Re: BUG: bad usercopy in memdup_user

[Tobin C. Harding]

Adding Linus

On Tue, Dec 19, 2017 at 03:12:05PM +0100, Dmitry Vyukov wrote:
> On Tue, Dec 19, 2017 at 3:08 PM, Tetsuo Handa
> <penguin-kernel@xxxxxxxxxxxxxxxxxxx> wrote:
> > Dmitry Vyukov wrote:
> >> On Tue, Dec 19, 2017 at 2:22 PM, Matthew Wilcox <willy@xxxxxxxxxxxxx> wrote:
> >> >> > >> This BUG is reporting
> >> >> > >>
> >> >> > >> [   26.089789] usercopy: kernel memory overwrite attempt detected to 0000000022a5b430 (kmalloc-1024) (1024 bytes)
> >> >> > >>
> >> >> > >> line. But isn't 0000000022a5b430 strange for kmalloc(1024, GFP_KERNEL)ed kernel address?
> >> >> > >
> >> >> > > The address is hashed (see the %p threads for 4.15).
> >> >> >
> >> >> >
> >> >> > +Tobin, is there a way to disable hashing entirely? The only
> >> >> > designation of syzbot is providing crash reports to kernel developers
> >> >> > with as much info as possible. It's fine for it to leak whatever.
> >> >>
> >> >> We have new specifier %px to print addresses in hex if leaking info is
> >> >> not a worry.
> >> >
> >> > Could we have a way to know that the printed address is hashed and not just
> >> > a pointer getting completely scrogged?  Perhaps prefix it with ... a hash!
> >> > So this line would look like:
> >> >
> >> > [   26.089789] usercopy: kernel memory overwrite attempt detected to #0000000022a5b430 (kmalloc-1024) (1024 bytes)
> >> >
> >> > Or does that miss the point of hashing the address, so the attacker
> >> > thinks its a real address?
> >>
> >> If we do something with this, I would suggest that we just disable
> >> hashing. Any of the concerns that lead to hashed pointers are not
> >> applicable in this context, moreover they are harmful, cause confusion
> >> and make it harder to debug these bugs. That perfectly can be an
> >> opt-in CONFIG_DEBUG_INSECURE_BLA_BLA_BLA.
> >>
> > Why not a kernel command line option? Hashing by default.
> 
> 
> Would work for continuous testing systems too.
> I just thought that since it has security implications, a config would
> be more reliable. Say if a particular distribution builds kernel
> without this config, then there is no way to enable it on the fly,
> intentionally or not.

I wasn't the architect behind the hashing, I've cc'd Linus in the event
he wants to correct me. I believe that some of the benefit of hashing
was to shake things up and force people to think about this issue. If we
implement a method of disabling hashing (command-line parameter or
CONFIG_) at this stage then we risk loosing this benefit since one has
to assume that people will just take the easy option and disable
it. Though perhaps after things settle a bit we could implement this
without the risk?

thanks,
Tobin.