BeamFS RFC: a Reed-Solomon file system for Linux

Aurelien Desbrieres sent BeamFS, a new Linux file system, to kernel developers as an 18-patch request for comments (RFC) on October 19, 2026. It is built to repair damaged data on a single disk, not just detect it. That is a job Linux file systems usually leave to a second copy of the data.

What BeamFS does

The RFC cover letter describes BeamFS as "a block-device filesystem that corrects corrupted bytes in place, on a single device." It was posted to linux-fsdevel, the mailing list where kernel file system work is reviewed.

The repair comes from Reed-Solomon codes. These are error-correcting codes that add extra check bytes to data, so a reader can rebuild bytes that were flipped.

BeamFS stores every data block as sixteen interleaved RS(255,239) codewords. Each codeword holds 239 bytes of data plus check bytes, and can correct up to eight bad symbols. Every read decodes the block. If BeamFS cannot vouch for the result, the read fails with an error (-EIO) instead of returning wrong bytes.

The protection is not free. In the default layout, "288 bytes of every 4096 (7%) are overhead," the cover letter says.

Where it came from

BeamFS grew out of FTRFS, an earlier proposal from the same author that stalled earlier this year, Phoronix reports. LWN's write-up of the FTRFS RFC, sent on April 13, 2026, described it as "dependable storage in radiation-intensive environments." It was aimed at embedded Linux in space and other harsh settings.

Phoronix says FTRFS first targeted satellite on-board computers and radiotherapy rooms. BeamFS keeps that focus on one storage device in radiation-prone places. Its README calls it an "EM-resilient Linux filesystem," meaning one that survives bit errors caused by electromagnetic effects. It has its own on-disk format, though, so it is a new file system rather than an FTRFS revision.

The series is large: 27 files changed and 18,045 lines added. The author says the code was prepared with Claude, Anthropic's AI model, under the author's direction. Every patch records that in an Assisted-by tag, the label kernel maintainers recently agreed to drop.

What the tests show

The cover letter reports these results for BeamFS 0.1.26 on Linux 7.3-rc5:

Test	Result
xfstests auto group, x86-64	122 of the 122 tests that ran passed
xfstests auto group, aarch64	122 of the 123 tests that ran passed
Separate run, qemux86-64	128 of 128 passed
Separate run, qemuarm64	129 of 129 passed
Emulated single-bit errors	Never returned wrong data

In the error tests, BeamFS returned the correct file in every case but one, where it refused the read. Ext4 and ext3 returned silently corrupted data, the project's README says. Btrfs refused the read.

The author lists the gaps too. No performance was measured. At most two bad symbols reached any codeword, so the eight-symbol limit was never tested. BeamFS has no journal and no copy-on-write, and "an unclean shutdown is left to fsck.beamfs."

The README is blunter: version 0.1.26 has known defects that can lose data, weaken security or crash the kernel. It advises against storing irreplaceable data on it.

What this means for developers

BeamFS is a proposal, not a merged feature. The series does not even build on mainline Linux until a separate pending kernel change is merged. Here is what is worth doing with it:

Do not use it for real data yet. The author's own README warns about data loss and kernel crashes. Test it only in a virtual machine with throwaway files.
Know what it adds. File systems like btrfs can spot a damaged block, but fixing it usually needs a second copy of the data. BeamFS tries to fix the block from its own check bytes, on one device. That suits small embedded systems that cannot carry a mirror.
Watch the two open questions. The author asks whether correction inside a file system "for one small device" belongs in fs/ at all. The second question is: "Is a journal a precondition for further review, or can it follow?" The answers will decide whether BeamFS goes anywhere.
Follow the AI-assisted review. BeamFS lands days after MeshXFS, another file system written with AI help. How maintainers handle an 18,045-line AI-assisted series will show what bar such code has to clear.

This article was first published on Tech AI Wire.