Kernel developer Cong Wang has introduced DAXFS, a new read-only filesystem designed for direct access to shared memory in Linux. Built on the Direct Access infrastructure, it bypasses traditional page caching to reduce memory overhead. The proposal aims to enable efficient sharing of data across kernels and devices.
Cong Wang, a kernel developer at Multikernel, announced DAXFS on the Linux kernel mailing lists. This filesystem leverages the Linux kernel's Direct Access (DAX) infrastructure, which allows direct access to shared physical memory without relying on the conventional block I/O stack.
Unlike established options such as RAMFS or TMPFS, which use the page cache and allocate memory per instance, DAXFS maps contiguous physical memory regions straight into the filesystem. As a result, file reads become direct memory loads, minimizing overhead in scenarios where read-only data might otherwise duplicate in RAM.
DAXFS supports memory from devices like GPUs, FPGAs, or those connected via CXL, using the dma-buf API for read-only access to device-backed data. Its design is deliberately simple: it employs a self-contained, read-only image format and skips runtime allocations, intricate metadata handling, and device-specific code.
Potential applications include sharing container or Docker base images across separate kernel instances via shared memory, avoiding network transfers in CXL memory-pooling setups, cutting RAM use by pairing DAXFS with OverlayFS for containers, and offering zero-copy access to static data in accelerator memory.
The code resides on GitHub as an experimental proposal. Wang noted that any integration into the mainline kernel would require community discussion and review. The announcement appeared on January 25, 2026.
