Enhancing dma-buf for User-Space Read and Write Operations

Introduction to dma-buf

The Linux kernel's dma-buf subsystem is a framework designed to facilitate the sharing of memory buffers between different drivers and devices. Its primary purpose is to enable efficient device-to-device I/O, particularly in scenarios where data must move between hardware components such as GPUs, network adapters, or storage controllers without unnecessary copying through CPU memory. By providing a unified mechanism for buffer management, dma-buf helps reduce overhead and improve performance in specialized workloads like graphics rendering, video processing, and high-speed networking.

Enhancing dma-buf for User-Space Read and Write Operations

Despite its utility, dma-buf has traditionally been limited in its support for certain I/O operations. Most notably, it has not natively accommodated read and write requests initiated from user space. This limitation has prompted ongoing efforts to extend the subsystem and make it more versatile for emerging use cases in storage and filesystem contexts.

Current Usage and Limitations

In its current form, dma-buf is typically employed for sharing memory that is already mapped into kernel or device address spaces. Drivers can export a buffer, and other drivers can import it, often relying on mmap() to provide user-space access to the shared memory. This works well for scenarios where the buffer is accessed in a random-access manner—such as frame buffers for graphics—but it falls short when the application requires streaming I/O, like reading from or writing to a file.

For user-space applications that need to perform read or write operations on a dma-buf, the current approach involves multiple layers of abstraction: copying data into a separate user-space buffer, performing the operation, and then copying back. This defeats the purpose of zero-copy sharing and introduces unnecessary latency. The lack of direct read/write support also complicates integration with standard file I/O APIs, making dma-buf less attractive for storage or network-oriented workloads.

Recognizing these gaps, the kernel community has been exploring ways to enhance dma-buf so that it can handle read and write operations more naturally, directly from user space.

The Session at the 2026 Linux Storage, Filesystem, Memory Management, and BPF Summit

At the 2026 Linux Storage, Filesystem, Memory Management, and BPF Summit (LSMMBPF), Pavel Begunkov, assisted by Kanchan Joshi, led a joint session spanning both the storage and memory-management tracks. The session aimed to address two key goals: improving efficiency in existing dma-buf usage and extending the subsystem to support user-space-initiated read and write operations.

The discussion covered several technical challenges and potential solutions. One of the primary topics was how to adapt the internal buffer management to allow asynchronous I/O operations that respect the lifetime and access patterns of shared buffers. Another focus was on the necessary API additions—both kernel-side and user-space—to enable seamless integration with modern I/O mechanisms like io_uring.

Begunkov and Joshi presented their ongoing work on a prototype that demonstrates how dma-buf can be used as a target for read and write operations. The approach leverages existing infrastructure while introducing minimal changes to the core subsystem, aiming to preserve backward compatibility and avoid breaking current users.

Proposed Improvements

The session outlined several concrete proposals:

Direct read/write operations: New dma_buf_read() and dma_buf_write() callbacks in the export operations structure, allowing drivers to implement custom I/O paths without requiring intermediate copies.
Integration with io_uring: Enabling the use of dma-buf file descriptors with io_uring for asynchronous read/write requests, thereby reducing context switches and improving throughput for high-performance workloads.
Memory registration handling: Addressing the need to pin or unpin buffer pages efficiently when operations cross device boundaries, while still allowing for dynamic reclamation.
User-space API extension: Implementing a new DMA_BUF_IOCTL interface that allows applications to submit read/write commands directly on a dma-buf fd, similar to how pwrite() works on regular files.

These improvements are designed to maintain the core advantages of dma-buf—zero-copy sharing and device-to-device efficiency—while adding the flexibility needed for general-purpose file I/O.

Implications for User-Space I/O

If implemented, these enhancements could significantly simplify the design of applications that require both device-shared buffers and standard file access. For example, a video transcoding pipeline could use a dma-buf as a shared input/output buffer, then directly read its content from a storage device or write it to a network socket without additional memory copies. Similarly, data analytics frameworks that rely on remote direct memory access (RDMA) could leverage dma-buf to offload I/O operations to hardware while keeping data in the same buffer pool.

From a performance standpoint, the elimination of extra copy steps reduces latency and CPU overhead, making dma-buf a compelling choice for high-throughput, low-latency systems. The integration with io_uring further boosts efficiency by allowing applications to submit batches of I/O requests without system call overhead.

However, the session also highlighted challenges: ensuring safe concurrent access, managing buffer lifetimes across device boundaries, and maintaining compatibility with existing dma-buf users like graphics drivers. The community is working on incremental steps, with the prototypes aiming for inclusion in a future kernel release.

Conclusion

The joint effort by Pavel Begunkov and Kanchan Joshi at the 2026 LSMMBPF Summit marks an important step toward making dma-buf a first-class citizen for read and write operations. By extending the subsystem to support user-space-initiated I/O, the Linux kernel can offer a more unified and efficient framework for sharing memory across devices and applications. As development continues, these changes promise to bridge the gap between specialized device buffers and general-purpose file I/O, unlocking new possibilities for high-performance computing and storage systems.

Tags: