Abstract: Direct assignment of I/O devices (Direct I/O) is the best performant I/O virtualization method. However, it requires the hypervisor to statically pin the entire guest memory, thereby hindering the efficiency of memory management. This problem can be fixed by presenting a virtual IOMMU (vIOMMU). Emulation of its DMA remapping capability carries sufficient information about guest DMA buffers, allowing the hypervisor to do fine-grained pinning of guest memory. However, established vIOMMUs are not widely used by commodity guests due to the emulation cost, thus cannot reliably eliminate static pinning in direct I/O. We propose and implement coIOMMU, a new vIOMMU architecture for efficient memory management with a cooperative DMA buffer tracking mechanism. The new mechanism provides a dedicated interface for hypervisor and guest to exchange DMA buffer information over a shared DMA tracking table (DTT), orthogonal to the costly DMA remapping interface. We also explore two techniques: smart pinning and lazy unpinning, to minimize the impact on the performance of direct I/O. Our evaluation results show that coIOMMU dramatically improves the efficiency of memory management in wide direct I/O usages with negligible cost. Moreover, the desired semantics of DMA remapping can be sustained when cooperative tracking is enabled alongside. Overall, we believe that coIOMMU can serve as a reliable solution for efficient memory management in direct I/O.