tree d5cbb92283322a1b1f3f5510a2cc73e5feff0444
parent b85ea95d086471afb4ad062012a4d73cd328fa86
author Al Viro <viro@zeniv.linux.org.uk> 1697494115 +0100
committer Al Viro <viro@zeniv.linux.org.uk> 1699998952 -0500

selinux: saner handling of policy reloads

On policy reload selinuxfs replaces two subdirectories (/booleans
and /class) with new variants.  Unfortunately, that's done with
serious abuses of directory locking.

1) lock_rename() should be done to parents, not to objects being
exchanged

2) there's a bunch of reasons why it should not be done for directories
that do not have a common ancestor; most of those do not apply to
selinuxfs, but even in the best case the proof is subtle and brittle.

3) failure halfway through the creation of /class will leak
names and values arrays.

4) use of d_genocide() is also rather brittle; it's probably not much of
a bug per se, but e.g. an overmount of /sys/fs/selinuxfs/classes/shm/index
with any regular file will end up with leaked mount on policy reload.
Sure, don't do it, but...

Let's stop messing with disconnected directories; just create
a temporary (/.swapover) with no permissions for anyone (on the
level of ->permission() returing -EPERM, no matter who's calling
it) and build the new /booleans and /class in there; then
lock_rename on root and that temporary directory and d_exchange()
old and new both for class and booleans.  Then unlock and use
simple_recursive_removal() to take the temporary out; it's much
more robust.

And instead of bothering with separate pathways for freeing
new (on failure halfway through) and old (on success) names/values,
do all freeing in one place.  With temporaries swapped with the
old ones when we are past all possible failures.

The only user-visible difference is that /.swapover shows up
(but isn't possible to open, look up into, etc.) for the
duration of policy reload.

Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>