IB/mlx5: Page faults handling infrastructure

* Refactor MR registration and cleanup, and fix reg_pages accounting.
* Create a work queue to handle page fault events in a kthread context.
* Register a fault handler to get events from the core for each QP.

The registered fault handler is empty in this patch, and only a later
patch implements it.

Signed-off-by: Sagi Grimberg <[email protected]>
Signed-off-by: Shachar Raindel <[email protected]>
Signed-off-by: Haggai Eran <[email protected]>
Signed-off-by: Roland Dreier <[email protected]>

Author: haggaie

drivers/infiniband/hw/mlx5/main.c

@@ -864,7 +864,7 @@ static ssize_t show_reg_pages(struct device *device,
 	struct mlx5_ib_dev *dev =
 		container_of(device, struct mlx5_ib_dev, ib_dev.dev);
 
-	return sprintf(buf, "%d\n", dev->mdev->priv.reg_pages);
+	return sprintf(buf, "%d\n", atomic_read(&dev->mdev->priv.reg_pages));
 }
 
 static ssize_t show_hca(struct device *device, struct device_attribute *attr,
@@ -1389,16 +1389,19 @@ static void *mlx5_ib_add(struct mlx5_core_dev *mdev)
 		goto err_eqs;
 
 	mutex_init(&dev->cap_mask_mutex);
-	spin_lock_init(&dev->mr_lock);
 
 	err = create_dev_resources(&dev->devr);
 	if (err)
 		goto err_eqs;
 
-	err = ib_register_device(&dev->ib_dev, NULL);
+	err = mlx5_ib_odp_init_one(dev);
 	if (err)
 		goto err_rsrc;
 
+	err = ib_register_device(&dev->ib_dev, NULL);
+	if (err)
+		goto err_odp;
+
 	err = create_umr_res(dev);
 	if (err)
 		goto err_dev;
@@ -1420,6 +1423,9 @@ static void *mlx5_ib_add(struct mlx5_core_dev *mdev)
 err_dev:
 	ib_unregister_device(&dev->ib_dev);
 
+err_odp:
+	mlx5_ib_odp_remove_one(dev);
+
 err_rsrc:
 	destroy_dev_resources(&dev->devr);
 
@@ -1435,8 +1441,10 @@ static void *mlx5_ib_add(struct mlx5_core_dev *mdev)
 static void mlx5_ib_remove(struct mlx5_core_dev *mdev, void *context)
 {
 	struct mlx5_ib_dev *dev = context;
+
 	ib_unregister_device(&dev->ib_dev);
 	destroy_umrc_res(dev);
+	mlx5_ib_odp_remove_one(dev);
 	destroy_dev_resources(&dev->devr);
 	free_comp_eqs(dev);
 	ib_dealloc_device(&dev->ib_dev);
@@ -1450,15 +1458,30 @@ static struct mlx5_interface mlx5_ib_interface = {
 
 static int __init mlx5_ib_init(void)
 {
+	int err;
+
 	if (deprecated_prof_sel != 2)
 		pr_warn("prof_sel is deprecated for mlx5_ib, set it for mlx5_core\n");
 
-	return mlx5_register_interface(&mlx5_ib_interface);
+	err = mlx5_ib_odp_init();
+	if (err)
+		return err;
+
+	err = mlx5_register_interface(&mlx5_ib_interface);
+	if (err)
+		goto clean_odp;
+
+	return err;
+
+clean_odp:
+	mlx5_ib_odp_cleanup();
+	return err;
 }
 
 static void __exit mlx5_ib_cleanup(void)
 {
 	mlx5_unregister_interface(&mlx5_ib_interface);
+	mlx5_ib_odp_cleanup();
 }
 
 module_init(mlx5_ib_init);

drivers/infiniband/hw/mlx5/mlx5_ib.h

@@ -149,6 +149,29 @@ enum {
 	MLX5_QP_EMPTY
 };
 
+/*
+ * Connect-IB can trigger up to four concurrent pagefaults
+ * per-QP.
+ */
+enum mlx5_ib_pagefault_context {
+	MLX5_IB_PAGEFAULT_RESPONDER_READ,
+	MLX5_IB_PAGEFAULT_REQUESTOR_READ,
+	MLX5_IB_PAGEFAULT_RESPONDER_WRITE,
+	MLX5_IB_PAGEFAULT_REQUESTOR_WRITE,
+	MLX5_IB_PAGEFAULT_CONTEXTS
+};
+
+static inline enum mlx5_ib_pagefault_context
+	mlx5_ib_get_pagefault_context(struct mlx5_pagefault *pagefault)
+{
+	return pagefault->flags & (MLX5_PFAULT_REQUESTOR | MLX5_PFAULT_WRITE);
+}
+
+struct mlx5_ib_pfault {
+	struct work_struct	work;
+	struct mlx5_pagefault	mpfault;
+};
+
 struct mlx5_ib_qp {
 	struct ib_qp		ibqp;
 	struct mlx5_core_qp	mqp;
@@ -194,6 +217,21 @@ struct mlx5_ib_qp {
 
 	/* Store signature errors */
 	bool			signature_en;
+
+#ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
+	/*
+	 * A flag that is true for QP's that are in a state that doesn't
+	 * allow page faults, and shouldn't schedule any more faults.
+	 */
+	int                     disable_page_faults;
+	/*
+	 * The disable_page_faults_lock protects a QP's disable_page_faults
+	 * field, allowing for a thread to atomically check whether the QP
+	 * allows page faults, and if so schedule a page fault.
+	 */
+	spinlock_t              disable_page_faults_lock;
+	struct mlx5_ib_pfault	pagefaults[MLX5_IB_PAGEFAULT_CONTEXTS];
+#endif
 };
 
 struct mlx5_ib_cq_buf {
@@ -392,13 +430,17 @@ struct mlx5_ib_dev {
 	struct umr_common		umrc;
 	/* sync used page count stats
 	 */
-	spinlock_t			mr_lock;
 	struct mlx5_ib_resources	devr;
 	struct mlx5_mr_cache		cache;
 	struct timer_list		delay_timer;
 	int				fill_delay;
 #ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
 	struct ib_odp_caps	odp_caps;
+	/*
+	 * Sleepable RCU that prevents destruction of MRs while they are still
+	 * being used by a page fault handler.
+	 */
+	struct srcu_struct      mr_srcu;
 #endif
 };
 
@@ -575,12 +617,33 @@ int mlx5_ib_check_mr_status(struct ib_mr *ibmr, u32 check_mask,
 			    struct ib_mr_status *mr_status);
 
 #ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
+extern struct workqueue_struct *mlx5_ib_page_fault_wq;
+
 int mlx5_ib_internal_query_odp_caps(struct mlx5_ib_dev *dev);
-#else
+void mlx5_ib_mr_pfault_handler(struct mlx5_ib_qp *qp,
+			       struct mlx5_ib_pfault *pfault);
+void mlx5_ib_odp_create_qp(struct mlx5_ib_qp *qp);
+int mlx5_ib_odp_init_one(struct mlx5_ib_dev *ibdev);
+void mlx5_ib_odp_remove_one(struct mlx5_ib_dev *ibdev);
+int __init mlx5_ib_odp_init(void);
+void mlx5_ib_odp_cleanup(void);
+void mlx5_ib_qp_disable_pagefaults(struct mlx5_ib_qp *qp);
+void mlx5_ib_qp_enable_pagefaults(struct mlx5_ib_qp *qp);
+
+#else /* CONFIG_INFINIBAND_ON_DEMAND_PAGING */
 static inline int mlx5_ib_internal_query_odp_caps(struct mlx5_ib_dev *dev)
 {
 	return 0;
 }
+
+static inline void mlx5_ib_odp_create_qp(struct mlx5_ib_qp *qp)		{}
+static inline int mlx5_ib_odp_init_one(struct mlx5_ib_dev *ibdev) { return 0; }
+static inline void mlx5_ib_odp_remove_one(struct mlx5_ib_dev *ibdev)	{}
+static inline int mlx5_ib_odp_init(void) { return 0; }
+static inline void mlx5_ib_odp_cleanup(void)				{}
+static inline void mlx5_ib_qp_disable_pagefaults(struct mlx5_ib_qp *qp) {}
+static inline void mlx5_ib_qp_enable_pagefaults(struct mlx5_ib_qp *qp)  {}
+
 #endif /* CONFIG_INFINIBAND_ON_DEMAND_PAGING */
 
 static inline void init_query_mad(struct ib_smp *mad)

drivers/infiniband/hw/mlx5/mr.c

@@ -52,6 +52,8 @@ static __be64 mlx5_ib_update_mtt_emergency_buffer[
 static DEFINE_MUTEX(mlx5_ib_update_mtt_emergency_buffer_mutex);
 #endif
 
+static int clean_mr(struct mlx5_ib_mr *mr);
+
 static int order2idx(struct mlx5_ib_dev *dev, int order)
 {
 	struct mlx5_mr_cache *cache = &dev->cache;
@@ -1049,6 +1051,10 @@ struct ib_mr *mlx5_ib_reg_user_mr(struct ib_pd *pd, u64 start, u64 length,
 			mlx5_ib_dbg(dev, "cache empty for order %d", order);
 			mr = NULL;
 		}
+	} else if (access_flags & IB_ACCESS_ON_DEMAND) {
+		err = -EINVAL;
+		pr_err("Got MR registration for ODP MR > 512MB, not supported for Connect-IB");
+		goto error;
 	}
 
 	if (!mr)
@@ -1064,9 +1070,7 @@ struct ib_mr *mlx5_ib_reg_user_mr(struct ib_pd *pd, u64 start, u64 length,
 
 	mr->umem = umem;
 	mr->npages = npages;
-	spin_lock(&dev->mr_lock);
-	dev->mdev->priv.reg_pages += npages;
-	spin_unlock(&dev->mr_lock);
+	atomic_add(npages, &dev->mdev->priv.reg_pages);
 	mr->ibmr.lkey = mr->mmr.key;
 	mr->ibmr.rkey = mr->mmr.key;
 
@@ -1110,12 +1114,9 @@ static int unreg_umr(struct mlx5_ib_dev *dev, struct mlx5_ib_mr *mr)
 	return err;
 }
 
-int mlx5_ib_dereg_mr(struct ib_mr *ibmr)
+static int clean_mr(struct mlx5_ib_mr *mr)
 {
-	struct mlx5_ib_dev *dev = to_mdev(ibmr->device);
-	struct mlx5_ib_mr *mr = to_mmr(ibmr);
-	struct ib_umem *umem = mr->umem;
-	int npages = mr->npages;
+	struct mlx5_ib_dev *dev = to_mdev(mr->ibmr.device);
 	int umred = mr->umred;
 	int err;
 
@@ -1135,16 +1136,32 @@ int mlx5_ib_dereg_mr(struct ib_mr *ibmr)
 		free_cached_mr(dev, mr);
 	}
 
+	if (!umred)
+		kfree(mr);
+
+	return 0;
+}
+
+int mlx5_ib_dereg_mr(struct ib_mr *ibmr)
+{
+	struct mlx5_ib_dev *dev = to_mdev(ibmr->device);
+	struct mlx5_ib_mr *mr = to_mmr(ibmr);
+	int npages = mr->npages;
+	struct ib_umem *umem = mr->umem;
+
+#ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
+	if (umem)
+		/* Wait for all running page-fault handlers to finish. */
+		synchronize_srcu(&dev->mr_srcu);
+#endif
+
+	clean_mr(mr);
+
 	if (umem) {
 		ib_umem_release(umem);
-		spin_lock(&dev->mr_lock);
-		dev->mdev->priv.reg_pages -= npages;
-		spin_unlock(&dev->mr_lock);
+		atomic_sub(npages, &dev->mdev->priv.reg_pages);
 	}
 
-	if (!umred)
-		kfree(mr);
-
 	return 0;
 }