Re: [kernel-hardening] Re: [PATCH v7 2/2] security: tty: make TIOCSTI ioctl require CAP_SYS_ADMIN

[Serge E. Hallyn]

Quoting Matt Brown (matt@xxxxxxxxx):
> On 6/2/17 11:36 AM, Serge E. Hallyn wrote:
> > Quoting Matt Brown (matt@xxxxxxxxx):
> >> On 6/1/17 5:24 PM, Alan Cox wrote:
> >>>> There's a difference between "bugs" and "security bugs". Letting
> >>>
> >>> Not really, it's merely a matter of severity of result. A non security
> >>> bug that hoses your hard disk is to anyone but security nutcases at
> >>> least as bad as a security hole.
> >>>
> >>>> security bugs continue to get exploited because we want to flush out
> >>>> bugs seems insensitive to the people getting attacked. I'd rather
> >>>> protect against a class of bug than have to endless fix each bug.
> >>>
> >>> The others are security bugs too to varying degree
> >>>
> >>>>> I'm not against doing something to protect the container folks, but that
> >>>>> something as with Android is a whitelist of ioctls. And if we need to do
> >>>>> this with a kernel hook lets do it properly.
> >>>>>
> >>>>> Remember the namespace of the tty on creation
> >>>>> If the magic security flag is set then
> >>>>>         Apply a whitelist to *any* tty ioctl call where the ns doesn't
> >>>>>                 match
> >>>>>
> >>>>> and we might as well just take the Android whitelist since they've kindly
> >>>>> built it for us all!
> >>>>>
> >>>>> In the tty layer it ends up being something around 10 lines of code and
> >>>>> some other file somewhere in security/ that's just a switch or similar
> >>>>> with the whitelisted ioctl codes in it.
> >>>>>
> >>>>> That (or a similar SELinux ruleset) would actually fix the problem.
> >>>>> SELinux would be better because it can also apply the rules when doing
> >>>>> things like su/sudo/...  
> >>>>
> >>>> Just to play devil's advocate, wouldn't such a system continue to not
> >>>> address your physical-console concerns? I wouldn't want to limit the
> >>>
> >>> It would for the cases that a whitelist and container check covers -
> >>> because the whitelist wouldn't allow you to do anything but boring stuff
> >>> on the tty. TIOCSTI is just one of a whole range of differently stupid
> >>> and annoying opportunities. Containers do not and should not be able to
> >>> set the keymap, change the video mode, use console selection, make funny
> >>> beepy noises, access video I/O registers and all the other stuff like
> >>> that. Nothing is going to break if we have a fairly conservative
> >>> whitelist.
> >>>
> >>>> protection to only containers (but it's a good start), since it
> >>>> wouldn't protect people not using containers that still have a
> >>>> privileged TTY attached badly somewhere.
> >>>
> >>> How are you going to magically fix the problem. I'm not opposed to fixing
> >>> the real problem but right now it appears to be a product of wishful
> >>> thinking not programming. What's the piece of security code that
> >>> magically discerns the fact you are running something untrusted at the
> >>> other end of your tty. SELinux can do it via labelling but I don't see
> >>> any generic automatic way for the kernel to magically work out when to
> >>> whitelist and when not to. If there is a better magic rule than
> >>> differing-namespace then provide the code.
> >>>
> >>> You can't just disable TIOCSTI, it has users deal with it. You can
> >>> get away with disabling it for namespace crossing I think but if you do
> >>> that you need to disable a pile of others.
> >>>
> >>> (If it breaks containers blocking TIOCSTI then we need to have a good
> >>> look at algorithms for deciding when to flush the input queue on exiting
> >>> a container or somesuch)
> >>>
> >>>> If you're talking about wholistic SELinux policy, sure, I could
> >>>> imagine a wholistic fix. But for the tons of people without a
> >>>> comprehensive SELinux policy, the proposed protection continues to
> >>>> make sense.
> >>>
> >>> No it doesn't. It's completely useless unless you actually bother to
> >>> address the other exploit opportunities.
> >>>
> >>> Right now the proposal is a hack to do 
> >>>
> >>> 	if (TIOCSTI && different_namespace && magic_flag)
> >>>
> >>
> >>
> >> This is not what my patch does. Mine is like:
> >>
> >> 	if (TIOCSTI && !ns_capable(tty->owner_user_ns, CAP_SYS_ADMIN) &&
> >> 		magic_flag)
> >>
> >> in other words:
> >> 	if (TIOCSTI && (different_owner_user_ns || !CAP_SYS_ADMIN) &&
> >> 		magic_flag)
> >>
> >> can you specify what you mean by different_namespace? which namespace?
> > 
> > I think you're focusing on the wrong thing.  Your capable check (apart
> > from the fact that I think I've been convinced CAP_SYS_ADMIN is wrong)
> > is fine.
> 

I'm cc:ing linux-api here because really we're designing an interesting
API.

> Can't we also have a sysctl that toggles if CAP_SYS_ADMIN is involved in
> this whitelist check?  Otherwise someone might leave things out of the
> whitelist just because they want to use those ioctls as a privileged
> process.

I'm not quite sure what you're asking for here.  Let me offer a precise
strawman design.  I'm sure there are problems with it, it's just a starting
point.

system-wide whitelist (for now 'may_push_chars') is full by default.

By default, nothing changes - you can use those on your own tty, need
CAP_SYS_ADMIN against init_user_ns otherwise.

Introduce a new CAP_TTY_PRIVILEGED.

When may_push_chars is removed from the whitelist, you lose the ability
to use TIOCSTI on a tty - even your own - if you do not have CAP_TTY_PRIVILEGED
against the tty's user_ns.

>  Also restricting a privileged user from ioctls with this
> whitelist approach is going to be pointless because, if the whitelist is
> configurable from userspace, they will just be able to modify the
> whitelist.
> 
> > 
> > The key point is to not only check for TIOCSTI, but instead check for
> > a whitelisted ioctl.
> > 
> > What would the whitelist look like?  Should configuing that be the way
> > that you enable/disable, instead of the sysctl in this patchset?  So
> > by default the whitelist includes all ioctls (no change), but things
> > like sandboxes/sudo/container-starts can clear out the whitelist?
> > 
> 
> I'm fine with moving this to an LSM that whitelists ioctls. I also want

Right -  what else would go into the whitelist?  may_mmap?

> to understand what a whitelist would like look and how you would
> configure it? Does a sysctl that is a list of allowed ioctls work? I
> don't want to just have a static whitelist that you can't change without
> recompiling your kernel.
> 
> just running a sysctl -a on a linux box shows me one thing that looks
> like a list: net.core.flow_limit_cpu_bitmap