{net,IB}/mlx5: Move Page fault EQ and ODP logic to RDMA

Use the new generic EQ API to move all ODP RDMA data structures and logic
form mlx5 core driver into mlx5_ib driver.

Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
Reviewed-by: Leon Romanovsky <leonro@mellanox.com>
Reviewed-by: Tariq Toukan <tariqt@mellanox.com>
Acked-by: Jason Gunthorpe <jgg@mellanox.com>
Signed-off-by: Leon Romanovsky <leonro@mellanox.com>

3 files changed, 286 insertions, 20 deletions

diff --git a/drivers/infiniband/hw/mlx5/main.c b/drivers/infiniband/hw/mlx5/main.c
index 6fbc0cba1bac..fcf4a0328a90 100644
--- a/drivers/infiniband/hw/mlx5/main.c
+++ b/drivers/infiniband/hw/mlx5/main.c
@@ -6040,6 +6040,11 @@ static int mlx5_ib_stage_odp_init(struct mlx5_ib_dev *dev)
 	return mlx5_ib_odp_init_one(dev);
 }
 
+void mlx5_ib_stage_odp_cleanup(struct mlx5_ib_dev *dev)
+{
+	mlx5_ib_odp_cleanup_one(dev);
+}
+
 int mlx5_ib_stage_counters_init(struct mlx5_ib_dev *dev)
 {
 	if (MLX5_CAP_GEN(dev->mdev, max_qp_cnt)) {
@@ -6225,7 +6230,7 @@ static const struct mlx5_ib_profile pf_profile = {
 		     mlx5_ib_stage_dev_res_cleanup),
 	STAGE_CREATE(MLX5_IB_STAGE_ODP,
 		     mlx5_ib_stage_odp_init,
-		     NULL),
+		     mlx5_ib_stage_odp_cleanup),
 	STAGE_CREATE(MLX5_IB_STAGE_COUNTERS,
 		     mlx5_ib_stage_counters_init,
 		     mlx5_ib_stage_counters_cleanup),
@@ -6395,9 +6400,6 @@ static struct mlx5_interface mlx5_ib_interface = {
 	.add            = mlx5_ib_add,
 	.remove         = mlx5_ib_remove,
 	.event          = mlx5_ib_event,
-#ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
-	.pfault		= mlx5_ib_pfault,
-#endif
 	.protocol	= MLX5_INTERFACE_PROTOCOL_IB,
 };
 
diff --git a/drivers/infiniband/hw/mlx5/mlx5_ib.h b/drivers/infiniband/hw/mlx5/mlx5_ib.h
index b651a7a6fde9..27999fd32356 100644
--- a/drivers/infiniband/hw/mlx5/mlx5_ib.h
+++ b/drivers/infiniband/hw/mlx5/mlx5_ib.h
@@ -880,6 +880,15 @@ struct mlx5_ib_lb_state {
 	bool			enabled;
 };
 
+struct mlx5_ib_pf_eq {
+	struct mlx5_ib_dev *dev;
+	struct mlx5_eq *core;
+	struct work_struct work;
+	spinlock_t lock; /* Pagefaults spinlock */
+	struct workqueue_struct *wq;
+	mempool_t *pool;
+};
+
 struct mlx5_ib_dev {
 	struct ib_device		ib_dev;
 	const struct uverbs_object_tree_def *driver_trees[7];
@@ -902,6 +911,8 @@ struct mlx5_ib_dev {
 #ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
 	struct ib_odp_caps	odp_caps;
 	u64			odp_max_size;
+	struct mlx5_ib_pf_eq	odp_pf_eq;
+
 	/*
 	 * Sleepable RCU that prevents destruction of MRs while they are still
 	 * being used by a page fault handler.
@@ -1158,9 +1169,8 @@ struct ib_mr *mlx5_ib_reg_dm_mr(struct ib_pd *pd, struct ib_dm *dm,
 
 #ifdef CONFIG_INFINIBAND_ON_DEMAND_PAGING
 void mlx5_ib_internal_fill_odp_caps(struct mlx5_ib_dev *dev);
-void mlx5_ib_pfault(struct mlx5_core_dev *mdev, void *context,
-		    struct mlx5_pagefault *pfault);
 int mlx5_ib_odp_init_one(struct mlx5_ib_dev *ibdev);
+void mlx5_ib_odp_cleanup_one(struct mlx5_ib_dev *ibdev);
 int __init mlx5_ib_odp_init(void);
 void mlx5_ib_odp_cleanup(void);
 void mlx5_ib_invalidate_range(struct ib_umem_odp *umem_odp, unsigned long start,
@@ -1175,6 +1185,7 @@ static inline void mlx5_ib_internal_fill_odp_caps(struct mlx5_ib_dev *dev)
 }
 
 static inline int mlx5_ib_odp_init_one(struct mlx5_ib_dev *ibdev) { return 0; }
+static inline void mlx5_ib_odp_cleanup_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_odp_init_mr_cache_entry(struct mlx5_cache_ent *ent) {}
diff --git a/drivers/infiniband/hw/mlx5/odp.c b/drivers/infiniband/hw/mlx5/odp.c
index 7d784b40e017..416d141322a0 100644
--- a/drivers/infiniband/hw/mlx5/odp.c
+++ b/drivers/infiniband/hw/mlx5/odp.c
@@ -37,6 +37,46 @@
 #include "mlx5_ib.h"
 #include "cmd.h"
 
+#include <linux/mlx5/eq.h>
+
+/* Contains the details of a pagefault. */
+struct mlx5_pagefault {
+	u32			bytes_committed;
+	u32			token;
+	u8			event_subtype;
+	u8			type;
+	union {
+		/* Initiator or send message responder pagefault details. */
+		struct {
+			/* Received packet size, only valid for responders. */
+			u32	packet_size;
+			/*
+			 * Number of resource holding WQE, depends on type.
+			 */
+			u32	wq_num;
+			/*
+			 * WQE index. Refers to either the send queue or
+			 * receive queue, according to event_subtype.
+			 */
+			u16	wqe_index;
+		} wqe;
+		/* RDMA responder pagefault details */
+		struct {
+			u32	r_key;
+			/*
+			 * Received packet size, minimal size page fault
+			 * resolution required for forward progress.
+			 */
+			u32	packet_size;
+			u32	rdma_op_len;
+			u64	rdma_va;
+		} rdma;
+	};
+
+	struct mlx5_ib_pf_eq	*eq;
+	struct work_struct	work;
+};
+
 #define MAX_PREFETCH_LEN (4*1024*1024U)
 
 /* Timeout in ms to wait for an active mmu notifier to complete when handling
@@ -304,14 +344,20 @@ static void mlx5_ib_page_fault_resume(struct mlx5_ib_dev *dev,
 {
 	int wq_num = pfault->event_subtype == MLX5_PFAULT_SUBTYPE_WQE ?
 		     pfault->wqe.wq_num : pfault->token;
-	int ret = mlx5_core_page_fault_resume(dev->mdev,
-					      pfault->token,
-					      wq_num,
-					      pfault->type,
-					      error);
-	if (ret)
-		mlx5_ib_err(dev, "Failed to resolve the page fault on WQ 0x%x\n",
-			    wq_num);
+	u32 out[MLX5_ST_SZ_DW(page_fault_resume_out)] = { };
+	u32 in[MLX5_ST_SZ_DW(page_fault_resume_in)]   = { };
+	int err;
+
+	MLX5_SET(page_fault_resume_in, in, opcode, MLX5_CMD_OP_PAGE_FAULT_RESUME);
+	MLX5_SET(page_fault_resume_in, in, page_fault_type, pfault->type);
+	MLX5_SET(page_fault_resume_in, in, token, pfault->token);
+	MLX5_SET(page_fault_resume_in, in, wq_number, wq_num);
+	MLX5_SET(page_fault_resume_in, in, error, !!error);
+
+	err = mlx5_cmd_exec(dev->mdev, in, sizeof(in), out, sizeof(out));
+	if (err)
+		mlx5_ib_err(dev, "Failed to resolve the page fault on WQ 0x%x err %d\n",
+			    wq_num, err);
 }
 
 static struct mlx5_ib_mr *implicit_mr_alloc(struct ib_pd *pd,
@@ -1196,10 +1242,8 @@ static void mlx5_ib_mr_rdma_pfault_handler(struct mlx5_ib_dev *dev,
 	}
 }
 
-void mlx5_ib_pfault(struct mlx5_core_dev *mdev, void *context,
-		    struct mlx5_pagefault *pfault)
+static void mlx5_ib_pfault(struct mlx5_ib_dev *dev, struct mlx5_pagefault *pfault)
 {
-	struct mlx5_ib_dev *dev = context;
 	u8 event_subtype = pfault->event_subtype;
 
 	switch (event_subtype) {
@@ -1216,6 +1260,203 @@ void mlx5_ib_pfault(struct mlx5_core_dev *mdev, void *context,
 	}
 }
 
+static void mlx5_ib_eqe_pf_action(struct work_struct *work)
+{
+	struct mlx5_pagefault *pfault = container_of(work,
+						     struct mlx5_pagefault,
+						     work);
+	struct mlx5_ib_pf_eq *eq = pfault->eq;
+
+	mlx5_ib_pfault(eq->dev, pfault);
+	mempool_free(pfault, eq->pool);
+}
+
+static void mlx5_ib_eq_pf_process(struct mlx5_ib_pf_eq *eq)
+{
+	struct mlx5_eqe_page_fault *pf_eqe;
+	struct mlx5_pagefault *pfault;
+	struct mlx5_eqe *eqe;
+	int cc = 0;
+
+	while ((eqe = mlx5_eq_get_eqe(eq->core, cc))) {
+		pfault = mempool_alloc(eq->pool, GFP_ATOMIC);
+		if (!pfault) {
+			schedule_work(&eq->work);
+			break;
+		}
+
+		pf_eqe = &eqe->data.page_fault;
+		pfault->event_subtype = eqe->sub_type;
+		pfault->bytes_committed = be32_to_cpu(pf_eqe->bytes_committed);
+
+		mlx5_ib_dbg(eq->dev,
+			    "PAGE_FAULT: subtype: 0x%02x, bytes_committed: 0x%06x\n",
+			    eqe->sub_type, pfault->bytes_committed);
+
+		switch (eqe->sub_type) {
+		case MLX5_PFAULT_SUBTYPE_RDMA:
+			/* RDMA based event */
+			pfault->type =
+				be32_to_cpu(pf_eqe->rdma.pftype_token) >> 24;
+			pfault->token =
+				be32_to_cpu(pf_eqe->rdma.pftype_token) &
+				MLX5_24BIT_MASK;
+			pfault->rdma.r_key =
+				be32_to_cpu(pf_eqe->rdma.r_key);
+			pfault->rdma.packet_size =
+				be16_to_cpu(pf_eqe->rdma.packet_length);
+			pfault->rdma.rdma_op_len =
+				be32_to_cpu(pf_eqe->rdma.rdma_op_len);
+			pfault->rdma.rdma_va =
+				be64_to_cpu(pf_eqe->rdma.rdma_va);
+			mlx5_ib_dbg(eq->dev,
+				    "PAGE_FAULT: type:0x%x, token: 0x%06x, r_key: 0x%08x\n",
+				    pfault->type, pfault->token,
+				    pfault->rdma.r_key);
+			mlx5_ib_dbg(eq->dev,
+				    "PAGE_FAULT: rdma_op_len: 0x%08x, rdma_va: 0x%016llx\n",
+				    pfault->rdma.rdma_op_len,
+				    pfault->rdma.rdma_va);
+			break;
+
+		case MLX5_PFAULT_SUBTYPE_WQE:
+			/* WQE based event */
+			pfault->type =
+				(be32_to_cpu(pf_eqe->wqe.pftype_wq) >> 24) & 0x7;
+			pfault->token =
+				be32_to_cpu(pf_eqe->wqe.token);
+			pfault->wqe.wq_num =
+				be32_to_cpu(pf_eqe->wqe.pftype_wq) &
+				MLX5_24BIT_MASK;
+			pfault->wqe.wqe_index =
+				be16_to_cpu(pf_eqe->wqe.wqe_index);
+			pfault->wqe.packet_size =
+				be16_to_cpu(pf_eqe->wqe.packet_length);
+			mlx5_ib_dbg(eq->dev,
+				    "PAGE_FAULT: type:0x%x, token: 0x%06x, wq_num: 0x%06x, wqe_index: 0x%04x\n",
+				    pfault->type, pfault->token,
+				    pfault->wqe.wq_num,
+				    pfault->wqe.wqe_index);
+			break;
+
+		default:
+			mlx5_ib_warn(eq->dev,
+				     "Unsupported page fault event sub-type: 0x%02hhx\n",
+				     eqe->sub_type);
+			/* Unsupported page faults should still be
+			 * resolved by the page fault handler
+			 */
+		}
+
+		pfault->eq = eq;
+		INIT_WORK(&pfault->work, mlx5_ib_eqe_pf_action);
+		queue_work(eq->wq, &pfault->work);
+
+		cc = mlx5_eq_update_cc(eq->core, ++cc);
+	}
+
+	mlx5_eq_update_ci(eq->core, cc, 1);
+}
+
+static irqreturn_t mlx5_ib_eq_pf_int(int irq, void *eq_ptr)
+{
+	struct mlx5_ib_pf_eq *eq = eq_ptr;
+	unsigned long flags;
+
+	if (spin_trylock_irqsave(&eq->lock, flags)) {
+		mlx5_ib_eq_pf_process(eq);
+		spin_unlock_irqrestore(&eq->lock, flags);
+	} else {
+		schedule_work(&eq->work);
+	}
+
+	return IRQ_HANDLED;
+}
+
+/* mempool_refill() was proposed but unfortunately wasn't accepted
+ * http://lkml.iu.edu/hypermail/linux/kernel/1512.1/05073.html
+ * Cheap workaround.
+ */
+static void mempool_refill(mempool_t *pool)
+{
+	while (pool->curr_nr < pool->min_nr)
+		mempool_free(mempool_alloc(pool, GFP_KERNEL), pool);
+}
+
+static void mlx5_ib_eq_pf_action(struct work_struct *work)
+{
+	struct mlx5_ib_pf_eq *eq =
+		container_of(work, struct mlx5_ib_pf_eq, work);
+
+	mempool_refill(eq->pool);
+
+	spin_lock_irq(&eq->lock);
+	mlx5_ib_eq_pf_process(eq);
+	spin_unlock_irq(&eq->lock);
+}
+
+enum {
+	MLX5_IB_NUM_PF_EQE	= 0x1000,
+	MLX5_IB_NUM_PF_DRAIN	= 64,
+};
+
+static int
+mlx5_ib_create_pf_eq(struct mlx5_ib_dev *dev, struct mlx5_ib_pf_eq *eq)
+{
+	struct mlx5_eq_param param = {};
+	int err;
+
+	INIT_WORK(&eq->work, mlx5_ib_eq_pf_action);
+	spin_lock_init(&eq->lock);
+	eq->dev = dev;
+
+	eq->pool = mempool_create_kmalloc_pool(MLX5_IB_NUM_PF_DRAIN,
+					       sizeof(struct mlx5_pagefault));
+	if (!eq->pool)
+		return -ENOMEM;
+
+	eq->wq = alloc_workqueue("mlx5_ib_page_fault",
+				 WQ_HIGHPRI | WQ_UNBOUND | WQ_MEM_RECLAIM,
+				 MLX5_NUM_CMD_EQE);
+	if (!eq->wq) {
+		err = -ENOMEM;
+		goto err_mempool;
+	}
+
+	param = (struct mlx5_eq_param) {
+		.index = MLX5_EQ_PFAULT_IDX,
+		.mask = 1 << MLX5_EVENT_TYPE_PAGE_FAULT,
+		.nent = MLX5_IB_NUM_PF_EQE,
+		.context = eq,
+		.handler = mlx5_ib_eq_pf_int
+	};
+	eq->core = mlx5_eq_create_generic(dev->mdev, "mlx5_ib_page_fault_eq", &param);
+	if (IS_ERR(eq->core)) {
+		err = PTR_ERR(eq->core);
+		goto err_wq;
+	}
+
+	return 0;
+err_wq:
+	destroy_workqueue(eq->wq);
+err_mempool:
+	mempool_destroy(eq->pool);
+	return err;
+}
+
+static int
+mlx5_ib_destroy_pf_eq(struct mlx5_ib_dev *dev, struct mlx5_ib_pf_eq *eq)
+{
+	int err;
+
+	err = mlx5_eq_destroy_generic(dev->mdev, eq->core);
+	cancel_work_sync(&eq->work);
+	destroy_workqueue(eq->wq);
+	mempool_destroy(eq->pool);
+
+	return err;
+}
+
 void mlx5_odp_init_mr_cache_entry(struct mlx5_cache_ent *ent)
 {
 	if (!(ent->dev->odp_caps.general_caps & IB_ODP_SUPPORT_IMPLICIT))
@@ -1244,7 +1485,7 @@ void mlx5_odp_init_mr_cache_entry(struct mlx5_cache_ent *ent)
 
 int mlx5_ib_odp_init_one(struct mlx5_ib_dev *dev)
 {
-	int ret;
+	int ret = 0;
 
 	if (dev->odp_caps.general_caps & IB_ODP_SUPPORT_IMPLICIT) {
 		ret = mlx5_cmd_null_mkey(dev->mdev, &dev->null_mkey);
@@ -1254,7 +1495,20 @@ int mlx5_ib_odp_init_one(struct mlx5_ib_dev *dev)
 		}
 	}
 
-	return 0;
+	if (!MLX5_CAP_GEN(dev->mdev, pg))
+		return ret;
+
+	ret = mlx5_ib_create_pf_eq(dev, &dev->odp_pf_eq);
+
+	return ret;
+}
+
+void mlx5_ib_odp_cleanup_one(struct mlx5_ib_dev *dev)
+{
+	if (!MLX5_CAP_GEN(dev->mdev, pg))
+		return;
+
+	mlx5_ib_destroy_pf_eq(dev, &dev->odp_pf_eq);
 }
 
 int mlx5_ib_odp_init(void)
@@ -1264,4 +1518,3 @@ int mlx5_ib_odp_init(void)
 
 	return 0;
 }
-