Brain tumors are among the most lethal cancers worldwide. Accurate segmentation of normal brain tissues and malignant tumor regions from 3D magnetic resonance imaging (MRI) is crucial for clinical diagnosis and surgical planning. However, missing modalities are a common issue in clinical practice, leading to loss of tumor details and disruption of inter-modal correlations, which significantly degrades segmentation performance. This paper proposes a Mixed-Scale Gated Attention Multi-Modal Fusion Network for brain tumor segmentation. The network first extracts feature pyramids through an encoder and adaptively fuses effective information at different scales. It then randomly selects features from one specific modality and fuses them with averaged features from other modalities. Finally, segmentation results are obtained through gated mixed-and-enhanced attention fusion. All three fusion techniques involved are capable of handling missing modalities, thereby improving segmentation outcomes. We evaluated our method on the BraTS2020 dataset. Under various missing-modality scenarios, the average Dice scores for the whole tumor, tumor core, and enhancing tumor reached 0.90048, 0.85113, and 0.65280, respectively. Comparative experiments with several recent methods demonstrate that our approach outperforms multiple mainstream segmentation methods, validating its superiority in handling missing-modality scenarios.