Re: [REGRESSION] Return change in 6.12.80+ with volatile mounting

On Wed, Apr 22, 2026 at 1:54 AM Samuel Karp <[email protected]> wrote:
>
> We can fix the test case for the "volatile" option and expect either
> "volatile" or "fsync=volatile". Existing use within containerd beyond
> the test code involves us producing the "volatile" option and passing
> it to the mount syscall (or _not_ passing it, depending on the
> situation). This seems like it should still work if "fsync=volatile"
> is an alias rather than a replacement.  However, containerd is not in
> full control of the mount options; as part of snapshot (layer)
> management external snapshotter implementations can provide mount
> options to containerd which containerd then passes to the kernel.  The
> protection in RemoveVolatileOption will break if a newer snapshotter
> produces "fsync=volatile" and provides it to an unpatched version of
> containerd.  I think the breakage scenario then becomes new kernel +
> new snapshotter (producing the new option) + old containerd.
>

Even if kernel always shows "volatile" in show mount options
the scenario you described will break containerd but that would be
a snapshotter regression (w.r.t. containerd) not a kernel regression.

As I explained before, overlayfs could add a new mount option
fsync=off which has similar "do not reuse this workdir" semantics.
I explained the correct fix to this situation.

Thanks,
Amir.