Cong Wang, a kernel developer at Multikernel, introduced a new Linux filesystem called DAXFS on the Linux kernel mailing lists. As the name suggests, DAXFS is built on the Linux kernel’s Direct Access (DAX) infrastructure, allowing filesystems to access shared physical memory directly without going through the traditional block I/O stack.
According to the announcement, unlike the well-known RAMFS or TMPFS, which rely on the page cache and allocate memory on a per-instance basis, DAXFS maps contiguous physical memory regions directly into the filesystem. File reads resolve to direct memory loads, reducing memory overhead in environments where identical read-only data would otherwise be replicated in RAM.
The new filesystem is designed to support memory exported by GPUs, FPGAs, or CXL-attached devices via the dma-buf API, enabling read-only access to data stored directly in device-backed memory.
Intentionally minimal in scope, DAXFS uses a self-contained, read-only image format and avoids runtime allocation, complex metadata management, and device-handling logic, making it suitable for certain use cases.
These include sharing a common container or Docker base image across independent kernel instances, using shared memory rather than network transfers in CXL memory-pooling scenarios, reducing RAM usage by combining a shared DAXFS base image with OverlayFS for containers, and providing zero-copy access to static data stored in accelerator memory.
The code is publicly available on GitHub, as the author has emphasized that DAXFS is an experimental proposal. Any path toward upstream inclusion would depend on further discussion and review within the Linux kernel development community.
For more information, see the message on the kernel’s mailing list.
