binfmt_misc: don't let an 'F' entry pin its own instance

In the Linux kernel, the following vulnerability has been resolved:

binfmt_misc: don't let an 'F' entry pin its own instance

An entry registered with 'F' opens its interpreter at registration time
and holds that file until the entry is freed. Any entry nobody removes
by hand only gets closed once the binfmt_misc superblock is shut down.
If the interpreter lives on a mount that keeps that superblock alive the
two pin each other:

binfmt_misc sb -> inode -> entry -> interp_file -> vfsmount -> binfmt_misc sb

TL;DR the file is never closed. Once the mount namespace is gone there
is nothing left to unregister through either.

There are two ways to trigger this bug:

- Point the interpreter at the instance itself. Its files are regular
files owned by the mounter and both bm_get_inode() and
simple_fill_super() leave i_op at empty_iops. So notify_change() falls
back to simple_setattr() and chmod +x works. We never set SB_I_NOEXEC
and so open_exec() accepts it.

- Use the instance as an overlayfs lower layer. The overlay superblock
holds a clone_private_mount() of every layer until it is destroyed and
that clone is in no namespace. So umount_tree() never reaches it.

That's a DoS. And it isn't only the superblock that leaks. It pins the
user namespace it was mounted in, so every iteration permanently eats
one of the caller's user namespace charges.

So let's just do the sane thing. SB_I_NOEXEC makes open_exec() fail on
the instance's own files and s_stack_depth makes overlayfs reject the
layer before it ever takes a clone. That also covers the ecryptfs and
fuse passthrough variants. What 'F' promises is unchanged.

The stable tag is narrower than the Fixes tags on purpose. Before
sandboxed mounts this needed global root against the single instance
everyone shares, and the change doesn't apply to those trees anyway.

Note that SB_I_NODEV is implicitly raised for userns mounts but raise it
explicitly here as well.