1 files changed, 278 insertions, 236 deletions

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;
 }