The commit that added this check did so in a very strange way - first
security_locked_down() is called, its value stored into retval, and if
it's nonzero, then an additional check is made for (change_irq ||
change_port), and if this is true, the function returns. However, if
the goto exit branch is not taken, the code keeps the retval value and
continues executing the function. Then, depending on whether
uport->ops->verify_port is set, the retval value may or may not be reset
to zero and eventually the error value from security_locked_down() may
abort the function a few lines below.
I will go out on a limb and assume that this isn't the intended behavior
and that an error value from security_locked_down() was supposed to
abort the function only in case (change_irq || change_port) is true.
Note that security_locked_down() should be called last in any series of
checks, since the SELinux implementation of this hook will do a check
against the policy and generate an audit record in case of denial. If
the operation was to carry on after calling security_locked_down(), then
the SELinux denial record would be bogus.
See commit 59438b46471a ("security,lockdown,selinux: implement SELinux
lockdown") for how SELinux implements this hook.
Fixes: 794edf30ee6c ("lockdown: Lock down TIOCSSERIAL")
Signed-off-by: Ondrej Mosnacek <[email protected]>
---
drivers/tty/serial/serial_core.c | 8 +++++---
1 file changed, 5 insertions(+), 3 deletions(-)
diff --git a/drivers/tty/serial/serial_core.c b/drivers/tty/serial/serial_core.c
index ba31e97d3d96..d7d8e7dbda60 100644
--- a/drivers/tty/serial/serial_core.c
+++ b/drivers/tty/serial/serial_core.c
@@ -865,9 +865,11 @@ static int uart_set_info(struct tty_struct *tty, struct tty_port *port,
goto check_and_exit;
}
- retval = security_locked_down(LOCKDOWN_TIOCSSERIAL);
- if (retval && (change_irq || change_port))
- goto exit;
+ if (change_irq || change_port) {
+ retval = security_locked_down(LOCKDOWN_TIOCSSERIAL);
+ if (retval)
+ goto exit;
+ }
/*
* Ask the low level driver to verify the settings.
--
2.31.1
On Fri, May 07, 2021 at 01:57:19PM +0200, Ondrej Mosnacek wrote:
> The commit that added this check did so in a very strange way - first
> security_locked_down() is called, its value stored into retval, and if
> it's nonzero, then an additional check is made for (change_irq ||
> change_port), and if this is true, the function returns. However, if
> the goto exit branch is not taken, the code keeps the retval value and
> continues executing the function. Then, depending on whether
> uport->ops->verify_port is set, the retval value may or may not be reset
> to zero and eventually the error value from security_locked_down() may
> abort the function a few lines below.
>
> I will go out on a limb and assume that this isn't the intended behavior
> and that an error value from security_locked_down() was supposed to
> abort the function only in case (change_irq || change_port) is true.
Are you _sure_ about this?
Verification from the authors and users of this odd feature might be
good to have, as I am loath to change how this works without them
weighing in here.
thanks,
greg k-h
On Fri, May 7, 2021 at 2:27 PM Greg Kroah-Hartman
<[email protected]> wrote:
> On Fri, May 07, 2021 at 01:57:19PM +0200, Ondrej Mosnacek wrote:
> > The commit that added this check did so in a very strange way - first
> > security_locked_down() is called, its value stored into retval, and if
> > it's nonzero, then an additional check is made for (change_irq ||
> > change_port), and if this is true, the function returns. However, if
> > the goto exit branch is not taken, the code keeps the retval value and
> > continues executing the function. Then, depending on whether
> > uport->ops->verify_port is set, the retval value may or may not be reset
> > to zero and eventually the error value from security_locked_down() may
> > abort the function a few lines below.
> >
> > I will go out on a limb and assume that this isn't the intended behavior
> > and that an error value from security_locked_down() was supposed to
> > abort the function only in case (change_irq || change_port) is true.
>
> Are you _sure_ about this?
>
> Verification from the authors and users of this odd feature might be
> good to have, as I am loath to change how this works without them
> weighing in here.
I'm not completely sure and I'm with you on not merging this without
feedback from people involved in the original patch and/or whoever
understands the logic in said function.
--
Ondrej Mosnacek
Software Engineer, Linux Security - SELinux kernel
Red Hat, Inc.
On Fri, May 07, 2021 at 01:57:19PM +0200, Ondrej Mosnacek wrote:
> The commit that added this check did so in a very strange way - first
> security_locked_down() is called, its value stored into retval, and if
> it's nonzero, then an additional check is made for (change_irq ||
> change_port), and if this is true, the function returns. However, if
> the goto exit branch is not taken, the code keeps the retval value and
> continues executing the function. Then, depending on whether
> uport->ops->verify_port is set, the retval value may or may not be reset
> to zero and eventually the error value from security_locked_down() may
> abort the function a few lines below.
>
> I will go out on a limb and assume that this isn't the intended behavior
> and that an error value from security_locked_down() was supposed to
> abort the function only in case (change_irq || change_port) is true.
>
> Note that security_locked_down() should be called last in any series of
> checks, since the SELinux implementation of this hook will do a check
> against the policy and generate an audit record in case of denial. If
> the operation was to carry on after calling security_locked_down(), then
> the SELinux denial record would be bogus.
>
> See commit 59438b46471a ("security,lockdown,selinux: implement SELinux
> lockdown") for how SELinux implements this hook.
>
> Fixes: 794edf30ee6c ("lockdown: Lock down TIOCSSERIAL")
> Signed-off-by: Ondrej Mosnacek <[email protected]>
> ---
> drivers/tty/serial/serial_core.c | 8 +++++---
> 1 file changed, 5 insertions(+), 3 deletions(-)
>
> diff --git a/drivers/tty/serial/serial_core.c b/drivers/tty/serial/serial_core.c
> index ba31e97d3d96..d7d8e7dbda60 100644
> --- a/drivers/tty/serial/serial_core.c
> +++ b/drivers/tty/serial/serial_core.c
> @@ -865,9 +865,11 @@ static int uart_set_info(struct tty_struct *tty, struct tty_port *port,
> goto check_and_exit;
> }
>
> - retval = security_locked_down(LOCKDOWN_TIOCSSERIAL);
> - if (retval && (change_irq || change_port))
> - goto exit;
> + if (change_irq || change_port) {
> + retval = security_locked_down(LOCKDOWN_TIOCSSERIAL);
> + if (retval)
> + goto exit;
> + }
>
> /*
> * Ask the low level driver to verify the settings.
Oops. Yeah, good catch -- I missed the kind of weird handling of retval
in this function when I originally reviewed it.
I think the goals of just covering IRQ/IO port changes originate from here:
https://lore.kernel.org/lkml/[email protected]/
And I think the "Reported-by: Greg KH" originates from here:
https://lore.kernel.org/lkml/[email protected]/
So, yes, I think your fix is correct.
Acked-by: Kees Cook <[email protected]>
--
Kees Cook