mm/secretmem: use refcount_t instead of atomic_t
authorJordy Zomer <jordy@jordyzomer.github.io>
Wed, 8 Sep 2021 02:56:18 +0000 (19:56 -0700)
committerLinus Torvalds <torvalds@linux-foundation.org>
Wed, 8 Sep 2021 18:50:24 +0000 (11:50 -0700)
commit110860541f443f950c1274f217a1a3e298670a33
treea4210a0c80f3793e0ac66806cc2e56ed6d9e2a86
parent41c961b9013ee9b6d0491f6926df546e37964b1f
mm/secretmem: use refcount_t instead of atomic_t

When a secret memory region is active, memfd_secret disables hibernation.
One of the goals is to keep the secret data from being written to
persistent-storage.

It accomplishes this by maintaining a reference count to
`secretmem_users`.  Once this reference is held your system can not be
hibernated due to the check in `hibernation_available()`.  However,
because `secretmem_users` is of type `atomic_t`, reference counter
overflows are possible.

As you can see there's an `atomic_inc` for each `memfd` that is opened in
the `memfd_secret` syscall.  If a local attacker succeeds to open 2^32
memfd's, the counter will wrap around to 0.  This implies that you may
hibernate again, even though there are still regions of this secret
memory, thereby bypassing the security check.

In an attempt to fix this I have used `refcount_t` instead of `atomic_t`
which prevents reference counter overflows.

Link: https://lkml.kernel.org/r/20210820043339.2151352-1-jordy@pwning.systems
Signed-off-by: Jordy Zomer <jordy@pwning.systems>
Cc: Kees Cook <keescook@chromium.org>,
Cc: Jordy Zomer <jordy@jordyzomer.github.io>
Cc: James Bottomley <James.Bottomley@HansenPartnership.com>
Cc: Mike Rapoport <rppt@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
mm/secretmem.c