Shrink filesystem

[hockdudu]

I've seen the documentation states that shrinking isn't supported¹, but I couldn't find any issue tracking this feature (is it planned?), so I'm creating this issue here.

Is this a limitation on bcachefs itself, or only on the userspace tools?

Footnotes

1. https://bcachefs-docs.readthedocs.io/en/latest/mgmt-fsresize.html ↩

[phedders]

+1 very much looking forward to this being possible. (Sign of a grown up FS for me, looking at you XFS... )

[matthiasgoergens]

The problem is that shrinking needs to move data around on disk, while extending just needs a change to the metadata.

From what I can tell, shrinking is definitely possible, it's just a lot of work to implement.

[JohnRTitor]

CC @koverstreet

Looking forward to see this feature available.

[koverstreet]

it's actually not that much work, move_data_phys() can move everything except journal, but there are still lots of projects that need doing and not enough engineers :)

…

On Mon, Oct 27, 2025, 3:12 AM Masum Reza ***@***.***> wrote: *JohnRTitor* left a comment (koverstreet/bcachefs-tools#268) <#268 (comment)> CC @koverstreet <https://github.com/koverstreet> Looking forward to see this feature available. — Reply to this email directly, view it on GitHub <#268 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAPGX3RAUD7RILCBVIMCYTD3ZXHWZAVCNFSM6AAAAACKJOVVL6VHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZTINBZHE4TEMRVGQ> . You are receiving this because you were mentioned.Message ID: ***@***.***>

[Komzpa]

Current related work:

• Tools-side bridge: device: let online resize delegate shrink to kernel #650 (device: let online resize delegate shrink to kernel). It is open, mergeStateStatus: CLEAN, and the full current CI matrix is green. It removes the tools-side precheck that rejected online shrinking, so bcachefs device resize can pass smaller target sizes to a mounted filesystem kernel that supports shrink. Offline shrink is still rejected by the bundled userspace implementation.
• Kernel-side implementation: bcachefs: implement online filesystem shrinking bcachefs#1073 is the current WIP for online shrinking; it is still open/dirty and has remaining kernel work before this issue can be considered implemented end-to-end.

So the tools command path is ready for kernel shrink support, but shrinking itself remains blocked on the kernel implementation.

[jullanggit]

Also here would be the in-kernel issue for the same topic: koverstreet/bcachefs#781

[Komzpa]

Tools-targeted WIP artifact is updated at #742.

Current status:

• koverstreet/bcachefs-tools#742 ports the online shrink implementation from koverstreet/bcachefs#1073 into bcachefs-tools/testing.
• koverstreet/bcachefs-tools#742 preserves the original Jul Lang / Codex commit stack instead of carrying the port as one squashed commit; tools-only adaptations are separate Komzpa commits.
• The latest tail commits fix shrink restart handling in discard retry and cached-tail invalidation paths. Earlier tail commits also let online smaller-size bcachefs device resize requests reach the resize worker and resize the tools-only buckets_nouse bitmap with the bucket arrays.
• It remains intentionally draft/WIP, because this is still experimental shrink code and should not be treated as safe for important filesystems yet.
• Current koverstreet/bcachefs-tools#742 head d432f6894649123cc1e6fde2157b967f2662094e is merge-clean with 17/17 visible GitHub CI checks passing. Local validation also passed git diff --check, BINDGEN=/home/kom/.cargo/bin/bindgen make -j32 bcachefs, ./target/release/bcachefs device resize --help, make -C /lib/modules/$(uname -r)/build M=$PWD/fs -j32 modules, and the full local shrink suite.
• Companion draft ktest coverage is koverstreet/ktest#85 at head 0bdec5d426122b740faefbda88c5d2fcbfe1f7e5, merge-clean with GitHub CI green. It also preserves the imported Jul Lang test commit stack, leaves out offline shrink with a separate online-scope adaptation commit because koverstreet/bcachefs-tools#742 still leaves offline shrink rejected, and adds a separate retarget-race fix.

So for this issue cluster, koverstreet/bcachefs-tools#742 is the current bcachefs-tools-side implementation artifact, koverstreet/ktest#85 is the current ktest proof artifact, and koverstreet/bcachefs#1073 remains the original/provenance implementation surface.