Merge branch 'mlx5-vfio' into mlx5-next

Merge net/mlx5 dependencies for device DMA logging.

Signed-off-by: Leon Romanovsky <leonro@nvidia.com>

7 files changed, 555 insertions, 325 deletions

diff --git a/drivers/infiniband/hw/mlx5/cq.c b/drivers/infiniband/hw/mlx5/cq.c
index 08371a80fdc2..be189e0525de 100644
--- a/drivers/infiniband/hw/mlx5/cq.c
+++ b/drivers/infiniband/hw/mlx5/cq.c
@@ -523,6 +523,10 @@ repoll:
 			    "Requestor" : "Responder", cq->mcq.cqn);
 		mlx5_ib_dbg(dev, "syndrome 0x%x, vendor syndrome 0x%x\n",
 			    err_cqe->syndrome, err_cqe->vendor_err_synd);
+		if (wc->status != IB_WC_WR_FLUSH_ERR &&
+		    (*cur_qp)->type == MLX5_IB_QPT_REG_UMR)
+			dev->umrc.state = MLX5_UMR_STATE_RECOVER;
+
 		if (opcode == MLX5_CQE_REQ_ERR) {
 			wq = &(*cur_qp)->sq;
 			wqe_ctr = be16_to_cpu(cqe64->wqe_counter);
diff --git a/drivers/infiniband/hw/mlx5/fs.c b/drivers/infiniband/hw/mlx5/fs.c
index 39ffb363ba0c..490ec308e309 100644
--- a/drivers/infiniband/hw/mlx5/fs.c
+++ b/drivers/infiniband/hw/mlx5/fs.c
@@ -679,7 +679,15 @@ enum flow_table_type {
 #define MLX5_FS_MAX_TYPES	 6
 #define MLX5_FS_MAX_ENTRIES	 BIT(16)
 
-static struct mlx5_ib_flow_prio *_get_prio(struct mlx5_flow_namespace *ns,
+static bool mlx5_ib_shared_ft_allowed(struct ib_device *device)
+{
+	struct mlx5_ib_dev *dev = to_mdev(device);
+
+	return MLX5_CAP_GEN(dev->mdev, shared_object_to_user_object_allowed);
+}
+
+static struct mlx5_ib_flow_prio *_get_prio(struct mlx5_ib_dev *dev,
+					   struct mlx5_flow_namespace *ns,
 					   struct mlx5_ib_flow_prio *prio,
 					   int priority,
 					   int num_entries, int num_groups,
@@ -688,6 +696,8 @@ static struct mlx5_ib_flow_prio *_get_prio(struct mlx5_flow_namespace *ns,
 	struct mlx5_flow_table_attr ft_attr = {};
 	struct mlx5_flow_table *ft;
 
+	if (mlx5_ib_shared_ft_allowed(&dev->ib_dev))
+		ft_attr.uid = MLX5_SHARED_RESOURCE_UID;
 	ft_attr.prio = priority;
 	ft_attr.max_fte = num_entries;
 	ft_attr.flags = flags;
@@ -784,8 +794,8 @@ static struct mlx5_ib_flow_prio *get_flow_table(struct mlx5_ib_dev *dev,
 
 	ft = prio->flow_table;
 	if (!ft)
-		return _get_prio(ns, prio, priority, max_table_size, num_groups,
-				 flags);
+		return _get_prio(dev, ns, prio, priority, max_table_size,
+				 num_groups, flags);
 
 	return prio;
 }
@@ -927,7 +937,7 @@ int mlx5_ib_fs_add_op_fc(struct mlx5_ib_dev *dev, u32 port_num,
 
 	prio = &dev->flow_db->opfcs[type];
 	if (!prio->flow_table) {
-		prio = _get_prio(ns, prio, priority,
+		prio = _get_prio(dev, ns, prio, priority,
 				 dev->num_ports * MAX_OPFC_RULES, 1, 0);
 		if (IS_ERR(prio)) {
 			err = PTR_ERR(prio);
@@ -1407,8 +1417,8 @@ free_ucmd:
 }
 
 static struct mlx5_ib_flow_prio *
-_get_flow_table(struct mlx5_ib_dev *dev,
-		struct mlx5_ib_flow_matcher *fs_matcher,
+_get_flow_table(struct mlx5_ib_dev *dev, u16 user_priority,
+		enum mlx5_flow_namespace_type ns_type,
 		bool mcast)
 {
 	struct mlx5_flow_namespace *ns = NULL;
@@ -1421,11 +1431,11 @@ _get_flow_table(struct mlx5_ib_dev *dev,
 	if (mcast)
 		priority = MLX5_IB_FLOW_MCAST_PRIO;
 	else
-		priority = ib_prio_to_core_prio(fs_matcher->priority, false);
+		priority = ib_prio_to_core_prio(user_priority, false);
 
 	esw_encap = mlx5_eswitch_get_encap_mode(dev->mdev) !=
 		DEVLINK_ESWITCH_ENCAP_MODE_NONE;
-	switch (fs_matcher->ns_type) {
+	switch (ns_type) {
 	case MLX5_FLOW_NAMESPACE_BYPASS:
 		max_table_size = BIT(
 			MLX5_CAP_FLOWTABLE_NIC_RX(dev->mdev, log_max_ft_size));
@@ -1452,17 +1462,17 @@ _get_flow_table(struct mlx5_ib_dev *dev,
 					       reformat_l3_tunnel_to_l2) &&
 		    esw_encap)
 			flags |= MLX5_FLOW_TABLE_TUNNEL_EN_REFORMAT;
-		priority = fs_matcher->priority;
+		priority = user_priority;
 		break;
 	case MLX5_FLOW_NAMESPACE_RDMA_RX:
 		max_table_size = BIT(
 			MLX5_CAP_FLOWTABLE_RDMA_RX(dev->mdev, log_max_ft_size));
-		priority = fs_matcher->priority;
+		priority = user_priority;
 		break;
 	case MLX5_FLOW_NAMESPACE_RDMA_TX:
 		max_table_size = BIT(
 			MLX5_CAP_FLOWTABLE_RDMA_TX(dev->mdev, log_max_ft_size));
-		priority = fs_matcher->priority;
+		priority = user_priority;
 		break;
 	default:
 		break;
@@ -1470,11 +1480,11 @@ _get_flow_table(struct mlx5_ib_dev *dev,
 
 	max_table_size = min_t(int, max_table_size, MLX5_FS_MAX_ENTRIES);
 
-	ns = mlx5_get_flow_namespace(dev->mdev, fs_matcher->ns_type);
+	ns = mlx5_get_flow_namespace(dev->mdev, ns_type);
 	if (!ns)
 		return ERR_PTR(-EOPNOTSUPP);
 
-	switch (fs_matcher->ns_type) {
+	switch (ns_type) {
 	case MLX5_FLOW_NAMESPACE_BYPASS:
 		prio = &dev->flow_db->prios[priority];
 		break;
@@ -1499,7 +1509,7 @@ _get_flow_table(struct mlx5_ib_dev *dev,
 	if (prio->flow_table)
 		return prio;
 
-	return _get_prio(ns, prio, priority, max_table_size,
+	return _get_prio(dev, ns, prio, priority, max_table_size,
 			 MLX5_FS_MAX_TYPES, flags);
 }
 
@@ -1618,7 +1628,8 @@ static struct mlx5_ib_flow_handler *raw_fs_rule_add(
 	mcast = raw_fs_is_multicast(fs_matcher, cmd_in);
 	mutex_lock(&dev->flow_db->lock);
 
-	ft_prio = _get_flow_table(dev, fs_matcher, mcast);
+	ft_prio = _get_flow_table(dev, fs_matcher->priority,
+				  fs_matcher->ns_type, mcast);
 	if (IS_ERR(ft_prio)) {
 		err = PTR_ERR(ft_prio);
 		goto unlock;
@@ -2015,6 +2026,23 @@ static int flow_matcher_cleanup(struct ib_uobject *uobject,
 	return 0;
 }
 
+static int steering_anchor_cleanup(struct ib_uobject *uobject,
+				   enum rdma_remove_reason why,
+				   struct uverbs_attr_bundle *attrs)
+{
+	struct mlx5_ib_steering_anchor *obj = uobject->object;
+
+	if (atomic_read(&obj->usecnt))
+		return -EBUSY;
+
+	mutex_lock(&obj->dev->flow_db->lock);
+	put_flow_table(obj->dev, obj->ft_prio, true);
+	mutex_unlock(&obj->dev->flow_db->lock);
+
+	kfree(obj);
+	return 0;
+}
+
 static int mlx5_ib_matcher_ns(struct uverbs_attr_bundle *attrs,
 			      struct mlx5_ib_flow_matcher *obj)
 {
@@ -2050,12 +2078,10 @@ static int mlx5_ib_matcher_ns(struct uverbs_attr_bundle *attrs,
 		if (err)
 			return err;
 
-		if (flags) {
-			mlx5_ib_ft_type_to_namespace(
+		if (flags)
+			return mlx5_ib_ft_type_to_namespace(
 				MLX5_IB_UAPI_FLOW_TABLE_TYPE_NIC_TX,
 				&obj->ns_type);
-			return 0;
-		}
 	}
 
 	obj->ns_type = MLX5_FLOW_NAMESPACE_BYPASS;
@@ -2121,6 +2147,75 @@ end:
 	return err;
 }
 
+static int UVERBS_HANDLER(MLX5_IB_METHOD_STEERING_ANCHOR_CREATE)(
+	struct uverbs_attr_bundle *attrs)
+{
+	struct ib_uobject *uobj = uverbs_attr_get_uobject(
+		attrs, MLX5_IB_ATTR_STEERING_ANCHOR_CREATE_HANDLE);
+	struct mlx5_ib_dev *dev = mlx5_udata_to_mdev(&attrs->driver_udata);
+	enum mlx5_ib_uapi_flow_table_type ib_uapi_ft_type;
+	enum mlx5_flow_namespace_type ns_type;
+	struct mlx5_ib_steering_anchor *obj;
+	struct mlx5_ib_flow_prio *ft_prio;
+	u16 priority;
+	u32 ft_id;
+	int err;
+
+	if (!capable(CAP_NET_RAW))
+		return -EPERM;
+
+	err = uverbs_get_const(&ib_uapi_ft_type, attrs,
+			       MLX5_IB_ATTR_STEERING_ANCHOR_FT_TYPE);
+	if (err)
+		return err;
+
+	err = mlx5_ib_ft_type_to_namespace(ib_uapi_ft_type, &ns_type);
+	if (err)
+		return err;
+
+	err = uverbs_copy_from(&priority, attrs,
+			       MLX5_IB_ATTR_STEERING_ANCHOR_PRIORITY);
+	if (err)
+		return err;
+
+	obj = kzalloc(sizeof(*obj), GFP_KERNEL);
+	if (!obj)
+		return -ENOMEM;
+
+	mutex_lock(&dev->flow_db->lock);
+	ft_prio = _get_flow_table(dev, priority, ns_type, 0);
+	if (IS_ERR(ft_prio)) {
+		mutex_unlock(&dev->flow_db->lock);
+		err = PTR_ERR(ft_prio);
+		goto free_obj;
+	}
+
+	ft_prio->refcount++;
+	ft_id = mlx5_flow_table_id(ft_prio->flow_table);
+	mutex_unlock(&dev->flow_db->lock);
+
+	err = uverbs_copy_to(attrs, MLX5_IB_ATTR_STEERING_ANCHOR_FT_ID,
+			     &ft_id, sizeof(ft_id));
+	if (err)
+		goto put_flow_table;
+
+	uobj->object = obj;
+	obj->dev = dev;
+	obj->ft_prio = ft_prio;
+	atomic_set(&obj->usecnt, 0);
+
+	return 0;
+
+put_flow_table:
+	mutex_lock(&dev->flow_db->lock);
+	put_flow_table(dev, ft_prio, true);
+	mutex_unlock(&dev->flow_db->lock);
+free_obj:
+	kfree(obj);
+
+	return err;
+}
+
 static struct ib_flow_action *
 mlx5_ib_create_modify_header(struct mlx5_ib_dev *dev,
 			     enum mlx5_ib_uapi_flow_table_type ft_type,
@@ -2477,6 +2572,35 @@ DECLARE_UVERBS_NAMED_OBJECT(MLX5_IB_OBJECT_FLOW_MATCHER,
 			    &UVERBS_METHOD(MLX5_IB_METHOD_FLOW_MATCHER_CREATE),
 			    &UVERBS_METHOD(MLX5_IB_METHOD_FLOW_MATCHER_DESTROY));
 
+DECLARE_UVERBS_NAMED_METHOD(
+	MLX5_IB_METHOD_STEERING_ANCHOR_CREATE,
+	UVERBS_ATTR_IDR(MLX5_IB_ATTR_STEERING_ANCHOR_CREATE_HANDLE,
+			MLX5_IB_OBJECT_STEERING_ANCHOR,
+			UVERBS_ACCESS_NEW,
+			UA_MANDATORY),
+	UVERBS_ATTR_CONST_IN(MLX5_IB_ATTR_STEERING_ANCHOR_FT_TYPE,
+			     enum mlx5_ib_uapi_flow_table_type,
+			     UA_MANDATORY),
+	UVERBS_ATTR_PTR_IN(MLX5_IB_ATTR_STEERING_ANCHOR_PRIORITY,
+			   UVERBS_ATTR_TYPE(u16),
+			   UA_MANDATORY),
+	UVERBS_ATTR_PTR_IN(MLX5_IB_ATTR_STEERING_ANCHOR_FT_ID,
+			   UVERBS_ATTR_TYPE(u32),
+			   UA_MANDATORY));
+
+DECLARE_UVERBS_NAMED_METHOD_DESTROY(
+	MLX5_IB_METHOD_STEERING_ANCHOR_DESTROY,
+	UVERBS_ATTR_IDR(MLX5_IB_ATTR_STEERING_ANCHOR_DESTROY_HANDLE,
+			MLX5_IB_OBJECT_STEERING_ANCHOR,
+			UVERBS_ACCESS_DESTROY,
+			UA_MANDATORY));
+
+DECLARE_UVERBS_NAMED_OBJECT(
+	MLX5_IB_OBJECT_STEERING_ANCHOR,
+	UVERBS_TYPE_ALLOC_IDR(steering_anchor_cleanup),
+	&UVERBS_METHOD(MLX5_IB_METHOD_STEERING_ANCHOR_CREATE),
+	&UVERBS_METHOD(MLX5_IB_METHOD_STEERING_ANCHOR_DESTROY));
+
 const struct uapi_definition mlx5_ib_flow_defs[] = {
 	UAPI_DEF_CHAIN_OBJ_TREE_NAMED(
 		MLX5_IB_OBJECT_FLOW_MATCHER),
@@ -2485,6 +2609,9 @@ const struct uapi_definition mlx5_ib_flow_defs[] = {
 		&mlx5_ib_fs),
 	UAPI_DEF_CHAIN_OBJ_TREE(UVERBS_OBJECT_FLOW_ACTION,
 				&mlx5_ib_flow_actions),
+	UAPI_DEF_CHAIN_OBJ_TREE_NAMED(
+		MLX5_IB_OBJECT_STEERING_ANCHOR,
+		UAPI_DEF_IS_OBJ_SUPPORTED(mlx5_ib_shared_ft_allowed)),
 	{},
 };
 
diff --git a/drivers/infiniband/hw/mlx5/main.c b/drivers/infiniband/hw/mlx5/main.c
index b68fddeac0f1..fc94a1b25485 100644
--- a/drivers/infiniband/hw/mlx5/main.c
+++ b/drivers/infiniband/hw/mlx5/main.c
@@ -2738,26 +2738,24 @@ static int set_has_smi_cap(struct mlx5_ib_dev *dev)
 	int err;
 	int port;
 
-	for (port = 1; port <= ARRAY_SIZE(dev->port_caps); port++) {
-		dev->port_caps[port - 1].has_smi = false;
-		if (MLX5_CAP_GEN(dev->mdev, port_type) ==
-		    MLX5_CAP_PORT_TYPE_IB) {
-			if (MLX5_CAP_GEN(dev->mdev, ib_virt)) {
-				err = mlx5_query_hca_vport_context(dev->mdev, 0,
-								   port, 0,
-								   &vport_ctx);
-				if (err) {
-					mlx5_ib_err(dev, "query_hca_vport_context for port=%d failed %d\n",
-						    port, err);
-					return err;
-				}
-				dev->port_caps[port - 1].has_smi =
-					vport_ctx.has_smi;
-			} else {
-				dev->port_caps[port - 1].has_smi = true;
-			}
+	if (MLX5_CAP_GEN(dev->mdev, port_type) != MLX5_CAP_PORT_TYPE_IB)
+		return 0;
+
+	for (port = 1; port <= dev->num_ports; port++) {
+		if (!MLX5_CAP_GEN(dev->mdev, ib_virt)) {
+			dev->port_caps[port - 1].has_smi = true;
+			continue;
 		}
+		err = mlx5_query_hca_vport_context(dev->mdev, 0, port, 0,
+						   &vport_ctx);
+		if (err) {
+			mlx5_ib_err(dev, "query_hca_vport_context for port=%d failed %d\n",
+				    port, err);
+			return err;
+		}
+		dev->port_caps[port - 1].has_smi = vport_ctx.has_smi;
 	}
+
 	return 0;
 }
 
@@ -4002,7 +4000,7 @@ static void mlx5_ib_stage_pre_ib_reg_umr_cleanup(struct mlx5_ib_dev *dev)
 {
 	int err;
 
-	err = mlx5_mr_cache_cleanup(dev);
+	err = mlx5_mkey_cache_cleanup(dev);
 	if (err)
 		mlx5_ib_warn(dev, "mr cache cleanup failed\n");
 
@@ -4022,7 +4020,7 @@ static int mlx5_ib_stage_post_ib_reg_umr_init(struct mlx5_ib_dev *dev)
 	if (ret)
 		return ret;
 
-	ret = mlx5_mr_cache_init(dev);
+	ret = mlx5_mkey_cache_init(dev);
 	if (ret) {
 		mlx5_ib_warn(dev, "mr cache init failed %d\n", ret);
 		mlx5r_umr_resource_cleanup(dev);
diff --git a/drivers/infiniband/hw/mlx5/mlx5_ib.h b/drivers/infiniband/hw/mlx5/mlx5_ib.h
index 998b67509a53..2e2ad3918385 100644
--- a/drivers/infiniband/hw/mlx5/mlx5_ib.h
+++ b/drivers/infiniband/hw/mlx5/mlx5_ib.h
@@ -259,6 +259,12 @@ struct mlx5_ib_flow_matcher {
 	u8			match_criteria_enable;
 };
 
+struct mlx5_ib_steering_anchor {
+	struct mlx5_ib_flow_prio *ft_prio;
+	struct mlx5_ib_dev *dev;
+	atomic_t usecnt;
+};
+
 struct mlx5_ib_pp {
 	u16 index;
 	struct mlx5_core_dev *mdev;
@@ -613,6 +619,7 @@ struct mlx5_ib_mkey {
 	unsigned int ndescs;
 	struct wait_queue_head wait;
 	refcount_t usecount;
+	struct mlx5_cache_ent *cache_ent;
 };
 
 #define MLX5_IB_MTT_PRESENT (MLX5_IB_MTT_READ | MLX5_IB_MTT_WRITE)
@@ -635,20 +642,9 @@ struct mlx5_ib_mr {
 	struct ib_mr ibmr;
 	struct mlx5_ib_mkey mmkey;
 
-	/* User MR data */
-	struct mlx5_cache_ent *cache_ent;
-	/* Everything after cache_ent is zero'd when MR allocated */
 	struct ib_umem *umem;
 
 	union {
-		/* Used only while the MR is in the cache */
-		struct {
-			u32 out[MLX5_ST_SZ_DW(create_mkey_out)];
-			struct mlx5_async_work cb_work;
-			/* Cache list element */
-			struct list_head list;
-		};
-
 		/* Used only by kernel MRs (umem == NULL) */
 		struct {
 			void *descs;
@@ -688,12 +684,6 @@ struct mlx5_ib_mr {
 	};
 };
 
-/* Zero the fields in the mr that are variant depending on usage */
-static inline void mlx5_clear_mr(struct mlx5_ib_mr *mr)
-{
-	memset_after(mr, 0, cache_ent);
-}
-
 static inline bool is_odp_mr(struct mlx5_ib_mr *mr)
 {
 	return IS_ENABLED(CONFIG_INFINIBAND_ON_DEMAND_PAGING) && mr->umem &&
@@ -717,21 +707,29 @@ struct mlx5_ib_umr_context {
 	struct completion	done;
 };
 
+enum {
+	MLX5_UMR_STATE_ACTIVE,
+	MLX5_UMR_STATE_RECOVER,
+	MLX5_UMR_STATE_ERR,
+};
+
 struct umr_common {
 	struct ib_pd	*pd;
 	struct ib_cq	*cq;
 	struct ib_qp	*qp;
-	/* control access to UMR QP
+	/* Protects from UMR QP overflow
 	 */
 	struct semaphore	sem;
+	/* Protects from using UMR while the UMR is not active
+	 */
+	struct mutex lock;
+	unsigned int state;
 };
 
 struct mlx5_cache_ent {
-	struct list_head	head;
-	/* sync access to the cahce entry
-	 */
-	spinlock_t		lock;
-
+	struct xarray		mkeys;
+	unsigned long		stored;
+	unsigned long		reserved;
 
 	char                    name[4];
 	u32                     order;
@@ -743,18 +741,11 @@ struct mlx5_cache_ent {
 	u8 fill_to_high_water:1;
 
 	/*
-	 * - available_mrs is the length of list head, ie the number of MRs
-	 *   available for immediate allocation.
-	 * - total_mrs is available_mrs plus all in use MRs that could be
-	 *   returned to the cache.
-	 * - limit is the low water mark for available_mrs, 2* limit is the
+	 * - limit is the low water mark for stored mkeys, 2* limit is the
 	 *   upper water mark.
-	 * - pending is the number of MRs currently being created
 	 */
-	u32 total_mrs;
-	u32 available_mrs;
+	u32 in_use;
 	u32 limit;
-	u32 pending;
 
 	/* Statistics */
 	u32                     miss;
@@ -763,9 +754,19 @@ struct mlx5_cache_ent {
 	struct delayed_work	dwork;
 };
 
-struct mlx5_mr_cache {
+struct mlx5r_async_create_mkey {
+	union {
+		u32 in[MLX5_ST_SZ_BYTES(create_mkey_in)];
+		u32 out[MLX5_ST_SZ_DW(create_mkey_out)];
+	};
+	struct mlx5_async_work cb_work;
+	struct mlx5_cache_ent *ent;
+	u32 mkey;
+};
+
+struct mlx5_mkey_cache {
 	struct workqueue_struct *wq;
-	struct mlx5_cache_ent	ent[MAX_MR_CACHE_ENTRIES];
+	struct mlx5_cache_ent	ent[MAX_MKEY_CACHE_ENTRIES];
 	struct dentry		*root;
 	unsigned long		last_add;
 };
@@ -1064,7 +1065,7 @@ struct mlx5_ib_dev {
 	struct mlx5_ib_resources	devr;
 
 	atomic_t			mkey_var;
-	struct mlx5_mr_cache		cache;
+	struct mlx5_mkey_cache		cache;
 	struct timer_list		delay_timer;
 	/* Prevents soft lock on massive reg MRs */
 	struct mutex			slow_path_mutex;
@@ -1309,8 +1310,8 @@ void mlx5_ib_populate_pas(struct ib_umem *umem, size_t page_size, __be64 *pas,
 			  u64 access_flags);
 void mlx5_ib_copy_pas(u64 *old, u64 *new, int step, int num);
 int mlx5_ib_get_cqe_size(struct ib_cq *ibcq);
-int mlx5_mr_cache_init(struct mlx5_ib_dev *dev);
-int mlx5_mr_cache_cleanup(struct mlx5_ib_dev *dev);
+int mlx5_mkey_cache_init(struct mlx5_ib_dev *dev);
+int mlx5_mkey_cache_cleanup(struct mlx5_ib_dev *dev);
 
 struct mlx5_ib_mr *mlx5_mr_cache_alloc(struct mlx5_ib_dev *dev,
 				       struct mlx5_cache_ent *ent,
@@ -1338,7 +1339,7 @@ int mlx5r_odp_create_eq(struct mlx5_ib_dev *dev, struct mlx5_ib_pf_eq *eq);
 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_odp_init_mr_cache_entry(struct mlx5_cache_ent *ent);
+void mlx5_odp_init_mkey_cache_entry(struct mlx5_cache_ent *ent);
 void mlx5_odp_populate_xlt(void *xlt, size_t idx, size_t nentries,
 			   struct mlx5_ib_mr *mr, int flags);
 
@@ -1357,7 +1358,7 @@ static inline int mlx5r_odp_create_eq(struct mlx5_ib_dev *dev,
 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) {}
+static inline void mlx5_odp_init_mkey_cache_entry(struct mlx5_cache_ent *ent) {}
 static inline void mlx5_odp_populate_xlt(void *xlt, size_t idx, size_t nentries,
 					 struct mlx5_ib_mr *mr, int flags) {}
 
diff --git a/drivers/infiniband/hw/mlx5/mr.c b/drivers/infiniband/hw/mlx5/mr.c
index aedfd7ff4846..129d531bd01b 100644
--- a/drivers/infiniband/hw/mlx5/mr.c
+++ b/drivers/infiniband/hw/mlx5/mr.c
@@ -82,15 +82,14 @@ static void set_mkc_access_pd_addr_fields(void *mkc, int acc, u64 start_addr,
 	MLX5_SET64(mkc, mkc, start_addr, start_addr);
 }
 
-static void assign_mkey_variant(struct mlx5_ib_dev *dev,
-				struct mlx5_ib_mkey *mkey, u32 *in)
+static void assign_mkey_variant(struct mlx5_ib_dev *dev, u32 *mkey, u32 *in)
 {
 	u8 key = atomic_inc_return(&dev->mkey_var);
 	void *mkc;
 
 	mkc = MLX5_ADDR_OF(create_mkey_in, in, memory_key_mkey_entry);
 	MLX5_SET(mkc, mkc, mkey_7_0, key);
-	mkey->key = key;
+	*mkey = key;
 }
 
 static int mlx5_ib_create_mkey(struct mlx5_ib_dev *dev,
@@ -98,7 +97,7 @@ static int mlx5_ib_create_mkey(struct mlx5_ib_dev *dev,
 {
 	int ret;
 
-	assign_mkey_variant(dev, mkey, in);
+	assign_mkey_variant(dev, &mkey->key, in);
 	ret = mlx5_core_create_mkey(dev->mdev, &mkey->key, in, inlen);
 	if (!ret)
 		init_waitqueue_head(&mkey->wait);
@@ -106,20 +105,21 @@ static int mlx5_ib_create_mkey(struct mlx5_ib_dev *dev,
 	return ret;
 }
 
-static int
-mlx5_ib_create_mkey_cb(struct mlx5_ib_dev *dev,
-		       struct mlx5_ib_mkey *mkey,
-		       struct mlx5_async_ctx *async_ctx,
-		       u32 *in, int inlen, u32 *out, int outlen,
-		       struct mlx5_async_work *context)
+static int mlx5_ib_create_mkey_cb(struct mlx5r_async_create_mkey *async_create)
 {
-	MLX5_SET(create_mkey_in, in, opcode, MLX5_CMD_OP_CREATE_MKEY);
-	assign_mkey_variant(dev, mkey, in);
-	return mlx5_cmd_exec_cb(async_ctx, in, inlen, out, outlen,
-				create_mkey_callback, context);
+	struct mlx5_ib_dev *dev = async_create->ent->dev;
+	size_t inlen = MLX5_ST_SZ_BYTES(create_mkey_in);
+	size_t outlen = MLX5_ST_SZ_BYTES(create_mkey_out);
+
+	MLX5_SET(create_mkey_in, async_create->in, opcode,
+		 MLX5_CMD_OP_CREATE_MKEY);
+	assign_mkey_variant(dev, &async_create->mkey, async_create->in);
+	return mlx5_cmd_exec_cb(&dev->async_ctx, async_create->in, inlen,
+				async_create->out, outlen, create_mkey_callback,
+				&async_create->cb_work);
 }
 
-static int mr_cache_max_order(struct mlx5_ib_dev *dev);
+static int mkey_cache_max_order(struct mlx5_ib_dev *dev);
 static void queue_adjust_cache_locked(struct mlx5_cache_ent *ent);
 
 static int destroy_mkey(struct mlx5_ib_dev *dev, struct mlx5_ib_mr *mr)
@@ -142,40 +142,132 @@ static void create_mkey_warn(struct mlx5_ib_dev *dev, int status, void *out)
 	mlx5_cmd_out_err(dev->mdev, MLX5_CMD_OP_CREATE_MKEY, 0, out);
 }
 
+
+static int push_mkey(struct mlx5_cache_ent *ent, bool limit_pendings,
+		     void *to_store)
+{
+	XA_STATE(xas, &ent->mkeys, 0);
+	void *curr;
+
+	xa_lock_irq(&ent->mkeys);
+	if (limit_pendings &&
+	    (ent->reserved - ent->stored) > MAX_PENDING_REG_MR) {
+		xa_unlock_irq(&ent->mkeys);
+		return -EAGAIN;
+	}
+	while (1) {
+		/*
+		 * This is cmpxchg (NULL, XA_ZERO_ENTRY) however this version
+		 * doesn't transparently unlock. Instead we set the xas index to
+		 * the current value of reserved every iteration.
+		 */
+		xas_set(&xas, ent->reserved);
+		curr = xas_load(&xas);
+		if (!curr) {
+			if (to_store && ent->stored == ent->reserved)
+				xas_store(&xas, to_store);
+			else
+				xas_store(&xas, XA_ZERO_ENTRY);
+			if (xas_valid(&xas)) {
+				ent->reserved++;
+				if (to_store) {
+					if (ent->stored != ent->reserved)
+						__xa_store(&ent->mkeys,
+							   ent->stored,
+							   to_store,
+							   GFP_KERNEL);
+					ent->stored++;
+					queue_adjust_cache_locked(ent);
+					WRITE_ONCE(ent->dev->cache.last_add,
+						   jiffies);
+				}
+			}
+		}
+		xa_unlock_irq(&ent->mkeys);
+
+		/*
+		 * Notice xas_nomem() must always be called as it cleans
+		 * up any cached allocation.
+		 */
+		if (!xas_nomem(&xas, GFP_KERNEL))
+			break;
+		xa_lock_irq(&ent->mkeys);
+	}
+	if (xas_error(&xas))
+		return xas_error(&xas);
+	if (WARN_ON(curr))
+		return -EINVAL;
+	return 0;
+}
+
+static void undo_push_reserve_mkey(struct mlx5_cache_ent *ent)
+{
+	void *old;
+
+	ent->reserved--;
+	old = __xa_erase(&ent->mkeys, ent->reserved);
+	WARN_ON(old);
+}
+
+static void push_to_reserved(struct mlx5_cache_ent *ent, u32 mkey)
+{
+	void *old;
+
+	old = __xa_store(&ent->mkeys, ent->stored, xa_mk_value(mkey), 0);
+	WARN_ON(old);
+	ent->stored++;
+}
+
+static u32 pop_stored_mkey(struct mlx5_cache_ent *ent)
+{
+	void *old, *xa_mkey;
+
+	ent->stored--;
+	ent->reserved--;
+
+	if (ent->stored == ent->reserved) {
+		xa_mkey = __xa_erase(&ent->mkeys, ent->stored);
+		WARN_ON(!xa_mkey);
+		return (u32)xa_to_value(xa_mkey);
+	}
+
+	xa_mkey = __xa_store(&ent->mkeys, ent->stored, XA_ZERO_ENTRY,
+			     GFP_KERNEL);
+	WARN_ON(!xa_mkey || xa_is_err(xa_mkey));
+	old = __xa_erase(&ent->mkeys, ent->reserved);
+	WARN_ON(old);
+	return (u32)xa_to_value(xa_mkey);
+}
+
 static void create_mkey_callback(int status, struct mlx5_async_work *context)
 {
-	struct mlx5_ib_mr *mr =
-		container_of(context, struct mlx5_ib_mr, cb_work);
-	struct mlx5_cache_ent *ent = mr->cache_ent;
+	struct mlx5r_async_create_mkey *mkey_out =
+		container_of(context, struct mlx5r_async_create_mkey, cb_work);
+	struct mlx5_cache_ent *ent = mkey_out->ent;
 	struct mlx5_ib_dev *dev = ent->dev;
 	unsigned long flags;
 
 	if (status) {
-		create_mkey_warn(dev, status, mr->out);
-		kfree(mr);
-		spin_lock_irqsave(&ent->lock, flags);
-		ent->pending--;
+		create_mkey_warn(dev, status, mkey_out->out);
+		kfree(mkey_out);
+		xa_lock_irqsave(&ent->mkeys, flags);
+		undo_push_reserve_mkey(ent);
 		WRITE_ONCE(dev->fill_delay, 1);
-		spin_unlock_irqrestore(&ent->lock, flags);
+		xa_unlock_irqrestore(&ent->mkeys, flags);
 		mod_timer(&dev->delay_timer, jiffies + HZ);
 		return;
 	}
 
-	mr->mmkey.type = MLX5_MKEY_MR;
-	mr->mmkey.key |= mlx5_idx_to_mkey(
-		MLX5_GET(create_mkey_out, mr->out, mkey_index));
-	init_waitqueue_head(&mr->mmkey.wait);
-
+	mkey_out->mkey |= mlx5_idx_to_mkey(
+		MLX5_GET(create_mkey_out, mkey_out->out, mkey_index));
 	WRITE_ONCE(dev->cache.last_add, jiffies);
 
-	spin_lock_irqsave(&ent->lock, flags);
-	list_add_tail(&mr->list, &ent->head);
-	ent->available_mrs++;
-	ent->total_mrs++;
+	xa_lock_irqsave(&ent->mkeys, flags);
+	push_to_reserved(ent, mkey_out->mkey);
 	/* If we are doing fill_to_high_water then keep going. */
 	queue_adjust_cache_locked(ent);
-	ent->pending--;
-	spin_unlock_irqrestore(&ent->lock, flags);
+	xa_unlock_irqrestore(&ent->mkeys, flags);
+	kfree(mkey_out);
 }
 
 static int get_mkc_octo_size(unsigned int access_mode, unsigned int ndescs)
@@ -197,15 +289,8 @@ static int get_mkc_octo_size(unsigned int access_mode, unsigned int ndescs)
 	return ret;
 }
 
-static struct mlx5_ib_mr *alloc_cache_mr(struct mlx5_cache_ent *ent, void *mkc)
+static void set_cache_mkc(struct mlx5_cache_ent *ent, void *mkc)
 {
-	struct mlx5_ib_mr *mr;
-
-	mr = kzalloc(sizeof(*mr), GFP_KERNEL);
-	if (!mr)
-		return NULL;
-	mr->cache_ent = ent;
-
 	set_mkc_access_pd_addr_fields(mkc, 0, 0, ent->dev->umrc.pd);
 	MLX5_SET(mkc, mkc, free, 1);
 	MLX5_SET(mkc, mkc, umr_en, 1);
@@ -215,133 +300,106 @@ static struct mlx5_ib_mr *alloc_cache_mr(struct mlx5_cache_ent *ent, void *mkc)
 	MLX5_SET(mkc, mkc, translations_octword_size,
 		 get_mkc_octo_size(ent->access_mode, ent->ndescs));
 	MLX5_SET(mkc, mkc, log_page_size, ent->page);
-	return mr;
 }
 
 /* Asynchronously schedule new MRs to be populated in the cache. */
 static int add_keys(struct mlx5_cache_ent *ent, unsigned int num)
 {
-	size_t inlen = MLX5_ST_SZ_BYTES(create_mkey_in);
-	struct mlx5_ib_mr *mr;
+	struct mlx5r_async_create_mkey *async_create;
 	void *mkc;
-	u32 *in;
 	int err = 0;
 	int i;
 
-	in = kzalloc(inlen, GFP_KERNEL);
-	if (!in)
-		return -ENOMEM;
-
-	mkc = MLX5_ADDR_OF(create_mkey_in, in, memory_key_mkey_entry);
 	for (i = 0; i < num; i++) {
-		mr = alloc_cache_mr(ent, mkc);
-		if (!mr) {
-			err = -ENOMEM;
-			break;
-		}
-		spin_lock_irq(&ent->lock);
-		if (ent->pending >= MAX_PENDING_REG_MR) {
-			err = -EAGAIN;
-			spin_unlock_irq(&ent->lock);
-			kfree(mr);
-			break;
-		}
-		ent->pending++;
-		spin_unlock_irq(&ent->lock);
-		err = mlx5_ib_create_mkey_cb(ent->dev, &mr->mmkey,
-					     &ent->dev->async_ctx, in, inlen,
-					     mr->out, sizeof(mr->out),
-					     &mr->cb_work);
+		async_create = kzalloc(sizeof(struct mlx5r_async_create_mkey),
+				       GFP_KERNEL);
+		if (!async_create)
+			return -ENOMEM;
+		mkc = MLX5_ADDR_OF(create_mkey_in, async_create->in,
+				   memory_key_mkey_entry);
+		set_cache_mkc(ent, mkc);
+		async_create->ent = ent;
+
+		err = push_mkey(ent, true, NULL);
+		if (err)
+			goto free_async_create;
+
+		err = mlx5_ib_create_mkey_cb(async_create);
 		if (err) {
-			spin_lock_irq(&ent->lock);
-			ent->pending--;
-			spin_unlock_irq(&ent->lock);
 			mlx5_ib_warn(ent->dev, "create mkey failed %d\n", err);
-			kfree(mr);
-			break;
+			goto err_undo_reserve;
 		}
 	}
 
-	kfree(in);
+	return 0;
+
+err_undo_reserve:
+	xa_lock_irq(&ent->mkeys);
+	undo_push_reserve_mkey(ent);
+	xa_unlock_irq(&ent->mkeys);
+free_async_create:
+	kfree(async_create);
 	return err;
 }
 
 /* Synchronously create a MR in the cache */
-static struct mlx5_ib_mr *create_cache_mr(struct mlx5_cache_ent *ent)
+static int create_cache_mkey(struct mlx5_cache_ent *ent, u32 *mkey)
 {
 	size_t inlen = MLX5_ST_SZ_BYTES(create_mkey_in);
-	struct mlx5_ib_mr *mr;
 	void *mkc;
 	u32 *in;
 	int err;
 
 	in = kzalloc(inlen, GFP_KERNEL);
 	if (!in)
-		return ERR_PTR(-ENOMEM);
+		return -ENOMEM;
 	mkc = MLX5_ADDR_OF(create_mkey_in, in, memory_key_mkey_entry);
+	set_cache_mkc(ent, mkc);
 
-	mr = alloc_cache_mr(ent, mkc);
-	if (!mr) {
-		err = -ENOMEM;
-		goto free_in;
-	}
-
-	err = mlx5_core_create_mkey(ent->dev->mdev, &mr->mmkey.key, in, inlen);
+	err = mlx5_core_create_mkey(ent->dev->mdev, mkey, in, inlen);
 	if (err)
-		goto free_mr;
+		goto free_in;
 
-	init_waitqueue_head(&mr->mmkey.wait);
-	mr->mmkey.type = MLX5_MKEY_MR;
 	WRITE_ONCE(ent->dev->cache.last_add, jiffies);
-	spin_lock_irq(&ent->lock);
-	ent->total_mrs++;
-	spin_unlock_irq(&ent->lock);
-	kfree(in);
-	return mr;
-free_mr:
-	kfree(mr);
 free_in:
 	kfree(in);
-	return ERR_PTR(err);
+	return err;
 }
 
 static void remove_cache_mr_locked(struct mlx5_cache_ent *ent)
 {
-	struct mlx5_ib_mr *mr;
+	u32 mkey;
 
-	lockdep_assert_held(&ent->lock);
-	if (list_empty(&ent->head))
+	lockdep_assert_held(&ent->mkeys.xa_lock);
+	if (!ent->stored)
 		return;
-	mr = list_first_entry(&ent->head, struct mlx5_ib_mr, list);
-	list_del(&mr->list);
-	ent->available_mrs--;
-	ent->total_mrs--;
-	spin_unlock_irq(&ent->lock);
-	mlx5_core_destroy_mkey(ent->dev->mdev, mr->mmkey.key);
-	kfree(mr);
-	spin_lock_irq(&ent->lock);
+	mkey = pop_stored_mkey(ent);
+	xa_unlock_irq(&ent->mkeys);
+	mlx5_core_destroy_mkey(ent->dev->mdev, mkey);
+	xa_lock_irq(&ent->mkeys);
 }
 
 static int resize_available_mrs(struct mlx5_cache_ent *ent, unsigned int target,
 				bool limit_fill)
+	 __acquires(&ent->mkeys) __releases(&ent->mkeys)
 {
 	int err;
 
-	lockdep_assert_held(&ent->lock);
+	lockdep_assert_held(&ent->mkeys.xa_lock);
 
 	while (true) {
 		if (limit_fill)
 			target = ent->limit * 2;
-		if (target == ent->available_mrs + ent->pending)
+		if (target == ent->reserved)
 			return 0;
-		if (target > ent->available_mrs + ent->pending) {
-			u32 todo = target - (ent->available_mrs + ent->pending);
+		if (target > ent->reserved) {
+			u32 todo = target - ent->reserved;
 
-			spin_unlock_irq(&ent->lock);
+			xa_unlock_irq(&ent->mkeys);
 			err = add_keys(ent, todo);
 			if (err == -EAGAIN)
 				usleep_range(3000, 5000);
-			spin_lock_irq(&ent->lock);
+			xa_lock_irq(&ent->mkeys);
 			if (err) {
 				if (err != -EAGAIN)
 					return err;
@@ -366,15 +424,15 @@ static ssize_t size_write(struct file *filp, const char __user *buf,
 
 	/*
 	 * Target is the new value of total_mrs the user requests, however we
-	 * cannot free MRs that are in use. Compute the target value for
-	 * available_mrs.
+	 * cannot free MRs that are in use. Compute the target value for stored
+	 * mkeys.
 	 */
-	spin_lock_irq(&ent->lock);
-	if (target < ent->total_mrs - ent->available_mrs) {
+	xa_lock_irq(&ent->mkeys);
+	if (target < ent->in_use) {
 		err = -EINVAL;
 		goto err_unlock;
 	}
-	target = target - (ent->total_mrs - ent->available_mrs);
+	target = target - ent->in_use;
 	if (target < ent->limit || target > ent->limit*2) {
 		err = -EINVAL;
 		goto err_unlock;
@@ -382,12 +440,12 @@ static ssize_t size_write(struct file *filp, const char __user *buf,
 	err = resize_available_mrs(ent, target, false);
 	if (err)
 		goto err_unlock;
-	spin_unlock_irq(&ent->lock);
+	xa_unlock_irq(&ent->mkeys);
 
 	return count;
 
 err_unlock:
-	spin_unlock_irq(&ent->lock);
+	xa_unlock_irq(&ent->mkeys);
 	return err;
 }
 
@@ -398,7 +456,7 @@ static ssize_t size_read(struct file *filp, char __user *buf, size_t count,
 	char lbuf[20];
 	int err;
 
-	err = snprintf(lbuf, sizeof(lbuf), "%d\n", ent->total_mrs);
+	err = snprintf(lbuf, sizeof(lbuf), "%ld\n", ent->stored + ent->in_use);
 	if (err < 0)
 		return err;
 
@@ -427,10 +485,10 @@ static ssize_t limit_write(struct file *filp, const char __user *buf,
 	 * Upon set we immediately fill the cache to high water mark implied by
 	 * the limit.
 	 */
-	spin_lock_irq(&ent->lock);
+	xa_lock_irq(&ent->mkeys);
 	ent->limit = var;
 	err = resize_available_mrs(ent, 0, true);
-	spin_unlock_irq(&ent->lock);
+	xa_unlock_irq(&ent->mkeys);
 	if (err)
 		return err;
 	return count;
@@ -457,17 +515,17 @@ static const struct file_operations limit_fops = {
 	.read	= limit_read,
 };
 
-static bool someone_adding(struct mlx5_mr_cache *cache)
+static bool someone_adding(struct mlx5_mkey_cache *cache)
 {
 	unsigned int i;
 
-	for (i = 0; i < MAX_MR_CACHE_ENTRIES; i++) {
+	for (i = 0; i < MAX_MKEY_CACHE_ENTRIES; i++) {
 		struct mlx5_cache_ent *ent = &cache->ent[i];
 		bool ret;
 
-		spin_lock_irq(&ent->lock);
-		ret = ent->available_mrs < ent->limit;
-		spin_unlock_irq(&ent->lock);
+		xa_lock_irq(&ent->mkeys);
+		ret = ent->stored < ent->limit;
+		xa_unlock_irq(&ent->mkeys);
 		if (ret)
 			return true;
 	}
@@ -481,26 +539,26 @@ static bool someone_adding(struct mlx5_mr_cache *cache)
  */
 static void queue_adjust_cache_locked(struct mlx5_cache_ent *ent)
 {
-	lockdep_assert_held(&ent->lock);
+	lockdep_assert_held(&ent->mkeys.xa_lock);
 
 	if (ent->disabled || READ_ONCE(ent->dev->fill_delay))
 		return;
-	if (ent->available_mrs < ent->limit) {
+	if (ent->stored < ent->limit) {
 		ent->fill_to_high_water = true;
 		mod_delayed_work(ent->dev->cache.wq, &ent->dwork, 0);
 	} else if (ent->fill_to_high_water &&
-		   ent->available_mrs + ent->pending < 2 * ent->limit) {
+		   ent->reserved < 2 * ent->limit) {
 		/*
 		 * Once we start populating due to hitting a low water mark
 		 * continue until we pass the high water mark.
 		 */
 		mod_delayed_work(ent->dev->cache.wq, &ent->dwork, 0);
-	} else if (ent->available_mrs == 2 * ent->limit) {
+	} else if (ent->stored == 2 * ent->limit) {
 		ent->fill_to_high_water = false;
-	} else if (ent->available_mrs > 2 * ent->limit) {
+	} else if (ent->stored > 2 * ent->limit) {
 		/* Queue deletion of excess entries */
 		ent->fill_to_high_water = false;
-		if (ent->pending)
+		if (ent->stored != ent->reserved)
 			queue_delayed_work(ent->dev->cache.wq, &ent->dwork,
 					   msecs_to_jiffies(1000));
 		else
@@ -511,25 +569,24 @@ static void queue_adjust_cache_locked(struct mlx5_cache_ent *ent)
 static void __cache_work_func(struct mlx5_cache_ent *ent)
 {
 	struct mlx5_ib_dev *dev = ent->dev;
-	struct mlx5_mr_cache *cache = &dev->cache;
+	struct mlx5_mkey_cache *cache = &dev->cache;
 	int err;
 
-	spin_lock_irq(&ent->lock);
+	xa_lock_irq(&ent->mkeys);
 	if (ent->disabled)
 		goto out;
 
-	if (ent->fill_to_high_water &&
-	    ent->available_mrs + ent->pending < 2 * ent->limit &&
+	if (ent->fill_to_high_water && ent->reserved < 2 * ent->limit &&
 	    !READ_ONCE(dev->fill_delay)) {
-		spin_unlock_irq(&ent->lock);
+		xa_unlock_irq(&ent->mkeys);
 		err = add_keys(ent, 1);
-		spin_lock_irq(&ent->lock);
+		xa_lock_irq(&ent->mkeys);
 		if (ent->disabled)
 			goto out;
 		if (err) {
 			/*
-			 * EAGAIN only happens if pending is positive, so we
-			 * will be rescheduled from reg_mr_callback(). The only
+			 * EAGAIN only happens if there are pending MRs, so we
+			 * will be rescheduled when storing them. The only
 			 * failure path here is ENOMEM.
 			 */
 			if (err != -EAGAIN) {
@@ -541,7 +598,7 @@ static void __cache_work_func(struct mlx5_cache_ent *ent)
 						   msecs_to_jiffies(1000));
 			}
 		}
-	} else if (ent->available_mrs > 2 * ent->limit) {
+	} else if (ent->stored > 2 * ent->limit) {
 		bool need_delay;
 
 		/*
@@ -556,11 +613,11 @@ static void __cache_work_func(struct mlx5_cache_ent *ent)
 		 * the garbage collection work to try to run in next cycle, in
 		 * order to free CPU resources to other tasks.
 		 */
-		spin_unlock_irq(&ent->lock);
+		xa_unlock_irq(&ent->mkeys);
 		need_delay = need_resched() || someone_adding(cache) ||
 			     !time_after(jiffies,
 					 READ_ONCE(cache->last_add) + 300 * HZ);
-		spin_lock_irq(&ent->lock);
+		xa_lock_irq(&ent->mkeys);
 		if (ent->disabled)
 			goto out;
 		if (need_delay) {
@@ -571,7 +628,7 @@ static void __cache_work_func(struct mlx5_cache_ent *ent)
 		queue_adjust_cache_locked(ent);
 	}
 out:
-	spin_unlock_irq(&ent->lock);
+	xa_unlock_irq(&ent->mkeys);
 }
 
 static void delayed_cache_work_func(struct work_struct *work)
@@ -587,73 +644,59 @@ struct mlx5_ib_mr *mlx5_mr_cache_alloc(struct mlx5_ib_dev *dev,
 				       int access_flags)
 {
 	struct mlx5_ib_mr *mr;
+	int err;
 
-	/* Matches access in alloc_cache_mr() */
 	if (!mlx5r_umr_can_reconfig(dev, 0, access_flags))
 		return ERR_PTR(-EOPNOTSUPP);
 
-	spin_lock_irq(&ent->lock);
-	if (list_empty(&ent->head)) {
+	mr = kzalloc(sizeof(*mr), GFP_KERNEL);
+	if (!mr)
+		return ERR_PTR(-ENOMEM);
+
+	xa_lock_irq(&ent->mkeys);
+	ent->in_use++;
+
+	if (!ent->stored) {
 		queue_adjust_cache_locked(ent);
 		ent->miss++;
-		spin_unlock_irq(&ent->lock);
-		mr = create_cache_mr(ent);
-		if (IS_ERR(mr))
-			return mr;
+		xa_unlock_irq(&ent->mkeys);
+		err = create_cache_mkey(ent, &mr->mmkey.key);
+		if (err) {
+			xa_lock_irq(&ent->mkeys);
+			ent->in_use--;
+			xa_unlock_irq(&ent->mkeys);
+			kfree(mr);
+			return ERR_PTR(err);
+		}
 	} else {
-		mr = list_first_entry(&ent->head, struct mlx5_ib_mr, list);
-		list_del(&mr->list);
-		ent->available_mrs--;
+		mr->mmkey.key = pop_stored_mkey(ent);
 		queue_adjust_cache_locked(ent);
-		spin_unlock_irq(&ent->lock);
-
-		mlx5_clear_mr(mr);
+		xa_unlock_irq(&ent->mkeys);
 	}
+	mr->mmkey.cache_ent = ent;
+	mr->mmkey.type = MLX5_MKEY_MR;
+	init_waitqueue_head(&mr->mmkey.wait);
 	return mr;
 }
 
-static void mlx5_mr_cache_free(struct mlx5_ib_dev *dev, struct mlx5_ib_mr *mr)
-{
-	struct mlx5_cache_ent *ent = mr->cache_ent;
-
-	WRITE_ONCE(dev->cache.last_add, jiffies);
-	spin_lock_irq(&ent->lock);
-	list_add_tail(&mr->list, &ent->head);
-	ent->available_mrs++;
-	queue_adjust_cache_locked(ent);
-	spin_unlock_irq(&ent->lock);
-}
-
 static void clean_keys(struct mlx5_ib_dev *dev, int c)
 {
-	struct mlx5_mr_cache *cache = &dev->cache;
+	struct mlx5_mkey_cache *cache = &dev->cache;
 	struct mlx5_cache_ent *ent = &cache->ent[c];
-	struct mlx5_ib_mr *tmp_mr;
-	struct mlx5_ib_mr *mr;
-	LIST_HEAD(del_list);
+	u32 mkey;
 
 	cancel_delayed_work(&ent->dwork);
-	while (1) {
-		spin_lock_irq(&ent->lock);
-		if (list_empty(&ent->head)) {
-			spin_unlock_irq(&ent->lock);
-			break;
-		}
-		mr = list_first_entry(&ent->head, struct mlx5_ib_mr, list);
-		list_move(&mr->list, &del_list);
-		ent->available_mrs--;
-		ent->total_mrs--;
-		spin_unlock_irq(&ent->lock);
-		mlx5_core_destroy_mkey(dev->mdev, mr->mmkey.key);
-	}
-
-	list_for_each_entry_safe(mr, tmp_mr, &del_list, list) {
-		list_del(&mr->list);
-		kfree(mr);
+	xa_lock_irq(&ent->mkeys);
+	while (ent->stored) {
+		mkey = pop_stored_mkey(ent);
+		xa_unlock_irq(&ent->mkeys);
+		mlx5_core_destroy_mkey(dev->mdev, mkey);
+		xa_lock_irq(&ent->mkeys);
 	}
+	xa_unlock_irq(&ent->mkeys);
 }
 
-static void mlx5_mr_cache_debugfs_cleanup(struct mlx5_ib_dev *dev)
+static void mlx5_mkey_cache_debugfs_cleanup(struct mlx5_ib_dev *dev)
 {
 	if (!mlx5_debugfs_root || dev->is_rep)
 		return;
@@ -662,9 +705,9 @@ static void mlx5_mr_cache_debugfs_cleanup(struct mlx5_ib_dev *dev)
 	dev->cache.root = NULL;
 }
 
-static void mlx5_mr_cache_debugfs_init(struct mlx5_ib_dev *dev)
+static void mlx5_mkey_cache_debugfs_init(struct mlx5_ib_dev *dev)
 {
-	struct mlx5_mr_cache *cache = &dev->cache;
+	struct mlx5_mkey_cache *cache = &dev->cache;
 	struct mlx5_cache_ent *ent;
 	struct dentry *dir;
 	int i;
@@ -674,13 +717,13 @@ static void mlx5_mr_cache_debugfs_init(struct mlx5_ib_dev *dev)
 
 	cache->root = debugfs_create_dir("mr_cache", mlx5_debugfs_get_dev_root(dev->mdev));
 
-	for (i = 0; i < MAX_MR_CACHE_ENTRIES; i++) {
+	for (i = 0; i < MAX_MKEY_CACHE_ENTRIES; i++) {
 		ent = &cache->ent[i];
 		sprintf(ent->name, "%d", ent->order);
 		dir = debugfs_create_dir(ent->name, cache->root);
 		debugfs_create_file("size", 0600, dir, ent, &size_fops);
 		debugfs_create_file("limit", 0600, dir, ent, &limit_fops);
-		debugfs_create_u32("cur", 0400, dir, &ent->available_mrs);
+		debugfs_create_ulong("cur", 0400, dir, &ent->stored);
 		debugfs_create_u32("miss", 0600, dir, &ent->miss);
 	}
 }
@@ -692,9 +735,9 @@ static void delay_time_func(struct timer_list *t)
 	WRITE_ONCE(dev->fill_delay, 0);
 }
 
-int mlx5_mr_cache_init(struct mlx5_ib_dev *dev)
+int mlx5_mkey_cache_init(struct mlx5_ib_dev *dev)
 {
-	struct mlx5_mr_cache *cache = &dev->cache;
+	struct mlx5_mkey_cache *cache = &dev->cache;
 	struct mlx5_cache_ent *ent;
 	int i;
 
@@ -707,22 +750,21 @@ int mlx5_mr_cache_init(struct mlx5_ib_dev *dev)
 
 	mlx5_cmd_init_async_ctx(dev->mdev, &dev->async_ctx);
 	timer_setup(&dev->delay_timer, delay_time_func, 0);
-	for (i = 0; i < MAX_MR_CACHE_ENTRIES; i++) {
+	for (i = 0; i < MAX_MKEY_CACHE_ENTRIES; i++) {
 		ent = &cache->ent[i];
-		INIT_LIST_HEAD(&ent->head);
-		spin_lock_init(&ent->lock);
+		xa_init_flags(&ent->mkeys, XA_FLAGS_LOCK_IRQ);
 		ent->order = i + 2;
 		ent->dev = dev;
 		ent->limit = 0;
 
 		INIT_DELAYED_WORK(&ent->dwork, delayed_cache_work_func);
 
-		if (i > MR_CACHE_LAST_STD_ENTRY) {
-			mlx5_odp_init_mr_cache_entry(ent);
+		if (i > MKEY_CACHE_LAST_STD_ENTRY) {
+			mlx5_odp_init_mkey_cache_entry(ent);
 			continue;
 		}
 
-		if (ent->order > mr_cache_max_order(dev))
+		if (ent->order > mkey_cache_max_order(dev))
 			continue;
 
 		ent->page = PAGE_SHIFT;
@@ -734,36 +776,36 @@ int mlx5_mr_cache_init(struct mlx5_ib_dev *dev)
 			ent->limit = dev->mdev->profile.mr_cache[i].limit;
 		else
 			ent->limit = 0;
-		spin_lock_irq(&ent->lock);
+		xa_lock_irq(&ent->mkeys);
 		queue_adjust_cache_locked(ent);
-		spin_unlock_irq(&ent->lock);
+		xa_unlock_irq(&ent->mkeys);
 	}
 
-	mlx5_mr_cache_debugfs_init(dev);
+	mlx5_mkey_cache_debugfs_init(dev);
 
 	return 0;
 }
 
-int mlx5_mr_cache_cleanup(struct mlx5_ib_dev *dev)
+int mlx5_mkey_cache_cleanup(struct mlx5_ib_dev *dev)
 {
 	unsigned int i;
 
 	if (!dev->cache.wq)
 		return 0;
 
-	for (i = 0; i < MAX_MR_CACHE_ENTRIES; i++) {
+	for (i = 0; i < MAX_MKEY_CACHE_ENTRIES; i++) {
 		struct mlx5_cache_ent *ent = &dev->cache.ent[i];
 
-		spin_lock_irq(&ent->lock);
+		xa_lock_irq(&ent->mkeys);
 		ent->disabled = true;
-		spin_unlock_irq(&ent->lock);
+		xa_unlock_irq(&ent->mkeys);
 		cancel_delayed_work_sync(&ent->dwork);
 	}
 
-	mlx5_mr_cache_debugfs_cleanup(dev);
+	mlx5_mkey_cache_debugfs_cleanup(dev);
 	mlx5_cmd_cleanup_async_ctx(&dev->async_ctx);
 
-	for (i = 0; i < MAX_MR_CACHE_ENTRIES; i++)
+	for (i = 0; i < MAX_MKEY_CACHE_ENTRIES; i++)
 		clean_keys(dev, i);
 
 	destroy_workqueue(dev->cache.wq);
@@ -830,22 +872,22 @@ static int get_octo_len(u64 addr, u64 len, int page_shift)
 	return (npages + 1) / 2;
 }
 
-static int mr_cache_max_order(struct mlx5_ib_dev *dev)
+static int mkey_cache_max_order(struct mlx5_ib_dev *dev)
 {
 	if (MLX5_CAP_GEN(dev->mdev, umr_extended_translation_offset))
-		return MR_CACHE_LAST_STD_ENTRY + 2;
+		return MKEY_CACHE_LAST_STD_ENTRY + 2;
 	return MLX5_MAX_UMR_SHIFT;
 }
 
-static struct mlx5_cache_ent *mr_cache_ent_from_order(struct mlx5_ib_dev *dev,
-						      unsigned int order)
+static struct mlx5_cache_ent *mkey_cache_ent_from_order(struct mlx5_ib_dev *dev,
+							unsigned int order)
 {
-	struct mlx5_mr_cache *cache = &dev->cache;
+	struct mlx5_mkey_cache *cache = &dev->cache;
 
 	if (order < cache->ent[0].order)
 		return &cache->ent[0];
 	order = order - cache->ent[0].order;
-	if (order > MR_CACHE_LAST_STD_ENTRY)
+	if (order > MKEY_CACHE_LAST_STD_ENTRY)
 		return NULL;
 	return &cache->ent[order];
 }
@@ -888,7 +930,7 @@ static struct mlx5_ib_mr *alloc_cacheable_mr(struct ib_pd *pd,
 						     0, iova);
 	if (WARN_ON(!page_size))
 		return ERR_PTR(-EINVAL);
-	ent = mr_cache_ent_from_order(
+	ent = mkey_cache_ent_from_order(
 		dev, order_base_2(ib_umem_num_dma_blocks(umem, page_size)));
 	/*
 	 * Matches access in alloc_cache_mr(). If the MR can't come from the
@@ -1320,7 +1362,7 @@ static bool can_use_umr_rereg_pas(struct mlx5_ib_mr *mr,
 	struct mlx5_ib_dev *dev = to_mdev(mr->ibmr.device);
 
 	/* We only track the allocated sizes of MRs from the cache */
-	if (!mr->cache_ent)
+	if (!mr->mmkey.cache_ent)
 		return false;
 	if (!mlx5r_umr_can_load_pas(dev, new_umem->length))
 		return false;
@@ -1329,7 +1371,7 @@ static bool can_use_umr_rereg_pas(struct mlx5_ib_mr *mr,
 		mlx5_umem_find_best_pgsz(new_umem, mkc, log_page_size, 0, iova);
 	if (WARN_ON(!*page_size))
 		return false;
-	return (1ULL << mr->cache_ent->order) >=
+	return (1ULL << mr->mmkey.cache_ent->order) >=
 	       ib_umem_num_dma_blocks(new_umem, *page_size);
 }
 
@@ -1570,15 +1612,17 @@ int mlx5_ib_dereg_mr(struct ib_mr *ibmr, struct ib_udata *udata)
 	}
 
 	/* Stop DMA */
-	if (mr->cache_ent) {
-		if (mlx5r_umr_revoke_mr(mr)) {
-			spin_lock_irq(&mr->cache_ent->lock);
-			mr->cache_ent->total_mrs--;
-			spin_unlock_irq(&mr->cache_ent->lock);
-			mr->cache_ent = NULL;
-		}
+	if (mr->mmkey.cache_ent) {
+		xa_lock_irq(&mr->mmkey.cache_ent->mkeys);
+		mr->mmkey.cache_ent->in_use--;
+		xa_unlock_irq(&mr->mmkey.cache_ent->mkeys);
+
+		if (mlx5r_umr_revoke_mr(mr) ||
+		    push_mkey(mr->mmkey.cache_ent, false,
+			      xa_mk_value(mr->mmkey.key)))
+			mr->mmkey.cache_ent = NULL;
 	}
-	if (!mr->cache_ent) {
+	if (!mr->mmkey.cache_ent) {
 		rc = destroy_mkey(to_mdev(mr->ibmr.device), mr);
 		if (rc)
 			return rc;
@@ -1595,12 +1639,10 @@ int mlx5_ib_dereg_mr(struct ib_mr *ibmr, struct ib_udata *udata)
 			mlx5_ib_free_odp_mr(mr);
 	}
 
-	if (mr->cache_ent) {
-		mlx5_mr_cache_free(dev, mr);
-	} else {
+	if (!mr->mmkey.cache_ent)
 		mlx5_free_priv_descs(mr);
-		kfree(mr);
-	}
+
+	kfree(mr);
 	return 0;
 }
 
diff --git a/drivers/infiniband/hw/mlx5/odp.c b/drivers/infiniband/hw/mlx5/odp.c
index 84da5674e1ab..e305bf1dc6c2 100644
--- a/drivers/infiniband/hw/mlx5/odp.c
+++ b/drivers/infiniband/hw/mlx5/odp.c
@@ -1588,7 +1588,7 @@ mlx5_ib_odp_destroy_eq(struct mlx5_ib_dev *dev, struct mlx5_ib_pf_eq *eq)
 	return err;
 }
 
-void mlx5_odp_init_mr_cache_entry(struct mlx5_cache_ent *ent)
+void mlx5_odp_init_mkey_cache_entry(struct mlx5_cache_ent *ent)
 {
 	if (!(ent->dev->odp_caps.general_caps & IB_ODP_SUPPORT_IMPLICIT))
 		return;
diff --git a/drivers/infiniband/hw/mlx5/umr.c b/drivers/infiniband/hw/mlx5/umr.c
index 3a48364c0918..e00b94d1b1ea 100644
--- a/drivers/infiniband/hw/mlx5/umr.c
+++ b/drivers/infiniband/hw/mlx5/umr.c
@@ -176,6 +176,7 @@ int mlx5r_umr_resource_init(struct mlx5_ib_dev *dev)
 	dev->umrc.pd = pd;
 
 	sema_init(&dev->umrc.sem, MAX_UMR_WR);
+	mutex_init(&dev->umrc.lock);
 
 	return 0;
 
@@ -195,6 +196,31 @@ void mlx5r_umr_resource_cleanup(struct mlx5_ib_dev *dev)
 	ib_dealloc_pd(dev->umrc.pd);
 }
 
+static int mlx5r_umr_recover(struct mlx5_ib_dev *dev)
+{
+	struct umr_common *umrc = &dev->umrc;
+	struct ib_qp_attr attr;
+	int err;
+
+	attr.qp_state = IB_QPS_RESET;
+	err = ib_modify_qp(umrc->qp, &attr, IB_QP_STATE);
+	if (err) {
+		mlx5_ib_dbg(dev, "Couldn't modify UMR QP\n");
+		goto err;
+	}
+
+	err = mlx5r_umr_qp_rst2rts(dev, umrc->qp);
+	if (err)
+		goto err;
+
+	umrc->state = MLX5_UMR_STATE_ACTIVE;
+	return 0;
+
+err:
+	umrc->state = MLX5_UMR_STATE_ERR;
+	return err;
+}
+
 static int mlx5r_umr_post_send(struct ib_qp *ibqp, u32 mkey, struct ib_cqe *cqe,
 			       struct mlx5r_umr_wqe *wqe, bool with_data)
 {
@@ -231,7 +257,7 @@ static int mlx5r_umr_post_send(struct ib_qp *ibqp, u32 mkey, struct ib_cqe *cqe,
 
 	id.ib_cqe = cqe;
 	mlx5r_finish_wqe(qp, ctrl, seg, size, cur_edge, idx, id.wr_id, 0,
-			 MLX5_FENCE_MODE_NONE, MLX5_OPCODE_UMR);
+			 MLX5_FENCE_MODE_INITIATOR_SMALL, MLX5_OPCODE_UMR);
 
 	mlx5r_ring_db(qp, 1, ctrl);
 
@@ -270,17 +296,49 @@ static int mlx5r_umr_post_send_wait(struct mlx5_ib_dev *dev, u32 mkey,
 	mlx5r_umr_init_context(&umr_context);
 
 	down(&umrc->sem);
-	err = mlx5r_umr_post_send(umrc->qp, mkey, &umr_context.cqe, wqe,
-				  with_data);
-	if (err)
-		mlx5_ib_warn(dev, "UMR post send failed, err %d\n", err);
-	else {
-		wait_for_completion(&umr_context.done);
-		if (umr_context.status != IB_WC_SUCCESS) {
-			mlx5_ib_warn(dev, "reg umr failed (%u)\n",
-				     umr_context.status);
+	while (true) {
+		mutex_lock(&umrc->lock);
+		if (umrc->state == MLX5_UMR_STATE_ERR) {
+			mutex_unlock(&umrc->lock);
 			err = -EFAULT;
+			break;
+		}
+
+		if (umrc->state == MLX5_UMR_STATE_RECOVER) {
+			mutex_unlock(&umrc->lock);
+			usleep_range(3000, 5000);
+			continue;
+		}
+
+		err = mlx5r_umr_post_send(umrc->qp, mkey, &umr_context.cqe, wqe,
+					  with_data);
+		mutex_unlock(&umrc->lock);
+		if (err) {
+			mlx5_ib_warn(dev, "UMR post send failed, err %d\n",
+				     err);
+			break;
 		}
+
+		wait_for_completion(&umr_context.done);
+
+		if (umr_context.status == IB_WC_SUCCESS)
+			break;
+
+		if (umr_context.status == IB_WC_WR_FLUSH_ERR)
+			continue;
+
+		WARN_ON_ONCE(1);
+		mlx5_ib_warn(dev,
+			"reg umr failed (%u). Trying to recover and resubmit the flushed WQEs\n",
+			umr_context.status);
+		mutex_lock(&umrc->lock);
+		err = mlx5r_umr_recover(dev);
+		mutex_unlock(&umrc->lock);
+		if (err)
+			mlx5_ib_warn(dev, "couldn't recover UMR, err %d\n",
+				     err);
+		err = -EFAULT;
+		break;
 	}
 	up(&umrc->sem);
 	return err;