On Tue, 24 Aug 2021 at 16:06, Mike Rapoport <rppt@xxxxxxxxxx> wrote: > > On Thu, Aug 19, 2021 at 10:33:49PM -0700, Kees Cook wrote: > > On Fri, Aug 20, 2021 at 06:33:38AM +0200, Jordy Zomer wrote: > > > 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. > > > > It's an unlikely condition to hit given max-open-fds, etc, but there's > > no reason to leave this weakness. Changing this to refcount_t is easy > > and better than using atomic_t. > > > > Reviewed-by: Kees Cook <keescook@xxxxxxxxxxxx> > > > > > 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. > > > > IMO, this hibernation check is also buggy, since it looks to be > > vulnerable to ToCToU: processes aren't frozen when > > hibernation_available() checks secretmem_users(), so a process could add > > one and fill it before the process freezer stops it. > > > > And of course, there's still the ptrace hole[1], which is think is quite > > serious as it renders the entire defense moot. > > I thought about what can be done here and could not come up with anything > better that prevent PTRACE on a process with secretmem, but this seems to > me too much from usability vs security POV. > > Protecting against root is always hard and secretmem anyway does not > provide 100% guarantee by itself but rather makes an accidental data leak > or non-target attack much harder. > > To be effective it also presumes that other hardening features are turned > on by the system administrator on production systems, so it's not > unrealistic to rely on ptrace being disabled. Hi, The issue existed before this change, but I think refcount_inc needs to be done before fd_install. After fd_install finishes, the fd can be used by userspace and we can have secret data in memory before the refcount_inc. A straightforward mis-use where a user will predict the returned fd in another thread before the syscall returns and will use it to store secret data is somewhat dubious because such a user just shoots themself in the foot. But a more interesting mis-used would be to close the predicted fd and decrement the refcount before the corresponding refcount_inc, this way one can briefly drop the refcount to zero while there are other users of secretmem.
