	}
+
+	list_for_each_entry(qp, &cq->list_recv_qp, cq_recv_list) {
+		sw_recv_comp(qp, num_entries, wc + *npolled, npolled);
+		if (*npolled >= num_entries)
+			return;
+	}
+}
+
 static int mlx5_poll_one(struct mlx5_ib_cq *cq,
 			 struct mlx5_ib_qp **cur_qp,
 			 struct ib_wc *wc)
@@ -594,12 +671,18 @@ int mlx5_ib_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *wc)
 {
 	struct mlx5_ib_cq *cq = to_mcq(ibcq);
 	struct mlx5_ib_qp *cur_qp = NULL;
+	struct mlx5_ib_dev *dev = to_mdev(cq->ibcq.device);
+	struct mlx5_core_dev *mdev = dev->mdev;
 	unsigned long flags;
 	int soft_polled = 0;
 	int npolled;
 	int err = 0;
 
 	spin_lock_irqsave(&cq->lock, flags);
+	if (mdev->state == MLX5_DEVICE_STATE_INTERNAL_ERROR) {
+		mlx5_ib_poll_sw_comp(cq, num_entries, wc, &npolled);
+		goto out;
+	}
 
 	if (unlikely(!list_empty(&cq->wc_list)))
 		soft_polled = poll_soft_wc(cq, num_entries, wc);
@@ -612,7 +695,7 @@ int mlx5_ib_poll_cq(struct ib_cq *ibcq, int num_entries, struct ib_wc *wc)
 
 	if (npolled)
 		mlx5_cq_set_ci(&cq->mcq);
-
+out:
 	spin_unlock_irqrestore(&cq->lock, flags);
 
 	if (err == 0 || err == -EAGAIN)
@@ -843,6 +926,8 @@ struct ib_cq *mlx5_ib_create_cq(struct ib_device *ibdev,
 	cq->resize_buf = NULL;
 	cq->resize_umem = NULL;
 	cq->create_flags = attr->flags;
+	INIT_LIST_HEAD(&cq->list_send_qp);
+	INIT_LIST_HEAD(&cq->list_recv_qp);
 
 	if (context) {
 		err = create_cq_user(dev, udata, context, cq, entries,

drivers/infiniband/hw/mlx5/main.c

@@ -1980,6 +1980,65 @@ static void pkey_change_handler(struct work_struct *work)
 	mutex_unlock(&ports->devr->mutex);
 }
 
+static void mlx5_ib_handle_internal_error(struct mlx5_ib_dev *ibdev)
+{
+	struct mlx5_ib_qp *mqp;
+	struct mlx5_ib_cq *send_mcq, *recv_mcq;
+	struct mlx5_core_cq *mcq;
+	struct list_head cq_armed_list;
+	unsigned long flags_qp;
+	unsigned long flags_cq;
+	unsigned long flags;
+
+	INIT_LIST_HEAD(&cq_armed_list);
+
+	/* Go over qp list reside on that ibdev, sync with create/destroy qp.*/
+	spin_lock_irqsave(&ibdev->reset_flow_resource_lock, flags);
+	list_for_each_entry(mqp, &ibdev->qp_list, qps_list) {
+		spin_lock_irqsave(&mqp->sq.lock, flags_qp);
+		if (mqp->sq.tail != mqp->sq.head) {
+			send_mcq = to_mcq(mqp->ibqp.send_cq);
+			spin_lock_irqsave(&send_mcq->lock, flags_cq);
+			if (send_mcq->mcq.comp &&
+			    mqp->ibqp.send_cq->comp_handler) {
+				if (!send_mcq->mcq.reset_notify_added) {
+					send_mcq->mcq.reset_notify_added = 1;
+					list_add_tail(&send_mcq->mcq.reset_notify,
+						      &cq_armed_list);
+				}
+			}
+			spin_unlock_irqrestore(&send_mcq->lock, flags_cq);
+		}
+		spin_unlock_irqrestore(&mqp->sq.lock, flags_qp);
+		spin_lock_irqsave(&mqp->rq.lock, flags_qp);
+		/* no handling is needed for SRQ */
+		if (!mqp->ibqp.srq) {
+			if (mqp->rq.tail != mqp->rq.head) {
+				recv_mcq = to_mcq(mqp->ibqp.recv_cq);
+				spin_lock_irqsave(&recv_mcq->lock, flags_cq);
+				if (recv_mcq->mcq.comp &&
+				    mqp->ibqp.recv_cq->comp_handler) {
+					if (!recv_mcq->mcq.reset_notify_added) {
+						recv_mcq->mcq.reset_notify_added = 1;
+						list_add_tail(&recv_mcq->mcq.reset_notify,
+							      &cq_armed_list);
+					}
+				}
+				spin_unlock_irqrestore(&recv_mcq->lock,
+						       flags_cq);
+			}
+		}
+		spin_unlock_irqrestore(&mqp->rq.lock, flags_qp);
+	}
+	/*At that point all inflight post send were put to be executed as of we
+	 * lock/unlock above locks Now need to arm all involved CQs.
+	 */
+	list_for_each_entry(mcq, &cq_armed_list, reset_notify) {
+		mcq->comp(mcq);
+	}
+	spin_unlock_irqrestore(&ibdev->reset_flow_resource_lock, flags);
+}
+
 static void mlx5_ib_event(struct mlx5_core_dev *dev, void *context,
 			  enum mlx5_dev_event event, unsigned long param)
 {
@@ -1992,6 +2051,7 @@ static void mlx5_ib_event(struct mlx5_core_dev *dev, void *context,
 	case MLX5_DEV_EVENT_SYS_ERROR:
 		ibdev->ib_active = false;
 		ibev.event = IB_EVENT_DEVICE_FATAL;
+		mlx5_ib_handle_internal_error(ibdev);
 		break;
 
 	case MLX5_DEV_EVENT_PORT_UP:
@@ -2595,6 +2655,8 @@ static void *mlx5_ib_add(struct mlx5_core_dev *mdev)
 
 	mutex_init(&dev->flow_db.lock);
 	mutex_init(&dev->cap_mask_mutex);
+	INIT_LIST_HEAD(&dev->qp_list);
+	spin_lock_init(&dev->reset_flow_resource_lock);
 
 	if (ll == IB_LINK_LAYER_ETHERNET) {
 		err = mlx5_enable_roce(dev);

drivers/infiniband/hw/mlx5/mlx5_ib.h

@@ -380,6 +380,9 @@ struct mlx5_ib_qp {
 	spinlock_t              disable_page_faults_lock;
 	struct mlx5_ib_pfault	pagefaults[MLX5_IB_PAGEFAULT_CONTEXTS];
 #endif
+	struct list_head	qps_list;
+	struct list_head	cq_recv_list;
+	struct list_head	cq_send_list;
 };
 
 struct mlx5_ib_cq_buf {
@@ -441,6 +444,8 @@ struct mlx5_ib_cq {
 	struct mlx5_ib_cq_buf  *resize_buf;
 	struct ib_umem	       *resize_umem;
 	int			cqe_size;
+	struct list_head	list_send_qp;
+	struct list_head	list_recv_qp;
 	u32			create_flags;
 	struct list_head	wc_list;
 	enum ib_cq_notify_flags notify_flags;
@@ -621,6 +626,9 @@ struct mlx5_ib_dev {
 	struct srcu_struct      mr_srcu;
 #endif
 	struct mlx5_ib_flow_db	flow_db;
+	/* protect resources needed as part of reset flow */
+	spinlock_t		reset_flow_resource_lock;
+	struct list_head	qp_list;
 };
 
 static inline struct mlx5_ib_cq *to_mibcq(struct mlx5_core_cq *mcq)

drivers/infiniband/hw/mlx5/mr.c

@@ -1193,12 +1193,16 @@ struct ib_mr *mlx5_ib_reg_user_mr(struct ib_pd *pd, u64 start, u64 length,
 
 static int unreg_umr(struct mlx5_ib_dev *dev, struct mlx5_ib_mr *mr)
 {
+	struct mlx5_core_dev *mdev = dev->mdev;
 	struct umr_common *umrc = &dev->umrc;
 	struct mlx5_ib_umr_context umr_context;
 	struct mlx5_umr_wr umrwr = {};
 	struct ib_send_wr *bad;
 	int err;
 
+	if (mdev->state == MLX5_DEVICE_STATE_INTERNAL_ERROR)
+		return 0;
+
 	mlx5_ib_init_umr_context(&umr_context);
 
 	umrwr.wr.wr_cqe = &umr_context.cqe;