Existing super-resolution (SR) models are typically trained on datasets captured under normal-light conditions. However, when dealing with images captured under low-light conditions, the degradation becomes more complex, and the difference in lighting conditions often leads to low-quality results when using standard SR models. Due to error accumulation, simply cascading low-light enhancement (LE) and SR algorithms may not result in satisfactory results. Therefore, in this paper, we tackle this issue by jointly considering SR and LE. We first propose a new dataset called DarkSR, which contains low-resolution (LR) RAW and sRGB images captured under the low-light conditions, along with the corresponding high-resolution (HR) sRGB images captured under normal-light conditions. Noticing the linear relationship between pixel values and scene radiance, as well as the high bit depth of RAW images, and considering the presence of ISP pipeline-related information in sRGB images, we introduce JSLNet, a dual-input network that effectively explores the complementary information from the low-light LR RAW and sRGB images. Extensive experiments demonstrate that compared to other state-of-the-art (SOTA) methods, our method achieves the best quality of results, while maintaining a relatively low computational burden. The code and dataset are available at  https://github.com/flyhu2/DarkSR .