RDMA/mlx5: Fix the flow of a miss in the allocation of a cache ODP MR

[ Upstream commit 2f0e60d ]

When an ODP MR cache entry is empty and trying to allocate it, increment
the ent->miss counter and call to queue_adjust_cache_locked() to verify
the entry is balanced.

Fixes: aad719d ("RDMA/mlx5: Allow MRs to be created in the cache synchronously")
Link: https://lore.kernel.org/r/09503e295276dcacc92cb1d8aef1ad0961c99dc1.1644947594.git.leonro@nvidia.com
Signed-off-by: Aharon Landau <aharonl@nvidia.com>
Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>
Signed-off-by: Sasha Levin <sashal@kernel.org>

drivers/infiniband/hw/mlx5/mr.c

@@ -585,6 +585,8 @@ struct mlx5_ib_mr *mlx5_mr_cache_alloc(struct mlx5_ib_dev *dev,
 	ent = &cache->ent[entry];
 	spin_lock_irq(&ent->lock);
 	if (list_empty(&ent->head)) {
+		queue_adjust_cache_locked(ent);
+		ent->miss++;
 		spin_unlock_irq(&ent->lock);
 		mr = create_cache_mr(ent);
 		if (IS_ERR(mr))