On Fri, Jul 24, 2020 at 11:34 AM Stephen Smalley
<stephen.smalley.work@xxxxxxxxx> wrote:
>
> On Fri, Jul 24, 2020 at 10:14 AM Mike Palmiotto
> <mike.palmiotto@xxxxxxxxxxxxxxx> 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@xxxxxxxxxxxxxxx>
> > ---
> > 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.
