Seth Forshee [off-list ref] writes:

Both of these filesystems already have use cases for mounting the
same super block from multiple user namespaces. For sysfs this
happens when using criu for snapshotting a container, where sysfs
is mnounted in the containers network ns but the hosts user ns.
The cgroup filesystem shares the same super block for all mounts
of the same hierarchy regardless of the namespace.

As a result, the restriction on mounting a super block from a
single user namespace creates regressions for existing uses of
these filesystems. For these specific filesystems this
restriction isn't really necessary since the backing store is
objects in kernel memory and thus the ids assigned from inodes
is not subject to translation relative to s_user_ns.

Add a new filesystem flag, FS_USERNS_SHARE_SB, which when set
causes sget_userns() to skip the check of s_user_ns. Set this
flag for the sysfs and cgroup filesystems to fix the
regressions.

So this one needs to be sget_userns(..., &init_user_ns, ...).
And not a new special case.

Apologies for not catching this earlier.

I am looking at folding all of this into the patch that introduces
sget_userns so that even bisects won't have regresssions.

We loose the ability to call mount -o remount and actually affect
these filesystems (which we don't have without s_user_ns) but we gain a
whole lot of simplicity, and we don't break the xattr and security label
on sysfs code.

Eric