[BUG][PATCH] audit: audit_log_start running on auditd should not stop

[Eric Paris]

On Fri, 2013-10-25 at 10:36 +0900, Toshiyuki Okajima wrote:

> systemd                                    |auditd
> -------------------------------------------+-----------------------------------
> ...                                        |
> -> audit_receive                           |...
>    -> mutex_lock(&audit_cmd_mutex)         |-> audit_receive
>       ... -> audit_log_start               |   -> mutex_lock(&audit_cmd_mutex)
>              -> wait_for_auditd            |      // wait for systemd
>                 -> schedule_timeout(60*HZ) |

Ugggh, definitely a problem.  Adding a similar hack to systemd really
does not seem like an acceptable answer.  It seems to me that in
audit_receive_msg()

case AUDIT_USER:
case AUDIT_FIRST_USER_MSG ... AUDIT_LAST_USER_MSG:
case AUDIT_FIRST_USER_MSG2 ... AUDIT_LAST_USER_MSG2:

we do not need to hold the audit_cmd_mutex.  So a quick and dirty patch
should be to just drop the mutex there (and we need to verify there
aren't issues running the audit_filter_user() without the lock).  That
will take care of systemd and anything USING audit.  It still means that
you could race with something configuring audit and auditd shutting
down.  Seems like a good quick and dirty 'fix' while we work on a better
fix...

To take care of that I think maybe we could drop the cmd_mutex every
time we call audit_log_start.  That's not necessarily going to be
pretty.  Maybe make a new switch at the top of the function which knows
which operations we are going to have to allocate an audit_buffer.  Drop
the lock, allocate the buffer, then retake the lock to finish running
audit_receive_msg()....

Maybe that second option isn't so hard and we can go directly after that
instead of just dealing with userspace audit messages?

Thoughts?