On 10/27/2020 10:29 AM, Aleksandr Nogikh wrote:

(resending the previous message in a plain/text mode)

On Mon, Oct 26, 2020 at 7:20 PM Casey Schaufler [off-list ref] wrote:
[...]

quoted

quoted

-     int RC = IRC;                                           \
-     do {                                                    \
+     int RC = lsm_hooks_inject_fail();                       \
+     if (RC == 0) {                                                          \

Injecting the failure here will prevent the loaded LSM hooks from
being called.

In this RFC, fault injection was intentionally placed before the code that
invokes LSM hooks. The reasoning was that it would simultaneously check
how the kernel code reacts to LSM denials and the effect of fault injections
on LSM modules.

quoted

quoted

              struct security_hook_list *P;                   \
+             RC = IRC;                                                               \
                                                              \
              hlist_for_each_entry(P, &security_hook_heads.FUNC, list) { \
                      RC = P->hook.FUNC(__VA_ARGS__);         \
                      if (RC != 0)                            \
                              break;                          \
              }                                               \
-     } while (0);                                            \
+     }                                                       \

Injecting the failure here would allow the loaded LSM hooks to
be called. It shouldn't make a difference, but hooks with side-effects
are always possible. I don't have an issue either way.

quoted

      RC;                                                     \
 })

Should we expect LSM modules to properly handle the cases when their
hooks with side effects were not invoked (unlike the selinux crash that
is described in the cover letter)? From the source code it seems that a
failure/denial from one module prevents the execution of the subsequent
hooks, so this looks like a realistic scenario.

Yes. Security modules have to accept the possibility that something
ahead of them in the stack will fail. This may be a DAC check, a
capability check or another security module.

If that is not true in general and depends on the specific active modules,
then it probably makes sense to introduce an option to control whether to
inject faults at the beginning of call_int_hook() or after the hooks have
been invoked.

If you want to do that you could implement it as an LSM. You could place it
anywhere in the stack that way. Based on what I see with the BPF lsm that might
be more work than it is worth.