Re: [PATCH v4] libselinux: use kernel status page by default

[Stephen Smalley <stephen.smalley.work@gmail.com>]

On Fri, Jul 24, 2020 at 3:43 PM Mike Palmiotto
<mike.palmiotto@crunchydata.com> wrote:
>
> On Fri, Jul 24, 2020 at 3:34 PM Mike Palmiotto
> <mike.palmiotto@crunchydata.com> wrote:
> >
> > On Fri, Jul 24, 2020 at 2:26 PM Stephen Smalley
> > <stephen.smalley.work@gmail.com> wrote:
> > >
> > > On Fri, Jul 24, 2020 at 12:23 PM Mike Palmiotto
> > > <mike.palmiotto@crunchydata.com> wrote:
> > > >
> > > > On Fri, Jul 24, 2020 at 12:09 PM Stephen Smalley
> > > > <stephen.smalley.work@gmail.com> wrote:
> > > > >
> > > > > On Fri, Jul 24, 2020 at 11:48 AM Stephen Smalley
> > > > > <stephen.smalley.work@gmail.com> wrote:
> > > > > >
> > > > > > On Fri, Jul 24, 2020 at 11:34 AM Stephen Smalley
> > > > > > <stephen.smalley.work@gmail.com> wrote:
> > > > > > >
> > > > > > > On Fri, Jul 24, 2020 at 10:14 AM Mike Palmiotto
> > > > > > > <mike.palmiotto@crunchydata.com> wrote:
> > > > > > > >
> > > > > > > > Commit bc2a8f418e3b ("libselinux: add selinux_status_* interfaces for
> > > > > > > > /selinux/status") introduced the sestatus mechanism, which allows for
> > > > > > > > mmap()'ing of the kernel status page as a replacement for avc_netlink.
> > > > > > > >
> > > > > > > > The mechanism was initially intended for use by userspace object
> > > > > > > > managers which were calculating access decisions within their
> > > > > > > > application and did not rely on the libselinux AVC implementation. In
> > > > > > > > order to properly make use of sestatus within avc_has_perm(), the status
> > > > > > > > mechanism needs to properly set avc internals during status events;
> > > > > > > > else, avc_enforcing is never updated upon sestatus changes.
> > > > > > > >
> > > > > > > > This commit introduces a new selinux_status_loop() function, which
> > > > > > > > replaces the default netlink-equivalent, avc_netlink_loop(). The
> > > > > > > > function watches the kernel status page until an error occurs, at which
> > > > > > > > point it will exit the thread. In the event that the status page cannot
> > > > > > > > be opened (on avc_open), the thread will continue to function as before
> > > > > > > > by using a fallback netlink socket.
> > > > > > > >
> > > > > > > > This allows us to replace the call to avc_netlink_open() in
> > > > > > > > avc_init_internal() with a call to selinux_status_open() and remove the
> > > > > > > > avc_netlink_check_nb() call from the critical code path in
> > > > > > > > avc_has_perm_noaudit(), as well as selinux_check_access().
> > > > > > > >
> > > > > > > > Userspace object managers that still need a netlink socket can call
> > > > > > > > avc_netlink_acquire_fd() to open open and/or obtain one.
> > > > > > > >
> > > > > > > > Update the manpage to reflect the new avc_netlink_acquire_fd()
> > > > > > > > functionality.
> > > > > > > >
> > > > > > > > Signed-off-by: Mike Palmiotto <mike.palmiotto@crunchydata.com>
> > > > > > > > ---
> > > > > > > > Testing:
> > > > > > > >   - dbus-daemon v1.12.8 on RHEL8.2
> > > > > > > >   - dbus-broker v22 on F32
> > > > > > >
> > > > > > > This looks good to me as far as the code is concerned.  However,
> > > > > > > installing the patched libselinux and rebooting, I notice that
> > > > > > > afterward I have dbus-daemon running on a Fedora rawhide instance and
> > > > > > > consuming nearly 100% CPU constantly.  I'm guessing it is sitting in
> > > > > > > the status loop. Not sure why there is a dbus-daemon instance running
> > > > > > > at all since dbus-broker seems to be the default in Fedora and
> > > > > > > systemctl shows dbus-daemon as disabled.  But if I revert to the stock
> > > > > > > libselinux, it stops hogging CPU.  Thoughts?
> > > > > >
> > > > > > Used gdb to attach to the separate thread and got a traceback before
> > > > > > and after the libselinux patch.
> > > > > > Sure enough, before it is performing a blocking poll() operation and
> > > > > > hence sleeping.  After it is spinning in
> > > > > > the status loop.
> > > > >
> > > > > So, options would appear to be:
> > > > > 1) Drop the usage of avc_using_threads altogether, i.e. even if the
> > > > > caller provided a thread callback, don't create another thread and
> > > > > just call selinux_status_updated() on every avc_has_perm_noaudit()
> > > > > unless avc_app_main_loop is set.  Rationale: dbus-daemon was only
> > > > > using threads to avoid the overhead of avc_netlink_check_nb() on every
> > > > > avc_has_perm_noaudit() call, and we've eliminated that via use of
> > > > > sestatus, hence we don't need to create a separate thread at all.
> > > > > -or-
> > > > > 2) If using threads, then create the netlink socket during avc_init
> > > > > and keep using the netlink loop for the thread.  This preserves the
> > > > > blocking behavior.
> > > > >
> > > > > #1 seems more optimal to me and gets rid of threading for dbus-daemon,
> > > > > which was something they didn't like anyway.
> > > >
> > > > Perhaps this is misguided, but it seems like avc_init is deprecated
> > > > and along with it the ability to even set a custom thread callback.
> > > > IOTW there does not appear to be a mechanism to set a thread callback
> > > > while using avc_open (only avc_init). Perhaps we can just get rid of
> > > > the default callback for avc_open and allow the (deprecated) avc_init
> > > > to continue using it as it does?
> > > >
> > > > Is this basically what you were proposing for #2? I think I'd be more
> > > > inclined to go with that approach, in case userspace object managers
> > > > are doing other things in their thread callback.
> > >
> > > I think that's the same as #2 if I understood you currently.  That's
> > > fine if you prefer it.
> > > So then programs using avc_init() with non-NULL thread callbacks
> > > (hence avc_using_threads == 1) will still create the netlink socket
> > > and start a thread running avc_netlink_loop().  And programs using
> > > avc_netlink_acquire_fd() will create the netlink socket if not already
> > > created and can use it however they want.  Everything else will move
> > > to using the status page.
> >
> > What do you think about moving avc_create_thread call (if
> > avc_using_threads is set) into avc_netlink_acquire_fd().
> >
> > That way, if the caller is using avc_init with a create_thread
> > callback, they can get their netlink socket and create the netlink
> > thread and everything will function as before. In theory, this would
> > also work for the sestatus netlink fallback.
>
> Alternatively, we could just move the thread creation into the
> sestatus fallback, since, as you pointed out, the only reason for
> creating a thread would be to avoid the avc_netlink_check_nb()
> overhead.

avc_netlink_acquire_fd() isn't called by dbus-daemon in its current
release used in Fedora/RHEL.
So adding it there won't help.  We could add it to
selinux_status_open().  Just need to make sure we don't call
avc_netlink_open() twice there (it is already called in the fallback
case) or make it safe to do so.