In the realm of Few-Shot Named Entity Recognition (Few-shot NER), task-decomposition methods have garnered significant attention. These methods decompose the named entity recognition task into two stages: namely entity extraction and entity classification, and have achieved remarkable success. Nevertheless, these methods are hampered by overly complex model training process and high time consumption, and may also result in the loss of significant knowledge among two stages. In this study, we propose a brand-new method called Unified One-stage Few-shot Named Entity Recognition (UOS-FSNER). This method combines candidate entity extraction and entity classification into a unified training framework by a new integrated labeling strategy, aiming to break through the limitations of task-decomposition methods and preserve the performance advantages of the task-decomposition methods as much as possible. Compared with those task-decomposition methods, our method significantly simplifies the training process, reduces the difficulty of model training and optimization, shortens training time and mitigates the knowledge loss issue caused by task decomposition. Experiments have demonstrated that UOS-FSNER not only can effectively overcome the limitations of task-decomposition methods but also further enhancing performance.